update ChangeLog

Change-Id: I60c273c650a305fe36564ccc5fb1c8d7ea18118f

.gitattributes

@@ -2,3 +2,4 @@
 .gitignore export-ignore
 .mailmap export-ignore
 *.pdf -text -diff
+*.ppm -text -diff

.gitignore

@@ -1,5 +1,6 @@
 *.l[ao]
 *.[ao]
+*.pc
 .deps
 .libs
 /aclocal.m4
@@ -14,12 +15,15 @@
 /libtool
 /ltmain.sh
 /missing
+/mkinstalldirs
 /stamp-h1
 Makefile
 Makefile.in
 examples/[cdv]webp
 examples/gif2webp
 examples/webpmux
+src/webp/config.h*
+src/webp/stamp-h1
 /output
 /doc/output
 *.idb

.mailmap

@@ -5,3 +5,4 @@ Pascal Massimino <pascal.massimino@gmail.com>
 Vikas Arora <vikasa@google.com>
 <vikasa@google.com> <vikasa@gmail.com>
 <vikasa@google.com> <vikaas.arora@gmail.com>
+<slobodan.prijic@imgtec.com> <Slobodan.Prijic@imgtec.com>

AUTHORS

@@ -1,18 +1,25 @@
 Contributors:
 - Charles Munger (clm at google dot com)
 - Christian Duvivier (cduvivier at google dot com)
+- Djordje Pesut (djordje dot pesut at imgtec dot com)
 - James Zern (jzern at google dot com)
 - Jan Engelhardt (jengelh at medozas dot de)
 - Johann (johann dot koenig at duck dot com)
+- Jovan Zelincevic (jovan dot zelincevic at imgtec dot com)
 - Jyrki Alakuijala (jyrki at google dot com)
+- levytamar82 (tamar dot levy at intel dot com)
 - Lou Quillio (louquillio at google dot com)
 - Mans Rullgard (mans at mansr dot com)
 - Martin Olsson (mnemo at minimum dot se)
 - Mikołaj Zalewski (mikolajz at google dot com)
 - Noel Chromium (noel at chromium dot org)
 - Pascal Massimino (pascal dot massimino at gmail dot com)
+- Paweł Hajdan, Jr (phajdan dot jr at chromium dot org)
 - Pierre Joye (pierre dot php at gmail dot com)
 - Scott LaVarnway (slavarnway at google dot com)
+- Scott Talbot (s at chikachow dot org)
+- Slobodan Prijic (slobodan dot prijic at imgtec dot com)
 - Somnath Banerjee (somnath dot banerjee at gmail dot com)
+- Timothy Gu (timothygu99 at gmail dot com)
 - Urvang Joshi (urvang at google dot com)
 - Vikas Arora (vikasa at google dot com)

Android.mk

@@ -3,33 +3,65 @@ LOCAL_PATH := $(call my-dir)
 WEBP_CFLAGS := -Wall -DANDROID -DHAVE_MALLOC_H -DHAVE_PTHREAD -DWEBP_USE_THREAD
 
 ifeq ($(APP_OPTIM),release)
-  WEBP_CFLAGS += -finline-functions -frename-registers -ffast-math -s \
+  WEBP_CFLAGS += -finline-functions -ffast-math \
                  -ffunction-sections -fdata-sections
+  ifeq ($(findstring clang,$(NDK_TOOLCHAIN_VERSION)),)
+    WEBP_CFLAGS += -frename-registers -s
+  endif
 endif
 
-include $(CLEAR_VARS)
+ifneq ($(findstring armeabi-v7a, $(TARGET_ARCH_ABI)),)
+  # Setting LOCAL_ARM_NEON will enable -mfpu=neon which may cause illegal
+  # instructions to be generated for armv7a code. Instead target the neon code
+  # specifically.
+  NEON := c.neon
+else
+  NEON := c
+endif
 
-LOCAL_SRC_FILES := \
+dec_srcs := \
     src/dec/alpha.c \
     src/dec/buffer.c \
     src/dec/frame.c \
     src/dec/idec.c \
     src/dec/io.c \
-    src/dec/layer.c \
     src/dec/quant.c \
     src/dec/tree.c \
     src/dec/vp8.c \
     src/dec/vp8l.c \
     src/dec/webp.c \
+
+demux_srcs := \
+    src/demux/demux.c \
+
+dsp_dec_srcs := \
+    src/dsp/alpha_processing.c \
+    src/dsp/alpha_processing_sse2.c \
     src/dsp/cpu.c \
     src/dsp/dec.c \
+    src/dsp/dec_clip_tables.c \
+    src/dsp/dec_mips32.c \
+    src/dsp/dec_neon.$(NEON) \
     src/dsp/dec_sse2.c \
-    src/dsp/enc.c \
-    src/dsp/enc_sse2.c \
     src/dsp/lossless.c \
+    src/dsp/lossless_mips32.c \
+    src/dsp/lossless_neon.$(NEON) \
+    src/dsp/lossless_sse2.c \
     src/dsp/upsampling.c \
+    src/dsp/upsampling_neon.$(NEON) \
     src/dsp/upsampling_sse2.c \
     src/dsp/yuv.c \
+    src/dsp/yuv_mips32.c \
+    src/dsp/yuv_sse2.c \
+
+dsp_enc_srcs := \
+    src/dsp/enc.c \
+    src/dsp/enc_avx2.c \
+    src/dsp/enc_mips32.c \
+    src/dsp/enc_neon.$(NEON) \
+    src/dsp/enc_sse2.c \
+
+enc_srcs := \
     src/enc/alpha.c \
     src/enc/analysis.c \
     src/enc/backward_references.c \
@@ -39,60 +71,145 @@ LOCAL_SRC_FILES := \
     src/enc/frame.c \
     src/enc/histogram.c \
     src/enc/iterator.c \
-    src/enc/layer.c \
     src/enc/picture.c \
+    src/enc/picture_csp.c \
+    src/enc/picture_psnr.c \
+    src/enc/picture_rescale.c \
+    src/enc/picture_tools.c \
     src/enc/quant.c \
     src/enc/syntax.c \
     src/enc/token.c \
     src/enc/tree.c \
     src/enc/vp8l.c \
     src/enc/webpenc.c \
-    src/utils/alpha_processing.c \
+
+mux_srcs := \
+    src/mux/muxedit.c \
+    src/mux/muxinternal.c \
+    src/mux/muxread.c \
+
+utils_dec_srcs := \
     src/utils/bit_reader.c \
-    src/utils/bit_writer.c \
     src/utils/color_cache.c \
     src/utils/filters.c \
     src/utils/huffman.c \
-    src/utils/huffman_encode.c \
-    src/utils/quant_levels.c \
     src/utils/quant_levels_dec.c \
     src/utils/random.c \
     src/utils/rescaler.c \
     src/utils/thread.c \
     src/utils/utils.c \
 
+utils_enc_srcs := \
+    src/utils/bit_writer.c \
+    src/utils/huffman_encode.c \
+    src/utils/quant_levels.c \
+
+################################################################################
+# libwebpdecoder
+
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := \
+    $(dec_srcs) \
+    $(dsp_dec_srcs) \
+    $(utils_dec_srcs) \
+
 LOCAL_CFLAGS := $(WEBP_CFLAGS)
 LOCAL_C_INCLUDES += $(LOCAL_PATH)/src
 
 # prefer arm over thumb mode for performance gains
 LOCAL_ARM_MODE := arm
 
-ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
-  # Setting LOCAL_ARM_NEON will enable -mfpu=neon which may cause illegal
-  # instructions to be generated for armv7a code. Instead target the neon code
-  # specifically.
-  LOCAL_SRC_FILES += src/dsp/dec_neon.c.neon
-  LOCAL_SRC_FILES += src/dsp/upsampling_neon.c.neon
-  LOCAL_SRC_FILES += src/dsp/enc_neon.c.neon
-endif
 LOCAL_STATIC_LIBRARIES := cpufeatures
 
-LOCAL_MODULE := webp
+LOCAL_MODULE := webpdecoder_static
 
 include $(BUILD_STATIC_LIBRARY)
 
+ifeq ($(ENABLE_SHARED),1)
+include $(CLEAR_VARS)
+
+LOCAL_WHOLE_STATIC_LIBRARIES := webpdecoder_static
+
+LOCAL_MODULE := webpdecoder
+
+include $(BUILD_SHARED_LIBRARY)
+endif  # ENABLE_SHARED=1
+
+################################################################################
+# libwebp
+
 include $(CLEAR_VARS)
 
 LOCAL_SRC_FILES := \
-    examples/dwebp.c \
-    examples/example_util.c \
+    $(dsp_enc_srcs) \
+    $(enc_srcs) \
+    $(utils_enc_srcs) \
 
 LOCAL_CFLAGS := $(WEBP_CFLAGS)
-LOCAL_C_INCLUDES := $(LOCAL_PATH)/src
-LOCAL_STATIC_LIBRARIES := webp
+LOCAL_C_INCLUDES += $(LOCAL_PATH)/src
 
-LOCAL_MODULE := dwebp
+# prefer arm over thumb mode for performance gains
+LOCAL_ARM_MODE := arm
 
-include $(BUILD_EXECUTABLE)
+LOCAL_WHOLE_STATIC_LIBRARIES := webpdecoder_static
+
+LOCAL_MODULE := webp
+
+ifeq ($(ENABLE_SHARED),1)
+  include $(BUILD_SHARED_LIBRARY)
+else
+  include $(BUILD_STATIC_LIBRARY)
+endif
+
+################################################################################
+# libwebpdemux
+
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := $(demux_srcs)
+
+LOCAL_CFLAGS := $(WEBP_CFLAGS)
+LOCAL_C_INCLUDES += $(LOCAL_PATH)/src
+
+# prefer arm over thumb mode for performance gains
+LOCAL_ARM_MODE := arm
+
+LOCAL_MODULE := webpdemux
+
+ifeq ($(ENABLE_SHARED),1)
+  LOCAL_SHARED_LIBRARIES := webp
+  include $(BUILD_SHARED_LIBRARY)
+else
+  LOCAL_STATIC_LIBRARIES := webp
+  include $(BUILD_STATIC_LIBRARY)
+endif
+
+################################################################################
+# libwebpmux
+
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := $(mux_srcs)
+
+LOCAL_CFLAGS := $(WEBP_CFLAGS)
+LOCAL_C_INCLUDES += $(LOCAL_PATH)/src
+
+# prefer arm over thumb mode for performance gains
+LOCAL_ARM_MODE := arm
+
+LOCAL_MODULE := webpmux
+
+ifeq ($(ENABLE_SHARED),1)
+  LOCAL_SHARED_LIBRARIES := webp
+  include $(BUILD_SHARED_LIBRARY)
+else
+  LOCAL_STATIC_LIBRARIES := webp
+  include $(BUILD_STATIC_LIBRARY)
+endif
+
+################################################################################
+
+include $(LOCAL_PATH)/examples/Android.mk
 
 $(call import-module,android/cpufeatures)

ChangeLog

@@ -1,3 +1,486 @@
+569fe57 update NEWS
+bd852f5 bump version to 0.4.3
+2d58b64 WebPPictureRescale: add a note about 0 width/height
+a0d8ca5 examples/Android.mk: add webpmux_example target
+34b1d29 Android.mk: add webpmux target
+7561988 Android.mk: add webpdemux target
+a987576 Android.mk: add webpdecoder{,_static} targets
+a6d4859 Android.mk: split source lists per-directory
+77544d5 fix iOS arm64 build with Xcode 6.3
+6dea157 doc/webp-container-spec: note MSB order for chunk diagrams
+f7cd57b doc/webp-container-spec: cosmetics
+1d6b250 vwebp: clear canvas at the beginning of each loop
+f97b3f8 webp-container-spec: clarify background clear on loop
+4ba83c1 vwebp: remove unnecessary static Help() prototype
+d34e8e3 vwebp/animation: display last frame on end-of-loop
+bbbc524 dec/vp8: clear 'dither_' on skipped blocks
+0339fa2 lossless_neon: enable subtract green for aarch64
+5a0c220 Regression fix for lossless decoding
+6e3a31d wicdec: (msvs) quiet some /analyze warnings
+b49a578 dwebp/WritePNG: mark png variables volatile
+0a4391a dwebp: include setjmp.h w/WEBP_HAVE_PNG
+90f1ec5 dwebp: correct sign in format strings
+b61ce86 VP8LEncodeStream: add an assert
+df1081b dsp/cpu: (msvs) add include for __cpuidex
+39aa055 dsp/cpu: (msvs) avoid immintrin.h on _M_ARM
+f814f42 dsp/cpu: add include for _xgetbv() w/MSVS
+8508ab9 cpu: fix AVX2 detection for gcc/clang targets
+5769623 fix handling of zero-sized partition #0 corner case
+b2e71a9 make the 'last_cpuinfo_used' variable names unique
+1273e84 add -Wformat-nonliteral and -Wformat-security
+3ae78eb multi-thread fix: lock each entry points with a static var
+5c1eeda webp-container-spec: remove references to fragments
+c5ceea4 enc_neon: fix building with non-Xcode clang (iOS)
+d0859d6 iosbuild: add x64_64 simulator support
+046732c WebPEncode: Support encoding same pic twice (even if modified)
+4426f50 webp/types.h: use inline for clang++/-std=c++11
+e297fc7 gif2webp: Use the default hint instead of WEBP_HINT_GRAPH.
+855fe43 Makefile.vc: add a 'legacy' RTLIBCFG option
+b7eb6d5 gif2webp: Support GIF_DISPOSE_RESTORE_PREVIOUS
+5691bdd gif2webp: Handle frames with odd offsets + disposal to background.
+8301da1 stopwatch.h: fix includes
+6a2209a update ChangeLog (tag: v0.4.2, origin/0.4.2, 0.4.2)
+36cad6a bit_reader.h: cosmetics: fix a typo
+e2ecae6 enc_mips32: workaround gcc-4.9 bug
+243e68d update ChangeLog (tag: v0.4.2-rc2)
+eec5f5f enc/vp8enci.h: update version number
+0c1b98d update NEWS
+69b0fc9 update AUTHORS
+857578a bump version to 0.4.2
+9129deb restore encode API compatibility
+f17b95e AssignSegments: quiet -Warray-bounds warning
+9c56c8a enc_neon: initialize vectors w/vdup_n_u32
+a008902 iosbuild: cleanup
+cc6de53 iosbuild: output autoconf req. on failure
+740d765 iosbuild: make iOS 6 the minimum requirement
+403023f iobuild.sh: only install .h files in Headers
+b65727b Premultiply with alpha during U/V downsampling
+8de0deb gif2webp: Background color correction
+f8b7d94 Amend the lossless spec according to issue #205, #206 and #224
+9102a7b Add a WebPExtractAlpha function to dsp
+e407b5d webpmux: simplify InitializeConfig()
+3e70e64 webpmux: fix indent
+be38f1a webpmux: fix exit status on numeric value parse error
+94dadcb webpmux: fix loop_count range check
+40b3a61 examples: warn on invalid numeric parameters
+b7d209a gif2webp: Handle frames with missing  graphic control extension
+bf0eb74 configure: simplify libpng-config invocation
+3740f7d Rectify bug in lossless incremental decoding.
+3ab0a37 make VP8LSetBitPos() set br->eos_ flag
+2e4312b Lossless decoding: fix eos_ flag condition
+e6609ac fix erroneous dec->status_ setting
+5692eae add a fallback to ALPHA_NO_COMPRESSION
+6ecd5bf ExUtilReadFromStdin: (windows) open stdin in bin mode
+4206ac6 webpmux: (windows) open stdout in binary mode
+d40e885 cwebp: (windows) open stdout in binary mode
+4aaf463 example_util: add ExUtilSetBinaryMode
+4c82ff7 webpmux man page: Clarify some title, descriptions and examples
+23d4fb3 dsp/lossless: workaround gcc-4.9 bug on arm
+5af7719 dsp.h: collect gcc/clang version test macros
+90d1124 enc_neon: enable QuantizeBlock for aarch64
+ee78e78 SmartRGBYUV: fix odd-width problem with pixel replication
+c9ac204 fix some MSVC64 warning about float conversion
+f4497a1 cpu: check for _MSC_VER before using msvc inline asm
+e2159fd faster RGB->YUV conversion function (~7% speedup)
+21abaa0 Add smart RGB->YUV conversion option -pre 4
+1a161e2 configure: add work around for gcc-4.9 aarch64 bug
+55b10de MIPS: mips32r2: added optimization for BSwap32
+76d2192 Update PATENTS to reflect s/VP8/WebM/g
+29a9db1 MIPS: detect mips32r6 and disable mips32r1 code
+245c4a6 Correctly use the AC_CANONICAL_* macros
+40aa8b6 cosmetics
+2ddcca5 cosmetics: remove some extraneous 'extern's
+f40dd7c vp8enci.h: cosmetics: fix '*' placement
+4610c9c bit_writer: cosmetics: rename kFlush() -> Flush()
+fc3c175 dsp: detect mips64 & disable mips32 code
+c1a7955 cwebp.1: restore quality description
+57a7e73 correct alpha_dithering_strength ABI check
+6c83157 correct WebPMemoryWriterClear ABI check
+8af2771 update ChangeLog (tag: v0.4.1, origin/0.4.1, 0.4.1)
+f59c0b4 iosbuild.sh: specify optimization flags
+8d34ea3 update ChangeLog (tag: v0.4.1-rc1)
+dbc3da6 makefile.unix: add vwebp.1 to the dist target
+89a7c83 update ChangeLog
+ffe67ee Merge "update NEWS for the next release" into 0.4.1
+2def1fe gif2webp: dust up the help message
+fb668d7 remove -noalphadither option from README/vwebp.1
+e49f693 update NEWS for the next release
+cd01358 Merge "update AUTHORS" into 0.4.1
+268d01e update AUTHORS
+85213b9 bump version to 0.4.1
+695f80a Merge "restore mux API compatibility" into 0.4.1
+862d296 restore mux API compatibility
+8f6f8c5 remove the !WEBP_REFERENCE_IMPLEMENTATION tweak in Put8x8uv
+d713a69 Merge changes If4debc15,I437a5d5f into 0.4.1
+c2fc52e restore encode API compatibility
+793368e restore decode API compatibility
+b8984f3 gif2webp: fix compile with giflib 5.1.0
+222f9b1 gif2webp: simplify giflib version checking
+d2cc61b Extend MakeARGB32() to accept Alpha channel.
+4595b62 Merge "use explicit size of kErrorMessages[] arrays"
+157de01 Merge "Actuate memory stats for PRINT_MEMORY_INFO"
+fbda2f4 JPEG decoder: delay conversion to YUV to WebPEncode() call
+0b747b1 use explicit size of kErrorMessages[] arrays
+3398d81 Actuate memory stats for PRINT_MEMORY_INFO
+6f3202b Merge "move WebPPictureInit to picture.c"
+6c347bb move WebPPictureInit to picture.c
+fb3acf1 fix configure message for multi-thread
+40b086f configure: check for _beginthreadex
+1549d62 reorder the YUVA->ARGB and ARGB->YUVA functions correctly
+c6461bf Merge "extract colorspace code from picture.c into picture_csp.c"
+736f2a1 extract colorspace code from picture.c into picture_csp.c
+645daa0 Merge "configure: check for -Wformat-security"
+abafed8 configure: check for -Wformat-security
+fbadb48 split monolithic picture.c into picture_{tools,psnr,rescale}.c
+c76f07e dec_neon/TransformAC3: initialize vector w/vcreate
+bb4fc05 gif2webp: Allow single-frame animations
+46fd44c thread: remove harmless race on status_ in End()
+5a1a726 Merge "configure: check for __builtin_bswapXX()"
+6781423 configure: check for __builtin_bswapXX()
+6450c48 configure: fix iOS builds
+6422e68 VP8LFillBitWindow: enable fast path for 32-bit builds
+4f7f52b VP8LFillBitWindow: respect WEBP_FORCE_ALIGNED
+e458bad endian_inl.h: implement htoleXX with BSwapXX
+f2664d1 endian_inl.h: add BSwap16
+6fbf534 Merge "configure: add --enable-aligned"
+dc0f479 configure: add --enable-aligned
+9cc69e2 Merge "configure: support WIC + OpenGL under mingw64"
+257adfb remove experimental YUV444 YUV422 and YUV400 code
+10f4257 configure: support WIC + OpenGL under mingw64
+380cca4 configure.ac: add AC_C_BIGENDIAN
+ee70a90 endian_inl.h: add BSwap64
+47779d4 endian_inl.h: add BSwap32
+d5104b1 utils: add endian_inl.h
+58ab622 Merge "make alpha-detection loop in IsKeyFrame() in good x/y order"
+9d56290 make alpha-detection loop in IsKeyFrame() in good x/y order
+516971b lossless: Remove unaligned read warning
+b8b596f Merge "configure.ac: add an autoconf version prerequisite"
+34b02f8 configure.ac: add an autoconf version prerequisite
+e59f536 neon: normalize vdup_n_* usage
+6ee7160 Merge changes I0da7b3d3,Idad2f278,I4accc305
+abc02f2 Merge "fix (uncompiled) typo"
+bc03670 neon: add INIT_VECTOR4
+6c1c632 neon: add INIT_VECTOR3
+dc7687e neon: add INIT_VECTOR2
+4536e7c add WebPMuxSetCanvasSize() to the mux API
+824eab1 fix (uncompiled) typo
+1f3e5f1 remove unused 'shift' argument and QFIX2 define
+8e86705 Merge "VP8LoadNewBytes: use __builtin_bswap32 if available"
+1b6a263 Merge "Fix handling of weird GIF with canvas dimension 0x0"
+1da3d46 VP8LoadNewBytes: use __builtin_bswap32 if available
+1582e40 Fix handling of weird GIF with canvas dimension 0x0
+b8811da Merge "rename interface -> winterface"
+db8b8b5 Fix logic in the GIF LOOP-detection parsing
+25aaddc rename interface -> winterface
+5584d9d make WebPSetWorkerInterface() check its arguments
+a9ef7ef Merge "cosmetics: update thread.h comments"
+c6af999 Merge "dust up the help message"
+0a8b886 dust up the help message
+a9cf319 cosmetics: update thread.h comments
+27bfeee QuantizeBlock SSE2 Optimization:
+2bc0dc3 Merge "webpmux: warn when odd frame offsets are used"
+3114ebe Merge changes Id8edd3c1,Id418eb96,Ide05e3be
+c072663 webpmux: warn when odd frame offsets are used
+c5c6b40 Merge "add alpha dithering for lossy"
+d514678 examples/Android.mk: add cwebp
+ca0fa7c Android.mk: move dwebp to examples/Android.mk
+73d8fca Android.mk: add ENABLE_SHARED flag
+6e93317 muxread: fix out of bounds read
+8b0f6a4 Makefile.vc: fix CFLAGS assignment w/HAVE_AVX2=1
+bbe32df add alpha dithering for lossy
+7902076 Merge "make error-code reporting consistent upon malloc failure"
+77bf441 make error-code reporting consistent upon malloc failure
+7a93c00 **/Makefile.am: remove unused AM_CPPFLAGS
+24e3080 Add an interface abstraction to the WebP worker thread implementation
+d6cd635 Merge "fix orig_rect==NULL case"
+2bfd1ff fix orig_rect==NULL case
+059e21c Merge "configure: move config.h to src/webp/config.h"
+f05fe00 properly report back encoding error code in WebPFrameCacheAddFrame()
+32b3137 configure: move config.h to src/webp/config.h
+90090d9 Merge changes I7c675e51,I84f7d785
+ae7661b makefiles: define WEBP_HAVE_AVX2 when appropriate
+69fce2e remove the special casing for res->first in VP8SetResidualCoeffs
+6e61a3a configure: test for -msse2
+b9d2efc rename upsampling_mips32.c to yuv_mips32.c
+bdfeeba dsp/yuv: move sse2 functions to yuv_sse2.c
+46b32e8 Merge "configure: set WEBP_HAVE_AVX2 when available"
+88305db Merge "VP8RandomBits2: prevent signed int overflow"
+73fee88 VP8RandomBits2: prevent signed int overflow
+db4860b enc_sse2: prevent signed int overflow
+3fdaf4d Merge "real fix for longjmp warning"
+385e334 real fix for longjmp warning
+230a055 configure: set WEBP_HAVE_AVX2 when available
+a2ac8a4 restore original value_/range_ field order
+5e2ee56 Merge "remove libwebpdspdecode dep on libwebpdsp_avx2"
+61362db remove libwebpdspdecode dep on libwebpdsp_avx2
+42c447a Merge "lossy bit-reader clean-up:"
+479ffd8 Merge "remove unused #include's"
+9754d39 Merge "strong filtering speed-up (~2-3% x86, ~1-2% for NEON)"
+158aff9 remove unused #include's
+09545ee lossy bit-reader clean-up:
+ea8b0a1 strong filtering speed-up (~2-3% x86, ~1-2% for NEON)
+6679f89 Optimize VP8SetResidualCoeffs.
+ac591cf fix for gcc-4.9 warnings about longjmp + local variables
+4dfa86b dsp/cpu: NaCl has no support for xgetbv
+4c39869 Merge "cwebp: fallback to native webp decode in WIC builds"
+33aa497 Merge "cwebp: add some missing newlines in longhelp output"
+c9b340a fix missing WebPInitAlphaProcessing call for premultiplied colorspace output
+57897ba Merge "lossless_neon: use vcreate_*() where appropriate"
+6aa4777 Merge "(enc|dec)_neon: use vcreate_*() where appropriate"
+0d346e4 Always reinit VP8TransformWHT instead of hard-coding
+7d039fc cwebp: fallback to native webp decode in WIC builds
+d471f42 cwebp: add some missing newlines in longhelp output
+bf0e003 lossless_neon: use vcreate_*() where appropriate
+9251c2f (enc|dec)_neon: use vcreate_*() where appropriate
+399b916 lossy decoding: correct alpha-rescaling for YUVA format
+78c12ed Merge "Makefile.vc: add rudimentary avx2 support"
+dc5b122 try to remove the spurious warning for static analysis
+ddfefd6 Makefile.vc: add rudimentary avx2 support
+a891164 Merge "simplify VP8LInitBitReader()"
+fdbcd44 simplify VP8LInitBitReader()
+7c00428 makefile.unix: add rudimentary avx2 support
+515e35c Merge "add stub dsp/enc_avx2.c"
+a05dc14 SSE2: yuv->rgb speed-up for point-sampling
+178e9a6 add stub dsp/enc_avx2.c
+1b99c09 Merge "configure: add a test for -mavx2"
+fe72807 configure: add a test for -mavx2
+e46a247 cpu: fix check for __cpuidex availability
+176fda2 fix the bit-writer for lossless in 32bit mode
+541784c dsp.h: add a check for AVX2 / define WEBP_USE_AVX2
+bdb151e dsp/cpu: add AVX2 detection
+ab9f2f8 Merge "revamp the point-sampling functions by processing a full plane"
+a2f8b28 revamp the point-sampling functions by processing a full plane
+ef07602 use decoder's DSP functions for autofilter
+2b5cb32 Merge "dsp/cpu: add AVX detection"
+df08e67 dsp/cpu: add AVX detection
+e2f405c Merge "clean-up and slight speed-up in-loop filtering SSE2"
+f60957b clean-up and slight speed-up in-loop filtering SSE2
+9fc3ae4 .gitattributes: treat .ppm as binary
+3da924b Merge "dsp/WEBP_USE_NEON: test for __aarch64__"
+c716449 Android.mk: always include *_neon.c in the build
+a577b23 dsp/WEBP_USE_NEON: test for __aarch64__
+54bfffc move RemapBitReader() from idec.c to bit_reader code
+34168ec Merge "remove all unused layer code"
+f1e7717 remove all unused layer code
+b0757db Code cleanup for VP8LGetHistoImageSymbols.
+5fe628d make the token page size be variable instead of fixed 8192
+f948d08 memory debug: allow setting pre-defined malloc failure points
+ca3d746 use block-based allocation for backward refs storage, and free-lists
+1ba61b0 enable NEON intrinsics in aarch64 builds
+b9d2bb6 dsp/neon.h: coalesce intrinsics-related defines
+b5c7525 iosbuild: add support for iOSv7/aarch64
+9383afd Reduce number of memory allocations while decoding lossless.
+888e63e Merge "dsp/lossless: prevent signed int overflow in left shift ops"
+8137f3e Merge "instrument memory allocation routines for debugging"
+2aa1873 instrument memory allocation routines for debugging
+d3bcf72 Don't allocate VP8LHashChain, but treat like automatic object
+bd6b861 dsp/lossless: prevent signed int overflow in left shift ops
+b7f19b8 Merge "dec/vp8l: prevent signed int overflow in left shift ops"
+29059d5 Merge "remove some uint64_t casts and use."
+e69a1df dec/vp8l: prevent signed int overflow in left shift ops
+cf5eb8a remove some uint64_t casts and use.
+38e2db3 MIPS: MIPS32r1: Added optimization for HistogramAdd.
+e0609ad dwebp: fix exit code on webp load failure
+bbd358a Merge "example_util.h: avoid forward declaring enums"
+8955da2 example_util.h: avoid forward declaring enums
+6d6865f Added SSE2 variants for Average2/3/4
+b3a616b make HistogramAdd() a pointer in dsp
+c8bbb63 dec_neon: relocate some inline-asm defines
+4e393bb dec_neon: enable intrinsics-only functions
+ba99a92 dec_neon: use positive tests for USE_INTRINSICS
+69058ff Merge "example_util: add ExUtilDecodeWebPIncremental"
+a7828e8 dec_neon: make WORK_AROUND_GCC conditional on version
+3f3d717 Merge "enc_neon: enable intrinsics-only functions"
+de3cb6c Merge "move LOCAL_GCC_VERSION def to dsp.h"
+1b2fe14 example_util: add ExUtilDecodeWebPIncremental
+ca49e7a Merge "enc_neon: move Transpose4x4 to dsp/neon.h"
+ad900ab Merge "fix warning about size_t -> int conversion"
+4825b43 fix warning about size_t -> int conversion
+42b35e0 enc_neon: enable intrinsics-only functions
+f937e01 move LOCAL_GCC_VERSION def to dsp.h
+5e1a17e enc_neon: move Transpose4x4 to dsp/neon.h
+c7b92a5 dec_neon: (WORK_AROUND_GCC) delete unused Load4x8
+8e5f90b Merge "make ExUtilLoadWebP() accept NULL bitstream param."
+05d4c1b Merge "cwebp: add webpdec"
+ddeb6ac cwebp: add webpdec
+35d7d09 Merge "Reduce memory footprint for encoding WebP lossless."
+0b89610 Reduce memory footprint for encoding WebP lossless.
+f0b65c9 make ExUtilLoadWebP() accept NULL bitstream param.
+9c0a60c Merge "dwebp: move webp decoding to example_util"
+1d62acf MIPS: MIPS32r1: Added optimization for HuffmanCost functions.
+4a0e739 dwebp: move webp decoding to example_util
+c022046 Merge "Bugfix: Incremental decode of lossy-alpha"
+8c7cd72 Bugfix: Incremental decode of lossy-alpha
+7955152 MIPS: fix error with number of registers.
+b1dabe3 Merge "Move the HuffmanCost() function to dsp lib"
+75b1200 Move the HuffmanCost() function to dsp lib
+2772b8b MIPS: fix assembler error revealed by clang's debug build
+6653b60 enc_mips32: fix unused symbol warning in debug
+8dec120 enc_mips32: disable ITransform(One) in debug builds
+98519dd enc_neon: convert Disto4x4 to intrinsics
+fe9317c cosmetics:
+953b074 enc_neon: cosmetics
+a9fc697 Merge "WIP: extract the float-calculation of HuffmanCost from loop"
+3f84b52 Merge "replace some mult-long (vmull_u8) with mult-long-accumulate (vmlal_u8)"
+4ae0533 MIPS: MIPS32r1: Added optimizations for ExtraCost functions.
+b30a04c WIP: extract the float-calculation of HuffmanCost from loop
+a8fe8ce Merge "NEON intrinsics version of CollectHistogram"
+95203d2 NEON intrinsics version of CollectHistogram
+7ca2e74 replace some mult-long (vmull_u8) with mult-long-accumulate (vmlal_u8)
+41c6efb fix lossless_neon.c
+8ff96a0 NEON intrinsics version of FTransform
+0214f4a Merge "MIPS: MIPS32r1: Added optimizations for FastLog2"
+baabf1e MIPS: MIPS32r1: Added optimizations for FastLog2
+3d49871 NEON functions for lossless coding
+3fe0291 MIPS: MIPS32r1: Added optimizations for SSE functions.
+c503b48 Merge "fix the gcc-4.6.0 bug by implementing alternative method"
+abe6f48 fix the gcc-4.6.0 bug by implementing alternative method
+5598bde enc_mips32.c: fix file mode
+2b1b4d5 MIPS: MIPS32r1: Add optimization for GetResidualCost
+f0a1f3c Merge "MIPS: MIPS32r1: Added optimization for FTransform"
+7231f61 MIPS: MIPS32r1: Added optimization for FTransform
+869eaf6  ~30% encoding speedup: use NEON for QuantizeBlock()
+f758af6 enc_neon: convert FTransformWHT to intrinsics
+7dad095 MIPS: MIPS32r1: Added optimization for Disto4x4 (TTransform)
+2298d5f MIPS: MIPS32r1: Added optimization for QuantizeBlock
+e88150c Merge "MIPS: MIPS32r1: Add optimization for ITransform"
+de693f2 lossless_neon: disable VP8LConvert* functions
+4143332 NEON intrinsics for encoding
+0ca2914 MIPS: MIPS32r1: Add optimization for ITransform
+71bca5e dec_neon: use vst_lane instead of vget_lane
+bf06105 Intrinsics NEON version of TransformOne
+19c6f1b Merge "dec_neon: use vld?_lane instead of vset?_lane"
+7a94c0c upsampling_neon: drop NEON suffix from local functions
+d14669c upsampling_sse2: drop SSE2 suffix from local functions
+2ca42a4 enc_sse2: drop SSE2 suffix from local functions
+d038e61 dec_sse2: drop SSE2 suffix from local functions
+fa52d75 dec_neon: use vld?_lane instead of vset?_lane
+c520e77 cosmetic: fix long line
+4b0f2da Merge "add intrinsics NEON code for chroma strong-filtering"
+e351ec0 add intrinsics NEON code for chroma strong-filtering
+aaf734b Merge "Add SSE2 version of forward cross-color transform"
+c90a902 Add SSE2 version of forward cross-color transform
+bc374ff Use histogram_bits to initalize transform_bits.
+2132992 Merge "Add strong filtering intrinsics (inner and outer edges)"
+5fbff3a Add strong filtering intrinsics (inner and outer edges)
+d4813f0 Add SSE2 function for Inverse Cross-color Transform
+2602956 dec_neon: add strong loopfilter intrinsics
+cca7d7e Merge "add intrinsics version of SimpleHFilter16NEON()"
+1a05dfa windows: fix dll builds
+d6c50d8 Merge "add some colorspace conversion functions in NEON"
+4fd7c82 SSE2 variants of Subtract-Green: Rectify loop condition
+97e5fac add some colorspace conversion functions in NEON
+b9a7a45 add intrinsics version of SimpleHFilter16NEON()
+daccbf4 add light filtering NEON intrinsics
+af44460 fix typo in STORE_WHT
+6af6b8e Tune HistogramCombineBin for large images.
+af93bdd use WebPSafe[CM]alloc/WebPSafeFree instead of [cm]alloc/free
+51f406a lossless_sse2: relocate VP8LDspInitSSE2 proto
+0f4f721 separate SSE2 lossless functions into its own file
+514fc25 VP8LConvertFromBGRA: use conversion function pointers
+6d2f352 dsp/dec: TransformDCUV: use VP8TransformDC
+defc8e1 Merge "fix out-of-bound read during alpha-plane decoding"
+fbed364 Merge "dsp: reuse wht transform from dec in encoder"
+d846708 Merge "Add SSE2 version of ARGB -> BGR/RGB/... conversion functions"
+207d03b fix out-of-bound read during alpha-plane decoding
+d1b33ad 2-5% faster trellis with clang/MacOS (and ~2-3% on ARM)
+369c26d Add SSE2 version of ARGB -> BGR/RGB/... conversion functions
+df230f2 dsp: reuse wht transform from dec in encoder
+80e218d Android.mk: fix build with APP_ABI=armeabi-v7a-hard
+59daf08 Merge "cosmetics:"
+5362200 cosmetics:
+3e7f34a AssignSegments: quiet array-bounds warning
+3c2ebf5 Merge "UpdateHistogramCost: avoid implicit double->float"
+cf821c8 UpdateHistogramCost: avoid implicit double->float
+312e638 Extend the search space for GetBestGreenRedToBlue
+1c58526 Fix few nits
+fef2270 Optimize and re-structure VP8LGetHistoImageSymbols
+068b14a Optimize lossless decoding.
+5f0cfa8 Do a binary search to get the optimum cache bits.
+24ca367 Merge "allow 'cwebp -o -' to emit output to stdout"
+e12f874 allow 'cwebp -o -' to emit output to stdout
+2bcad89 allow some more stdin/stout I/O
+84ed4b3 fix cwebp.1 typos after patch #69199
+65b99f1 add a -z option to cwebp, and WebPConfigLosslessPreset() function
+3017661 4-5% faster trellis by removing some unneeded calculations.
+687a58e histogram.c: reindent after b33e8a0
+06d456f Merge "~3-4% faster lossless encoding"
+c60de26 ~3-4% faster lossless encoding
+42eb06f Merge "few cosmetics after patch #69079"
+82af826 few cosmetics after patch #69079
+b33e8a0 Refactor code for HistogramCombine.
+ca1bfff Merge "5-10% encoding speedup with faster trellis (-m 6)"
+5aeeb08 5-10% encoding speedup with faster trellis (-m 6)
+82ae1bf cosmetics: normalize VP8GetCPUInfo checks
+e3dd924 Merge "Refactor GetBestPredictorForTile for future tuning."
+206cc1b Refactor GetBestPredictorForTile for future tuning.
+3cb8406 Merge "speed-up trellis quant (~5-10% overall speed-up)"
+b66f222 Merge "lossy encoding: ~3% speed-up"
+4287d0d speed-up trellis quant (~5-10% overall speed-up)
+390c8b3 lossy encoding: ~3% speed-up
+9a463c4 Merge "dec_neon: convert TransformWHT to intrinsics"
+e8605e9 Merge "dec_neon: add ConvertU8ToS16"
+4aa3e41 MIPS: MIPS32r1: rescaler bugfix
+c16cd99 Speed up lossless encoder.
+9d6b5ff dec_neon: convert TransformWHT to intrinsics
+2ff0aae dec_neon: add ConvertU8ToS16
+77a8f91 fix compilation with USE_YUVj flag
+4acbec1 Merge changes I3b240ffb,Ia9370283,Ia2d28728
+2719bb7 dec_neon: TransformAC3: work on packed vectors
+b7b60ca dec_neon: add SaturateAndStore4x4
+b7685d7 Rescale: let ImportRow / ExportRow be pointer-to-function
+e02f16e dec_neon.c: convert TransformDC to intrinsics
+9cba963 add missing file
+8992ddb use static clipping tables
+0235d5e 1-2% faster quantization in SSE2
+b2fbc36 fix VC12-x64 warning
+6e37cb9 Merge "cosmetics: backward_references.c: reindent after a7d2ee3"
+a42ea97 cosmetics: backward_references.c: reindent after a7d2ee3
+6c32744 Merge "fix missing __BIG_ENDIAN__ definition on some platform"
+a8b6aad fix missing __BIG_ENDIAN__ definition on some platform
+fde2904 Increase initial buffer size for VP8L Bit Writer.
+a7d2ee3 Optimize cache estimate logic.
+7fb6095 Merge "dec_neon.c: add TransformAC3"
+bf182e8 VP8LBitWriter: use a bit-accumulator
+3f40b4a Merge "MIPS: MIPS32r1: clang macro warning resolved"
+1684f4e WebP Decoder: Mark some truncated bitstreams as invalid
+acbedac MIPS: MIPS32r1: clang macro warning resolved
+228e487 dec_neon.c: add TransformAC3
+393f89b Android.mk: avoid gcc-specific flags with clang
+32aeaf1 revamp VP8LColorSpaceTransform() a bit
+0c7cc4c Merge "Don't dereference NULL, ensure HashChain fully initialized"
+391316f Don't dereference NULL, ensure HashChain fully initialized
+926ff40 WEBP_SWAP_16BIT_CSP: remove code dup
+1d1cd3b Fix decode bug for rgbA_4444/RGBA_4444 color-modes.
+939e70e update AUTHORS file
+8934a62 cosmetics: *_mips32.c
+dd438c9 MIPS: MIPS32r1: Optimization of some simple point-sampling functions. PATCH [6/6]
+5352091 Added support for calling sampling functions via pointers.
+d16c697 MIPS: MIPS32r1: Optimization of filter functions. PATCH [5/6]
+04336fc MIPS: MIPS32r1: Optimization of function TransformOne. PATCH [4/6]
+92d8fc7 MIPS: MIPS32r1: Optimization of function WebPRescalerImportRow. PATCH [3/6]
+bbc23ff parse one row of intra modes altogether
+a2f608f Merge "MIPS: MIPS32r1: Optimization of function WebPRescalerExportRow. [2/6]"
+8823085 MIPS: MIPS32r1: Optimization of function WebPRescalerExportRow. [2/6]
+c5a5b02 decode mt+incremental: fix segfault in debug builds
+9882b2f always use fast-analysis for all methods.
+000adac Merge "autoconf: update ax_pthread.m4"
+2d2fc37 update .gitignore
+5bf4255 Merge "Make it possible to avoid automagic dependencies"
+c1cb193 disable NEON for arm64 platform
+73a304e Make it possible to avoid automagic dependencies
+4d493f8 MIPS: MIPS32r1: Decoder bit reader function optimized. PATCH [1/6]
+c741183 make WebPCleanupTransparentArea work with argb picture
+5da1855 add a decoding option to flip image vertically
+00c3c4e Merge "add man/vwebp.1"
+2c6bb42 add man/vwebp.1
+ea59a8e Merge "Merge tag 'v0.4.0'"
+7574bed fix comments related to array sizes
+0b5a90f dwebp.1: fix option formatting
+effcb0f Merge tag 'v0.4.0'
+7c76255 autoconf: update ax_pthread.m4
+fff2a11 make -short work with -print_ssim, -print_psnr, etc.
+68e7901 update ChangeLog (tag: v0.4.0-rc1, tag: v0.4.0, origin/0.4.0, 0.4.0)
 256e433 update NEWS description with new general features
 2962534 Merge "gif2webp: don't use C99 %zu" into 0.4.0
 3b9f9dd gif2webp: don't use C99 %zu

Makefile.vc

@@ -27,7 +27,7 @@ PLATFORM_LDFLAGS = /SAFESEH
 NOLOGO     = /nologo
 CCNODBG    = cl.exe $(NOLOGO) /O2 /DNDEBUG
 CCDEBUG    = cl.exe $(NOLOGO) /Od /Gm /Zi /D_DEBUG /RTC1
-CFLAGS     = /Isrc $(NOLOGO) /W3 /EHsc /c /GS
+CFLAGS     = /Isrc $(NOLOGO) /W3 /EHsc /c
 CFLAGS     = $(CFLAGS) /DWIN32 /D_CRT_SECURE_NO_WARNINGS /DWIN32_LEAN_AND_MEAN
 CFLAGS     = $(CFLAGS) /DHAVE_WINCODEC_H /DWEBP_USE_THREAD
 LDFLAGS    = /LARGEADDRESSAWARE /MANIFEST /NXCOMPAT /DYNAMICBASE
@@ -44,11 +44,21 @@ OUTDIR = ..\obj\
 OUTDIR = $(OBJDIR)
 !ENDIF
 
+!IF "$(HAVE_AVX2)" == "1"
+CFLAGS = $(CFLAGS) /DWEBP_HAVE_AVX2
+AVX2_FLAGS = /arch:AVX2
+!ENDIF
+
 ##############################################################
 # Runtime library configuration
 !IF "$(RTLIBCFG)" == "static"
 RTLIB  = /MT
 RTLIBD = /MTd
+!ELSE IF "$(RTLIBCFG)" == "legacy"
+RTLIBCFG = static
+RTLIB  = /MT
+RTLIBD = /MTd
+CFLAGS = $(CFLAGS) /GS- /arch:IA32
 !ELSE
 RTLIB   = /MD
 RTLIBD  = /MDd
@@ -134,6 +144,7 @@ CFGSET = TRUE
 !MESSAGE - all                            - build (de)mux-based targets for CFG
 !MESSAGE
 !MESSAGE RTLIBCFG controls the runtime library linkage - 'static' or 'dynamic'.
+!MESSAGE   'legacy' will produce a Windows 2000 compatible library.
 !MESSAGE OBJDIR is the path where you like to build (obj, bins, etc.),
 !MESSAGE   defaults to ..\obj
 
@@ -156,7 +167,6 @@ DEC_OBJS = \
     $(DIROBJ)\dec\frame.obj \
     $(DIROBJ)\dec\idec.obj \
     $(DIROBJ)\dec\io.obj \
-    $(DIROBJ)\dec\layer.obj \
     $(DIROBJ)\dec\quant.obj \
     $(DIROBJ)\dec\tree.obj \
     $(DIROBJ)\dec\vp8.obj \
@@ -167,18 +177,29 @@ DEMUX_OBJS = \
     $(DIROBJ)\demux\demux.obj \
 
 DSP_DEC_OBJS = \
+    $(DIROBJ)\dsp\alpha_processing.obj \
+    $(DIROBJ)\dsp\alpha_processing_sse2.obj \
     $(DIROBJ)\dsp\cpu.obj \
     $(DIROBJ)\dsp\dec.obj \
+    $(DIROBJ)\dsp\dec_clip_tables.obj \
+    $(DIROBJ)\dsp\dec_mips32.obj \
     $(DIROBJ)\dsp\dec_neon.obj \
     $(DIROBJ)\dsp\dec_sse2.obj \
     $(DIROBJ)\dsp\lossless.obj \
+    $(DIROBJ)\dsp\lossless_mips32.obj \
+    $(DIROBJ)\dsp\lossless_neon.obj \
+    $(DIROBJ)\dsp\lossless_sse2.obj \
     $(DIROBJ)\dsp\upsampling.obj \
     $(DIROBJ)\dsp\upsampling_neon.obj \
     $(DIROBJ)\dsp\upsampling_sse2.obj \
     $(DIROBJ)\dsp\yuv.obj \
+    $(DIROBJ)\dsp\yuv_mips32.obj \
+    $(DIROBJ)\dsp\yuv_sse2.obj \
 
 DSP_ENC_OBJS = \
     $(DIROBJ)\dsp\enc.obj \
+    $(DIROBJ)\dsp\enc_avx2.obj \
+    $(DIROBJ)\dsp\enc_mips32.obj \
     $(DIROBJ)\dsp\enc_neon.obj \
     $(DIROBJ)\dsp\enc_sse2.obj \
 
@@ -187,6 +208,7 @@ EX_FORMAT_DEC_OBJS = \
     $(DIROBJ)\examples\metadata.obj \
     $(DIROBJ)\examples\pngdec.obj \
     $(DIROBJ)\examples\tiffdec.obj \
+    $(DIROBJ)\examples\webpdec.obj \
     $(DIROBJ)\examples\wicdec.obj \
 
 EX_UTIL_OBJS = \
@@ -202,8 +224,11 @@ ENC_OBJS = \
     $(DIROBJ)\enc\frame.obj \
     $(DIROBJ)\enc\histogram.obj \
     $(DIROBJ)\enc\iterator.obj \
-    $(DIROBJ)\enc\layer.obj \
     $(DIROBJ)\enc\picture.obj \
+    $(DIROBJ)\enc\picture_csp.obj \
+    $(DIROBJ)\enc\picture_psnr.obj \
+    $(DIROBJ)\enc\picture_rescale.obj \
+    $(DIROBJ)\enc\picture_tools.obj \
     $(DIROBJ)\enc\quant.obj \
     $(DIROBJ)\enc\syntax.obj \
     $(DIROBJ)\enc\token.obj \
@@ -217,7 +242,6 @@ MUX_OBJS = \
     $(DIROBJ)\mux\muxread.obj \
 
 UTILS_DEC_OBJS = \
-    $(DIROBJ)\utils\alpha_processing.obj \
     $(DIROBJ)\utils\bit_reader.obj \
     $(DIROBJ)\utils\color_cache.obj \
     $(DIROBJ)\utils\filters.obj \
@@ -309,8 +333,14 @@ $(DIROBJ)\$(DLLC): $(DIROBJ)\$(DLLINC)
 	@echo } >> $@
 
 .SUFFIXES: .c .obj .res .exe
+# File-specific flag builds. Note batch rules take precedence over wildcards,
+# so for now name each file individually.
+$(DIROBJ)\dsp\enc_avx2.obj: src\dsp\enc_avx2.c
+	$(CC) $(CFLAGS) $(AVX2_FLAGS) /Fd$(LIBWEBP_PDBNAME) /Fo$(DIROBJ)\dsp\ \
+	  src\dsp\$(@B).c
+# Batch rules
 {examples}.c{$(DIROBJ)\examples}.obj::
-	$(CC) $(CFLAGS) /Fd$(DIROBJ)\examples\ /Fo$(DIROBJ)\examples\  $<
+	$(CC) $(CFLAGS) /Fd$(DIROBJ)\examples\ /Fo$(DIROBJ)\examples\ $<
 {src\dec}.c{$(DIROBJ)\dec}.obj::
 	$(CC) $(CFLAGS) /Fd$(LIBWEBP_PDBNAME) /Fo$(DIROBJ)\dec\ $<
 {src\demux}.c{$(DIROBJ)\demux}.obj::

NEWS

@@ -1,3 +1,29 @@
+- 3/3/15: version 0.4.3
+  This is a binary compatible release.
+  * Android / gcc / iOS / MSVS build fixes and improvements
+  * lossless decode fix (issue #239 -- since 0.4.0)
+  * documentation / vwebp updates for animation
+  * multi-threading fix (issue #234)
+
+- 10/13/14: version 0.4.2
+  This is a binary compatible release.
+  * Android / gcc build fixes
+  * (Windows) fix reading from stdin and writing to stdout
+  * gif2webp: miscellaneous fixes
+  * fix 'alpha-leak' with lossy compression (issue #220)
+  * the lossless bitstream spec has been amended to reflect the current code
+
+- 7/24/14: version 0.4.1
+  This is a binary compatible release.
+  * AArch64 (arm64) & MIPS support/optimizations
+  * NEON assembly additions:
+    - ~25% faster lossy decode / encode (-m 4)
+    - ~10% faster lossless decode
+    - ~5-10% faster lossless encode (-m 3/4)
+  * dwebp/vwebp can read from stdin
+  * cwebp/gif2webp can write to stdout
+  * cwebp can read webp files; useful if storing sources as webp lossless
+
 - 12/19/13: version 0.4.0
   * improved gif2webp tool
   * numerous fixes, compression improvement and speed-up

PATENTS

@@ -1,22 +1,23 @@
 Additional IP Rights Grant (Patents)
+------------------------------------
 
-"This implementation" means the copyrightable works distributed by
-Google as part of the WebM Project.
+"These implementations" means the copyrightable works that implement the WebM
+codecs distributed by Google as part of the WebM Project.
 
-Google hereby grants to you a perpetual, worldwide, non-exclusive,
-no-charge, royalty-free, irrevocable (except as stated in this section)
-patent license to make, have made, use, offer to sell, sell, import,
-transfer, and otherwise run, modify and propagate the contents of this
-implementation of VP8, where such license applies only to those patent
-claims, both currently owned by Google and acquired in the future,
-licensable by Google that are necessarily infringed by this
-implementation of VP8. This grant does not include claims that would be
-infringed only as a consequence of further modification of this
-implementation. If you or your agent or exclusive licensee institute or
-order or agree to the institution of patent litigation against any
-entity (including a cross-claim or counterclaim in a lawsuit) alleging
-that this implementation of VP8 or any code incorporated within this
-implementation of VP8 constitutes direct or contributory patent
-infringement, or inducement of patent infringement, then any patent
-rights granted to you under this License for this implementation of VP8
-shall terminate as of the date such litigation is filed.
+Google hereby grants to you a perpetual, worldwide, non-exclusive, no-charge,
+royalty-free, irrevocable (except as stated in this section) patent license to
+make, have made, use, offer to sell, sell, import, transfer, and otherwise
+run, modify and propagate the contents of these implementations of WebM, where
+such license applies only to those patent claims, both currently owned by
+Google and acquired in the future, licensable by Google that are necessarily
+infringed by these implementations of WebM. This grant does not include claims
+that would be infringed only as a consequence of further modification of these
+implementations. If you or your agent or exclusive licensee institute or order
+or agree to the institution of patent litigation or any other patent
+enforcement activity against any entity (including a cross-claim or
+counterclaim in a lawsuit) alleging that any of these implementations of WebM
+or any code incorporated within any of these implementations of WebM
+constitutes direct or contributory patent infringement, or inducement of
+patent infringement, then any patent rights granted to you under this License
+for these implementations of WebM shall terminate as of the date such
+litigation is filed.

README

@@ -4,7 +4,7 @@
           \__\__/\____/\_____/__/ ____  ___
                 / _/ /    \    \ /  _ \/ _/
                /  \_/   / /   \ \   __/  \__
-               \____/____/\_____/_____/____/v0.4.0
+               \____/____/\_____/_____/____/v0.4.3
 
 Description:
 ============
@@ -140,28 +140,30 @@ A longer list of options is available using the -longhelp command line flag:
 Usage:
  cwebp [-preset <...>] [options] in_file [-o out_file]
 
-If input size (-s) for an image is not specified, it is assumed to be a PNG,
-JPEG or TIFF file.
-options:
+If input size (-s) for an image is not specified, it is
+assumed to be a PNG, JPEG, TIFF or WebP file.
+
+Options:
   -h / -help  ............ short help
   -H / -longhelp  ........ long help
   -q <float> ............. quality factor (0:small..100:big)
-  -alpha_q <int> ......... Transparency-compression quality (0..100).
-  -preset <string> ....... Preset setting, one of:
+  -alpha_q <int> ......... transparency-compression quality (0..100)
+  -preset <string> ....... preset setting, one of:
                             default, photo, picture,
                             drawing, icon, text
-     -preset must come first, as it overwrites other parameters.
+     -preset must come first, as it overwrites other parameters
+
   -m <int> ............... compression method (0=fast, 6=slowest)
   -segments <int> ........ number of segments to use (1..4)
-  -size <int> ............ Target size (in bytes)
-  -psnr <float> .......... Target PSNR (in dB. typically: 42)
+  -size <int> ............ target size (in bytes)
+  -psnr <float> .......... target PSNR (in dB. typically: 42)
 
-  -s <int> <int> ......... Input size (width x height) for YUV
-  -sns <int> ............. Spatial Noise Shaping (0:off, 100:max)
+  -s <int> <int> ......... input size (width x height) for YUV
+  -sns <int> ............. spatial noise shaping (0:off, 100:max)
   -f <int> ............... filter strength (0=off..100)
   -sharpness <int> ....... filter sharpness (0:most .. 7:least sharp)
-  -strong ................ use strong filter instead of simple (default).
-  -nostrong .............. use simple filter instead of strong.
+  -strong ................ use strong filter instead of simple (default)
+  -nostrong .............. use simple filter instead of strong
   -partition_limit <int> . limit quality to fit the 512k limit on
                            the first partition (0=no degradation ... 100=full)
   -pass <int> ............ analysis pass number (1..10)
@@ -169,41 +171,40 @@ options:
   -resize <w> <h> ........ resize picture (after any cropping)
   -mt .................... use multi-threading if available
   -low_memory ............ reduce memory usage (slower encoding)
-  -map <int> ............. print map of extra info.
-  -print_psnr ............ prints averaged PSNR distortion.
-  -print_ssim ............ prints averaged SSIM distortion.
-  -print_lsim ............ prints local-similarity distortion.
-  -d <file.pgm> .......... dump the compressed output (PGM file).
-  -alpha_method <int> .... Transparency-compression method (0..1)
-  -alpha_filter <string> . predictive filtering for alpha plane.
-                           One of: none, fast (default) or best.
-  -alpha_cleanup ......... Clean RGB values in transparent area.
-  -blend_alpha <hex> ..... Blend colors against background color
+  -map <int> ............. print map of extra info
+  -print_psnr ............ prints averaged PSNR distortion
+  -print_ssim ............ prints averaged SSIM distortion
+  -print_lsim ............ prints local-similarity distortion
+  -d <file.pgm> .......... dump the compressed output (PGM file)
+  -alpha_method <int> .... transparency-compression method (0..1)
+  -alpha_filter <string> . predictive filtering for alpha plane,
+                           one of: none, fast (default) or best
+  -alpha_cleanup ......... clean RGB values in transparent area
+  -blend_alpha <hex> ..... blend colors against background color
                            expressed as RGB values written in
                            hexadecimal, e.g. 0xc0e0d0 for red=0xc0
-                           green=0xe0 and blue=0xd0.
-  -noalpha ............... discard any transparency information.
-  -lossless .............. Encode image losslessly.
-  -hint <string> ......... Specify image characteristics hint.
-                           One of: photo, picture or graph
+                           green=0xe0 and blue=0xd0
+  -noalpha ............... discard any transparency information
+  -lossless .............. encode image losslessly
+  -hint <string> ......... specify image characteristics hint,
+                           one of: photo, picture or graph
 
   -metadata <string> ..... comma separated list of metadata to
                            copy from the input to the output if present.
                            Valid values: all, none (default), exif, icc, xmp
 
   -short ................. condense printed message
-  -quiet ................. don't print anything.
-  -version ............... print version number and exit.
-  -noasm ................. disable all assembly optimizations.
+  -quiet ................. don't print anything
+  -version ............... print version number and exit
+  -noasm ................. disable all assembly optimizations
   -v ..................... verbose, e.g. print encoding/decoding times
   -progress .............. report encoding progress
 
 Experimental Options:
-  -jpeg_like ............. Roughly match expected JPEG size.
-  -af .................... auto-adjust filter strength.
+  -jpeg_like ............. roughly match expected JPEG size
+  -af .................... auto-adjust filter strength
   -pre <int> ............. pre-processing filter
 
-
 The main options you might want to try in order to further tune the
 visual quality are:
  -preset
@@ -262,19 +263,19 @@ Use following options to convert into alternate image formats:
   -yuv ......... save the raw YUV samples in flat layout
 
  Other options are:
-  -version  .... print version number and exit.
-  -nofancy ..... don't use the fancy YUV420 upscaler.
-  -nofilter .... disable in-loop filtering.
-  -nodither .... disable dithering.
+  -version  .... print version number and exit
+  -nofancy ..... don't use the fancy YUV420 upscaler
+  -nofilter .... disable in-loop filtering
+  -nodither .... disable dithering
   -dither <d> .. dithering strength (in 0..100)
   -mt .......... use multi-threading
   -crop <x> <y> <w> <h> ... crop output with the given rectangle
   -scale <w> <h> .......... scale the output (*after* any cropping)
-  -alpha ....... only save the alpha plane.
+  -alpha ....... only save the alpha plane
   -incremental . use incremental decoding (useful for tests)
-  -h     ....... this help message.
+  -h     ....... this help message
   -v     ....... verbose (e.g. print encoding/decoding times)
-  -noasm ....... disable all assembly optimizations.
+  -noasm ....... disable all assembly optimizations
 
 Visualization tool:
 ===================
@@ -288,19 +289,19 @@ Usage: vwebp in_file [options]
 
 Decodes the WebP image file and visualize it using OpenGL
 Options are:
-  -version  .... print version number and exit.
-  -noicc ....... don't use the icc profile if present.
-  -nofancy ..... don't use the fancy YUV420 upscaler.
-  -nofilter .... disable in-loop filtering.
-  -dither <int>  dithering strength (0..100). Default=50.
-  -mt .......... use multi-threading.
-  -info ........ print info.
-  -h     ....... this help message.
+  -version  .... print version number and exit
+  -noicc ....... don't use the icc profile if present
+  -nofancy ..... don't use the fancy YUV420 upscaler
+  -nofilter .... disable in-loop filtering
+  -dither <int>  dithering strength (0..100), default=50
+  -mt .......... use multi-threading
+  -info ........ print info
+  -h     ....... this help message
 
 Keyboard shortcuts:
-  'c' ................ toggle use of color profile.
-  'i' ................ overlay file information.
-  'q' / 'Q' / ESC .... quit.
+  'c' ................ toggle use of color profile
+  'i' ................ overlay file information
+  'q' / 'Q' / ESC .... quit
 
 Building:
 ---------
@@ -336,24 +337,24 @@ vwebp.
 
 Usage:
  gif2webp [options] gif_file -o webp_file
-options:
+Options:
   -h / -help  ............ this help
-  -lossy ................. Encode image using lossy compression.
-  -mixed ................. For each frame in the image, pick lossy
-                           or lossless compression heuristically.
+  -lossy ................. encode image using lossy compression
+  -mixed ................. for each frame in the image, pick lossy
+                           or lossless compression heuristically
   -q <float> ............. quality factor (0:small..100:big)
   -m <int> ............... compression method (0=fast, 6=slowest)
-  -kmin <int> ............ Min distance between key frames
-  -kmax <int> ............ Max distance between key frames
+  -kmin <int> ............ min distance between key frames
+  -kmax <int> ............ max distance between key frames
   -f <int> ............... filter strength (0=off..100)
   -metadata <string> ..... comma separated list of metadata to
-                           copy from the input to the output if present.
+                           copy from the input to the output if present
                            Valid values: all, none, icc, xmp (default)
   -mt .................... use multi-threading if available
 
-  -version ............... print version number and exit.
-  -v ..................... verbose.
-  -quiet ................. don't print anything.
+  -version ............... print version number and exit
+  -v ..................... verbose
+  -quiet ................. don't print anything
 
 Building:
 ---------
@@ -442,15 +443,20 @@ The encoding flow looks like:
 
   // Set up a byte-output write method. WebPMemoryWriter, for instance.
   WebPMemoryWriter wrt;
+  WebPMemoryWriterInit(&wrt);     // initialize 'wrt'
+
   pic.writer = MyFileWriter;
   pic.custom_ptr = my_opaque_structure_to_make_MyFileWriter_work;
-  // initialize 'wrt' here...
 
   // Compress!
   int ok = WebPEncode(&config, &pic);   // ok = 0 => error occurred!
   WebPPictureFree(&pic);  // must be called independently of the 'ok' result.
 
   // output data should have been handled by the writer at that point.
+  // -> compressed data is the memory buffer described by wrt.mem / wrt.size
+
+  // deallocate the memory used by compressed data
+  WebPMemoryWriterClear(&wrt);
 
 -------------------------------------- END PSEUDO EXAMPLE
 

README.mux

@@ -1,7 +1,7 @@
           __   __  ____  ____  ____  __ __  _     __ __
          /  \\/  \/  _ \/  _ \/  _ \/  \  \/ \___/_ / _\
          \       /   __/  _  \   __/      /  /  (_/  /__
-          \__\__/\_____/_____/__/  \__//_/\_____/__/___/v0.2.0
+          \__\__/\_____/_____/__/  \__//_/\_____/__/___/v0.2.2
 
 
 Description:
@@ -33,35 +33,35 @@ Usage: webpmux -get GET_OPTIONS INPUT -o OUTPUT
        webpmux -version
 
 GET_OPTIONS:
- Extract relevant data.
-   icc       Get ICC profile.
-   exif      Get EXIF metadata.
-   xmp       Get XMP metadata.
-   frame n   Get nth frame.
+ Extract relevant data:
+   icc       get ICC profile
+   exif      get EXIF metadata
+   xmp       get XMP metadata
+   frame n   get nth frame
 
 SET_OPTIONS:
- Set color profile/metadata.
-   icc  file.icc     Set ICC profile.
-   exif file.exif    Set EXIF metadata.
-   xmp  file.xmp     Set XMP metadata.
+ Set color profile/metadata:
+   icc  file.icc     set ICC profile
+   exif file.exif    set EXIF metadata
+   xmp  file.xmp     set XMP metadata
    where:    'file.icc' contains the ICC profile to be set,
              'file.exif' contains the EXIF metadata to be set
              'file.xmp' contains the XMP metadata to be set
 
 STRIP_OPTIONS:
- Strip color profile/metadata.
-   icc       Strip ICC profile.
-   exif      Strip EXIF metadata.
-   xmp       Strip XMP metadata.
+ Strip color profile/metadata:
+   icc       strip ICC profile
+   exif      strip EXIF metadata
+   xmp       strip XMP metadata
 
 FRAME_OPTIONS(i):
- Create animation.
+ Create animation:
    file_i +di+[xi+yi[+mi[bi]]]
    where:    'file_i' is the i'th animation frame (WebP format),
-             'di' is the pause duration before next frame.
-             'xi','yi' specify the image offset for this frame.
-             'mi' is the dispose method for this frame (0 or 1).
-             'bi' is the blending method for this frame (+b or -b).
+             'di' is the pause duration before next frame,
+             'xi','yi' specify the image offset for this frame,
+             'mi' is the dispose method for this frame (0 or 1),
+             'bi' is the blending method for this frame (+b or -b)
 
 LOOP_COUNT:
  Number of times to repeat the animation.
@@ -72,7 +72,7 @@ BACKGROUND_COLOR:
   A,R,G,B
   where:    'A', 'R', 'G' and 'B' are integers in the range 0 to 255 specifying
             the Alpha, Red, Green and Blue component values respectively
-            [Default: 255,255,255,255].
+            [Default: 255,255,255,255]
 
 INPUT & OUTPUT are in WebP format.
 

configure.ac

@@ -1,7 +1,8 @@
-AC_INIT([libwebp], [0.4.0],
+AC_INIT([libwebp], [0.4.3],
         [http://code.google.com/p/webp/issues],,
         [http://developers.google.com/speed/webp])
-AC_CANONICAL_TARGET
+AC_CANONICAL_HOST
+AC_PREREQ([2.60])
 AM_INIT_AUTOMAKE([-Wall foreign subdir-objects])
 
 dnl === automake >= 1.12 requires this for 'unusual archivers' support.
@@ -14,6 +15,9 @@ AM_PROG_CC_C_O
 dnl === Enable less verbose output when building.
 m4_ifdef([AM_SILENT_RULES], [AM_SILENT_RULES([yes])])
 
+dnl == test endianness
+AC_C_BIGENDIAN
+
 dnl === SET_IF_UNSET(shell_var, value)
 dnl ===   Set the shell variable 'shell_var' to 'value' if it is unset.
 AC_DEFUN([SET_IF_UNSET], [test "${$1+set}" = "set" || $1=$2])
@@ -31,33 +35,78 @@ AC_ARG_WITH([pkgconfigdir], AS_HELP_STRING([--with-pkgconfigdir=DIR],
             [pkgconfigdir="$withval"], [pkgconfigdir='${libdir}/pkgconfig'])
 AC_SUBST([pkgconfigdir])
 
-dnl === TEST_AND_ADD_CFLAGS(flag)
-dnl ===   Checks whether $CC supports 'flag' and adds it to AM_CFLAGS on success.
+dnl === TEST_AND_ADD_CFLAGS(var, flag)
+dnl ===   Checks whether $CC supports 'flag' and adds it to 'var'
+dnl ===   on success.
 AC_DEFUN([TEST_AND_ADD_CFLAGS],
          [SAVED_CFLAGS="$CFLAGS"
-          CFLAGS="-Werror $1"
-          AC_MSG_CHECKING([whether $CC supports $1])
+          CFLAGS="-Werror $2"
+          AC_MSG_CHECKING([whether $CC supports $2])
           dnl Note AC_LANG_PROGRAM([]) uses an old-style main definition.
           AC_COMPILE_IFELSE([AC_LANG_SOURCE([int main(void) { return 0; }])],
                             [AC_MSG_RESULT([yes])]
                             dnl Simply append the variable avoiding a
                             dnl compatibility ifdef for AS_VAR_APPEND as this
                             dnl variable shouldn't grow all that large.
-                            [AM_CFLAGS="$AM_CFLAGS $1"],
+                            [$1="${$1} $2"],
                             [AC_MSG_RESULT([no])])
           CFLAGS="$SAVED_CFLAGS"])
-TEST_AND_ADD_CFLAGS([-Wall])
-TEST_AND_ADD_CFLAGS([-Wdeclaration-after-statement])
-TEST_AND_ADD_CFLAGS([-Wextra])
-TEST_AND_ADD_CFLAGS([-Wmissing-declarations])
-TEST_AND_ADD_CFLAGS([-Wmissing-prototypes])
-TEST_AND_ADD_CFLAGS([-Wold-style-definition])
-TEST_AND_ADD_CFLAGS([-Wshadow])
-TEST_AND_ADD_CFLAGS([-Wunused-but-set-variable])
-TEST_AND_ADD_CFLAGS([-Wunused])
-TEST_AND_ADD_CFLAGS([-Wvla])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wall])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wdeclaration-after-statement])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wextra])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wformat-nonliteral])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wformat-security])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wmissing-declarations])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wmissing-prototypes])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wold-style-definition])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wshadow])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wunused-but-set-variable])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wunused])
+TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-Wvla])
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62040
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61622
+AS_IF([test "$GCC" = "yes" ], [
+       gcc_version=`$CC -dumpversion`
+       gcc_wht_bug=""
+       case "$host_cpu" in
+         aarch64|arm64)
+          case "$gcc_version" in
+            4.9|4.9.0|4.9.1) gcc_wht_bug=yes ;;
+          esac
+       esac
+       AS_IF([test "$gcc_wht_bug" = "yes"], [
+              TEST_AND_ADD_CFLAGS([AM_CFLAGS], [-frename-registers])])])
 AC_SUBST([AM_CFLAGS])
 
+dnl === Check for machine specific flags
+TEST_AND_ADD_CFLAGS([AVX2_FLAGS], [-mavx2])
+AS_IF([test -n "$AVX2_FLAGS"], [
+  SAVED_CFLAGS=$CFLAGS
+  CFLAGS="$CFLAGS $AVX2_FLAGS"
+  AC_CHECK_HEADER([immintrin.h],
+                  [AC_DEFINE(WEBP_HAVE_AVX2, [1],
+                   [Set to 1 if AVX2 is supported])],
+                  [AVX2_FLAGS=""],
+                  dnl it's illegal to directly include avx2intrin.h, but it's
+                  dnl included conditionally in immintrin.h, tricky!
+                  [#ifndef __AVX2__
+                   #error avx2 is not enabled
+                   #endif
+                  ])
+  CFLAGS=$SAVED_CFLAGS])
+AC_SUBST([AVX2_FLAGS])
+
+TEST_AND_ADD_CFLAGS([SSE2_FLAGS], [-msse2])
+AS_IF([test -n "$SSE2_FLAGS"], [
+  SAVED_CFLAGS=$CFLAGS
+  CFLAGS="$CFLAGS $SSE2_FLAGS"
+  AC_CHECK_HEADER([emmintrin.h],
+                  [AC_DEFINE(WEBP_HAVE_SSE2, [1],
+                   [Set to 1 if SSE2 is supported])],
+                  [SSE2_FLAGS=""])
+  CFLAGS=$SAVED_CFLAGS])
+AC_SUBST([SSE2_FLAGS])
+
 dnl === CLEAR_LIBVARS([var_pfx])
 dnl ===   Clears <var_pfx>_{INCLUDES,LIBS}.
 AC_DEFUN([CLEAR_LIBVARS], [$1_INCLUDES=""; $1_LIBS=""])
@@ -93,6 +142,26 @@ AC_DEFUN([LIBCHECK_EPILOGUE],
           CPPFLAGS=$SAVED_CPPFLAGS
           LIBS=$SAVED_LIBS])
 
+dnl === Check for gcc builtins
+
+dnl === CHECK_FOR_BUILTIN([builtin], [param], [define])
+dnl ===   links a C AC_LANG_PROGRAM, with <builtin>(<param>)
+dnl ===   AC_DEFINE'ing <define> if successful.
+AC_DEFUN([CHECK_FOR_BUILTIN],
+         [AC_LANG_PUSH([C])
+          AC_MSG_CHECKING([for $1])
+          AC_LINK_IFELSE([AC_LANG_PROGRAM([], [$1($2)])],
+                         [AC_MSG_RESULT([yes])
+                          AC_DEFINE([$3], [1],
+                                    [Set to 1 if $1 is available])],
+                         [AC_MSG_RESULT([no])]),
+          AC_LANG_POP])
+
+dnl AC_CHECK_FUNC doesn't work with builtin's.
+CHECK_FOR_BUILTIN([__builtin_bswap16], [1u << 15], [HAVE_BUILTIN_BSWAP16])
+CHECK_FOR_BUILTIN([__builtin_bswap32], [1u << 31], [HAVE_BUILTIN_BSWAP32])
+CHECK_FOR_BUILTIN([__builtin_bswap64], [1ull << 63], [HAVE_BUILTIN_BSWAP64])
+
 dnl === Check for pthread support
 AC_ARG_ENABLE([threading],
               AS_HELP_STRING([--disable-threading],
@@ -106,212 +175,241 @@ if test "$enable_threading" = "yes"; then
               CFLAGS="$CFLAGS $PTHREAD_CFLAGS"
               CC="$PTHREAD_CC"
              ],
-             [enable_threading=no])
+             [AC_CHECK_FUNC([_beginthreadex],
+                            [AC_DEFINE([WEBP_USE_THREAD], [1],
+                                       [Undefine this to disable thread
+                                        support.])],
+                            [enable_threading=no])])
 fi
 AC_MSG_NOTICE([checking if threading is enabled... ${enable_threading-no}])
 
 dnl === check for OpenGL/GLUT support ===
-CLEAR_LIBVARS([GL])
-WITHLIB_OPTION([gl], [GL])
 
-LIBCHECK_PROLOGUE([GL])
+AC_ARG_ENABLE([gl], AS_HELP_STRING([--disable-gl],
+                                   [Disable detection of OpenGL support
+                                    @<:@default=auto@:>@]))
+AS_IF([test "x$enable_gl" != "xno"], [
+  CLEAR_LIBVARS([GL])
+  WITHLIB_OPTION([gl], [GL])
 
-glut_cflags="none"
-glut_ldflags="none"
-case $host_os in
-  darwin*)
-    # Special case for OSX builds. Append these to give the user a chance to
-    # override with --with-gl*
-    glut_cflags="$glut_cflags|-framework GLUT -framework OpenGL"
-    glut_ldflags="$glut_ldflags|-framework GLUT -framework OpenGL"
-    ;;
-esac
+  LIBCHECK_PROLOGUE([GL])
 
-GLUT_SAVED_CPPFLAGS="$CPPFLAGS"
-SAVED_IFS="$IFS"
-IFS="|"
-for flag in $glut_cflags; do
-  # restore IFS immediately as the autoconf macros may need the default.
-  IFS="$SAVED_IFS"
-  unset ac_cv_header_GL_glut_h
-  unset ac_cv_header_OpenGL_glut_h
-
-  case $flag in
-    none) ;;
-    *) CPPFLAGS="$flag $CPPFLAGS";;
+  glut_cflags="none"
+  glut_ldflags="none"
+  case $host_os in
+    darwin*)
+      # Special case for OSX builds. Append these to give the user a chance to
+      # override with --with-gl*
+      glut_cflags="$glut_cflags|-framework GLUT -framework OpenGL"
+      glut_ldflags="$glut_ldflags|-framework GLUT -framework OpenGL"
+      ;;
   esac
-  AC_CHECK_HEADERS([GL/glut.h GLUT/glut.h OpenGL/glut.h],
-                   [glut_headers=yes;
-                    test "$flag" = "none" || GL_INCLUDES="$CPPFLAGS";
-                    break])
-  CPPFLAGS="$GLUT_SAVED_CPPFLAGS"
-  test "$glut_headers" = "yes" && break
-done
-IFS="$SAVED_IFS"
 
-if test "$glut_headers" = "yes"; then
-  AC_LANG_PUSH([C])
-  GLUT_SAVED_LDFLAGS="$LDFLAGS"
+  GLUT_SAVED_CPPFLAGS="$CPPFLAGS"
   SAVED_IFS="$IFS"
   IFS="|"
-  for flag in $glut_ldflags; do
+  for flag in $glut_cflags; do
     # restore IFS immediately as the autoconf macros may need the default.
     IFS="$SAVED_IFS"
-    unset ac_cv_search_glBegin
+    unset ac_cv_header_GL_glut_h
+    unset ac_cv_header_OpenGL_glut_h
 
     case $flag in
       none) ;;
-      *) LDFLAGS="$flag $LDFLAGS";;
+      *) CPPFLAGS="$flag $CPPFLAGS";;
     esac
-
-    # find libGL
-    GL_SAVED_LIBS="$LIBS"
-    AC_SEARCH_LIBS([glBegin], [GL OpenGL])
-    LIBS="$GL_SAVED_LIBS"
-
-    # A direct link to libGL may not be necessary on e.g., linux.
-    GLUT_SAVED_LIBS="$LIBS"
-    for lib in "" "-lglut" "-lglut $ac_cv_search_glBegin"; do
-      LIBS="$lib"
-      AC_LINK_IFELSE(
-        [AC_LANG_PROGRAM([
-           #ifdef __cplusplus
-           # define EXTERN_C extern "C"
-           #else
-           # define EXTERN_C
-           #endif
-           EXTERN_C char glOrtho();
-           EXTERN_C char glutMainLoop();
-          ],[
-           glOrtho();
-           glutMainLoop();
-          ])
-        ],
-        AC_DEFINE(WEBP_HAVE_GL, [1],
-                  [Set to 1 if OpenGL is supported])
-        [glut_support=yes], []
-      )
-      if test "$glut_support" = "yes"; then
-        GL_LIBS="$LDFLAGS $lib"
-        break
-      fi
-    done
-    LIBS="$GLUT_SAVED_LIBS"
-    LDFLAGS="$GLUT_SAVED_LDFLAGS"
-    test "$glut_support" = "yes" && break
+    AC_CHECK_HEADERS([GL/glut.h GLUT/glut.h OpenGL/glut.h],
+                     [glut_headers=yes;
+                      test "$flag" = "none" || GL_INCLUDES="$CPPFLAGS";
+                      break])
+    CPPFLAGS="$GLUT_SAVED_CPPFLAGS"
+    test "$glut_headers" = "yes" && break
   done
   IFS="$SAVED_IFS"
-  AC_LANG_POP
-fi
 
-LIBCHECK_EPILOGUE([GL])
+  if test "$glut_headers" = "yes"; then
+    AC_LANG_PUSH([C])
+    GLUT_SAVED_LDFLAGS="$LDFLAGS"
+    SAVED_IFS="$IFS"
+    IFS="|"
+    for flag in $glut_ldflags; do
+      # restore IFS immediately as the autoconf macros may need the default.
+      IFS="$SAVED_IFS"
+      unset ac_cv_search_glBegin
 
-if test "$glut_support" = "yes" -a "$enable_libwebpdemux" = "yes"; then
-  build_vwebp=yes
-fi
+      case $flag in
+        none) ;;
+        *) LDFLAGS="$flag $LDFLAGS";;
+      esac
+
+      # find libGL
+      GL_SAVED_LIBS="$LIBS"
+      AC_SEARCH_LIBS([glBegin], [GL OpenGL opengl32])
+      LIBS="$GL_SAVED_LIBS"
+
+      # A direct link to libGL may not be necessary on e.g., linux.
+      GLUT_SAVED_LIBS="$LIBS"
+      for lib in "" "-lglut" "-lglut $ac_cv_search_glBegin"; do
+        LIBS="$lib"
+        AC_LINK_IFELSE(
+          [AC_LANG_PROGRAM([
+             #ifdef __cplusplus
+             # define EXTERN_C extern "C"
+             #else
+             # define EXTERN_C
+             #endif
+             EXTERN_C char glOrtho();
+             EXTERN_C char glutMainLoop();
+            ],[
+             glOrtho();
+             glutMainLoop();
+            ])
+          ],
+          AC_DEFINE(WEBP_HAVE_GL, [1],
+                    [Set to 1 if OpenGL is supported])
+          [glut_support=yes], []
+        )
+        if test "$glut_support" = "yes"; then
+          GL_LIBS="$LDFLAGS $lib"
+          break
+        fi
+      done
+      LIBS="$GLUT_SAVED_LIBS"
+      LDFLAGS="$GLUT_SAVED_LDFLAGS"
+      test "$glut_support" = "yes" && break
+    done
+    IFS="$SAVED_IFS"
+    AC_LANG_POP
+  fi
+
+  LIBCHECK_EPILOGUE([GL])
+
+  if test "$glut_support" = "yes" -a "$enable_libwebpdemux" = "yes"; then
+    build_vwebp=yes
+  fi
+])
 AM_CONDITIONAL([BUILD_VWEBP], [test "$build_vwebp" = "yes"])
 
 dnl === check for PNG support ===
 
-CLEAR_LIBVARS([PNG])
-AC_PATH_PROGS(LIBPNG_CONFIG,
-              [libpng-config libpng15-config libpng14-config libpng12-config])
-if test -n "$LIBPNG_CONFIG"; then
-  PNG_INCLUDES=`$LIBPNG_CONFIG --cflags`
-  PNG_PREFIX=`$LIBPNG_CONFIG --prefix`
-  if test "${PNG_PREFIX}/lib" != "/usr/lib" ; then
-    PNG_LIBS="-L${PNG_PREFIX}/lib"
+AC_ARG_ENABLE([png], AS_HELP_STRING([--disable-png],
+                                    [Disable detection of PNG format support
+                                     @<:@default=auto@:>@]))
+AS_IF([test "x$enable_png" != "xno"], [
+  CLEAR_LIBVARS([PNG])
+  AC_PATH_PROGS([LIBPNG_CONFIG],
+                [libpng-config libpng16-config libpng15-config libpng14-config \
+                 libpng12-config])
+  if test -n "$LIBPNG_CONFIG"; then
+    PNG_INCLUDES=`$LIBPNG_CONFIG --cflags`
+    PNG_LIBS="`$LIBPNG_CONFIG --ldflags`"
   fi
-  PNG_LIBS="$PNG_LIBS `$LIBPNG_CONFIG --libs`"
-fi
 
-WITHLIB_OPTION([png], [PNG])
+  WITHLIB_OPTION([png], [PNG])
 
-LIBCHECK_PROLOGUE([PNG])
-AC_CHECK_HEADER(png.h,
-  AC_SEARCH_LIBS(png_get_libpng_ver, [png],
-                 [test "$ac_cv_search_png_get_libpng_ver" = "none required" \
-                    || PNG_LIBS="$PNG_LIBS $ac_cv_search_png_get_libpng_ver"
-                  PNG_INCLUDES="$PNG_INCLUDES -DWEBP_HAVE_PNG"
-                  AC_DEFINE(WEBP_HAVE_PNG, [1],
-                            [Set to 1 if PNG library is installed])
-                  png_support=yes
-                 ],
-                 [AC_MSG_WARN(Optional png library not found)
-                  PNG_LIBS=""
-                  PNG_INCLUDES=""
-                 ],
-                 [$MATH_LIBS]),
-  [AC_MSG_WARN(png library not available - no png.h)
-   PNG_LIBS=""
-   PNG_INCLUDES=""
-  ],
-)
-LIBCHECK_EPILOGUE([PNG])
+  LIBCHECK_PROLOGUE([PNG])
+  AC_CHECK_HEADER(png.h,
+    AC_SEARCH_LIBS(png_get_libpng_ver, [png],
+                   [test "$ac_cv_search_png_get_libpng_ver" = "none required" \
+                      || PNG_LIBS="$PNG_LIBS $ac_cv_search_png_get_libpng_ver"
+                    PNG_INCLUDES="$PNG_INCLUDES -DWEBP_HAVE_PNG"
+                    AC_DEFINE(WEBP_HAVE_PNG, [1],
+                              [Set to 1 if PNG library is installed])
+                    png_support=yes
+                   ],
+                   [AC_MSG_WARN(Optional png library not found)
+                    PNG_LIBS=""
+                    PNG_INCLUDES=""
+                   ],
+                   [$MATH_LIBS]),
+    [AC_MSG_WARN(png library not available - no png.h)
+     PNG_LIBS=""
+     PNG_INCLUDES=""
+    ],
+  )
+  LIBCHECK_EPILOGUE([PNG])
+])
 
 dnl === check for JPEG support ===
 
-CLEAR_LIBVARS([JPEG])
-WITHLIB_OPTION([jpeg], [JPEG])
+AC_ARG_ENABLE([jpeg],
+              AS_HELP_STRING([--disable-jpeg],
+                             [Disable detection of JPEG format support
+                              @<:@default=auto@:>@]))
+AS_IF([test "x$enable_jpeg" != "xno"], [
+  CLEAR_LIBVARS([JPEG])
+  WITHLIB_OPTION([jpeg], [JPEG])
 
-LIBCHECK_PROLOGUE([JPEG])
-AC_CHECK_HEADER(jpeglib.h,
-  AC_CHECK_LIB(jpeg, jpeg_set_defaults,
-               [JPEG_LIBS="$JPEG_LIBS -ljpeg"
-                JPEG_INCLUDES="$JPEG_INCLUDES -DWEBP_HAVE_JPEG"
-                AC_DEFINE(WEBP_HAVE_JPEG, [1],
-                          [Set to 1 if JPEG library is installed])
-                jpeg_support=yes
-               ],
-               AC_MSG_WARN(Optional jpeg library not found),
-               [$MATH_LIBS]),
-  AC_MSG_WARN(jpeg library not available - no jpeglib.h)
-)
-LIBCHECK_EPILOGUE([JPEG])
+  LIBCHECK_PROLOGUE([JPEG])
+  AC_CHECK_HEADER(jpeglib.h,
+    AC_CHECK_LIB(jpeg, jpeg_set_defaults,
+                 [JPEG_LIBS="$JPEG_LIBS -ljpeg"
+                  JPEG_INCLUDES="$JPEG_INCLUDES -DWEBP_HAVE_JPEG"
+                  AC_DEFINE(WEBP_HAVE_JPEG, [1],
+                            [Set to 1 if JPEG library is installed])
+                  jpeg_support=yes
+                 ],
+                 AC_MSG_WARN(Optional jpeg library not found),
+                 [$MATH_LIBS]),
+    AC_MSG_WARN(jpeg library not available - no jpeglib.h)
+  )
+  LIBCHECK_EPILOGUE([JPEG])
+])
 
 dnl === check for TIFF support ===
 
-CLEAR_LIBVARS([TIFF])
-WITHLIB_OPTION([tiff], [TIFF])
+AC_ARG_ENABLE([tiff],
+              AS_HELP_STRING([--disable-tiff],
+                             [Disable detection of TIFF format support
+                              @<:@default=auto@:>@]))
+AS_IF([test "x$enable_tiff" != "xno"], [
+  CLEAR_LIBVARS([TIFF])
+  WITHLIB_OPTION([tiff], [TIFF])
 
-LIBCHECK_PROLOGUE([TIFF])
-AC_CHECK_HEADER(tiffio.h,
-  AC_CHECK_LIB(tiff, TIFFGetVersion,
-               [TIFF_LIBS="$TIFF_LIBS -ltiff"
-                TIFF_INCLUDES="$TIFF_INCLUDES -DWEBP_HAVE_TIFF"
-                AC_DEFINE(WEBP_HAVE_TIFF, [1],
-                          [Set to 1 if TIFF library is installed])
-                tiff_support=yes
-               ],
-               AC_MSG_WARN(Optional tiff library not found),
-               [$MATH_LIBS]),
-  AC_MSG_WARN(tiff library not available - no tiffio.h)
-)
-LIBCHECK_EPILOGUE([TIFF])
+  LIBCHECK_PROLOGUE([TIFF])
+  AC_CHECK_HEADER(tiffio.h,
+    AC_CHECK_LIB(tiff, TIFFGetVersion,
+                 [TIFF_LIBS="$TIFF_LIBS -ltiff"
+                  TIFF_INCLUDES="$TIFF_INCLUDES -DWEBP_HAVE_TIFF"
+                  AC_DEFINE(WEBP_HAVE_TIFF, [1],
+                            [Set to 1 if TIFF library is installed])
+                  tiff_support=yes
+                 ],
+                 AC_MSG_WARN(Optional tiff library not found),
+                 [$MATH_LIBS]),
+    AC_MSG_WARN(tiff library not available - no tiffio.h)
+  )
+  LIBCHECK_EPILOGUE([TIFF])
+])
 
 dnl === check for GIF support ===
 
-CLEAR_LIBVARS([GIF])
-WITHLIB_OPTION([gif], [GIF])
+AC_ARG_ENABLE([gif], AS_HELP_STRING([--disable-gif],
+                                    [Disable detection of GIF format support
+                                     @<:@default=auto@:>@]))
+AS_IF([test "x$enable_gif" != "xno"], [
+  CLEAR_LIBVARS([GIF])
+  WITHLIB_OPTION([gif], [GIF])
 
-LIBCHECK_PROLOGUE([GIF])
-AC_CHECK_HEADER(gif_lib.h,
-  AC_CHECK_LIB([gif], [DGifOpenFileHandle],
-               [GIF_LIBS="$GIF_LIBS -lgif"
-                AC_DEFINE(WEBP_HAVE_GIF, [1],
-                          [Set to 1 if GIF library is installed])
-                gif_support=yes
-               ],
-               AC_MSG_WARN(Optional gif library not found),
-               [$MATH_LIBS]),
-  AC_MSG_WARN(gif library not available - no gif_lib.h)
-)
-LIBCHECK_EPILOGUE([GIF])
+  LIBCHECK_PROLOGUE([GIF])
+  AC_CHECK_HEADER(gif_lib.h,
+    AC_CHECK_LIB([gif], [DGifOpenFileHandle],
+                 [GIF_LIBS="$GIF_LIBS -lgif"
+                  AC_DEFINE(WEBP_HAVE_GIF, [1],
+                            [Set to 1 if GIF library is installed])
+                  gif_support=yes
+                 ],
+                 AC_MSG_WARN(Optional gif library not found),
+                 [$MATH_LIBS]),
+    AC_MSG_WARN(gif library not available - no gif_lib.h)
+  )
+  LIBCHECK_EPILOGUE([GIF])
 
-if test "$gif_support" = "yes" -a \
-        "$enable_libwebpmux" = "yes"; then
-  build_gif2webp=yes
-fi
+  if test "$gif_support" = "yes" -a \
+          "$enable_libwebpmux" = "yes"; then
+    build_gif2webp=yes
+  fi
+])
 AM_CONDITIONAL([BUILD_GIF2WEBP], [test "${build_gif2webp}" = "yes"])
 
 dnl === check for WIC support ===
@@ -322,7 +420,9 @@ AC_ARG_ENABLE([wic],
                               @<:@default=auto@:>@]),,
               [enable_wic=yes])
 
-if test "$target_os" = "mingw32" -a "$enable_wic" = "yes"; then
+case $host_os in
+mingw*)
+if test "$enable_wic" = "yes"; then
   AC_CHECK_HEADERS([wincodec.h shlwapi.h windows.h])
   if test "$ac_cv_header_wincodec_h" = "yes"; then
     AC_MSG_CHECKING(for Windows Imaging Component support)
@@ -362,6 +462,20 @@ if test "$target_os" = "mingw32" -a "$enable_wic" = "yes"; then
     AC_MSG_RESULT(${wic_support-no})
   fi
 fi
+esac
+
+dnl === If --enable-aligned is defined, define WEBP_FORCE_ALIGNED
+
+AC_MSG_CHECKING(if --enable-aligned option is specified)
+AC_ARG_ENABLE([aligned],
+              AS_HELP_STRING([--enable-aligned],
+                             [Force aligned memory operations in non-dsp code
+                              (may be slower)]))
+if test "$enable_aligned" = "yes"; then
+  AC_DEFINE(WEBP_FORCE_ALIGNED, [1],
+            [Define to 1 to force aligned memory operations])
+fi
+AC_MSG_RESULT(${enable_aligned-no})
 
 dnl === If --enable-swap-16bit-csp is defined, add -DWEBP_SWAP_16BIT_CSP
 
@@ -416,7 +530,7 @@ AM_CONDITIONAL([BUILD_LIBWEBPDECODER], [test "$enable_libwebpdecoder" = "yes"])
 dnl =========================
 
 AC_CONFIG_MACRO_DIR([m4])
-AC_CONFIG_HEADERS([config.h])
+AC_CONFIG_HEADERS([src/webp/config.h])
 AC_CONFIG_FILES([Makefile src/Makefile man/Makefile \
                  examples/Makefile src/dec/Makefile \
                  src/enc/Makefile src/dsp/Makefile \
@@ -434,7 +548,7 @@ WebP Configuration Summary
 
 Shared libraries: ${enable_shared}
 Static libraries: ${enable_static}
-Threaded decode: ${enable_threading-no}
+Threading support: ${enable_threading-no}
 libwebp: yes
 libwebpdecoder: ${enable_libwebpdecoder-no}
 libwebpdemux: ${enable_libwebpdemux-no}

doc/webp-container-spec.txt

@@ -46,25 +46,16 @@ for:
   * **Animation.** An image may have multiple frames with pauses between them,
     making it an animation.
 
-  * **Image Fragmentation.** A single bitstream in WebP has an inherent
-    limitation for width or height of 2^14 pixels, and, when using VP8, a 512
-    KiB limit on the size of the first compressed partition. To support larger
-    images, the format supports images that are composed of multiple fragments,
-    each encoded as a separate bitstream. All fragments logically form a single
-    image: they have common metadata, color profile, etc. Image fragmentation
-    may also improve efficiency for larger images, e.g., grass can be encoded
-    differently than sky.
-
 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
 document are to be interpreted as described in [RFC 2119][].
 
+Bit numbering in chunk diagrams starts at `0` for the most significant bit
+('MSB 0') as described in [RFC 1166][].
+
 **Note:** Out of the features mentioned above, lossy compression, lossless
 compression, transparency, metadata, color profile and animation are finalized
-and are to be considered stable. On the other hand, image fragmentation is
-experimental as of now, and is open to discussion, feedback and comments.
-The same is indicated using annotation "_status: experimental_" in the relevant
-sections of this document.
+and are to be considered stable.
 
 Terminology & Basics
 ------------------------
@@ -79,7 +70,7 @@ Below are additional terms used throughout this document:
 _Reader/Writer_
 
 : Code that reads WebP files is referred to as a _reader_, while code that
-writes them is referred to as a _writer_.
+  writes them is referred to as a _writer_.
 
 _uint16_
 
@@ -101,10 +92,12 @@ _FourCC_
 _1-based_
 
 : An unsigned integer field storing values offset by `-1`. e.g., Such a field
-would store value _25_ as _24_.
+  would store value _25_ as _24_.
 
-RIFF file format
+
+RIFF File Format
 ----------------
+
 The WebP file format is based on the RIFF (resource interchange file format)
 document format.
 
@@ -144,7 +137,8 @@ _ChunkHeader('ABCD')_
 chunks that apply to any RIFF file format, while FourCCs specific to a file
 format are all lowercase. WebP does not follow this convention.
 
-WebP file header
+
+WebP File Header
 ----------------
 
      0                   1                   2                   3
@@ -164,8 +158,8 @@ WebP file header
 File Size: 32 bits (_uint32_)
 
 : The size of the file in bytes starting at offset 8. The maximum value of
-this field is 2^32 minus 10 bytes and thus the size of the whole file is at
-most 4GiB minus 2 bytes.
+  this field is 2^32 minus 10 bytes and thus the size of the whole file is at
+  most 4GiB minus 2 bytes.
 
 'WEBP': 32 bits
 
@@ -177,7 +171,8 @@ the 'WEBP' FourCC. The file SHOULD NOT contain anything after it. As the size
 of any chunk is even, the size given by the RIFF header is also even. The
 contents of individual chunks will be described in the following sections.
 
-Simple file format (lossy)
+
+Simple File Format (Lossy)
 --------------------------
 
 This layout SHOULD be used if the image requires _lossy_ encoding and does not
@@ -215,7 +210,8 @@ width and height. That is assumed to be the width and height of the canvas.
 The VP8 specification describes how to decode the image into Y'CbCr
 format. To convert to RGB, Rec. 601 SHOULD be used.
 
-Simple file format (lossless)
+
+Simple File Format (Lossless)
 -----------------------------
 
 **Note:** Older readers may not support files using the lossless format.
@@ -253,7 +249,8 @@ The current specification of the VP8L bitstream can be found at
 contains the VP8L image width and height. That is assumed to be the width
 and height of the canvas.
 
-Extended file format
+
+Extended File Format
 --------------------
 
 **Note:** Older readers may not support files using the extended format.
@@ -278,10 +275,6 @@ For a _still image_, the _image data_ consists of a single frame, whereas for
 an _animated image_, it consists of multiple frames. More details about frames
 can be found in the [Animation](#animation) section.
 
-Moreover, each frame can be fragmented or non-fragmented, as will be described
-in the [Extended WebP file header](#extended_header) section. More details about
-fragments can be found in the [Fragments](#fragments) section.
-
 All chunks SHOULD be placed in the same order as listed above. If a chunk
 appears in the wrong place, the file is invalid, but readers MAY parse the
 file, ignoring the chunks that come too late.
@@ -302,7 +295,7 @@ Extended WebP file header:
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |                      ChunkHeader('VP8X')                      |
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
-    |Rsv|I|L|E|X|A|F|                   Reserved                    |
+    |Rsv|I|L|E|X|A|R|                   Reserved                    |
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |          Canvas Width Minus One               |             ...
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
@@ -320,7 +313,7 @@ ICC profile (I): 1 bit
 Alpha (L): 1 bit
 
 : Set if any of the frames of the image contain transparency information
-("alpha").
+  ("alpha").
 
 EXIF metadata (E): 1 bit
 
@@ -333,11 +326,11 @@ XMP metadata (X): 1 bit
 Animation (A): 1 bit
 
 : Set if this is an animated image. Data in 'ANIM' and 'ANMF' chunks should be
-used to control the animation.
+  used to control the animation.
 
-Image Fragmentation (F): 1 bit _\[status: experimental\]_
+Reserved (R): 1 bit
 
-: Set if any of the frames in the image are represented by fragments.
+: SHOULD be `0`.
 
 Reserved: 24 bits
 
@@ -382,9 +375,9 @@ animation.
 Background Color: 32 bits (_uint32_)
 
 : The default background color of the canvas in \[Blue, Green, Red, Alpha\]
-byte order. This color MAY be used to fill the unused space on the canvas around
-the frames, as well as the transparent pixels of the first frame. Background
-color is also used when disposal method is `1`.
+  byte order. This color MAY be used to fill the unused space on the canvas
+  around the frames, as well as the transparent pixels of the first frame.
+  Background color is also used when disposal method is `1`.
 
 **Note**:
 
@@ -394,6 +387,9 @@ color is also used when disposal method is `1`.
   * Viewer applications SHOULD treat the background color value as a hint, and
     are not required to use it.
 
+  * The canvas is cleared at the start of each loop. The background color MAY be
+    used to achieve this.
+
 Loop Count: 16 bits (_uint16_)
 
 : The number of times to loop the animation. `0` means infinitely.
@@ -402,7 +398,6 @@ This chunk MUST appear if the _Animation_ flag in the VP8X chunk is set.
 If the _Animation_ flag is not set and this chunk is present, it
 SHOULD be ignored.
 
-
 ANMF chunk:
 
 For animated images, this chunk contains information about a _single_ frame.
@@ -445,8 +440,8 @@ Frame Height Minus One: 24 bits (_uint24_)
 Frame Duration: 24 bits (_uint24_)
 
 : The time to wait before displaying the next frame, in 1 millisecond units.
-In particular, frame duration of 0 is useful when one wants to update multiple
-areas of the canvas at once during the animation.
+  In particular, frame duration of 0 is useful when one wants to update
+  multiple areas of the canvas at once during the animation.
 
 Reserved: 6 bits
 
@@ -454,28 +449,28 @@ Reserved: 6 bits
 
 Blending method (B): 1 bit
 
-: Indicates how transparent pixels of _the current frame_ are to be blended with
-corresponding pixels of the previous canvas:
+: Indicates how transparent pixels of _the current frame_ are to be blended
+  with corresponding pixels of the previous canvas:
 
-  * `0`: Use alpha blending. After disposing of the previous frame, render the
-    current frame on the canvas using [alpha-blending](#alpha-blending). If the
-    current frame does not have an alpha channel, assume alpha value of 255,
-    effectively replacing the rectangle.
+    * `0`: Use alpha blending. After disposing of the previous frame, render the
+      current frame on the canvas using [alpha-blending](#alpha-blending). If
+      the current frame does not have an alpha channel, assume alpha value of
+      255, effectively replacing the rectangle.
 
-  * `1`: Do not blend. After disposing of the previous frame, render the
-    current frame on the canvas by overwriting the rectangle covered by the
-    current frame.
+    * `1`: Do not blend. After disposing of the previous frame, render the
+      current frame on the canvas by overwriting the rectangle covered by the
+      current frame.
 
 Disposal method (D): 1 bit
 
-: Indicates how _the current frame_ is to be treated after it has been displayed
-(before rendering the next frame) on the canvas:
+: Indicates how _the current frame_ is to be treated after it has been
+  displayed (before rendering the next frame) on the canvas:
 
-  * `0`: Do not dispose. Leave the canvas as is.
+    * `0`: Do not dispose. Leave the canvas as is.
 
-  * `1`: Dispose to background color. Fill the _rectangle_ on the canvas covered
-    by the _current frame_ with background color specified in the
-    [ANIM chunk](#anim_chunk).
+    * `1`: Dispose to background color. Fill the _rectangle_ on the canvas
+      covered by the _current frame_ with background color specified in the
+      [ANIM chunk](#anim_chunk).
 
 **Notes**:
 
@@ -506,9 +501,7 @@ Disposal method (D): 1 bit
 
 Frame Data: _Chunk Size_ - `16` bytes
 
-: For a fragmented frame, it consists of multiple [fragment chunks](#fragments).
-
-: For a non-fragmented frame, it consists of:
+: Consists of:
 
   * An optional [alpha subchunk](#alpha) for the frame.
 
@@ -519,49 +512,6 @@ Frame Data: _Chunk Size_ - `16` bytes
 **Note**: The 'ANMF' payload, _Frame Data_ above, consists of individual
 _padded_ chunks as described by the [RIFF file format](#riff-file-format).
 
-#### Fragments _\[status: experimental\]_
-
-For images that are represented by fragments, this chunk contains data for
-a single fragment. If the _Image Fragmentation Flag_ is not set, then this chunk
-SHOULD NOT be present.
-
-     0                   1                   2                   3
-     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
-    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
-    |                      ChunkHeader('FRGM')                      |
-    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
-    |                  Fragment X                   |             ...
-    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
-    ...       Fragment Y            |         Fragment Data         |
-    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
-
-Fragment X: 24 bits (_uint24_)
-
-: The X coordinate of the upper left corner of the fragment is `Fragment X * 2`
-
-Fragment Y: 24 bits (_uint24_)
-
-: The Y coordinate of the upper left corner of the fragment is `Fragment Y * 2`
-
-Fragment Data: _Chunk Size_ - `6` bytes
-
-: It contains:
-
-  * An optional [alpha subchunk](#alpha) for the fragment.
-  * The [bitstream subchunk](#bitstream-vp8vp8l) for the fragment.
-  * An optional list of [unknown chunks](#unknown-chunks).
-
-Note: The width and height of the fragment is obtained from the bitstream
-subchunk.
-
-The fragments of a frame SHOULD have the following properties:
-
-  * They collectively cover the whole frame.
-
-  * No pair of fragments have any overlapping region on the frame.
-
-  * No portion of any fragment should be located outside of the canvas.
-
 #### Alpha
 
      0                   1                   2                   3
@@ -579,20 +529,20 @@ Reserved (Rsv): 2 bits
 Pre-processing (P): 2 bits
 
 : These INFORMATIVE bits are used to signal the pre-processing that has
-been performed during compression. The decoder can use this information to
-e.g. dither the values or smooth the gradients prior to display.
+  been performed during compression. The decoder can use this information to
+  e.g. dither the values or smooth the gradients prior to display.
 
-  * `0`: no pre-processing
-  * `1`: level reduction
+    * `0`: no pre-processing
+    * `1`: level reduction
 
 Filtering method (F): 2 bits
 
 : The filtering method used:
 
-  * `0`: None.
-  * `1`: Horizontal filter.
-  * `2`: Vertical filter.
-  * `3`: Gradient filter.
+    * `0`: None.
+    * `1`: Horizontal filter.
+    * `2`: Vertical filter.
+    * `3`: Gradient filter.
 
 For each pixel, filtering is performed using the following calculations.
 Assume the alpha values surrounding the current `X` position are labeled as:
@@ -636,15 +586,15 @@ Compression method (C): 2 bits
 
 : The compression method used:
 
-  * `0`: No compression.
-  * `1`: Compressed using the WebP lossless format.
+    * `0`: No compression.
+    * `1`: Compressed using the WebP lossless format.
 
 Alpha bitstream: _Chunk Size_ - `1` bytes
 
 : Encoded alpha bitstream.
 
-This optional chunk contains encoded alpha data for this frame/fragment. A
-frame/fragment containing a 'VP8L' chunk SHOULD NOT contain this chunk.
+This optional chunk contains encoded alpha data for this frame. A frame
+containing a 'VP8L' chunk SHOULD NOT contain this chunk.
 
 **Rationale**: The transparency information is already part of the 'VP8L'
 chunk.
@@ -675,15 +625,15 @@ compression method is '0') or compressed using the lossless format
 
 #### Bitstream (VP8/VP8L)
 
-This chunk contains compressed bitstream data for a single frame/fragment.
+This chunk contains compressed bitstream data for a single frame.
 
 A bitstream chunk may be either (i) a VP8 chunk, using "VP8 " (note the
 significant fourth-character space) as its tag _or_ (ii) a VP8L chunk, using
 "VP8L" as its tag.
 
 The formats of VP8 and VP8L chunks are as described in sections
-[Simple file format (lossy)](#simple-file-format-lossy)
-and [Simple file format (lossless)](#simple-file-format-lossless) respectively.
+[Simple File Format (Lossy)](#simple-file-format-lossy)
+and [Simple File Format (Lossless)](#simple-file-format-lossless) respectively.
 
 #### Color profile
 
@@ -731,7 +681,6 @@ EXIF Metadata: _Chunk Size_ bytes
 
 : image metadata in EXIF format.
 
-
 XMP chunk:
 
      0                   1                   2                   3
@@ -762,47 +711,17 @@ A file MAY contain unknown chunks:
 
   * At the end of the file as described in [Extended WebP file
     header](#extended_header) section.
-  * At the end of FRGM and ANMF chunks as described in [Fragments](#fragments)
-    and [Animation](#animation) sections.
+  * At the end of ANMF chunks as described in the
+    [Animation](#animation) section.
 
 Readers SHOULD ignore these chunks. Writers SHOULD preserve them in their
 original order (unless they specifically intend to modify these chunks).
 
-### Assembling the Canvas from fragments/frames
+### Assembling the Canvas from frames
 
-Here we provide an overview of how a reader should assemble a canvas in case
-of a fragmented-image and in case of an animated image. The notation
-_VP8X.field_ means the field in the 'VP8X' chunk with the same description.
-
-Displaying a _fragmented image_ canvas MUST be equivalent to the following
-pseudocode: _\[status: experimental\]_
-
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-assert VP8X.flags.hasFragments
-canvas ← new black image of size VP8X.canvasWidth x VP8X.canvasHeight.
-frgm_params ← nil
-for chunk in image_data:
-    assert chunk.tag is "FRGM"
-    frgm_params.fragmentX = Fragment X
-    frgm_params.fragmentY = Fragment Y
-    for subchunk in 'Fragment Data':
-        if subchunk.tag == "ALPH":
-            assert alpha subchunks not found in 'Fragment Data' earlier
-            frgm_params.alpha = alpha_data
-        else if subchunk.tag == "VP8 " OR subchunk.tag == "VP8L":
-            assert bitstream subchunks not found in 'Fragment Data' earlier
-            frgm_params.bitstream = bitstream_data
-    frgm_params.fragmentWidth = Width extracted from bitstream subchunk
-    frgm_params.fragmentHeight = Height extracted from bitstream subchunk
-    assert VP8X.canvasWidth >=
-        frgm_params.fragmentX + frgm_params.fragmentWidth
-    assert VP8X.canvasHeight >=
-        frgm_params.fragmentY + frgm_params.fragmentHeight
-    assert fragment has the properties mentioned in "Image Fragments" section.
-    render fragment with frame_params.alpha and frame_params.bitstream on canvas
-    with top-left corner in (frgm_params.fragmentX, frgm_params.fragmentY).
-canvas contains the decoded canvas.
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Here we provide an overview of how a reader should assemble a canvas in the
+case of an animated image. The notation _VP8X.field_ means the field in the
+'VP8X' chunk with the same description.
 
 Displaying an _animated image_ canvas MUST be equivalent to the following
 pseudocode:
@@ -810,28 +729,25 @@ pseudocode:
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 assert VP8X.flags.hasAnimation
 canvas ← new image of size VP8X.canvasWidth x VP8X.canvasHeight with
-background color ANIM.background_color.
+         background color ANIM.background_color.
 loop_count ← ANIM.loopCount
 dispose_method ← ANIM.disposeMethod
 if loop_count == 0:
     loop_count = ∞
 frame_params ← nil
-for loop = 0, ..., loop_count - 1
-    assert next chunk in image_data is ANMF
-    frame_params.frameX = Frame X
-    frame_params.frameY = Frame Y
-    frame_params.frameWidth = Frame Width Minus One + 1
-    frame_params.frameHeight = Frame Height Minus One + 1
-    frame_params.frameDuration = Frame Duration
-    assert VP8X.canvasWidth >= frame_params.frameX + frame_params.frameWidth
-    assert VP8X.canvasHeight >= frame_params.frameY + frame_params.frameHeight
-    if VP8X.flags.hasFragments and first subchunk in 'Frame Data' is FRGM
-        // Fragmented frame.
-        frame_params.{bitstream,alpha} = canvas decoded from subchunks in
-                                         'Frame Data' as per the pseudocode for
-                                         _fragmented image_ above.
-    else
-        // Non-fragmented frame.
+assert next chunk in image_data is ANMF
+for loop = 0..loop_count - 1
+    clear canvas to ANIM.background_color or application defined color
+    until eof or non-ANMF chunk
+        frame_params.frameX = Frame X
+        frame_params.frameY = Frame Y
+        frame_params.frameWidth = Frame Width Minus One + 1
+        frame_params.frameHeight = Frame Height Minus One + 1
+        frame_params.frameDuration = Frame Duration
+        frame_right = frame_params.frameX + frame_params.frameWidth
+        frame_bottom = frame_params.frameY + frame_params.frameHeight
+        assert VP8X.canvasWidth >= frame_right
+        assert VP8X.canvasHeight >= frame_bottom
         for subchunk in 'Frame Data':
             if subchunk.tag == "ALPH":
                 assert alpha subchunks not found in 'Frame Data' earlier
@@ -839,14 +755,15 @@ for loop = 0, ..., loop_count - 1
             else if subchunk.tag == "VP8 " OR subchunk.tag == "VP8L":
                 assert bitstream subchunks not found in 'Frame Data' earlier
                 frame_params.bitstream = bitstream_data
-    render frame with frame_params.alpha and frame_params.bitstream on canvas
-    with top-left corner in (frame_params.frameX, frame_params.frameY), using
-    dispose method dispose_method.
-    Show the contents of the image for frame_params.frameDuration * 1ms.
-canvas contains the decoded canvas.
+        render frame with frame_params.alpha and frame_params.bitstream on
+            canvas with top-left corner at (frame_params.frameX,
+            frame_params.frameY), using dispose method dispose_method.
+        canvas contains the decoded image.
+        Show the contents of the canvas for frame_params.frameDuration * 1ms.
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-Example file layouts
+
+Example File Layouts
 --------------------
 
 A lossy encoded image with alpha may look as follows:
@@ -878,17 +795,6 @@ RIFF/WEBP
 +- XMP  (metadata)
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-A fragmented image may look as follows:
-
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-RIFF/WEBP
-+- VP8X (descriptions of features used)
-+- FRGM (fragment1 parameters + data)
-+- FRGM (fragment2 parameters + data)
-+- FRGM (fragment3 parameters + data)
-+- FRGM (fragment4 parameters + data)
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
 An animated image with EXIF metadata may look as follows:
 
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -906,4 +812,5 @@ RIFF/WEBP
 [webpllspec]: https://gerrit.chromium.org/gerrit/gitweb?p=webm/libwebp.git;a=blob;f=doc/webp-lossless-bitstream-spec.txt;hb=master
 [iccspec]: http://www.color.org/icc_specs2.xalter
 [metadata]: http://www.metadataworkinggroup.org/pdf/mwg_guidance.pdf
+[rfc 1166]: http://tools.ietf.org/html/rfc1166
 [rfc 2119]: http://tools.ietf.org/html/rfc2119

doc/webp-lossless-bitstream-spec.txt

@@ -14,6 +14,7 @@ Specification for WebP Lossless Bitstream
 
 _Jyrki Alakuijala, Ph.D., Google, Inc., 2012-06-19_
 
+Paragraphs marked as \[AMENDED\] were amended on 2014-09-16.
 
 Abstract
 --------
@@ -172,8 +173,8 @@ It should be set to 0 when all alpha values are 255 in the picture, and
 int alpha_is_used = ReadBits(1);
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-The version_number is a 3 bit code that must be discarded by the decoder
-at this time. Complying encoders write a 3-bit value 0.
+The version_number is a 3 bit code that must be set to 0. Any other value
+should be treated as an error. \[AMENDED\]
 
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 int version_number = ReadBits(3);
@@ -330,7 +331,7 @@ uint32 Select(uint32 L, uint32 T, uint32 TL) {
            abs(pGreen - GREEN(T)) + abs(pBlue - BLUE(T));
 
   // Return either left or top, the one closer to the prediction.
-  if (pL <= pT) {
+  if (pL < pT) {     // \[AMENDED\]
     return L;
   } else {
     return T;
@@ -542,6 +543,9 @@ color.
 argb = color_table[GREEN(argb)];
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
+If the index is equal or larger than color_table_size, the argb color value
+should be set to 0x00000000 (transparent black).  \[AMENDED\]
+
 When the color table is small (equal to or less than 16 colors), several
 pixels are bundled into a single pixel. The pixel bundling packs several
 (2, 4, or 8) pixels into a single pixel, reducing the image width

examples/Android.mk

@@ -0,0 +1,71 @@
+LOCAL_PATH := $(call my-dir)
+
+################################################################################
+# libexample_util
+
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := \
+    example_util.c \
+
+LOCAL_CFLAGS := $(WEBP_CFLAGS)
+LOCAL_C_INCLUDES := $(LOCAL_PATH)/../src
+
+LOCAL_MODULE := example_util
+
+include $(BUILD_STATIC_LIBRARY)
+
+################################################################################
+# cwebp
+
+include $(CLEAR_VARS)
+
+# Note: to enable jpeg/png encoding the sources from AOSP can be used with
+# minor modification to their Android.mk files.
+LOCAL_SRC_FILES := \
+    cwebp.c \
+    jpegdec.c \
+    metadata.c \
+    pngdec.c \
+    tiffdec.c \
+    webpdec.c \
+
+LOCAL_CFLAGS := $(WEBP_CFLAGS)
+LOCAL_C_INCLUDES := $(LOCAL_PATH)/../src
+LOCAL_STATIC_LIBRARIES := example_util webp
+
+LOCAL_MODULE := cwebp
+
+include $(BUILD_EXECUTABLE)
+
+################################################################################
+# dwebp
+
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := \
+    dwebp.c \
+
+LOCAL_CFLAGS := $(WEBP_CFLAGS)
+LOCAL_C_INCLUDES := $(LOCAL_PATH)/../src
+LOCAL_STATIC_LIBRARIES := example_util webp
+
+LOCAL_MODULE := dwebp
+
+include $(BUILD_EXECUTABLE)
+
+################################################################################
+# webpmux
+
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := \
+    webpmux.c \
+
+LOCAL_CFLAGS := $(WEBP_CFLAGS)
+LOCAL_C_INCLUDES := $(LOCAL_PATH)/../src
+LOCAL_STATIC_LIBRARIES := example_util webpmux webp
+
+LOCAL_MODULE := webpmux_example
+
+include $(BUILD_EXECUTABLE)

examples/Makefile.am

@@ -1,4 +1,4 @@
-AM_CPPFLAGS = -I$(top_srcdir)/src
+AM_CPPFLAGS = -I$(top_builddir)/src -I$(top_srcdir)/src
 
 bin_PROGRAMS = dwebp cwebp
 if BUILD_VWEBP
@@ -14,7 +14,7 @@ endif
 
 noinst_LTLIBRARIES = libexampleutil.la
 
-libexampleutil_la_SOURCES = example_util.c example_util.h
+libexampleutil_la_SOURCES = example_util.c example_util.h stopwatch.h
 
 dwebp_SOURCES = dwebp.c stopwatch.h
 dwebp_CPPFLAGS  = $(AM_CPPFLAGS) $(USE_EXPERIMENTAL_CODE)
@@ -25,10 +25,12 @@ cwebp_SOURCES  = cwebp.c metadata.c metadata.h stopwatch.h
 cwebp_SOURCES += jpegdec.c jpegdec.h
 cwebp_SOURCES += pngdec.c pngdec.h
 cwebp_SOURCES += tiffdec.c tiffdec.h
+cwebp_SOURCES += webpdec.c webpdec.h
 cwebp_SOURCES += wicdec.c wicdec.h
 cwebp_CPPFLAGS  = $(AM_CPPFLAGS) $(USE_EXPERIMENTAL_CODE)
 cwebp_CPPFLAGS += $(JPEG_INCLUDES) $(PNG_INCLUDES) $(TIFF_INCLUDES)
-cwebp_LDADD = ../src/libwebp.la $(JPEG_LIBS) $(PNG_LIBS) $(TIFF_LIBS)
+cwebp_LDADD  = libexampleutil.la ../src/libwebp.la
+cwebp_LDADD += $(JPEG_LIBS) $(PNG_LIBS) $(TIFF_LIBS)
 
 gif2webp_SOURCES = gif2webp.c gif2webp_util.c
 gif2webp_CPPFLAGS = $(AM_CPPFLAGS) $(USE_EXPERIMENTAL_CODE) $(GIF_INCLUDES)

examples/cwebp.c

@@ -17,17 +17,19 @@
 #include <string.h>
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
 #include "webp/encode.h"
 
+#include "./example_util.h"
 #include "./metadata.h"
 #include "./stopwatch.h"
 
 #include "./jpegdec.h"
 #include "./pngdec.h"
 #include "./tiffdec.h"
+#include "./webpdec.h"
 #include "./wicdec.h"
 
 #ifndef WEBP_DLL
@@ -93,6 +95,9 @@ static int ReadPicture(const char* const filename, WebPPicture* const pic,
   } else {
     // If no size specified, try to decode it using WIC.
     ok = ReadPictureWithWIC(filename, pic, keep_alpha, metadata);
+    if (!ok) {
+      ok = ReadWebP(filename, pic, keep_alpha, metadata);
+    }
   }
   if (!ok) {
     fprintf(stderr, "Error! Could not process file %s\n", filename);
@@ -106,26 +111,30 @@ typedef enum {
   PNG_ = 0,
   JPEG_,
   TIFF_,  // 'TIFF' clashes with libtiff
+  WEBP_,
   UNSUPPORTED
 } InputFileFormat;
 
 static InputFileFormat GetImageType(FILE* in_file) {
   InputFileFormat format = UNSUPPORTED;
-  uint32_t magic;
-  uint8_t buf[4];
+  uint32_t magic1, magic2;
+  uint8_t buf[12];
 
-  if ((fread(&buf[0], 4, 1, in_file) != 1) ||
+  if ((fread(&buf[0], 12, 1, in_file) != 1) ||
       (fseek(in_file, 0, SEEK_SET) != 0)) {
     return format;
   }
 
-  magic = ((uint32_t)buf[0] << 24) | (buf[1] << 16) | (buf[2] << 8) | buf[3];
-  if (magic == 0x89504E47U) {
+  magic1 = ((uint32_t)buf[0] << 24) | (buf[1] << 16) | (buf[2] << 8) | buf[3];
+  magic2 = ((uint32_t)buf[8] << 24) | (buf[9] << 16) | (buf[10] << 8) | buf[11];
+  if (magic1 == 0x89504E47U) {
     format = PNG_;
-  } else if (magic >= 0xFFD8FF00U && magic <= 0xFFD8FFFFU) {
+  } else if (magic1 >= 0xFFD8FF00U && magic1 <= 0xFFD8FFFFU) {
     format = JPEG_;
-  } else if (magic == 0x49492A00 || magic == 0x4D4D002A) {
+  } else if (magic1 == 0x49492A00 || magic1 == 0x4D4D002A) {
     format = TIFF_;
+  } else if (magic1 == 0x52494646 && magic2 == 0x57454250) {
+    format = WEBP_;
   }
   return format;
 }
@@ -148,6 +157,8 @@ static int ReadPicture(const char* const filename, WebPPicture* const pic,
       ok = ReadJPEG(in_file, pic, metadata);
     } else if (format == TIFF_) {
       ok = ReadTIFF(filename, pic, keep_alpha, metadata);
+    } else if (format == WEBP_) {
+      ok = ReadWebP(filename, pic, keep_alpha, metadata);
     }
   } else {
     // If image size is specified, infer it as YUV format.
@@ -266,10 +277,6 @@ static void PrintExtraInfoLossy(const WebPPicture* const pic, int short_output,
         fprintf(stderr, "             transparency:   %6d (%.1f dB)\n",
                 stats->alpha_data_size, stats->PSNR[4]);
       }
-      if (stats->layer_data_size) {
-        fprintf(stderr, "             enhancement:    %6d\n",
-                stats->layer_data_size);
-      }
       fprintf(stderr, " Residuals bytes  "
                       "|segment 1|segment 2|segment 3"
                       "|segment 4|  total\n");
@@ -298,6 +305,9 @@ static void PrintExtraInfoLossy(const WebPPicture* const pic, int short_output,
       PrintFullLosslessInfo(stats, "alpha");
     }
   }
+}
+
+static void PrintMapInfo(const WebPPicture* const pic) {
   if (pic->extra_info != NULL) {
     const int mb_w = (pic->width + 15) / 16;
     const int mb_h = (pic->height + 15) / 16;
@@ -307,18 +317,18 @@ static void PrintExtraInfoLossy(const WebPPicture* const pic, int short_output,
       for (x = 0; x < mb_w; ++x) {
         const int c = pic->extra_info[x + y * mb_w];
         if (type == 1) {   // intra4/intra16
-          printf("%c", "+."[c]);
+          fprintf(stderr, "%c", "+."[c]);
         } else if (type == 2) {    // segments
-          printf("%c", ".-*X"[c]);
+          fprintf(stderr, "%c", ".-*X"[c]);
         } else if (type == 3) {    // quantizers
-          printf("%.2d ", c);
+          fprintf(stderr, "%.2d ", c);
         } else if (type == 6 || type == 7) {
-          printf("%3d ", c);
+          fprintf(stderr, "%3d ", c);
         } else {
-          printf("0x%.2x ", c);
+          fprintf(stderr, "0x%.2x ", c);
         }
       }
-      printf("\n");
+      fprintf(stderr, "\n");
     }
   }
 }
@@ -531,9 +541,8 @@ static int WriteWebPWithMetadata(FILE* const out,
 //------------------------------------------------------------------------------
 
 static int ProgressReport(int percent, const WebPPicture* const picture) {
-  printf("[%s]: %3d %%      \r",
-         (char*)picture->user_data, percent);
-  fflush(stdout);
+  fprintf(stderr, "[%s]: %3d %%      \r",
+          (char*)picture->user_data, percent);
   return 1;  // all ok
 }
 
@@ -550,35 +559,39 @@ static void HelpShort(void) {
 static void HelpLong(void) {
   printf("Usage:\n");
   printf(" cwebp [-preset <...>] [options] in_file [-o out_file]\n\n");
-  printf("If input size (-s) for an image is not specified, "
-         "it is assumed to be a PNG, JPEG or TIFF file.\n");
+  printf("If input size (-s) for an image is not specified, it is\n"
+         "assumed to be a PNG, JPEG, TIFF or WebP file.\n");
 #ifdef HAVE_WINCODEC_H
-  printf("Windows builds can take as input any of the files handled by WIC\n");
+  printf("Windows builds can take as input any of the files handled by WIC.\n");
 #endif
-  printf("options:\n");
+  printf("\nOptions:\n");
   printf("  -h / -help  ............ short help\n");
   printf("  -H / -longhelp  ........ long help\n");
   printf("  -q <float> ............. quality factor (0:small..100:big)\n");
-  printf("  -alpha_q <int> ......... Transparency-compression quality "
-         "(0..100).\n");
-  printf("  -preset <string> ....... Preset setting, one of:\n");
+  printf("  -alpha_q <int> ......... transparency-compression quality "
+         "(0..100)\n");
+  printf("  -preset <string> ....... preset setting, one of:\n");
   printf("                            default, photo, picture,\n");
   printf("                            drawing, icon, text\n");
-  printf("     -preset must come first, as it overwrites other parameters.");
+  printf("     -preset must come first, as it overwrites other parameters\n");
+#if WEBP_ENCODER_ABI_VERSION > 0x0202
+  printf("  -z <int> ............... activates lossless preset with given\n"
+         "                           level in [0:fast, ..., 9:slowest]\n");
+#endif
   printf("\n");
   printf("  -m <int> ............... compression method (0=fast, 6=slowest)\n");
   printf("  -segments <int> ........ number of segments to use (1..4)\n");
-  printf("  -size <int> ............ Target size (in bytes)\n");
-  printf("  -psnr <float> .......... Target PSNR (in dB. typically: 42)\n");
+  printf("  -size <int> ............ target size (in bytes)\n");
+  printf("  -psnr <float> .......... target PSNR (in dB. typically: 42)\n");
   printf("\n");
-  printf("  -s <int> <int> ......... Input size (width x height) for YUV\n");
-  printf("  -sns <int> ............. Spatial Noise Shaping (0:off, 100:max)\n");
+  printf("  -s <int> <int> ......... input size (width x height) for YUV\n");
+  printf("  -sns <int> ............. spatial noise shaping (0:off, 100:max)\n");
   printf("  -f <int> ............... filter strength (0=off..100)\n");
   printf("  -sharpness <int> ....... "
          "filter sharpness (0:most .. 7:least sharp)\n");
   printf("  -strong ................ use strong filter instead "
-                                     "of simple (default).\n");
-  printf("  -nostrong .............. use simple filter instead of strong.\n");
+                                     "of simple (default)\n");
+  printf("  -nostrong .............. use simple filter instead of strong\n");
   printf("  -partition_limit <int> . limit quality to fit the 512k limit on\n");
   printf("                           "
          "the first partition (0=no degradation ... 100=full)\n");
@@ -587,26 +600,23 @@ static void HelpLong(void) {
   printf("  -resize <w> <h> ........ resize picture (after any cropping)\n");
   printf("  -mt .................... use multi-threading if available\n");
   printf("  -low_memory ............ reduce memory usage (slower encoding)\n");
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-  printf("  -444 / -422 / -gray ..... Change colorspace\n");
-#endif
-  printf("  -map <int> ............. print map of extra info.\n");
-  printf("  -print_psnr ............ prints averaged PSNR distortion.\n");
-  printf("  -print_ssim ............ prints averaged SSIM distortion.\n");
-  printf("  -print_lsim ............ prints local-similarity distortion.\n");
-  printf("  -d <file.pgm> .......... dump the compressed output (PGM file).\n");
-  printf("  -alpha_method <int> .... Transparency-compression method (0..1)\n");
-  printf("  -alpha_filter <string> . predictive filtering for alpha plane.\n");
-  printf("                           One of: none, fast (default) or best.\n");
-  printf("  -alpha_cleanup ......... Clean RGB values in transparent area.\n");
-  printf("  -blend_alpha <hex> ..... Blend colors against background color\n"
+  printf("  -map <int> ............. print map of extra info\n");
+  printf("  -print_psnr ............ prints averaged PSNR distortion\n");
+  printf("  -print_ssim ............ prints averaged SSIM distortion\n");
+  printf("  -print_lsim ............ prints local-similarity distortion\n");
+  printf("  -d <file.pgm> .......... dump the compressed output (PGM file)\n");
+  printf("  -alpha_method <int> .... transparency-compression method (0..1)\n");
+  printf("  -alpha_filter <string> . predictive filtering for alpha plane,\n");
+  printf("                           one of: none, fast (default) or best\n");
+  printf("  -alpha_cleanup ......... clean RGB values in transparent area\n");
+  printf("  -blend_alpha <hex> ..... blend colors against background color\n"
          "                           expressed as RGB values written in\n"
          "                           hexadecimal, e.g. 0xc0e0d0 for red=0xc0\n"
-         "                           green=0xe0 and blue=0xd0.\n");
-  printf("  -noalpha ............... discard any transparency information.\n");
-  printf("  -lossless .............. Encode image losslessly.\n");
-  printf("  -hint <string> ......... Specify image characteristics hint.\n");
-  printf("                           One of: photo, picture or graph\n");
+         "                           green=0xe0 and blue=0xd0\n");
+  printf("  -noalpha ............... discard any transparency information\n");
+  printf("  -lossless .............. encode image losslessly\n");
+  printf("  -hint <string> ......... specify image characteristics hint,\n");
+  printf("                           one of: photo, picture or graph\n");
 
   printf("\n");
   printf("  -metadata <string> ..... comma separated list of metadata to\n");
@@ -617,18 +627,18 @@ static void HelpLong(void) {
 
   printf("\n");
   printf("  -short ................. condense printed message\n");
-  printf("  -quiet ................. don't print anything.\n");
-  printf("  -version ............... print version number and exit.\n");
+  printf("  -quiet ................. don't print anything\n");
+  printf("  -version ............... print version number and exit\n");
 #ifndef WEBP_DLL
-  printf("  -noasm ................. disable all assembly optimizations.\n");
+  printf("  -noasm ................. disable all assembly optimizations\n");
 #endif
   printf("  -v ..................... verbose, e.g. print encoding/decoding "
          "times\n");
   printf("  -progress .............. report encoding progress\n");
   printf("\n");
   printf("Experimental Options:\n");
-  printf("  -jpeg_like ............. Roughly match expected JPEG size.\n");
-  printf("  -af .................... auto-adjust filter strength.\n");
+  printf("  -jpeg_like ............. roughly match expected JPEG size\n");
+  printf("  -af .................... auto-adjust filter strength\n");
   printf("  -pre <int> ............. pre-processing filter\n");
   printf("\n");
 }
@@ -636,7 +646,7 @@ static void HelpLong(void) {
 //------------------------------------------------------------------------------
 // Error messages
 
-static const char* const kErrorMessages[] = {
+static const char* const kErrorMessages[VP8_ENC_ERROR_LAST] = {
   "OK",
   "OUT_OF_MEMORY: Out of memory allocating objects",
   "BITSTREAM_OUT_OF_MEMORY: Out of memory re-allocating byte buffer",
@@ -669,6 +679,10 @@ int main(int argc, const char *argv[]) {
   uint32_t background_color = 0xffffffu;
   int crop = 0, crop_x = 0, crop_y = 0, crop_w = 0, crop_h = 0;
   int resize_w = 0, resize_h = 0;
+#if WEBP_ENCODER_ABI_VERSION > 0x0202
+  int lossless_preset = 6;
+  int use_lossless_preset = -1;  // -1=unset, 0=don't use, 1=use it
+#endif
   int show_progress = 0;
   int keep_metadata = 0;
   int metadata_written = 0;
@@ -696,6 +710,7 @@ int main(int argc, const char *argv[]) {
   }
 
   for (c = 1; c < argc; ++c) {
+    int parse_error = 0;
     if (!strcmp(argv[c], "-h") || !strcmp(argv[c], "-help")) {
       HelpShort();
       return 0;
@@ -717,23 +732,33 @@ int main(int argc, const char *argv[]) {
       config.show_compressed = 1;
       print_distortion = 2;
     } else if (!strcmp(argv[c], "-short")) {
-      short_output++;
+      ++short_output;
     } else if (!strcmp(argv[c], "-s") && c < argc - 2) {
-      picture.width = strtol(argv[++c], NULL, 0);
-      picture.height = strtol(argv[++c], NULL, 0);
+      picture.width = ExUtilGetInt(argv[++c], 0, &parse_error);
+      picture.height = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-m") && c < argc - 1) {
-      config.method = strtol(argv[++c], NULL, 0);
+      config.method = ExUtilGetInt(argv[++c], 0, &parse_error);
+#if WEBP_ENCODER_ABI_VERSION > 0x0202
+      use_lossless_preset = 0;   // disable -z option
+#endif
     } else if (!strcmp(argv[c], "-q") && c < argc - 1) {
-      config.quality = (float)strtod(argv[++c], NULL);
+      config.quality = ExUtilGetFloat(argv[++c], &parse_error);
+#if WEBP_ENCODER_ABI_VERSION > 0x0202
+      use_lossless_preset = 0;   // disable -z option
+    } else if (!strcmp(argv[c], "-z") && c < argc - 1) {
+      lossless_preset = ExUtilGetInt(argv[++c], 0, &parse_error);
+      if (use_lossless_preset != 0) use_lossless_preset = 1;
+#endif
     } else if (!strcmp(argv[c], "-alpha_q") && c < argc - 1) {
-      config.alpha_quality = strtol(argv[++c], NULL, 0);
+      config.alpha_quality = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-alpha_method") && c < argc - 1) {
-      config.alpha_compression = strtol(argv[++c], NULL, 0);
+      config.alpha_compression = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-alpha_cleanup")) {
       keep_alpha = keep_alpha ? 2 : 0;
     } else if (!strcmp(argv[c], "-blend_alpha") && c < argc - 1) {
       blend_alpha = 1;
-      background_color = strtol(argv[++c], NULL, 16);  // <- parses '0x' prefix
+      // background color is given in hex with an optional '0x' prefix
+      background_color = ExUtilGetInt(argv[++c], 16, &parse_error);
       background_color = background_color & 0x00ffffffu;
     } else if (!strcmp(argv[c], "-alpha_filter") && c < argc - 1) {
       ++c;
@@ -764,13 +789,13 @@ int main(int argc, const char *argv[]) {
         goto Error;
       }
     } else if (!strcmp(argv[c], "-size") && c < argc - 1) {
-      config.target_size = strtol(argv[++c], NULL, 0);
+      config.target_size = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-psnr") && c < argc - 1) {
-      config.target_PSNR = (float)strtod(argv[++c], NULL);
+      config.target_PSNR = ExUtilGetFloat(argv[++c], &parse_error);
     } else if (!strcmp(argv[c], "-sns") && c < argc - 1) {
-      config.sns_strength = strtol(argv[++c], NULL, 0);
+      config.sns_strength = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-f") && c < argc - 1) {
-      config.filter_strength = strtol(argv[++c], NULL, 0);
+      config.filter_strength = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-af")) {
       config.autofilter = 1;
     } else if (!strcmp(argv[c], "-jpeg_like")) {
@@ -784,34 +809,26 @@ int main(int argc, const char *argv[]) {
     } else if (!strcmp(argv[c], "-nostrong")) {
       config.filter_type = 0;
     } else if (!strcmp(argv[c], "-sharpness") && c < argc - 1) {
-      config.filter_sharpness = strtol(argv[++c], NULL, 0);
+      config.filter_sharpness = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-pass") && c < argc - 1) {
-      config.pass = strtol(argv[++c], NULL, 0);
+      config.pass = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-pre") && c < argc - 1) {
-      config.preprocessing = strtol(argv[++c], NULL, 0);
+      config.preprocessing = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-segments") && c < argc - 1) {
-      config.segments = strtol(argv[++c], NULL, 0);
+      config.segments = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-partition_limit") && c < argc - 1) {
-      config.partition_limit = strtol(argv[++c], NULL, 0);
+      config.partition_limit = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-map") && c < argc - 1) {
-      picture.extra_info_type = strtol(argv[++c], NULL, 0);
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-    } else if (!strcmp(argv[c], "-444")) {
-      picture.colorspace = WEBP_YUV444;
-    } else if (!strcmp(argv[c], "-422")) {
-      picture.colorspace = WEBP_YUV422;
-    } else if (!strcmp(argv[c], "-gray")) {
-      picture.colorspace = WEBP_YUV400;
-#endif
+      picture.extra_info_type = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-crop") && c < argc - 4) {
       crop = 1;
-      crop_x = strtol(argv[++c], NULL, 0);
-      crop_y = strtol(argv[++c], NULL, 0);
-      crop_w = strtol(argv[++c], NULL, 0);
-      crop_h = strtol(argv[++c], NULL, 0);
+      crop_x = ExUtilGetInt(argv[++c], 0, &parse_error);
+      crop_y = ExUtilGetInt(argv[++c], 0, &parse_error);
+      crop_w = ExUtilGetInt(argv[++c], 0, &parse_error);
+      crop_h = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-resize") && c < argc - 2) {
-      resize_w = strtol(argv[++c], NULL, 0);
-      resize_h = strtol(argv[++c], NULL, 0);
+      resize_w = ExUtilGetInt(argv[++c], 0, &parse_error);
+      resize_h = ExUtilGetInt(argv[++c], 0, &parse_error);
 #ifndef WEBP_DLL
     } else if (!strcmp(argv[c], "-noasm")) {
       VP8GetCPUInfo = NULL;
@@ -906,6 +923,11 @@ int main(int argc, const char *argv[]) {
     } else {
       in_file = argv[c];
     }
+
+    if (parse_error) {
+      HelpLong();
+      return -1;
+    }
   }
   if (in_file == NULL) {
     fprintf(stderr, "No input file specified!\n");
@@ -913,6 +935,15 @@ int main(int argc, const char *argv[]) {
     goto Error;
   }
 
+#if WEBP_ENCODER_ABI_VERSION > 0x0202
+  if (use_lossless_preset == 1) {
+    if (!WebPConfigLosslessPreset(&config, lossless_preset)) {
+      fprintf(stderr, "Invalid lossless preset (-z %d)\n", lossless_preset);
+      goto Error;
+    }
+  }
+#endif
+
   // Check for unsupported command line options for lossless mode and log
   // warning for such options.
   if (!quiet && config.lossless == 1) {
@@ -956,8 +987,9 @@ int main(int argc, const char *argv[]) {
   }
 
   // Open the output
-  if (out_file) {
-    out = fopen(out_file, "wb");
+  if (out_file != NULL) {
+    const int use_stdout = !strcmp(out_file, "-");
+    out = use_stdout ? ExUtilSetBinaryMode(stdout) : fopen(out_file, "wb");
     if (out == NULL) {
       fprintf(stderr, "Error! Cannot open output file '%s'\n", out_file);
       goto Error;
@@ -1058,47 +1090,60 @@ int main(int argc, const char *argv[]) {
   }
 
   if (!quiet) {
-    if (config.lossless) {
-      PrintExtraInfoLossless(&picture, short_output, in_file);
-    } else {
-      PrintExtraInfoLossy(&picture, short_output, config.low_memory, in_file);
+    if (!short_output || print_distortion < 0) {
+      if (config.lossless) {
+        PrintExtraInfoLossless(&picture, short_output, in_file);
+      } else {
+        PrintExtraInfoLossy(&picture, short_output, config.low_memory, in_file);
+      }
+    }
+    if (!short_output && picture.extra_info_type > 0) {
+      PrintMapInfo(&picture);
+    }
+    if (print_distortion >= 0) {    // print distortion
+      static const char* distortion_names[] = { "PSNR", "SSIM", "LSIM" };
+      float values[5];
+      // Comparison is performed in YUVA colorspace.
+      if (original_picture.use_argb &&
+          !WebPPictureARGBToYUVA(&original_picture, WEBP_YUV420A)) {
+       fprintf(stderr, "Error while converting original picture to YUVA.\n");
+        goto Error;
+      }
+      if (picture.use_argb &&
+          !WebPPictureARGBToYUVA(&picture, WEBP_YUV420A)) {
+        fprintf(stderr, "Error while converting compressed picture to YUVA.\n");
+        goto Error;
+      }
+      if (!WebPPictureDistortion(&picture, &original_picture,
+                                 print_distortion, values)) {
+        fprintf(stderr, "Error while computing the distortion.\n");
+        goto Error;
+      }
+      if (!short_output) {
+        fprintf(stderr, "%s: Y:%.2f U:%.2f V:%.2f A:%.2f  Total:%.2f\n",
+                distortion_names[print_distortion],
+                values[0], values[1], values[2], values[3], values[4]);
+      } else {
+        fprintf(stderr, "%7d %.4f\n", picture.stats->coded_size, values[4]);
+      }
     }
     if (!short_output) {
       PrintMetadataInfo(&metadata, metadata_written);
     }
   }
-  if (!quiet && !short_output && print_distortion >= 0) {  // print distortion
-    static const char* distortion_names[] = { "PSNR", "SSIM", "LSIM" };
-    float values[5];
-    // Comparison is performed in YUVA colorspace.
-    if (original_picture.use_argb &&
-        !WebPPictureARGBToYUVA(&original_picture, WEBP_YUV420A)) {
-      fprintf(stderr, "Error while converting original picture to YUVA.\n");
-      goto Error;
-    }
-    if (picture.use_argb &&
-        !WebPPictureARGBToYUVA(&picture, WEBP_YUV420A)) {
-      fprintf(stderr, "Error while converting compressed picture to YUVA.\n");
-      goto Error;
-    }
-    if (!WebPPictureDistortion(&picture, &original_picture,
-                               print_distortion, values)) {
-      fprintf(stderr, "Error while computing the distortion.\n");
-      goto Error;
-    }
-    fprintf(stderr, "%s: Y:%.2f U:%.2f V:%.2f A:%.2f  Total:%.2f\n",
-            distortion_names[print_distortion],
-            values[0], values[1], values[2], values[3], values[4]);
-  }
   return_value = 0;
 
  Error:
+#if WEBP_ENCODER_ABI_VERSION > 0x0203
+  WebPMemoryWriterClear(&memory_writer);
+#else
   free(memory_writer.mem);
+#endif
   free(picture.extra_info);
   MetadataFree(&metadata);
   WebPPictureFree(&picture);
   WebPPictureFree(&original_picture);
-  if (out != NULL) {
+  if (out != NULL && out != stdout) {
     fclose(out);
   }
 

examples/dwebp.c

@@ -17,11 +17,12 @@
 #include <string.h>
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
 #ifdef WEBP_HAVE_PNG
 #include <png.h>
+#include <setjmp.h>   // note: this must be included *after* png.h
 #endif
 
 #ifdef HAVE_WINCODEC_H
@@ -38,11 +39,6 @@
 #include <wincodec.h>
 #endif
 
-#if defined(_WIN32)
-#include <fcntl.h>   // for _O_BINARY
-#include <io.h>      // for _setmode()
-#endif
-
 #include "webp/decode.h"
 #include "./example_util.h"
 #include "./stopwatch.h"
@@ -197,8 +193,8 @@ static int WritePNG(FILE* out_file, const WebPDecBuffer* const buffer) {
   uint8_t* const rgb = buffer->u.RGBA.rgba;
   const int stride = buffer->u.RGBA.stride;
   const int has_alpha = (buffer->colorspace == MODE_RGBA);
-  png_structp png;
-  png_infop info;
+  volatile png_structp png;
+  volatile png_infop info;
   png_uint_32 y;
 
   png = png_create_write_struct(PNG_LIBPNG_VER_STRING,
@@ -208,11 +204,11 @@ static int WritePNG(FILE* out_file, const WebPDecBuffer* const buffer) {
   }
   info = png_create_info_struct(png);
   if (info == NULL) {
-    png_destroy_write_struct(&png, NULL);
+    png_destroy_write_struct((png_structpp)&png, NULL);
     return 0;
   }
   if (setjmp(png_jmpbuf(png))) {
-    png_destroy_write_struct(&png, &info);
+    png_destroy_write_struct((png_structpp)&png, (png_infopp)&info);
     return 0;
   }
   png_init_io(png, out_file);
@@ -226,7 +222,7 @@ static int WritePNG(FILE* out_file, const WebPDecBuffer* const buffer) {
     png_write_rows(png, &row, 1);
   }
   png_write_end(png, info);
-  png_destroy_write_struct(&png, &info);
+  png_destroy_write_struct((png_structpp)&png, (png_infopp)&info);
   return 1;
 }
 #else    // !HAVE_WINCODEC_H && !WEBP_HAVE_PNG
@@ -249,10 +245,10 @@ static int WritePPM(FILE* fout, const WebPDecBuffer* const buffer, int alpha) {
   uint32_t y;
 
   if (alpha) {
-    fprintf(fout, "P7\nWIDTH %d\nHEIGHT %d\nDEPTH 4\nMAXVAL 255\n"
+    fprintf(fout, "P7\nWIDTH %u\nHEIGHT %u\nDEPTH 4\nMAXVAL 255\n"
                   "TUPLTYPE RGB_ALPHA\nENDHDR\n", width, height);
   } else {
-    fprintf(fout, "P6\n%d %d\n255\n", width, height);
+    fprintf(fout, "P6\n%u %u\n255\n", width, height);
   }
   for (y = 0; y < height; ++y) {
     if (fwrite(rgb + y * stride, width, bytes_per_px, fout) != bytes_per_px) {
@@ -409,7 +405,7 @@ static int WriteAlphaPlane(FILE* fout, const WebPDecBuffer* const buffer) {
   const int a_stride = buffer->u.YUVA.a_stride;
   uint32_t y;
   assert(a != NULL);
-  fprintf(fout, "P5\n%d %d\n255\n", width, height);
+  fprintf(fout, "P5\n%u %u\n255\n", width, height);
   for (y = 0; y < height; ++y) {
     if (fwrite(a + y * a_stride, width, 1, fout) != 1) {
       return 0;
@@ -482,15 +478,8 @@ static int SaveOutput(const WebPDecBuffer* const buffer,
   needs_open_file = (format != PNG);
 #endif
 
-#if defined(_WIN32)
-  if (use_stdout && _setmode(_fileno(stdout), _O_BINARY) == -1) {
-    fprintf(stderr, "Failed to reopen stdout in O_BINARY mode.\n");
-    return -1;
-  }
-#endif
-
   if (needs_open_file) {
-    fout = use_stdout ? stdout : fopen(out_file, "wb");
+    fout = use_stdout ? ExUtilSetBinaryMode(stdout) : fopen(out_file, "wb");
     if (fout == NULL) {
       fprintf(stderr, "Error opening output file %s\n", out_file);
       return 0;
@@ -552,29 +541,30 @@ static void Help(void) {
          "  -yuv ......... save the raw YUV samples in flat layout\n"
          "\n"
          " Other options are:\n"
-         "  -version  .... print version number and exit.\n"
-         "  -nofancy ..... don't use the fancy YUV420 upscaler.\n"
-         "  -nofilter .... disable in-loop filtering.\n"
-         "  -nodither .... disable dithering.\n"
+         "  -version  .... print version number and exit\n"
+         "  -nofancy ..... don't use the fancy YUV420 upscaler\n"
+         "  -nofilter .... disable in-loop filtering\n"
+         "  -nodither .... disable dithering\n"
          "  -dither <d> .. dithering strength (in 0..100)\n"
+#if WEBP_DECODER_ABI_VERSION > 0x0204
+         "  -alpha_dither  use alpha-plane dithering if needed\n"
+#endif
          "  -mt .......... use multi-threading\n"
          "  -crop <x> <y> <w> <h> ... crop output with the given rectangle\n"
          "  -scale <w> <h> .......... scale the output (*after* any cropping)\n"
-         "  -alpha ....... only save the alpha plane.\n"
+#if WEBP_DECODER_ABI_VERSION > 0x0203
+         "  -flip ........ flip the output vertically\n"
+#endif
+         "  -alpha ....... only save the alpha plane\n"
          "  -incremental . use incremental decoding (useful for tests)\n"
-         "  -h     ....... this help message.\n"
+         "  -h     ....... this help message\n"
          "  -v     ....... verbose (e.g. print encoding/decoding times)\n"
 #ifndef WEBP_DLL
-         "  -noasm ....... disable all assembly optimizations.\n"
+         "  -noasm ....... disable all assembly optimizations\n"
 #endif
         );
 }
 
-static const char* const kStatusMessages[] = {
-  "OK", "OUT_OF_MEMORY", "INVALID_PARAM", "BITSTREAM_ERROR",
-  "UNSUPPORTED_FEATURE", "SUSPENDED", "USER_ABORT", "NOT_ENOUGH_DATA"
-};
-
 static const char* const kFormatType[] = {
   "unspecified", "lossy", "lossless"
 };
@@ -597,6 +587,7 @@ int main(int argc, const char *argv[]) {
   }
 
   for (c = 1; c < argc; ++c) {
+    int parse_error = 0;
     if (!strcmp(argv[c], "-h") || !strcmp(argv[c], "-help")) {
       Help();
       return 0;
@@ -627,20 +618,29 @@ int main(int argc, const char *argv[]) {
       format = YUV;
     } else if (!strcmp(argv[c], "-mt")) {
       config.options.use_threads = 1;
+#if WEBP_DECODER_ABI_VERSION > 0x0204
+    } else if (!strcmp(argv[c], "-alpha_dither")) {
+      config.options.alpha_dithering_strength = 100;
+#endif
     } else if (!strcmp(argv[c], "-nodither")) {
       config.options.dithering_strength = 0;
     } else if (!strcmp(argv[c], "-dither") && c < argc - 1) {
-      config.options.dithering_strength = strtol(argv[++c], NULL, 0);
+      config.options.dithering_strength =
+          ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-crop") && c < argc - 4) {
       config.options.use_cropping = 1;
-      config.options.crop_left   = strtol(argv[++c], NULL, 0);
-      config.options.crop_top    = strtol(argv[++c], NULL, 0);
-      config.options.crop_width  = strtol(argv[++c], NULL, 0);
-      config.options.crop_height = strtol(argv[++c], NULL, 0);
+      config.options.crop_left   = ExUtilGetInt(argv[++c], 0, &parse_error);
+      config.options.crop_top    = ExUtilGetInt(argv[++c], 0, &parse_error);
+      config.options.crop_width  = ExUtilGetInt(argv[++c], 0, &parse_error);
+      config.options.crop_height = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-scale") && c < argc - 2) {
       config.options.use_scaling = 1;
-      config.options.scaled_width  = strtol(argv[++c], NULL, 0);
-      config.options.scaled_height = strtol(argv[++c], NULL, 0);
+      config.options.scaled_width  = ExUtilGetInt(argv[++c], 0, &parse_error);
+      config.options.scaled_height = ExUtilGetInt(argv[++c], 0, &parse_error);
+#if WEBP_DECODER_ABI_VERSION > 0x0203
+    } else if (!strcmp(argv[c], "-flip")) {
+      config.options.flip = 1;
+#endif
     } else if (!strcmp(argv[c], "-v")) {
       verbose = 1;
 #ifndef WEBP_DLL
@@ -659,6 +659,11 @@ int main(int argc, const char *argv[]) {
     } else {
       in_file = argv[c];
     }
+
+    if (parse_error) {
+      Help();
+      return -1;
+    }
   }
 
   if (in_file == NULL) {
@@ -668,27 +673,11 @@ int main(int argc, const char *argv[]) {
   }
 
   {
-    Stopwatch stop_watch;
     VP8StatusCode status = VP8_STATUS_OK;
     size_t data_size = 0;
     const uint8_t* data = NULL;
-
-    if (!ExUtilReadFile(in_file, &data, &data_size)) return -1;
-
-    if (verbose) {
-      StopwatchReset(&stop_watch);
-    }
-
-    status = WebPGetFeatures(data, data_size, bitstream);
-    if (status != VP8_STATUS_OK) {
-      goto end;
-    }
-
-    if (bitstream->has_animation) {
-      fprintf(stderr,
-              "Error! Decoding of an animated WebP file is not supported.\n"
-              "       Use webpmux to extract the individual frames or\n"
-              "       vwebp to view this image.\n");
+    if (!ExUtilLoadWebP(in_file, &data, &data_size, bitstream)) {
+      return -1;
     }
 
     switch (format) {
@@ -724,31 +713,16 @@ int main(int argc, const char *argv[]) {
         return -1;
     }
 
-    // Decoding call.
-    if (!incremental) {
-      status = WebPDecode(data, data_size, &config);
+    if (incremental) {
+      status = ExUtilDecodeWebPIncremental(data, data_size, verbose, &config);
     } else {
-      WebPIDecoder* const idec = WebPIDecode(data, data_size, &config);
-      if (idec == NULL) {
-        fprintf(stderr, "Failed during WebPINewDecoder().\n");
-        status = VP8_STATUS_OUT_OF_MEMORY;
-        goto end;
-      } else {
-        status = WebPIUpdate(idec, data, data_size);
-        WebPIDelete(idec);
-      }
+      status = ExUtilDecodeWebP(data, data_size, verbose, &config);
     }
 
-    if (verbose) {
-      const double decode_time = StopwatchReadAndReset(&stop_watch);
-      fprintf(stderr, "Time to decode picture: %.3fs\n", decode_time);
-    }
- end:
     free((void*)data);
     ok = (status == VP8_STATUS_OK);
     if (!ok) {
-      fprintf(stderr, "Decoding of %s failed.\n", in_file);
-      fprintf(stderr, "Status: %d (%s)\n", status, kStatusMessages[status]);
+      ExUtilPrintWebPError(in_file, status);
       goto Exit;
     }
   }

examples/example_util.c

@@ -11,20 +11,104 @@
 //
 
 #include "./example_util.h"
+
+#if defined(_WIN32)
+#include <fcntl.h>   // for _O_BINARY
+#include <io.h>      // for _setmode()
+#endif
 #include <stdio.h>
 #include <stdlib.h>
+#include <string.h>
+
+#include "webp/decode.h"
+#include "./stopwatch.h"
+
+//------------------------------------------------------------------------------
+// String parsing
+
+uint32_t ExUtilGetUInt(const char* const v, int base, int* const error) {
+  char* end = NULL;
+  const uint32_t n = (v != NULL) ? (uint32_t)strtoul(v, &end, base) : 0u;
+  if (end == v && error != NULL && !*error) {
+    *error = 1;
+    fprintf(stderr, "Error! '%s' is not an integer.\n",
+            (v != NULL) ? v : "(null)");
+  }
+  return n;
+}
+
+int ExUtilGetInt(const char* const v, int base, int* const error) {
+  return (int)ExUtilGetUInt(v, base, error);
+}
+
+float ExUtilGetFloat(const char* const v, int* const error) {
+  char* end = NULL;
+  const float f = (v != NULL) ? (float)strtod(v, &end) : 0.f;
+  if (end == v && error != NULL && !*error) {
+    *error = 1;
+    fprintf(stderr, "Error! '%s' is not a floating point number.\n",
+            (v != NULL) ? v : "(null)");
+  }
+  return f;
+}
 
 // -----------------------------------------------------------------------------
 // File I/O
 
+FILE* ExUtilSetBinaryMode(FILE* file) {
+#if defined(_WIN32)
+  if (_setmode(_fileno(file), _O_BINARY) == -1) {
+    fprintf(stderr, "Failed to reopen file in O_BINARY mode.\n");
+    return NULL;
+  }
+#endif
+  return file;
+}
+
+int ExUtilReadFromStdin(const uint8_t** data, size_t* data_size) {
+  static const size_t kBlockSize = 16384;  // default initial size
+  size_t max_size = 0;
+  size_t size = 0;
+  uint8_t* input = NULL;
+
+  if (data == NULL || data_size == NULL) return 0;
+  *data = NULL;
+  *data_size = 0;
+
+  if (!ExUtilSetBinaryMode(stdin)) return 0;
+
+  while (!feof(stdin)) {
+    // We double the buffer size each time and read as much as possible.
+    const size_t extra_size = (max_size == 0) ? kBlockSize : max_size;
+    void* const new_data = realloc(input, max_size + extra_size);
+    if (new_data == NULL) goto Error;
+    input = (uint8_t*)new_data;
+    max_size += extra_size;
+    size += fread(input + size, 1, extra_size, stdin);
+    if (size < max_size) break;
+  }
+  if (ferror(stdin)) goto Error;
+  *data = input;
+  *data_size = size;
+  return 1;
+
+ Error:
+  free(input);
+  fprintf(stderr, "Could not read from stdin\n");
+  return 0;
+}
+
 int ExUtilReadFile(const char* const file_name,
                    const uint8_t** data, size_t* data_size) {
   int ok;
   void* file_data;
   size_t file_size;
   FILE* in;
+  const int from_stdin = (file_name == NULL) || !strcmp(file_name, "-");
 
-  if (file_name == NULL || data == NULL || data_size == NULL) return 0;
+  if (from_stdin) return ExUtilReadFromStdin(data, data_size);
+
+  if (data == NULL || data_size == NULL) return 0;
   *data = NULL;
   *data_size = 0;
 
@@ -56,17 +140,119 @@ int ExUtilWriteFile(const char* const file_name,
                     const uint8_t* data, size_t data_size) {
   int ok;
   FILE* out;
+  const int to_stdout = (file_name == NULL) || !strcmp(file_name, "-");
 
-  if (file_name == NULL || data == NULL) {
+  if (data == NULL) {
     return 0;
   }
-  out = fopen(file_name, "wb");
+  out = to_stdout ? stdout : fopen(file_name, "wb");
   if (out == NULL) {
     fprintf(stderr, "Error! Cannot open output file '%s'\n", file_name);
     return 0;
   }
   ok = (fwrite(data, data_size, 1, out) == 1);
-  fclose(out);
+  if (out != stdout) fclose(out);
   return ok;
 }
 
+//------------------------------------------------------------------------------
+// WebP decoding
+
+static const char* const kStatusMessages[VP8_STATUS_NOT_ENOUGH_DATA + 1] = {
+  "OK", "OUT_OF_MEMORY", "INVALID_PARAM", "BITSTREAM_ERROR",
+  "UNSUPPORTED_FEATURE", "SUSPENDED", "USER_ABORT", "NOT_ENOUGH_DATA"
+};
+
+static void PrintAnimationWarning(const WebPDecoderConfig* const config) {
+  if (config->input.has_animation) {
+    fprintf(stderr,
+            "Error! Decoding of an animated WebP file is not supported.\n"
+            "       Use webpmux to extract the individual frames or\n"
+            "       vwebp to view this image.\n");
+  }
+}
+
+void ExUtilPrintWebPError(const char* const in_file, int status) {
+  fprintf(stderr, "Decoding of %s failed.\n", in_file);
+  fprintf(stderr, "Status: %d", status);
+  if (status >= VP8_STATUS_OK && status <= VP8_STATUS_NOT_ENOUGH_DATA) {
+    fprintf(stderr, "(%s)", kStatusMessages[status]);
+  }
+  fprintf(stderr, "\n");
+}
+
+int ExUtilLoadWebP(const char* const in_file,
+                   const uint8_t** data, size_t* data_size,
+                   WebPBitstreamFeatures* bitstream) {
+  VP8StatusCode status;
+  WebPBitstreamFeatures local_features;
+  if (!ExUtilReadFile(in_file, data, data_size)) return 0;
+
+  if (bitstream == NULL) {
+    bitstream = &local_features;
+  }
+
+  status = WebPGetFeatures(*data, *data_size, bitstream);
+  if (status != VP8_STATUS_OK) {
+    free((void*)*data);
+    *data = NULL;
+    *data_size = 0;
+    ExUtilPrintWebPError(in_file, status);
+    return 0;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+VP8StatusCode ExUtilDecodeWebP(const uint8_t* const data, size_t data_size,
+                               int verbose, WebPDecoderConfig* const config) {
+  Stopwatch stop_watch;
+  VP8StatusCode status = VP8_STATUS_OK;
+  if (config == NULL) return VP8_STATUS_INVALID_PARAM;
+
+  PrintAnimationWarning(config);
+
+  StopwatchReset(&stop_watch);
+
+  // Decoding call.
+  status = WebPDecode(data, data_size, config);
+
+  if (verbose) {
+    const double decode_time = StopwatchReadAndReset(&stop_watch);
+    fprintf(stderr, "Time to decode picture: %.3fs\n", decode_time);
+  }
+  return status;
+}
+
+VP8StatusCode ExUtilDecodeWebPIncremental(
+    const uint8_t* const data, size_t data_size,
+    int verbose, WebPDecoderConfig* const config) {
+  Stopwatch stop_watch;
+  VP8StatusCode status = VP8_STATUS_OK;
+  if (config == NULL) return VP8_STATUS_INVALID_PARAM;
+
+  PrintAnimationWarning(config);
+
+  StopwatchReset(&stop_watch);
+
+  // Decoding call.
+  {
+    WebPIDecoder* const idec = WebPIDecode(data, data_size, config);
+    if (idec == NULL) {
+      fprintf(stderr, "Failed during WebPINewDecoder().\n");
+      return VP8_STATUS_OUT_OF_MEMORY;
+    } else {
+      status = WebPIUpdate(idec, data, data_size);
+      WebPIDelete(idec);
+    }
+  }
+
+  if (verbose) {
+    const double decode_time = StopwatchReadAndReset(&stop_watch);
+    fprintf(stderr, "Time to decode picture: %.3fs\n", decode_time);
+  }
+  return status;
+}
+
+// -----------------------------------------------------------------------------

examples/example_util.h

@@ -13,22 +13,75 @@
 #ifndef WEBP_EXAMPLES_EXAMPLE_UTIL_H_
 #define WEBP_EXAMPLES_EXAMPLE_UTIL_H_
 
-#include "webp/types.h"
+#include <stdio.h>
+#include "webp/decode.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+//------------------------------------------------------------------------------
+// String parsing
+
+// Parses 'v' using strto(ul|l|d)(). If error is non-NULL, '*error' is set to
+// true on failure while on success it is left unmodified to allow chaining of
+// calls. An error is only printed on the first occurrence.
+uint32_t ExUtilGetUInt(const char* const v, int base, int* const error);
+int ExUtilGetInt(const char* const v, int base, int* const error);
+float ExUtilGetFloat(const char* const v, int* const error);
+
+//------------------------------------------------------------------------------
+// File I/O
+
+// Reopen file in binary (O_BINARY) mode.
+// Returns 'file' on success, NULL otherwise.
+FILE* ExUtilSetBinaryMode(FILE* file);
+
 // Allocates storage for entire file 'file_name' and returns contents and size
 // in 'data' and 'data_size'. Returns 1 on success, 0 otherwise. '*data' should
 // be deleted using free().
+// If 'file_name' is NULL or equal to "-", input is read from stdin by calling
+// the function ExUtilReadFromStdin().
 int ExUtilReadFile(const char* const file_name,
                    const uint8_t** data, size_t* data_size);
 
+// Same as ExUtilReadFile(), but reads until EOF from stdin instead.
+int ExUtilReadFromStdin(const uint8_t** data, size_t* data_size);
+
 // Write a data segment into a file named 'file_name'. Returns true if ok.
+// If 'file_name' is NULL or equal to "-", output is written to stdout.
 int ExUtilWriteFile(const char* const file_name,
                     const uint8_t* data, size_t data_size);
 
+//------------------------------------------------------------------------------
+// WebP decoding
+
+// Prints an informative error message regarding decode failure of 'in_file'.
+// 'status' is treated as a VP8StatusCode and if valid will be printed as a
+// text string.
+void ExUtilPrintWebPError(const char* const in_file, int status);
+
+// Reads a WebP from 'in_file', returning the contents and size in 'data' and
+// 'data_size'. If not NULL, 'bitstream' is populated using WebPGetFeatures().
+// Returns true on success.
+int ExUtilLoadWebP(const char* const in_file,
+                   const uint8_t** data, size_t* data_size,
+                   WebPBitstreamFeatures* bitstream);
+
+// Decodes the WebP contained in 'data'.
+// 'config' is a structure previously initialized by WebPInitDecoderConfig().
+// 'config->output' should have the desired colorspace selected. 'verbose' will
+// cause decode timing to be reported.
+// Returns the decoder status. On success 'config->output' will contain the
+// decoded picture.
+VP8StatusCode ExUtilDecodeWebP(const uint8_t* const data, size_t data_size,
+                               int verbose, WebPDecoderConfig* const config);
+
+// Same as ExUtilDecodeWebP(), but using the incremental decoder.
+VP8StatusCode ExUtilDecodeWebPIncremental(
+    const uint8_t* const data, size_t data_size,
+    int verbose, WebPDecoderConfig* const config);
+
 #ifdef __cplusplus
 }    // extern "C"
 #endif

examples/gif2webp.c

@@ -17,7 +17,7 @@
 #include <string.h>
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
 #ifdef WEBP_HAVE_GIF
@@ -28,6 +28,16 @@
 #include "./example_util.h"
 #include "./gif2webp_util.h"
 
+// GIFLIB_MAJOR is only defined in libgif >= 4.2.0.
+#if defined(GIFLIB_MAJOR) && defined(GIFLIB_MINOR)
+# define LOCAL_GIF_VERSION ((GIFLIB_MAJOR << 8) | GIFLIB_MINOR)
+# define LOCAL_GIF_PREREQ(maj, min) \
+    (LOCAL_GIF_VERSION >= (((maj) << 8) | (min)))
+#else
+# define LOCAL_GIF_VERSION 0
+# define LOCAL_GIF_PREREQ(maj, min) 0
+#endif
+
 #define GIF_TRANSPARENT_MASK 0x01
 #define GIF_DISPOSE_MASK     0x07
 #define GIF_DISPOSE_SHIFT    2
@@ -36,7 +46,7 @@
 
 //------------------------------------------------------------------------------
 
-static int transparent_index = -1;  // Index of transparent color in the map.
+static int transparent_index = -1;  // Opaque frame by default.
 
 static void SanitizeKeyFrameIntervals(size_t* const kmin_ptr,
                                       size_t* const kmax_ptr) {
@@ -103,12 +113,12 @@ static void Remap(const uint8_t* const src, const GifFileType* const gif,
 static int ReadFrame(GifFileType* const gif, WebPFrameRect* const gif_rect,
                      WebPPicture* const webp_frame) {
   WebPPicture sub_image;
-  const GifImageDesc image_desc = gif->Image;
+  const GifImageDesc* const image_desc = &gif->Image;
   uint32_t* dst = NULL;
   uint8_t* tmp = NULL;
   int ok = 0;
   WebPFrameRect rect = {
-      image_desc.Left, image_desc.Top, image_desc.Width, image_desc.Height
+      image_desc->Left, image_desc->Top, image_desc->Width, image_desc->Height
   };
   *gif_rect = rect;
 
@@ -124,7 +134,7 @@ static int ReadFrame(GifFileType* const gif, WebPFrameRect* const gif_rect,
   tmp = (uint8_t*)malloc(rect.width * sizeof(*tmp));
   if (tmp == NULL) goto End;
 
-  if (image_desc.Interlace) {  // Interlaced image.
+  if (image_desc->Interlace) {  // Interlaced image.
     // We need 4 passes, with the following offsets and jumps.
     const int interlace_offsets[] = { 0, 4, 2, 1 };
     const int interlace_jumps[]   = { 8, 8, 4, 2 };
@@ -153,30 +163,29 @@ static int ReadFrame(GifFileType* const gif, WebPFrameRect* const gif_rect,
   return ok;
 }
 
-static int GetBackgroundColor(const ColorMapObject* const color_map,
-                              int bgcolor_idx, uint32_t* const bgcolor) {
+static void GetBackgroundColor(const ColorMapObject* const color_map,
+                               int bgcolor_idx, uint32_t* const bgcolor) {
   if (transparent_index != -1 && bgcolor_idx == transparent_index) {
     *bgcolor = WEBP_UTIL_TRANSPARENT_COLOR;  // Special case.
-    return 1;
   } else if (color_map == NULL || color_map->Colors == NULL
              || bgcolor_idx >= color_map->ColorCount) {
-    return 0;  // Invalid color map or index.
+    *bgcolor = WHITE_COLOR;
+    fprintf(stderr,
+            "GIF decode warning: invalid background color index. Assuming "
+            "white background.\n");
   } else {
     const GifColorType color = color_map->Colors[bgcolor_idx];
     *bgcolor = (0xff        << 24)
              | (color.Red   << 16)
              | (color.Green <<  8)
              | (color.Blue  <<  0);
-    return 1;
   }
 }
 
 static void DisplayGifError(const GifFileType* const gif, int gif_error) {
-  // GIFLIB_MAJOR is only defined in libgif >= 4.2.0.
   // libgif 4.2.0 has retired PrintGifError() and added GifErrorString().
-#if defined(GIFLIB_MAJOR) && defined(GIFLIB_MINOR) && \
-        ((GIFLIB_MAJOR == 4 && GIFLIB_MINOR >= 2) || GIFLIB_MAJOR > 4)
-#if GIFLIB_MAJOR >= 5
+#if LOCAL_GIF_PREREQ(4,2)
+#if LOCAL_GIF_PREREQ(5,0)
   // Static string actually, hence the const char* cast.
   const char* error_str = (const char*)GifErrorString(
       (gif == NULL) ? gif_error : gif->Error);
@@ -194,7 +203,7 @@ static void DisplayGifError(const GifFileType* const gif, int gif_error) {
 #endif
 }
 
-static const char* const kErrorMessages[] = {
+static const char* const kErrorMessages[-WEBP_MUX_NOT_ENOUGH_DATA + 1] = {
   "WEBP_MUX_NOT_FOUND", "WEBP_MUX_INVALID_ARGUMENT", "WEBP_MUX_BAD_DATA",
   "WEBP_MUX_MEMORY_ERROR", "WEBP_MUX_NOT_ENOUGH_DATA"
 };
@@ -215,26 +224,26 @@ enum {
 static void Help(void) {
   printf("Usage:\n");
   printf(" gif2webp [options] gif_file -o webp_file\n");
-  printf("options:\n");
+  printf("Options:\n");
   printf("  -h / -help  ............ this help\n");
-  printf("  -lossy ................. Encode image using lossy compression.\n");
-  printf("  -mixed ................. For each frame in the image, pick lossy\n"
-         "                           or lossless compression heuristically.\n");
+  printf("  -lossy ................. encode image using lossy compression\n");
+  printf("  -mixed ................. for each frame in the image, pick lossy\n"
+         "                           or lossless compression heuristically\n");
   printf("  -q <float> ............. quality factor (0:small..100:big)\n");
   printf("  -m <int> ............... compression method (0=fast, 6=slowest)\n");
-  printf("  -kmin <int> ............ Min distance between key frames\n");
-  printf("  -kmax <int> ............ Max distance between key frames\n");
+  printf("  -kmin <int> ............ min distance between key frames\n");
+  printf("  -kmax <int> ............ max distance between key frames\n");
   printf("  -f <int> ............... filter strength (0=off..100)\n");
   printf("  -metadata <string> ..... comma separated list of metadata to\n");
   printf("                           ");
-  printf("copy from the input to the output if present.\n");
+  printf("copy from the input to the output if present\n");
   printf("                           "
          "Valid values: all, none, icc, xmp (default)\n");
   printf("  -mt .................... use multi-threading if available\n");
   printf("\n");
-  printf("  -version ............... print version number and exit.\n");
-  printf("  -v ..................... verbose.\n");
-  printf("  -quiet ................. don't print anything.\n");
+  printf("  -version ............... print version number and exit\n");
+  printf("  -v ..................... verbose\n");
+  printf("  -quiet ................. don't print anything\n");
   printf("\n");
 }
 
@@ -250,7 +259,8 @@ int main(int argc, const char *argv[]) {
   GifFileType* gif = NULL;
   WebPConfig config;
   WebPPicture frame;
-  WebPMuxFrameInfo info;
+  int duration = 0;
+  FrameDisposeMethod orig_dispose = FRAME_DISPOSE_NONE;
   WebPMuxAnimParams anim = { WHITE_COLOR, 0 };
   WebPFrameCache* cache = NULL;
 
@@ -270,17 +280,11 @@ int main(int argc, const char *argv[]) {
   size_t kmax = 0;
   int allow_mixed = 0;   // If true, each frame can be lossy or lossless.
 
-  memset(&info, 0, sizeof(info));
-  info.id = WEBP_CHUNK_ANMF;
-  info.dispose_method = WEBP_MUX_DISPOSE_BACKGROUND;
-  info.blend_method = WEBP_MUX_BLEND;
-
   if (!WebPConfigInit(&config) || !WebPPictureInit(&frame)) {
     fprintf(stderr, "Error! Version mismatch!\n");
     return -1;
   }
   config.lossless = 1;  // Use lossless compression by default.
-  config.image_hint = WEBP_HINT_GRAPH;   // always low-color
 
   if (argc == 1) {
     Help();
@@ -288,6 +292,7 @@ int main(int argc, const char *argv[]) {
   }
 
   for (c = 1; c < argc; ++c) {
+    int parse_error = 0;
     if (!strcmp(argv[c], "-h") || !strcmp(argv[c], "-help")) {
       Help();
       return 0;
@@ -299,17 +304,17 @@ int main(int argc, const char *argv[]) {
       allow_mixed = 1;
       config.lossless = 0;
     } else if (!strcmp(argv[c], "-q") && c < argc - 1) {
-      config.quality = (float)strtod(argv[++c], NULL);
+      config.quality = ExUtilGetFloat(argv[++c], &parse_error);
     } else if (!strcmp(argv[c], "-m") && c < argc - 1) {
-      config.method = strtol(argv[++c], NULL, 0);
+      config.method = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-kmax") && c < argc - 1) {
-      kmax = strtoul(argv[++c], NULL, 0);
+      kmax = ExUtilGetUInt(argv[++c], 0, &parse_error);
       default_kmax = 0;
     } else if (!strcmp(argv[c], "-kmin") && c < argc - 1) {
-      kmin = strtoul(argv[++c], NULL, 0);
+      kmin = ExUtilGetUInt(argv[++c], 0, &parse_error);
       default_kmin = 0;
     } else if (!strcmp(argv[c], "-f") && c < argc - 1) {
-      config.filter_strength = strtol(argv[++c], NULL, 0);
+      config.filter_strength = ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-metadata") && c < argc - 1) {
       static const struct {
         const char* option;
@@ -373,6 +378,11 @@ int main(int argc, const char *argv[]) {
     } else {
       in_file = argv[c];
     }
+
+    if (parse_error) {
+      Help();
+      return -1;
+    }
   }
 
   // Appropriate default kmin, kmax values for lossy and lossless.
@@ -396,24 +406,13 @@ int main(int argc, const char *argv[]) {
   }
 
   // Start the decoder object
-#if defined(GIFLIB_MAJOR) && (GIFLIB_MAJOR >= 5)
-  // There was an API change in version 5.0.0.
+#if LOCAL_GIF_PREREQ(5,0)
   gif = DGifOpenFileName(in_file, &gif_error);
 #else
   gif = DGifOpenFileName(in_file);
 #endif
   if (gif == NULL) goto End;
 
-  // Allocate current buffer
-  frame.width = gif->SWidth;
-  frame.height = gif->SHeight;
-  frame.use_argb = 1;
-  if (!WebPPictureAlloc(&frame)) goto End;
-
-  // Initialize cache
-  cache = WebPFrameCacheNew(frame.width, frame.height, kmin, kmax, allow_mixed);
-  if (cache == NULL) goto End;
-
   mux = WebPMuxNew();
   if (mux == NULL) {
     fprintf(stderr, "ERROR: could not create a mux object.\n");
@@ -429,13 +428,66 @@ int main(int argc, const char *argv[]) {
     switch (type) {
       case IMAGE_DESC_RECORD_TYPE: {
         WebPFrameRect gif_rect;
+        GifImageDesc* const image_desc = &gif->Image;
 
         if (!DGifGetImageDesc(gif)) goto End;
+
+        // Fix some broken GIF global headers that report
+        // 0 x 0 screen dimension.
+        if (is_first_frame) {
+          if (verbose) {
+            printf("Canvas screen: %d x %d\n", gif->SWidth, gif->SHeight);
+          }
+          if (gif->SWidth == 0 || gif->SHeight == 0) {
+            image_desc->Left = 0;
+            image_desc->Top = 0;
+            gif->SWidth = image_desc->Width;
+            gif->SHeight = image_desc->Height;
+            if (gif->SWidth <= 0 || gif->SHeight <= 0) {
+              goto End;
+            }
+            if (verbose) {
+              printf("Fixed canvas screen dimension to: %d x %d\n",
+                     gif->SWidth, gif->SHeight);
+            }
+          }
+#if WEBP_MUX_ABI_VERSION > 0x0101
+          // Set definitive canvas size.
+          err = WebPMuxSetCanvasSize(mux, gif->SWidth, gif->SHeight);
+          if (err != WEBP_MUX_OK) {
+            fprintf(stderr, "Invalid canvas size %d x %d\n",
+                    gif->SWidth, gif->SHeight);
+            goto End;
+          }
+#endif
+          // Allocate current buffer.
+          frame.width = gif->SWidth;
+          frame.height = gif->SHeight;
+          frame.use_argb = 1;
+          if (!WebPPictureAlloc(&frame)) goto End;
+          WebPUtilClearPic(&frame, NULL);
+
+          // Initialize cache.
+          cache = WebPFrameCacheNew(frame.width, frame.height,
+                                    kmin, kmax, allow_mixed);
+          if (cache == NULL) goto End;
+
+          // Background color.
+          GetBackgroundColor(gif->SColorMap, gif->SBackGroundColor,
+                             &anim.bgcolor);
+        }
+        // Some even more broken GIF can have sub-rect with zero width/height.
+        if (image_desc->Width == 0 || image_desc->Height == 0) {
+          image_desc->Width = gif->SWidth;
+          image_desc->Height = gif->SHeight;
+        }
+
         if (!ReadFrame(gif, &gif_rect, &frame)) {
           goto End;
         }
 
-        if (!WebPFrameCacheAddFrame(cache, &config, &gif_rect, &frame, &info)) {
+        if (!WebPFrameCacheAddFrame(cache, &config, &gif_rect, orig_dispose,
+                                    duration, &frame)) {
           fprintf(stderr, "Error! Cannot encode frame as WebP\n");
           fprintf(stderr, "Error code: %d\n", frame.error_code);
         }
@@ -447,6 +499,13 @@ int main(int argc, const char *argv[]) {
           goto End;
         }
         is_first_frame = 0;
+
+        // In GIF, graphic control extensions are optional for a frame, so we
+        // may not get one before reading the next frame. To handle this case,
+        // we reset frame properties to reasonable defaults for the next frame.
+        orig_dispose = FRAME_DISPOSE_NONE;
+        duration = 0;
+        transparent_index = -1;  // Opaque frame by default.
         break;
       }
       case EXTENSION_RECORD_TYPE: {
@@ -464,30 +523,21 @@ int main(int argc, const char *argv[]) {
             const int dispose = (flags >> GIF_DISPOSE_SHIFT) & GIF_DISPOSE_MASK;
             const int delay = data[2] | (data[3] << 8);  // In 10 ms units.
             if (data[0] != 4) goto End;
-            info.duration = delay * 10;  // Duration is in 1 ms units for WebP.
-            if (dispose == 3) {
-              static int warning_printed = 0;
-              if (!warning_printed) {
-                fprintf(stderr, "WARNING: GIF_DISPOSE_RESTORE unsupported.\n");
-                warning_printed = 1;
-              }
-              // failsafe. TODO(urvang): emulate the correct behaviour by
-              // recoding the whole frame.
-              info.dispose_method = WEBP_MUX_DISPOSE_BACKGROUND;
-            } else {
-              info.dispose_method =
-                  (dispose == 2) ? WEBP_MUX_DISPOSE_BACKGROUND
-                                 : WEBP_MUX_DISPOSE_NONE;
+            duration = delay * 10;  // Duration is in 1 ms units for WebP.
+            switch (dispose) {
+              case 3:
+                orig_dispose = FRAME_DISPOSE_RESTORE_PREVIOUS;
+                break;
+              case 2:
+                orig_dispose = FRAME_DISPOSE_BACKGROUND;
+                break;
+              case 1:
+              case 0:
+              default:
+                orig_dispose = FRAME_DISPOSE_NONE;
+                break;
             }
             transparent_index = (flags & GIF_TRANSPARENT_MASK) ? data[4] : -1;
-            if (is_first_frame) {
-              if (!GetBackgroundColor(gif->SColorMap, gif->SBackGroundColor,
-                                      &anim.bgcolor)) {
-                fprintf(stderr, "GIF decode warning: invalid background color "
-                                "index. Assuming white background.\n");
-              }
-              WebPUtilClearPic(&frame, NULL);
-            }
             break;
           }
           case PLAINTEXT_EXT_FUNC_CODE: {
@@ -495,13 +545,16 @@ int main(int argc, const char *argv[]) {
           }
           case APPLICATION_EXT_FUNC_CODE: {
             if (data[0] != 11) break;    // Chunk is too short
-            if (!memcmp(data + 1, "NETSCAPE2.0", 11)) {
+            if (!memcmp(data + 1, "NETSCAPE2.0", 11) ||
+                !memcmp(data + 1, "ANIMEXTS1.0", 11)) {
               // Recognize and parse Netscape2.0 NAB extension for loop count.
               if (DGifGetExtensionNext(gif, &data) == GIF_ERROR) goto End;
               if (data == NULL) goto End;  // Loop count sub-block missing.
-              if (data[0] != 3 && data[1] != 1) break;   // wrong size/marker
+              if (data[0] < 3 || data[1] != 1) break;   // wrong size/marker
               anim.loop_count = data[2] | (data[3] << 8);
-              if (verbose) printf("Loop count: %d\n", anim.loop_count);
+              if (verbose) {
+                fprintf(stderr, "Loop count: %d\n", anim.loop_count);
+              }
             } else {  // An extension containing metadata.
               // We only store the first encountered chunk of each type, and
               // only if requested by the user.
@@ -555,7 +608,8 @@ int main(int argc, const char *argv[]) {
                 // Add metadata chunk.
                 err = WebPMuxSetChunk(mux, fourccs[is_icc], &metadata, 1);
                 if (verbose) {
-                  printf("%s size: %d\n", features[is_icc], (int)metadata.size);
+                  fprintf(stderr, "%s size: %d\n",
+                          features[is_icc], (int)metadata.size);
                 }
                 WebPDataClear(&metadata);
                 if (err != WEBP_MUX_OK) {
@@ -621,11 +675,11 @@ int main(int argc, const char *argv[]) {
       goto End;
     }
     if (!quiet) {
-      printf("Saved output file: %s\n", out_file);
+      fprintf(stderr, "Saved output file: %s\n", out_file);
     }
   } else {
     if (!quiet) {
-      printf("Nothing written; use -o flag to save the result.\n");
+      fprintf(stderr, "Nothing written; use -o flag to save the result.\n");
     }
   }
 
@@ -644,7 +698,11 @@ int main(int argc, const char *argv[]) {
     DisplayGifError(gif, gif_error);
   }
   if (gif != NULL) {
+#if LOCAL_GIF_PREREQ(5,1)
+    DGifCloseFile(gif, &gif_error);
+#else
     DGifCloseFile(gif);
+#endif
   }
 
   return !ok;

examples/gif2webp_util.h

@@ -29,6 +29,13 @@ extern "C" {
 
 struct WebPPicture;
 
+// Includes all disposal methods, even the ones not supported by WebP bitstream.
+typedef enum FrameDisposeMethod {
+  FRAME_DISPOSE_NONE,
+  FRAME_DISPOSE_BACKGROUND,
+  FRAME_DISPOSE_RESTORE_PREVIOUS
+} FrameDisposeMethod;
+
 typedef struct {
   int x_offset, y_offset, width, height;
 } WebPFrameRect;
@@ -53,14 +60,15 @@ WebPFrameCache* WebPFrameCacheNew(int width, int height,
 // Release all the frame data from 'cache' and free 'cache'.
 void WebPFrameCacheDelete(WebPFrameCache* const cache);
 
-// Given an image described by 'frame', 'info' and 'orig_rect', optimize it for
-// WebP, encode it and add it to 'cache'.
-// This takes care of frame disposal too, according to 'info->dispose_method'.
+// Given an image described by 'frame', 'rect', 'dispose_method' and 'duration',
+// optimize it for WebP, encode it and add it to 'cache'. 'rect' can be NULL.
+// This takes care of frame disposal too, according to 'dispose_method'.
+// Returns false in case of error (and sets frame->error_code accordingly).
 int WebPFrameCacheAddFrame(WebPFrameCache* const cache,
                            const WebPConfig* const config,
-                           const WebPFrameRect* const orig_rect,
-                           WebPPicture* const frame,
-                           WebPMuxFrameInfo* const info);
+                           const WebPFrameRect* const rect,
+                           FrameDisposeMethod dispose_method, int duration,
+                           WebPPicture* const frame);
 
 // Flush the *ready* frames from cache and add them to 'mux'. If 'verbose' is
 // true, prints the information about these frames.

examples/jpegdec.c

@@ -12,7 +12,7 @@
 #include "./jpegdec.h"
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
 #include <stdio.h>
@@ -211,30 +211,31 @@ static void my_error_exit(j_common_ptr dinfo) {
 int ReadJPEG(FILE* in_file, WebPPicture* const pic, Metadata* const metadata) {
   int ok = 0;
   int stride, width, height;
-  struct jpeg_decompress_struct dinfo;
+  volatile struct jpeg_decompress_struct dinfo;
   struct my_error_mgr jerr;
-  uint8_t* rgb = NULL;
+  uint8_t* volatile rgb = NULL;
   JSAMPROW buffer[1];
 
+  memset((j_decompress_ptr)&dinfo, 0, sizeof(dinfo));   // for setjmp sanity
   dinfo.err = jpeg_std_error(&jerr.pub);
   jerr.pub.error_exit = my_error_exit;
 
   if (setjmp(jerr.setjmp_buffer)) {
  Error:
     MetadataFree(metadata);
-    jpeg_destroy_decompress(&dinfo);
+    jpeg_destroy_decompress((j_decompress_ptr)&dinfo);
     goto End;
   }
 
-  jpeg_create_decompress(&dinfo);
-  jpeg_stdio_src(&dinfo, in_file);
-  if (metadata != NULL) SaveMetadataMarkers(&dinfo);
-  jpeg_read_header(&dinfo, TRUE);
+  jpeg_create_decompress((j_decompress_ptr)&dinfo);
+  jpeg_stdio_src((j_decompress_ptr)&dinfo, in_file);
+  if (metadata != NULL) SaveMetadataMarkers((j_decompress_ptr)&dinfo);
+  jpeg_read_header((j_decompress_ptr)&dinfo, TRUE);
 
   dinfo.out_color_space = JCS_RGB;
   dinfo.do_fancy_upsampling = TRUE;
 
-  jpeg_start_decompress(&dinfo);
+  jpeg_start_decompress((j_decompress_ptr)&dinfo);
 
   if (dinfo.output_components != 3) {
     goto Error;
@@ -251,26 +252,27 @@ int ReadJPEG(FILE* in_file, WebPPicture* const pic, Metadata* const metadata) {
   buffer[0] = (JSAMPLE*)rgb;
 
   while (dinfo.output_scanline < dinfo.output_height) {
-    if (jpeg_read_scanlines(&dinfo, buffer, 1) != 1) {
+    if (jpeg_read_scanlines((j_decompress_ptr)&dinfo, buffer, 1) != 1) {
       goto End;
     }
     buffer[0] += stride;
   }
 
   if (metadata != NULL) {
-    ok = ExtractMetadataFromJPEG(&dinfo, metadata);
+    ok = ExtractMetadataFromJPEG((j_decompress_ptr)&dinfo, metadata);
     if (!ok) {
       fprintf(stderr, "Error extracting JPEG metadata!\n");
       goto Error;
     }
   }
 
-  jpeg_finish_decompress(&dinfo);
-  jpeg_destroy_decompress(&dinfo);
+  jpeg_finish_decompress((j_decompress_ptr)&dinfo);
+  jpeg_destroy_decompress((j_decompress_ptr)&dinfo);
 
   // WebP conversion.
   pic->width = width;
   pic->height = height;
+  pic->use_argb = 1;      // store raw RGB samples
   ok = WebPPictureImportRGB(pic, rgb, stride);
   if (!ok) goto Error;
 

examples/pngdec.c

@@ -12,7 +12,7 @@
 #include "./pngdec.h"
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
 #include <stdio.h>
@@ -190,9 +190,9 @@ static int ExtractMetadataFromPNG(png_structp png,
 
 int ReadPNG(FILE* in_file, WebPPicture* const pic, int keep_alpha,
             Metadata* const metadata) {
-  png_structp png;
-  png_infop info = NULL;
-  png_infop end_info = NULL;
+  volatile png_structp png;
+  volatile png_infop info = NULL;
+  volatile png_infop end_info = NULL;
   int color_type, bit_depth, interlaced;
   int has_alpha;
   int num_passes;
@@ -200,7 +200,7 @@ int ReadPNG(FILE* in_file, WebPPicture* const pic, int keep_alpha,
   int ok = 0;
   png_uint_32 width, height, y;
   int stride;
-  uint8_t* rgb = NULL;
+  uint8_t* volatile rgb = NULL;
 
   png = png_create_read_struct(PNG_LIBPNG_VER_STRING, 0, 0, 0);
   if (png == NULL) {
@@ -211,7 +211,6 @@ int ReadPNG(FILE* in_file, WebPPicture* const pic, int keep_alpha,
   if (setjmp(png_jmpbuf(png))) {
  Error:
     MetadataFree(metadata);
-    png_destroy_read_struct(&png, &info, &end_info);
     goto End;
   }
 
@@ -228,7 +227,9 @@ int ReadPNG(FILE* in_file, WebPPicture* const pic, int keep_alpha,
 
   png_set_strip_16(png);
   png_set_packing(png);
-  if (color_type == PNG_COLOR_TYPE_PALETTE) png_set_palette_to_rgb(png);
+  if (color_type == PNG_COLOR_TYPE_PALETTE) {
+    png_set_palette_to_rgb(png);
+  }
   if (color_type == PNG_COLOR_TYPE_GRAY ||
       color_type == PNG_COLOR_TYPE_GRAY_ALPHA) {
     if (bit_depth < 8) {
@@ -255,7 +256,7 @@ int ReadPNG(FILE* in_file, WebPPicture* const pic, int keep_alpha,
   if (rgb == NULL) goto Error;
   for (p = 0; p < num_passes; ++p) {
     for (y = 0; y < height; ++y) {
-      png_bytep row = rgb + y * stride;
+      png_bytep row = (png_bytep)(rgb + y * stride);
       png_read_rows(png, &row, NULL, 1);
     }
   }
@@ -267,8 +268,6 @@ int ReadPNG(FILE* in_file, WebPPicture* const pic, int keep_alpha,
     goto Error;
   }
 
-  png_destroy_read_struct(&png, &info, &end_info);
-
   pic->width = width;
   pic->height = height;
   pic->use_argb = 1;
@@ -280,6 +279,10 @@ int ReadPNG(FILE* in_file, WebPPicture* const pic, int keep_alpha,
   }
 
  End:
+  if (png != NULL) {
+    png_destroy_read_struct((png_structpp)&png,
+                            (png_infopp)&info, (png_infopp)&end_info);
+  }
   free(rgb);
   return ok;
 }

examples/stopwatch.h

@@ -14,6 +14,8 @@
 #ifndef WEBP_EXAMPLES_STOPWATCH_H_
 #define WEBP_EXAMPLES_STOPWATCH_H_
 
+#include "webp/types.h"
+
 #if defined _WIN32 && !defined __GNUC__
 #include <windows.h>
 
@@ -37,6 +39,7 @@ static WEBP_INLINE double StopwatchReadAndReset(Stopwatch* watch) {
 
 
 #else    /* !_WIN32 */
+#include <string.h>  // memcpy
 #include <sys/time.h>
 
 typedef struct timeval Stopwatch;
@@ -46,10 +49,13 @@ static WEBP_INLINE void StopwatchReset(Stopwatch* watch) {
 }
 
 static WEBP_INLINE double StopwatchReadAndReset(Stopwatch* watch) {
-  const struct timeval old_value = *watch;
+  struct timeval old_value;
+  double delta_sec, delta_usec;
+  memcpy(&old_value, watch, sizeof(old_value));
   gettimeofday(watch, NULL);
-  return watch->tv_sec - old_value.tv_sec +
-      (watch->tv_usec - old_value.tv_usec) / 1000000.0;
+  delta_sec = (double)watch->tv_sec - old_value.tv_sec;
+  delta_usec = (double)watch->tv_usec - old_value.tv_usec;
+  return delta_sec + delta_usec / 1000000.0;
 }
 
 #endif   /* _WIN32 */

examples/tiffdec.c

@@ -12,7 +12,7 @@
 #include "./tiffdec.h"
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
 #include <stdio.h>
@@ -97,7 +97,7 @@ int ReadTIFF(const char* const filename,
       pic->width = width;
       pic->height = height;
       // TIFF data is ABGR
-#ifdef __BIG_ENDIAN__
+#ifdef WORDS_BIGENDIAN
       TIFFSwabArrayOfLong(raster, width * height);
 #endif
       pic->use_argb = 1;

examples/vwebp.c

@@ -11,7 +11,7 @@
 //
 // Author: Skal (pascal.massimino@gmail.com)
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
 #include <stdio.h>
@@ -42,8 +42,6 @@
 #define snprintf _snprintf
 #endif
 
-static void Help(void);
-
 // Unfortunate global variables. Gathered into a struct for comfort.
 static struct {
   int has_animation;
@@ -82,6 +80,16 @@ static void ClearParams(void) {
   kParams.dmux = NULL;
 }
 
+// Sets the previous frame to the dimensions of the canvas and has it dispose
+// to background to cause the canvas to be cleared.
+static void ClearPreviousFrame(void) {
+  WebPIterator* const prev = &kParams.prev_frame;
+  prev->width = kParams.canvas_width;
+  prev->height = kParams.canvas_height;
+  prev->x_offset = prev->y_offset = 0;
+  prev->dispose_method = WEBP_MUX_DISPOSE_BACKGROUND;
+}
+
 // -----------------------------------------------------------------------------
 // Color profile handling
 static int ApplyColorProfile(const WebPData* const profile,
@@ -181,6 +189,8 @@ static void decode_callback(int what) {
         if (WebPDemuxGetFrame(kParams.dmux, 1, curr)) {
           --kParams.loop_count;
           kParams.done = (kParams.loop_count == 0);
+          if (kParams.done) return;
+          ClearPreviousFrame();
         } else {
           kParams.decoding_error = 1;
           kParams.done = 1;
@@ -372,19 +382,22 @@ static void Help(void) {
   printf("Usage: vwebp in_file [options]\n\n"
          "Decodes the WebP image file and visualize it using OpenGL\n"
          "Options are:\n"
-         "  -version  .... print version number and exit.\n"
-         "  -noicc ....... don't use the icc profile if present.\n"
-         "  -nofancy ..... don't use the fancy YUV420 upscaler.\n"
-         "  -nofilter .... disable in-loop filtering.\n"
-         "  -dither <int>  dithering strength (0..100). Default=50.\n"
-         "  -mt .......... use multi-threading.\n"
-         "  -info ........ print info.\n"
-         "  -h     ....... this help message.\n"
+         "  -version  .... print version number and exit\n"
+         "  -noicc ....... don't use the icc profile if present\n"
+         "  -nofancy ..... don't use the fancy YUV420 upscaler\n"
+         "  -nofilter .... disable in-loop filtering\n"
+         "  -dither <int>  dithering strength (0..100), default=50\n"
+#if WEBP_DECODER_ABI_VERSION > 0x0204
+         "  -noalphadither disable alpha plane dithering\n"
+#endif
+         "  -mt .......... use multi-threading\n"
+         "  -info ........ print info\n"
+         "  -h     ....... this help message\n"
          "\n"
          "Keyboard shortcuts:\n"
-         "  'c' ................ toggle use of color profile.\n"
-         "  'i' ................ overlay file information.\n"
-         "  'q' / 'Q' / ESC .... quit.\n"
+         "  'c' ................ toggle use of color profile\n"
+         "  'i' ................ overlay file information\n"
+         "  'q' / 'Q' / ESC .... quit\n"
         );
 }
 
@@ -392,16 +405,19 @@ int main(int argc, char *argv[]) {
   int c;
   WebPDecoderConfig* const config = &kParams.config;
   WebPIterator* const curr = &kParams.curr_frame;
-  WebPIterator* const prev = &kParams.prev_frame;
 
   if (!WebPInitDecoderConfig(config)) {
     fprintf(stderr, "Library version mismatch!\n");
     return -1;
   }
   config->options.dithering_strength = 50;
+#if WEBP_DECODER_ABI_VERSION > 0x0204
+  config->options.alpha_dithering_strength = 100;
+#endif
   kParams.use_color_profile = 1;
 
   for (c = 1; c < argc; ++c) {
+    int parse_error = 0;
     if (!strcmp(argv[c], "-h") || !strcmp(argv[c], "-help")) {
       Help();
       return 0;
@@ -411,8 +427,13 @@ int main(int argc, char *argv[]) {
       config->options.no_fancy_upsampling = 1;
     } else if (!strcmp(argv[c], "-nofilter")) {
       config->options.bypass_filtering = 1;
+#if WEBP_DECODER_ABI_VERSION > 0x0204
+    } else if (!strcmp(argv[c], "-noalphadither")) {
+      config->options.alpha_dithering_strength = 0;
+#endif
     } else if (!strcmp(argv[c], "-dither") && c + 1 < argc) {
-      config->options.dithering_strength = strtol(argv[++c], NULL, 0);
+      config->options.dithering_strength =
+          ExUtilGetInt(argv[++c], 0, &parse_error);
     } else if (!strcmp(argv[c], "-info")) {
       kParams.print_info = 1;
     } else if (!strcmp(argv[c], "-version")) {
@@ -435,6 +456,11 @@ int main(int argc, char *argv[]) {
     } else {
       kParams.file_name = argv[c];
     }
+
+    if (parse_error) {
+      Help();
+      return -1;
+    }
   }
 
   if (kParams.file_name == NULL) {
@@ -469,10 +495,7 @@ int main(int argc, char *argv[]) {
     printf("Canvas: %d x %d\n", kParams.canvas_width, kParams.canvas_height);
   }
 
-  prev->width = kParams.canvas_width;
-  prev->height = kParams.canvas_height;
-  prev->x_offset = prev->y_offset = 0;
-  prev->dispose_method = WEBP_MUX_DISPOSE_BACKGROUND;
+  ClearPreviousFrame();
 
   memset(&kParams.iccp, 0, sizeof(kParams.iccp));
   kParams.has_color_profile =

examples/webpdec.c

@@ -0,0 +1,67 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// WebP decode.
+
+#include "./webpdec.h"
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "webp/decode.h"
+#include "webp/encode.h"
+#include "./example_util.h"
+#include "./metadata.h"
+
+int ReadWebP(const char* const in_file, WebPPicture* const pic,
+             int keep_alpha, Metadata* const metadata) {
+  int ok = 0;
+  size_t data_size = 0;
+  const uint8_t* data = NULL;
+  VP8StatusCode status = VP8_STATUS_OK;
+  WebPDecoderConfig config;
+  WebPDecBuffer* const output_buffer = &config.output;
+  WebPBitstreamFeatures* const bitstream = &config.input;
+
+  // TODO(jzern): add Exif/XMP/ICC extraction.
+  if (metadata != NULL) {
+    fprintf(stderr, "Warning: metadata extraction from WebP is unsupported.\n");
+  }
+
+  if (!WebPInitDecoderConfig(&config)) {
+    fprintf(stderr, "Library version mismatch!\n");
+    return 0;
+  }
+
+  if (ExUtilLoadWebP(in_file, &data, &data_size, bitstream)) {
+    const int has_alpha = keep_alpha && bitstream->has_alpha;
+    output_buffer->colorspace = has_alpha ? MODE_RGBA : MODE_RGB;
+
+    status = ExUtilDecodeWebP(data, data_size, 0, &config);
+    if (status == VP8_STATUS_OK) {
+      const uint8_t* const rgba = output_buffer->u.RGBA.rgba;
+      const int stride = output_buffer->u.RGBA.stride;
+      pic->width = output_buffer->width;
+      pic->height = output_buffer->height;
+      pic->use_argb = 1;
+      ok = has_alpha ? WebPPictureImportRGBA(pic, rgba, stride)
+                     : WebPPictureImportRGB(pic, rgba, stride);
+    }
+  }
+
+  if (status != VP8_STATUS_OK) {
+    ExUtilPrintWebPError(in_file, status);
+  }
+
+  free((void*)data);
+  WebPFreeDecBuffer(output_buffer);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------

examples/webpdec.h

@@ -0,0 +1,33 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// WebP decode.
+
+#ifndef WEBP_EXAMPLES_WEBPDEC_H_
+#define WEBP_EXAMPLES_WEBPDEC_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct Metadata;
+struct WebPPicture;
+
+// Reads a WebP from 'in_file', returning the decoded output in 'pic'.
+// If 'keep_alpha' is true and the WebP has an alpha channel, the output is
+// RGBA otherwise it will be RGB.
+// Returns true on success.
+int ReadWebP(const char* const in_file, struct WebPPicture* const pic,
+             int keep_alpha, struct Metadata* const metadata);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // WEBP_EXAMPLES_WEBPDEC_H_

examples/webpmux.c

@@ -46,7 +46,7 @@
 */
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
 #include <assert.h>
@@ -130,7 +130,7 @@ static int CountOccurrences(const char* arglist[], int list_length,
   return num_occurences;
 }
 
-static const char* const kErrorMessages[] = {
+static const char* const kErrorMessages[-WEBP_MUX_NOT_ENOUGH_DATA + 1] = {
   "WEBP_MUX_NOT_FOUND", "WEBP_MUX_INVALID_ARGUMENT", "WEBP_MUX_BAD_DATA",
   "WEBP_MUX_MEMORY_ERROR", "WEBP_MUX_NOT_ENOUGH_DATA"
 };
@@ -303,51 +303,51 @@ static void PrintHelp(void) {
 
   printf("\n");
   printf("GET_OPTIONS:\n");
-  printf(" Extract relevant data.\n");
-  printf("   icc       Get ICC profile.\n");
-  printf("   exif      Get EXIF metadata.\n");
-  printf("   xmp       Get XMP metadata.\n");
+  printf(" Extract relevant data:\n");
+  printf("   icc       get ICC profile\n");
+  printf("   exif      get EXIF metadata\n");
+  printf("   xmp       get XMP metadata\n");
 #ifdef WEBP_EXPERIMENTAL_FEATURES
-  printf("   frgm n    Get nth fragment.\n");
+  printf("   frgm n    get nth fragment\n");
 #endif
-  printf("   frame n   Get nth frame.\n");
+  printf("   frame n   get nth frame\n");
 
   printf("\n");
   printf("SET_OPTIONS:\n");
-  printf(" Set color profile/metadata.\n");
-  printf("   icc  file.icc     Set ICC profile.\n");
-  printf("   exif file.exif    Set EXIF metadata.\n");
-  printf("   xmp  file.xmp     Set XMP metadata.\n");
+  printf(" Set color profile/metadata:\n");
+  printf("   icc  file.icc     set ICC profile\n");
+  printf("   exif file.exif    set EXIF metadata\n");
+  printf("   xmp  file.xmp     set XMP metadata\n");
   printf("   where:    'file.icc' contains the ICC profile to be set,\n");
   printf("             'file.exif' contains the EXIF metadata to be set\n");
   printf("             'file.xmp' contains the XMP metadata to be set\n");
 
   printf("\n");
   printf("STRIP_OPTIONS:\n");
-  printf(" Strip color profile/metadata.\n");
-  printf("   icc       Strip ICC profile.\n");
-  printf("   exif      Strip EXIF metadata.\n");
-  printf("   xmp       Strip XMP metadata.\n");
+  printf(" Strip color profile/metadata:\n");
+  printf("   icc       strip ICC profile\n");
+  printf("   exif      strip EXIF metadata\n");
+  printf("   xmp       strip XMP metadata\n");
 
 #ifdef WEBP_EXPERIMENTAL_FEATURES
   printf("\n");
   printf("FRAGMENT_OPTIONS(i):\n");
-  printf(" Create fragmented image.\n");
+  printf(" Create fragmented image:\n");
   printf("   file_i +xi+yi\n");
   printf("   where:    'file_i' is the i'th fragment (WebP format),\n");
-  printf("             'xi','yi' specify the image offset for this fragment."
+  printf("             'xi','yi' specify the image offset for this fragment"
          "\n");
 #endif
 
   printf("\n");
   printf("FRAME_OPTIONS(i):\n");
-  printf(" Create animation.\n");
+  printf(" Create animation:\n");
   printf("   file_i +di+[xi+yi[+mi[bi]]]\n");
   printf("   where:    'file_i' is the i'th animation frame (WebP format),\n");
-  printf("             'di' is the pause duration before next frame.\n");
-  printf("             'xi','yi' specify the image offset for this frame.\n");
-  printf("             'mi' is the dispose method for this frame (0 or 1).\n");
-  printf("             'bi' is the blending method for this frame (+b or -b)."
+  printf("             'di' is the pause duration before next frame,\n");
+  printf("             'xi','yi' specify the image offset for this frame,\n");
+  printf("             'mi' is the dispose method for this frame (0 or 1),\n");
+  printf("             'bi' is the blending method for this frame (+b or -b)"
          "\n");
 
   printf("\n");
@@ -363,7 +363,7 @@ static void PrintHelp(void) {
          "specifying\n");
   printf("            the Alpha, Red, Green and Blue component values "
          "respectively\n");
-  printf("            [Default: 255,255,255,255].\n");
+  printf("            [Default: 255,255,255,255]\n");
 
   printf("\nINPUT & OUTPUT are in WebP format.\n");
 
@@ -371,6 +371,14 @@ static void PrintHelp(void) {
   printf(" and is assumed to be\nvalid.\n");
 }
 
+static void WarnAboutOddOffset(const WebPMuxFrameInfo* const info) {
+  if ((info->x_offset | info->y_offset) & 1) {
+    fprintf(stderr, "Warning: odd offsets will be snapped to even values"
+            " (%d, %d) -> (%d, %d)\n", info->x_offset, info->y_offset,
+            info->x_offset & ~1, info->y_offset & ~1);
+  }
+}
+
 static int ReadFileToWebPData(const char* const filename,
                               WebPData* const webp_data) {
   const uint8_t* data;
@@ -394,8 +402,9 @@ static int CreateMux(const char* const filename, WebPMux** mux) {
 
 static int WriteData(const char* filename, const WebPData* const webpdata) {
   int ok = 0;
-  FILE* fout = strcmp(filename, "-") ? fopen(filename, "wb") : stdout;
-  if (!fout) {
+  FILE* fout = strcmp(filename, "-") ? fopen(filename, "wb")
+                                     : ExUtilSetBinaryMode(stdout);
+  if (fout == NULL) {
     fprintf(stderr, "Error opening output WebP file %s!\n", filename);
     return 0;
   }
@@ -444,6 +453,9 @@ static int ParseFrameArgs(const char* args, WebPMuxFrameInfo* const info) {
     default:
       return 0;
   }
+
+  WarnAboutOddOffset(info);
+
   // Note: The sanity of the following conversion is checked by
   // WebPMuxPushFrame().
   info->dispose_method = (WebPMuxAnimDispose)dispose_method;
@@ -456,7 +468,10 @@ static int ParseFrameArgs(const char* args, WebPMuxFrameInfo* const info) {
 }
 
 static int ParseFragmentArgs(const char* args, WebPMuxFrameInfo* const info) {
-  return (sscanf(args, "+%d+%d", &info->x_offset, &info->y_offset) == 2);
+  const int ok =
+      (sscanf(args, "+%d+%d", &info->x_offset, &info->y_offset) == 2);
+  if (ok) WarnAboutOddOffset(info);
+  return ok;
 }
 
 static int ParseBgcolorArgs(const char* args, uint32_t* const bgcolor) {
@@ -473,7 +488,7 @@ static int ParseBgcolorArgs(const char* args, uint32_t* const bgcolor) {
 static void DeleteConfig(WebPMuxConfig* config) {
   if (config != NULL) {
     free(config->feature_.args_);
-    free(config);
+    memset(config, 0, sizeof(*config));
   }
 }
 
@@ -776,33 +791,27 @@ static int ValidateConfig(WebPMuxConfig* config) {
 
 // Create config object from command-line arguments.
 static int InitializeConfig(int argc, const char* argv[],
-                            WebPMuxConfig** config) {
+                            WebPMuxConfig* config) {
   int num_feature_args = 0;
   int ok = 1;
 
   assert(config != NULL);
-  *config = NULL;
+  memset(config, 0, sizeof(*config));
 
   // Validate command-line arguments.
   if (!ValidateCommandLine(argc, argv, &num_feature_args)) {
     ERROR_GOTO1("Exiting due to command-line parsing error.\n", Err1);
   }
 
-  // Allocate memory.
-  *config = (WebPMuxConfig*)calloc(1, sizeof(**config));
-  if (*config == NULL) {
-    ERROR_GOTO1("ERROR: Memory allocation error.\n", Err1);
-  }
-  (*config)->feature_.arg_count_ = num_feature_args;
-  (*config)->feature_.args_ =
-      (FeatureArg*)calloc(num_feature_args, sizeof(FeatureArg));
-  if ((*config)->feature_.args_ == NULL) {
+  config->feature_.arg_count_ = num_feature_args;
+  config->feature_.args_ =
+      (FeatureArg*)calloc(num_feature_args, sizeof(*config->feature_.args_));
+  if (config->feature_.args_ == NULL) {
     ERROR_GOTO1("ERROR: Memory allocation error.\n", Err1);
   }
 
   // Parse command-line.
-  if (!ParseCommandLine(argc, argv, *config) ||
-      !ValidateConfig(*config)) {
+  if (!ParseCommandLine(argc, argv, config) || !ValidateConfig(config)) {
     ERROR_GOTO1("Exiting due to command-line parsing error.\n", Err1);
   }
 
@@ -824,14 +833,16 @@ static int GetFrameFragment(const WebPMux* mux,
   WebPMux* mux_single = NULL;
   long num = 0;
   int ok = 1;
+  int parse_error = 0;
   const WebPChunkId id = is_frame ? WEBP_CHUNK_ANMF : WEBP_CHUNK_FRGM;
   WebPMuxFrameInfo info;
   WebPDataInit(&info.bitstream);
 
-  num = strtol(config->feature_.args_[0].params_, NULL, 10);
+  num = ExUtilGetInt(config->feature_.args_[0].params_, 10, &parse_error);
   if (num < 0) {
     ERROR_GOTO1("ERROR: Frame/Fragment index must be non-negative.\n", ErrGet);
   }
+  if (parse_error) goto ErrGet;
 
   err = WebPMuxGetFrame(mux, num, &info);
   if (err == WEBP_MUX_OK && info.id != id) err = WEBP_MUX_NOT_FOUND;
@@ -857,7 +868,7 @@ static int GetFrameFragment(const WebPMux* mux,
  ErrGet:
   WebPDataClear(&info.bitstream);
   WebPMuxDelete(mux_single);
-  return ok;
+  return ok && !parse_error;
 }
 
 // Read and process config.
@@ -919,16 +930,19 @@ static int Process(const WebPMuxConfig* config) {
                 break;
               }
               case SUBTYPE_LOOP: {
-                const long loop_count =
-                    strtol(feature->args_[i].params_, NULL, 10);
-                if (loop_count != (int)loop_count) {
+                int parse_error = 0;
+                const int loop_count =
+                    ExUtilGetInt(feature->args_[i].params_, 10, &parse_error);
+                if (loop_count < 0 || loop_count > 65535) {
                   // Note: This is only a 'necessary' condition for loop_count
                   // to be valid. The 'sufficient' conditioned in checked in
                   // WebPMuxSetAnimationParams() method called later.
                   ERROR_GOTO1("ERROR: Loop count must be in the range 0 to "
                               "65535.\n", Err2);
                 }
-                params.loop_count = (int)loop_count;
+                ok = !parse_error;
+                if (!ok) goto Err2;
+                params.loop_count = loop_count;
                 break;
               }
               case SUBTYPE_ANMF: {
@@ -1028,8 +1042,8 @@ static int Process(const WebPMuxConfig* config) {
                       ErrorString(err), kDescriptions[feature->type_], Err2);
         }
       } else {
-          ERROR_GOTO1("ERROR: Invalid feature for action 'strip'.\n", Err2);
-          break;
+        ERROR_GOTO1("ERROR: Invalid feature for action 'strip'.\n", Err2);
+        break;
       }
       ok = WriteWebP(mux, config->output_);
       break;
@@ -1055,14 +1069,14 @@ static int Process(const WebPMuxConfig* config) {
 // Main.
 
 int main(int argc, const char* argv[]) {
-  WebPMuxConfig* config;
+  WebPMuxConfig config;
   int ok = InitializeConfig(argc - 1, argv + 1, &config);
   if (ok) {
-    ok = Process(config);
+    ok = Process(&config);
   } else {
     PrintHelp();
   }
-  DeleteConfig(config);
+  DeleteConfig(&config);
   return !ok;
 }
 

examples/wicdec.c

@@ -12,9 +12,10 @@
 #include "./wicdec.h"
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "webp/config.h"
 #endif
 
+#include <assert.h>
 #include <stdio.h>
 
 #ifdef HAVE_WINCODEC_H
@@ -109,6 +110,7 @@ static HRESULT ExtractICCP(IWICImagingFactory* const factory,
     IFS(IWICBitmapFrameDecode_GetColorContexts(frame,
                                                count, color_contexts,
                                                &num_color_contexts));
+    assert(FAILED(hr) || num_color_contexts <= count);
     for (i = 0; SUCCEEDED(hr) && i < num_color_contexts; ++i) {
       WICColorContextType type;
       IFS(IWICColorContext_GetType(color_contexts[i], &type));
@@ -116,7 +118,7 @@ static HRESULT ExtractICCP(IWICImagingFactory* const factory,
         UINT size;
         IFS(IWICColorContext_GetProfileBytes(color_contexts[i],
                                              0, NULL, &size));
-        if (size > 0) {
+        if (SUCCEEDED(hr) && size > 0) {
           iccp->bytes = (uint8_t*)malloc(size);
           if (iccp->bytes == NULL) {
             hr = E_OUTOFMEMORY;
@@ -261,7 +263,7 @@ int ReadPictureWithWIC(const char* const filename,
   IFS(IWICBitmapFrameDecode_GetPixelFormat(frame, &src_pixel_format));
   IFS(IWICBitmapDecoder_GetContainerFormat(decoder, &src_container_format));
 
-  if (keep_alpha) {
+  if (SUCCEEDED(hr) && keep_alpha) {
     const GUID** guid;
     for (guid = kAlphaContainers; *guid != NULL; ++guid) {
       if (IsEqualGUID(MAKE_REFGUID(src_container_format),

iosbuild.sh

@@ -12,31 +12,37 @@
 set -e
 
 # Extract the latest SDK version from the final field of the form: iphoneosX.Y
-declare -r SDK=$(xcodebuild -showsdks \
+readonly SDK=$(xcodebuild -showsdks \
   | grep iphoneos | sort | tail -n 1 | awk '{print substr($NF, 9)}'
 )
 # Extract Xcode version.
-declare -r XCODE=$(xcodebuild -version | grep Xcode | cut -d " " -f2)
+readonly XCODE=$(xcodebuild -version | grep Xcode | cut -d " " -f2)
+if [[ -z "${XCODE}" ]]; then
+  echo "Xcode not available"
+  exit 1
+fi
 
-declare -r OLDPATH=${PATH}
+readonly OLDPATH=${PATH}
 
 # Add iPhoneOS-V6 to the list of platforms below if you need armv6 support.
 # Note that iPhoneOS-V6 support is not available with the iOS6 SDK.
-declare -r PLATFORMS="iPhoneSimulator iPhoneOS-V7 iPhoneOS-V7s"
-declare -r SRCDIR=$(dirname $0)
-declare -r TOPDIR=$(pwd)
-declare -r BUILDDIR="${TOPDIR}/iosbuild"
-declare -r TARGETDIR="${TOPDIR}/WebP.framework"
-declare -r DEVELOPER=$(xcode-select --print-path)
-declare -r PLATFORMSROOT="${DEVELOPER}/Platforms"
-declare -r LIPO=$(xcrun -sdk iphoneos${SDK} -find lipo)
+PLATFORMS="iPhoneSimulator iPhoneSimulator64"
+PLATFORMS+=" iPhoneOS-V7 iPhoneOS-V7s iPhoneOS-V7-arm64"
+readonly PLATFORMS
+readonly SRCDIR=$(dirname $0)
+readonly TOPDIR=$(pwd)
+readonly BUILDDIR="${TOPDIR}/iosbuild"
+readonly TARGETDIR="${TOPDIR}/WebP.framework"
+readonly DEVELOPER=$(xcode-select --print-path)
+readonly PLATFORMSROOT="${DEVELOPER}/Platforms"
+readonly LIPO=$(xcrun -sdk iphoneos${SDK} -find lipo)
 LIBLIST=''
 
 if [[ -z "${SDK}" ]]; then
   echo "iOS SDK not available"
   exit 1
-elif [[ ${SDK} < 4.0 ]]; then
-  echo "You need iOS SDK version 4.0 or above"
+elif [[ ${SDK} < 6.0 ]]; then
+  echo "You need iOS SDK version 6.0 or above"
   exit 1
 else
   echo "iOS SDK Version ${SDK}"
@@ -47,10 +53,25 @@ rm -rf ${TARGETDIR}
 mkdir -p ${BUILDDIR}
 mkdir -p ${TARGETDIR}/Headers/
 
-[[ -e ${SRCDIR}/configure ]] || (cd ${SRCDIR} && sh autogen.sh)
+if [[ ! -e ${SRCDIR}/configure ]]; then
+  if ! (cd ${SRCDIR} && sh autogen.sh); then
+    cat <<EOT
+Error creating configure script!
+This script requires the autoconf/automake and libtool to build. MacPorts can
+be used to obtain these:
+http://www.macports.org/install.php
+EOT
+    exit 1
+  fi
+fi
 
 for PLATFORM in ${PLATFORMS}; do
-  if [[ "${PLATFORM}" == "iPhoneOS-V7s" ]]; then
+  ARCH2=""
+  if [[ "${PLATFORM}" == "iPhoneOS-V7-arm64" ]]; then
+    PLATFORM="iPhoneOS"
+    ARCH="aarch64"
+    ARCH2="arm64"
+  elif [[ "${PLATFORM}" == "iPhoneOS-V7s" ]]; then
     PLATFORM="iPhoneOS"
     ARCH="armv7s"
   elif [[ "${PLATFORM}" == "iPhoneOS-V7" ]]; then
@@ -59,6 +80,9 @@ for PLATFORM in ${PLATFORMS}; do
   elif [[ "${PLATFORM}" == "iPhoneOS-V6" ]]; then
     PLATFORM="iPhoneOS"
     ARCH="armv6"
+  elif [[ "${PLATFORM}" == "iPhoneSimulator64" ]]; then
+    PLATFORM="iPhoneSimulator"
+    ARCH="x86_64"
   else
     ARCH="i386"
   fi
@@ -66,30 +90,20 @@ for PLATFORM in ${PLATFORMS}; do
   ROOTDIR="${BUILDDIR}/${PLATFORM}-${SDK}-${ARCH}"
   mkdir -p "${ROOTDIR}"
 
-  SDKROOT="${PLATFORMSROOT}/${PLATFORM}.platform/Developer/SDKs/${PLATFORM}${SDK}.sdk/"
-  CFLAGS="-arch ${ARCH} -pipe -isysroot ${SDKROOT}"
-  LDFLAGS="-arch ${ARCH} -pipe -isysroot ${SDKROOT}"
+  DEVROOT="${DEVELOPER}/Toolchains/XcodeDefault.xctoolchain"
+  SDKROOT="${PLATFORMSROOT}/"
+  SDKROOT+="${PLATFORM}.platform/Developer/SDKs/${PLATFORM}${SDK}.sdk/"
+  CFLAGS="-arch ${ARCH2:-${ARCH}} -pipe -isysroot ${SDKROOT} -O3 -DNDEBUG"
+  CFLAGS+=" -miphoneos-version-min=6.0"
 
-  if [[ -z "${XCODE}" ]]; then
-    echo "XCODE not available"
-    exit 1
-  elif [[ ${SDK} < 5.0.0 ]]; then
-    DEVROOT="${PLATFORMSROOT}/${PLATFORM}.platform/Developer/"
-  else
-    DEVROOT="${DEVELOPER}/Toolchains/XcodeDefault.xctoolchain"
-    CFLAGS+=" -miphoneos-version-min=5.0"
-    LDFLAGS+=" -miphoneos-version-min=5.0"
-  fi
-
-  export CFLAGS
-  export LDFLAGS
-  export CXXFLAGS=${CFLAGS}
+  set -x
   export PATH="${DEVROOT}/usr/bin:${OLDPATH}"
-
   ${SRCDIR}/configure --host=${ARCH}-apple-darwin --prefix=${ROOTDIR} \
     --build=$(${SRCDIR}/config.guess) \
     --disable-shared --enable-static \
-    --enable-libwebpdecoder --enable-swap-16bit-csp
+    --enable-libwebpdecoder --enable-swap-16bit-csp \
+    CFLAGS="${CFLAGS}"
+  set +x
 
   # run make only in the src/ directory to create libwebpdecoder.a
   cd src/
@@ -104,5 +118,5 @@ for PLATFORM in ${PLATFORMS}; do
   export PATH=${OLDPATH}
 done
 
-cp -a ${SRCDIR}/src/webp/* ${TARGETDIR}/Headers/
+cp -a ${SRCDIR}/src/webp/*.h ${TARGETDIR}/Headers/
 ${LIPO} -create ${LIBLIST} -output ${TARGETDIR}/WebP

m4/ax_pthread.m4

@@ -82,7 +82,7 @@
 #   modified version of the Autoconf Macro, you may extend this special
 #   exception to the GPL to apply to your modified version as well.
 
-#serial 18
+#serial 21
 
 AU_ALIAS([ACX_PTHREAD], [AX_PTHREAD])
 AC_DEFUN([AX_PTHREAD], [
@@ -103,8 +103,8 @@ if test x"$PTHREAD_LIBS$PTHREAD_CFLAGS" != x; then
         save_LIBS="$LIBS"
         LIBS="$PTHREAD_LIBS $LIBS"
         AC_MSG_CHECKING([for pthread_join in LIBS=$PTHREAD_LIBS with CFLAGS=$PTHREAD_CFLAGS])
-        AC_TRY_LINK_FUNC(pthread_join, ax_pthread_ok=yes)
-        AC_MSG_RESULT($ax_pthread_ok)
+        AC_TRY_LINK_FUNC([pthread_join], [ax_pthread_ok=yes])
+        AC_MSG_RESULT([$ax_pthread_ok])
         if test x"$ax_pthread_ok" = xno; then
                 PTHREAD_LIBS=""
                 PTHREAD_CFLAGS=""
@@ -164,6 +164,20 @@ case ${host_os} in
         ;;
 esac
 
+# Clang doesn't consider unrecognized options an error unless we specify
+# -Werror. We throw in some extra Clang-specific options to ensure that
+# this doesn't happen for GCC, which also accepts -Werror.
+
+AC_MSG_CHECKING([if compiler needs -Werror to reject unknown flags])
+save_CFLAGS="$CFLAGS"
+ax_pthread_extra_flags="-Werror"
+CFLAGS="$CFLAGS $ax_pthread_extra_flags -Wunknown-warning-option -Wsizeof-array-argument"
+AC_COMPILE_IFELSE([AC_LANG_PROGRAM([int foo(void);],[foo()])],
+                  [AC_MSG_RESULT([yes])],
+                  [ax_pthread_extra_flags=
+                   AC_MSG_RESULT([no])])
+CFLAGS="$save_CFLAGS"
+
 if test x"$ax_pthread_ok" = xno; then
 for flag in $ax_pthread_flags; do
 
@@ -178,7 +192,7 @@ for flag in $ax_pthread_flags; do
                 ;;
 
                 pthread-config)
-                AC_CHECK_PROG(ax_pthread_config, pthread-config, yes, no)
+                AC_CHECK_PROG([ax_pthread_config], [pthread-config], [yes], [no])
                 if test x"$ax_pthread_config" = xno; then continue; fi
                 PTHREAD_CFLAGS="`pthread-config --cflags`"
                 PTHREAD_LIBS="`pthread-config --ldflags` `pthread-config --libs`"
@@ -193,7 +207,7 @@ for flag in $ax_pthread_flags; do
         save_LIBS="$LIBS"
         save_CFLAGS="$CFLAGS"
         LIBS="$PTHREAD_LIBS $LIBS"
-        CFLAGS="$CFLAGS $PTHREAD_CFLAGS"
+        CFLAGS="$CFLAGS $PTHREAD_CFLAGS $ax_pthread_extra_flags"
 
         # Check for various functions.  We must include pthread.h,
         # since some functions may be macros.  (On the Sequent, we
@@ -219,7 +233,7 @@ for flag in $ax_pthread_flags; do
         LIBS="$save_LIBS"
         CFLAGS="$save_CFLAGS"
 
-        AC_MSG_RESULT($ax_pthread_ok)
+        AC_MSG_RESULT([$ax_pthread_ok])
         if test "x$ax_pthread_ok" = xyes; then
                 break;
         fi
@@ -245,9 +259,9 @@ if test "x$ax_pthread_ok" = xyes; then
                 [attr_name=$attr; break],
                 [])
         done
-        AC_MSG_RESULT($attr_name)
+        AC_MSG_RESULT([$attr_name])
         if test "$attr_name" != PTHREAD_CREATE_JOINABLE; then
-            AC_DEFINE_UNQUOTED(PTHREAD_CREATE_JOINABLE, $attr_name,
+            AC_DEFINE_UNQUOTED([PTHREAD_CREATE_JOINABLE], [$attr_name],
                                [Define to necessary symbol if this constant
                                 uses a non-standard name on your system.])
         fi
@@ -261,45 +275,54 @@ if test "x$ax_pthread_ok" = xyes; then
             if test "$GCC" = "yes"; then
                 flag="-D_REENTRANT"
             else
+                # TODO: What about Clang on Solaris?
                 flag="-mt -D_REENTRANT"
             fi
             ;;
         esac
-        AC_MSG_RESULT(${flag})
+        AC_MSG_RESULT([$flag])
         if test "x$flag" != xno; then
             PTHREAD_CFLAGS="$flag $PTHREAD_CFLAGS"
         fi
 
         AC_CACHE_CHECK([for PTHREAD_PRIO_INHERIT],
-            ax_cv_PTHREAD_PRIO_INHERIT, [
-                AC_LINK_IFELSE([
-                    AC_LANG_PROGRAM([[#include <pthread.h>]], [[int i = PTHREAD_PRIO_INHERIT;]])],
+            [ax_cv_PTHREAD_PRIO_INHERIT], [
+                AC_LINK_IFELSE([AC_LANG_PROGRAM([[#include <pthread.h>]],
+                                                [[int i = PTHREAD_PRIO_INHERIT;]])],
                     [ax_cv_PTHREAD_PRIO_INHERIT=yes],
                     [ax_cv_PTHREAD_PRIO_INHERIT=no])
             ])
         AS_IF([test "x$ax_cv_PTHREAD_PRIO_INHERIT" = "xyes"],
-            AC_DEFINE([HAVE_PTHREAD_PRIO_INHERIT], 1, [Have PTHREAD_PRIO_INHERIT.]))
+            [AC_DEFINE([HAVE_PTHREAD_PRIO_INHERIT], [1], [Have PTHREAD_PRIO_INHERIT.])])
 
         LIBS="$save_LIBS"
         CFLAGS="$save_CFLAGS"
 
-        # More AIX lossage: must compile with xlc_r or cc_r
-        if test x"$GCC" != xyes; then
-          AC_CHECK_PROGS(PTHREAD_CC, xlc_r cc_r, ${CC})
-        else
-          PTHREAD_CC=$CC
+        # More AIX lossage: compile with *_r variant
+        if test "x$GCC" != xyes; then
+            case $host_os in
+                aix*)
+                AS_CASE(["x/$CC"],
+                  [x*/c89|x*/c89_128|x*/c99|x*/c99_128|x*/cc|x*/cc128|x*/xlc|x*/xlc_v6|x*/xlc128|x*/xlc128_v6],
+                  [#handle absolute path differently from PATH based program lookup
+                   AS_CASE(["x$CC"],
+                     [x/*],
+                     [AS_IF([AS_EXECUTABLE_P([${CC}_r])],[PTHREAD_CC="${CC}_r"])],
+                     [AC_CHECK_PROGS([PTHREAD_CC],[${CC}_r],[$CC])])])
+                ;;
+            esac
         fi
-else
-        PTHREAD_CC="$CC"
 fi
 
-AC_SUBST(PTHREAD_LIBS)
-AC_SUBST(PTHREAD_CFLAGS)
-AC_SUBST(PTHREAD_CC)
+test -n "$PTHREAD_CC" || PTHREAD_CC="$CC"
+
+AC_SUBST([PTHREAD_LIBS])
+AC_SUBST([PTHREAD_CFLAGS])
+AC_SUBST([PTHREAD_CC])
 
 # Finally, execute ACTION-IF-FOUND/ACTION-IF-NOT-FOUND:
 if test x"$ax_pthread_ok" = xyes; then
-        ifelse([$1],,AC_DEFINE(HAVE_PTHREAD,1,[Define if you have POSIX threads libraries and header files.]),[$1])
+        ifelse([$1],,[AC_DEFINE([HAVE_PTHREAD],[1],[Define if you have POSIX threads libraries and header files.])],[$1])
         :
 else
         ax_pthread_ok=no

makefile.unix

@@ -67,8 +67,20 @@ EXTRA_FLAGS += -Wmissing-prototypes
 EXTRA_FLAGS += -Wmissing-declarations
 EXTRA_FLAGS += -Wdeclaration-after-statement
 EXTRA_FLAGS += -Wshadow
+EXTRA_FLAGS += -Wformat-security -Wformat-nonliteral
+
 # EXTRA_FLAGS += -Wvla
 
+# AVX2-specific flags:
+ifeq ($(HAVE_AVX2), 1)
+EXTRA_FLAGS += -DWEBP_HAVE_AVX2
+src/dsp/%_avx2.o: EXTRA_FLAGS += -mavx2
+endif
+
+# NEON-specific flags:
+# EXTRA_FLAGS += -march=armv7-a -mfloat-abi=hard -mfpu=neon -mtune=cortex-a8
+# -> seems to make the overall lib slower: -fno-split-wide-types
+
 #### Nothing should normally be changed below this line ####
 
 AR = ar
@@ -87,7 +99,6 @@ DEC_OBJS = \
     src/dec/frame.o \
     src/dec/idec.o \
     src/dec/io.o \
-    src/dec/layer.o \
     src/dec/quant.o \
     src/dec/tree.o \
     src/dec/vp8.o \
@@ -98,18 +109,29 @@ DEMUX_OBJS = \
     src/demux/demux.o \
 
 DSP_DEC_OBJS = \
+    src/dsp/alpha_processing.o \
+    src/dsp/alpha_processing_sse2.o \
     src/dsp/cpu.o \
     src/dsp/dec.o \
+    src/dsp/dec_clip_tables.o \
+    src/dsp/dec_mips32.o \
     src/dsp/dec_neon.o \
     src/dsp/dec_sse2.o \
     src/dsp/lossless.o \
+    src/dsp/lossless_mips32.o \
+    src/dsp/lossless_neon.o \
+    src/dsp/lossless_sse2.o \
     src/dsp/upsampling.o \
     src/dsp/upsampling_neon.o \
     src/dsp/upsampling_sse2.o \
     src/dsp/yuv.o \
+    src/dsp/yuv_mips32.o \
+    src/dsp/yuv_sse2.o \
 
 DSP_ENC_OBJS = \
     src/dsp/enc.o \
+    src/dsp/enc_avx2.o \
+    src/dsp/enc_mips32.o \
     src/dsp/enc_neon.o \
     src/dsp/enc_sse2.o \
 
@@ -123,8 +145,11 @@ ENC_OBJS = \
     src/enc/frame.o \
     src/enc/histogram.o \
     src/enc/iterator.o \
-    src/enc/layer.o \
     src/enc/picture.o \
+    src/enc/picture_csp.o \
+    src/enc/picture_psnr.o \
+    src/enc/picture_rescale.o \
+    src/enc/picture_tools.o \
     src/enc/quant.o \
     src/enc/syntax.o \
     src/enc/token.o \
@@ -137,6 +162,7 @@ EX_FORMAT_DEC_OBJS = \
     examples/metadata.o \
     examples/pngdec.o \
     examples/tiffdec.o \
+    examples/webpdec.o \
 
 EX_UTIL_OBJS = \
     examples/example_util.o \
@@ -150,7 +176,6 @@ MUX_OBJS = \
     src/mux/muxread.o \
 
 UTILS_DEC_OBJS = \
-    src/utils/alpha_processing.o \
     src/utils/bit_reader.o \
     src/utils/color_cache.o \
     src/utils/filters.o \
@@ -188,10 +213,15 @@ HDRS = \
     src/dec/webpi.h \
     src/dsp/dsp.h \
     src/dsp/lossless.h \
+    src/dsp/neon.h \
     src/dsp/yuv.h \
+    src/dsp/yuv_tables_sse2.h \
+    src/enc/backward_references.h \
     src/enc/cost.h \
+    src/enc/histogram.h \
     src/enc/vp8enci.h \
-    src/utils/alpha_processing.h \
+    src/enc/vp8li.h \
+    src/mux/muxi.h \
     src/utils/bit_reader.h \
     src/utils/bit_writer.h \
     src/utils/color_cache.h \
@@ -203,6 +233,7 @@ HDRS = \
     src/utils/random.h \
     src/utils/rescaler.h \
     src/utils/thread.h \
+    src/utils/utils.h \
     src/webp/format_constants.h \
     $(HDRS_INSTALLED) \
 
@@ -222,6 +253,14 @@ all: ex $(EXTRA_EXAMPLES)
 
 $(EX_FORMAT_DEC_OBJS): %.o: %.h
 
+# special dependencies:
+#   tree.c/vp8.c/bit_reader.c <-> bit_reader_inl.h, endian_inl.h
+#   bit_writer.c <-> endian_inl.h
+src/dec/tree.o: src/utils/bit_reader_inl.h src/utils/endian_inl.h
+src/dec/vp8.o: src/utils/bit_reader_inl.h src/utils/endian_inl.h
+src/utils/bit_reader.o: src/utils/bit_reader_inl.h src/utils/endian_inl.h
+src/utils/bit_writer.o: src/utils/endian_inl.h
+
 %.o: %.c $(HDRS)
 	$(CC) $(CFLAGS) $(CPPFLAGS) -c $< -o $@
 
@@ -241,7 +280,7 @@ examples/gif2webp: examples/gif2webp.o
 examples/vwebp: examples/vwebp.o
 examples/webpmux: examples/webpmux.o
 
-examples/cwebp: src/libwebp.a
+examples/cwebp: examples/libexample_util.a src/libwebp.a
 examples/cwebp: EXTRA_LIBS += $(CWEBP_LIBS)
 examples/dwebp: examples/libexample_util.a src/libwebpdecoder.a
 examples/dwebp: EXTRA_LIBS += $(DWEBP_LIBS)
@@ -270,7 +309,7 @@ dist: all
 	$(INSTALL) -m644 src/demux/libwebpdemux.a $(DESTDIR)/lib
 	$(INSTALL) -m644 src/mux/libwebpmux.a $(DESTDIR)/lib
 	umask 022; \
-	for m in man/[cd]webp.1 man/gif2webp.1 man/webpmux.1; do \
+	for m in man/[cdv]webp.1 man/gif2webp.1 man/webpmux.1; do \
 	  basenam=$$(basename $$m .1); \
 	  $(GROFF) -t -e -man -T utf8 $$m \
 	    | $(COL) -bx >$(DESTDIR)/doc/$${basenam}.txt; \

man/Makefile.am

@@ -5,4 +5,7 @@ endif
 if BUILD_GIF2WEBP
   man_MANS += gif2webp.1
 endif
+if BUILD_VWEBP
+  man_MANS += vwebp.1
+endif
 EXTRA_DIST = $(man_MANS)

man/cwebp.1

@@ -1,5 +1,5 @@
 .\"                                      Hey, EMACS: -*- nroff -*-
-.TH CWEBP 1 "December 12, 2013"
+.TH CWEBP 1 "Oct 13, 2014"
 .SH NAME
 cwebp \- compress an image file to a WebP file
 .SH SYNOPSIS
@@ -12,13 +12,19 @@ This manual page documents the
 command.
 .PP
 \fBcwebp\fP compresses an image using the WebP format.
-Input format can be either PNG, JPEG, TIFF or raw Y'CbCr samples.
+Input format can be either PNG, JPEG, TIFF, WebP or raw Y'CbCr samples.
 .SH OPTIONS
 The basic options are:
 .TP
 .BI \-o " string
 Specify the name of the output WebP file. If omitted, \fBcwebp\fP will
 perform compression but only report statistics.
+Using "\-" as output name will direct output to 'stdout'.
+.TP
+.BI \-\- " string
+Explicitly specify the input file. This option is useful if the input
+file starts with an '\-' for instance. This option must appear \fBlast\fP.
+Any other options afterward will be ignored.
 .TP
 .B \-h, \-help
 A short usage summary.
@@ -39,6 +45,15 @@ with lower quality. Best quality is achieved by using a value of 100.
 In case of lossless compression (specified by the \-lossless option), a small
 factor enables faster compression speed, but produces a larger file. Maximum
 compression is achieved by using a value of 100.
+.\" TODO(jzern): restore post-v0.4.1
+.\" .TP
+.\" .BI \-z " int
+.\" Switch on \fBlossless\fP compression mode with the specified level between 0
+.\" and 9, with level 0 being the fastest, 9 being the slowest. Fast mode
+.\" produces larger file size than slower ones. A good default is \-z 6.
+.\" This option is actually a shortcut for some predefined settings for quality
+.\" and method. If options \-q  or \-m are subsequently used, they will invalidate
+.\" the effect of this \-z option.
 .TP
 .BI \-alpha_q " int
 Specify the compression factor for alpha compression between 0 and 100.

man/dwebp.1

@@ -1,5 +1,5 @@
 .\"                                      Hey, EMACS: -*- nroff -*-
-.TH DWEBP 1 "December 12, 2013"
+.TH DWEBP 1 "July 22, 2014"
 .SH NAME
 dwebp \- decompress a WebP file to an image file
 .SH SYNOPSIS
@@ -25,6 +25,12 @@ Print the version number (as major.minor.revision) and exit.
 Specify the name of the output file (as PNG format by default).
 Using "-" as output name will direct output to 'stdout'.
 .TP
+.BI \-\- " string
+Explicitly specify the input file. This option is useful if the input
+file starts with an '\-' for instance. This option must appear \fBlast\fP.
+Any other options afterward will be ignored. If the input file is "\-",
+the data will be read from \fIstdin\fP instead of a file.
+.TP
 .B \-bmp
 Change the output format to uncompressed BMP.
 .TP
@@ -57,10 +63,16 @@ Don't use the in-loop filtering process even if it is required by
 the bitstream. This may produce visible blocks on the non-compliant output,
 but it will make the decoding faster.
 .TP
-.B \-dither " strength
+.BI \-dither " strength
 Specify a dithering \fBstrength\fP between 0 and 100. Dithering is a
 post-processing effect applied to chroma components in lossy compression.
 It helps by smoothing gradients and avoiding banding artifacts.
+.\" TODO(jzern): restore post-v0.4.1
+.\" .TP
+.\" .BI \-alpha_dither
+.\" If the compressed file contains a transparency plane that was quantized
+.\" during compression, this flag will allow dithering the reconstructed plane
+.\" in order to generate smoother transparency gradients.
 .TP
 .B \-nodither
 Disable all dithering (default).
@@ -75,6 +87,10 @@ This cropping area must be fully contained within the source rectangle.
 The top-left corner will be snapped to even coordinates if needed.
 This option is meant to reduce the memory needed for cropping large images.
 Note: the cropping is applied \fIbefore\fP any scaling.
+.\" TODO(jzern): restore post-v0.4.1
+.\" .TP
+.\" .B \-flip
+.\" Flip decoded image vertically (can be useful for OpenGL textures for instance).
 .TP
 .BI \-scale " width height
 Rescale the decoded picture to dimension \fBwidth\fP x \fBheight\fP. This
@@ -101,6 +117,8 @@ dwebp picture.webp \-o output.png
 dwebp picture.webp \-ppm \-o output.ppm
 .br
 dwebp \-o output.ppm \-\- \-\-\-picture.webp
+.br
+cat picture.webp | dwebp \-o \- \-\- \- > output.ppm
 
 .SH AUTHORS
 \fBdwebp\fP was written by the WebP team.

man/gif2webp.1

@@ -1,5 +1,5 @@
 .\"                                      Hey, EMACS: -*- nroff -*-
-.TH GIF2WEBP 1 "December 17, 2013"
+.TH GIF2WEBP 1 "March 7, 2014"
 .SH NAME
 gif2webp \- Convert a GIF image to WebP
 .SH SYNOPSIS
@@ -18,6 +18,7 @@ The basic options are:
 .BI \-o " string
 Specify the name of the output WebP file. If omitted, \fBgif2webp\fP will
 perform conversion but only report statistics.
+Using "\-" as output name will direct output to 'stdout'.
 .TP
 .B \-h, \-help
 Usage information.

man/vwebp.1

@@ -0,0 +1,91 @@
+.\"                                      Hey, EMACS: -*- nroff -*-
+.TH VWEBP 1 "July 23, 2014"
+.SH NAME
+vwebp \- decompress a WebP file and display it in a window
+.SH SYNOPSIS
+.B vwebp
+.RI [ options ] " input_file.webp
+.br
+.SH DESCRIPTION
+This manual page documents the
+.B vwebp
+command.
+.PP
+\fBvwebp\fP decompresses a WebP file and displays it in a window using OpenGL.
+.SH OPTIONS
+.TP
+.B \-h
+Print usage summary.
+.TP
+.B \-version
+Print version number and exit.
+.TP
+.B \-noicc
+Don't use the ICC profile if present.
+.TP
+.B \-nofancy
+Don't use the fancy YUV420 upscaler.
+.TP
+.B \-nofilter
+Disable in-loop filtering.
+.TP
+.BI \-dither " strength
+Specify a dithering \fBstrength\fP between 0 and 100. Dithering is a
+post-processing effect applied to chroma components in lossy compression.
+It helps by smoothing gradients and avoiding banding artifacts. Default: 50.
+.\" TODO(jzern): restore post-v0.4.1
+.\" .TP
+.\" .BI \-noalphadither
+.\" By default, quantized transparency planes are dithered during decompression,
+.\" to smooth the gradients. This flag will prevent this dithering.
+.TP
+.B \-mt
+Use multi-threading for decoding, if possible.
+.TP
+.B \-info
+Display image information on top of the decoded image.
+.TP
+.BI \-\- " string
+Explicitly specify the input file. This option is useful if the input
+file starts with an '\-' for instance. This option must appear \fBlast\fP.
+Any other options afterward will be ignored. If the input file is "\-",
+the data will be read from \fIstdin\fP instead of a file.
+.TP
+
+.SH KEYBOARD SHORTCUTS
+.TP
+.B 'c'
+Toggle use of color profile.
+.TP
+.B 'i'
+Overlay file information.
+.TP
+.B 'q' / 'Q' / ESC
+Quit.
+
+.SH BUGS
+Please report all bugs to our issue tracker:
+http://code.google.com/p/webp/issues
+.br
+Patches welcome! See this page to get started:
+http://www.webmproject.org/code/contribute/submitting-patches/
+
+.SH EXAMPLES
+vwebp picture.webp
+.br
+vwebp picture.webp -mt -dither 0
+.br
+vwebp \-\- \-\-\-picture.webp
+
+.SH AUTHORS
+\fBvwebp\fP was written by the WebP team.
+.br
+The latest source tree is available at http://www.webmproject.org/code
+.PP
+This manual page was written for the Debian project (and may be used by others).
+
+.SH SEE ALSO
+.BR dwebp (1)
+.br
+Please refer to http://developers.google.com/speed/webp/ for additional
+information.

man/webpmux.1

@@ -1,7 +1,8 @@
 .\"                                      Hey, EMACS: -*- nroff -*-
-.TH WEBPMUX 1 "December 17, 2013"
+.TH WEBPMUX 1 "August 28, 2014"
 .SH NAME
-webpmux \- command line tool to create WebP Mux/container file.
+webpmux \- create animated WebP files from non\-animated WebP images, extract
+frames from animated WebP images, and manage XMP/EXIF metadata and ICC profile.
 .SH SYNOPSIS
 .B webpmux \-get
 .I GET_OPTIONS
@@ -45,8 +46,8 @@ This manual page documents the
 .B webpmux
 command.
 .PP
-\fBwebpmux\fP can be used to create a WebP container file
-and extract/strip relevant data from the container file.
+\fBwebpmux\fP can be used to create/extract from animated WebP files, as well as
+to add/extract/strip XMP/EXIF metadata and ICC profile.
 .SH OPTIONS
 .SS GET_OPTIONS (\-get):
 .TP
@@ -60,7 +61,7 @@ Get EXIF metadata.
 Get XMP metadata.
 .TP
 .BI frame " n
-Get nth frame.
+Get nth frame from an animated image. (n = 0 has a special meaning: last frame).
 
 .SS SET_OPTIONS (\-set)
 .TP
@@ -91,12 +92,13 @@ Strip EXIF metadata.
 Strip XMP metadata.
 
 .SS FRAME_OPTIONS (\-frame)
+Create an animated WebP file from multiple (non\-animated) WebP images.
 .TP
 .I file_i +di[+xi+yi[+mi[bi]]]
 Where: 'file_i' is the i'th frame (WebP format), 'xi','yi' specify the image
 offset for this frame, 'di' is the pause duration before next frame, 'mi' is
 the dispose method for this frame (0 for NONE or 1 for BACKGROUND) and 'bi' is
-the blending method for this frame (+b for BLEND or -b for NO_BLEND).
+the blending method for this frame (+b for BLEND or \-b for NO_BLEND).
 Argument 'bi' can be omitted and will default to +b (BLEND).
 Also, 'mi' can be omitted if 'bi' is omitted and will default to 0 (NONE).
 Finally, if 'mi' and 'bi' are omitted then 'xi' and 'yi' can be omitted and will
@@ -130,37 +132,61 @@ Please report all bugs to our issue tracker:
 http://code.google.com/p/webp/issues
 .br
 Patches welcome! See this page to get started:
-http://www.webmproject.org/code/contribute/submitting-patches/
+http://www.webmproject.org/code/contribute/submitting\-patches/
 
 .SH EXAMPLES
+.P
+Add ICC profile:
+.br
 webpmux \-set icc image_profile.icc in.webp \-o icc_container.webp
+.P
+Extract ICC profile:
 .br
 webpmux \-get icc icc_container.webp \-o image_profile.icc
+.P
+Strip ICC profile:
 .br
 webpmux \-strip icc icc_container.webp \-o without_icc.webp
+.P
+Add XMP metadata:
 .br
 webpmux \-set xmp image_metadata.xmp in.webp \-o xmp_container.webp
+.P
+Extract XMP metadata:
 .br
 webpmux \-get xmp xmp_container.webp \-o image_metadata.xmp
+.P
+Strip XMP metadata:
 .br
 webpmux \-strip xmp xmp_container.webp \-o without_xmp.webp
+.P
+Add EXIF metadata:
 .br
 webpmux \-set exif image_metadata.exif in.webp \-o exif_container.webp
+.P
+Extract EXIF metadata:
 .br
 webpmux \-get exif exif_container.webp \-o image_metadata.exif
+.P
+Strip EXIF metadata:
 .br
 webpmux \-strip exif exif_container.webp \-o without_exif.webp
+.P
+Create an animated WebP file from 3 (non\-animated) WebP images:
 .br
-webpmux \-frame anim_1.webp +100 \-frame anim_2.webp +100+50+50
+webpmux \-frame 1.webp +100 \-frame 2.webp +100+50+50
 .br
 .RS 8
-\-frame anim_2.webp +100+50+50+1+b \-loop 10 \-bgcolor 255,255,255,255
+\-frame 3.webp +100+50+50+1+b \-loop 10 \-bgcolor 255,255,255,255
 .br
-.RS 8
 \-o anim_container.webp
 .RE
+.P
+Get the 2nd frame from an animated WebP file:
 .br
 webpmux \-get frame 2 anim_container.webp \-o frame_2.webp
+.P
+Using \-get/\-set/\-strip with input file name starting with '\-':
 .br
 webpmux \-set icc image_profile.icc \-o icc_container.webp \-\- \-\-\-in.webp
 .br

src/.gitignore

@@ -1 +0,0 @@
-/*.pc

src/Makefile.am

@@ -8,7 +8,6 @@ if WANT_DEMUX
   SUBDIRS += demux
 endif
 
-AM_CPPFLAGS = -I$(top_srcdir)/src
 lib_LTLIBRARIES = libwebp.la
 
 if BUILD_LIBWEBPDECODER
@@ -36,7 +35,7 @@ libwebp_la_LIBADD += utils/libwebputils.la
 # other than the ones listed on the command line, i.e., after linking, it will
 # not have unresolved symbols. Some platforms (Windows among them) require all
 # symbols in shared libraries to be resolved at library creation.
-libwebp_la_LDFLAGS = -no-undefined -version-info 5:0:0
+libwebp_la_LDFLAGS = -no-undefined -version-info 5:3:0
 libwebpincludedir = $(includedir)/webp
 pkgconfig_DATA = libwebp.pc
 
@@ -48,7 +47,7 @@ if BUILD_LIBWEBPDECODER
   libwebpdecoder_la_LIBADD += dsp/libwebpdspdecode.la
   libwebpdecoder_la_LIBADD += utils/libwebputilsdecode.la
 
-  libwebpdecoder_la_LDFLAGS = -no-undefined -version-info 1:0:0
+  libwebpdecoder_la_LDFLAGS = -no-undefined -version-info 1:3:0
   pkgconfig_DATA += libwebpdecoder.pc
 endif
 

src/dec/Makefile.am

@@ -1,4 +1,3 @@
-AM_CPPFLAGS = -I$(top_srcdir)/src
 noinst_LTLIBRARIES = libwebpdecode.la
 
 libwebpdecode_la_SOURCES =
@@ -9,7 +8,6 @@ libwebpdecode_la_SOURCES += decode_vp8.h
 libwebpdecode_la_SOURCES += frame.c
 libwebpdecode_la_SOURCES += idec.c
 libwebpdecode_la_SOURCES += io.c
-libwebpdecode_la_SOURCES += layer.c
 libwebpdecode_la_SOURCES += quant.c
 libwebpdecode_la_SOURCES += tree.c
 libwebpdecode_la_SOURCES += vp8.c

src/dec/alpha.c

@@ -16,13 +16,14 @@
 #include "./vp8i.h"
 #include "./vp8li.h"
 #include "../utils/quant_levels_dec.h"
+#include "../utils/utils.h"
 #include "../webp/format_constants.h"
 
 //------------------------------------------------------------------------------
 // ALPHDecoder object.
 
 ALPHDecoder* ALPHNew(void) {
-  ALPHDecoder* const dec = (ALPHDecoder*)calloc(1, sizeof(*dec));
+  ALPHDecoder* const dec = (ALPHDecoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
   return dec;
 }
 
@@ -30,7 +31,7 @@ void ALPHDelete(ALPHDecoder* const dec) {
   if (dec != NULL) {
     VP8LDelete(dec->vp8l_dec_);
     dec->vp8l_dec_ = NULL;
-    free(dec);
+    WebPSafeFree(dec);
   }
 }
 
@@ -107,12 +108,6 @@ static int ALPHDecode(VP8Decoder* const dec, int row, int num_rows) {
     unfilter_func(width, height, width, row, num_rows, output);
   }
 
-  if (alph_dec->pre_processing_ == ALPHA_PREPROCESSED_LEVELS) {
-    if (!DequantizeLevels(output, width, height, row, num_rows)) {
-      return 0;
-    }
-  }
-
   if (row + num_rows == dec->pic_hdr_.height_) {
     dec->is_alpha_decoded_ = 1;
   }
@@ -142,12 +137,22 @@ const uint8_t* VP8DecompressAlphaRows(VP8Decoder* const dec,
       dec->alph_dec_ = NULL;
       return NULL;
     }
+    // if we allowed use of alpha dithering, check whether it's needed at all
+    if (dec->alph_dec_->pre_processing_ != ALPHA_PREPROCESSED_LEVELS) {
+      dec->alpha_dithering_ = 0;  // disable dithering
+    } else {
+      num_rows = height;          // decode everything in one pass
+    }
   }
 
   if (!dec->is_alpha_decoded_) {
     int ok = 0;
     assert(dec->alph_dec_ != NULL);
     ok = ALPHDecode(dec, row, num_rows);
+    if (ok && dec->alpha_dithering_ > 0) {
+      ok = WebPDequantizeLevels(dec->alpha_plane_, width, height,
+                                dec->alpha_dithering_);
+    }
     if (!ok || dec->is_alpha_decoded_) {
       ALPHDelete(dec->alph_dec_);
       dec->alph_dec_ = NULL;
@@ -158,4 +163,3 @@ const uint8_t* VP8DecompressAlphaRows(VP8Decoder* const dec,
   // Return a pointer to the current decoded row.
   return dec->alpha_plane_ + row * width;
 }
-

src/dec/buffer.c

@@ -42,29 +42,34 @@ static VP8StatusCode CheckDecBuffer(const WebPDecBuffer* const buffer) {
     ok = 0;
   } else if (!WebPIsRGBMode(mode)) {   // YUV checks
     const WebPYUVABuffer* const buf = &buffer->u.YUVA;
-    const uint64_t y_size = (uint64_t)buf->y_stride * height;
-    const uint64_t u_size = (uint64_t)buf->u_stride * ((height + 1) / 2);
-    const uint64_t v_size = (uint64_t)buf->v_stride * ((height + 1) / 2);
-    const uint64_t a_size = (uint64_t)buf->a_stride * height;
+    const int y_stride = abs(buf->y_stride);
+    const int u_stride = abs(buf->u_stride);
+    const int v_stride = abs(buf->v_stride);
+    const int a_stride = abs(buf->a_stride);
+    const uint64_t y_size = (uint64_t)y_stride * height;
+    const uint64_t u_size = (uint64_t)u_stride * ((height + 1) / 2);
+    const uint64_t v_size = (uint64_t)v_stride * ((height + 1) / 2);
+    const uint64_t a_size = (uint64_t)a_stride * height;
     ok &= (y_size <= buf->y_size);
     ok &= (u_size <= buf->u_size);
     ok &= (v_size <= buf->v_size);
-    ok &= (buf->y_stride >= width);
-    ok &= (buf->u_stride >= (width + 1) / 2);
-    ok &= (buf->v_stride >= (width + 1) / 2);
+    ok &= (y_stride >= width);
+    ok &= (u_stride >= (width + 1) / 2);
+    ok &= (v_stride >= (width + 1) / 2);
     ok &= (buf->y != NULL);
     ok &= (buf->u != NULL);
     ok &= (buf->v != NULL);
     if (mode == MODE_YUVA) {
-      ok &= (buf->a_stride >= width);
+      ok &= (a_stride >= width);
       ok &= (a_size <= buf->a_size);
       ok &= (buf->a != NULL);
     }
   } else {    // RGB checks
     const WebPRGBABuffer* const buf = &buffer->u.RGBA;
-    const uint64_t size = (uint64_t)buf->stride * height;
+    const int stride = abs(buf->stride);
+    const uint64_t size = (uint64_t)stride * height;
     ok &= (size <= buf->size);
-    ok &= (buf->stride >= width * kModeBpp[mode]);
+    ok &= (stride >= width * kModeBpp[mode]);
     ok &= (buf->rgba != NULL);
   }
   return ok ? VP8_STATUS_OK : VP8_STATUS_INVALID_PARAM;
@@ -131,9 +136,35 @@ static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
   return CheckDecBuffer(buffer);
 }
 
+VP8StatusCode WebPFlipBuffer(WebPDecBuffer* const buffer) {
+  if (buffer == NULL) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  if (WebPIsRGBMode(buffer->colorspace)) {
+    WebPRGBABuffer* const buf = &buffer->u.RGBA;
+    buf->rgba += (buffer->height - 1) * buf->stride;
+    buf->stride = -buf->stride;
+  } else {
+    WebPYUVABuffer* const buf = &buffer->u.YUVA;
+    const int H = buffer->height;
+    buf->y += (H - 1) * buf->y_stride;
+    buf->y_stride = -buf->y_stride;
+    buf->u += ((H - 1) >> 1) * buf->u_stride;
+    buf->u_stride = -buf->u_stride;
+    buf->v += ((H - 1) >> 1) * buf->v_stride;
+    buf->v_stride = -buf->v_stride;
+    if (buf->a != NULL) {
+      buf->a += (H - 1) * buf->a_stride;
+      buf->a_stride = -buf->a_stride;
+    }
+  }
+  return VP8_STATUS_OK;
+}
+
 VP8StatusCode WebPAllocateDecBuffer(int w, int h,
                                     const WebPDecoderOptions* const options,
                                     WebPDecBuffer* const out) {
+  VP8StatusCode status;
   if (out == NULL || w <= 0 || h <= 0) {
     return VP8_STATUS_INVALID_PARAM;
   }
@@ -160,8 +191,17 @@ VP8StatusCode WebPAllocateDecBuffer(int w, int h,
   out->width = w;
   out->height = h;
 
-  // Then, allocate buffer for real
-  return AllocateBuffer(out);
+  // Then, allocate buffer for real.
+  status = AllocateBuffer(out);
+  if (status != VP8_STATUS_OK) return status;
+
+#if WEBP_DECODER_ABI_VERSION > 0x0203
+  // Use the stride trick if vertical flip is needed.
+  if (options != NULL && options->flip) {
+    status = WebPFlipBuffer(out);
+  }
+#endif
+  return status;
 }
 
 //------------------------------------------------------------------------------
@@ -178,8 +218,9 @@ int WebPInitDecBufferInternal(WebPDecBuffer* buffer, int version) {
 
 void WebPFreeDecBuffer(WebPDecBuffer* buffer) {
   if (buffer != NULL) {
-    if (!buffer->is_external_memory)
-      free(buffer->private_memory);
+    if (!buffer->is_external_memory) {
+      WebPSafeFree(buffer->private_memory);
+    }
     buffer->private_memory = NULL;
   }
 }

src/dec/frame.c

@@ -177,6 +177,15 @@ void VP8InitDithering(const WebPDecoderOptions* const options,
         dec->dither_ = 1;
       }
     }
+#if WEBP_DECODER_ABI_VERSION > 0x0204
+    // potentially allow alpha dithering
+    dec->alpha_dithering_ = options->alpha_dithering_strength;
+    if (dec->alpha_dithering_ > 100) {
+      dec->alpha_dithering_ = 100;
+    } else if (dec->alpha_dithering_ < 0) {
+      dec->alpha_dithering_ = 0;
+    }
+#endif
   }
 }
 
@@ -347,7 +356,7 @@ int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io) {
   } else {
     WebPWorker* const worker = &dec->worker_;
     // Finish previous job *before* updating context
-    ok &= WebPWorkerSync(worker);
+    ok &= WebPGetWorkerInterface()->Sync(worker);
     assert(worker->status_ == OK);
     if (ok) {   // spawn a new deblocking/output job
       ctx->io_ = *io;
@@ -367,7 +376,8 @@ int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io) {
         ctx->f_info_ = dec->f_info_;
         dec->f_info_ = tmp;
       }
-      WebPWorkerLaunch(worker);    // (reconstruct)+filter in parallel
+      // (reconstruct)+filter in parallel
+      WebPGetWorkerInterface()->Launch(worker);
       if (++dec->cache_id_ == dec->num_caches_) {
         dec->cache_id_ = 0;
       }
@@ -437,7 +447,7 @@ VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io) {
 int VP8ExitCritical(VP8Decoder* const dec, VP8Io* const io) {
   int ok = 1;
   if (dec->mt_method_ > 0) {
-    ok = WebPWorkerSync(&dec->worker_);
+    ok = WebPGetWorkerInterface()->Sync(&dec->worker_);
   }
 
   if (io->teardown != NULL) {
@@ -478,7 +488,7 @@ static int InitThreadContext(VP8Decoder* const dec) {
   dec->cache_id_ = 0;
   if (dec->mt_method_ > 0) {
     WebPWorker* const worker = &dec->worker_;
-    if (!WebPWorkerReset(worker)) {
+    if (!WebPGetWorkerInterface()->Reset(worker)) {
       return VP8SetError(dec, VP8_STATUS_OUT_OF_MEMORY,
                          "thread initialization failed.");
     }
@@ -502,7 +512,7 @@ int VP8GetThreadMethod(const WebPDecoderOptions* const options,
   (void)headers;
   (void)width;
   (void)height;
-  assert(!headers->is_lossless);
+  assert(headers == NULL || !headers->is_lossless);
 #if defined(WEBP_USE_THREAD)
   if (width < MIN_WIDTH_FOR_THREADS) return 0;
   // TODO(skal): tune the heuristic further
@@ -549,7 +559,7 @@ static int AllocateMemory(VP8Decoder* const dec) {
 
   if (needed != (size_t)needed) return 0;  // check for overflow
   if (needed > dec->mem_size_) {
-    free(dec->mem_);
+    WebPSafeFree(dec->mem_);
     dec->mem_size_ = 0;
     dec->mem_ = WebPSafeMalloc(needed, sizeof(uint8_t));
     if (dec->mem_ == NULL) {

src/dec/idec.c

@@ -72,28 +72,20 @@ struct WebPIDecoder {
   MemBuffer mem_;          // input memory buffer.
   WebPDecBuffer output_;   // output buffer (when no external one is supplied)
   size_t chunk_size_;      // Compressed VP8/VP8L size extracted from Header.
+
+  int last_mb_y_;          // last row reached for intra-mode decoding
 };
 
 // MB context to restore in case VP8DecodeMB() fails
 typedef struct {
   VP8MB left_;
   VP8MB info_;
-  uint8_t intra_t_[4];
-  uint8_t intra_l_[4];
-  VP8BitReader br_;
   VP8BitReader token_br_;
 } MBContext;
 
 //------------------------------------------------------------------------------
 // MemBuffer: incoming data handling
 
-static void RemapBitReader(VP8BitReader* const br, ptrdiff_t offset) {
-  if (br->buf_ != NULL) {
-    br->buf_ += offset;
-    br->buf_end_ += offset;
-  }
-}
-
 static WEBP_INLINE size_t MemDataSize(const MemBuffer* mem) {
   return (mem->end_ - mem->start_);
 }
@@ -130,12 +122,12 @@ static void DoRemap(WebPIDecoder* const idec, ptrdiff_t offset) {
       if (offset != 0) {
         int p;
         for (p = 0; p <= last_part; ++p) {
-          RemapBitReader(dec->parts_ + p, offset);
+          VP8RemapBitReader(dec->parts_ + p, offset);
         }
         // Remap partition #0 data pointer to new offset, but only in MAP
         // mode (in APPEND mode, partition #0 is copied into a fixed memory).
         if (mem->mode_ == MEM_MODE_MAP) {
-          RemapBitReader(&dec->br_, offset);
+          VP8RemapBitReader(&dec->br_, offset);
         }
       }
       assert(last_part >= 0);
@@ -189,7 +181,7 @@ static int AppendToMemBuffer(WebPIDecoder* const idec,
         (uint8_t*)WebPSafeMalloc(extra_size, sizeof(*new_buf));
     if (new_buf == NULL) return 0;
     memcpy(new_buf, old_base, current_size);
-    free(mem->buf_);
+    WebPSafeFree(mem->buf_);
     mem->buf_ = new_buf;
     mem->buf_size_ = (size_t)extra_size;
     mem->start_ = new_mem_start;
@@ -231,8 +223,8 @@ static void InitMemBuffer(MemBuffer* const mem) {
 static void ClearMemBuffer(MemBuffer* const mem) {
   assert(mem);
   if (mem->mode_ == MEM_MODE_APPEND) {
-    free(mem->buf_);
-    free((void*)mem->part0_buf_);
+    WebPSafeFree(mem->buf_);
+    WebPSafeFree((void*)mem->part0_buf_);
   }
 }
 
@@ -246,35 +238,36 @@ static int CheckMemBufferMode(MemBuffer* const mem, MemBufferMode expected) {
   return 1;
 }
 
+// To be called last.
+static VP8StatusCode FinishDecoding(WebPIDecoder* const idec) {
+#if WEBP_DECODER_ABI_VERSION > 0x0203
+  const WebPDecoderOptions* const options = idec->params_.options;
+  WebPDecBuffer* const output = idec->params_.output;
+
+  idec->state_ = STATE_DONE;
+  if (options != NULL && options->flip) {
+    return WebPFlipBuffer(output);
+  }
+#endif
+  idec->state_ = STATE_DONE;
+  return VP8_STATUS_OK;
+}
+
 //------------------------------------------------------------------------------
 // Macroblock-decoding contexts
 
 static void SaveContext(const VP8Decoder* dec, const VP8BitReader* token_br,
                         MBContext* const context) {
-  const VP8BitReader* const br = &dec->br_;
-  const VP8MB* const left = dec->mb_info_ - 1;
-  const VP8MB* const info = dec->mb_info_ + dec->mb_x_;
-
-  context->left_ = *left;
-  context->info_ = *info;
-  context->br_ = *br;
+  context->left_ = dec->mb_info_[-1];
+  context->info_ = dec->mb_info_[dec->mb_x_];
   context->token_br_ = *token_br;
-  memcpy(context->intra_t_, dec->intra_t_ + 4 * dec->mb_x_, 4);
-  memcpy(context->intra_l_, dec->intra_l_, 4);
 }
 
 static void RestoreContext(const MBContext* context, VP8Decoder* const dec,
                            VP8BitReader* const token_br) {
-  VP8BitReader* const br = &dec->br_;
-  VP8MB* const left = dec->mb_info_ - 1;
-  VP8MB* const info = dec->mb_info_ + dec->mb_x_;
-
-  *left = context->left_;
-  *info = context->info_;
-  *br = context->br_;
+  dec->mb_info_[-1] = context->left_;
+  dec->mb_info_[dec->mb_x_] = context->info_;
   *token_br = context->token_br_;
-  memcpy(dec->intra_t_ + 4 * dec->mb_x_, context->intra_t_, 4);
-  memcpy(dec->intra_l_, context->intra_l_, 4);
 }
 
 //------------------------------------------------------------------------------
@@ -310,6 +303,7 @@ static VP8StatusCode DecodeWebPHeaders(WebPIDecoder* const idec) {
 
   headers.data = data;
   headers.data_size = curr_size;
+  headers.have_all_data = 0;
   status = WebPParseHeaders(&headers);
   if (status == VP8_STATUS_NOT_ENOUGH_DATA) {
     return VP8_STATUS_SUSPENDED;  // We haven't found a VP8 chunk yet.
@@ -363,30 +357,33 @@ static VP8StatusCode DecodeVP8FrameHeader(WebPIDecoder* const idec) {
 }
 
 // Partition #0
-static int CopyParts0Data(WebPIDecoder* const idec) {
+static VP8StatusCode CopyParts0Data(WebPIDecoder* const idec) {
   VP8Decoder* const dec = (VP8Decoder*)idec->dec_;
   VP8BitReader* const br = &dec->br_;
-  const size_t psize = br->buf_end_ - br->buf_;
+  const size_t part_size = br->buf_end_ - br->buf_;
   MemBuffer* const mem = &idec->mem_;
   assert(!idec->is_lossless_);
   assert(mem->part0_buf_ == NULL);
-  assert(psize > 0);
-  assert(psize <= mem->part0_size_);  // Format limit: no need for runtime check
+  // the following is a format limitation, no need for runtime check:
+  assert(part_size <= mem->part0_size_);
+  if (part_size == 0) {   // can't have zero-size partition #0
+    return VP8_STATUS_BITSTREAM_ERROR;
+  }
   if (mem->mode_ == MEM_MODE_APPEND) {
     // We copy and grab ownership of the partition #0 data.
-    uint8_t* const part0_buf = (uint8_t*)malloc(psize);
+    uint8_t* const part0_buf = (uint8_t*)WebPSafeMalloc(1ULL, part_size);
     if (part0_buf == NULL) {
-      return 0;
+      return VP8_STATUS_OUT_OF_MEMORY;
     }
-    memcpy(part0_buf, br->buf_, psize);
+    memcpy(part0_buf, br->buf_, part_size);
     mem->part0_buf_ = part0_buf;
     br->buf_ = part0_buf;
-    br->buf_end_ = part0_buf + psize;
+    br->buf_end_ = part0_buf + part_size;
   } else {
     // Else: just keep pointers to the partition #0's data in dec_->br_.
   }
-  mem->start_ += psize;
-  return 1;
+  mem->start_ += part_size;
+  return VP8_STATUS_OK;
 }
 
 static VP8StatusCode DecodePartition0(WebPIDecoder* const idec) {
@@ -420,8 +417,10 @@ static VP8StatusCode DecodePartition0(WebPIDecoder* const idec) {
   dec->mt_method_ = VP8GetThreadMethod(params->options, NULL,
                                        io->width, io->height);
   VP8InitDithering(params->options, dec);
-  if (!CopyParts0Data(idec)) {
-    return IDecError(idec, VP8_STATUS_OUT_OF_MEMORY);
+
+  dec->status_ = CopyParts0Data(idec);
+  if (dec->status_ != VP8_STATUS_OK) {
+    return IDecError(idec, dec->status_);
   }
 
   // Finish setting up the decoding parameters. Will call io->setup().
@@ -446,16 +445,26 @@ static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
 
   assert(dec->ready_);
   for (; dec->mb_y_ < dec->mb_h_; ++dec->mb_y_) {
-    VP8BitReader* token_br = &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
+    if (idec->last_mb_y_ != dec->mb_y_) {
+      if (!VP8ParseIntraModeRow(&dec->br_, dec)) {
+        // note: normally, error shouldn't occur since we already have the whole
+        // partition0 available here in DecodeRemaining(). Reaching EOF while
+        // reading intra modes really means a BITSTREAM_ERROR.
+        return IDecError(idec, VP8_STATUS_BITSTREAM_ERROR);
+      }
+      idec->last_mb_y_ = dec->mb_y_;
+    }
     for (; dec->mb_x_ < dec->mb_w_; ++dec->mb_x_) {
+      VP8BitReader* const token_br =
+          &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
       MBContext context;
       SaveContext(dec, token_br, &context);
       if (!VP8DecodeMB(dec, token_br)) {
-        RestoreContext(&context, dec, token_br);
         // We shouldn't fail when MAX_MB data was available
         if (dec->num_parts_ == 1 && MemDataSize(&idec->mem_) > MAX_MB_SIZE) {
           return IDecError(idec, VP8_STATUS_BITSTREAM_ERROR);
         }
+        RestoreContext(&context, dec, token_br);
         return VP8_STATUS_SUSPENDED;
       }
       // Release buffer only if there is only one partition
@@ -476,9 +485,7 @@ static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
     return IDecError(idec, VP8_STATUS_USER_ABORT);
   }
   dec->ready_ = 0;
-  idec->state_ = STATE_DONE;
-
-  return VP8_STATUS_OK;
+  return FinishDecoding(idec);
 }
 
 static VP8StatusCode ErrorStatusLossless(WebPIDecoder* const idec,
@@ -527,12 +534,16 @@ static VP8StatusCode DecodeVP8LData(WebPIDecoder* const idec) {
   }
 
   if (!VP8LDecodeImage(dec)) {
+    // The decoding is called after all the data-bytes are aggregated. Change
+    // the error to VP8_BITSTREAM_ERROR in case lossless decoder fails to decode
+    // all the pixels (VP8_STATUS_SUSPENDED).
+    if (dec->status_ == VP8_STATUS_SUSPENDED) {
+      dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+    }
     return ErrorStatusLossless(idec, dec->status_);
   }
 
-  idec->state_ = STATE_DONE;
-
-  return VP8_STATUS_OK;
+  return FinishDecoding(idec);
 }
 
   // Main decoding loop
@@ -568,7 +579,7 @@ static VP8StatusCode IDecode(WebPIDecoder* idec) {
 // Public functions
 
 WebPIDecoder* WebPINewDecoder(WebPDecBuffer* output_buffer) {
-  WebPIDecoder* idec = (WebPIDecoder*)calloc(1, sizeof(*idec));
+  WebPIDecoder* idec = (WebPIDecoder*)WebPSafeCalloc(1ULL, sizeof(*idec));
   if (idec == NULL) {
     return NULL;
   }
@@ -576,6 +587,8 @@ WebPIDecoder* WebPINewDecoder(WebPDecBuffer* output_buffer) {
   idec->state_ = STATE_WEBP_HEADER;
   idec->chunk_size_ = 0;
 
+  idec->last_mb_y_ = -1;
+
   InitMemBuffer(&idec->mem_);
   WebPInitDecBuffer(&idec->output_);
   VP8InitIo(&idec->io_);
@@ -625,7 +638,7 @@ void WebPIDelete(WebPIDecoder* idec) {
   }
   ClearMemBuffer(&idec->mem_);
   WebPFreeDecBuffer(&idec->output_);
-  free(idec);
+  WebPSafeFree(idec);
 }
 
 //------------------------------------------------------------------------------

src/dec/io.c

@@ -17,6 +17,7 @@
 #include "./webpi.h"
 #include "../dsp/dsp.h"
 #include "../dsp/yuv.h"
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 // Main YUV<->RGB conversion functions
@@ -44,27 +45,13 @@ static int EmitYUV(const VP8Io* const io, WebPDecParams* const p) {
 
 // Point-sampling U/V sampler.
 static int EmitSampledRGB(const VP8Io* const io, WebPDecParams* const p) {
-  WebPDecBuffer* output = p->output;
-  const WebPRGBABuffer* const buf = &output->u.RGBA;
-  uint8_t* dst = buf->rgba + io->mb_y * buf->stride;
-  const uint8_t* y_src = io->y;
-  const uint8_t* u_src = io->u;
-  const uint8_t* v_src = io->v;
-  const WebPSampleLinePairFunc sample = WebPSamplers[output->colorspace];
-  const int mb_w = io->mb_w;
-  const int last = io->mb_h - 1;
-  int j;
-  for (j = 0; j < last; j += 2) {
-    sample(y_src, y_src + io->y_stride, u_src, v_src,
-           dst, dst + buf->stride, mb_w);
-    y_src += 2 * io->y_stride;
-    u_src += io->uv_stride;
-    v_src += io->uv_stride;
-    dst += 2 * buf->stride;
-  }
-  if (j == last) {  // Just do the last line twice
-    sample(y_src, y_src, u_src, v_src, dst, dst, mb_w);
-  }
+  WebPDecBuffer* const output = p->output;
+  WebPRGBABuffer* const buf = &output->u.RGBA;
+  uint8_t* const dst = buf->rgba + io->mb_y * buf->stride;
+  WebPSamplerProcessPlane(io->y, io->y_stride,
+                          io->u, io->v, io->uv_stride,
+                          dst, buf->stride, io->mb_w, io->mb_h,
+                          WebPSamplers[output->colorspace]);
   return io->mb_h;
 }
 
@@ -250,7 +237,11 @@ static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
     int num_rows;
     const int start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
     uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
+#ifdef WEBP_SWAP_16BIT_CSP
+    uint8_t* alpha_dst = base_rgba;
+#else
     uint8_t* alpha_dst = base_rgba + 1;
+#endif
     uint32_t alpha_mask = 0x0f;
     int i, j;
 
@@ -289,7 +280,17 @@ static int Rescale(const uint8_t* src, int src_stride,
 static int EmitRescaledYUV(const VP8Io* const io, WebPDecParams* const p) {
   const int mb_h = io->mb_h;
   const int uv_mb_h = (mb_h + 1) >> 1;
-  const int num_lines_out = Rescale(io->y, io->y_stride, mb_h, &p->scaler_y);
+  WebPRescaler* const scaler = &p->scaler_y;
+  int num_lines_out = 0;
+  if (WebPIsAlphaMode(p->output->colorspace) && io->a != NULL) {
+    // Before rescaling, we premultiply the luma directly into the io->y
+    // internal buffer. This is OK since these samples are not used for
+    // intra-prediction (the top samples are saved in cache_y_/u_/v_).
+    // But we need to cast the const away, though.
+    WebPMultRows((uint8_t*)io->y, io->y_stride,
+                 io->a, io->width, io->mb_w, mb_h, 0);
+  }
+  num_lines_out = Rescale(io->y, io->y_stride, mb_h, scaler);
   Rescale(io->u, io->uv_stride, uv_mb_h, &p->scaler_u);
   Rescale(io->v, io->uv_stride, uv_mb_h, &p->scaler_v);
   return num_lines_out;
@@ -297,7 +298,14 @@ static int EmitRescaledYUV(const VP8Io* const io, WebPDecParams* const p) {
 
 static int EmitRescaledAlphaYUV(const VP8Io* const io, WebPDecParams* const p) {
   if (io->a != NULL) {
-    Rescale(io->a, io->width, io->mb_h, &p->scaler_a);
+    const WebPYUVABuffer* const buf = &p->output->u.YUVA;
+    uint8_t* dst_y = buf->y + p->last_y * buf->y_stride;
+    const uint8_t* src_a = buf->a + p->last_y * buf->a_stride;
+    const int num_lines_out = Rescale(io->a, io->width, io->mb_h, &p->scaler_a);
+    if (num_lines_out > 0) {   // unmultiply the Y
+      WebPMultRows(dst_y, buf->y_stride, src_a, buf->a_stride,
+                   p->scaler_a.dst_width, num_lines_out, 1);
+    }
   }
   return 0;
 }
@@ -316,11 +324,11 @@ static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
   size_t tmp_size;
   int32_t* work;
 
-  tmp_size = work_size + 2 * uv_work_size;
+  tmp_size = (work_size + 2 * uv_work_size) * sizeof(*work);
   if (has_alpha) {
-    tmp_size += work_size;
+    tmp_size += work_size * sizeof(*work);
   }
-  p->memory = calloc(1, tmp_size * sizeof(*work));
+  p->memory = WebPSafeCalloc(1ULL, tmp_size);
   if (p->memory == NULL) {
     return 0;   // memory error
   }
@@ -347,6 +355,7 @@ static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
                      io->mb_w, out_width, io->mb_h, out_height,
                      work + work_size + 2 * uv_work_size);
     p->emit_alpha = EmitRescaledAlphaYUV;
+    WebPInitAlphaProcessing();
   }
   return 1;
 }
@@ -366,9 +375,9 @@ static int ExportRGB(WebPDecParams* const p, int y_pos) {
          WebPRescalerHasPendingOutput(&p->scaler_u)) {
     assert(p->last_y + y_pos + num_lines_out < p->output->height);
     assert(p->scaler_u.y_accum == p->scaler_v.y_accum);
-    WebPRescalerExportRow(&p->scaler_y);
-    WebPRescalerExportRow(&p->scaler_u);
-    WebPRescalerExportRow(&p->scaler_v);
+    WebPRescalerExportRow(&p->scaler_y, 0);
+    WebPRescalerExportRow(&p->scaler_u, 0);
+    WebPRescalerExportRow(&p->scaler_v, 0);
     convert(p->scaler_y.dst, p->scaler_u.dst, p->scaler_v.dst,
             dst, p->scaler_y.dst_width);
     dst += buf->stride;
@@ -416,7 +425,7 @@ static int ExportAlpha(WebPDecParams* const p, int y_pos) {
   while (WebPRescalerHasPendingOutput(&p->scaler_a)) {
     int i;
     assert(p->last_y + y_pos + num_lines_out < p->output->height);
-    WebPRescalerExportRow(&p->scaler_a);
+    WebPRescalerExportRow(&p->scaler_a, 0);
     for (i = 0; i < width; ++i) {
       const uint32_t alpha_value = p->scaler_a.dst[i];
       dst[4 * i] = alpha_value;
@@ -435,7 +444,11 @@ static int ExportAlpha(WebPDecParams* const p, int y_pos) {
 static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
   uint8_t* const base_rgba = buf->rgba + (p->last_y + y_pos) * buf->stride;
+#ifdef WEBP_SWAP_16BIT_CSP
+  uint8_t* alpha_dst = base_rgba;
+#else
   uint8_t* alpha_dst = base_rgba + 1;
+#endif
   int num_lines_out = 0;
   const WEBP_CSP_MODE colorspace = p->output->colorspace;
   const int width = p->scaler_a.dst_width;
@@ -445,7 +458,7 @@ static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
   while (WebPRescalerHasPendingOutput(&p->scaler_a)) {
     int i;
     assert(p->last_y + y_pos + num_lines_out < p->output->height);
-    WebPRescalerExportRow(&p->scaler_a);
+    WebPRescalerExportRow(&p->scaler_a, 0);
     for (i = 0; i < width; ++i) {
       // Fill in the alpha value (converted to 4 bits).
       const uint32_t alpha_value = p->scaler_a.dst[i] >> 4;
@@ -484,7 +497,7 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
   const size_t work_size = 2 * out_width;   // scratch memory for one rescaler
   int32_t* work;  // rescalers work area
   uint8_t* tmp;   // tmp storage for scaled YUV444 samples before RGB conversion
-  size_t tmp_size1, tmp_size2;
+  size_t tmp_size1, tmp_size2, total_size;
 
   tmp_size1 = 3 * work_size;
   tmp_size2 = 3 * out_width;
@@ -492,7 +505,8 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
     tmp_size1 += work_size;
     tmp_size2 += out_width;
   }
-  p->memory = calloc(1, tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp));
+  total_size = tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp);
+  p->memory = WebPSafeCalloc(1ULL, total_size);
   if (p->memory == NULL) {
     return 0;   // memory error
   }
@@ -524,6 +538,7 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
     } else {
       p->emit_alpha_row = ExportAlpha;
     }
+    WebPInitAlphaProcessing();
   }
   return 1;
 }
@@ -544,7 +559,9 @@ static int CustomSetup(VP8Io* io) {
   if (!WebPIoInitFromOptions(p->options, io, is_alpha ? MODE_YUV : MODE_YUVA)) {
     return 0;
   }
-
+  if (is_alpha && WebPIsPremultipliedMode(colorspace)) {
+    WebPInitUpsamplers();
+  }
   if (io->use_scaling) {
     const int ok = is_rgb ? InitRGBRescaler(io, p) : InitYUVRescaler(io, p);
     if (!ok) {
@@ -553,10 +570,10 @@ static int CustomSetup(VP8Io* io) {
   } else {
     if (is_rgb) {
       p->emit = EmitSampledRGB;   // default
-#ifdef FANCY_UPSAMPLING
       if (io->fancy_upsampling) {
+#ifdef FANCY_UPSAMPLING
         const int uv_width = (io->mb_w + 1) >> 1;
-        p->memory = malloc(io->mb_w + 2 * uv_width);
+        p->memory = WebPSafeMalloc(1ULL, (size_t)(io->mb_w + 2 * uv_width));
         if (p->memory == NULL) {
           return 0;   // memory error.
         }
@@ -565,18 +582,22 @@ static int CustomSetup(VP8Io* io) {
         p->tmp_v = p->tmp_u + uv_width;
         p->emit = EmitFancyRGB;
         WebPInitUpsamplers();
-      }
 #endif
+      } else {
+        WebPInitSamplers();
+      }
     } else {
       p->emit = EmitYUV;
     }
     if (is_alpha) {  // need transparency output
-      if (WebPIsPremultipliedMode(colorspace)) WebPInitPremultiply();
       p->emit_alpha =
           (colorspace == MODE_RGBA_4444 || colorspace == MODE_rgbA_4444) ?
               EmitAlphaRGBA4444
           : is_rgb ? EmitAlphaRGB
           : EmitAlphaYUV;
+      if (is_rgb) {
+        WebPInitAlphaProcessing();
+      }
     }
   }
 
@@ -610,7 +631,7 @@ static int CustomPut(const VP8Io* io) {
 
 static void CustomTeardown(const VP8Io* io) {
   WebPDecParams* const p = (WebPDecParams*)io->opaque;
-  free(p->memory);
+  WebPSafeFree(p->memory);
   p->memory = NULL;
 }
 
@@ -625,4 +646,3 @@ void WebPInitCustomIo(WebPDecParams* const params, VP8Io* const io) {
 }
 
 //------------------------------------------------------------------------------
-

src/dec/tree.c

@@ -11,7 +11,8 @@
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
-#include "vp8i.h"
+#include "./vp8i.h"
+#include "../utils/bit_reader_inl.h"
 
 #define USE_GENERIC_TREE
 
@@ -278,10 +279,23 @@ void VP8ResetProba(VP8Proba* const proba) {
   // proba->bands_[][] is initialized later
 }
 
-void VP8ParseIntraMode(VP8BitReader* const br, VP8Decoder* const dec) {
-  uint8_t* const top = dec->intra_t_ + 4 * dec->mb_x_;
+static void ParseIntraMode(VP8BitReader* const br,
+                           VP8Decoder* const dec, int mb_x) {
+  uint8_t* const top = dec->intra_t_ + 4 * mb_x;
   uint8_t* const left = dec->intra_l_;
-  VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+  VP8MBData* const block = dec->mb_data_ + mb_x;
+
+  // Note: we don't save segment map (yet), as we don't expect
+  // to decode more than 1 keyframe.
+  if (dec->segment_hdr_.update_map_) {
+    // Hardcoded tree parsing
+    block->segment_ = !VP8GetBit(br, dec->proba_.segments_[0])
+                    ? VP8GetBit(br, dec->proba_.segments_[1])
+                    : 2 + VP8GetBit(br, dec->proba_.segments_[2]);
+  } else {
+    block->segment_ = 0;  // default for intra
+  }
+  if (dec->use_skip_proba_) block->skip_ = VP8GetBit(br, dec->skip_p_);
 
   block->is_i4x4_ = !VP8GetBit(br, 145);   // decide for B_PRED first
   if (!block->is_i4x4_) {
@@ -332,6 +346,14 @@ void VP8ParseIntraMode(VP8BitReader* const br, VP8Decoder* const dec) {
                  : VP8GetBit(br, 183) ? TM_PRED : H_PRED;
 }
 
+int VP8ParseIntraModeRow(VP8BitReader* const br, VP8Decoder* const dec) {
+  int mb_x;
+  for (mb_x = 0; mb_x < dec->mb_w_; ++mb_x) {
+    ParseIntraMode(br, dec, mb_x);
+  }
+  return !dec->br_.eof_;
+}
+
 //------------------------------------------------------------------------------
 // Paragraph 13
 

src/dec/vp8.c

@@ -17,7 +17,8 @@
 #include "./vp8i.h"
 #include "./vp8li.h"
 #include "./webpi.h"
-#include "../utils/bit_reader.h"
+#include "../utils/bit_reader_inl.h"
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 
@@ -44,10 +45,10 @@ int VP8InitIoInternal(VP8Io* const io, int version) {
 }
 
 VP8Decoder* VP8New(void) {
-  VP8Decoder* const dec = (VP8Decoder*)calloc(1, sizeof(*dec));
+  VP8Decoder* const dec = (VP8Decoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
   if (dec != NULL) {
     SetOk(dec);
-    WebPWorkerInit(&dec->worker_);
+    WebPGetWorkerInterface()->Init(&dec->worker_);
     dec->ready_ = 0;
     dec->num_parts_ = 1;
   }
@@ -68,7 +69,7 @@ const char* VP8StatusMessage(VP8Decoder* const dec) {
 void VP8Delete(VP8Decoder* const dec) {
   if (dec != NULL) {
     VP8Clear(dec);
-    free(dec);
+    WebPSafeFree(dec);
   }
 }
 
@@ -317,7 +318,6 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
 
     VP8ResetProba(&dec->proba_);
     ResetSegmentHeader(&dec->segment_hdr_);
-    dec->segment_ = 0;    // default for intra
   }
 
   // Check if we have all the partition #0 available, and initialize dec->br_
@@ -363,28 +363,6 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
 
   VP8ParseProba(br, dec);
 
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-  // Extensions
-  if (dec->pic_hdr_.colorspace_) {
-    const size_t kTrailerSize = 8;
-    const uint8_t kTrailerMarker = 0x01;
-    const uint8_t* ext_buf = buf - kTrailerSize;
-    size_t size;
-
-    if (frm_hdr->partition_length_ < kTrailerSize ||
-        ext_buf[kTrailerSize - 1] != kTrailerMarker) {
-      return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
-                         "RIFF: Inconsistent extra information.");
-    }
-
-    // Layer
-    size = (ext_buf[0] << 0) | (ext_buf[1] << 8) | (ext_buf[2] << 16);
-    dec->layer_data_size_ = size;
-    dec->layer_data_ = NULL;  // will be set later
-    dec->layer_colorspace_ = ext_buf[3];
-  }
-#endif
-
   // sanitized state
   dec->ready_ = 1;
   return 1;
@@ -479,8 +457,8 @@ static int ParseResiduals(VP8Decoder* const dec,
                           VP8MB* const mb, VP8BitReader* const token_br) {
   VP8BandProbas (* const bands)[NUM_BANDS] = dec->proba_.bands_;
   const VP8BandProbas* ac_proba;
-  const VP8QuantMatrix* const q = &dec->dqm_[dec->segment_];
   VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+  const VP8QuantMatrix* const q = &dec->dqm_[block->segment_];
   int16_t* dst = block->coeffs_;
   VP8MB* const left_mb = dec->mb_info_ - 1;
   uint8_t tnz, lnz;
@@ -570,26 +548,10 @@ static int ParseResiduals(VP8Decoder* const dec,
 // Main loop
 
 int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br) {
-  VP8BitReader* const br = &dec->br_;
   VP8MB* const left = dec->mb_info_ - 1;
   VP8MB* const mb = dec->mb_info_ + dec->mb_x_;
   VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
-  int skip;
-
-  // Note: we don't save segment map (yet), as we don't expect
-  // to decode more than 1 keyframe.
-  if (dec->segment_hdr_.update_map_) {
-    // Hardcoded tree parsing
-    dec->segment_ = !VP8GetBit(br, dec->proba_.segments_[0]) ?
-        VP8GetBit(br, dec->proba_.segments_[1]) :
-        2 + VP8GetBit(br, dec->proba_.segments_[2]);
-  }
-  skip = dec->use_skip_proba_ ? VP8GetBit(br, dec->skip_p_) : 0;
-
-  VP8ParseIntraMode(br, dec);
-  if (br->eof_) {
-    return 0;
-  }
+  int skip = dec->use_skip_proba_ ? block->skip_ : 0;
 
   if (!skip) {
     skip = ParseResiduals(dec, mb, token_br);
@@ -600,11 +562,12 @@ int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br) {
     }
     block->non_zero_y_ = 0;
     block->non_zero_uv_ = 0;
+    block->dither_ = 0;
   }
 
   if (dec->filter_type_ > 0) {  // store filter info
     VP8FInfo* const finfo = dec->f_info_ + dec->mb_x_;
-    *finfo = dec->fstrengths_[dec->segment_][block->is_i4x4_];
+    *finfo = dec->fstrengths_[block->segment_][block->is_i4x4_];
     finfo->f_inner_ |= !skip;
   }
 
@@ -624,6 +587,10 @@ static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
     // Parse bitstream for this row.
     VP8BitReader* const token_br =
         &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
+    if (!VP8ParseIntraModeRow(&dec->br_, dec)) {
+      return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
+                         "Premature end-of-partition0 encountered.");
+    }
     for (; dec->mb_x_ < dec->mb_w_; ++dec->mb_x_) {
       if (!VP8DecodeMB(dec, token_br)) {
         return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
@@ -638,18 +605,9 @@ static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
     }
   }
   if (dec->mt_method_ > 0) {
-    if (!WebPWorkerSync(&dec->worker_)) return 0;
+    if (!WebPGetWorkerInterface()->Sync(&dec->worker_)) return 0;
   }
 
-  // Finish
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-  if (dec->layer_data_size_ > 0) {
-    if (!VP8DecodeLayer(dec)) {
-      return 0;
-    }
-  }
-#endif
-
   return 1;
 }
 
@@ -697,12 +655,10 @@ void VP8Clear(VP8Decoder* const dec) {
   if (dec == NULL) {
     return;
   }
-  if (dec->mt_method_ > 0) {
-    WebPWorkerEnd(&dec->worker_);
-  }
+  WebPGetWorkerInterface()->End(&dec->worker_);
   ALPHDelete(dec->alph_dec_);
   dec->alph_dec_ = NULL;
-  free(dec->mem_);
+  WebPSafeFree(dec->mem_);
   dec->mem_ = NULL;
   dec->mem_size_ = 0;
   memset(&dec->br_, 0, sizeof(dec->br_));

src/dec/vp8i.h

@@ -31,7 +31,7 @@ extern "C" {
 // version numbers
 #define DEC_MAJ_VERSION 0
 #define DEC_MIN_VERSION 4
-#define DEC_REV_VERSION 0
+#define DEC_REV_VERSION 3
 
 // intra prediction modes
 enum { B_DC_PRED = 0,   // 4x4 modes
@@ -195,6 +195,8 @@ typedef struct {
   uint32_t non_zero_y_;
   uint32_t non_zero_uv_;
   uint8_t dither_;      // local dithering strength (deduced from non_zero_*)
+  uint8_t skip_;
+  uint8_t segment_;
 } VP8MBData;
 
 // Persistent information needed by the parallel processing
@@ -265,7 +267,6 @@ struct VP8Decoder {
   uint8_t* intra_t_;      // top intra modes values: 4 * mb_w_
   uint8_t  intra_l_[4];   // left intra modes values
 
-  uint8_t segment_;       // segment of the currently parsed block
   VP8TopSamples* yuv_t_;  // top y/u/v samples
 
   VP8MB* mb_info_;        // contextual macroblock info (mb_w_ + 1)
@@ -295,12 +296,8 @@ struct VP8Decoder {
   const uint8_t* alpha_data_;     // compressed alpha data (if present)
   size_t alpha_data_size_;
   int is_alpha_decoded_;  // true if alpha_data_ is decoded in alpha_plane_
-  uint8_t* alpha_plane_;        // output. Persistent, contains the whole data.
-
-  // extensions
-  int layer_colorspace_;
-  const uint8_t* layer_data_;   // compressed layer data (if present)
-  size_t layer_data_size_;
+  uint8_t* alpha_plane_;  // output. Persistent, contains the whole data.
+  int alpha_dithering_;   // derived from decoding options (0=off, 100=full).
 };
 
 //------------------------------------------------------------------------------
@@ -313,7 +310,8 @@ int VP8SetError(VP8Decoder* const dec,
 // in tree.c
 void VP8ResetProba(VP8Proba* const proba);
 void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec);
-void VP8ParseIntraMode(VP8BitReader* const br,  VP8Decoder* const dec);
+// parses one row of intra mode data in partition 0, returns !eof
+int VP8ParseIntraModeRow(VP8BitReader* const br, VP8Decoder* const dec);
 
 // in quant.c
 void VP8ParseQuant(VP8Decoder* const dec);
@@ -347,9 +345,6 @@ int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br);
 const uint8_t* VP8DecompressAlphaRows(VP8Decoder* const dec,
                                       int row, int num_rows);
 
-// in layer.c
-int VP8DecodeLayer(VP8Decoder* const dec);
-
 //------------------------------------------------------------------------------
 
 #ifdef __cplusplus

src/dec/vp8l.c

@@ -12,13 +12,13 @@
 // Authors: Vikas Arora (vikaas.arora@gmail.com)
 //          Jyrki Alakuijala (jyrki@google.com)
 
-#include <stdio.h>
 #include <stdlib.h>
+
 #include "./alphai.h"
 #include "./vp8li.h"
+#include "../dsp/dsp.h"
 #include "../dsp/lossless.h"
 #include "../dsp/yuv.h"
-#include "../utils/alpha_processing.h"
 #include "../utils/huffman.h"
 #include "../utils/utils.h"
 
@@ -187,9 +187,10 @@ static int ReadHuffmanCodeLengths(
   int max_symbol;
   int prev_code_len = DEFAULT_CODE_LENGTH;
   HuffmanTree tree;
+  int huff_codes[NUM_CODE_LENGTH_CODES] = { 0 };
 
-  if (!HuffmanTreeBuildImplicit(&tree, code_length_code_lengths,
-                                NUM_CODE_LENGTH_CODES)) {
+  if (!VP8LHuffmanTreeBuildImplicit(&tree, code_length_code_lengths,
+                                    huff_codes, NUM_CODE_LENGTH_CODES)) {
     dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
     return 0;
   }
@@ -232,11 +233,15 @@ static int ReadHuffmanCodeLengths(
   ok = 1;
 
  End:
-  HuffmanTreeRelease(&tree);
+  VP8LHuffmanTreeFree(&tree);
+  if (!ok) dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
   return ok;
 }
 
+// 'code_lengths' is pre-allocated temporary buffer, used for creating Huffman
+// tree.
 static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
+                           int* const code_lengths, int* const huff_codes,
                            HuffmanTree* const tree) {
   int ok = 0;
   VP8LBitReader* const br = &dec->br_;
@@ -245,7 +250,6 @@ static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
   if (simple_code) {  // Read symbols, codes & code lengths directly.
     int symbols[2];
     int codes[2];
-    int code_lengths[2];
     const int num_symbols = VP8LReadBits(br, 1) + 1;
     const int first_symbol_len_code = VP8LReadBits(br, 1);
     // The first code is either 1 bit or 8 bit code.
@@ -258,10 +262,9 @@ static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
       codes[1] = 1;
       code_lengths[1] = num_symbols - 1;
     }
-    ok = HuffmanTreeBuildExplicit(tree, code_lengths, codes, symbols,
-                                  alphabet_size, num_symbols);
+    ok = VP8LHuffmanTreeBuildExplicit(tree, code_lengths, codes, symbols,
+                                      alphabet_size, num_symbols);
   } else {  // Decode Huffman-coded code lengths.
-    int* code_lengths = NULL;
     int i;
     int code_length_code_lengths[NUM_CODE_LENGTH_CODES] = { 0 };
     const int num_codes = VP8LReadBits(br, 4) + 4;
@@ -270,22 +273,15 @@ static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
       return 0;
     }
 
-    code_lengths =
-        (int*)WebPSafeCalloc((uint64_t)alphabet_size, sizeof(*code_lengths));
-    if (code_lengths == NULL) {
-      dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
-      return 0;
-    }
+    memset(code_lengths, 0, alphabet_size * sizeof(*code_lengths));
 
     for (i = 0; i < num_codes; ++i) {
       code_length_code_lengths[kCodeLengthCodeOrder[i]] = VP8LReadBits(br, 3);
     }
     ok = ReadHuffmanCodeLengths(dec, code_length_code_lengths, alphabet_size,
                                 code_lengths);
-    if (ok) {
-      ok = HuffmanTreeBuildImplicit(tree, code_lengths, alphabet_size);
-    }
-    free(code_lengths);
+    ok = ok && VP8LHuffmanTreeBuildImplicit(tree, code_lengths, huff_codes,
+                                            alphabet_size);
   }
   ok = ok && !br->error_;
   if (!ok) {
@@ -295,19 +291,6 @@ static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
   return 1;
 }
 
-static void DeleteHtreeGroups(HTreeGroup* htree_groups, int num_htree_groups) {
-  if (htree_groups != NULL) {
-    int i, j;
-    for (i = 0; i < num_htree_groups; ++i) {
-      HuffmanTree* const htrees = htree_groups[i].htrees_;
-      for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
-        HuffmanTreeRelease(&htrees[j]);
-      }
-    }
-    free(htree_groups);
-  }
-}
-
 static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
                             int color_cache_bits, int allow_recursion) {
   int i, j;
@@ -316,6 +299,9 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
   uint32_t* huffman_image = NULL;
   HTreeGroup* htree_groups = NULL;
   int num_htree_groups = 1;
+  int max_alphabet_size = 0;
+  int* code_lengths = NULL;
+  int* huff_codes = NULL;
 
   if (allow_recursion && VP8LReadBits(br, 1)) {
     // use meta Huffman codes.
@@ -341,11 +327,24 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
 
   if (br->error_) goto Error;
 
-  assert(num_htree_groups <= 0x10000);
-  htree_groups =
-      (HTreeGroup*)WebPSafeCalloc((uint64_t)num_htree_groups,
-                                  sizeof(*htree_groups));
-  if (htree_groups == NULL) {
+  // Find maximum alphabet size for the htree group.
+  for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
+    int alphabet_size = kAlphabetSize[j];
+    if (j == 0 && color_cache_bits > 0) {
+      alphabet_size += 1 << color_cache_bits;
+    }
+    if (max_alphabet_size < alphabet_size) {
+      max_alphabet_size = alphabet_size;
+    }
+  }
+
+  htree_groups = VP8LHtreeGroupsNew(num_htree_groups);
+  code_lengths =
+      (int*)WebPSafeCalloc((uint64_t)max_alphabet_size, sizeof(*code_lengths));
+  huff_codes =
+      (int*)WebPSafeMalloc((uint64_t)max_alphabet_size, sizeof(*huff_codes));
+
+  if (htree_groups == NULL || code_lengths == NULL || huff_codes == NULL) {
     dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
     goto Error;
   }
@@ -354,12 +353,18 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
     HuffmanTree* const htrees = htree_groups[i].htrees_;
     for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
       int alphabet_size = kAlphabetSize[j];
+      HuffmanTree* const htree = htrees + j;
       if (j == 0 && color_cache_bits > 0) {
         alphabet_size += 1 << color_cache_bits;
       }
-      if (!ReadHuffmanCode(alphabet_size, dec, htrees + j)) goto Error;
+      if (!ReadHuffmanCode(alphabet_size, dec, code_lengths, huff_codes,
+                           htree)) {
+        goto Error;
+      }
     }
   }
+  WebPSafeFree(huff_codes);
+  WebPSafeFree(code_lengths);
 
   // All OK. Finalize pointers and return.
   hdr->huffman_image_ = huffman_image;
@@ -368,8 +373,10 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
   return 1;
 
  Error:
-  free(huffman_image);
-  DeleteHtreeGroups(htree_groups, num_htree_groups);
+  WebPSafeFree(huff_codes);
+  WebPSafeFree(code_lengths);
+  WebPSafeFree(huffman_image);
+  VP8LHtreeGroupsFree(htree_groups, num_htree_groups);
   return 0;
 }
 
@@ -420,7 +427,7 @@ static int Export(WebPRescaler* const rescaler, WEBP_CSP_MODE colorspace,
   int num_lines_out = 0;
   while (WebPRescalerHasPendingOutput(rescaler)) {
     uint8_t* const dst = rgba + num_lines_out * rgba_stride;
-    WebPRescalerExportRow(rescaler);
+    WebPRescalerExportRow(rescaler, 0);
     WebPMultARGBRow(src, dst_width, 1);
     VP8LConvertFromBGRA(src, dst_width, colorspace, dst);
     ++num_lines_out;
@@ -468,6 +475,7 @@ static int EmitRows(WEBP_CSP_MODE colorspace,
 //------------------------------------------------------------------------------
 // Export to YUVA
 
+// TODO(skal): should be in yuv.c
 static void ConvertToYUVA(const uint32_t* const src, int width, int y_pos,
                           const WebPDecBuffer* const output) {
   const WebPYUVABuffer* const buf = &output->u.YUVA;
@@ -537,7 +545,7 @@ static int ExportYUVA(const VP8LDecoder* const dec, int y_pos) {
   const int dst_width = rescaler->dst_width;
   int num_lines_out = 0;
   while (WebPRescalerHasPendingOutput(rescaler)) {
-    WebPRescalerExportRow(rescaler);
+    WebPRescalerExportRow(rescaler, 0);
     WebPMultARGBRow(src, dst_width, 1);
     ConvertToYUVA(src, dst_width, y_pos, dec->output_);
     ++y_pos;
@@ -740,6 +748,7 @@ static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
   const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES;
   const int mask = hdr->huffman_mask_;
   assert(htree_group != NULL);
+  assert(pos < end);
   assert(last_row <= height);
   assert(Is8bOptimizable(hdr));
 
@@ -793,6 +802,7 @@ static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
       ok = 0;
       goto End;
     }
+    assert(br->eos_ == VP8LIsEndOfStream(br));
     ok = !br->error_;
     if (!ok) goto End;
   }
@@ -830,6 +840,7 @@ static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
       (hdr->color_cache_size_ > 0) ? &hdr->color_cache_ : NULL;
   const int mask = hdr->huffman_mask_;
   assert(htree_group != NULL);
+  assert(src < src_end);
   assert(src_last <= src_end);
 
   while (!br->eos_ && src < src_last) {
@@ -849,7 +860,7 @@ static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
       VP8LFillBitWindow(br);
       blue = ReadSymbol(&htree_group->htrees_[BLUE], br);
       alpha = ReadSymbol(&htree_group->htrees_[ALPHA], br);
-      *src = (alpha << 24) | (red << 16) | (green << 8) | blue;
+      *src = ((uint32_t)alpha << 24) | (red << 16) | (green << 8) | blue;
     AdvanceByOne:
       ++src;
       ++col;
@@ -889,7 +900,7 @@ static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
           process_func(dec, row);
         }
       }
-      if (src < src_last) {
+      if (src < src_end) {
         if (col & mask) htree_group = GetHtreeGroupForPos(hdr, col, row);
         if (color_cache != NULL) {
           while (last_cached < src) {
@@ -909,6 +920,7 @@ static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
       ok = 0;
       goto End;
     }
+    assert(br->eos_ == VP8LIsEndOfStream(br));
     ok = !br->error_;
     if (!ok) goto End;
   }
@@ -931,7 +943,7 @@ static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
 // VP8LTransform
 
 static void ClearTransform(VP8LTransform* const transform) {
-  free(transform->data_);
+  WebPSafeFree(transform->data_);
   transform->data_ = NULL;
 }
 
@@ -955,7 +967,7 @@ static int ExpandColorMap(int num_colors, VP8LTransform* const transform) {
     }
     for (; i < 4 * final_num_colors; ++i)
       new_data[i] = 0;  // black tail.
-    free(transform->data_);
+    WebPSafeFree(transform->data_);
     transform->data_ = new_color_map;
   }
   return 1;
@@ -1025,8 +1037,8 @@ static void InitMetadata(VP8LMetadata* const hdr) {
 static void ClearMetadata(VP8LMetadata* const hdr) {
   assert(hdr);
 
-  free(hdr->huffman_image_);
-  DeleteHtreeGroups(hdr->htree_groups_, hdr->num_htree_groups_);
+  WebPSafeFree(hdr->huffman_image_);
+  VP8LHtreeGroupsFree(hdr->htree_groups_, hdr->num_htree_groups_);
   VP8LColorCacheClear(&hdr->color_cache_);
   InitMetadata(hdr);
 }
@@ -1035,7 +1047,7 @@ static void ClearMetadata(VP8LMetadata* const hdr) {
 // VP8LDecoder
 
 VP8LDecoder* VP8LNew(void) {
-  VP8LDecoder* const dec = (VP8LDecoder*)calloc(1, sizeof(*dec));
+  VP8LDecoder* const dec = (VP8LDecoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
   if (dec == NULL) return NULL;
   dec->status_ = VP8_STATUS_OK;
   dec->action_ = READ_DIM;
@@ -1051,7 +1063,7 @@ void VP8LClear(VP8LDecoder* const dec) {
   if (dec == NULL) return;
   ClearMetadata(&dec->hdr_);
 
-  free(dec->pixels_);
+  WebPSafeFree(dec->pixels_);
   dec->pixels_ = NULL;
   for (i = 0; i < dec->next_transform_; ++i) {
     ClearTransform(&dec->transforms_[i]);
@@ -1059,7 +1071,7 @@ void VP8LClear(VP8LDecoder* const dec) {
   dec->next_transform_ = 0;
   dec->transforms_seen_ = 0;
 
-  free(dec->rescaler_memory);
+  WebPSafeFree(dec->rescaler_memory);
   dec->rescaler_memory = NULL;
 
   dec->output_ = NULL;   // leave no trace behind
@@ -1068,7 +1080,7 @@ void VP8LClear(VP8LDecoder* const dec) {
 void VP8LDelete(VP8LDecoder* const dec) {
   if (dec != NULL) {
     VP8LClear(dec);
-    free(dec);
+    WebPSafeFree(dec);
   }
 }
 
@@ -1155,7 +1167,7 @@ static int DecodeImageStream(int xsize, int ysize,
  End:
 
   if (!ok) {
-    free(data);
+    WebPSafeFree(data);
     ClearMetadata(hdr);
     // If not enough data (br.eos_) resulted in BIT_STREAM_ERROR, update the
     // status appropriately.
@@ -1294,6 +1306,10 @@ int VP8LDecodeAlphaImageStream(ALPHDecoder* const alph_dec, int last_row) {
   assert(dec->action_ == READ_DATA);
   assert(last_row <= dec->height_);
 
+  if (dec->last_pixel_ == dec->width_ * dec->height_) {
+    return 1;  // done
+  }
+
   // Decode (with special row processing).
   return alph_dec->use_8b_decode ?
       DecodeAlphaData(dec, (uint8_t*)dec->pixels_, dec->width_, dec->height_,
@@ -1341,6 +1357,10 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
   // Sanity checks.
   if (dec == NULL) return 0;
 
+  dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+  assert(dec->hdr_.htree_groups_ != NULL);
+  assert(dec->hdr_.num_htree_groups_ > 0);
+
   io = dec->io_;
   assert(io != NULL);
   params = (WebPDecParams*)io->opaque;
@@ -1358,6 +1378,11 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
 
   if (io->use_scaling && !AllocateAndInitRescaler(dec, io)) goto Err;
 
+  if (io->use_scaling || WebPIsPremultipliedMode(dec->output_->colorspace)) {
+    // need the alpha-multiply functions for premultiplied output or rescaling
+    WebPInitAlphaProcessing();
+  }
+
   // Decode.
   dec->action_ = READ_DATA;
   if (!DecodeImageData(dec, dec->pixels_, dec->width_, dec->height_,
@@ -1377,4 +1402,3 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
 }
 
 //------------------------------------------------------------------------------
-

src/dec/vp8li.h

@@ -20,7 +20,6 @@
 #include "../utils/bit_reader.h"
 #include "../utils/color_cache.h"
 #include "../utils/huffman.h"
-#include "../webp/format_constants.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -41,10 +40,6 @@ struct VP8LTransform {
   uint32_t              *data_;   // transform data.
 };
 
-typedef struct {
-  HuffmanTree htrees_[HUFFMAN_CODES_PER_META_CODE];
-} HTreeGroup;
-
 typedef struct {
   int             color_cache_size_;
   VP8LColorCache  color_cache_;

src/dec/webp.c

@@ -52,13 +52,14 @@ static WEBP_INLINE uint32_t get_le32(const uint8_t* const data) {
 }
 
 // Validates the RIFF container (if detected) and skips over it.
-// If a RIFF container is detected,
-// Returns VP8_STATUS_BITSTREAM_ERROR for invalid header, and
-//         VP8_STATUS_OK otherwise.
+// If a RIFF container is detected, returns:
+//     VP8_STATUS_BITSTREAM_ERROR for invalid header,
+//     VP8_STATUS_NOT_ENOUGH_DATA for truncated data if have_all_data is true,
+// and VP8_STATUS_OK otherwise.
 // In case there are not enough bytes (partial RIFF container), return 0 for
 // *riff_size. Else return the RIFF size extracted from the header.
 static VP8StatusCode ParseRIFF(const uint8_t** const data,
-                               size_t* const data_size,
+                               size_t* const data_size, int have_all_data,
                                size_t* const riff_size) {
   assert(data != NULL);
   assert(data_size != NULL);
@@ -77,6 +78,9 @@ static VP8StatusCode ParseRIFF(const uint8_t** const data,
       if (size > MAX_CHUNK_PAYLOAD) {
         return VP8_STATUS_BITSTREAM_ERROR;
       }
+      if (have_all_data && (size > *data_size - CHUNK_HEADER_SIZE)) {
+        return VP8_STATUS_NOT_ENOUGH_DATA;  // Truncated bitstream.
+      }
       // We have a RIFF container. Skip it.
       *riff_size = size;
       *data += RIFF_HEADER_SIZE;
@@ -223,9 +227,8 @@ static VP8StatusCode ParseOptionalChunks(const uint8_t** const data,
 // extracted from the VP8/VP8L chunk header.
 // The flag '*is_lossless' is set to 1 in case of VP8L chunk / raw VP8L data.
 static VP8StatusCode ParseVP8Header(const uint8_t** const data_ptr,
-                                    size_t* const data_size,
-                                    size_t riff_size,
-                                    size_t* const chunk_size,
+                                    size_t* const data_size, int have_all_data,
+                                    size_t riff_size, size_t* const chunk_size,
                                     int* const is_lossless) {
   const uint8_t* const data = *data_ptr;
   const int is_vp8 = !memcmp(data, "VP8 ", TAG_SIZE);
@@ -248,6 +251,9 @@ static VP8StatusCode ParseVP8Header(const uint8_t** const data_ptr,
     if ((riff_size >= minimal_size) && (size > riff_size - minimal_size)) {
       return VP8_STATUS_BITSTREAM_ERROR;  // Inconsistent size information.
     }
+    if (have_all_data && (size > *data_size - CHUNK_HEADER_SIZE)) {
+      return VP8_STATUS_NOT_ENOUGH_DATA;  // Truncated bitstream.
+    }
     // Skip over CHUNK_HEADER_SIZE bytes from VP8/VP8L Header.
     *chunk_size = size;
     *data_ptr += CHUNK_HEADER_SIZE;
@@ -291,6 +297,7 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
   int found_vp8x = 0;
   int animation_present = 0;
   int fragments_present = 0;
+  const int have_all_data = (headers != NULL) ? headers->have_all_data : 0;
 
   VP8StatusCode status;
   WebPHeaderStructure hdrs;
@@ -303,7 +310,7 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
   hdrs.data_size = data_size;
 
   // Skip over RIFF header.
-  status = ParseRIFF(&data, &data_size, &hdrs.riff_size);
+  status = ParseRIFF(&data, &data_size, have_all_data, &hdrs.riff_size);
   if (status != VP8_STATUS_OK) {
     return status;   // Wrong RIFF header / insufficient data.
   }
@@ -353,7 +360,7 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
   }
 
   // Skip over VP8/VP8L header.
-  status = ParseVP8Header(&data, &data_size, hdrs.riff_size,
+  status = ParseVP8Header(&data, &data_size, have_all_data, hdrs.riff_size,
                           &hdrs.compressed_size, &hdrs.is_lossless);
   if (status != VP8_STATUS_OK) {
     goto ReturnWidthHeight;  // Wrong VP8/VP8L chunk-header / insufficient data.
@@ -452,6 +459,7 @@ static VP8StatusCode DecodeInto(const uint8_t* const data, size_t data_size,
 
   headers.data = data;
   headers.data_size = data_size;
+  headers.have_all_data = 1;
   status = WebPParseHeaders(&headers);   // Process Pre-VP8 chunks.
   if (status != VP8_STATUS_OK) {
     return status;
@@ -512,6 +520,12 @@ static VP8StatusCode DecodeInto(const uint8_t* const data, size_t data_size,
   if (status != VP8_STATUS_OK) {
     WebPFreeDecBuffer(params->output);
   }
+
+#if WEBP_DECODER_ABI_VERSION > 0x0203
+  if (params->options != NULL && params->options->flip) {
+    status = WebPFlipBuffer(params->output);
+  }
+#endif
   return status;
 }
 
@@ -776,9 +790,9 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
     h = options->crop_height;
     x = options->crop_left;
     y = options->crop_top;
-    if (!WebPIsRGBMode(src_colorspace)) {   // only snap for YUV420 or YUV422
+    if (!WebPIsRGBMode(src_colorspace)) {   // only snap for YUV420
       x &= ~1;
-      y &= ~1;    // TODO(later): only for YUV420, not YUV422.
+      y &= ~1;
     }
     if (x < 0 || y < 0 || w <= 0 || h <= 0 || x + w > W || y + h > H) {
       return 0;  // out of frame boundary error

src/dec/webpi.h

@@ -54,6 +54,7 @@ void WebPResetDecParams(WebPDecParams* const params);
 typedef struct {
   const uint8_t* data;         // input buffer
   size_t data_size;            // input buffer size
+  int have_all_data;           // true if all data is known to be available
   size_t offset;               // offset to main data chunk (VP8 or VP8L)
   const uint8_t* alpha_data;   // points to alpha chunk (if present)
   size_t alpha_data_size;      // alpha chunk size
@@ -93,10 +94,15 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
 // dimension / etc.). If *options is not NULL, also verify that the options'
 // parameters are valid and apply them to the width/height dimensions of the
 // output buffer. This takes cropping / scaling / rotation into account.
+// Also incorporates the options->flip flag to flip the buffer parameters if
+// needed.
 VP8StatusCode WebPAllocateDecBuffer(int width, int height,
                                     const WebPDecoderOptions* const options,
                                     WebPDecBuffer* const buffer);
 
+// Flip buffer vertically by negating the various strides.
+VP8StatusCode WebPFlipBuffer(WebPDecBuffer* const buffer);
+
 // Copy 'src' into 'dst' buffer, making sure 'dst' is not marked as owner of the
 // memory (still held by 'src').
 void WebPCopyDecBuffer(const WebPDecBuffer* const src,
@@ -105,8 +111,6 @@ void WebPCopyDecBuffer(const WebPDecBuffer* const src,
 // Copy and transfer ownership from src to dst (beware of parameter order!)
 void WebPGrabDecBuffer(WebPDecBuffer* const src, WebPDecBuffer* const dst);
 
-
-
 //------------------------------------------------------------------------------
 
 #ifdef __cplusplus

src/demux/Makefile.am

@@ -1,4 +1,3 @@
-AM_CPPFLAGS = -I$(top_srcdir)/src
 lib_LTLIBRARIES = libwebpdemux.la
 
 libwebpdemux_la_SOURCES =
@@ -10,6 +9,6 @@ libwebpdemuxinclude_HEADERS += ../webp/mux_types.h
 libwebpdemuxinclude_HEADERS += ../webp/types.h
 
 libwebpdemux_la_LIBADD = ../libwebp.la
-libwebpdemux_la_LDFLAGS = -no-undefined -version-info 1:0:0
+libwebpdemux_la_LDFLAGS = -no-undefined -version-info 1:2:0
 libwebpdemuxincludedir = $(includedir)/webp
 pkgconfig_DATA = libwebpdemux.pc

src/demux/demux.c

@@ -11,7 +11,7 @@
 //
 
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "../webp/config.h"
 #endif
 
 #include <assert.h>
@@ -25,7 +25,7 @@
 
 #define DMUX_MAJ_VERSION 0
 #define DMUX_MIN_VERSION 2
-#define DMUX_REV_VERSION 0
+#define DMUX_REV_VERSION 2
 
 typedef struct {
   size_t start_;        // start location of the data
@@ -289,7 +289,7 @@ static ParseStatus NewFrame(const MemBuffer* const mem,
   if (actual_size < min_size) return PARSE_ERROR;
   if (MemDataSize(mem) < min_size)  return PARSE_NEED_MORE_DATA;
 
-  *frame = (Frame*)calloc(1, sizeof(**frame));
+  *frame = (Frame*)WebPSafeCalloc(1ULL, sizeof(**frame));
   return (*frame == NULL) ? PARSE_ERROR : PARSE_OK;
 }
 
@@ -317,7 +317,7 @@ static ParseStatus ParseAnimationFrame(
       (bits & 1) ? WEBP_MUX_DISPOSE_BACKGROUND : WEBP_MUX_DISPOSE_NONE;
   frame->blend_method_ = (bits & 2) ? WEBP_MUX_NO_BLEND : WEBP_MUX_BLEND;
   if (frame->width_ * (uint64_t)frame->height_ >= MAX_IMAGE_AREA) {
-    free(frame);
+    WebPSafeFree(frame);
     return PARSE_ERROR;
   }
 
@@ -333,7 +333,7 @@ static ParseStatus ParseAnimationFrame(
     }
   }
 
-  if (!added_frame) free(frame);
+  if (!added_frame) WebPSafeFree(frame);
   return status;
 }
 
@@ -368,7 +368,7 @@ static ParseStatus ParseFragment(WebPDemuxer* const dmux,
     }
   }
 
-  if (!added_fragment) free(frame);
+  if (!added_fragment) WebPSafeFree(frame);
   return status;
 }
 #endif  // WEBP_EXPERIMENTAL_FEATURES
@@ -379,7 +379,7 @@ static ParseStatus ParseFragment(WebPDemuxer* const dmux,
 // Returns true on success, false otherwise.
 static int StoreChunk(WebPDemuxer* const dmux,
                       size_t start_offset, uint32_t size) {
-  Chunk* const chunk = (Chunk*)calloc(1, sizeof(*chunk));
+  Chunk* const chunk = (Chunk*)WebPSafeCalloc(1ULL, sizeof(*chunk));
   if (chunk == NULL) return 0;
 
   chunk->data_.offset_ = start_offset;
@@ -427,7 +427,7 @@ static ParseStatus ParseSingleImage(WebPDemuxer* const dmux) {
   if (SizeIsInvalid(mem, min_size)) return PARSE_ERROR;
   if (MemDataSize(mem) < min_size) return PARSE_NEED_MORE_DATA;
 
-  frame = (Frame*)calloc(1, sizeof(*frame));
+  frame = (Frame*)WebPSafeCalloc(1ULL, sizeof(*frame));
   if (frame == NULL) return PARSE_ERROR;
 
   // For the single image case we allow parsing of a partial frame, but we need
@@ -458,7 +458,7 @@ static ParseStatus ParseSingleImage(WebPDemuxer* const dmux) {
     }
   }
 
-  if (!image_added) free(frame);
+  if (!image_added) WebPSafeFree(frame);
   return status;
 }
 
@@ -729,7 +729,7 @@ WebPDemuxer* WebPDemuxInternal(const WebPData* data, int allow_partial,
   partial = (mem.buf_size_ < mem.riff_end_);
   if (!allow_partial && partial) return NULL;
 
-  dmux = (WebPDemuxer*)calloc(1, sizeof(*dmux));
+  dmux = (WebPDemuxer*)WebPSafeCalloc(1ULL, sizeof(*dmux));
   if (dmux == NULL) return NULL;
   InitDemux(dmux, &mem);
 
@@ -761,14 +761,14 @@ void WebPDemuxDelete(WebPDemuxer* dmux) {
   for (f = dmux->frames_; f != NULL;) {
     Frame* const cur_frame = f;
     f = f->next_;
-    free(cur_frame);
+    WebPSafeFree(cur_frame);
   }
   for (c = dmux->chunks_; c != NULL;) {
     Chunk* const cur_chunk = c;
     c = c->next_;
-    free(cur_chunk);
+    WebPSafeFree(cur_chunk);
   }
-  free(dmux);
+  WebPSafeFree(dmux);
 }
 
 // -----------------------------------------------------------------------------

src/dsp/Makefile.am

@@ -1,5 +1,5 @@
-AM_CPPFLAGS = -I$(top_srcdir)/src
-noinst_LTLIBRARIES = libwebpdsp.la
+noinst_LTLIBRARIES = libwebpdsp.la libwebpdsp_avx2.la
+noinst_LTLIBRARIES += libwebpdsp_sse2.la libwebpdspdecode_sse2.la
 
 if BUILD_LIBWEBPDECODER
   noinst_LTLIBRARIES += libwebpdspdecode.la
@@ -9,23 +9,49 @@ common_HEADERS = ../webp/types.h
 commondir = $(includedir)/webp
 
 COMMON_SOURCES =
+COMMON_SOURCES += alpha_processing.c
 COMMON_SOURCES += cpu.c
 COMMON_SOURCES += dec.c
+COMMON_SOURCES += dec_clip_tables.c
+COMMON_SOURCES += dec_mips32.c
 COMMON_SOURCES += dec_neon.c
-COMMON_SOURCES += dec_sse2.c
 COMMON_SOURCES += dsp.h
 COMMON_SOURCES += lossless.c
 COMMON_SOURCES += lossless.h
+COMMON_SOURCES += lossless_mips32.c
+COMMON_SOURCES += lossless_neon.c
+COMMON_SOURCES += neon.h
 COMMON_SOURCES += upsampling.c
 COMMON_SOURCES += upsampling_neon.c
-COMMON_SOURCES += upsampling_sse2.c
 COMMON_SOURCES += yuv.c
 COMMON_SOURCES += yuv.h
+COMMON_SOURCES += yuv_mips32.c
 
 ENC_SOURCES =
 ENC_SOURCES += enc.c
+ENC_SOURCES += enc_mips32.c
 ENC_SOURCES += enc_neon.c
-ENC_SOURCES += enc_sse2.c
+
+libwebpdsp_avx2_la_SOURCES =
+libwebpdsp_avx2_la_SOURCES += enc_avx2.c
+libwebpdsp_avx2_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
+libwebpdsp_avx2_la_CFLAGS = $(AM_CFLAGS) $(AVX2_FLAGS)
+
+libwebpdspdecode_sse2_la_SOURCES =
+libwebpdspdecode_sse2_la_SOURCES += alpha_processing_sse2.c
+libwebpdspdecode_sse2_la_SOURCES += dec_sse2.c
+libwebpdspdecode_sse2_la_SOURCES += lossless_sse2.c
+libwebpdspdecode_sse2_la_SOURCES += upsampling_sse2.c
+libwebpdspdecode_sse2_la_SOURCES += yuv_sse2.c
+libwebpdspdecode_sse2_la_SOURCES += yuv_tables_sse2.h
+libwebpdspdecode_sse2_la_CPPFLAGS = $(libwebpdsp_sse2_la_CPPFLAGS)
+libwebpdspdecode_sse2_la_CFLAGS = $(libwebpdsp_sse2_la_CFLAGS)
+
+libwebpdsp_sse2_la_SOURCES =
+libwebpdsp_sse2_la_SOURCES += enc_sse2.c
+libwebpdsp_sse2_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
+libwebpdsp_sse2_la_CFLAGS = $(AM_CFLAGS) $(SSE2_FLAGS)
+libwebpdsp_sse2_la_LIBADD = libwebpdspdecode_sse2.la
 
 libwebpdsp_la_SOURCES = $(COMMON_SOURCES) $(ENC_SOURCES)
 
@@ -33,12 +59,14 @@ noinst_HEADERS =
 noinst_HEADERS += ../dec/decode_vp8.h
 noinst_HEADERS += ../webp/decode.h
 
-libwebpdsp_la_LDFLAGS = -lm
 libwebpdsp_la_CPPFLAGS = $(USE_EXPERIMENTAL_CODE) $(USE_SWAP_16BIT_CSP)
+libwebpdsp_la_LDFLAGS = -lm
+libwebpdsp_la_LIBADD = libwebpdsp_avx2.la libwebpdsp_sse2.la
 
 if BUILD_LIBWEBPDECODER
   libwebpdspdecode_la_SOURCES = $(COMMON_SOURCES)
 
-  libwebpdspdecode_la_LDFLAGS = $(libwebpdsp_la_LDFLAGS)
   libwebpdspdecode_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
+  libwebpdspdecode_la_LDFLAGS = $(libwebpdsp_la_LDFLAGS)
+  libwebpdspdecode_la_LIBADD = libwebpdspdecode_sse2.la
 endif

src/dsp/alpha_processing.c

@@ -12,7 +12,7 @@
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include <assert.h>
-#include "./alpha_processing.h"
+#include "./dsp.h"
 
 // Tables can be faster on some platform but incur some extra binary size (~2k).
 // #define USE_TABLES_FOR_ALPHA_MULT
@@ -134,7 +134,7 @@ static WEBP_INLINE uint32_t GetScale(uint32_t a, int inverse) {
 
 #endif    // USE_TABLES_FOR_ALPHA_MULT
 
-void WebPMultARGBRow(uint32_t* const ptr, int width, int inverse) {
+static void MultARGBRow(uint32_t* const ptr, int width, int inverse) {
   int x;
   for (x = 0; x < width; ++x) {
     const uint32_t argb = ptr[x];
@@ -154,17 +154,8 @@ void WebPMultARGBRow(uint32_t* const ptr, int width, int inverse) {
   }
 }
 
-void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows,
-                      int inverse) {
-  int n;
-  for (n = 0; n < num_rows; ++n) {
-    WebPMultARGBRow((uint32_t*)ptr, width, inverse);
-    ptr += stride;
-  }
-}
-
-void WebPMultRow(uint8_t* const ptr, const uint8_t* const alpha,
-                 int width, int inverse) {
+static void MultRow(uint8_t* const ptr, const uint8_t* const alpha,
+                    int width, int inverse) {
   int x;
   for (x = 0; x < width; ++x) {
     const uint32_t a = alpha[x];
@@ -179,6 +170,26 @@ void WebPMultRow(uint8_t* const ptr, const uint8_t* const alpha,
   }
 }
 
+#undef KINV_255
+#undef HALF
+#undef MFIX
+
+void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse);
+void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha,
+                    int width, int inverse);
+
+//------------------------------------------------------------------------------
+// Generic per-plane calls
+
+void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows,
+                      int inverse) {
+  int n;
+  for (n = 0; n < num_rows; ++n) {
+    WebPMultARGBRow((uint32_t*)ptr, width, inverse);
+    ptr += stride;
+  }
+}
+
 void WebPMultRows(uint8_t* ptr, int stride,
                   const uint8_t* alpha, int alpha_stride,
                   int width, int num_rows, int inverse) {
@@ -190,7 +201,135 @@ void WebPMultRows(uint8_t* ptr, int stride,
   }
 }
 
-#undef KINV_255
-#undef HALF
-#undef MFIX
+//------------------------------------------------------------------------------
+// Premultiplied modes
 
+// non dithered-modes
+
+// (x * a * 32897) >> 23 is bit-wise equivalent to (int)(x * a / 255.)
+// for all 8bit x or a. For bit-wise equivalence to (int)(x * a / 255. + .5),
+// one can use instead: (x * a * 65793 + (1 << 23)) >> 24
+#if 1     // (int)(x * a / 255.)
+#define MULTIPLIER(a)   ((a) * 32897U)
+#define PREMULTIPLY(x, m) (((x) * (m)) >> 23)
+#else     // (int)(x * a / 255. + .5)
+#define MULTIPLIER(a) ((a) * 65793U)
+#define PREMULTIPLY(x, m) (((x) * (m) + (1U << 23)) >> 24)
+#endif
+
+static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_first,
+                               int w, int h, int stride) {
+  while (h-- > 0) {
+    uint8_t* const rgb = rgba + (alpha_first ? 1 : 0);
+    const uint8_t* const alpha = rgba + (alpha_first ? 0 : 3);
+    int i;
+    for (i = 0; i < w; ++i) {
+      const uint32_t a = alpha[4 * i];
+      if (a != 0xff) {
+        const uint32_t mult = MULTIPLIER(a);
+        rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult);
+        rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult);
+        rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult);
+      }
+    }
+    rgba += stride;
+  }
+}
+#undef MULTIPLIER
+#undef PREMULTIPLY
+
+// rgbA4444
+
+#define MULTIPLIER(a)  ((a) * 0x1111)    // 0x1111 ~= (1 << 16) / 15
+
+static WEBP_INLINE uint8_t dither_hi(uint8_t x) {
+  return (x & 0xf0) | (x >> 4);
+}
+
+static WEBP_INLINE uint8_t dither_lo(uint8_t x) {
+  return (x & 0x0f) | (x << 4);
+}
+
+static WEBP_INLINE uint8_t multiply(uint8_t x, uint32_t m) {
+  return (x * m) >> 16;
+}
+
+static WEBP_INLINE void ApplyAlphaMultiply4444(uint8_t* rgba4444,
+                                               int w, int h, int stride,
+                                               int rg_byte_pos /* 0 or 1 */) {
+  while (h-- > 0) {
+    int i;
+    for (i = 0; i < w; ++i) {
+      const uint32_t rg = rgba4444[2 * i + rg_byte_pos];
+      const uint32_t ba = rgba4444[2 * i + (rg_byte_pos ^ 1)];
+      const uint8_t a = ba & 0x0f;
+      const uint32_t mult = MULTIPLIER(a);
+      const uint8_t r = multiply(dither_hi(rg), mult);
+      const uint8_t g = multiply(dither_lo(rg), mult);
+      const uint8_t b = multiply(dither_hi(ba), mult);
+      rgba4444[2 * i + rg_byte_pos] = (r & 0xf0) | ((g >> 4) & 0x0f);
+      rgba4444[2 * i + (rg_byte_pos ^ 1)] = (b & 0xf0) | a;
+    }
+    rgba4444 += stride;
+  }
+}
+#undef MULTIPLIER
+
+static void ApplyAlphaMultiply_16b(uint8_t* rgba4444,
+                                   int w, int h, int stride) {
+#ifdef WEBP_SWAP_16BIT_CSP
+  ApplyAlphaMultiply4444(rgba4444, w, h, stride, 1);
+#else
+  ApplyAlphaMultiply4444(rgba4444, w, h, stride, 0);
+#endif
+}
+
+static int ExtractAlpha(const uint8_t* argb, int argb_stride,
+                        int width, int height,
+                        uint8_t* alpha, int alpha_stride) {
+  uint8_t alpha_mask = 0xff;
+  int i, j;
+
+  for (j = 0; j < height; ++j) {
+    for (i = 0; i < width; ++i) {
+      const uint8_t alpha_value = argb[4 * i];
+      alpha[i] = alpha_value;
+      alpha_mask &= alpha_value;
+    }
+    argb += argb_stride;
+    alpha += alpha_stride;
+  }
+  return (alpha_mask == 0xff);
+}
+
+void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int);
+void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int);
+int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
+
+//------------------------------------------------------------------------------
+// Init function
+
+extern void WebPInitAlphaProcessingSSE2(void);
+
+static volatile VP8CPUInfo alpha_processing_last_cpuinfo_used =
+    (VP8CPUInfo)&alpha_processing_last_cpuinfo_used;
+
+void WebPInitAlphaProcessing(void) {
+  if (alpha_processing_last_cpuinfo_used == VP8GetCPUInfo) return;
+
+  WebPMultARGBRow = MultARGBRow;
+  WebPMultRow = MultRow;
+  WebPApplyAlphaMultiply = ApplyAlphaMultiply;
+  WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply_16b;
+  WebPExtractAlpha = ExtractAlpha;
+
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitAlphaProcessingSSE2();
+    }
+#endif
+  }
+  alpha_processing_last_cpuinfo_used = VP8GetCPUInfo;
+}

src/dsp/alpha_processing_sse2.c

@@ -0,0 +1,77 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Utilities for processing transparent channel.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_SSE2)
+#include <emmintrin.h>
+
+//------------------------------------------------------------------------------
+
+static int ExtractAlpha(const uint8_t* argb, int argb_stride,
+                        int width, int height,
+                        uint8_t* alpha, int alpha_stride) {
+  // alpha_and stores an 'and' operation of all the alpha[] values. The final
+  // value is not 0xff if any of the alpha[] is not equal to 0xff.
+  uint32_t alpha_and = 0xff;
+  int i, j;
+  const __m128i a_mask = _mm_set1_epi32(0xffu);  // to preserve alpha
+  const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u);
+  __m128i all_alphas = all_0xff;
+
+  // We must be able to access 3 extra bytes after the last written byte
+  // 'src[4 * width - 4]', because we don't know if alpha is the first or the
+  // last byte of the quadruplet.
+  const int limit = (width - 1) & ~7;
+
+  for (j = 0; j < height; ++j) {
+    const __m128i* src = (const __m128i*)argb;
+    for (i = 0; i < limit; i += 8) {
+      // load 32 argb bytes
+      const __m128i a0 = _mm_loadu_si128(src + 0);
+      const __m128i a1 = _mm_loadu_si128(src + 1);
+      const __m128i b0 = _mm_and_si128(a0, a_mask);
+      const __m128i b1 = _mm_and_si128(a1, a_mask);
+      const __m128i c0 = _mm_packs_epi32(b0, b1);
+      const __m128i d0 = _mm_packus_epi16(c0, c0);
+      // store
+      _mm_storel_epi64((__m128i*)&alpha[i], d0);
+      // accumulate eight alpha 'and' in parallel
+      all_alphas = _mm_and_si128(all_alphas, d0);
+      src += 2;
+    }
+    for (; i < width; ++i) {
+      const uint32_t alpha_value = argb[4 * i];
+      alpha[i] = alpha_value;
+      alpha_and &= alpha_value;
+    }
+    argb += argb_stride;
+    alpha += alpha_stride;
+  }
+  // Combine the eight alpha 'and' into a 8-bit mask.
+  alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff));
+  return (alpha_and == 0xff);
+}
+
+#endif   // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// Init function
+
+extern void WebPInitAlphaProcessingSSE2(void);
+
+void WebPInitAlphaProcessingSSE2(void) {
+#if defined(WEBP_USE_SSE2)
+  WebPExtractAlpha = ExtractAlpha;
+#endif
+}

src/dsp/cpu.c

@@ -29,19 +29,54 @@ static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) {
     "cpuid\n"
     "xchg %%edi, %%ebx\n"
     : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type));
+    : "a"(info_type), "c"(0));
 }
 #elif defined(__i386__) || defined(__x86_64__)
 static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) {
   __asm__ volatile (
     "cpuid\n"
     : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type));
+    : "a"(info_type), "c"(0));
 }
+#elif (defined(_M_X64) || defined(_M_IX86)) && \
+      defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 150030729  // >= VS2008 SP1
+#include <intrin.h>
+#define GetCPUInfo(info, type) __cpuidex(info, type, 0)  // set ecx=0
 #elif defined(WEBP_MSC_SSE2)
 #define GetCPUInfo __cpuid
 #endif
 
+// NaCl has no support for xgetbv or the raw opcode.
+#if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__))
+static WEBP_INLINE uint64_t xgetbv(void) {
+  const uint32_t ecx = 0;
+  uint32_t eax, edx;
+  // Use the raw opcode for xgetbv for compatibility with older toolchains.
+  __asm__ volatile (
+    ".byte 0x0f, 0x01, 0xd0\n"
+    : "=a"(eax), "=d"(edx) : "c" (ecx));
+  return ((uint64_t)edx << 32) | eax;
+}
+#elif (defined(_M_X64) || defined(_M_IX86)) && \
+      defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219  // >= VS2010 SP1
+#include <immintrin.h>
+#define xgetbv() _xgetbv(0)
+#elif defined(_MSC_VER) && defined(_M_IX86)
+static WEBP_INLINE uint64_t xgetbv(void) {
+  uint32_t eax_, edx_;
+  __asm {
+    xor ecx, ecx  // ecx = 0
+    // Use the raw opcode for xgetbv for compatibility with older toolchains.
+    __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0
+    mov eax_, eax
+    mov edx_, edx
+  }
+  return ((uint64_t)edx_ << 32) | eax_;
+}
+#else
+#define xgetbv() 0U  // no AVX for older x64 or unrecognized toolchains.
+#endif
+
 #if defined(__i386__) || defined(__x86_64__) || defined(WEBP_MSC_SSE2)
 static int x86CPUInfo(CPUFeature feature) {
   int cpu_info[4];
@@ -52,10 +87,23 @@ static int x86CPUInfo(CPUFeature feature) {
   if (feature == kSSE3) {
     return 0 != (cpu_info[2] & 0x00000001);
   }
+  if (feature == kAVX) {
+    // bits 27 (OSXSAVE) & 28 (256-bit AVX)
+    if ((cpu_info[2] & 0x18000000) == 0x18000000) {
+      // XMM state and YMM state enabled by the OS.
+      return (xgetbv() & 0x6) == 0x6;
+    }
+  }
+  if (feature == kAVX2) {
+    if (x86CPUInfo(kAVX)) {
+      GetCPUInfo(cpu_info, 7);
+      return ((cpu_info[1] & 0x00000020) == 0x00000020);
+    }
+  }
   return 0;
 }
 VP8CPUInfo VP8GetCPUInfo = x86CPUInfo;
-#elif defined(WEBP_ANDROID_NEON)
+#elif defined(WEBP_ANDROID_NEON)  // NB: needs to be before generic NEON test.
 static int AndroidCPUInfo(CPUFeature feature) {
   const AndroidCpuFamily cpu_family = android_getCpuFamily();
   const uint64_t cpu_features = android_getCpuFeatures();
@@ -66,7 +114,7 @@ static int AndroidCPUInfo(CPUFeature feature) {
   return 0;
 }
 VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo;
-#elif defined(__ARM_NEON__)
+#elif defined(WEBP_USE_NEON)
 // define a dummy function to enable turning off NEON at runtime by setting
 // VP8DecGetCPUInfo = NULL
 static int armCPUInfo(CPUFeature feature) {
@@ -74,6 +122,12 @@ static int armCPUInfo(CPUFeature feature) {
   return 1;
 }
 VP8CPUInfo VP8GetCPUInfo = armCPUInfo;
+#elif defined(WEBP_USE_MIPS32)
+static int mipsCPUInfo(CPUFeature feature) {
+  (void)feature;
+  return 1;
+}
+VP8CPUInfo VP8GetCPUInfo = mipsCPUInfo;
 #else
 VP8CPUInfo VP8GetCPUInfo = NULL;
 #endif

src/dsp/dec.c

@@ -15,37 +15,6 @@
 #include "../dec/vp8i.h"
 
 //------------------------------------------------------------------------------
-// run-time tables (~4k)
-
-static uint8_t abs0[255 + 255 + 1];     // abs(i)
-static uint8_t abs1[255 + 255 + 1];     // abs(i)>>1
-static int8_t sclip1[1020 + 1020 + 1];  // clips [-1020, 1020] to [-128, 127]
-static int8_t sclip2[112 + 112 + 1];    // clips [-112, 112] to [-16, 15]
-static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
-
-// We declare this variable 'volatile' to prevent instruction reordering
-// and make sure it's set to true _last_ (so as to be thread-safe)
-static volatile int tables_ok = 0;
-
-static void DspInitTables(void) {
-  if (!tables_ok) {
-    int i;
-    for (i = -255; i <= 255; ++i) {
-      abs0[255 + i] = (i < 0) ? -i : i;
-      abs1[255 + i] = abs0[255 + i] >> 1;
-    }
-    for (i = -1020; i <= 1020; ++i) {
-      sclip1[1020 + i] = (i < -128) ? -128 : (i > 127) ? 127 : i;
-    }
-    for (i = -112; i <= 112; ++i) {
-      sclip2[112 + i] = (i < -16) ? -16 : (i > 15) ? 15 : i;
-    }
-    for (i = -255; i <= 255 + 255; ++i) {
-      clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i;
-    }
-    tables_ok = 1;
-  }
-}
 
 static WEBP_INLINE uint8_t clip_8b(int v) {
   return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
@@ -146,10 +115,10 @@ static void TransformDC(const int16_t *in, uint8_t* dst) {
 }
 
 static void TransformDCUV(const int16_t* in, uint8_t* dst) {
-  if (in[0 * 16]) TransformDC(in + 0 * 16, dst);
-  if (in[1 * 16]) TransformDC(in + 1 * 16, dst + 4);
-  if (in[2 * 16]) TransformDC(in + 2 * 16, dst + 4 * BPS);
-  if (in[3 * 16]) TransformDC(in + 3 * 16, dst + 4 * BPS + 4);
+  if (in[0 * 16]) VP8TransformDC(in + 0 * 16, dst);
+  if (in[1 * 16]) VP8TransformDC(in + 1 * 16, dst + 4);
+  if (in[2 * 16]) VP8TransformDC(in + 2 * 16, dst + 4 * BPS);
+  if (in[3 * 16]) VP8TransformDC(in + 3 * 16, dst + 4 * BPS + 4);
 }
 
 #undef STORE
@@ -184,7 +153,7 @@ static void TransformWHT(const int16_t* in, int16_t* out) {
   }
 }
 
-void (*VP8TransformWHT)(const int16_t* in, int16_t* out) = TransformWHT;
+void (*VP8TransformWHT)(const int16_t* in, int16_t* out);
 
 //------------------------------------------------------------------------------
 // Intra predictions
@@ -193,7 +162,7 @@ void (*VP8TransformWHT)(const int16_t* in, int16_t* out) = TransformWHT;
 
 static WEBP_INLINE void TrueMotion(uint8_t *dst, int size) {
   const uint8_t* top = dst - BPS;
-  const uint8_t* const clip0 = clip1 + 255 - top[-1];
+  const uint8_t* const clip0 = VP8kclip1 - top[-1];
   int y;
   for (y = 0; y < size; ++y) {
     const uint8_t* const clip = clip0 + dst[-1];
@@ -448,14 +417,9 @@ static void HE8uv(uint8_t *dst) {    // horizontal
 // helper for chroma-DC predictions
 static WEBP_INLINE void Put8x8uv(uint8_t value, uint8_t* dst) {
   int j;
-#ifndef WEBP_REFERENCE_IMPLEMENTATION
-  const uint64_t v = (uint64_t)value * 0x0101010101010101ULL;
   for (j = 0; j < 8; ++j) {
-    *(uint64_t*)(dst + j * BPS) = v;
+    memset(dst + j * BPS, value, 8);
   }
-#else
-  for (j = 0; j < 8; ++j) memset(dst + j * BPS, value, 8);
-#endif
 }
 
 static void DC8uv(uint8_t *dst) {     // DC
@@ -512,61 +476,62 @@ const VP8PredFunc VP8PredChroma8[NUM_B_DC_MODES] = {
 // 4 pixels in, 2 pixels out
 static WEBP_INLINE void do_filter2(uint8_t* p, int step) {
   const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
-  const int a = 3 * (q0 - p0) + sclip1[1020 + p1 - q1];
-  const int a1 = sclip2[112 + ((a + 4) >> 3)];
-  const int a2 = sclip2[112 + ((a + 3) >> 3)];
-  p[-step] = clip1[255 + p0 + a2];
-  p[    0] = clip1[255 + q0 - a1];
+  const int a = 3 * (q0 - p0) + VP8ksclip1[p1 - q1];  // in [-893,892]
+  const int a1 = VP8ksclip2[(a + 4) >> 3];            // in [-16,15]
+  const int a2 = VP8ksclip2[(a + 3) >> 3];
+  p[-step] = VP8kclip1[p0 + a2];
+  p[    0] = VP8kclip1[q0 - a1];
 }
 
 // 4 pixels in, 4 pixels out
 static WEBP_INLINE void do_filter4(uint8_t* p, int step) {
   const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
   const int a = 3 * (q0 - p0);
-  const int a1 = sclip2[112 + ((a + 4) >> 3)];
-  const int a2 = sclip2[112 + ((a + 3) >> 3)];
+  const int a1 = VP8ksclip2[(a + 4) >> 3];
+  const int a2 = VP8ksclip2[(a + 3) >> 3];
   const int a3 = (a1 + 1) >> 1;
-  p[-2*step] = clip1[255 + p1 + a3];
-  p[-  step] = clip1[255 + p0 + a2];
-  p[      0] = clip1[255 + q0 - a1];
-  p[   step] = clip1[255 + q1 - a3];
+  p[-2*step] = VP8kclip1[p1 + a3];
+  p[-  step] = VP8kclip1[p0 + a2];
+  p[      0] = VP8kclip1[q0 - a1];
+  p[   step] = VP8kclip1[q1 - a3];
 }
 
 // 6 pixels in, 6 pixels out
 static WEBP_INLINE void do_filter6(uint8_t* p, int step) {
   const int p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step];
   const int q0 = p[0], q1 = p[step], q2 = p[2*step];
-  const int a = sclip1[1020 + 3 * (q0 - p0) + sclip1[1020 + p1 - q1]];
+  const int a = VP8ksclip1[3 * (q0 - p0) + VP8ksclip1[p1 - q1]];
+  // a is in [-128,127], a1 in [-27,27], a2 in [-18,18] and a3 in [-9,9]
   const int a1 = (27 * a + 63) >> 7;  // eq. to ((3 * a + 7) * 9) >> 7
   const int a2 = (18 * a + 63) >> 7;  // eq. to ((2 * a + 7) * 9) >> 7
   const int a3 = (9  * a + 63) >> 7;  // eq. to ((1 * a + 7) * 9) >> 7
-  p[-3*step] = clip1[255 + p2 + a3];
-  p[-2*step] = clip1[255 + p1 + a2];
-  p[-  step] = clip1[255 + p0 + a1];
-  p[      0] = clip1[255 + q0 - a1];
-  p[   step] = clip1[255 + q1 - a2];
-  p[ 2*step] = clip1[255 + q2 - a3];
+  p[-3*step] = VP8kclip1[p2 + a3];
+  p[-2*step] = VP8kclip1[p1 + a2];
+  p[-  step] = VP8kclip1[p0 + a1];
+  p[      0] = VP8kclip1[q0 - a1];
+  p[   step] = VP8kclip1[q1 - a2];
+  p[ 2*step] = VP8kclip1[q2 - a3];
 }
 
 static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) {
   const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
-  return (abs0[255 + p1 - p0] > thresh) || (abs0[255 + q1 - q0] > thresh);
+  return (VP8kabs0[p1 - p0] > thresh) || (VP8kabs0[q1 - q0] > thresh);
 }
 
-static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int thresh) {
-  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
-  return (2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) <= thresh;
+static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int t) {
+  const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return ((4 * VP8kabs0[p0 - q0] + VP8kabs0[p1 - q1]) <= t);
 }
 
 static WEBP_INLINE int needs_filter2(const uint8_t* p,
                                      int step, int t, int it) {
-  const int p3 = p[-4*step], p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step];
-  const int q0 = p[0], q1 = p[step], q2 = p[2*step], q3 = p[3*step];
-  if ((2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) > t)
-    return 0;
-  return abs0[255 + p3 - p2] <= it && abs0[255 + p2 - p1] <= it &&
-         abs0[255 + p1 - p0] <= it && abs0[255 + q3 - q2] <= it &&
-         abs0[255 + q2 - q1] <= it && abs0[255 + q1 - q0] <= it;
+  const int p3 = p[-4 * step], p2 = p[-3 * step], p1 = p[-2 * step];
+  const int p0 = p[-step], q0 = p[0];
+  const int q1 = p[step], q2 = p[2 * step], q3 = p[3 * step];
+  if ((4 * VP8kabs0[p0 - q0] + VP8kabs0[p1 - q1]) > t) return 0;
+  return VP8kabs0[p3 - p2] <= it && VP8kabs0[p2 - p1] <= it &&
+         VP8kabs0[p1 - p0] <= it && VP8kabs0[q3 - q2] <= it &&
+         VP8kabs0[q2 - q1] <= it && VP8kabs0[q1 - q0] <= it;
 }
 
 //------------------------------------------------------------------------------
@@ -574,8 +539,9 @@ static WEBP_INLINE int needs_filter2(const uint8_t* p,
 
 static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
   int i;
+  const int thresh2 = 2 * thresh + 1;
   for (i = 0; i < 16; ++i) {
-    if (needs_filter(p + i, stride, thresh)) {
+    if (needs_filter(p + i, stride, thresh2)) {
       do_filter2(p + i, stride);
     }
   }
@@ -583,8 +549,9 @@ static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
 
 static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
   int i;
+  const int thresh2 = 2 * thresh + 1;
   for (i = 0; i < 16; ++i) {
-    if (needs_filter(p + i * stride, 1, thresh)) {
+    if (needs_filter(p + i * stride, 1, thresh2)) {
       do_filter2(p + i * stride, 1);
     }
   }
@@ -612,8 +579,9 @@ static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
 static WEBP_INLINE void FilterLoop26(uint8_t* p,
                                      int hstride, int vstride, int size,
                                      int thresh, int ithresh, int hev_thresh) {
+  const int thresh2 = 2 * thresh + 1;
   while (size-- > 0) {
-    if (needs_filter2(p, hstride, thresh, ithresh)) {
+    if (needs_filter2(p, hstride, thresh2, ithresh)) {
       if (hev(p, hstride, hev_thresh)) {
         do_filter2(p, hstride);
       } else {
@@ -627,8 +595,9 @@ static WEBP_INLINE void FilterLoop26(uint8_t* p,
 static WEBP_INLINE void FilterLoop24(uint8_t* p,
                                      int hstride, int vstride, int size,
                                      int thresh, int ithresh, int hev_thresh) {
+  const int thresh2 = 2 * thresh + 1;
   while (size-- > 0) {
-    if (needs_filter2(p, hstride, thresh, ithresh)) {
+    if (needs_filter2(p, hstride, thresh2, ithresh)) {
       if (hev(p, hstride, hev_thresh)) {
         do_filter2(p, hstride);
       } else {
@@ -717,10 +686,17 @@ VP8SimpleFilterFunc VP8SimpleHFilter16i;
 
 extern void VP8DspInitSSE2(void);
 extern void VP8DspInitNEON(void);
+extern void VP8DspInitMIPS32(void);
+
+static volatile VP8CPUInfo dec_last_cpuinfo_used =
+    (VP8CPUInfo)&dec_last_cpuinfo_used;
 
 void VP8DspInit(void) {
-  DspInitTables();
+  if (dec_last_cpuinfo_used == VP8GetCPUInfo) return;
 
+  VP8InitClipTables();
+
+  VP8TransformWHT = TransformWHT;
   VP8Transform = TransformTwo;
   VP8TransformUV = TransformUV;
   VP8TransformDC = TransformDC;
@@ -741,7 +717,7 @@ void VP8DspInit(void) {
   VP8SimpleHFilter16i = SimpleHFilter16i;
 
   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
-  if (VP8GetCPUInfo) {
+  if (VP8GetCPUInfo != NULL) {
 #if defined(WEBP_USE_SSE2)
     if (VP8GetCPUInfo(kSSE2)) {
       VP8DspInitSSE2();
@@ -750,7 +726,11 @@ void VP8DspInit(void) {
     if (VP8GetCPUInfo(kNEON)) {
       VP8DspInitNEON();
     }
+#elif defined(WEBP_USE_MIPS32)
+    if (VP8GetCPUInfo(kMIPS32)) {
+      VP8DspInitMIPS32();
+    }
 #endif
   }
+  dec_last_cpuinfo_used = VP8GetCPUInfo;
 }
-

src/dsp/dec_clip_tables.c

@@ -0,0 +1,366 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Clipping tables for filtering
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./dsp.h"
+
+#define USE_STATIC_TABLES     // undefine to have run-time table initialization
+
+#ifdef USE_STATIC_TABLES
+
+static const uint8_t abs0[255 + 255 + 1] = {
+  0xff, 0xfe, 0xfd, 0xfc, 0xfb, 0xfa, 0xf9, 0xf8, 0xf7, 0xf6, 0xf5, 0xf4,
+  0xf3, 0xf2, 0xf1, 0xf0, 0xef, 0xee, 0xed, 0xec, 0xeb, 0xea, 0xe9, 0xe8,
+  0xe7, 0xe6, 0xe5, 0xe4, 0xe3, 0xe2, 0xe1, 0xe0, 0xdf, 0xde, 0xdd, 0xdc,
+  0xdb, 0xda, 0xd9, 0xd8, 0xd7, 0xd6, 0xd5, 0xd4, 0xd3, 0xd2, 0xd1, 0xd0,
+  0xcf, 0xce, 0xcd, 0xcc, 0xcb, 0xca, 0xc9, 0xc8, 0xc7, 0xc6, 0xc5, 0xc4,
+  0xc3, 0xc2, 0xc1, 0xc0, 0xbf, 0xbe, 0xbd, 0xbc, 0xbb, 0xba, 0xb9, 0xb8,
+  0xb7, 0xb6, 0xb5, 0xb4, 0xb3, 0xb2, 0xb1, 0xb0, 0xaf, 0xae, 0xad, 0xac,
+  0xab, 0xaa, 0xa9, 0xa8, 0xa7, 0xa6, 0xa5, 0xa4, 0xa3, 0xa2, 0xa1, 0xa0,
+  0x9f, 0x9e, 0x9d, 0x9c, 0x9b, 0x9a, 0x99, 0x98, 0x97, 0x96, 0x95, 0x94,
+  0x93, 0x92, 0x91, 0x90, 0x8f, 0x8e, 0x8d, 0x8c, 0x8b, 0x8a, 0x89, 0x88,
+  0x87, 0x86, 0x85, 0x84, 0x83, 0x82, 0x81, 0x80, 0x7f, 0x7e, 0x7d, 0x7c,
+  0x7b, 0x7a, 0x79, 0x78, 0x77, 0x76, 0x75, 0x74, 0x73, 0x72, 0x71, 0x70,
+  0x6f, 0x6e, 0x6d, 0x6c, 0x6b, 0x6a, 0x69, 0x68, 0x67, 0x66, 0x65, 0x64,
+  0x63, 0x62, 0x61, 0x60, 0x5f, 0x5e, 0x5d, 0x5c, 0x5b, 0x5a, 0x59, 0x58,
+  0x57, 0x56, 0x55, 0x54, 0x53, 0x52, 0x51, 0x50, 0x4f, 0x4e, 0x4d, 0x4c,
+  0x4b, 0x4a, 0x49, 0x48, 0x47, 0x46, 0x45, 0x44, 0x43, 0x42, 0x41, 0x40,
+  0x3f, 0x3e, 0x3d, 0x3c, 0x3b, 0x3a, 0x39, 0x38, 0x37, 0x36, 0x35, 0x34,
+  0x33, 0x32, 0x31, 0x30, 0x2f, 0x2e, 0x2d, 0x2c, 0x2b, 0x2a, 0x29, 0x28,
+  0x27, 0x26, 0x25, 0x24, 0x23, 0x22, 0x21, 0x20, 0x1f, 0x1e, 0x1d, 0x1c,
+  0x1b, 0x1a, 0x19, 0x18, 0x17, 0x16, 0x15, 0x14, 0x13, 0x12, 0x11, 0x10,
+  0x0f, 0x0e, 0x0d, 0x0c, 0x0b, 0x0a, 0x09, 0x08, 0x07, 0x06, 0x05, 0x04,
+  0x03, 0x02, 0x01, 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08,
+  0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x10, 0x11, 0x12, 0x13, 0x14,
+  0x15, 0x16, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, 0x20,
+  0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27, 0x28, 0x29, 0x2a, 0x2b, 0x2c,
+  0x2d, 0x2e, 0x2f, 0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37, 0x38,
+  0x39, 0x3a, 0x3b, 0x3c, 0x3d, 0x3e, 0x3f, 0x40, 0x41, 0x42, 0x43, 0x44,
+  0x45, 0x46, 0x47, 0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f, 0x50,
+  0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5a, 0x5b, 0x5c,
+  0x5d, 0x5e, 0x5f, 0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68,
+  0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x70, 0x71, 0x72, 0x73, 0x74,
+  0x75, 0x76, 0x77, 0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x80,
+  0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89, 0x8a, 0x8b, 0x8c,
+  0x8d, 0x8e, 0x8f, 0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98,
+  0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f, 0xa0, 0xa1, 0xa2, 0xa3, 0xa4,
+  0xa5, 0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf, 0xb0,
+  0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xbb, 0xbc,
+  0xbd, 0xbe, 0xbf, 0xc0, 0xc1, 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc7, 0xc8,
+  0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf, 0xd0, 0xd1, 0xd2, 0xd3, 0xd4,
+  0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xdb, 0xdc, 0xdd, 0xde, 0xdf, 0xe0,
+  0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea, 0xeb, 0xec,
+  0xed, 0xee, 0xef, 0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+  0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff
+};
+
+static const int8_t sclip1[1020 + 1020 + 1] = {
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+  0x80, 0x80, 0x80, 0x80, 0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
+  0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f, 0x90, 0x91, 0x92, 0x93,
+  0x94, 0x95, 0x96, 0x97, 0x98, 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f,
+  0xa0, 0xa1, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xab,
+  0xac, 0xad, 0xae, 0xaf, 0xb0, 0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7,
+  0xb8, 0xb9, 0xba, 0xbb, 0xbc, 0xbd, 0xbe, 0xbf, 0xc0, 0xc1, 0xc2, 0xc3,
+  0xc4, 0xc5, 0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf,
+  0xd0, 0xd1, 0xd2, 0xd3, 0xd4, 0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xdb,
+  0xdc, 0xdd, 0xde, 0xdf, 0xe0, 0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7,
+  0xe8, 0xe9, 0xea, 0xeb, 0xec, 0xed, 0xee, 0xef, 0xf0, 0xf1, 0xf2, 0xf3,
+  0xf4, 0xf5, 0xf6, 0xf7, 0xf8, 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff,
+  0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b,
+  0x0c, 0x0d, 0x0e, 0x0f, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
+  0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, 0x20, 0x21, 0x22, 0x23,
+  0x24, 0x25, 0x26, 0x27, 0x28, 0x29, 0x2a, 0x2b, 0x2c, 0x2d, 0x2e, 0x2f,
+  0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37, 0x38, 0x39, 0x3a, 0x3b,
+  0x3c, 0x3d, 0x3e, 0x3f, 0x40, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47,
+  0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f, 0x50, 0x51, 0x52, 0x53,
+  0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5a, 0x5b, 0x5c, 0x5d, 0x5e, 0x5f,
+  0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69, 0x6a, 0x6b,
+  0x6c, 0x6d, 0x6e, 0x6f, 0x70, 0x71, 0x72, 0x73, 0x74, 0x75, 0x76, 0x77,
+  0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+  0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f
+};
+
+static const int8_t sclip2[112 + 112 + 1] = {
+  0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
+  0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
+  0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
+  0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
+  0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
+  0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
+  0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
+  0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0,
+  0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8, 0xf9, 0xfa, 0xfb,
+  0xfc, 0xfd, 0xfe, 0xff, 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
+  0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+  0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+  0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+  0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+  0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+  0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+  0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+  0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+  0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f
+};
+
+static const uint8_t clip1[255 + 511 + 1] = {
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08,
+  0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x10, 0x11, 0x12, 0x13, 0x14,
+  0x15, 0x16, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, 0x20,
+  0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27, 0x28, 0x29, 0x2a, 0x2b, 0x2c,
+  0x2d, 0x2e, 0x2f, 0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37, 0x38,
+  0x39, 0x3a, 0x3b, 0x3c, 0x3d, 0x3e, 0x3f, 0x40, 0x41, 0x42, 0x43, 0x44,
+  0x45, 0x46, 0x47, 0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f, 0x50,
+  0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5a, 0x5b, 0x5c,
+  0x5d, 0x5e, 0x5f, 0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68,
+  0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x70, 0x71, 0x72, 0x73, 0x74,
+  0x75, 0x76, 0x77, 0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x80,
+  0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89, 0x8a, 0x8b, 0x8c,
+  0x8d, 0x8e, 0x8f, 0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98,
+  0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f, 0xa0, 0xa1, 0xa2, 0xa3, 0xa4,
+  0xa5, 0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf, 0xb0,
+  0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xbb, 0xbc,
+  0xbd, 0xbe, 0xbf, 0xc0, 0xc1, 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc7, 0xc8,
+  0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf, 0xd0, 0xd1, 0xd2, 0xd3, 0xd4,
+  0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xdb, 0xdc, 0xdd, 0xde, 0xdf, 0xe0,
+  0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea, 0xeb, 0xec,
+  0xed, 0xee, 0xef, 0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+  0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
+};
+
+#else
+
+// uninitialized tables
+static uint8_t abs0[255 + 255 + 1];
+static int8_t sclip1[1020 + 1020 + 1];
+static int8_t sclip2[112 + 112 + 1];
+static uint8_t clip1[255 + 511 + 1];
+
+// We declare this variable 'volatile' to prevent instruction reordering
+// and make sure it's set to true _last_ (so as to be thread-safe)
+static volatile int tables_ok = 0;
+
+#endif
+
+const int8_t* const VP8ksclip1 = &sclip1[1020];
+const int8_t* const VP8ksclip2 = &sclip2[112];
+const uint8_t* const VP8kclip1 = &clip1[255];
+const uint8_t* const VP8kabs0 = &abs0[255];
+
+void VP8InitClipTables(void) {
+#if !defined(USE_STATIC_TABLES)
+  int i;
+  if (!tables_ok) {
+    for (i = -255; i <= 255; ++i) {
+      abs0[255 + i] = (i < 0) ? -i : i;
+    }
+    for (i = -1020; i <= 1020; ++i) {
+      sclip1[1020 + i] = (i < -128) ? -128 : (i > 127) ? 127 : i;
+    }
+    for (i = -112; i <= 112; ++i) {
+      sclip2[112 + i] = (i < -16) ? -16 : (i > 15) ? 15 : i;
+    }
+    for (i = -255; i <= 255 + 255; ++i) {
+      clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i;
+    }
+    tables_ok = 1;
+  }
+#endif    // USE_STATIC_TABLES
+}

src/dsp/dec_mips32.c

@@ -0,0 +1,578 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// MIPS version of dsp functions
+//
+// Author(s):  Djordje Pesut    (djordje.pesut@imgtec.com)
+//             Jovan Zelincevic (jovan.zelincevic@imgtec.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_MIPS32)
+
+static const int kC1 = 20091 + (1 << 16);
+static const int kC2 = 35468;
+
+static WEBP_INLINE int abs_mips32(int x) {
+  const int sign = x >> 31;
+  return (x ^ sign) - sign;
+}
+
+// 4 pixels in, 2 pixels out
+static WEBP_INLINE void do_filter2(uint8_t* p, int step) {
+  const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  const int a = 3 * (q0 - p0) + VP8ksclip1[p1 - q1];
+  const int a1 = VP8ksclip2[(a + 4) >> 3];
+  const int a2 = VP8ksclip2[(a + 3) >> 3];
+  p[-step] = VP8kclip1[p0 + a2];
+  p[    0] = VP8kclip1[q0 - a1];
+}
+
+// 4 pixels in, 4 pixels out
+static WEBP_INLINE void do_filter4(uint8_t* p, int step) {
+  const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  const int a = 3 * (q0 - p0);
+  const int a1 = VP8ksclip2[(a + 4) >> 3];
+  const int a2 = VP8ksclip2[(a + 3) >> 3];
+  const int a3 = (a1 + 1) >> 1;
+  p[-2 * step] = VP8kclip1[p1 + a3];
+  p[-    step] = VP8kclip1[p0 + a2];
+  p[        0] = VP8kclip1[q0 - a1];
+  p[     step] = VP8kclip1[q1 - a3];
+}
+
+// 6 pixels in, 6 pixels out
+static WEBP_INLINE void do_filter6(uint8_t* p, int step) {
+  const int p2 = p[-3 * step], p1 = p[-2 * step], p0 = p[-step];
+  const int q0 = p[0], q1 = p[step], q2 = p[2 * step];
+  const int a = VP8ksclip1[3 * (q0 - p0) + VP8ksclip1[p1 - q1]];
+  const int a1 = (27 * a + 63) >> 7;  // eq. to ((3 * a + 7) * 9) >> 7
+  const int a2 = (18 * a + 63) >> 7;  // eq. to ((2 * a + 7) * 9) >> 7
+  const int a3 = (9  * a + 63) >> 7;  // eq. to ((1 * a + 7) * 9) >> 7
+  p[-3 * step] = VP8kclip1[p2 + a3];
+  p[-2 * step] = VP8kclip1[p1 + a2];
+  p[-    step] = VP8kclip1[p0 + a1];
+  p[        0] = VP8kclip1[q0 - a1];
+  p[     step] = VP8kclip1[q1 - a2];
+  p[ 2 * step] = VP8kclip1[q2 - a3];
+}
+
+static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) {
+  const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return (abs_mips32(p1 - p0) > thresh) || (abs_mips32(q1 - q0) > thresh);
+}
+
+static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int thresh) {
+  const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return ((2 * abs_mips32(p0 - q0) + (abs_mips32(p1 - q1) >> 1)) <= thresh);
+}
+
+static WEBP_INLINE int needs_filter2(const uint8_t* p,
+                                     int step, int t, int it) {
+  const int p3 = p[-4 * step], p2 = p[-3 * step];
+  const int p1 = p[-2 * step], p0 = p[-step];
+  const int q0 = p[0], q1 = p[step], q2 = p[2 * step], q3 = p[3 * step];
+  if ((2 * abs_mips32(p0 - q0) + (abs_mips32(p1 - q1) >> 1)) > t) {
+    return 0;
+  }
+  return abs_mips32(p3 - p2) <= it && abs_mips32(p2 - p1) <= it &&
+         abs_mips32(p1 - p0) <= it && abs_mips32(q3 - q2) <= it &&
+         abs_mips32(q2 - q1) <= it && abs_mips32(q1 - q0) <= it;
+}
+
+static WEBP_INLINE void FilterLoop26(uint8_t* p,
+                                     int hstride, int vstride, int size,
+                                     int thresh, int ithresh, int hev_thresh) {
+  while (size-- > 0) {
+    if (needs_filter2(p, hstride, thresh, ithresh)) {
+      if (hev(p, hstride, hev_thresh)) {
+        do_filter2(p, hstride);
+      } else {
+        do_filter6(p, hstride);
+      }
+    }
+    p += vstride;
+  }
+}
+
+static WEBP_INLINE void FilterLoop24(uint8_t* p,
+                                     int hstride, int vstride, int size,
+                                     int thresh, int ithresh, int hev_thresh) {
+  while (size-- > 0) {
+    if (needs_filter2(p, hstride, thresh, ithresh)) {
+      if (hev(p, hstride, hev_thresh)) {
+        do_filter2(p, hstride);
+      } else {
+        do_filter4(p, hstride);
+      }
+    }
+    p += vstride;
+  }
+}
+
+// on macroblock edges
+static void VFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop26(p, stride, 1, 16, thresh, ithresh, hev_thresh);
+}
+
+static void HFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop26(p, 1, stride, 16, thresh, ithresh, hev_thresh);
+}
+
+// 8-pixels wide variant, for chroma filtering
+static void VFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
+  FilterLoop26(u, stride, 1, 8, thresh, ithresh, hev_thresh);
+  FilterLoop26(v, stride, 1, 8, thresh, ithresh, hev_thresh);
+}
+
+static void HFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
+  FilterLoop26(u, 1, stride, 8, thresh, ithresh, hev_thresh);
+  FilterLoop26(v, 1, stride, 8, thresh, ithresh, hev_thresh);
+}
+
+static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
+  FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
+}
+
+static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
+  FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
+}
+
+// on three inner edges
+static void VFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh);
+  }
+}
+
+static void HFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Simple In-loop filtering (Paragraph 15.2)
+
+static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    if (needs_filter(p + i, stride, thresh)) {
+      do_filter2(p + i, stride);
+    }
+  }
+}
+
+static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    if (needs_filter(p + i * stride, 1, thresh)) {
+      do_filter2(p + i * stride, 1);
+    }
+  }
+}
+
+static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    SimpleVFilter16(p, stride, thresh);
+  }
+}
+
+static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    SimpleHFilter16(p, stride, thresh);
+  }
+}
+
+static void TransformOne(const int16_t* in, uint8_t* dst) {
+  int temp0, temp1, temp2, temp3, temp4;
+  int temp5, temp6, temp7, temp8, temp9;
+  int temp10, temp11, temp12, temp13, temp14;
+  int temp15, temp16, temp17, temp18;
+  int16_t* p_in = (int16_t*)in;
+
+  // loops unrolled and merged to avoid usage of tmp buffer
+  // and to reduce number of stalls. MUL macro is written
+  // in assembler and inlined
+  __asm__ volatile(
+    "lh       %[temp0],  0(%[in])                      \n\t"
+    "lh       %[temp8],  16(%[in])                     \n\t"
+    "lh       %[temp4],  8(%[in])                      \n\t"
+    "lh       %[temp12], 24(%[in])                     \n\t"
+    "addu     %[temp16], %[temp0],  %[temp8]           \n\t"
+    "subu     %[temp0],  %[temp0],  %[temp8]           \n\t"
+    "mul      %[temp8],  %[temp4],  %[kC2]             \n\t"
+    "mul      %[temp17], %[temp12], %[kC1]             \n\t"
+    "mul      %[temp4],  %[temp4],  %[kC1]             \n\t"
+    "mul      %[temp12], %[temp12], %[kC2]             \n\t"
+    "lh       %[temp1],  2(%[in])                      \n\t"
+    "lh       %[temp5],  10(%[in])                     \n\t"
+    "lh       %[temp9],  18(%[in])                     \n\t"
+    "lh       %[temp13], 26(%[in])                     \n\t"
+    "sra      %[temp8],  %[temp8],  16                 \n\t"
+    "sra      %[temp17], %[temp17], 16                 \n\t"
+    "sra      %[temp4],  %[temp4],  16                 \n\t"
+    "sra      %[temp12], %[temp12], 16                 \n\t"
+    "lh       %[temp2],  4(%[in])                      \n\t"
+    "lh       %[temp6],  12(%[in])                     \n\t"
+    "lh       %[temp10], 20(%[in])                     \n\t"
+    "lh       %[temp14], 28(%[in])                     \n\t"
+    "subu     %[temp17], %[temp8],  %[temp17]          \n\t"
+    "addu     %[temp4],  %[temp4],  %[temp12]          \n\t"
+    "addu     %[temp8],  %[temp16], %[temp4]           \n\t"
+    "subu     %[temp4],  %[temp16], %[temp4]           \n\t"
+    "addu     %[temp16], %[temp1],  %[temp9]           \n\t"
+    "subu     %[temp1],  %[temp1],  %[temp9]           \n\t"
+    "lh       %[temp3],  6(%[in])                      \n\t"
+    "lh       %[temp7],  14(%[in])                     \n\t"
+    "lh       %[temp11], 22(%[in])                     \n\t"
+    "lh       %[temp15], 30(%[in])                     \n\t"
+    "addu     %[temp12], %[temp0],  %[temp17]          \n\t"
+    "subu     %[temp0],  %[temp0],  %[temp17]          \n\t"
+    "mul      %[temp9],  %[temp5],  %[kC2]             \n\t"
+    "mul      %[temp17], %[temp13], %[kC1]             \n\t"
+    "mul      %[temp5],  %[temp5],  %[kC1]             \n\t"
+    "mul      %[temp13], %[temp13], %[kC2]             \n\t"
+    "sra      %[temp9],  %[temp9],  16                 \n\t"
+    "sra      %[temp17], %[temp17], 16                 \n\t"
+    "subu     %[temp17], %[temp9],  %[temp17]          \n\t"
+    "sra      %[temp5],  %[temp5],  16                 \n\t"
+    "sra      %[temp13], %[temp13], 16                 \n\t"
+    "addu     %[temp5],  %[temp5],  %[temp13]          \n\t"
+    "addu     %[temp13], %[temp1],  %[temp17]          \n\t"
+    "subu     %[temp1],  %[temp1],  %[temp17]          \n\t"
+    "mul      %[temp17], %[temp14], %[kC1]             \n\t"
+    "mul      %[temp14], %[temp14], %[kC2]             \n\t"
+    "addu     %[temp9],  %[temp16], %[temp5]           \n\t"
+    "subu     %[temp5],  %[temp16], %[temp5]           \n\t"
+    "addu     %[temp16], %[temp2],  %[temp10]          \n\t"
+    "subu     %[temp2],  %[temp2],  %[temp10]          \n\t"
+    "mul      %[temp10], %[temp6],  %[kC2]             \n\t"
+    "mul      %[temp6],  %[temp6],  %[kC1]             \n\t"
+    "sra      %[temp17], %[temp17], 16                 \n\t"
+    "sra      %[temp14], %[temp14], 16                 \n\t"
+    "sra      %[temp10], %[temp10], 16                 \n\t"
+    "sra      %[temp6],  %[temp6],  16                 \n\t"
+    "subu     %[temp17], %[temp10], %[temp17]          \n\t"
+    "addu     %[temp6],  %[temp6],  %[temp14]          \n\t"
+    "addu     %[temp10], %[temp16], %[temp6]           \n\t"
+    "subu     %[temp6],  %[temp16], %[temp6]           \n\t"
+    "addu     %[temp14], %[temp2],  %[temp17]          \n\t"
+    "subu     %[temp2],  %[temp2],  %[temp17]          \n\t"
+    "mul      %[temp17], %[temp15], %[kC1]             \n\t"
+    "mul      %[temp15], %[temp15], %[kC2]             \n\t"
+    "addu     %[temp16], %[temp3],  %[temp11]          \n\t"
+    "subu     %[temp3],  %[temp3],  %[temp11]          \n\t"
+    "mul      %[temp11], %[temp7],  %[kC2]             \n\t"
+    "mul      %[temp7],  %[temp7],  %[kC1]             \n\t"
+    "addiu    %[temp8],  %[temp8],  4                  \n\t"
+    "addiu    %[temp12], %[temp12], 4                  \n\t"
+    "addiu    %[temp0],  %[temp0],  4                  \n\t"
+    "addiu    %[temp4],  %[temp4],  4                  \n\t"
+    "sra      %[temp17], %[temp17], 16                 \n\t"
+    "sra      %[temp15], %[temp15], 16                 \n\t"
+    "sra      %[temp11], %[temp11], 16                 \n\t"
+    "sra      %[temp7],  %[temp7],  16                 \n\t"
+    "subu     %[temp17], %[temp11], %[temp17]          \n\t"
+    "addu     %[temp7],  %[temp7],  %[temp15]          \n\t"
+    "addu     %[temp15], %[temp3],  %[temp17]          \n\t"
+    "subu     %[temp3],  %[temp3],  %[temp17]          \n\t"
+    "addu     %[temp11], %[temp16], %[temp7]           \n\t"
+    "subu     %[temp7],  %[temp16], %[temp7]           \n\t"
+    "addu     %[temp16], %[temp8],  %[temp10]          \n\t"
+    "subu     %[temp8],  %[temp8],  %[temp10]          \n\t"
+    "mul      %[temp10], %[temp9],  %[kC2]             \n\t"
+    "mul      %[temp17], %[temp11], %[kC1]             \n\t"
+    "mul      %[temp9],  %[temp9],  %[kC1]             \n\t"
+    "mul      %[temp11], %[temp11], %[kC2]             \n\t"
+    "sra      %[temp10], %[temp10], 16                 \n\t"
+    "sra      %[temp17], %[temp17], 16                 \n\t"
+    "sra      %[temp9],  %[temp9],  16                 \n\t"
+    "sra      %[temp11], %[temp11], 16                 \n\t"
+    "subu     %[temp17], %[temp10], %[temp17]          \n\t"
+    "addu     %[temp11], %[temp9],  %[temp11]          \n\t"
+    "addu     %[temp10], %[temp12], %[temp14]          \n\t"
+    "subu     %[temp12], %[temp12], %[temp14]          \n\t"
+    "mul      %[temp14], %[temp13], %[kC2]             \n\t"
+    "mul      %[temp9],  %[temp15], %[kC1]             \n\t"
+    "mul      %[temp13], %[temp13], %[kC1]             \n\t"
+    "mul      %[temp15], %[temp15], %[kC2]             \n\t"
+    "sra      %[temp14], %[temp14], 16                 \n\t"
+    "sra      %[temp9],  %[temp9],  16                 \n\t"
+    "sra      %[temp13], %[temp13], 16                 \n\t"
+    "sra      %[temp15], %[temp15], 16                 \n\t"
+    "subu     %[temp9],  %[temp14], %[temp9]           \n\t"
+    "addu     %[temp15], %[temp13], %[temp15]          \n\t"
+    "addu     %[temp14], %[temp0],  %[temp2]           \n\t"
+    "subu     %[temp0],  %[temp0],  %[temp2]           \n\t"
+    "mul      %[temp2],  %[temp1],  %[kC2]             \n\t"
+    "mul      %[temp13], %[temp3],  %[kC1]             \n\t"
+    "mul      %[temp1],  %[temp1],  %[kC1]             \n\t"
+    "mul      %[temp3],  %[temp3],  %[kC2]             \n\t"
+    "sra      %[temp2],  %[temp2],  16                 \n\t"
+    "sra      %[temp13], %[temp13], 16                 \n\t"
+    "sra      %[temp1],  %[temp1],  16                 \n\t"
+    "sra      %[temp3],  %[temp3],  16                 \n\t"
+    "subu     %[temp13], %[temp2],  %[temp13]          \n\t"
+    "addu     %[temp3],  %[temp1],  %[temp3]           \n\t"
+    "addu     %[temp2],  %[temp4],  %[temp6]           \n\t"
+    "subu     %[temp4],  %[temp4],  %[temp6]           \n\t"
+    "mul      %[temp6],  %[temp5],  %[kC2]             \n\t"
+    "mul      %[temp1],  %[temp7],  %[kC1]             \n\t"
+    "mul      %[temp5],  %[temp5],  %[kC1]             \n\t"
+    "mul      %[temp7],  %[temp7],  %[kC2]             \n\t"
+    "sra      %[temp6],  %[temp6],  16                 \n\t"
+    "sra      %[temp1],  %[temp1],  16                 \n\t"
+    "sra      %[temp5],  %[temp5],  16                 \n\t"
+    "sra      %[temp7],  %[temp7],  16                 \n\t"
+    "subu     %[temp1],  %[temp6],  %[temp1]           \n\t"
+    "addu     %[temp7],  %[temp5],  %[temp7]           \n\t"
+    "addu     %[temp5],  %[temp16], %[temp11]          \n\t"
+    "subu     %[temp16], %[temp16], %[temp11]          \n\t"
+    "addu     %[temp11], %[temp8],  %[temp17]          \n\t"
+    "subu     %[temp8],  %[temp8],  %[temp17]          \n\t"
+    "sra      %[temp5],  %[temp5],  3                  \n\t"
+    "sra      %[temp16], %[temp16], 3                  \n\t"
+    "sra      %[temp11], %[temp11], 3                  \n\t"
+    "sra      %[temp8],  %[temp8],  3                  \n\t"
+    "addu     %[temp17], %[temp10], %[temp15]          \n\t"
+    "subu     %[temp10], %[temp10], %[temp15]          \n\t"
+    "addu     %[temp15], %[temp12], %[temp9]           \n\t"
+    "subu     %[temp12], %[temp12], %[temp9]           \n\t"
+    "sra      %[temp17], %[temp17], 3                  \n\t"
+    "sra      %[temp10], %[temp10], 3                  \n\t"
+    "sra      %[temp15], %[temp15], 3                  \n\t"
+    "sra      %[temp12], %[temp12], 3                  \n\t"
+    "addu     %[temp9],  %[temp14], %[temp3]           \n\t"
+    "subu     %[temp14], %[temp14], %[temp3]           \n\t"
+    "addu     %[temp3],  %[temp0],  %[temp13]          \n\t"
+    "subu     %[temp0],  %[temp0],  %[temp13]          \n\t"
+    "sra      %[temp9],  %[temp9],  3                  \n\t"
+    "sra      %[temp14], %[temp14], 3                  \n\t"
+    "sra      %[temp3],  %[temp3],  3                  \n\t"
+    "sra      %[temp0],  %[temp0],  3                  \n\t"
+    "addu     %[temp13], %[temp2],  %[temp7]           \n\t"
+    "subu     %[temp2],  %[temp2],  %[temp7]           \n\t"
+    "addu     %[temp7],  %[temp4],  %[temp1]           \n\t"
+    "subu     %[temp4],  %[temp4],  %[temp1]           \n\t"
+    "sra      %[temp13], %[temp13], 3                  \n\t"
+    "sra      %[temp2],  %[temp2],  3                  \n\t"
+    "sra      %[temp7],  %[temp7],  3                  \n\t"
+    "sra      %[temp4],  %[temp4],  3                  \n\t"
+    "addiu    %[temp6],  $zero,     255                \n\t"
+    "lbu      %[temp1],  0(%[dst])                     \n\t"
+    "addu     %[temp1],  %[temp1],  %[temp5]           \n\t"
+    "sra      %[temp5],  %[temp1],  8                  \n\t"
+    "sra      %[temp18], %[temp1],  31                 \n\t"
+    "beqz     %[temp5],  1f                            \n\t"
+    "xor      %[temp1],  %[temp1],  %[temp1]           \n\t"
+    "movz     %[temp1],  %[temp6],  %[temp18]          \n\t"
+  "1:                                                  \n\t"
+    "lbu      %[temp18], 1(%[dst])                     \n\t"
+    "sb       %[temp1],  0(%[dst])                     \n\t"
+    "addu     %[temp18], %[temp18], %[temp11]          \n\t"
+    "sra      %[temp11], %[temp18], 8                  \n\t"
+    "sra      %[temp1],  %[temp18], 31                 \n\t"
+    "beqz     %[temp11], 2f                            \n\t"
+    "xor      %[temp18], %[temp18], %[temp18]          \n\t"
+    "movz     %[temp18], %[temp6],  %[temp1]           \n\t"
+  "2:                                                  \n\t"
+    "lbu      %[temp1],  2(%[dst])                     \n\t"
+    "sb       %[temp18], 1(%[dst])                     \n\t"
+    "addu     %[temp1],  %[temp1],  %[temp8]           \n\t"
+    "sra      %[temp8],  %[temp1],  8                  \n\t"
+    "sra      %[temp18], %[temp1],  31                 \n\t"
+    "beqz     %[temp8],  3f                            \n\t"
+    "xor      %[temp1],  %[temp1],  %[temp1]           \n\t"
+    "movz     %[temp1],  %[temp6],  %[temp18]          \n\t"
+  "3:                                                  \n\t"
+    "lbu      %[temp18], 3(%[dst])                     \n\t"
+    "sb       %[temp1],  2(%[dst])                     \n\t"
+    "addu     %[temp18], %[temp18], %[temp16]          \n\t"
+    "sra      %[temp16], %[temp18], 8                  \n\t"
+    "sra      %[temp1],  %[temp18], 31                 \n\t"
+    "beqz     %[temp16], 4f                            \n\t"
+    "xor      %[temp18], %[temp18], %[temp18]          \n\t"
+    "movz     %[temp18], %[temp6],  %[temp1]           \n\t"
+  "4:                                                  \n\t"
+    "sb       %[temp18], 3(%[dst])                     \n\t"
+    "lbu      %[temp5],  32(%[dst])                    \n\t"
+    "lbu      %[temp8],  33(%[dst])                    \n\t"
+    "lbu      %[temp11], 34(%[dst])                    \n\t"
+    "lbu      %[temp16], 35(%[dst])                    \n\t"
+    "addu     %[temp5],  %[temp5],  %[temp17]          \n\t"
+    "addu     %[temp8],  %[temp8],  %[temp15]          \n\t"
+    "addu     %[temp11], %[temp11], %[temp12]          \n\t"
+    "addu     %[temp16], %[temp16], %[temp10]          \n\t"
+    "sra      %[temp18], %[temp5],  8                  \n\t"
+    "sra      %[temp1],  %[temp5],  31                 \n\t"
+    "beqz     %[temp18], 5f                            \n\t"
+    "xor      %[temp5],  %[temp5],  %[temp5]           \n\t"
+    "movz     %[temp5],  %[temp6],  %[temp1]           \n\t"
+  "5:                                                  \n\t"
+    "sra      %[temp18], %[temp8],  8                  \n\t"
+    "sra      %[temp1],  %[temp8],  31                 \n\t"
+    "beqz     %[temp18], 6f                            \n\t"
+    "xor      %[temp8],  %[temp8],  %[temp8]           \n\t"
+    "movz     %[temp8],  %[temp6],  %[temp1]           \n\t"
+  "6:                                                  \n\t"
+    "sra      %[temp18], %[temp11], 8                  \n\t"
+    "sra      %[temp1],  %[temp11], 31                 \n\t"
+    "sra      %[temp17], %[temp16], 8                  \n\t"
+    "sra      %[temp15], %[temp16], 31                 \n\t"
+    "beqz     %[temp18], 7f                            \n\t"
+    "xor      %[temp11], %[temp11], %[temp11]          \n\t"
+    "movz     %[temp11], %[temp6],  %[temp1]           \n\t"
+  "7:                                                  \n\t"
+    "beqz     %[temp17], 8f                            \n\t"
+    "xor      %[temp16], %[temp16], %[temp16]          \n\t"
+    "movz     %[temp16], %[temp6],  %[temp15]          \n\t"
+  "8:                                                  \n\t"
+    "sb       %[temp5],  32(%[dst])                    \n\t"
+    "sb       %[temp8],  33(%[dst])                    \n\t"
+    "sb       %[temp11], 34(%[dst])                    \n\t"
+    "sb       %[temp16], 35(%[dst])                    \n\t"
+    "lbu      %[temp5],  64(%[dst])                    \n\t"
+    "lbu      %[temp8],  65(%[dst])                    \n\t"
+    "lbu      %[temp11], 66(%[dst])                    \n\t"
+    "lbu      %[temp16], 67(%[dst])                    \n\t"
+    "addu     %[temp5],  %[temp5],  %[temp9]           \n\t"
+    "addu     %[temp8],  %[temp8],  %[temp3]           \n\t"
+    "addu     %[temp11], %[temp11], %[temp0]           \n\t"
+    "addu     %[temp16], %[temp16], %[temp14]          \n\t"
+    "sra      %[temp18], %[temp5],  8                  \n\t"
+    "sra      %[temp1],  %[temp5],  31                 \n\t"
+    "sra      %[temp17], %[temp8],  8                  \n\t"
+    "sra      %[temp15], %[temp8],  31                 \n\t"
+    "sra      %[temp12], %[temp11], 8                  \n\t"
+    "sra      %[temp10], %[temp11], 31                 \n\t"
+    "sra      %[temp9],  %[temp16], 8                  \n\t"
+    "sra      %[temp3],  %[temp16], 31                 \n\t"
+    "beqz     %[temp18], 9f                            \n\t"
+    "xor      %[temp5],  %[temp5],  %[temp5]           \n\t"
+    "movz     %[temp5],  %[temp6],  %[temp1]           \n\t"
+  "9:                                                  \n\t"
+    "beqz     %[temp17], 10f                           \n\t"
+    "xor      %[temp8],  %[temp8],  %[temp8]           \n\t"
+    "movz     %[temp8],  %[temp6],  %[temp15]          \n\t"
+  "10:                                                 \n\t"
+    "beqz     %[temp12], 11f                           \n\t"
+    "xor      %[temp11], %[temp11], %[temp11]          \n\t"
+    "movz     %[temp11], %[temp6],  %[temp10]          \n\t"
+  "11:                                                 \n\t"
+    "beqz     %[temp9],  12f                           \n\t"
+    "xor      %[temp16], %[temp16], %[temp16]          \n\t"
+    "movz     %[temp16], %[temp6],  %[temp3]           \n\t"
+  "12:                                                 \n\t"
+    "sb       %[temp5],  64(%[dst])                    \n\t"
+    "sb       %[temp8],  65(%[dst])                    \n\t"
+    "sb       %[temp11], 66(%[dst])                    \n\t"
+    "sb       %[temp16], 67(%[dst])                    \n\t"
+    "lbu      %[temp5],  96(%[dst])                    \n\t"
+    "lbu      %[temp8],  97(%[dst])                    \n\t"
+    "lbu      %[temp11], 98(%[dst])                    \n\t"
+    "lbu      %[temp16], 99(%[dst])                    \n\t"
+    "addu     %[temp5],  %[temp5],  %[temp13]          \n\t"
+    "addu     %[temp8],  %[temp8],  %[temp7]           \n\t"
+    "addu     %[temp11], %[temp11], %[temp4]           \n\t"
+    "addu     %[temp16], %[temp16], %[temp2]           \n\t"
+    "sra      %[temp18], %[temp5],  8                  \n\t"
+    "sra      %[temp1],  %[temp5],  31                 \n\t"
+    "sra      %[temp17], %[temp8],  8                  \n\t"
+    "sra      %[temp15], %[temp8],  31                 \n\t"
+    "sra      %[temp12], %[temp11], 8                  \n\t"
+    "sra      %[temp10], %[temp11], 31                 \n\t"
+    "sra      %[temp9],  %[temp16], 8                  \n\t"
+    "sra      %[temp3],  %[temp16], 31                 \n\t"
+    "beqz     %[temp18], 13f                           \n\t"
+    "xor      %[temp5],  %[temp5],  %[temp5]           \n\t"
+    "movz     %[temp5],  %[temp6],  %[temp1]           \n\t"
+  "13:                                                 \n\t"
+    "beqz     %[temp17], 14f                           \n\t"
+    "xor      %[temp8],  %[temp8],  %[temp8]           \n\t"
+    "movz     %[temp8],  %[temp6],  %[temp15]          \n\t"
+  "14:                                                 \n\t"
+    "beqz     %[temp12], 15f                           \n\t"
+    "xor      %[temp11], %[temp11], %[temp11]          \n\t"
+    "movz     %[temp11], %[temp6],  %[temp10]          \n\t"
+  "15:                                                 \n\t"
+    "beqz     %[temp9],  16f                           \n\t"
+    "xor      %[temp16], %[temp16], %[temp16]          \n\t"
+    "movz     %[temp16], %[temp6],  %[temp3]           \n\t"
+  "16:                                                 \n\t"
+    "sb       %[temp5],  96(%[dst])                    \n\t"
+    "sb       %[temp8],  97(%[dst])                    \n\t"
+    "sb       %[temp11], 98(%[dst])                    \n\t"
+    "sb       %[temp16], 99(%[dst])                    \n\t"
+
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8),
+      [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11),
+      [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14),
+      [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17),
+      [temp18]"=&r"(temp18)
+    : [in]"r"(p_in), [kC1]"r"(kC1), [kC2]"r"(kC2), [dst]"r"(dst)
+    : "memory", "hi", "lo"
+  );
+}
+
+static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
+  TransformOne(in, dst);
+  if (do_two) {
+    TransformOne(in + 16, dst + 4);
+  }
+}
+
+#endif  // WEBP_USE_MIPS32
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8DspInitMIPS32(void);
+
+void VP8DspInitMIPS32(void) {
+#if defined(WEBP_USE_MIPS32)
+  VP8InitClipTables();
+
+  VP8Transform = TransformTwo;
+
+  VP8VFilter16 = VFilter16;
+  VP8HFilter16 = HFilter16;
+  VP8VFilter8 = VFilter8;
+  VP8HFilter8 = HFilter8;
+  VP8VFilter16i = VFilter16i;
+  VP8HFilter16i = HFilter16i;
+  VP8VFilter8i = VFilter8i;
+  VP8HFilter8i = HFilter8i;
+
+  VP8SimpleVFilter16 = SimpleVFilter16;
+  VP8SimpleHFilter16 = SimpleHFilter16;
+  VP8SimpleVFilter16i = SimpleVFilter16i;
+  VP8SimpleHFilter16i = SimpleHFilter16i;
+#endif  // WEBP_USE_MIPS32
+}

src/dsp/dec_sse2.c

@@ -26,7 +26,7 @@
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)
 
-static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
+static void Transform(const int16_t* in, uint8_t* dst, int do_two) {
   // This implementation makes use of 16-bit fixed point versions of two
   // multiply constants:
   //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@@ -246,7 +246,7 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
 
 #if defined(USE_TRANSFORM_AC3)
 #define MUL(a, b) (((a) * (b)) >> 16)
-static void TransformAC3SSE2(const int16_t* in, uint8_t* dst) {
+static void TransformAC3(const int16_t* in, uint8_t* dst) {
   static const int kC1 = 20091 + (1 << 16);
   static const int kC2 = 35468;
   const __m128i A = _mm_set1_epi16(in[0] + 4);
@@ -298,20 +298,15 @@ static void TransformAC3SSE2(const int16_t* in, uint8_t* dst) {
     _mm_subs_epu8((q), (p)),                                                   \
     _mm_subs_epu8((p), (q)))
 
-// Shift each byte of "a" by N bits while preserving by the sign bit.
-//
-// It first shifts the lower bytes of the words and then the upper bytes and
-// then merges the results together.
-#define SIGNED_SHIFT_N(a, N) {                                                 \
-  __m128i t = a;                                                               \
-  t = _mm_slli_epi16(t, 8);                                                    \
-  t = _mm_srai_epi16(t, N);                                                    \
-  t = _mm_srli_epi16(t, 8);                                                    \
-                                                                               \
-  a = _mm_srai_epi16(a, N + 8);                                                \
-  a = _mm_slli_epi16(a, 8);                                                    \
-                                                                               \
-  a = _mm_or_si128(t, a);                                                      \
+// Shift each byte of "x" by 3 bits while preserving by the sign bit.
+static WEBP_INLINE void SignedShift8b(__m128i* const x) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i signs = _mm_cmpgt_epi8(zero, *x);
+  const __m128i lo_0 = _mm_unpacklo_epi8(*x, signs);  // s8 -> s16 sign extend
+  const __m128i hi_0 = _mm_unpackhi_epi8(*x, signs);
+  const __m128i lo_1 = _mm_srai_epi16(lo_0, 3);
+  const __m128i hi_1 = _mm_srai_epi16(hi_0, 3);
+  *x = _mm_packs_epi16(lo_1, hi_1);
 }
 
 #define FLIP_SIGN_BIT2(a, b) {                                                 \
@@ -324,103 +319,123 @@ static void TransformAC3SSE2(const int16_t* in, uint8_t* dst) {
   FLIP_SIGN_BIT2(c, d);                                                        \
 }
 
-#define GET_NOTHEV(p1, p0, q0, q1, hev_thresh, not_hev) {                      \
-  const __m128i zero = _mm_setzero_si128();                                    \
-  const __m128i t_1 = MM_ABS(p1, p0);                                          \
-  const __m128i t_2 = MM_ABS(q1, q0);                                          \
-                                                                               \
-  const __m128i h = _mm_set1_epi8(hev_thresh);                                 \
-  const __m128i t_3 = _mm_subs_epu8(t_1, h);  /* abs(p1 - p0) - hev_tresh */   \
-  const __m128i t_4 = _mm_subs_epu8(t_2, h);  /* abs(q1 - q0) - hev_tresh */   \
-                                                                               \
-  not_hev = _mm_or_si128(t_3, t_4);                                            \
-  not_hev = _mm_cmpeq_epi8(not_hev, zero); /* not_hev <= t1 && not_hev <= t2 */\
+// input/output is uint8_t
+static WEBP_INLINE void GetNotHEV(const __m128i* const p1,
+                                  const __m128i* const p0,
+                                  const __m128i* const q0,
+                                  const __m128i* const q1,
+                                  int hev_thresh, __m128i* const not_hev) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t_1 = MM_ABS(*p1, *p0);
+  const __m128i t_2 = MM_ABS(*q1, *q0);
+
+  const __m128i h = _mm_set1_epi8(hev_thresh);
+  const __m128i t_3 = _mm_subs_epu8(t_1, h);  // abs(p1 - p0) - hev_tresh
+  const __m128i t_4 = _mm_subs_epu8(t_2, h);  // abs(q1 - q0) - hev_tresh
+
+  *not_hev = _mm_or_si128(t_3, t_4);
+  *not_hev = _mm_cmpeq_epi8(*not_hev, zero);  // not_hev <= t1 && not_hev <= t2
 }
 
-#define GET_BASE_DELTA(p1, p0, q0, q1, o) {                                    \
-  const __m128i qp0 = _mm_subs_epi8(q0, p0);  /* q0 - p0 */                    \
-  o = _mm_subs_epi8(p1, q1);            /* p1 - q1 */                          \
-  o = _mm_adds_epi8(o, qp0);            /* p1 - q1 + 1 * (q0 - p0) */          \
-  o = _mm_adds_epi8(o, qp0);            /* p1 - q1 + 2 * (q0 - p0) */          \
-  o = _mm_adds_epi8(o, qp0);            /* p1 - q1 + 3 * (q0 - p0) */          \
+// input pixels are int8_t
+static WEBP_INLINE void GetBaseDelta(const __m128i* const p1,
+                                     const __m128i* const p0,
+                                     const __m128i* const q0,
+                                     const __m128i* const q1,
+                                     __m128i* const delta) {
+  // beware of addition order, for saturation!
+  const __m128i p1_q1 = _mm_subs_epi8(*p1, *q1);   // p1 - q1
+  const __m128i q0_p0 = _mm_subs_epi8(*q0, *p0);   // q0 - p0
+  const __m128i s1 = _mm_adds_epi8(p1_q1, q0_p0);  // p1 - q1 + 1 * (q0 - p0)
+  const __m128i s2 = _mm_adds_epi8(q0_p0, s1);     // p1 - q1 + 2 * (q0 - p0)
+  const __m128i s3 = _mm_adds_epi8(q0_p0, s2);     // p1 - q1 + 3 * (q0 - p0)
+  *delta = s3;
 }
 
-#define DO_SIMPLE_FILTER(p0, q0, fl) {                                         \
-  const __m128i three = _mm_set1_epi8(3);                                      \
-  const __m128i four = _mm_set1_epi8(4);                                       \
-  __m128i v3 = _mm_adds_epi8(fl, three);                                       \
-  __m128i v4 = _mm_adds_epi8(fl, four);                                        \
-                                                                               \
-  /* Do +4 side */                                                             \
-  SIGNED_SHIFT_N(v4, 3);                /* v4 >> 3  */                         \
-  q0 = _mm_subs_epi8(q0, v4);           /* q0 -= v4 */                         \
-                                                                               \
-  /* Now do +3 side */                                                         \
-  SIGNED_SHIFT_N(v3, 3);                /* v3 >> 3  */                         \
-  p0 = _mm_adds_epi8(p0, v3);           /* p0 += v3 */                         \
+// input and output are int8_t
+static WEBP_INLINE void DoSimpleFilter(__m128i* const p0, __m128i* const q0,
+                                       const __m128i* const fl) {
+  const __m128i k3 = _mm_set1_epi8(3);
+  const __m128i k4 = _mm_set1_epi8(4);
+  __m128i v3 = _mm_adds_epi8(*fl, k3);
+  __m128i v4 = _mm_adds_epi8(*fl, k4);
+
+  SignedShift8b(&v4);                  // v4 >> 3
+  SignedShift8b(&v3);                  // v3 >> 3
+  *q0 = _mm_subs_epi8(*q0, v4);        // q0 -= v4
+  *p0 = _mm_adds_epi8(*p0, v3);        // p0 += v3
 }
 
 // Updates values of 2 pixels at MB edge during complex filtering.
 // Update operations:
 // q = q - delta and p = p + delta; where delta = [(a_hi >> 7), (a_lo >> 7)]
-#define UPDATE_2PIXELS(pi, qi, a_lo, a_hi) {                                   \
-  const __m128i a_lo7 = _mm_srai_epi16(a_lo, 7);                               \
-  const __m128i a_hi7 = _mm_srai_epi16(a_hi, 7);                               \
-  const __m128i delta = _mm_packs_epi16(a_lo7, a_hi7);                         \
-  pi = _mm_adds_epi8(pi, delta);                                               \
-  qi = _mm_subs_epi8(qi, delta);                                               \
+// Pixels 'pi' and 'qi' are int8_t on input, uint8_t on output (sign flip).
+static WEBP_INLINE void Update2Pixels(__m128i* const pi, __m128i* const qi,
+                                      const __m128i* const a0_lo,
+                                      const __m128i* const a0_hi) {
+  const __m128i a1_lo = _mm_srai_epi16(*a0_lo, 7);
+  const __m128i a1_hi = _mm_srai_epi16(*a0_hi, 7);
+  const __m128i delta = _mm_packs_epi16(a1_lo, a1_hi);
+  const __m128i sign_bit = _mm_set1_epi8(0x80);
+  *pi = _mm_adds_epi8(*pi, delta);
+  *qi = _mm_subs_epi8(*qi, delta);
+  FLIP_SIGN_BIT2(*pi, *qi);
 }
 
-static void NeedsFilter(const __m128i* p1, const __m128i* p0, const __m128i* q0,
-                        const __m128i* q1, int thresh, __m128i *mask) {
-  __m128i t1 = MM_ABS(*p1, *q1);        // abs(p1 - q1)
-  *mask = _mm_set1_epi8(0xFE);
-  t1 = _mm_and_si128(t1, *mask);        // set lsb of each byte to zero
-  t1 = _mm_srli_epi16(t1, 1);           // abs(p1 - q1) / 2
+// input pixels are uint8_t
+static WEBP_INLINE void NeedsFilter(const __m128i* const p1,
+                                    const __m128i* const p0,
+                                    const __m128i* const q0,
+                                    const __m128i* const q1,
+                                    int thresh, __m128i* const mask) {
+  const __m128i m_thresh = _mm_set1_epi8(thresh);
+  const __m128i t1 = MM_ABS(*p1, *q1);        // abs(p1 - q1)
+  const __m128i kFE = _mm_set1_epi8(0xFE);
+  const __m128i t2 = _mm_and_si128(t1, kFE);  // set lsb of each byte to zero
+  const __m128i t3 = _mm_srli_epi16(t2, 1);   // abs(p1 - q1) / 2
 
-  *mask = MM_ABS(*p0, *q0);             // abs(p0 - q0)
-  *mask = _mm_adds_epu8(*mask, *mask);  // abs(p0 - q0) * 2
-  *mask = _mm_adds_epu8(*mask, t1);     // abs(p0 - q0) * 2 + abs(p1 - q1) / 2
+  const __m128i t4 = MM_ABS(*p0, *q0);        // abs(p0 - q0)
+  const __m128i t5 = _mm_adds_epu8(t4, t4);   // abs(p0 - q0) * 2
+  const __m128i t6 = _mm_adds_epu8(t5, t3);   // abs(p0-q0)*2 + abs(p1-q1)/2
 
-  t1 = _mm_set1_epi8(thresh);
-  *mask = _mm_subs_epu8(*mask, t1);     // mask <= thresh
-  *mask = _mm_cmpeq_epi8(*mask, _mm_setzero_si128());
+  const __m128i t7 = _mm_subs_epu8(t6, m_thresh);  // mask <= m_thresh
+  *mask = _mm_cmpeq_epi8(t7, _mm_setzero_si128());
 }
 
 //------------------------------------------------------------------------------
 // Edge filtering functions
 
 // Applies filter on 2 pixels (p0 and q0)
-static WEBP_INLINE void DoFilter2(const __m128i* p1, __m128i* p0, __m128i* q0,
-                                  const __m128i* q1, int thresh) {
+static WEBP_INLINE void DoFilter2(__m128i* const p1, __m128i* const p0,
+                                  __m128i* const q0, __m128i* const q1,
+                                  int thresh) {
   __m128i a, mask;
   const __m128i sign_bit = _mm_set1_epi8(0x80);
+  // convert p1/q1 to int8_t (for GetBaseDelta)
   const __m128i p1s = _mm_xor_si128(*p1, sign_bit);
   const __m128i q1s = _mm_xor_si128(*q1, sign_bit);
 
   NeedsFilter(p1, p0, q0, q1, thresh, &mask);
 
-  // convert to signed values
   FLIP_SIGN_BIT2(*p0, *q0);
-
-  GET_BASE_DELTA(p1s, *p0, *q0, q1s, a);
+  GetBaseDelta(&p1s, p0, q0, &q1s, &a);
   a = _mm_and_si128(a, mask);     // mask filter values we don't care about
-  DO_SIMPLE_FILTER(*p0, *q0, a);
-
-  // unoffset
+  DoSimpleFilter(p0, q0, &a);
   FLIP_SIGN_BIT2(*p0, *q0);
 }
 
 // Applies filter on 4 pixels (p1, p0, q0 and q1)
-static WEBP_INLINE void DoFilter4(__m128i* p1, __m128i *p0,
-                                  __m128i* q0, __m128i* q1,
-                                  const __m128i* mask, int hev_thresh) {
+static WEBP_INLINE void DoFilter4(__m128i* const p1, __m128i* const p0,
+                                  __m128i* const q0, __m128i* const q1,
+                                  const __m128i* const mask, int hev_thresh) {
+  const __m128i sign_bit = _mm_set1_epi8(0x80);
+  const __m128i k64 = _mm_set1_epi8(0x40);
+  const __m128i zero = _mm_setzero_si128();
   __m128i not_hev;
   __m128i t1, t2, t3;
-  const __m128i sign_bit = _mm_set1_epi8(0x80);
 
   // compute hev mask
-  GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev);
+  GetNotHEV(p1, p0, q0, q1, hev_thresh, &not_hev);
 
   // convert to signed values
   FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
@@ -433,92 +448,83 @@ static WEBP_INLINE void DoFilter4(__m128i* p1, __m128i *p0,
   t1 = _mm_adds_epi8(t1, t2);          // hev(p1 - q1) + 3 * (q0 - p0)
   t1 = _mm_and_si128(t1, *mask);       // mask filter values we don't care about
 
-  // Do +4 side
-  t2 = _mm_set1_epi8(4);
-  t2 = _mm_adds_epi8(t1, t2);        // 3 * (q0 - p0) + (p1 - q1) + 4
-  SIGNED_SHIFT_N(t2, 3);             // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3
-  t3 = t2;                           // save t2
-  *q0 = _mm_subs_epi8(*q0, t2);      // q0 -= t2
-
-  // Now do +3 side
   t2 = _mm_set1_epi8(3);
-  t2 = _mm_adds_epi8(t1, t2);        // +3 instead of +4
-  SIGNED_SHIFT_N(t2, 3);             // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3
+  t3 = _mm_set1_epi8(4);
+  t2 = _mm_adds_epi8(t1, t2);        // 3 * (q0 - p0) + (p1 - q1) + 3
+  t3 = _mm_adds_epi8(t1, t3);        // 3 * (q0 - p0) + (p1 - q1) + 4
+  SignedShift8b(&t2);                // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3
+  SignedShift8b(&t3);                // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3
   *p0 = _mm_adds_epi8(*p0, t2);      // p0 += t2
+  *q0 = _mm_subs_epi8(*q0, t3);      // q0 -= t3
+  FLIP_SIGN_BIT2(*p0, *q0);
 
-  t2 = _mm_set1_epi8(1);
-  t3 = _mm_adds_epi8(t3, t2);
-  SIGNED_SHIFT_N(t3, 1);             // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 4
+  // this is equivalent to signed (a + 1) >> 1 calculation
+  t2 = _mm_add_epi8(t3, sign_bit);
+  t3 = _mm_avg_epu8(t2, zero);
+  t3 = _mm_sub_epi8(t3, k64);
 
   t3 = _mm_and_si128(not_hev, t3);   // if !hev
   *q1 = _mm_subs_epi8(*q1, t3);      // q1 -= t3
   *p1 = _mm_adds_epi8(*p1, t3);      // p1 += t3
-
-  // unoffset
-  FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
+  FLIP_SIGN_BIT2(*p1, *q1);
 }
 
 // Applies filter on 6 pixels (p2, p1, p0, q0, q1 and q2)
-static WEBP_INLINE void DoFilter6(__m128i *p2, __m128i* p1, __m128i *p0,
-                                  __m128i* q0, __m128i* q1, __m128i *q2,
-                                  const __m128i* mask, int hev_thresh) {
-  __m128i a, not_hev;
+static WEBP_INLINE void DoFilter6(__m128i* const p2, __m128i* const p1,
+                                  __m128i* const p0, __m128i* const q0,
+                                  __m128i* const q1, __m128i* const q2,
+                                  const __m128i* const mask, int hev_thresh) {
+  const __m128i zero = _mm_setzero_si128();
   const __m128i sign_bit = _mm_set1_epi8(0x80);
+  __m128i a, not_hev;
 
   // compute hev mask
-  GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev);
+  GetNotHEV(p1, p0, q0, q1, hev_thresh, &not_hev);
 
-  // convert to signed values
   FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
   FLIP_SIGN_BIT2(*p2, *q2);
-
-  GET_BASE_DELTA(*p1, *p0, *q0, *q1, a);
+  GetBaseDelta(p1, p0, q0, q1, &a);
 
   { // do simple filter on pixels with hev
     const __m128i m = _mm_andnot_si128(not_hev, *mask);
     const __m128i f = _mm_and_si128(a, m);
-    DO_SIMPLE_FILTER(*p0, *q0, f);
+    DoSimpleFilter(p0, q0, &f);
   }
+
   { // do strong filter on pixels with not hev
-    const __m128i zero = _mm_setzero_si128();
-    const __m128i nine = _mm_set1_epi16(0x0900);
-    const __m128i sixty_three = _mm_set1_epi16(63);
+    const __m128i k9 = _mm_set1_epi16(0x0900);
+    const __m128i k63 = _mm_set1_epi16(63);
 
     const __m128i m = _mm_and_si128(not_hev, *mask);
     const __m128i f = _mm_and_si128(a, m);
+
     const __m128i f_lo = _mm_unpacklo_epi8(zero, f);
     const __m128i f_hi = _mm_unpackhi_epi8(zero, f);
 
-    const __m128i f9_lo = _mm_mulhi_epi16(f_lo, nine);   // Filter (lo) * 9
-    const __m128i f9_hi = _mm_mulhi_epi16(f_hi, nine);   // Filter (hi) * 9
-    const __m128i f18_lo = _mm_add_epi16(f9_lo, f9_lo);  // Filter (lo) * 18
-    const __m128i f18_hi = _mm_add_epi16(f9_hi, f9_hi);  // Filter (hi) * 18
+    const __m128i f9_lo = _mm_mulhi_epi16(f_lo, k9);    // Filter (lo) * 9
+    const __m128i f9_hi = _mm_mulhi_epi16(f_hi, k9);    // Filter (hi) * 9
 
-    const __m128i a2_lo = _mm_add_epi16(f9_lo, sixty_three);  // Filter * 9 + 63
-    const __m128i a2_hi = _mm_add_epi16(f9_hi, sixty_three);  // Filter * 9 + 63
+    const __m128i a2_lo = _mm_add_epi16(f9_lo, k63);    // Filter * 9 + 63
+    const __m128i a2_hi = _mm_add_epi16(f9_hi, k63);    // Filter * 9 + 63
 
-    const __m128i a1_lo = _mm_add_epi16(f18_lo, sixty_three);  // F... * 18 + 63
-    const __m128i a1_hi = _mm_add_epi16(f18_hi, sixty_three);  // F... * 18 + 63
+    const __m128i a1_lo = _mm_add_epi16(a2_lo, f9_lo);  // Filter * 18 + 63
+    const __m128i a1_hi = _mm_add_epi16(a2_hi, f9_hi);  // Filter * 18 + 63
 
-    const __m128i a0_lo = _mm_add_epi16(f18_lo, a2_lo);  // Filter * 27 + 63
-    const __m128i a0_hi = _mm_add_epi16(f18_hi, a2_hi);  // Filter * 27 + 63
+    const __m128i a0_lo = _mm_add_epi16(a1_lo, f9_lo);  // Filter * 27 + 63
+    const __m128i a0_hi = _mm_add_epi16(a1_hi, f9_hi);  // Filter * 27 + 63
 
-    UPDATE_2PIXELS(*p2, *q2, a2_lo, a2_hi);
-    UPDATE_2PIXELS(*p1, *q1, a1_lo, a1_hi);
-    UPDATE_2PIXELS(*p0, *q0, a0_lo, a0_hi);
+    Update2Pixels(p2, q2, &a2_lo, &a2_hi);
+    Update2Pixels(p1, q1, &a1_lo, &a1_hi);
+    Update2Pixels(p0, q0, &a0_lo, &a0_hi);
   }
-
-  // unoffset
-  FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
-  FLIP_SIGN_BIT2(*p2, *q2);
 }
 
 // reads 8 rows across a vertical edge.
 //
 // TODO(somnath): Investigate _mm_shuffle* also see if it can be broken into
 // two Load4x4() to avoid code duplication.
-static WEBP_INLINE void Load8x4(const uint8_t* b, int stride,
-                                __m128i* p, __m128i* q) {
+static WEBP_INLINE void Load8x4(const uint8_t* const b, int stride,
+                                __m128i* const p, __m128i* const q) {
   __m128i t1, t2;
 
   // Load 0th, 1st, 4th and 5th rows
@@ -557,10 +563,11 @@ static WEBP_INLINE void Load8x4(const uint8_t* b, int stride,
   *q = _mm_unpackhi_epi32(t1, t2);
 }
 
-static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8,
+static WEBP_INLINE void Load16x4(const uint8_t* const r0,
+                                 const uint8_t* const r8,
                                  int stride,
-                                 __m128i* p1, __m128i* p0,
-                                 __m128i* q0, __m128i* q1) {
+                                 __m128i* const p1, __m128i* const p0,
+                                 __m128i* const q0, __m128i* const q1) {
   __m128i t1, t2;
   // Assume the pixels around the edge (|) are numbered as follows
   //                00 01 | 02 03
@@ -592,7 +599,7 @@ static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8,
   *q1 = _mm_unpackhi_epi64(t2, *q1);
 }
 
-static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) {
+static WEBP_INLINE void Store4x4(__m128i* const x, uint8_t* dst, int stride) {
   int i;
   for (i = 0; i < 4; ++i, dst += stride) {
     *((int32_t*)dst) = _mm_cvtsi128_si32(*x);
@@ -601,48 +608,51 @@ static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) {
 }
 
 // Transpose back and store
-static WEBP_INLINE void Store16x4(uint8_t* r0, uint8_t* r8, int stride,
-                                  __m128i* p1, __m128i* p0,
-                                  __m128i* q0, __m128i* q1) {
-  __m128i t1;
+static WEBP_INLINE void Store16x4(const __m128i* const p1,
+                                  const __m128i* const p0,
+                                  const __m128i* const q0,
+                                  const __m128i* const q1,
+                                  uint8_t* r0, uint8_t* r8,
+                                  int stride) {
+  __m128i t1, p1_s, p0_s, q0_s, q1_s;
 
   // p0 = 71 70 61 60 51 50 41 40 31 30 21 20 11 10 01 00
   // p1 = f1 f0 e1 e0 d1 d0 c1 c0 b1 b0 a1 a0 91 90 81 80
   t1 = *p0;
-  *p0 = _mm_unpacklo_epi8(*p1, t1);
-  *p1 = _mm_unpackhi_epi8(*p1, t1);
+  p0_s = _mm_unpacklo_epi8(*p1, t1);
+  p1_s = _mm_unpackhi_epi8(*p1, t1);
 
   // q0 = 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02
   // q1 = f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82
   t1 = *q0;
-  *q0 = _mm_unpacklo_epi8(t1, *q1);
-  *q1 = _mm_unpackhi_epi8(t1, *q1);
+  q0_s = _mm_unpacklo_epi8(t1, *q1);
+  q1_s = _mm_unpackhi_epi8(t1, *q1);
 
   // p0 = 33 32 31 30 23 22 21 20 13 12 11 10 03 02 01 00
   // q0 = 73 72 71 70 63 62 61 60 53 52 51 50 43 42 41 40
-  t1 = *p0;
-  *p0 = _mm_unpacklo_epi16(t1, *q0);
-  *q0 = _mm_unpackhi_epi16(t1, *q0);
+  t1 = p0_s;
+  p0_s = _mm_unpacklo_epi16(t1, q0_s);
+  q0_s = _mm_unpackhi_epi16(t1, q0_s);
 
   // p1 = b3 b2 b1 b0 a3 a2 a1 a0 93 92 91 90 83 82 81 80
   // q1 = f3 f2 f1 f0 e3 e2 e1 e0 d3 d2 d1 d0 c3 c2 c1 c0
-  t1 = *p1;
-  *p1 = _mm_unpacklo_epi16(t1, *q1);
-  *q1 = _mm_unpackhi_epi16(t1, *q1);
+  t1 = p1_s;
+  p1_s = _mm_unpacklo_epi16(t1, q1_s);
+  q1_s = _mm_unpackhi_epi16(t1, q1_s);
 
-  Store4x4(p0, r0, stride);
+  Store4x4(&p0_s, r0, stride);
   r0 += 4 * stride;
-  Store4x4(q0, r0, stride);
+  Store4x4(&q0_s, r0, stride);
 
-  Store4x4(p1, r8, stride);
+  Store4x4(&p1_s, r8, stride);
   r8 += 4 * stride;
-  Store4x4(q1, r8, stride);
+  Store4x4(&q1_s, r8, stride);
 }
 
 //------------------------------------------------------------------------------
 // Simple In-loop filtering (Paragraph 15.2)
 
-static void SimpleVFilter16SSE2(uint8_t* p, int stride, int thresh) {
+static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
   // Load
   __m128i p1 = _mm_loadu_si128((__m128i*)&p[-2 * stride]);
   __m128i p0 = _mm_loadu_si128((__m128i*)&p[-stride]);
@@ -653,49 +663,49 @@ static void SimpleVFilter16SSE2(uint8_t* p, int stride, int thresh) {
 
   // Store
   _mm_storeu_si128((__m128i*)&p[-stride], p0);
-  _mm_storeu_si128((__m128i*)p, q0);
+  _mm_storeu_si128((__m128i*)&p[0], q0);
 }
 
-static void SimpleHFilter16SSE2(uint8_t* p, int stride, int thresh) {
+static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
   __m128i p1, p0, q0, q1;
 
   p -= 2;  // beginning of p1
 
-  Load16x4(p, p + 8 * stride,  stride, &p1, &p0, &q0, &q1);
+  Load16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1);
   DoFilter2(&p1, &p0, &q0, &q1, thresh);
-  Store16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1);
+  Store16x4(&p1, &p0, &q0, &q1, p, p + 8 * stride, stride);
 }
 
-static void SimpleVFilter16iSSE2(uint8_t* p, int stride, int thresh) {
+static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
   int k;
   for (k = 3; k > 0; --k) {
     p += 4 * stride;
-    SimpleVFilter16SSE2(p, stride, thresh);
+    SimpleVFilter16(p, stride, thresh);
   }
 }
 
-static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) {
+static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
   int k;
   for (k = 3; k > 0; --k) {
     p += 4;
-    SimpleHFilter16SSE2(p, stride, thresh);
+    SimpleHFilter16(p, stride, thresh);
   }
 }
 
 //------------------------------------------------------------------------------
 // Complex In-loop filtering (Paragraph 15.3)
 
-#define MAX_DIFF1(p3, p2, p1, p0, m) {                                         \
-  m = MM_ABS(p3, p2);                                                          \
-  m = _mm_max_epu8(m, MM_ABS(p2, p1));                                         \
-  m = _mm_max_epu8(m, MM_ABS(p1, p0));                                         \
-}
-
-#define MAX_DIFF2(p3, p2, p1, p0, m) {                                         \
+#define MAX_DIFF1(p3, p2, p1, p0, m) do {                                      \
+  m = MM_ABS(p1, p0);                                                          \
   m = _mm_max_epu8(m, MM_ABS(p3, p2));                                         \
   m = _mm_max_epu8(m, MM_ABS(p2, p1));                                         \
+} while (0)
+
+#define MAX_DIFF2(p3, p2, p1, p0, m) do {                                      \
   m = _mm_max_epu8(m, MM_ABS(p1, p0));                                         \
-}
+  m = _mm_max_epu8(m, MM_ABS(p3, p2));                                         \
+  m = _mm_max_epu8(m, MM_ABS(p2, p1));                                         \
+} while (0)
 
 #define LOAD_H_EDGES4(p, stride, e1, e2, e3, e4) {                             \
   e1 = _mm_loadu_si128((__m128i*)&(p)[0 * stride]);                            \
@@ -704,10 +714,11 @@ static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) {
   e4 = _mm_loadu_si128((__m128i*)&(p)[3 * stride]);                            \
 }
 
-#define LOADUV_H_EDGE(p, u, v, stride) {                                       \
-  p = _mm_loadl_epi64((__m128i*)&(u)[(stride)]);                               \
-  p = _mm_unpacklo_epi64(p, _mm_loadl_epi64((__m128i*)&(v)[(stride)]));        \
-}
+#define LOADUV_H_EDGE(p, u, v, stride) do {                                    \
+  const __m128i U = _mm_loadl_epi64((__m128i*)&(u)[(stride)]);                 \
+  const __m128i V = _mm_loadl_epi64((__m128i*)&(v)[(stride)]);                 \
+  p = _mm_unpacklo_epi64(U, V);                                                \
+} while (0)
 
 #define LOADUV_H_EDGES4(u, v, stride, e1, e2, e3, e4) {                        \
   LOADUV_H_EDGE(e1, u, v, 0 * stride);                                         \
@@ -722,18 +733,23 @@ static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) {
   _mm_storel_epi64((__m128i*)&v[(stride)], p);                                 \
 }
 
-#define COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask) {               \
-  __m128i fl_yes;                                                              \
-  const __m128i it = _mm_set1_epi8(ithresh);                                   \
-  mask = _mm_subs_epu8(mask, it);                                              \
-  mask = _mm_cmpeq_epi8(mask, _mm_setzero_si128());                            \
-  NeedsFilter(&p1, &p0, &q0, &q1, thresh, &fl_yes);                            \
-  mask = _mm_and_si128(mask, fl_yes);                                          \
+static WEBP_INLINE void ComplexMask(const __m128i* const p1,
+                                    const __m128i* const p0,
+                                    const __m128i* const q0,
+                                    const __m128i* const q1,
+                                    int thresh, int ithresh,
+                                    __m128i* const mask) {
+  const __m128i it = _mm_set1_epi8(ithresh);
+  const __m128i diff = _mm_subs_epu8(*mask, it);
+  const __m128i thresh_mask = _mm_cmpeq_epi8(diff, _mm_setzero_si128());
+  __m128i filter_mask;
+  NeedsFilter(p1, p0, q0, q1, thresh, &filter_mask);
+  *mask = _mm_and_si128(thresh_mask, filter_mask);
 }
 
 // on macroblock edges
-static void VFilter16SSE2(uint8_t* p, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void VFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
   __m128i t1;
   __m128i mask;
   __m128i p2, p1, p0, q0, q1, q2;
@@ -746,20 +762,20 @@ static void VFilter16SSE2(uint8_t* p, int stride,
   LOAD_H_EDGES4(p, stride, q0, q1, q2, t1);
   MAX_DIFF2(t1, q2, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
 
   // Store
   _mm_storeu_si128((__m128i*)&p[-3 * stride], p2);
   _mm_storeu_si128((__m128i*)&p[-2 * stride], p1);
   _mm_storeu_si128((__m128i*)&p[-1 * stride], p0);
-  _mm_storeu_si128((__m128i*)&p[0 * stride], q0);
-  _mm_storeu_si128((__m128i*)&p[1 * stride], q1);
-  _mm_storeu_si128((__m128i*)&p[2 * stride], q2);
+  _mm_storeu_si128((__m128i*)&p[+0 * stride], q0);
+  _mm_storeu_si128((__m128i*)&p[+1 * stride], q1);
+  _mm_storeu_si128((__m128i*)&p[+2 * stride], q2);
 }
 
-static void HFilter16SSE2(uint8_t* p, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void HFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i p3, p2, p1, p0, q0, q1, q2, q3;
 
@@ -770,71 +786,78 @@ static void HFilter16SSE2(uint8_t* p, int stride,
   Load16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3);  // q0, q1, q2, q3
   MAX_DIFF2(q3, q2, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
 
-  Store16x4(b, b + 8 * stride, stride, &p3, &p2, &p1, &p0);
-  Store16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3);
+  Store16x4(&p3, &p2, &p1, &p0, b, b + 8 * stride, stride);
+  Store16x4(&q0, &q1, &q2, &q3, p, p + 8 * stride, stride);
 }
 
 // on three inner edges
-static void VFilter16iSSE2(uint8_t* p, int stride,
-                           int thresh, int ithresh, int hev_thresh) {
+static void VFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
   int k;
-  __m128i mask;
-  __m128i t1, t2, p1, p0, q0, q1;
+  __m128i p3, p2, p1, p0;   // loop invariants
+
+  LOAD_H_EDGES4(p, stride, p3, p2, p1, p0);  // prologue
 
   for (k = 3; k > 0; --k) {
-    // Load p3, p2, p1, p0
-    LOAD_H_EDGES4(p, stride, t2, t1, p1, p0);
-    MAX_DIFF1(t2, t1, p1, p0, mask);
-
+    __m128i mask, tmp1, tmp2;
+    uint8_t* const b = p + 2 * stride;   // beginning of p1
     p += 4 * stride;
 
-    // Load q0, q1, q2, q3
-    LOAD_H_EDGES4(p, stride, q0, q1, t1, t2);
-    MAX_DIFF2(t2, t1, q1, q0, mask);
+    MAX_DIFF1(p3, p2, p1, p0, mask);   // compute partial mask
+    LOAD_H_EDGES4(p, stride, p3, p2, tmp1, tmp2);
+    MAX_DIFF2(p3, p2, tmp1, tmp2, mask);
 
-    COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
-    DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
+    // p3 and p2 are not just temporary variables here: they will be
+    // re-used for next span. And q2/q3 will become p1/p0 accordingly.
+    ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask);
+    DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh);
 
     // Store
-    _mm_storeu_si128((__m128i*)&p[-2 * stride], p1);
-    _mm_storeu_si128((__m128i*)&p[-1 * stride], p0);
-    _mm_storeu_si128((__m128i*)&p[0 * stride], q0);
-    _mm_storeu_si128((__m128i*)&p[1 * stride], q1);
+    _mm_storeu_si128((__m128i*)&b[0 * stride], p1);
+    _mm_storeu_si128((__m128i*)&b[1 * stride], p0);
+    _mm_storeu_si128((__m128i*)&b[2 * stride], p3);
+    _mm_storeu_si128((__m128i*)&b[3 * stride], p2);
+
+    // rotate samples
+    p1 = tmp1;
+    p0 = tmp2;
   }
 }
 
-static void HFilter16iSSE2(uint8_t* p, int stride,
-                           int thresh, int ithresh, int hev_thresh) {
+static void HFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
   int k;
-  uint8_t* b;
-  __m128i mask;
-  __m128i t1, t2, p1, p0, q0, q1;
+  __m128i p3, p2, p1, p0;   // loop invariants
+
+  Load16x4(p, p + 8 * stride, stride, &p3, &p2, &p1, &p0);  // prologue
 
   for (k = 3; k > 0; --k) {
-    b = p;
-    Load16x4(b, b + 8 * stride, stride, &t2, &t1, &p1, &p0);  // p3, p2, p1, p0
-    MAX_DIFF1(t2, t1, p1, p0, mask);
+    __m128i mask, tmp1, tmp2;
+    uint8_t* const b = p + 2;   // beginning of p1
 
-    b += 4;  // beginning of q0
-    Load16x4(b, b + 8 * stride, stride, &q0, &q1, &t1, &t2);  // q0, q1, q2, q3
-    MAX_DIFF2(t2, t1, q1, q0, mask);
+    p += 4;  // beginning of q0 (and next span)
 
-    COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
-    DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
+    MAX_DIFF1(p3, p2, p1, p0, mask);   // compute partial mask
+    Load16x4(p, p + 8 * stride, stride, &p3, &p2, &tmp1, &tmp2);
+    MAX_DIFF2(p3, p2, tmp1, tmp2, mask);
 
-    b -= 2;  // beginning of p1
-    Store16x4(b, b + 8 * stride, stride, &p1, &p0, &q0, &q1);
+    ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask);
+    DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh);
 
-    p += 4;
+    Store16x4(&p1, &p0, &p3, &p2, b, b + 8 * stride, stride);
+
+    // rotate samples
+    p1 = tmp1;
+    p0 = tmp2;
   }
 }
 
 // 8-pixels wide variant, for chroma filtering
-static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
-                         int thresh, int ithresh, int hev_thresh) {
+static void VFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, p2, p1, p0, q0, q1, q2;
 
@@ -846,7 +869,7 @@ static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
   LOADUV_H_EDGES4(u, v, stride, q0, q1, q2, t1);
   MAX_DIFF2(t1, q2, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
 
   // Store
@@ -858,8 +881,8 @@ static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
   STOREUV(q2, u, v, 2 * stride);
 }
 
-static void HFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
-                         int thresh, int ithresh, int hev_thresh) {
+static void HFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i p3, p2, p1, p0, q0, q1, q2, q3;
 
@@ -871,15 +894,15 @@ static void HFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
   Load16x4(u, v, stride, &q0, &q1, &q2, &q3);    // q0, q1, q2, q3
   MAX_DIFF2(q3, q2, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
 
-  Store16x4(tu, tv, stride, &p3, &p2, &p1, &p0);
-  Store16x4(u, v, stride, &q0, &q1, &q2, &q3);
+  Store16x4(&p3, &p2, &p1, &p0, tu, tv, stride);
+  Store16x4(&q0, &q1, &q2, &q3, u, v, stride);
 }
 
-static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, t2, p1, p0, q0, q1;
 
@@ -894,7 +917,7 @@ static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
   LOADUV_H_EDGES4(u, v, stride, q0, q1, t1, t2);
   MAX_DIFF2(t2, t1, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
 
   // Store
@@ -904,8 +927,8 @@ static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
   STOREUV(q1, u, v, 1 * stride);
 }
 
-static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, t2, p1, p0, q0, q1;
   Load16x4(u, v, stride, &t2, &t1, &p1, &p0);   // p3, p2, p1, p0
@@ -916,12 +939,12 @@ static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
   Load16x4(u, v, stride, &q0, &q1, &t1, &t2);  // q0, q1, q2, q3
   MAX_DIFF2(t2, t1, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
 
   u -= 2;  // beginning of p1
   v -= 2;
-  Store16x4(u, v, stride, &p1, &p0, &q0, &q1);
+  Store16x4(&p1, &p0, &q0, &q1, u, v, stride);
 }
 
 #endif   // WEBP_USE_SSE2
@@ -933,24 +956,23 @@ extern void VP8DspInitSSE2(void);
 
 void VP8DspInitSSE2(void) {
 #if defined(WEBP_USE_SSE2)
-  VP8Transform = TransformSSE2;
+  VP8Transform = Transform;
 #if defined(USE_TRANSFORM_AC3)
-  VP8TransformAC3 = TransformAC3SSE2;
+  VP8TransformAC3 = TransformAC3;
 #endif
 
-  VP8VFilter16 = VFilter16SSE2;
-  VP8HFilter16 = HFilter16SSE2;
-  VP8VFilter8 = VFilter8SSE2;
-  VP8HFilter8 = HFilter8SSE2;
-  VP8VFilter16i = VFilter16iSSE2;
-  VP8HFilter16i = HFilter16iSSE2;
-  VP8VFilter8i = VFilter8iSSE2;
-  VP8HFilter8i = HFilter8iSSE2;
+  VP8VFilter16 = VFilter16;
+  VP8HFilter16 = HFilter16;
+  VP8VFilter8 = VFilter8;
+  VP8HFilter8 = HFilter8;
+  VP8VFilter16i = VFilter16i;
+  VP8HFilter16i = HFilter16i;
+  VP8VFilter8i = VFilter8i;
+  VP8HFilter8i = HFilter8i;
 
-  VP8SimpleVFilter16 = SimpleVFilter16SSE2;
-  VP8SimpleHFilter16 = SimpleHFilter16SSE2;
-  VP8SimpleVFilter16i = SimpleVFilter16iSSE2;
-  VP8SimpleHFilter16i = SimpleHFilter16iSSE2;
+  VP8SimpleVFilter16 = SimpleVFilter16;
+  VP8SimpleHFilter16 = SimpleHFilter16;
+  VP8SimpleVFilter16i = SimpleVFilter16i;
+  VP8SimpleHFilter16i = SimpleHFilter16i;
 #endif   // WEBP_USE_SSE2
 }
-

src/dsp/dsp.h

@@ -14,6 +14,10 @@
 #ifndef WEBP_DSP_DSP_H_
 #define WEBP_DSP_DSP_H_
 
+#ifdef HAVE_CONFIG_H
+#include "../webp/config.h"
+#endif
+
 #include "../webp/types.h"
 
 #ifdef __cplusplus
@@ -23,27 +27,63 @@ extern "C" {
 //------------------------------------------------------------------------------
 // CPU detection
 
+#if defined(__GNUC__)
+# define LOCAL_GCC_VERSION ((__GNUC__ << 8) | __GNUC_MINOR__)
+# define LOCAL_GCC_PREREQ(maj, min) \
+    (LOCAL_GCC_VERSION >= (((maj) << 8) | (min)))
+#else
+# define LOCAL_GCC_VERSION 0
+# define LOCAL_GCC_PREREQ(maj, min) 0
+#endif
+
+#ifdef __clang__
+# define LOCAL_CLANG_VERSION ((__clang_major__ << 8) | __clang_minor__)
+# define LOCAL_CLANG_PREREQ(maj, min) \
+    (LOCAL_CLANG_VERSION >= (((maj) << 8) | (min)))
+#else
+# define LOCAL_CLANG_VERSION 0
+# define LOCAL_CLANG_PREREQ(maj, min) 0
+#endif  // __clang__
+
 #if defined(_MSC_VER) && _MSC_VER > 1310 && \
     (defined(_M_X64) || defined(_M_IX86))
 #define WEBP_MSC_SSE2  // Visual C++ SSE2 targets
 #endif
 
-#if defined(__SSE2__) || defined(WEBP_MSC_SSE2)
+// WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp
+// files without intrinsics, allowing the corresponding Init() to be called.
+// Files containing intrinsics will need to be built targeting the instruction
+// set so should succeed on one of the earlier tests.
+#if defined(__SSE2__) || defined(WEBP_MSC_SSE2) || defined(WEBP_HAVE_SSE2)
 #define WEBP_USE_SSE2
 #endif
 
+#if defined(__AVX2__) || defined(WEBP_HAVE_AVX2)
+#define WEBP_USE_AVX2
+#endif
+
 #if defined(__ANDROID__) && defined(__ARM_ARCH_7A__)
 #define WEBP_ANDROID_NEON  // Android targets that might support NEON
 #endif
 
-#if defined(__ARM_NEON__) || defined(WEBP_ANDROID_NEON)
+#if defined(__ARM_NEON__) || defined(WEBP_ANDROID_NEON) || defined(__aarch64__)
 #define WEBP_USE_NEON
 #endif
 
+#if defined(__mips__) && !defined(__mips64) && (__mips_isa_rev < 6)
+#define WEBP_USE_MIPS32
+#if (__mips_isa_rev >= 2)
+#define WEBP_USE_MIPS32_R2
+#endif
+#endif
+
 typedef enum {
   kSSE2,
   kSSE3,
-  kNEON
+  kAVX,
+  kAVX2,
+  kNEON,
+  kMIPS32
 } CPUFeature;
 // returns true if the CPU supports the feature.
 typedef int (*VP8CPUInfo)(CPUFeature feature);
@@ -61,7 +101,6 @@ typedef void (*VP8Fdct)(const uint8_t* src, const uint8_t* ref, int16_t* out);
 typedef void (*VP8WHT)(const int16_t* in, int16_t* out);
 extern VP8Idct VP8ITransform;
 extern VP8Fdct VP8FTransform;
-extern VP8WHT VP8ITransformWHT;
 extern VP8WHT VP8FTransformWHT;
 // Predictions
 // *dst is the destination block. *top and *left can be NULL.
@@ -83,7 +122,7 @@ extern VP8BlockCopy VP8Copy4x4;
 // Quantization
 struct VP8Matrix;   // forward declaration
 typedef int (*VP8QuantizeBlock)(int16_t in[16], int16_t out[16],
-                                int n, const struct VP8Matrix* const mtx);
+                                const struct VP8Matrix* const mtx);
 extern VP8QuantizeBlock VP8EncQuantizeBlock;
 
 // specific to 2nd transform:
@@ -121,6 +160,13 @@ extern const VP8PredFunc VP8PredLuma16[/* NUM_B_DC_MODES */];
 extern const VP8PredFunc VP8PredChroma8[/* NUM_B_DC_MODES */];
 extern const VP8PredFunc VP8PredLuma4[/* NUM_BMODES */];
 
+// clipping tables (for filtering)
+extern const int8_t* const VP8ksclip1;  // clips [-1020, 1020] to [-128, 127]
+extern const int8_t* const VP8ksclip2;  // clips [-112, 112] to [-16, 15]
+extern const uint8_t* const VP8kclip1;  // clips [-255,511] to [0,255]
+extern const uint8_t* const VP8kabs0;   // abs(x) for x in [-255,255]
+void VP8InitClipTables(void);           // must be called first
+
 // simple filter (only for luma)
 typedef void (*VP8SimpleFilterFunc)(uint8_t* p, int stride, int thresh);
 extern VP8SimpleFilterFunc VP8SimpleVFilter16;
@@ -166,21 +212,20 @@ typedef void (*WebPUpsampleLinePairFunc)(
 // Fancy upsampling functions to convert YUV to RGB(A) modes
 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 
-// Initializes SSE2 version of the fancy upsamplers.
-void WebPInitUpsamplersSSE2(void);
-
-// NEON version
-void WebPInitUpsamplersNEON(void);
-
 #endif    // FANCY_UPSAMPLING
 
-// Point-sampling methods.
-typedef void (*WebPSampleLinePairFunc)(
-    const uint8_t* top_y, const uint8_t* bottom_y,
-    const uint8_t* u, const uint8_t* v,
-    uint8_t* top_dst, uint8_t* bottom_dst, int len);
+// Per-row point-sampling methods.
+typedef void (*WebPSamplerRowFunc)(const uint8_t* y,
+                                   const uint8_t* u, const uint8_t* v,
+                                   uint8_t* dst, int len);
+// Generic function to apply 'WebPSamplerRowFunc' to the whole plane:
+void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
+                             const uint8_t* u, const uint8_t* v, int uv_stride,
+                             uint8_t* dst, int dst_stride,
+                             int width, int height, WebPSamplerRowFunc func);
 
-extern const WebPSampleLinePairFunc WebPSamplers[/* MODE_LAST */];
+// Sampling functions to convert rows of YUV to RGB(A)
+extern WebPSamplerRowFunc WebPSamplers[/* MODE_LAST */];
 
 // General function for converting two lines of ARGB or RGBA.
 // 'alpha_is_last' should be true if 0xff000000 is stored in memory as
@@ -194,11 +239,14 @@ typedef void (*WebPYUV444Converter)(const uint8_t* y,
 
 extern const WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
 
-// Main function to be called
+// Must be called before using the WebPUpsamplers[] (and for premultiplied
+// colorspaces like rgbA, rgbA4444, etc)
 void WebPInitUpsamplers(void);
+// Must be called before using WebPSamplers[]
+void WebPInitSamplers(void);
 
 //------------------------------------------------------------------------------
-// Pre-multiply planes with alpha values
+// Utilities for processing transparent channel.
 
 // Apply alpha pre-multiply on an rgba, bgra or argb plane of size w * h.
 // alpha_first should be 0 for argb, 1 for rgba or bgra (where alpha is last).
@@ -209,13 +257,34 @@ extern void (*WebPApplyAlphaMultiply)(
 extern void (*WebPApplyAlphaMultiply4444)(
     uint8_t* rgba4444, int w, int h, int stride);
 
+// Extract the alpha values from 32b values in argb[] and pack them into alpha[]
+// (this is the opposite of WebPDispatchAlpha).
+// Returns true if there's only trivial 0xff alpha values.
+extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride,
+                               int width, int height,
+                               uint8_t* alpha, int alpha_stride);
+
+// Pre-Multiply operation transforms x into x * A / 255  (where x=Y,R,G or B).
+// Un-Multiply operation transforms x into x * 255 / A.
+
+// Pre-Multiply or Un-Multiply (if 'inverse' is true) argb values in a row.
+extern void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse);
+
+// Same a WebPMultARGBRow(), but for several rows.
+void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows,
+                      int inverse);
+
+// Same for a row of single values, with side alpha values.
+extern void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha,
+                           int width, int inverse);
+
+// Same a WebPMultRow(), but for several 'num_rows' rows.
+void WebPMultRows(uint8_t* ptr, int stride,
+                  const uint8_t* alpha, int alpha_stride,
+                  int width, int num_rows, int inverse);
+
 // To be called first before using the above.
-void WebPInitPremultiply(void);
-
-void WebPInitPremultiplySSE2(void);   // should not be called directly.
-void WebPInitPremultiplyNEON(void);
-
-//------------------------------------------------------------------------------
+void WebPInitAlphaProcessing(void);
 
 #ifdef __cplusplus
 }    // extern "C"

src/dsp/enc.c

@@ -159,33 +159,6 @@ static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
   }
 }
 
-static void ITransformWHT(const int16_t* in, int16_t* out) {
-  int tmp[16];
-  int i;
-  for (i = 0; i < 4; ++i) {
-    const int a0 = in[0 + i] + in[12 + i];
-    const int a1 = in[4 + i] + in[ 8 + i];
-    const int a2 = in[4 + i] - in[ 8 + i];
-    const int a3 = in[0 + i] - in[12 + i];
-    tmp[0  + i] = a0 + a1;
-    tmp[8  + i] = a0 - a1;
-    tmp[4  + i] = a3 + a2;
-    tmp[12 + i] = a3 - a2;
-  }
-  for (i = 0; i < 4; ++i) {
-    const int dc = tmp[0 + i * 4] + 3;    // w/ rounder
-    const int a0 = dc             + tmp[3 + i * 4];
-    const int a1 = tmp[1 + i * 4] + tmp[2 + i * 4];
-    const int a2 = tmp[1 + i * 4] - tmp[2 + i * 4];
-    const int a3 = dc             - tmp[3 + i * 4];
-    out[ 0] = (a0 + a1) >> 3;
-    out[16] = (a3 + a2) >> 3;
-    out[32] = (a0 - a1) >> 3;
-    out[48] = (a3 - a2) >> 3;
-    out += 64;
-  }
-}
-
 static void FTransformWHT(const int16_t* in, int16_t* out) {
   // input is 12b signed
   int32_t tmp[16];
@@ -627,21 +600,23 @@ static const uint8_t kZigzag[16] = {
 
 // Simple quantization
 static int QuantizeBlock(int16_t in[16], int16_t out[16],
-                         int n, const VP8Matrix* const mtx) {
+                         const VP8Matrix* const mtx) {
   int last = -1;
-  for (; n < 16; ++n) {
+  int n;
+  for (n = 0; n < 16; ++n) {
     const int j = kZigzag[n];
     const int sign = (in[j] < 0);
-    const int coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
+    const uint32_t coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
     if (coeff > mtx->zthresh_[j]) {
-      const int Q = mtx->q_[j];
-      const int iQ = mtx->iq_[j];
-      const int B = mtx->bias_[j];
-      out[n] = QUANTDIV(coeff, iQ, B);
-      if (out[n] > MAX_LEVEL) out[n] = MAX_LEVEL;
-      if (sign) out[n] = -out[n];
-      in[j] = out[n] * Q;
-      if (out[n]) last = n;
+      const uint32_t Q = mtx->q_[j];
+      const uint32_t iQ = mtx->iq_[j];
+      const uint32_t B = mtx->bias_[j];
+      int level = QUANTDIV(coeff, iQ, B);
+      if (level > MAX_LEVEL) level = MAX_LEVEL;
+      if (sign) level = -level;
+      in[j] = level * Q;
+      out[n] = level;
+      if (level) last = n;
     } else {
       out[n] = 0;
       in[j] = 0;
@@ -656,17 +631,18 @@ static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
   for (n = 0; n < 16; ++n) {
     const int j = kZigzag[n];
     const int sign = (in[j] < 0);
-    const int coeff = sign ? -in[j] : in[j];
+    const uint32_t coeff = sign ? -in[j] : in[j];
     assert(mtx->sharpen_[j] == 0);
     if (coeff > mtx->zthresh_[j]) {
-      const int Q = mtx->q_[j];
-      const int iQ = mtx->iq_[j];
-      const int B = mtx->bias_[j];
-      out[n] = QUANTDIV(coeff, iQ, B);
-      if (out[n] > MAX_LEVEL) out[n] = MAX_LEVEL;
-      if (sign) out[n] = -out[n];
-      in[j] = out[n] * Q;
-      if (out[n]) last = n;
+      const uint32_t Q = mtx->q_[j];
+      const uint32_t iQ = mtx->iq_[j];
+      const uint32_t B = mtx->bias_[j];
+      int level = QUANTDIV(coeff, iQ, B);
+      if (level > MAX_LEVEL) level = MAX_LEVEL;
+      if (sign) level = -level;
+      in[j] = level * Q;
+      out[n] = level;
+      if (level) last = n;
     } else {
       out[n] = 0;
       in[j] = 0;
@@ -697,7 +673,6 @@ static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); }
 VP8CHisto VP8CollectHistogram;
 VP8Idct VP8ITransform;
 VP8Fdct VP8FTransform;
-VP8WHT VP8ITransformWHT;
 VP8WHT VP8FTransformWHT;
 VP8Intra4Preds VP8EncPredLuma4;
 VP8IntraPreds VP8EncPredLuma16;
@@ -713,16 +688,23 @@ VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
 VP8BlockCopy VP8Copy4x4;
 
 extern void VP8EncDspInitSSE2(void);
+extern void VP8EncDspInitAVX2(void);
 extern void VP8EncDspInitNEON(void);
+extern void VP8EncDspInitMIPS32(void);
+
+static volatile VP8CPUInfo enc_last_cpuinfo_used =
+    (VP8CPUInfo)&enc_last_cpuinfo_used;
 
 void VP8EncDspInit(void) {
+  if (enc_last_cpuinfo_used == VP8GetCPUInfo) return;
+
+  VP8DspInit();  // common inverse transforms
   InitTables();
 
   // default C implementations
   VP8CollectHistogram = CollectHistogram;
   VP8ITransform = ITransform;
   VP8FTransform = FTransform;
-  VP8ITransformWHT = ITransformWHT;
   VP8FTransformWHT = FTransformWHT;
   VP8EncPredLuma4 = Intra4Preds;
   VP8EncPredLuma16 = Intra16Preds;
@@ -738,16 +720,28 @@ void VP8EncDspInit(void) {
   VP8Copy4x4 = Copy4x4;
 
   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
-  if (VP8GetCPUInfo) {
+  if (VP8GetCPUInfo != NULL) {
 #if defined(WEBP_USE_SSE2)
     if (VP8GetCPUInfo(kSSE2)) {
       VP8EncDspInitSSE2();
     }
-#elif defined(WEBP_USE_NEON)
+#endif
+#if defined(WEBP_USE_AVX2)
+    if (VP8GetCPUInfo(kAVX2)) {
+      VP8EncDspInitAVX2();
+    }
+#endif
+#if defined(WEBP_USE_NEON)
     if (VP8GetCPUInfo(kNEON)) {
       VP8EncDspInitNEON();
     }
+#endif
+#if defined(WEBP_USE_MIPS32)
+    if (VP8GetCPUInfo(kMIPS32)) {
+      VP8EncDspInitMIPS32();
+    }
 #endif
   }
+  enc_last_cpuinfo_used = VP8GetCPUInfo;
 }
 

src/dsp/enc_avx2.c

@@ -1,4 +1,4 @@
-// Copyright 2011 Google Inc. All Rights Reserved.
+// Copyright 2014 Google Inc. All Rights Reserved.
 //
 // Use of this source code is governed by a BSD-style license
 // that can be found in the COPYING file in the root of the source
@@ -7,24 +7,18 @@
 // be found in the AUTHORS file in the root of the source tree.
 // -----------------------------------------------------------------------------
 //
-// Enhancement layer (for YUV444/422)
-//
-// Author: Skal (pascal.massimino@gmail.com)
+// AVX2 version of speed-critical encoding functions.
 
-#include <assert.h>
-#include <stdlib.h>
+#include "./dsp.h"
 
-#include "./vp8i.h"
+#if defined(WEBP_USE_AVX2)
+
+#endif  // WEBP_USE_AVX2
 
 //------------------------------------------------------------------------------
+// Entry point
 
-int VP8DecodeLayer(VP8Decoder* const dec) {
-  assert(dec);
-  assert(dec->layer_data_size_ > 0);
-  (void)dec;
+extern void VP8EncDspInitAVX2(void);
 
-  // TODO: handle enhancement layer here.
-
-  return 1;
+void VP8EncDspInitAVX2(void) {
 }
-

src/dsp/enc_mips32.c

@@ -0,0 +1,776 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// MIPS version of speed-critical encoding functions.
+//
+// Author(s): Djordje Pesut    (djordje.pesut@imgtec.com)
+//            Jovan Zelincevic (jovan.zelincevic@imgtec.com)
+//            Slobodan Prijic  (slobodan.prijic@imgtec.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_MIPS32)
+
+#include "../enc/vp8enci.h"
+#include "../enc/cost.h"
+
+#if defined(__GNUC__) && defined(__ANDROID__) && LOCAL_GCC_VERSION == 0x409
+#define WORK_AROUND_GCC
+#endif
+
+static const int kC1 = 20091 + (1 << 16);
+static const int kC2 = 35468;
+
+// macro for one vertical pass in ITransformOne
+// MUL macro inlined
+// temp0..temp15 holds tmp[0]..tmp[15]
+// A..D - offsets in bytes to load from in buffer
+// TEMP0..TEMP3 - registers for corresponding tmp elements
+// TEMP4..TEMP5 - temporary registers
+#define VERTICAL_PASS(A, B, C, D, TEMP4, TEMP0, TEMP1, TEMP2, TEMP3)        \
+  "lh      %[temp16],      "#A"(%[temp20])                 \n\t"            \
+  "lh      %[temp18],      "#B"(%[temp20])                 \n\t"            \
+  "lh      %[temp17],      "#C"(%[temp20])                 \n\t"            \
+  "lh      %[temp19],      "#D"(%[temp20])                 \n\t"            \
+  "addu    %["#TEMP4"],    %[temp16],      %[temp18]       \n\t"            \
+  "subu    %[temp16],      %[temp16],      %[temp18]       \n\t"            \
+  "mul     %["#TEMP0"],    %[temp17],      %[kC2]          \n\t"            \
+  "mul     %[temp18],      %[temp19],      %[kC1]          \n\t"            \
+  "mul     %[temp17],      %[temp17],      %[kC1]          \n\t"            \
+  "mul     %[temp19],      %[temp19],      %[kC2]          \n\t"            \
+  "sra     %["#TEMP0"],    %["#TEMP0"],    16              \n\n"            \
+  "sra     %[temp18],      %[temp18],      16              \n\n"            \
+  "sra     %[temp17],      %[temp17],      16              \n\n"            \
+  "sra     %[temp19],      %[temp19],      16              \n\n"            \
+  "subu    %["#TEMP2"],    %["#TEMP0"],    %[temp18]       \n\t"            \
+  "addu    %["#TEMP3"],    %[temp17],      %[temp19]       \n\t"            \
+  "addu    %["#TEMP0"],    %["#TEMP4"],    %["#TEMP3"]     \n\t"            \
+  "addu    %["#TEMP1"],    %[temp16],      %["#TEMP2"]     \n\t"            \
+  "subu    %["#TEMP2"],    %[temp16],      %["#TEMP2"]     \n\t"            \
+  "subu    %["#TEMP3"],    %["#TEMP4"],    %["#TEMP3"]     \n\t"
+
+// macro for one horizontal pass in ITransformOne
+// MUL and STORE macros inlined
+// a = clip_8b(a) is replaced with: a = max(a, 0); a = min(a, 255)
+// temp0..temp15 holds tmp[0]..tmp[15]
+// A..D - offsets in bytes to load from ref and store to dst buffer
+// TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements
+#define HORIZONTAL_PASS(A, B, C, D, TEMP0, TEMP4, TEMP8, TEMP12)            \
+  "addiu   %["#TEMP0"],    %["#TEMP0"],    4               \n\t"            \
+  "addu    %[temp16],      %["#TEMP0"],    %["#TEMP8"]     \n\t"            \
+  "subu    %[temp17],      %["#TEMP0"],    %["#TEMP8"]     \n\t"            \
+  "mul     %["#TEMP0"],    %["#TEMP4"],    %[kC2]          \n\t"            \
+  "mul     %["#TEMP8"],    %["#TEMP12"],   %[kC1]          \n\t"            \
+  "mul     %["#TEMP4"],    %["#TEMP4"],    %[kC1]          \n\t"            \
+  "mul     %["#TEMP12"],   %["#TEMP12"],   %[kC2]          \n\t"            \
+  "sra     %["#TEMP0"],    %["#TEMP0"],    16              \n\t"            \
+  "sra     %["#TEMP8"],    %["#TEMP8"],    16              \n\t"            \
+  "sra     %["#TEMP4"],    %["#TEMP4"],    16              \n\t"            \
+  "sra     %["#TEMP12"],   %["#TEMP12"],   16              \n\t"            \
+  "subu    %[temp18],      %["#TEMP0"],    %["#TEMP8"]     \n\t"            \
+  "addu    %[temp19],      %["#TEMP4"],    %["#TEMP12"]    \n\t"            \
+  "addu    %["#TEMP0"],    %[temp16],      %[temp19]       \n\t"            \
+  "addu    %["#TEMP4"],    %[temp17],      %[temp18]       \n\t"            \
+  "subu    %["#TEMP8"],    %[temp17],      %[temp18]       \n\t"            \
+  "subu    %["#TEMP12"],   %[temp16],      %[temp19]       \n\t"            \
+  "lw      %[temp20],      0(%[args])                      \n\t"            \
+  "sra     %["#TEMP0"],    %["#TEMP0"],    3               \n\t"            \
+  "sra     %["#TEMP4"],    %["#TEMP4"],    3               \n\t"            \
+  "sra     %["#TEMP8"],    %["#TEMP8"],    3               \n\t"            \
+  "sra     %["#TEMP12"],   %["#TEMP12"],   3               \n\t"            \
+  "lbu     %[temp16],      "#A"(%[temp20])                 \n\t"            \
+  "lbu     %[temp17],      "#B"(%[temp20])                 \n\t"            \
+  "lbu     %[temp18],      "#C"(%[temp20])                 \n\t"            \
+  "lbu     %[temp19],      "#D"(%[temp20])                 \n\t"            \
+  "addu    %["#TEMP0"],    %[temp16],      %["#TEMP0"]     \n\t"            \
+  "addu    %["#TEMP4"],    %[temp17],      %["#TEMP4"]     \n\t"            \
+  "addu    %["#TEMP8"],    %[temp18],      %["#TEMP8"]     \n\t"            \
+  "addu    %["#TEMP12"],   %[temp19],      %["#TEMP12"]    \n\t"            \
+  "slt     %[temp16],      %["#TEMP0"],    $zero           \n\t"            \
+  "slt     %[temp17],      %["#TEMP4"],    $zero           \n\t"            \
+  "slt     %[temp18],      %["#TEMP8"],    $zero           \n\t"            \
+  "slt     %[temp19],      %["#TEMP12"],   $zero           \n\t"            \
+  "movn    %["#TEMP0"],    $zero,          %[temp16]       \n\t"            \
+  "movn    %["#TEMP4"],    $zero,          %[temp17]       \n\t"            \
+  "movn    %["#TEMP8"],    $zero,          %[temp18]       \n\t"            \
+  "movn    %["#TEMP12"],   $zero,          %[temp19]       \n\t"            \
+  "addiu   %[temp20],      $zero,          255             \n\t"            \
+  "slt     %[temp16],      %["#TEMP0"],    %[temp20]       \n\t"            \
+  "slt     %[temp17],      %["#TEMP4"],    %[temp20]       \n\t"            \
+  "slt     %[temp18],      %["#TEMP8"],    %[temp20]       \n\t"            \
+  "slt     %[temp19],      %["#TEMP12"],   %[temp20]       \n\t"            \
+  "movz    %["#TEMP0"],    %[temp20],      %[temp16]       \n\t"            \
+  "movz    %["#TEMP4"],    %[temp20],      %[temp17]       \n\t"            \
+  "lw      %[temp16],      8(%[args])                      \n\t"            \
+  "movz    %["#TEMP8"],    %[temp20],      %[temp18]       \n\t"            \
+  "movz    %["#TEMP12"],   %[temp20],      %[temp19]       \n\t"            \
+  "sb      %["#TEMP0"],    "#A"(%[temp16])                 \n\t"            \
+  "sb      %["#TEMP4"],    "#B"(%[temp16])                 \n\t"            \
+  "sb      %["#TEMP8"],    "#C"(%[temp16])                 \n\t"            \
+  "sb      %["#TEMP12"],   "#D"(%[temp16])                 \n\t"
+
+// Does one or two inverse transforms.
+static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
+                                      uint8_t* dst) {
+  int temp0, temp1, temp2, temp3, temp4, temp5, temp6;
+  int temp7, temp8, temp9, temp10, temp11, temp12, temp13;
+  int temp14, temp15, temp16, temp17, temp18, temp19, temp20;
+  const int* args[3] = {(const int*)ref, (const int*)in, (const int*)dst};
+
+  __asm__ volatile(
+    "lw      %[temp20],      4(%[args])                      \n\t"
+    VERTICAL_PASS(0, 16,  8, 24, temp4,  temp0,  temp1,  temp2,  temp3)
+    VERTICAL_PASS(2, 18, 10, 26, temp8,  temp4,  temp5,  temp6,  temp7)
+    VERTICAL_PASS(4, 20, 12, 28, temp12, temp8,  temp9,  temp10, temp11)
+    VERTICAL_PASS(6, 22, 14, 30, temp20, temp12, temp13, temp14, temp15)
+
+    HORIZONTAL_PASS( 0,  1,  2,  3, temp0, temp4, temp8,  temp12)
+    HORIZONTAL_PASS(16, 17, 18, 19, temp1, temp5, temp9,  temp13)
+    HORIZONTAL_PASS(32, 33, 34, 35, temp2, temp6, temp10, temp14)
+    HORIZONTAL_PASS(48, 49, 50, 51, temp3, temp7, temp11, temp15)
+
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8),
+      [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11),
+      [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14),
+      [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17),
+      [temp18]"=&r"(temp18), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20)
+    : [args]"r"(args), [kC1]"r"(kC1), [kC2]"r"(kC2)
+    : "memory", "hi", "lo"
+  );
+}
+
+static void ITransform(const uint8_t* ref, const int16_t* in,
+                       uint8_t* dst, int do_two) {
+  ITransformOne(ref, in, dst);
+  if (do_two) {
+    ITransformOne(ref + 4, in + 16, dst + 4);
+  }
+}
+
+#undef VERTICAL_PASS
+#undef HORIZONTAL_PASS
+
+// macro for one pass through for loop in QuantizeBlock
+// QUANTDIV macro inlined
+// J - offset in bytes (kZigzag[n] * 2)
+// K - offset in bytes (kZigzag[n] * 4)
+// N - offset in bytes (n * 2)
+#define QUANTIZE_ONE(J, K, N)                                               \
+  "lh           %[temp0],       "#J"(%[ppin])                       \n\t"   \
+  "lhu          %[temp1],       "#J"(%[ppsharpen])                  \n\t"   \
+  "lw           %[temp2],       "#K"(%[ppzthresh])                  \n\t"   \
+  "sra          %[sign],        %[temp0],           15              \n\t"   \
+  "xor          %[coeff],       %[temp0],           %[sign]         \n\t"   \
+  "subu         %[coeff],       %[coeff],           %[sign]         \n\t"   \
+  "addu         %[coeff],       %[coeff],           %[temp1]        \n\t"   \
+  "slt          %[temp4],       %[temp2],           %[coeff]        \n\t"   \
+  "addiu        %[temp5],       $zero,              0               \n\t"   \
+  "addiu        %[level],       $zero,              0               \n\t"   \
+  "beqz         %[temp4],       2f                                  \n\t"   \
+  "lhu          %[temp1],       "#J"(%[ppiq])                       \n\t"   \
+  "lw           %[temp2],       "#K"(%[ppbias])                     \n\t"   \
+  "lhu          %[temp3],       "#J"(%[ppq])                        \n\t"   \
+  "mul          %[level],       %[coeff],           %[temp1]        \n\t"   \
+  "addu         %[level],       %[level],           %[temp2]        \n\t"   \
+  "sra          %[level],       %[level],           17              \n\t"   \
+  "slt          %[temp4],       %[max_level],       %[level]        \n\t"   \
+  "movn         %[level],       %[max_level],       %[temp4]        \n\t"   \
+  "xor          %[level],       %[level],           %[sign]         \n\t"   \
+  "subu         %[level],       %[level],           %[sign]         \n\t"   \
+  "mul          %[temp5],       %[level],           %[temp3]        \n\t"   \
+"2:                                                                 \n\t"   \
+  "sh           %[temp5],       "#J"(%[ppin])                       \n\t"   \
+  "sh           %[level],       "#N"(%[pout])                       \n\t"
+
+static int QuantizeBlock(int16_t in[16], int16_t out[16],
+                         const VP8Matrix* const mtx) {
+  int temp0, temp1, temp2, temp3, temp4, temp5;
+  int sign, coeff, level, i;
+  int max_level = MAX_LEVEL;
+
+  int16_t* ppin             = &in[0];
+  int16_t* pout             = &out[0];
+  const uint16_t* ppsharpen = &mtx->sharpen_[0];
+  const uint32_t* ppzthresh = &mtx->zthresh_[0];
+  const uint16_t* ppq       = &mtx->q_[0];
+  const uint16_t* ppiq      = &mtx->iq_[0];
+  const uint32_t* ppbias    = &mtx->bias_[0];
+
+  __asm__ volatile(
+    QUANTIZE_ONE( 0,  0,  0)
+    QUANTIZE_ONE( 2,  4,  2)
+    QUANTIZE_ONE( 8, 16,  4)
+    QUANTIZE_ONE(16, 32,  6)
+    QUANTIZE_ONE(10, 20,  8)
+    QUANTIZE_ONE( 4,  8, 10)
+    QUANTIZE_ONE( 6, 12, 12)
+    QUANTIZE_ONE(12, 24, 14)
+    QUANTIZE_ONE(18, 36, 16)
+    QUANTIZE_ONE(24, 48, 18)
+    QUANTIZE_ONE(26, 52, 20)
+    QUANTIZE_ONE(20, 40, 22)
+    QUANTIZE_ONE(14, 28, 24)
+    QUANTIZE_ONE(22, 44, 26)
+    QUANTIZE_ONE(28, 56, 28)
+    QUANTIZE_ONE(30, 60, 30)
+
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
+      [temp2]"=&r"(temp2), [temp3]"=&r"(temp3),
+      [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [sign]"=&r"(sign), [coeff]"=&r"(coeff),
+      [level]"=&r"(level)
+    : [pout]"r"(pout), [ppin]"r"(ppin),
+      [ppiq]"r"(ppiq), [max_level]"r"(max_level),
+      [ppbias]"r"(ppbias), [ppzthresh]"r"(ppzthresh),
+      [ppsharpen]"r"(ppsharpen), [ppq]"r"(ppq)
+    : "memory", "hi", "lo"
+  );
+
+  // moved out from macro to increase possibility for earlier breaking
+  for (i = 15; i >= 0; i--) {
+    if (out[i]) return 1;
+  }
+  return 0;
+}
+
+#undef QUANTIZE_ONE
+
+// macro for one horizontal pass in Disto4x4 (TTransform)
+// two calls of function TTransform are merged into single one
+// A..D - offsets in bytes to load from a and b buffers
+// E..H - offsets in bytes to store first results to tmp buffer
+// E1..H1 - offsets in bytes to store second results to tmp buffer
+#define HORIZONTAL_PASS(A, B, C, D, E, F, G, H, E1, F1, G1, H1)   \
+  "lbu    %[temp0],  "#A"(%[a])              \n\t"                \
+  "lbu    %[temp1],  "#B"(%[a])              \n\t"                \
+  "lbu    %[temp2],  "#C"(%[a])              \n\t"                \
+  "lbu    %[temp3],  "#D"(%[a])              \n\t"                \
+  "lbu    %[temp4],  "#A"(%[b])              \n\t"                \
+  "lbu    %[temp5],  "#B"(%[b])              \n\t"                \
+  "lbu    %[temp6],  "#C"(%[b])              \n\t"                \
+  "lbu    %[temp7],  "#D"(%[b])              \n\t"                \
+  "addu   %[temp8],  %[temp0],    %[temp2]   \n\t"                \
+  "subu   %[temp0],  %[temp0],    %[temp2]   \n\t"                \
+  "addu   %[temp2],  %[temp1],    %[temp3]   \n\t"                \
+  "subu   %[temp1],  %[temp1],    %[temp3]   \n\t"                \
+  "addu   %[temp3],  %[temp4],    %[temp6]   \n\t"                \
+  "subu   %[temp4],  %[temp4],    %[temp6]   \n\t"                \
+  "addu   %[temp6],  %[temp5],    %[temp7]   \n\t"                \
+  "subu   %[temp5],  %[temp5],    %[temp7]   \n\t"                \
+  "addu   %[temp7],  %[temp8],    %[temp2]   \n\t"                \
+  "subu   %[temp2],  %[temp8],    %[temp2]   \n\t"                \
+  "addu   %[temp8],  %[temp0],    %[temp1]   \n\t"                \
+  "subu   %[temp0],  %[temp0],    %[temp1]   \n\t"                \
+  "addu   %[temp1],  %[temp3],    %[temp6]   \n\t"                \
+  "subu   %[temp3],  %[temp3],    %[temp6]   \n\t"                \
+  "addu   %[temp6],  %[temp4],    %[temp5]   \n\t"                \
+  "subu   %[temp4],  %[temp4],    %[temp5]   \n\t"                \
+  "sw     %[temp7],  "#E"(%[tmp])            \n\t"                \
+  "sw     %[temp2],  "#H"(%[tmp])            \n\t"                \
+  "sw     %[temp8],  "#F"(%[tmp])            \n\t"                \
+  "sw     %[temp0],  "#G"(%[tmp])            \n\t"                \
+  "sw     %[temp1],  "#E1"(%[tmp])           \n\t"                \
+  "sw     %[temp3],  "#H1"(%[tmp])           \n\t"                \
+  "sw     %[temp6],  "#F1"(%[tmp])           \n\t"                \
+  "sw     %[temp4],  "#G1"(%[tmp])           \n\t"
+
+// macro for one vertical pass in Disto4x4 (TTransform)
+// two calls of function TTransform are merged into single one
+// since only one accu is available in mips32r1 instruction set
+//   first is done second call of function TTransform and after
+//   that first one.
+//   const int sum1 = TTransform(a, w);
+//   const int sum2 = TTransform(b, w);
+//   return abs(sum2 - sum1) >> 5;
+//   (sum2 - sum1) is calculated with madds (sub2) and msubs (sub1)
+// A..D - offsets in bytes to load first results from tmp buffer
+// A1..D1 - offsets in bytes to load second results from tmp buffer
+// E..H - offsets in bytes to load from w buffer
+#define VERTICAL_PASS(A, B, C, D, A1, B1, C1, D1, E, F, G, H)     \
+  "lw     %[temp0],  "#A1"(%[tmp])           \n\t"                \
+  "lw     %[temp1],  "#C1"(%[tmp])           \n\t"                \
+  "lw     %[temp2],  "#B1"(%[tmp])           \n\t"                \
+  "lw     %[temp3],  "#D1"(%[tmp])           \n\t"                \
+  "addu   %[temp8],  %[temp0],    %[temp1]   \n\t"                \
+  "subu   %[temp0],  %[temp0],    %[temp1]   \n\t"                \
+  "addu   %[temp1],  %[temp2],    %[temp3]   \n\t"                \
+  "subu   %[temp2],  %[temp2],    %[temp3]   \n\t"                \
+  "addu   %[temp3],  %[temp8],    %[temp1]   \n\t"                \
+  "subu   %[temp8],  %[temp8],    %[temp1]   \n\t"                \
+  "addu   %[temp1],  %[temp0],    %[temp2]   \n\t"                \
+  "subu   %[temp0],  %[temp0],    %[temp2]   \n\t"                \
+  "sra    %[temp4],  %[temp3],    31         \n\t"                \
+  "sra    %[temp5],  %[temp1],    31         \n\t"                \
+  "sra    %[temp6],  %[temp0],    31         \n\t"                \
+  "sra    %[temp7],  %[temp8],    31         \n\t"                \
+  "xor    %[temp3],  %[temp3],    %[temp4]   \n\t"                \
+  "xor    %[temp1],  %[temp1],    %[temp5]   \n\t"                \
+  "xor    %[temp0],  %[temp0],    %[temp6]   \n\t"                \
+  "xor    %[temp8],  %[temp8],    %[temp7]   \n\t"                \
+  "subu   %[temp3],  %[temp3],    %[temp4]   \n\t"                \
+  "subu   %[temp1],  %[temp1],    %[temp5]   \n\t"                \
+  "subu   %[temp0],  %[temp0],    %[temp6]   \n\t"                \
+  "subu   %[temp8],  %[temp8],    %[temp7]   \n\t"                \
+  "lhu    %[temp4],  "#E"(%[w])              \n\t"                \
+  "lhu    %[temp5],  "#F"(%[w])              \n\t"                \
+  "lhu    %[temp6],  "#G"(%[w])              \n\t"                \
+  "lhu    %[temp7],  "#H"(%[w])              \n\t"                \
+  "madd   %[temp4],  %[temp3]                \n\t"                \
+  "madd   %[temp5],  %[temp1]                \n\t"                \
+  "madd   %[temp6],  %[temp0]                \n\t"                \
+  "madd   %[temp7],  %[temp8]                \n\t"                \
+  "lw     %[temp0],  "#A"(%[tmp])            \n\t"                \
+  "lw     %[temp1],  "#C"(%[tmp])            \n\t"                \
+  "lw     %[temp2],  "#B"(%[tmp])            \n\t"                \
+  "lw     %[temp3],  "#D"(%[tmp])            \n\t"                \
+  "addu   %[temp8],  %[temp0],    %[temp1]   \n\t"                \
+  "subu   %[temp0],  %[temp0],    %[temp1]   \n\t"                \
+  "addu   %[temp1],  %[temp2],    %[temp3]   \n\t"                \
+  "subu   %[temp2],  %[temp2],    %[temp3]   \n\t"                \
+  "addu   %[temp3],  %[temp8],    %[temp1]   \n\t"                \
+  "subu   %[temp1],  %[temp8],    %[temp1]   \n\t"                \
+  "addu   %[temp8],  %[temp0],    %[temp2]   \n\t"                \
+  "subu   %[temp0],  %[temp0],    %[temp2]   \n\t"                \
+  "sra    %[temp2],  %[temp3],    31         \n\t"                \
+  "xor    %[temp3],  %[temp3],    %[temp2]   \n\t"                \
+  "subu   %[temp3],  %[temp3],    %[temp2]   \n\t"                \
+  "msub   %[temp4],  %[temp3]                \n\t"                \
+  "sra    %[temp2],  %[temp8],    31         \n\t"                \
+  "sra    %[temp3],  %[temp0],    31         \n\t"                \
+  "sra    %[temp4],  %[temp1],    31         \n\t"                \
+  "xor    %[temp8],  %[temp8],    %[temp2]   \n\t"                \
+  "xor    %[temp0],  %[temp0],    %[temp3]   \n\t"                \
+  "xor    %[temp1],  %[temp1],    %[temp4]   \n\t"                \
+  "subu   %[temp8],  %[temp8],    %[temp2]   \n\t"                \
+  "subu   %[temp0],  %[temp0],    %[temp3]   \n\t"                \
+  "subu   %[temp1],  %[temp1],    %[temp4]   \n\t"                \
+  "msub   %[temp5],  %[temp8]                \n\t"                \
+  "msub   %[temp6],  %[temp0]                \n\t"                \
+  "msub   %[temp7],  %[temp1]                \n\t"
+
+static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
+                    const uint16_t* const w) {
+  int tmp[32];
+  int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8;
+
+  __asm__ volatile(
+    HORIZONTAL_PASS( 0,  1,  2,  3,    0,  4,  8, 12,    64,  68,  72,  76)
+    HORIZONTAL_PASS(16, 17, 18, 19,   16, 20, 24, 28,    80,  84,  88,  92)
+    HORIZONTAL_PASS(32, 33, 34, 35,   32, 36, 40, 44,    96, 100, 104, 108)
+    HORIZONTAL_PASS(48, 49, 50, 51,   48, 52, 56, 60,   112, 116, 120, 124)
+    "mthi   $zero                             \n\t"
+    "mtlo   $zero                             \n\t"
+    VERTICAL_PASS( 0, 16, 32, 48,     64, 80,  96, 112,   0,  8, 16, 24)
+    VERTICAL_PASS( 4, 20, 36, 52,     68, 84, 100, 116,   2, 10, 18, 26)
+    VERTICAL_PASS( 8, 24, 40, 56,     72, 88, 104, 120,   4, 12, 20, 28)
+    VERTICAL_PASS(12, 28, 44, 60,     76, 92, 108, 124,   6, 14, 22, 30)
+    "mflo   %[temp0]                          \n\t"
+    "sra    %[temp1],  %[temp0],  31          \n\t"
+    "xor    %[temp0],  %[temp0],  %[temp1]    \n\t"
+    "subu   %[temp0],  %[temp0],  %[temp1]    \n\t"
+    "sra    %[temp0],  %[temp0],  5           \n\t"
+
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8)
+    : [a]"r"(a), [b]"r"(b), [w]"r"(w), [tmp]"r"(tmp)
+    : "memory", "hi", "lo"
+  );
+
+  return temp0;
+}
+
+#undef VERTICAL_PASS
+#undef HORIZONTAL_PASS
+
+static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
+                      const uint16_t* const w) {
+  int D = 0;
+  int x, y;
+  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
+    for (x = 0; x < 16; x += 4) {
+      D += Disto4x4(a + x + y, b + x + y, w);
+    }
+  }
+  return D;
+}
+
+// macro for one horizontal pass in FTransform
+// temp0..temp15 holds tmp[0]..tmp[15]
+// A..D - offsets in bytes to load from src and ref buffers
+// TEMP0..TEMP3 - registers for corresponding tmp elements
+#define HORIZONTAL_PASS(A, B, C, D, TEMP0, TEMP1, TEMP2, TEMP3) \
+  "lw     %["#TEMP1"],  0(%[args])                     \n\t"    \
+  "lw     %["#TEMP2"],  4(%[args])                     \n\t"    \
+  "lbu    %[temp16],    "#A"(%["#TEMP1"])              \n\t"    \
+  "lbu    %[temp17],    "#A"(%["#TEMP2"])              \n\t"    \
+  "lbu    %[temp18],    "#B"(%["#TEMP1"])              \n\t"    \
+  "lbu    %[temp19],    "#B"(%["#TEMP2"])              \n\t"    \
+  "subu   %[temp20],    %[temp16],    %[temp17]        \n\t"    \
+  "lbu    %[temp16],    "#C"(%["#TEMP1"])              \n\t"    \
+  "lbu    %[temp17],    "#C"(%["#TEMP2"])              \n\t"    \
+  "subu   %["#TEMP0"],  %[temp18],    %[temp19]        \n\t"    \
+  "lbu    %[temp18],    "#D"(%["#TEMP1"])              \n\t"    \
+  "lbu    %[temp19],    "#D"(%["#TEMP2"])              \n\t"    \
+  "subu   %["#TEMP1"],  %[temp16],    %[temp17]        \n\t"    \
+  "subu   %["#TEMP2"],  %[temp18],    %[temp19]        \n\t"    \
+  "addu   %["#TEMP3"],  %[temp20],    %["#TEMP2"]      \n\t"    \
+  "subu   %["#TEMP2"],  %[temp20],    %["#TEMP2"]      \n\t"    \
+  "addu   %[temp20],    %["#TEMP0"],  %["#TEMP1"]      \n\t"    \
+  "subu   %["#TEMP0"],  %["#TEMP0"],  %["#TEMP1"]      \n\t"    \
+  "mul    %[temp16],    %["#TEMP2"],  %[c5352]         \n\t"    \
+  "mul    %[temp17],    %["#TEMP2"],  %[c2217]         \n\t"    \
+  "mul    %[temp18],    %["#TEMP0"],  %[c5352]         \n\t"    \
+  "mul    %[temp19],    %["#TEMP0"],  %[c2217]         \n\t"    \
+  "addu   %["#TEMP1"],  %["#TEMP3"],  %[temp20]        \n\t"    \
+  "subu   %[temp20],    %["#TEMP3"],  %[temp20]        \n\t"    \
+  "sll    %["#TEMP0"],  %["#TEMP1"],  3                \n\t"    \
+  "sll    %["#TEMP2"],  %[temp20],    3                \n\t"    \
+  "addiu  %[temp16],    %[temp16],    1812             \n\t"    \
+  "addiu  %[temp17],    %[temp17],    937              \n\t"    \
+  "addu   %[temp16],    %[temp16],    %[temp19]        \n\t"    \
+  "subu   %[temp17],    %[temp17],    %[temp18]        \n\t"    \
+  "sra    %["#TEMP1"],  %[temp16],    9                \n\t"    \
+  "sra    %["#TEMP3"],  %[temp17],    9                \n\t"
+
+// macro for one vertical pass in FTransform
+// temp0..temp15 holds tmp[0]..tmp[15]
+// A..D - offsets in bytes to store to out buffer
+// TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements
+#define VERTICAL_PASS(A, B, C, D, TEMP0, TEMP4, TEMP8, TEMP12)  \
+  "addu   %[temp16],    %["#TEMP0"],  %["#TEMP12"]     \n\t"    \
+  "subu   %[temp19],    %["#TEMP0"],  %["#TEMP12"]     \n\t"    \
+  "addu   %[temp17],    %["#TEMP4"],  %["#TEMP8"]      \n\t"    \
+  "subu   %[temp18],    %["#TEMP4"],  %["#TEMP8"]      \n\t"    \
+  "mul    %["#TEMP8"],  %[temp19],    %[c2217]         \n\t"    \
+  "mul    %["#TEMP12"], %[temp18],    %[c2217]         \n\t"    \
+  "mul    %["#TEMP4"],  %[temp19],    %[c5352]         \n\t"    \
+  "mul    %[temp18],    %[temp18],    %[c5352]         \n\t"    \
+  "addiu  %[temp16],    %[temp16],    7                \n\t"    \
+  "addu   %["#TEMP0"],  %[temp16],    %[temp17]        \n\t"    \
+  "sra    %["#TEMP0"],  %["#TEMP0"],  4                \n\t"    \
+  "addu   %["#TEMP12"], %["#TEMP12"], %["#TEMP4"]      \n\t"    \
+  "subu   %["#TEMP4"],  %[temp16],    %[temp17]        \n\t"    \
+  "sra    %["#TEMP4"],  %["#TEMP4"],  4                \n\t"    \
+  "addiu  %["#TEMP8"],  %["#TEMP8"],  30000            \n\t"    \
+  "addiu  %["#TEMP12"], %["#TEMP12"], 12000            \n\t"    \
+  "addiu  %["#TEMP8"],  %["#TEMP8"],  21000            \n\t"    \
+  "subu   %["#TEMP8"],  %["#TEMP8"],  %[temp18]        \n\t"    \
+  "sra    %["#TEMP12"], %["#TEMP12"], 16               \n\t"    \
+  "sra    %["#TEMP8"],  %["#TEMP8"],  16               \n\t"    \
+  "addiu  %[temp16],    %["#TEMP12"], 1                \n\t"    \
+  "movn   %["#TEMP12"], %[temp16],    %[temp19]        \n\t"    \
+  "sh     %["#TEMP0"],  "#A"(%[temp20])                \n\t"    \
+  "sh     %["#TEMP4"],  "#C"(%[temp20])                \n\t"    \
+  "sh     %["#TEMP8"],  "#D"(%[temp20])                \n\t"    \
+  "sh     %["#TEMP12"], "#B"(%[temp20])                \n\t"
+
+static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
+  int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8;
+  int temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16;
+  int temp17, temp18, temp19, temp20;
+  const int c2217 = 2217;
+  const int c5352 = 5352;
+  const int* const args[3] =
+      { (const int*)src, (const int*)ref, (const int*)out };
+
+  __asm__ volatile(
+    HORIZONTAL_PASS( 0,  1,  2,  3, temp0,  temp1,  temp2,  temp3)
+    HORIZONTAL_PASS(16, 17, 18, 19, temp4,  temp5,  temp6,  temp7)
+    HORIZONTAL_PASS(32, 33, 34, 35, temp8,  temp9,  temp10, temp11)
+    HORIZONTAL_PASS(48, 49, 50, 51, temp12, temp13, temp14, temp15)
+    "lw   %[temp20],    8(%[args])                     \n\t"
+    VERTICAL_PASS(0,  8, 16, 24, temp0, temp4, temp8,  temp12)
+    VERTICAL_PASS(2, 10, 18, 26, temp1, temp5, temp9,  temp13)
+    VERTICAL_PASS(4, 12, 20, 28, temp2, temp6, temp10, temp14)
+    VERTICAL_PASS(6, 14, 22, 30, temp3, temp7, temp11, temp15)
+
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8),
+      [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11),
+      [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14),
+      [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17),
+      [temp18]"=&r"(temp18), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20)
+    : [args]"r"(args), [c2217]"r"(c2217), [c5352]"r"(c5352)
+    : "memory", "hi", "lo"
+  );
+}
+
+#undef VERTICAL_PASS
+#undef HORIZONTAL_PASS
+
+// Forward declaration.
+extern int VP8GetResidualCostMIPS32(int ctx0, const VP8Residual* const res);
+
+int VP8GetResidualCostMIPS32(int ctx0, const VP8Residual* const res) {
+  int n = res->first;
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  int p0 = res->prob[n][ctx0][0];
+  const uint16_t* t = res->cost[n][ctx0];
+  int cost;
+  const int const_2 = 2;
+  const int const_255 = 255;
+  const int const_max_level = MAX_VARIABLE_LEVEL;
+  int res_cost;
+  int res_prob;
+  int res_coeffs;
+  int res_last;
+  int v_reg;
+  int b_reg;
+  int ctx_reg;
+  int cost_add, temp_1, temp_2, temp_3;
+
+  if (res->last < 0) {
+    return VP8BitCost(0, p0);
+  }
+
+  cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0;
+
+  res_cost = (int)res->cost;
+  res_prob = (int)res->prob;
+  res_coeffs = (int)res->coeffs;
+  res_last = (int)res->last;
+
+  __asm__ volatile(
+    ".set   push                                                           \n\t"
+    ".set   noreorder                                                      \n\t"
+
+    "sll    %[temp_1],     %[n],              1                            \n\t"
+    "addu   %[res_coeffs], %[res_coeffs],     %[temp_1]                    \n\t"
+    "slt    %[temp_2],     %[n],              %[res_last]                  \n\t"
+    "bnez   %[temp_2],     1f                                              \n\t"
+    " li    %[cost_add],   0                                               \n\t"
+    "b      2f                                                             \n\t"
+    " nop                                                                  \n\t"
+  "1:                                                                      \n\t"
+    "lh     %[v_reg],      0(%[res_coeffs])                                \n\t"
+    "addu   %[b_reg],      %[n],              %[VP8EncBands]               \n\t"
+    "move   %[temp_1],     %[const_max_level]                              \n\t"
+    "addu   %[cost],       %[cost],           %[cost_add]                  \n\t"
+    "negu   %[temp_2],     %[v_reg]                                        \n\t"
+    "slti   %[temp_3],     %[v_reg],          0                            \n\t"
+    "movn   %[v_reg],      %[temp_2],         %[temp_3]                    \n\t"
+    "lbu    %[b_reg],      1(%[b_reg])                                     \n\t"
+    "li     %[cost_add],   0                                               \n\t"
+
+    "sltiu  %[temp_3],     %[v_reg],          2                            \n\t"
+    "move   %[ctx_reg],    %[v_reg]                                        \n\t"
+    "movz   %[ctx_reg],    %[const_2],        %[temp_3]                    \n\t"
+    //  cost += VP8LevelCost(t, v);
+    "slt    %[temp_3],     %[v_reg],          %[const_max_level]           \n\t"
+    "movn   %[temp_1],     %[v_reg],          %[temp_3]                    \n\t"
+    "sll    %[temp_2],     %[v_reg],          1                            \n\t"
+    "addu   %[temp_2],     %[temp_2],         %[VP8LevelFixedCosts]        \n\t"
+    "lhu    %[temp_2],     0(%[temp_2])                                    \n\t"
+    "sll    %[temp_1],     %[temp_1],         1                            \n\t"
+    "addu   %[temp_1],     %[temp_1],         %[t]                         \n\t"
+    "lhu    %[temp_3],     0(%[temp_1])                                    \n\t"
+    "addu   %[cost],       %[cost],           %[temp_2]                    \n\t"
+
+    //  t = res->cost[b][ctx];
+    "sll    %[temp_1],     %[ctx_reg],        7                            \n\t"
+    "sll    %[temp_2],     %[ctx_reg],        3                            \n\t"
+    "addu   %[cost],       %[cost],           %[temp_3]                    \n\t"
+    "addu   %[temp_1],     %[temp_1],         %[temp_2]                    \n\t"
+    "sll    %[temp_2],     %[b_reg],          3                            \n\t"
+    "sll    %[temp_3],     %[b_reg],          5                            \n\t"
+    "sub    %[temp_2],     %[temp_3],         %[temp_2]                    \n\t"
+    "sll    %[temp_3],     %[temp_2],         4                            \n\t"
+    "addu   %[temp_1],     %[temp_1],         %[temp_3]                    \n\t"
+    "addu   %[temp_2],     %[temp_2],         %[res_cost]                  \n\t"
+    "addiu  %[n],          %[n],              1                            \n\t"
+    "addu   %[t],          %[temp_1],         %[temp_2]                    \n\t"
+    "slt    %[temp_1],     %[n],              %[res_last]                  \n\t"
+    "bnez   %[temp_1],     1b                                              \n\t"
+    " addiu %[res_coeffs], %[res_coeffs],     2                            \n\t"
+   "2:                                                                     \n\t"
+
+    ".set   pop                                                            \n\t"
+    : [cost]"+r"(cost), [t]"+r"(t), [n]"+r"(n), [v_reg]"=&r"(v_reg),
+      [ctx_reg]"=&r"(ctx_reg), [b_reg]"=&r"(b_reg), [cost_add]"=&r"(cost_add),
+      [temp_1]"=&r"(temp_1), [temp_2]"=&r"(temp_2), [temp_3]"=&r"(temp_3)
+    : [const_2]"r"(const_2), [const_255]"r"(const_255), [res_last]"r"(res_last),
+      [VP8EntropyCost]"r"(VP8EntropyCost), [VP8EncBands]"r"(VP8EncBands),
+      [const_max_level]"r"(const_max_level), [res_prob]"r"(res_prob),
+      [VP8LevelFixedCosts]"r"(VP8LevelFixedCosts), [res_coeffs]"r"(res_coeffs),
+      [res_cost]"r"(res_cost)
+    : "memory"
+  );
+
+  // Last coefficient is always non-zero
+  {
+    const int v = abs(res->coeffs[n]);
+    assert(v != 0);
+    cost += VP8LevelCost(t, v);
+    if (n < 15) {
+      const int b = VP8EncBands[n + 1];
+      const int ctx = (v == 1) ? 1 : 2;
+      const int last_p0 = res->prob[b][ctx][0];
+      cost += VP8BitCost(0, last_p0);
+    }
+  }
+  return cost;
+}
+
+#define GET_SSE_INNER(A, B, C, D)                               \
+  "lbu     %[temp0],    "#A"(%[a])                   \n\t"      \
+  "lbu     %[temp1],    "#A"(%[b])                   \n\t"      \
+  "lbu     %[temp2],    "#B"(%[a])                   \n\t"      \
+  "lbu     %[temp3],    "#B"(%[b])                   \n\t"      \
+  "lbu     %[temp4],    "#C"(%[a])                   \n\t"      \
+  "lbu     %[temp5],    "#C"(%[b])                   \n\t"      \
+  "lbu     %[temp6],    "#D"(%[a])                   \n\t"      \
+  "lbu     %[temp7],    "#D"(%[b])                   \n\t"      \
+  "subu    %[temp0],    %[temp0],     %[temp1]       \n\t"      \
+  "subu    %[temp2],    %[temp2],     %[temp3]       \n\t"      \
+  "subu    %[temp4],    %[temp4],     %[temp5]       \n\t"      \
+  "subu    %[temp6],    %[temp6],     %[temp7]       \n\t"      \
+  "madd    %[temp0],    %[temp0]                     \n\t"      \
+  "madd    %[temp2],    %[temp2]                     \n\t"      \
+  "madd    %[temp4],    %[temp4]                     \n\t"      \
+  "madd    %[temp6],    %[temp6]                     \n\t"
+
+#define GET_SSE(A, B, C, D)               \
+  GET_SSE_INNER(A, A + 1, A + 2, A + 3)   \
+  GET_SSE_INNER(B, B + 1, B + 2, B + 3)   \
+  GET_SSE_INNER(C, C + 1, C + 2, C + 3)   \
+  GET_SSE_INNER(D, D + 1, D + 2, D + 3)
+
+#if !defined(WORK_AROUND_GCC)
+static int SSE16x16(const uint8_t* a, const uint8_t* b) {
+  int count;
+  int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
+
+  __asm__ volatile(
+     "mult   $zero,    $zero                            \n\t"
+
+     GET_SSE(  0,   4,   8,  12)
+     GET_SSE( 16,  20,  24,  28)
+     GET_SSE( 32,  36,  40,  44)
+     GET_SSE( 48,  52,  56,  60)
+     GET_SSE( 64,  68,  72,  76)
+     GET_SSE( 80,  84,  88,  92)
+     GET_SSE( 96, 100, 104, 108)
+     GET_SSE(112, 116, 120, 124)
+     GET_SSE(128, 132, 136, 140)
+     GET_SSE(144, 148, 152, 156)
+     GET_SSE(160, 164, 168, 172)
+     GET_SSE(176, 180, 184, 188)
+     GET_SSE(192, 196, 200, 204)
+     GET_SSE(208, 212, 216, 220)
+     GET_SSE(224, 228, 232, 236)
+     GET_SSE(240, 244, 248, 252)
+
+    "mflo    %[count]                                   \n\t"
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count)
+    : [a]"r"(a), [b]"r"(b)
+    : "memory", "hi" , "lo"
+  );
+  return count;
+}
+
+static int SSE16x8(const uint8_t* a, const uint8_t* b) {
+  int count;
+  int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
+
+  __asm__ volatile(
+     "mult   $zero,    $zero                            \n\t"
+
+     GET_SSE(  0,   4,   8,  12)
+     GET_SSE( 16,  20,  24,  28)
+     GET_SSE( 32,  36,  40,  44)
+     GET_SSE( 48,  52,  56,  60)
+     GET_SSE( 64,  68,  72,  76)
+     GET_SSE( 80,  84,  88,  92)
+     GET_SSE( 96, 100, 104, 108)
+     GET_SSE(112, 116, 120, 124)
+
+    "mflo    %[count]                                   \n\t"
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count)
+    : [a]"r"(a), [b]"r"(b)
+    : "memory", "hi" , "lo"
+  );
+  return count;
+}
+
+static int SSE8x8(const uint8_t* a, const uint8_t* b) {
+  int count;
+  int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
+
+  __asm__ volatile(
+     "mult   $zero,    $zero                            \n\t"
+
+     GET_SSE( 0,   4,  16,  20)
+     GET_SSE(32,  36,  48,  52)
+     GET_SSE(64,  68,  80,  84)
+     GET_SSE(96, 100, 112, 116)
+
+    "mflo    %[count]                                   \n\t"
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count)
+    : [a]"r"(a), [b]"r"(b)
+    : "memory", "hi" , "lo"
+  );
+  return count;
+}
+
+static int SSE4x4(const uint8_t* a, const uint8_t* b) {
+  int count;
+  int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
+
+  __asm__ volatile(
+     "mult   $zero,    $zero                            \n\t"
+
+     GET_SSE(0, 16, 32, 48)
+
+    "mflo    %[count]                                   \n\t"
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count)
+    : [a]"r"(a), [b]"r"(b)
+    : "memory", "hi" , "lo"
+  );
+  return count;
+}
+
+#endif  // WORK_AROUND_GCC
+
+#undef GET_SSE_MIPS32
+#undef GET_SSE_MIPS32_INNER
+
+#endif  // WEBP_USE_MIPS32
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8EncDspInitMIPS32(void);
+
+void VP8EncDspInitMIPS32(void) {
+#if defined(WEBP_USE_MIPS32)
+  VP8ITransform = ITransform;
+  VP8EncQuantizeBlock = QuantizeBlock;
+  VP8TDisto4x4 = Disto4x4;
+  VP8TDisto16x16 = Disto16x16;
+  VP8FTransform = FTransform;
+#if !defined(WORK_AROUND_GCC)
+  VP8SSE16x16 = SSE16x16;
+  VP8SSE8x8 = SSE8x8;
+  VP8SSE16x8 = SSE16x8;
+  VP8SSE4x4 = SSE4x4;
+#endif
+#endif  // WEBP_USE_MIPS32
+}

src/dsp/enc_neon.c

@@ -15,18 +15,122 @@
 
 #if defined(WEBP_USE_NEON)
 
+#include <assert.h>
+
+#include "./neon.h"
 #include "../enc/vp8enci.h"
 
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)
 
 // Inverse transform.
-// This code is pretty much the same as TransformOneNEON in the decoder, except
+// This code is pretty much the same as TransformOne in the dec_neon.c, except
 // for subtraction to *ref. See the comments there for algorithmic explanations.
+
+static const int16_t kC1 = 20091;
+static const int16_t kC2 = 17734;  // half of kC2, actually. See comment above.
+
+// This code works but is *slower* than the inlined-asm version below
+// (with gcc-4.6). So we disable it for now. Later, it'll be conditional to
+// USE_INTRINSICS define.
+// With gcc-4.8, it's a little faster speed than inlined-assembly.
+#if defined(USE_INTRINSICS)
+
+// Treats 'v' as an uint8x8_t and zero extends to an int16x8_t.
+static WEBP_INLINE int16x8_t ConvertU8ToS16(uint32x2_t v) {
+  return vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(v)));
+}
+
+// Performs unsigned 8b saturation on 'dst01' and 'dst23' storing the result
+// to the corresponding rows of 'dst'.
+static WEBP_INLINE void SaturateAndStore4x4(uint8_t* const dst,
+                                            const int16x8_t dst01,
+                                            const int16x8_t dst23) {
+  // Unsigned saturate to 8b.
+  const uint8x8_t dst01_u8 = vqmovun_s16(dst01);
+  const uint8x8_t dst23_u8 = vqmovun_s16(dst23);
+
+  // Store the results.
+  vst1_lane_u32((uint32_t*)(dst + 0 * BPS), vreinterpret_u32_u8(dst01_u8), 0);
+  vst1_lane_u32((uint32_t*)(dst + 1 * BPS), vreinterpret_u32_u8(dst01_u8), 1);
+  vst1_lane_u32((uint32_t*)(dst + 2 * BPS), vreinterpret_u32_u8(dst23_u8), 0);
+  vst1_lane_u32((uint32_t*)(dst + 3 * BPS), vreinterpret_u32_u8(dst23_u8), 1);
+}
+
+static WEBP_INLINE void Add4x4(const int16x8_t row01, const int16x8_t row23,
+                               const uint8_t* const ref, uint8_t* const dst) {
+  uint32x2_t dst01 = vdup_n_u32(0);
+  uint32x2_t dst23 = vdup_n_u32(0);
+
+  // Load the source pixels.
+  dst01 = vld1_lane_u32((uint32_t*)(ref + 0 * BPS), dst01, 0);
+  dst23 = vld1_lane_u32((uint32_t*)(ref + 2 * BPS), dst23, 0);
+  dst01 = vld1_lane_u32((uint32_t*)(ref + 1 * BPS), dst01, 1);
+  dst23 = vld1_lane_u32((uint32_t*)(ref + 3 * BPS), dst23, 1);
+
+  {
+    // Convert to 16b.
+    const int16x8_t dst01_s16 = ConvertU8ToS16(dst01);
+    const int16x8_t dst23_s16 = ConvertU8ToS16(dst23);
+
+    // Descale with rounding.
+    const int16x8_t out01 = vrsraq_n_s16(dst01_s16, row01, 3);
+    const int16x8_t out23 = vrsraq_n_s16(dst23_s16, row23, 3);
+    // Add the inverse transform.
+    SaturateAndStore4x4(dst, out01, out23);
+  }
+}
+
+static WEBP_INLINE void Transpose8x2(const int16x8_t in0, const int16x8_t in1,
+                                     int16x8x2_t* const out) {
+  // a0 a1 a2 a3 | b0 b1 b2 b3   => a0 b0 c0 d0 | a1 b1 c1 d1
+  // c0 c1 c2 c3 | d0 d1 d2 d3      a2 b2 c2 d2 | a3 b3 c3 d3
+  const int16x8x2_t tmp0 = vzipq_s16(in0, in1);   // a0 c0 a1 c1 a2 c2 ...
+                                                  // b0 d0 b1 d1 b2 d2 ...
+  *out = vzipq_s16(tmp0.val[0], tmp0.val[1]);
+}
+
+static WEBP_INLINE void TransformPass(int16x8x2_t* const rows) {
+  // {rows} = in0 | in4
+  //          in8 | in12
+  // B1 = in4 | in12
+  const int16x8_t B1 =
+      vcombine_s16(vget_high_s16(rows->val[0]), vget_high_s16(rows->val[1]));
+  // C0 = kC1 * in4 | kC1 * in12
+  // C1 = kC2 * in4 | kC2 * in12
+  const int16x8_t C0 = vsraq_n_s16(B1, vqdmulhq_n_s16(B1, kC1), 1);
+  const int16x8_t C1 = vqdmulhq_n_s16(B1, kC2);
+  const int16x4_t a = vqadd_s16(vget_low_s16(rows->val[0]),
+                                vget_low_s16(rows->val[1]));   // in0 + in8
+  const int16x4_t b = vqsub_s16(vget_low_s16(rows->val[0]),
+                                vget_low_s16(rows->val[1]));   // in0 - in8
+  // c = kC2 * in4 - kC1 * in12
+  // d = kC1 * in4 + kC2 * in12
+  const int16x4_t c = vqsub_s16(vget_low_s16(C1), vget_high_s16(C0));
+  const int16x4_t d = vqadd_s16(vget_low_s16(C0), vget_high_s16(C1));
+  const int16x8_t D0 = vcombine_s16(a, b);      // D0 = a | b
+  const int16x8_t D1 = vcombine_s16(d, c);      // D1 = d | c
+  const int16x8_t E0 = vqaddq_s16(D0, D1);      // a+d | b+c
+  const int16x8_t E_tmp = vqsubq_s16(D0, D1);   // a-d | b-c
+  const int16x8_t E1 = vcombine_s16(vget_high_s16(E_tmp), vget_low_s16(E_tmp));
+  Transpose8x2(E0, E1, rows);
+}
+
+static void ITransformOne(const uint8_t* ref,
+                          const int16_t* in, uint8_t* dst) {
+  int16x8x2_t rows;
+  INIT_VECTOR2(rows, vld1q_s16(in + 0), vld1q_s16(in + 8));
+  TransformPass(&rows);
+  TransformPass(&rows);
+  Add4x4(rows.val[0], rows.val[1], ref, dst);
+}
+
+#else
+
 static void ITransformOne(const uint8_t* ref,
                           const int16_t* in, uint8_t* dst) {
   const int kBPS = BPS;
-  const int16_t kC1C2[] = { 20091, 17734, 0, 0 };  // kC1 / (kC2 >> 1) / 0 / 0
+  const int16_t kC1C2[] = { kC1, kC2, 0, 0 };
 
   __asm__ volatile (
     "vld1.16         {q1, q2}, [%[in]]           \n"
@@ -137,6 +241,8 @@ static void ITransformOne(const uint8_t* ref,
   );
 }
 
+#endif    // USE_INTRINSICS
+
 static void ITransform(const uint8_t* ref,
                        const int16_t* in, uint8_t* dst, int do_two) {
   ITransformOne(ref, in, dst);
@@ -145,76 +251,102 @@ static void ITransform(const uint8_t* ref,
   }
 }
 
-// Same code as dec_neon.c
-static void ITransformWHT(const int16_t* in, int16_t* out) {
-  const int kStep = 32;  // The store is only incrementing the pointer as if we
-                         // had stored a single byte.
-  __asm__ volatile (
-    // part 1
-    // load data into q0, q1
-    "vld1.16         {q0, q1}, [%[in]]           \n"
-
-    "vaddl.s16       q2, d0, d3                  \n" // a0 = in[0] + in[12]
-    "vaddl.s16       q3, d1, d2                  \n" // a1 = in[4] + in[8]
-    "vsubl.s16       q4, d1, d2                  \n" // a2 = in[4] - in[8]
-    "vsubl.s16       q5, d0, d3                  \n" // a3 = in[0] - in[12]
-
-    "vadd.s32        q0, q2, q3                  \n" // tmp[0] = a0 + a1
-    "vsub.s32        q2, q2, q3                  \n" // tmp[8] = a0 - a1
-    "vadd.s32        q1, q5, q4                  \n" // tmp[4] = a3 + a2
-    "vsub.s32        q3, q5, q4                  \n" // tmp[12] = a3 - a2
-
-    // Transpose
-    // q0 = tmp[0, 4, 8, 12], q1 = tmp[2, 6, 10, 14]
-    // q2 = tmp[1, 5, 9, 13], q3 = tmp[3, 7, 11, 15]
-    "vswp            d1, d4                      \n" // vtrn.64 q0, q2
-    "vswp            d3, d6                      \n" // vtrn.64 q1, q3
-    "vtrn.32         q0, q1                      \n"
-    "vtrn.32         q2, q3                      \n"
-
-    "vmov.s32        q4, #3                      \n" // dc = 3
-    "vadd.s32        q0, q0, q4                  \n" // dc = tmp[0] + 3
-    "vadd.s32        q6, q0, q3                  \n" // a0 = dc + tmp[3]
-    "vadd.s32        q7, q1, q2                  \n" // a1 = tmp[1] + tmp[2]
-    "vsub.s32        q8, q1, q2                  \n" // a2 = tmp[1] - tmp[2]
-    "vsub.s32        q9, q0, q3                  \n" // a3 = dc - tmp[3]
-
-    "vadd.s32        q0, q6, q7                  \n"
-    "vshrn.s32       d0, q0, #3                  \n" // (a0 + a1) >> 3
-    "vadd.s32        q1, q9, q8                  \n"
-    "vshrn.s32       d1, q1, #3                  \n" // (a3 + a2) >> 3
-    "vsub.s32        q2, q6, q7                  \n"
-    "vshrn.s32       d2, q2, #3                  \n" // (a0 - a1) >> 3
-    "vsub.s32        q3, q9, q8                  \n"
-    "vshrn.s32       d3, q3, #3                  \n" // (a3 - a2) >> 3
-
-    // set the results to output
-    "vst1.16         d0[0], [%[out]], %[kStep]      \n"
-    "vst1.16         d1[0], [%[out]], %[kStep]      \n"
-    "vst1.16         d2[0], [%[out]], %[kStep]      \n"
-    "vst1.16         d3[0], [%[out]], %[kStep]      \n"
-    "vst1.16         d0[1], [%[out]], %[kStep]      \n"
-    "vst1.16         d1[1], [%[out]], %[kStep]      \n"
-    "vst1.16         d2[1], [%[out]], %[kStep]      \n"
-    "vst1.16         d3[1], [%[out]], %[kStep]      \n"
-    "vst1.16         d0[2], [%[out]], %[kStep]      \n"
-    "vst1.16         d1[2], [%[out]], %[kStep]      \n"
-    "vst1.16         d2[2], [%[out]], %[kStep]      \n"
-    "vst1.16         d3[2], [%[out]], %[kStep]      \n"
-    "vst1.16         d0[3], [%[out]], %[kStep]      \n"
-    "vst1.16         d1[3], [%[out]], %[kStep]      \n"
-    "vst1.16         d2[3], [%[out]], %[kStep]      \n"
-    "vst1.16         d3[3], [%[out]], %[kStep]      \n"
-
-    : [out] "+r"(out)  // modified registers
-    : [in] "r"(in), [kStep] "r"(kStep)  // constants
-    : "memory", "q0", "q1", "q2", "q3", "q4",
-      "q5", "q6", "q7", "q8", "q9" // clobbered
-  );
+// Load all 4x4 pixels into a single uint8x16_t variable.
+static uint8x16_t Load4x4(const uint8_t* src) {
+  uint32x4_t out = vdupq_n_u32(0);
+  out = vld1q_lane_u32((const uint32_t*)(src + 0 * BPS), out, 0);
+  out = vld1q_lane_u32((const uint32_t*)(src + 1 * BPS), out, 1);
+  out = vld1q_lane_u32((const uint32_t*)(src + 2 * BPS), out, 2);
+  out = vld1q_lane_u32((const uint32_t*)(src + 3 * BPS), out, 3);
+  return vreinterpretq_u8_u32(out);
 }
 
 // Forward transform.
 
+#if defined(USE_INTRINSICS)
+
+static WEBP_INLINE void Transpose4x4_S16(const int16x4_t A, const int16x4_t B,
+                                         const int16x4_t C, const int16x4_t D,
+                                         int16x8_t* const out01,
+                                         int16x8_t* const out32) {
+  const int16x4x2_t AB = vtrn_s16(A, B);
+  const int16x4x2_t CD = vtrn_s16(C, D);
+  const int32x2x2_t tmp02 = vtrn_s32(vreinterpret_s32_s16(AB.val[0]),
+                                     vreinterpret_s32_s16(CD.val[0]));
+  const int32x2x2_t tmp13 = vtrn_s32(vreinterpret_s32_s16(AB.val[1]),
+                                     vreinterpret_s32_s16(CD.val[1]));
+  *out01 = vreinterpretq_s16_s64(
+      vcombine_s64(vreinterpret_s64_s32(tmp02.val[0]),
+                   vreinterpret_s64_s32(tmp13.val[0])));
+  *out32 = vreinterpretq_s16_s64(
+      vcombine_s64(vreinterpret_s64_s32(tmp13.val[1]),
+                   vreinterpret_s64_s32(tmp02.val[1])));
+}
+
+static WEBP_INLINE int16x8_t DiffU8ToS16(const uint8x8_t a,
+                                         const uint8x8_t b) {
+  return vreinterpretq_s16_u16(vsubl_u8(a, b));
+}
+
+static void FTransform(const uint8_t* src, const uint8_t* ref,
+                       int16_t* out) {
+  int16x8_t d0d1, d3d2;   // working 4x4 int16 variables
+  {
+    const uint8x16_t S0 = Load4x4(src);
+    const uint8x16_t R0 = Load4x4(ref);
+    const int16x8_t D0D1 = DiffU8ToS16(vget_low_u8(S0), vget_low_u8(R0));
+    const int16x8_t D2D3 = DiffU8ToS16(vget_high_u8(S0), vget_high_u8(R0));
+    const int16x4_t D0 = vget_low_s16(D0D1);
+    const int16x4_t D1 = vget_high_s16(D0D1);
+    const int16x4_t D2 = vget_low_s16(D2D3);
+    const int16x4_t D3 = vget_high_s16(D2D3);
+    Transpose4x4_S16(D0, D1, D2, D3, &d0d1, &d3d2);
+  }
+  {    // 1rst pass
+    const int32x4_t kCst937 = vdupq_n_s32(937);
+    const int32x4_t kCst1812 = vdupq_n_s32(1812);
+    const int16x8_t a0a1 = vaddq_s16(d0d1, d3d2);   // d0+d3 | d1+d2   (=a0|a1)
+    const int16x8_t a3a2 = vsubq_s16(d0d1, d3d2);   // d0-d3 | d1-d2   (=a3|a2)
+    const int16x8_t a0a1_2 = vshlq_n_s16(a0a1, 3);
+    const int16x4_t tmp0 = vadd_s16(vget_low_s16(a0a1_2),
+                                    vget_high_s16(a0a1_2));
+    const int16x4_t tmp2 = vsub_s16(vget_low_s16(a0a1_2),
+                                    vget_high_s16(a0a1_2));
+    const int32x4_t a3_2217 = vmull_n_s16(vget_low_s16(a3a2), 2217);
+    const int32x4_t a2_2217 = vmull_n_s16(vget_high_s16(a3a2), 2217);
+    const int32x4_t a2_p_a3 = vmlal_n_s16(a2_2217, vget_low_s16(a3a2), 5352);
+    const int32x4_t a3_m_a2 = vmlsl_n_s16(a3_2217, vget_high_s16(a3a2), 5352);
+    const int16x4_t tmp1 = vshrn_n_s32(vaddq_s32(a2_p_a3, kCst1812), 9);
+    const int16x4_t tmp3 = vshrn_n_s32(vaddq_s32(a3_m_a2, kCst937), 9);
+    Transpose4x4_S16(tmp0, tmp1, tmp2, tmp3, &d0d1, &d3d2);
+  }
+  {    // 2nd pass
+    // the (1<<16) addition is for the replacement: a3!=0  <-> 1-(a3==0)
+    const int32x4_t kCst12000 = vdupq_n_s32(12000 + (1 << 16));
+    const int32x4_t kCst51000 = vdupq_n_s32(51000);
+    const int16x8_t a0a1 = vaddq_s16(d0d1, d3d2);   // d0+d3 | d1+d2   (=a0|a1)
+    const int16x8_t a3a2 = vsubq_s16(d0d1, d3d2);   // d0-d3 | d1-d2   (=a3|a2)
+    const int16x4_t a0_k7 = vadd_s16(vget_low_s16(a0a1), vdup_n_s16(7));
+    const int16x4_t out0 = vshr_n_s16(vadd_s16(a0_k7, vget_high_s16(a0a1)), 4);
+    const int16x4_t out2 = vshr_n_s16(vsub_s16(a0_k7, vget_high_s16(a0a1)), 4);
+    const int32x4_t a3_2217 = vmull_n_s16(vget_low_s16(a3a2), 2217);
+    const int32x4_t a2_2217 = vmull_n_s16(vget_high_s16(a3a2), 2217);
+    const int32x4_t a2_p_a3 = vmlal_n_s16(a2_2217, vget_low_s16(a3a2), 5352);
+    const int32x4_t a3_m_a2 = vmlsl_n_s16(a3_2217, vget_high_s16(a3a2), 5352);
+    const int16x4_t tmp1 = vaddhn_s32(a2_p_a3, kCst12000);
+    const int16x4_t out3 = vaddhn_s32(a3_m_a2, kCst51000);
+    const int16x4_t a3_eq_0 =
+        vreinterpret_s16_u16(vceq_s16(vget_low_s16(a3a2), vdup_n_s16(0)));
+    const int16x4_t out1 = vadd_s16(tmp1, a3_eq_0);
+    vst1_s16(out +  0, out0);
+    vst1_s16(out +  4, out1);
+    vst1_s16(out +  8, out2);
+    vst1_s16(out + 12, out3);
+  }
+}
+
+#else
+
 // adapted from vp8/encoder/arm/neon/shortfdct_neon.asm
 static const int16_t kCoeff16[] = {
   5352,  5352,  5352, 5352, 2217,  2217,  2217, 2217
@@ -339,69 +471,76 @@ static void FTransform(const uint8_t* src, const uint8_t* ref,
   );
 }
 
-static void FTransformWHT(const int16_t* in, int16_t* out) {
-  const int kStep = 32;
-  __asm__ volatile (
-    // d0 = in[0 * 16] , d1 = in[1 * 16]
-    // d2 = in[2 * 16] , d3 = in[3 * 16]
-    "vld1.16         d0[0], [%[in]], %[kStep]   \n"
-    "vld1.16         d1[0], [%[in]], %[kStep]   \n"
-    "vld1.16         d2[0], [%[in]], %[kStep]   \n"
-    "vld1.16         d3[0], [%[in]], %[kStep]   \n"
-    "vld1.16         d0[1], [%[in]], %[kStep]   \n"
-    "vld1.16         d1[1], [%[in]], %[kStep]   \n"
-    "vld1.16         d2[1], [%[in]], %[kStep]   \n"
-    "vld1.16         d3[1], [%[in]], %[kStep]   \n"
-    "vld1.16         d0[2], [%[in]], %[kStep]   \n"
-    "vld1.16         d1[2], [%[in]], %[kStep]   \n"
-    "vld1.16         d2[2], [%[in]], %[kStep]   \n"
-    "vld1.16         d3[2], [%[in]], %[kStep]   \n"
-    "vld1.16         d0[3], [%[in]], %[kStep]   \n"
-    "vld1.16         d1[3], [%[in]], %[kStep]   \n"
-    "vld1.16         d2[3], [%[in]], %[kStep]   \n"
-    "vld1.16         d3[3], [%[in]], %[kStep]   \n"
+#endif
 
-    "vaddl.s16       q2, d0, d2                 \n" // a0=(in[0*16]+in[2*16])
-    "vaddl.s16       q3, d1, d3                 \n" // a1=(in[1*16]+in[3*16])
-    "vsubl.s16       q4, d1, d3                 \n" // a2=(in[1*16]-in[3*16])
-    "vsubl.s16       q5, d0, d2                 \n" // a3=(in[0*16]-in[2*16])
+#define LOAD_LANE_16b(VALUE, LANE) do {             \
+  (VALUE) = vld1_lane_s16(src, (VALUE), (LANE));    \
+  src += stride;                                    \
+} while (0)
 
-    "vqadd.s32       q6, q2, q3                 \n" // a0 + a1
-    "vqadd.s32       q7, q5, q4                 \n" // a3 + a2
-    "vqsub.s32       q8, q5, q4                 \n" // a3 - a2
-    "vqsub.s32       q9, q2, q3                 \n" // a0 - a1
+static void FTransformWHT(const int16_t* src, int16_t* out) {
+  const int stride = 16;
+  const int16x4_t zero = vdup_n_s16(0);
+  int32x4x4_t tmp0;
+  int16x4x4_t in;
+  INIT_VECTOR4(in, zero, zero, zero, zero);
+  LOAD_LANE_16b(in.val[0], 0);
+  LOAD_LANE_16b(in.val[1], 0);
+  LOAD_LANE_16b(in.val[2], 0);
+  LOAD_LANE_16b(in.val[3], 0);
+  LOAD_LANE_16b(in.val[0], 1);
+  LOAD_LANE_16b(in.val[1], 1);
+  LOAD_LANE_16b(in.val[2], 1);
+  LOAD_LANE_16b(in.val[3], 1);
+  LOAD_LANE_16b(in.val[0], 2);
+  LOAD_LANE_16b(in.val[1], 2);
+  LOAD_LANE_16b(in.val[2], 2);
+  LOAD_LANE_16b(in.val[3], 2);
+  LOAD_LANE_16b(in.val[0], 3);
+  LOAD_LANE_16b(in.val[1], 3);
+  LOAD_LANE_16b(in.val[2], 3);
+  LOAD_LANE_16b(in.val[3], 3);
 
-    // Transpose
-    // q6 = tmp[0, 1,  2,  3] ; q7 = tmp[ 4,  5,  6,  7]
-    // q8 = tmp[8, 9, 10, 11] ; q9 = tmp[12, 13, 14, 15]
-    "vswp            d13, d16                   \n" // vtrn.64 q0, q2
-    "vswp            d15, d18                   \n" // vtrn.64 q1, q3
-    "vtrn.32         q6, q7                     \n"
-    "vtrn.32         q8, q9                     \n"
+  {
+    // a0 = in[0 * 16] + in[2 * 16]
+    // a1 = in[1 * 16] + in[3 * 16]
+    // a2 = in[1 * 16] - in[3 * 16]
+    // a3 = in[0 * 16] - in[2 * 16]
+    const int32x4_t a0 = vaddl_s16(in.val[0], in.val[2]);
+    const int32x4_t a1 = vaddl_s16(in.val[1], in.val[3]);
+    const int32x4_t a2 = vsubl_s16(in.val[1], in.val[3]);
+    const int32x4_t a3 = vsubl_s16(in.val[0], in.val[2]);
+    tmp0.val[0] = vaddq_s32(a0, a1);
+    tmp0.val[1] = vaddq_s32(a3, a2);
+    tmp0.val[2] = vsubq_s32(a3, a2);
+    tmp0.val[3] = vsubq_s32(a0, a1);
+  }
+  {
+    const int32x4x4_t tmp1 = Transpose4x4(tmp0);
+    // a0 = tmp[0 + i] + tmp[ 8 + i]
+    // a1 = tmp[4 + i] + tmp[12 + i]
+    // a2 = tmp[4 + i] - tmp[12 + i]
+    // a3 = tmp[0 + i] - tmp[ 8 + i]
+    const int32x4_t a0 = vaddq_s32(tmp1.val[0], tmp1.val[2]);
+    const int32x4_t a1 = vaddq_s32(tmp1.val[1], tmp1.val[3]);
+    const int32x4_t a2 = vsubq_s32(tmp1.val[1], tmp1.val[3]);
+    const int32x4_t a3 = vsubq_s32(tmp1.val[0], tmp1.val[2]);
+    const int32x4_t b0 = vhaddq_s32(a0, a1);  // (a0 + a1) >> 1
+    const int32x4_t b1 = vhaddq_s32(a3, a2);  // (a3 + a2) >> 1
+    const int32x4_t b2 = vhsubq_s32(a3, a2);  // (a3 - a2) >> 1
+    const int32x4_t b3 = vhsubq_s32(a0, a1);  // (a0 - a1) >> 1
+    const int16x4_t out0 = vmovn_s32(b0);
+    const int16x4_t out1 = vmovn_s32(b1);
+    const int16x4_t out2 = vmovn_s32(b2);
+    const int16x4_t out3 = vmovn_s32(b3);
 
-    "vqadd.s32       q0, q6, q8                 \n" // a0 = tmp[0] + tmp[8]
-    "vqadd.s32       q1, q7, q9                 \n" // a1 = tmp[4] + tmp[12]
-    "vqsub.s32       q2, q7, q9                 \n" // a2 = tmp[4] - tmp[12]
-    "vqsub.s32       q3, q6, q8                 \n" // a3 = tmp[0] - tmp[8]
-
-    "vqadd.s32       q4, q0, q1                 \n" // b0 = a0 + a1
-    "vqadd.s32       q5, q3, q2                 \n" // b1 = a3 + a2
-    "vqsub.s32       q6, q3, q2                 \n" // b2 = a3 - a2
-    "vqsub.s32       q7, q0, q1                 \n" // b3 = a0 - a1
-
-    "vshrn.s32       d18, q4, #1                \n" // b0 >> 1
-    "vshrn.s32       d19, q5, #1                \n" // b1 >> 1
-    "vshrn.s32       d20, q6, #1                \n" // b2 >> 1
-    "vshrn.s32       d21, q7, #1                \n" // b3 >> 1
-
-    "vst1.16         {q9, q10}, [%[out]]        \n"
-
-    : [in] "+r"(in)
-    : [kStep] "r"(kStep), [out] "r"(out)
-    : "memory", "q0", "q1", "q2", "q3", "q4", "q5",
-      "q6", "q7", "q8", "q9", "q10"       // clobbered
-  ) ;
+    vst1_s16(out +  0, out0);
+    vst1_s16(out +  4, out1);
+    vst1_s16(out +  8, out2);
+    vst1_s16(out + 12, out3);
+  }
 }
+#undef LOAD_LANE_16b
 
 //------------------------------------------------------------------------------
 // Texture distortion
@@ -409,9 +548,136 @@ static void FTransformWHT(const int16_t* in, int16_t* out) {
 // We try to match the spectral content (weighted) between source and
 // reconstructed samples.
 
+// This code works but is *slower* than the inlined-asm version below
+// (with gcc-4.6). So we disable it for now. Later, it'll be conditional to
+// USE_INTRINSICS define.
+// With gcc-4.8, it's only slightly slower than the inlined.
+#if defined(USE_INTRINSICS)
+
+// Zero extend an uint16x4_t 'v' to an int32x4_t.
+static WEBP_INLINE int32x4_t ConvertU16ToS32(uint16x4_t v) {
+  return vreinterpretq_s32_u32(vmovl_u16(v));
+}
+
+// Does a regular 4x4 transpose followed by an adjustment of the upper columns
+// in the inner rows to restore the source order of differences,
+// i.e., a0 - a1 | a3 - a2.
+static WEBP_INLINE int32x4x4_t DistoTranspose4x4(const int32x4x4_t rows) {
+  int32x4x4_t out = Transpose4x4(rows);
+  // restore source order in the columns containing differences.
+  const int32x2_t r1h = vget_high_s32(out.val[1]);
+  const int32x2_t r2h = vget_high_s32(out.val[2]);
+  out.val[1] = vcombine_s32(vget_low_s32(out.val[1]), r2h);
+  out.val[2] = vcombine_s32(vget_low_s32(out.val[2]), r1h);
+  return out;
+}
+
+static WEBP_INLINE int32x4x4_t DistoHorizontalPass(const uint8x8_t r0r1,
+                                                   const uint8x8_t r2r3) {
+  // a0 = in[0] + in[2] | a1 = in[1] + in[3]
+  const uint16x8_t a0a1 = vaddl_u8(r0r1, r2r3);
+  // a3 = in[0] - in[2] | a2 = in[1] - in[3]
+  const uint16x8_t a3a2 = vsubl_u8(r0r1, r2r3);
+  const int32x4_t tmp0 = vpaddlq_s16(vreinterpretq_s16_u16(a0a1));  // a0 + a1
+  const int32x4_t tmp1 = vpaddlq_s16(vreinterpretq_s16_u16(a3a2));  // a3 + a2
+  // no pairwise subtraction; reorder to perform tmp[2]/tmp[3] calculations.
+  // a0a0 a3a3 a0a0 a3a3 a0a0 a3a3 a0a0 a3a3
+  // a1a1 a2a2 a1a1 a2a2 a1a1 a2a2 a1a1 a2a2
+  const int16x8x2_t transpose =
+      vtrnq_s16(vreinterpretq_s16_u16(a0a1), vreinterpretq_s16_u16(a3a2));
+  // tmp[3] = a0 - a1 | tmp[2] = a3 - a2
+  const int32x4_t tmp32_1 = vsubl_s16(vget_low_s16(transpose.val[0]),
+                                      vget_low_s16(transpose.val[1]));
+  const int32x4_t tmp32_2 = vsubl_s16(vget_high_s16(transpose.val[0]),
+                                      vget_high_s16(transpose.val[1]));
+  // [0]: tmp[3] [1]: tmp[2]
+  const int32x4x2_t split = vtrnq_s32(tmp32_1, tmp32_2);
+  const int32x4x4_t res = { { tmp0, tmp1, split.val[1], split.val[0] } };
+  return res;
+}
+
+static WEBP_INLINE int32x4x4_t DistoVerticalPass(const int32x4x4_t rows) {
+  // a0 = tmp[0 + i] + tmp[8 + i];
+  const int32x4_t a0 = vaddq_s32(rows.val[0], rows.val[1]);
+  // a1 = tmp[4 + i] + tmp[12+ i];
+  const int32x4_t a1 = vaddq_s32(rows.val[2], rows.val[3]);
+  // a2 = tmp[4 + i] - tmp[12+ i];
+  const int32x4_t a2 = vsubq_s32(rows.val[2], rows.val[3]);
+  // a3 = tmp[0 + i] - tmp[8 + i];
+  const int32x4_t a3 = vsubq_s32(rows.val[0], rows.val[1]);
+  const int32x4_t b0 = vqabsq_s32(vaddq_s32(a0, a1));  // abs(a0 + a1)
+  const int32x4_t b1 = vqabsq_s32(vaddq_s32(a3, a2));  // abs(a3 + a2)
+  const int32x4_t b2 = vabdq_s32(a3, a2);              // abs(a3 - a2)
+  const int32x4_t b3 = vabdq_s32(a0, a1);              // abs(a0 - a1)
+  const int32x4x4_t res = { { b0, b1, b2, b3 } };
+  return res;
+}
+
+// Calculate the weighted sum of the rows in 'b'.
+static WEBP_INLINE int64x1_t DistoSum(const int32x4x4_t b,
+                                      const int32x4_t w0, const int32x4_t w1,
+                                      const int32x4_t w2, const int32x4_t w3) {
+  const int32x4_t s0 = vmulq_s32(w0, b.val[0]);
+  const int32x4_t s1 = vmlaq_s32(s0, w1, b.val[1]);
+  const int32x4_t s2 = vmlaq_s32(s1, w2, b.val[2]);
+  const int32x4_t s3 = vmlaq_s32(s2, w3, b.val[3]);
+  const int64x2_t sum1 = vpaddlq_s32(s3);
+  const int64x1_t sum2 = vadd_s64(vget_low_s64(sum1), vget_high_s64(sum1));
+  return sum2;
+}
+
+#define LOAD_LANE_32b(src, VALUE, LANE) \
+    (VALUE) = vld1q_lane_u32((const uint32_t*)(src), (VALUE), (LANE))
+
+// Hadamard transform
+// Returns the weighted sum of the absolute value of transformed coefficients.
+static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
+                    const uint16_t* const w) {
+  uint32x4_t d0d1 = { 0, 0, 0, 0 };
+  uint32x4_t d2d3 = { 0, 0, 0, 0 };
+  LOAD_LANE_32b(a + 0 * BPS, d0d1, 0);  // a00 a01 a02 a03
+  LOAD_LANE_32b(a + 1 * BPS, d0d1, 1);  // a10 a11 a12 a13
+  LOAD_LANE_32b(b + 0 * BPS, d0d1, 2);  // b00 b01 b02 b03
+  LOAD_LANE_32b(b + 1 * BPS, d0d1, 3);  // b10 b11 b12 b13
+  LOAD_LANE_32b(a + 2 * BPS, d2d3, 0);  // a20 a21 a22 a23
+  LOAD_LANE_32b(a + 3 * BPS, d2d3, 1);  // a30 a31 a32 a33
+  LOAD_LANE_32b(b + 2 * BPS, d2d3, 2);  // b20 b21 b22 b23
+  LOAD_LANE_32b(b + 3 * BPS, d2d3, 3);  // b30 b31 b32 b33
+
+  {
+    // a00 a01 a20 a21 a10 a11 a30 a31 b00 b01 b20 b21 b10 b11 b30 b31
+    // a02 a03 a22 a23 a12 a13 a32 a33 b02 b03 b22 b23 b12 b13 b32 b33
+    const uint16x8x2_t tmp =
+        vtrnq_u16(vreinterpretq_u16_u32(d0d1), vreinterpretq_u16_u32(d2d3));
+    const uint8x16_t d0d1u8 = vreinterpretq_u8_u16(tmp.val[0]);
+    const uint8x16_t d2d3u8 = vreinterpretq_u8_u16(tmp.val[1]);
+    const int32x4x4_t hpass_a = DistoHorizontalPass(vget_low_u8(d0d1u8),
+                                                    vget_low_u8(d2d3u8));
+    const int32x4x4_t hpass_b = DistoHorizontalPass(vget_high_u8(d0d1u8),
+                                                    vget_high_u8(d2d3u8));
+    const int32x4x4_t tmp_a = DistoTranspose4x4(hpass_a);
+    const int32x4x4_t tmp_b = DistoTranspose4x4(hpass_b);
+    const int32x4x4_t vpass_a = DistoVerticalPass(tmp_a);
+    const int32x4x4_t vpass_b = DistoVerticalPass(tmp_b);
+    const int32x4_t w0 = ConvertU16ToS32(vld1_u16(w + 0));
+    const int32x4_t w1 = ConvertU16ToS32(vld1_u16(w + 4));
+    const int32x4_t w2 = ConvertU16ToS32(vld1_u16(w + 8));
+    const int32x4_t w3 = ConvertU16ToS32(vld1_u16(w + 12));
+    const int64x1_t sum1 = DistoSum(vpass_a, w0, w1, w2, w3);
+    const int64x1_t sum2 = DistoSum(vpass_b, w0, w1, w2, w3);
+    const int32x2_t diff = vabd_s32(vreinterpret_s32_s64(sum1),
+                                    vreinterpret_s32_s64(sum2));
+    const int32x2_t res = vshr_n_s32(diff, 5);
+    return vget_lane_s32(res, 0);
+  }
+}
+
+#undef LOAD_LANE_32b
+
+#else
+
 // Hadamard transform
 // Returns the weighted sum of the absolute value of transformed coefficients.
-// This uses a TTransform helper function in C
 static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
                     const uint16_t* const w) {
   const int kBPS = BPS;
@@ -598,6 +864,8 @@ static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
   return sum;
 }
 
+#endif  // USE_INTRINSICS
+
 static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
                       const uint16_t* const w) {
   int D = 0;
@@ -610,6 +878,179 @@ static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
   return D;
 }
 
+//------------------------------------------------------------------------------
+
+static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
+                             int start_block, int end_block,
+                             VP8Histogram* const histo) {
+  const uint16x8_t max_coeff_thresh = vdupq_n_u16(MAX_COEFF_THRESH);
+  int j;
+  for (j = start_block; j < end_block; ++j) {
+    int16_t out[16];
+    FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
+    {
+      int k;
+      const int16x8_t a0 = vld1q_s16(out + 0);
+      const int16x8_t b0 = vld1q_s16(out + 8);
+      const uint16x8_t a1 = vreinterpretq_u16_s16(vabsq_s16(a0));
+      const uint16x8_t b1 = vreinterpretq_u16_s16(vabsq_s16(b0));
+      const uint16x8_t a2 = vshrq_n_u16(a1, 3);
+      const uint16x8_t b2 = vshrq_n_u16(b1, 3);
+      const uint16x8_t a3 = vminq_u16(a2, max_coeff_thresh);
+      const uint16x8_t b3 = vminq_u16(b2, max_coeff_thresh);
+      vst1q_s16(out + 0, vreinterpretq_s16_u16(a3));
+      vst1q_s16(out + 8, vreinterpretq_s16_u16(b3));
+      // Convert coefficients to bin.
+      for (k = 0; k < 16; ++k) {
+        histo->distribution[out[k]]++;
+      }
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+
+static WEBP_INLINE void AccumulateSSE16(const uint8_t* const a,
+                                        const uint8_t* const b,
+                                        uint32x4_t* const sum) {
+  const uint8x16_t a0 = vld1q_u8(a);
+  const uint8x16_t b0 = vld1q_u8(b);
+  const uint8x16_t abs_diff = vabdq_u8(a0, b0);
+  uint16x8_t prod = vmull_u8(vget_low_u8(abs_diff), vget_low_u8(abs_diff));
+  prod = vmlal_u8(prod, vget_high_u8(abs_diff), vget_high_u8(abs_diff));
+  *sum = vpadalq_u16(*sum, prod);      // pair-wise add and accumulate
+}
+
+// Horizontal sum of all four uint32_t values in 'sum'.
+static int SumToInt(uint32x4_t sum) {
+  const uint64x2_t sum2 = vpaddlq_u32(sum);
+  const uint64_t sum3 = vgetq_lane_u64(sum2, 0) + vgetq_lane_u64(sum2, 1);
+  return (int)sum3;
+}
+
+static int SSE16x16(const uint8_t* a, const uint8_t* b) {
+  uint32x4_t sum = vdupq_n_u32(0);
+  int y;
+  for (y = 0; y < 16; ++y) {
+    AccumulateSSE16(a + y * BPS, b + y * BPS, &sum);
+  }
+  return SumToInt(sum);
+}
+
+static int SSE16x8(const uint8_t* a, const uint8_t* b) {
+  uint32x4_t sum = vdupq_n_u32(0);
+  int y;
+  for (y = 0; y < 8; ++y) {
+    AccumulateSSE16(a + y * BPS, b + y * BPS, &sum);
+  }
+  return SumToInt(sum);
+}
+
+static int SSE8x8(const uint8_t* a, const uint8_t* b) {
+  uint32x4_t sum = vdupq_n_u32(0);
+  int y;
+  for (y = 0; y < 8; ++y) {
+    const uint8x8_t a0 = vld1_u8(a + y * BPS);
+    const uint8x8_t b0 = vld1_u8(b + y * BPS);
+    const uint8x8_t abs_diff = vabd_u8(a0, b0);
+    const uint16x8_t prod = vmull_u8(abs_diff, abs_diff);
+    sum = vpadalq_u16(sum, prod);
+  }
+  return SumToInt(sum);
+}
+
+static int SSE4x4(const uint8_t* a, const uint8_t* b) {
+  const uint8x16_t a0 = Load4x4(a);
+  const uint8x16_t b0 = Load4x4(b);
+  const uint8x16_t abs_diff = vabdq_u8(a0, b0);
+  uint16x8_t prod = vmull_u8(vget_low_u8(abs_diff), vget_low_u8(abs_diff));
+  prod = vmlal_u8(prod, vget_high_u8(abs_diff), vget_high_u8(abs_diff));
+  return SumToInt(vpaddlq_u16(prod));
+}
+
+//------------------------------------------------------------------------------
+
+// Compilation with gcc-4.6.x is problematic for now.
+#if !defined(WORK_AROUND_GCC)
+
+static int16x8_t Quantize(int16_t* const in,
+                          const VP8Matrix* const mtx, int offset) {
+  const uint16x8_t sharp = vld1q_u16(&mtx->sharpen_[offset]);
+  const uint16x8_t q = vld1q_u16(&mtx->q_[offset]);
+  const uint16x8_t iq = vld1q_u16(&mtx->iq_[offset]);
+  const uint32x4_t bias0 = vld1q_u32(&mtx->bias_[offset + 0]);
+  const uint32x4_t bias1 = vld1q_u32(&mtx->bias_[offset + 4]);
+
+  const int16x8_t a = vld1q_s16(in + offset);                // in
+  const uint16x8_t b = vreinterpretq_u16_s16(vabsq_s16(a));  // coeff = abs(in)
+  const int16x8_t sign = vshrq_n_s16(a, 15);                 // sign
+  const uint16x8_t c = vaddq_u16(b, sharp);                  // + sharpen
+  const uint32x4_t m0 = vmull_u16(vget_low_u16(c), vget_low_u16(iq));
+  const uint32x4_t m1 = vmull_u16(vget_high_u16(c), vget_high_u16(iq));
+  const uint32x4_t m2 = vhaddq_u32(m0, bias0);
+  const uint32x4_t m3 = vhaddq_u32(m1, bias1);     // (coeff * iQ + bias) >> 1
+  const uint16x8_t c0 = vcombine_u16(vshrn_n_u32(m2, 16),
+                                     vshrn_n_u32(m3, 16));   // QFIX=17 = 16+1
+  const uint16x8_t c1 = vminq_u16(c0, vdupq_n_u16(MAX_LEVEL));
+  const int16x8_t c2 = veorq_s16(vreinterpretq_s16_u16(c1), sign);
+  const int16x8_t c3 = vsubq_s16(c2, sign);                  // restore sign
+  const int16x8_t c4 = vmulq_s16(c3, vreinterpretq_s16_u16(q));
+  vst1q_s16(in + offset, c4);
+  assert(QFIX == 17);  // this function can't work as is if QFIX != 16+1
+  return c3;
+}
+
+static const uint8_t kShuffles[4][8] = {
+  { 0,   1,  2,  3,  8,  9, 16, 17 },
+  { 10, 11,  4,  5,  6,  7, 12, 13 },
+  { 18, 19, 24, 25, 26, 27, 20, 21 },
+  { 14, 15, 22, 23, 28, 29, 30, 31 }
+};
+
+static int QuantizeBlock(int16_t in[16], int16_t out[16],
+                         const VP8Matrix* const mtx) {
+  const int16x8_t out0 = Quantize(in, mtx, 0);
+  const int16x8_t out1 = Quantize(in, mtx, 8);
+  uint8x8x4_t shuffles;
+  // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
+  // non-standard versions there.
+#if defined(__APPLE__) && defined(__aarch64__) && \
+    defined(__apple_build_version__) && (__apple_build_version__< 6020037)
+  uint8x16x2_t all_out;
+  INIT_VECTOR2(all_out, vreinterpretq_u8_s16(out0), vreinterpretq_u8_s16(out1));
+  INIT_VECTOR4(shuffles,
+               vtbl2q_u8(all_out, vld1_u8(kShuffles[0])),
+               vtbl2q_u8(all_out, vld1_u8(kShuffles[1])),
+               vtbl2q_u8(all_out, vld1_u8(kShuffles[2])),
+               vtbl2q_u8(all_out, vld1_u8(kShuffles[3])));
+#else
+  uint8x8x4_t all_out;
+  INIT_VECTOR4(all_out,
+               vreinterpret_u8_s16(vget_low_s16(out0)),
+               vreinterpret_u8_s16(vget_high_s16(out0)),
+               vreinterpret_u8_s16(vget_low_s16(out1)),
+               vreinterpret_u8_s16(vget_high_s16(out1)));
+  INIT_VECTOR4(shuffles,
+               vtbl4_u8(all_out, vld1_u8(kShuffles[0])),
+               vtbl4_u8(all_out, vld1_u8(kShuffles[1])),
+               vtbl4_u8(all_out, vld1_u8(kShuffles[2])),
+               vtbl4_u8(all_out, vld1_u8(kShuffles[3])));
+#endif
+  // Zigzag reordering
+  vst1_u8((uint8_t*)(out +  0), shuffles.val[0]);
+  vst1_u8((uint8_t*)(out +  4), shuffles.val[1]);
+  vst1_u8((uint8_t*)(out +  8), shuffles.val[2]);
+  vst1_u8((uint8_t*)(out + 12), shuffles.val[3]);
+  // test zeros
+  if (*(uint64_t*)(out +  0) != 0) return 1;
+  if (*(uint64_t*)(out +  4) != 0) return 1;
+  if (*(uint64_t*)(out +  8) != 0) return 1;
+  if (*(uint64_t*)(out + 12) != 0) return 1;
+  return 0;
+}
+
+#endif   // !WORK_AROUND_GCC
+
 #endif   // WEBP_USE_NEON
 
 //------------------------------------------------------------------------------
@@ -622,11 +1063,17 @@ void VP8EncDspInitNEON(void) {
   VP8ITransform = ITransform;
   VP8FTransform = FTransform;
 
-  VP8ITransformWHT = ITransformWHT;
   VP8FTransformWHT = FTransformWHT;
 
   VP8TDisto4x4 = Disto4x4;
   VP8TDisto16x16 = Disto16x16;
+  VP8CollectHistogram = CollectHistogram;
+  VP8SSE16x16 = SSE16x16;
+  VP8SSE16x8 = SSE16x8;
+  VP8SSE8x8 = SSE8x8;
+  VP8SSE4x4 = SSE4x4;
+#if !defined(WORK_AROUND_GCC)
+  VP8EncQuantizeBlock = QuantizeBlock;
+#endif
 #endif   // WEBP_USE_NEON
 }
-

src/dsp/enc_sse2.c

@@ -17,7 +17,9 @@
 #include <stdlib.h>  // for abs()
 #include <emmintrin.h>
 
+#include "../enc/cost.h"
 #include "../enc/vp8enci.h"
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 // Quite useful macro for debugging. Left here for convenience.
@@ -52,9 +54,9 @@ static void PrintReg(const __m128i r, const char* const name, int size) {
 // Compute susceptibility based on DCT-coeff histograms:
 // the higher, the "easier" the macroblock is to compress.
 
-static void CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
-                                 int start_block, int end_block,
-                                 VP8Histogram* const histo) {
+static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
+                             int start_block, int end_block,
+                             VP8Histogram* const histo) {
   const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
   int j;
   for (j = start_block; j < end_block; ++j) {
@@ -98,8 +100,8 @@ static void CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
 // Transforms (Paragraph 14.4)
 
 // Does one or two inverse transforms.
-static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
-                           int do_two) {
+static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
+                       int do_two) {
   // This implementation makes use of 16-bit fixed point versions of two
   // multiply constants:
   //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@@ -318,8 +320,7 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
   }
 }
 
-static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
-                           int16_t* out) {
+static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i seven = _mm_set1_epi16(7);
   const __m128i k937 = _mm_set1_epi32(937);
@@ -444,14 +445,14 @@ static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
     // -> f1 = f1 + 1 - (a3 == 0)
     const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
 
-    _mm_storel_epi64((__m128i*)&out[ 0], d0);
-    _mm_storel_epi64((__m128i*)&out[ 4], g1);
-    _mm_storel_epi64((__m128i*)&out[ 8], d2);
-    _mm_storel_epi64((__m128i*)&out[12], f3);
+    const __m128i d0_g1 = _mm_unpacklo_epi64(d0, g1);
+    const __m128i d2_f3 = _mm_unpacklo_epi64(d2, f3);
+    _mm_storeu_si128((__m128i*)&out[0], d0_g1);
+    _mm_storeu_si128((__m128i*)&out[8], d2_f3);
   }
 }
 
-static void FTransformWHTSSE2(const int16_t* in, int16_t* out) {
+static void FTransformWHT(const int16_t* in, int16_t* out) {
   int32_t tmp[16];
   int i;
   for (i = 0; i < 4; ++i, in += 64) {
@@ -487,8 +488,8 @@ static void FTransformWHTSSE2(const int16_t* in, int16_t* out) {
 //------------------------------------------------------------------------------
 // Metric
 
-static int SSE_Nx4SSE2(const uint8_t* a, const uint8_t* b,
-                       int num_quads, int do_16) {
+static int SSE_Nx4(const uint8_t* a, const uint8_t* b,
+                   int num_quads, int do_16) {
   const __m128i zero = _mm_setzero_si128();
   __m128i sum1 = zero;
   __m128i sum2 = zero;
@@ -565,19 +566,19 @@ static int SSE_Nx4SSE2(const uint8_t* a, const uint8_t* b,
   }
 }
 
-static int SSE16x16SSE2(const uint8_t* a, const uint8_t* b) {
-  return SSE_Nx4SSE2(a, b, 4, 1);
+static int SSE16x16(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4(a, b, 4, 1);
 }
 
-static int SSE16x8SSE2(const uint8_t* a, const uint8_t* b) {
-  return SSE_Nx4SSE2(a, b, 2, 1);
+static int SSE16x8(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4(a, b, 2, 1);
 }
 
-static int SSE8x8SSE2(const uint8_t* a, const uint8_t* b) {
-  return SSE_Nx4SSE2(a, b, 2, 0);
+static int SSE8x8(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4(a, b, 2, 0);
 }
 
-static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) {
+static int SSE4x4(const uint8_t* a, const uint8_t* b) {
   const __m128i zero = _mm_setzero_si128();
 
   // Load values. Note that we read 8 pixels instead of 4,
@@ -634,8 +635,8 @@ static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) {
 // Hadamard transform
 // Returns the difference between the weighted sum of the absolute value of
 // transformed coefficients.
-static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
-                          const uint16_t* const w) {
+static int TTransform(const uint8_t* inA, const uint8_t* inB,
+                      const uint16_t* const w) {
   int32_t sum[4];
   __m128i tmp_0, tmp_1, tmp_2, tmp_3;
   const __m128i zero = _mm_setzero_si128();
@@ -782,19 +783,19 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
   return sum[0] + sum[1] + sum[2] + sum[3];
 }
 
-static int Disto4x4SSE2(const uint8_t* const a, const uint8_t* const b,
-                        const uint16_t* const w) {
-  const int diff_sum = TTransformSSE2(a, b, w);
+static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
+                    const uint16_t* const w) {
+  const int diff_sum = TTransform(a, b, w);
   return abs(diff_sum) >> 5;
 }
 
-static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b,
-                          const uint16_t* const w) {
+static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
+                      const uint16_t* const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
     for (x = 0; x < 16; x += 4) {
-      D += Disto4x4SSE2(a + x + y, b + x + y, w);
+      D += Disto4x4(a + x + y, b + x + y, w);
     }
   }
   return D;
@@ -804,9 +805,9 @@ static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b,
 // Quantization
 //
 
-// Simple quantization
-static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
-                             int n, const VP8Matrix* const mtx) {
+static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16],
+                                       const uint16_t* const sharpen,
+                                       const VP8Matrix* const mtx) {
   const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
   const __m128i zero = _mm_setzero_si128();
   __m128i coeff0, coeff8;
@@ -818,18 +819,14 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   //                  we can use _mm_load_si128 instead of _mm_loadu_si128.
   __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
   __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
-  const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[0]);
-  const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[8]);
   const __m128i iq0 = _mm_loadu_si128((__m128i*)&mtx->iq_[0]);
   const __m128i iq8 = _mm_loadu_si128((__m128i*)&mtx->iq_[8]);
-  const __m128i bias0 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]);
-  const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]);
   const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]);
   const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]);
 
-  // sign(in) = in >> 15  (0x0000 if positive, 0xffff if negative)
-  const __m128i sign0 = _mm_srai_epi16(in0, 15);
-  const __m128i sign8 = _mm_srai_epi16(in8, 15);
+  // extract sign(in)  (0x0000 if positive, 0xffff if negative)
+  const __m128i sign0 = _mm_cmpgt_epi16(zero, in0);
+  const __m128i sign8 = _mm_cmpgt_epi16(zero, in8);
 
   // coeff = abs(in) = (in ^ sign) - sign
   coeff0 = _mm_xor_si128(in0, sign0);
@@ -838,32 +835,35 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   coeff8 = _mm_sub_epi16(coeff8, sign8);
 
   // coeff = abs(in) + sharpen
-  coeff0 = _mm_add_epi16(coeff0, sharpen0);
-  coeff8 = _mm_add_epi16(coeff8, sharpen8);
+  if (sharpen != NULL) {
+    const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&sharpen[0]);
+    const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&sharpen[8]);
+    coeff0 = _mm_add_epi16(coeff0, sharpen0);
+    coeff8 = _mm_add_epi16(coeff8, sharpen8);
+  }
 
-  // out = (coeff * iQ + B) >> QFIX;
+  // out = (coeff * iQ + B) >> QFIX
   {
     // doing calculations with 32b precision (QFIX=17)
     // out = (coeff * iQ)
-    __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
-    __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
-    __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
-    __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
+    const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
+    const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
+    const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
+    const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
     __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
     __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
     __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
     __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
-    // expand bias from 16b to 32b
-    __m128i bias_00 = _mm_unpacklo_epi16(bias0, zero);
-    __m128i bias_04 = _mm_unpackhi_epi16(bias0, zero);
-    __m128i bias_08 = _mm_unpacklo_epi16(bias8, zero);
-    __m128i bias_12 = _mm_unpackhi_epi16(bias8, zero);
     // out = (coeff * iQ + B)
+    const __m128i bias_00 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]);
+    const __m128i bias_04 = _mm_loadu_si128((__m128i*)&mtx->bias_[4]);
+    const __m128i bias_08 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]);
+    const __m128i bias_12 = _mm_loadu_si128((__m128i*)&mtx->bias_[12]);
     out_00 = _mm_add_epi32(out_00, bias_00);
     out_04 = _mm_add_epi32(out_04, bias_04);
     out_08 = _mm_add_epi32(out_08, bias_08);
     out_12 = _mm_add_epi32(out_12, bias_12);
-    // out = (coeff * iQ + B) >> QFIX;
+    // out = QUANTDIV(coeff, iQ, B, QFIX)
     out_00 = _mm_srai_epi32(out_00, QFIX);
     out_04 = _mm_srai_epi32(out_04, QFIX);
     out_08 = _mm_srai_epi32(out_08, QFIX);
@@ -916,19 +916,44 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   }
 
   // detect if all 'out' values are zeroes or not
-  {
-    int32_t tmp[4];
-    _mm_storeu_si128((__m128i*)tmp, packed_out);
-    if (n) {
-      tmp[0] &= ~0xff;
-    }
-    return (tmp[3] || tmp[2] || tmp[1] || tmp[0]);
-  }
+  return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
 }
 
-static int QuantizeBlockWHTSSE2(int16_t in[16], int16_t out[16],
-                                const VP8Matrix* const mtx) {
-  return QuantizeBlockSSE2(in, out, 0, mtx);
+static int QuantizeBlock(int16_t in[16], int16_t out[16],
+                         const VP8Matrix* const mtx) {
+  return DoQuantizeBlock(in, out, &mtx->sharpen_[0], mtx);
+}
+
+static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
+                            const VP8Matrix* const mtx) {
+  return DoQuantizeBlock(in, out, NULL, mtx);
+}
+
+// Forward declaration.
+void VP8SetResidualCoeffsSSE2(const int16_t* const coeffs,
+                              VP8Residual* const res);
+
+void VP8SetResidualCoeffsSSE2(const int16_t* const coeffs,
+                              VP8Residual* const res) {
+  const __m128i c0 = _mm_loadu_si128((const __m128i*)coeffs);
+  const __m128i c1 = _mm_loadu_si128((const __m128i*)(coeffs + 8));
+  // Use SSE to compare 8 values with a single instruction.
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i m0 = _mm_cmpeq_epi16(c0, zero);
+  const __m128i m1 = _mm_cmpeq_epi16(c1, zero);
+  // Get the comparison results as a bitmask, consisting of two times 16 bits:
+  // two identical bits for each result. Concatenate both bitmasks to get a
+  // single 32 bit value. Negate the mask to get the position of entries that
+  // are not equal to zero. We don't need to mask out least significant bits
+  // according to res->first, since coeffs[0] is 0 if res->first > 0
+  const uint32_t mask =
+      ~(((uint32_t)_mm_movemask_epi8(m1) << 16) | _mm_movemask_epi8(m0));
+  // The position of the most significant non-zero bit indicates the position of
+  // the last non-zero value. Divide the result by two because __movemask_epi8
+  // operates on 8 bit values instead of 16 bit values.
+  assert(res->first == 0 || coeffs[0] == 0);
+  res->last = mask ? (BitsLog2Floor(mask) >> 1) : -1;
+  res->coeffs = coeffs;
 }
 
 #endif   // WEBP_USE_SSE2
@@ -940,18 +965,18 @@ extern void VP8EncDspInitSSE2(void);
 
 void VP8EncDspInitSSE2(void) {
 #if defined(WEBP_USE_SSE2)
-  VP8CollectHistogram = CollectHistogramSSE2;
-  VP8EncQuantizeBlock = QuantizeBlockSSE2;
-  VP8EncQuantizeBlockWHT = QuantizeBlockWHTSSE2;
-  VP8ITransform = ITransformSSE2;
-  VP8FTransform = FTransformSSE2;
-  VP8FTransformWHT = FTransformWHTSSE2;
-  VP8SSE16x16 = SSE16x16SSE2;
-  VP8SSE16x8 = SSE16x8SSE2;
-  VP8SSE8x8 = SSE8x8SSE2;
-  VP8SSE4x4 = SSE4x4SSE2;
-  VP8TDisto4x4 = Disto4x4SSE2;
-  VP8TDisto16x16 = Disto16x16SSE2;
+  VP8CollectHistogram = CollectHistogram;
+  VP8EncQuantizeBlock = QuantizeBlock;
+  VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
+  VP8ITransform = ITransform;
+  VP8FTransform = FTransform;
+  VP8FTransformWHT = FTransformWHT;
+  VP8SSE16x16 = SSE16x16;
+  VP8SSE16x8 = SSE16x8;
+  VP8SSE8x8 = SSE8x8;
+  VP8SSE4x4 = SSE4x4;
+  VP8TDisto4x4 = Disto4x4;
+  VP8TDisto16x16 = Disto16x16;
 #endif   // WEBP_USE_SSE2
 }
 

src/dsp/lossless.h

@@ -18,26 +18,58 @@
 #include "../webp/types.h"
 #include "../webp/decode.h"
 
+#include "../enc/histogram.h"
+#include "../utils/utils.h"
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 //------------------------------------------------------------------------------
-//
+// Signatures and generic function-pointers
 
-typedef uint32_t (*VP8LPredClampedAddSubFunc)(uint32_t c0, uint32_t c1,
-                                              uint32_t c2);
-typedef uint32_t (*VP8LPredSelectFunc)(uint32_t c0, uint32_t c1, uint32_t c2);
-typedef void (*VP8LSubtractGreenFromBlueAndRedFunc)(uint32_t* argb_data,
-                                                    int num_pixs);
-typedef void (*VP8LAddGreenToBlueAndRedFunc)(uint32_t* data_start,
-                                             const uint32_t* data_end);
+typedef uint32_t (*VP8LPredictorFunc)(uint32_t left, const uint32_t* const top);
+extern VP8LPredictorFunc VP8LPredictors[16];
 
-extern VP8LPredClampedAddSubFunc VP8LClampedAddSubtractFull;
-extern VP8LPredClampedAddSubFunc VP8LClampedAddSubtractHalf;
-extern VP8LPredSelectFunc VP8LSelect;
-extern VP8LSubtractGreenFromBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
-extern VP8LAddGreenToBlueAndRedFunc VP8LAddGreenToBlueAndRed;
+typedef void (*VP8LProcessBlueAndRedFunc)(uint32_t* argb_data, int num_pixels);
+extern VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
+extern VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed;
+
+typedef struct {
+  // Note: the members are uint8_t, so that any negative values are
+  // automatically converted to "mod 256" values.
+  uint8_t green_to_red_;
+  uint8_t green_to_blue_;
+  uint8_t red_to_blue_;
+} VP8LMultipliers;
+typedef void (*VP8LTransformColorFunc)(const VP8LMultipliers* const m,
+                                       uint32_t* argb_data, int num_pixels);
+extern VP8LTransformColorFunc VP8LTransformColor;
+extern VP8LTransformColorFunc VP8LTransformColorInverse;
+
+typedef void (*VP8LConvertFunc)(const uint32_t* src, int num_pixels,
+                                uint8_t* dst);
+extern VP8LConvertFunc VP8LConvertBGRAToRGB;
+extern VP8LConvertFunc VP8LConvertBGRAToRGBA;
+extern VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
+extern VP8LConvertFunc VP8LConvertBGRAToRGB565;
+extern VP8LConvertFunc VP8LConvertBGRAToBGR;
+
+// Expose some C-only fallback functions
+void VP8LTransformColor_C(const VP8LMultipliers* const m,
+                          uint32_t* data, int num_pixels);
+void VP8LTransformColorInverse_C(const VP8LMultipliers* const m,
+                                 uint32_t* data, int num_pixels);
+
+void VP8LConvertBGRAToRGB_C(const uint32_t* src, int num_pixels, uint8_t* dst);
+void VP8LConvertBGRAToRGBA_C(const uint32_t* src, int num_pixels, uint8_t* dst);
+void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src,
+                                 int num_pixels, uint8_t* dst);
+void VP8LConvertBGRAToRGB565_C(const uint32_t* src,
+                               int num_pixels, uint8_t* dst);
+void VP8LConvertBGRAToBGR_C(const uint32_t* src, int num_pixels, uint8_t* dst);
+void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels);
+void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels);
 
 // Must be called before calling any of the above methods.
 void VP8LDspInit(void);
@@ -66,7 +98,7 @@ void VP8LResidualImage(int width, int height, int bits,
                        uint32_t* const argb, uint32_t* const argb_scratch,
                        uint32_t* const image);
 
-void VP8LColorSpaceTransform(int width, int height, int bits, int step,
+void VP8LColorSpaceTransform(int width, int height, int bits, int quality,
                              uint32_t* const argb, uint32_t* image);
 
 //------------------------------------------------------------------------------
@@ -85,58 +117,55 @@ static WEBP_INLINE uint32_t VP8LSubSampleSize(uint32_t size,
   return (size + (1 << sampling_bits) - 1) >> sampling_bits;
 }
 
+// -----------------------------------------------------------------------------
 // Faster logarithm for integers. Small values use a look-up table.
 #define LOG_LOOKUP_IDX_MAX 256
 extern const float kLog2Table[LOG_LOOKUP_IDX_MAX];
 extern const float kSLog2Table[LOG_LOOKUP_IDX_MAX];
-float VP8LFastLog2Slow(int v);
-float VP8LFastSLog2Slow(int v);
-static WEBP_INLINE float VP8LFastLog2(int v) {
+typedef float (*VP8LFastLog2SlowFunc)(uint32_t v);
+
+extern VP8LFastLog2SlowFunc VP8LFastLog2Slow;
+extern VP8LFastLog2SlowFunc VP8LFastSLog2Slow;
+
+static WEBP_INLINE float VP8LFastLog2(uint32_t v) {
   return (v < LOG_LOOKUP_IDX_MAX) ? kLog2Table[v] : VP8LFastLog2Slow(v);
 }
 // Fast calculation of v * log2(v) for integer input.
-static WEBP_INLINE float VP8LFastSLog2(int v) {
+static WEBP_INLINE float VP8LFastSLog2(uint32_t v) {
   return (v < LOG_LOOKUP_IDX_MAX) ? kSLog2Table[v] : VP8LFastSLog2Slow(v);
 }
 
+// -----------------------------------------------------------------------------
+// Huffman-cost related functions.
+
+typedef double (*VP8LCostFunc)(const uint32_t* population, int length);
+typedef double (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y,
+                                       int length);
+
+extern VP8LCostFunc VP8LExtraCost;
+extern VP8LCostCombinedFunc VP8LExtraCostCombined;
+
+typedef struct {        // small struct to hold counters
+  int counts[2];        // index: 0=zero steak, 1=non-zero streak
+  int streaks[2][2];    // [zero/non-zero][streak<3 / streak>=3]
+} VP8LStreaks;
+
+typedef VP8LStreaks (*VP8LCostCountFunc)(const uint32_t* population,
+                                         int length);
+typedef VP8LStreaks (*VP8LCostCombinedCountFunc)(const uint32_t* X,
+                                                 const uint32_t* Y, int length);
+
+extern VP8LCostCountFunc VP8LHuffmanCostCount;
+extern VP8LCostCombinedCountFunc VP8LHuffmanCostCombinedCount;
+
+typedef void (*VP8LHistogramAddFunc)(const VP8LHistogram* const a,
+                                     const VP8LHistogram* const b,
+                                     VP8LHistogram* const out);
+extern VP8LHistogramAddFunc VP8LHistogramAdd;
+
 // -----------------------------------------------------------------------------
 // PrefixEncode()
 
-// use GNU builtins where available.
-#if defined(__GNUC__) && \
-    ((__GNUC__ == 3 && __GNUC_MINOR__ >= 4) || __GNUC__ >= 4)
-static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
-  return 31 ^ __builtin_clz(n);
-}
-#elif defined(_MSC_VER) && _MSC_VER > 1310 && \
-      (defined(_M_X64) || defined(_M_IX86))
-#include <intrin.h>
-#pragma intrinsic(_BitScanReverse)
-
-static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
-  unsigned long first_set_bit;
-  _BitScanReverse(&first_set_bit, n);
-  return first_set_bit;
-}
-#else
-// Returns (int)floor(log2(n)). n must be > 0.
-static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
-  int log = 0;
-  uint32_t value = n;
-  int i;
-
-  for (i = 4; i >= 0; --i) {
-    const int shift = (1 << i);
-    const uint32_t x = value >> shift;
-    if (x != 0) {
-      value = x;
-      log += shift;
-    }
-  }
-  return log;
-}
-#endif
-
 static WEBP_INLINE int VP8LBitsLog2Ceiling(uint32_t n) {
   const int log_floor = BitsLog2Floor(n);
   if (n == (n & ~(n - 1)))  // zero or a power of two.

src/dsp/lossless_mips32.c

@@ -0,0 +1,416 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// MIPS version of lossless functions
+//
+// Author(s):  Djordje Pesut    (djordje.pesut@imgtec.com)
+//             Jovan Zelincevic (jovan.zelincevic@imgtec.com)
+
+#include "./dsp.h"
+#include "./lossless.h"
+
+#if defined(WEBP_USE_MIPS32)
+
+#include <assert.h>
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define APPROX_LOG_WITH_CORRECTION_MAX  65536
+#define APPROX_LOG_MAX                   4096
+#define LOG_2_RECIPROCAL 1.44269504088896338700465094007086
+
+static float FastSLog2Slow(uint32_t v) {
+  assert(v >= LOG_LOOKUP_IDX_MAX);
+  if (v < APPROX_LOG_WITH_CORRECTION_MAX) {
+    uint32_t log_cnt, y, correction;
+    const int c24 = 24;
+    const float v_f = (float)v;
+    uint32_t temp;
+
+    // Xf = 256 = 2^8
+    // log_cnt is index of leading one in upper 24 bits
+    __asm__ volatile(
+      "clz      %[log_cnt], %[v]                      \n\t"
+      "addiu    %[y],       $zero,        1           \n\t"
+      "subu     %[log_cnt], %[c24],       %[log_cnt]  \n\t"
+      "sllv     %[y],       %[y],         %[log_cnt]  \n\t"
+      "srlv     %[temp],    %[v],         %[log_cnt]  \n\t"
+      : [log_cnt]"=&r"(log_cnt), [y]"=&r"(y),
+        [temp]"=r"(temp)
+      : [c24]"r"(c24), [v]"r"(v)
+    );
+
+    // vf = (2^log_cnt) * Xf; where y = 2^log_cnt and Xf < 256
+    // Xf = floor(Xf) * (1 + (v % y) / v)
+    // log2(Xf) = log2(floor(Xf)) + log2(1 + (v % y) / v)
+    // The correction factor: log(1 + d) ~ d; for very small d values, so
+    // log2(1 + (v % y) / v) ~ LOG_2_RECIPROCAL * (v % y)/v
+    // LOG_2_RECIPROCAL ~ 23/16
+
+    // (v % y) = (v % 2^log_cnt) = v & (2^log_cnt - 1)
+    correction = (23 * (v & (y - 1))) >> 4;
+    return v_f * (kLog2Table[temp] + log_cnt) + correction;
+  } else {
+    return (float)(LOG_2_RECIPROCAL * v * log((double)v));
+  }
+}
+
+static float FastLog2Slow(uint32_t v) {
+  assert(v >= LOG_LOOKUP_IDX_MAX);
+  if (v < APPROX_LOG_WITH_CORRECTION_MAX) {
+    uint32_t log_cnt, y;
+    const int c24 = 24;
+    double log_2;
+    uint32_t temp;
+
+    __asm__ volatile(
+      "clz      %[log_cnt], %[v]                      \n\t"
+      "addiu    %[y],       $zero,        1           \n\t"
+      "subu     %[log_cnt], %[c24],       %[log_cnt]  \n\t"
+      "sllv     %[y],       %[y],         %[log_cnt]  \n\t"
+      "srlv     %[temp],    %[v],         %[log_cnt]  \n\t"
+      : [log_cnt]"=&r"(log_cnt), [y]"=&r"(y),
+        [temp]"=r"(temp)
+      : [c24]"r"(c24), [v]"r"(v)
+    );
+
+    log_2 = kLog2Table[temp] + log_cnt;
+    if (v >= APPROX_LOG_MAX) {
+      // Since the division is still expensive, add this correction factor only
+      // for large values of 'v'.
+
+      const uint32_t correction = (23 * (v & (y - 1))) >> 4;
+      log_2 += (double)correction / v;
+    }
+    return (float)log_2;
+  } else {
+    return (float)(LOG_2_RECIPROCAL * log((double)v));
+  }
+}
+
+// C version of this function:
+//   int i = 0;
+//   int64_t cost = 0;
+//   const uint32_t* pop = &population[4];
+//   const uint32_t* LoopEnd = &population[length];
+//   while (pop != LoopEnd) {
+//     ++i;
+//     cost += i * *pop;
+//     cost += i * *(pop + 1);
+//     pop += 2;
+//   }
+//   return (double)cost;
+static double ExtraCost(const uint32_t* const population, int length) {
+  int i, temp0, temp1;
+  const uint32_t* pop = &population[4];
+  const uint32_t* const LoopEnd = &population[length];
+
+  __asm__ volatile(
+    "mult   $zero,    $zero                  \n\t"
+    "xor    %[i],     %[i],       %[i]       \n\t"
+    "beq    %[pop],   %[LoopEnd], 2f         \n\t"
+  "1:                                        \n\t"
+    "lw     %[temp0], 0(%[pop])              \n\t"
+    "lw     %[temp1], 4(%[pop])              \n\t"
+    "addiu  %[i],     %[i],       1          \n\t"
+    "addiu  %[pop],   %[pop],     8          \n\t"
+    "madd   %[i],     %[temp0]               \n\t"
+    "madd   %[i],     %[temp1]               \n\t"
+    "bne    %[pop],   %[LoopEnd], 1b         \n\t"
+  "2:                                        \n\t"
+    "mfhi   %[temp0]                         \n\t"
+    "mflo   %[temp1]                         \n\t"
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
+      [i]"=&r"(i), [pop]"+r"(pop)
+    : [LoopEnd]"r"(LoopEnd)
+    : "memory", "hi", "lo"
+  );
+
+  return (double)((int64_t)temp0 << 32 | temp1);
+}
+
+// C version of this function:
+//   int i = 0;
+//   int64_t cost = 0;
+//   const uint32_t* pX = &X[4];
+//   const uint32_t* pY = &Y[4];
+//   const uint32_t* LoopEnd = &X[length];
+//   while (pX != LoopEnd) {
+//     const uint32_t xy0 = *pX + *pY;
+//     const uint32_t xy1 = *(pX + 1) + *(pY + 1);
+//     ++i;
+//     cost += i * xy0;
+//     cost += i * xy1;
+//     pX += 2;
+//     pY += 2;
+//   }
+//   return (double)cost;
+static double ExtraCostCombined(const uint32_t* const X,
+                                const uint32_t* const Y, int length) {
+  int i, temp0, temp1, temp2, temp3;
+  const uint32_t* pX = &X[4];
+  const uint32_t* pY = &Y[4];
+  const uint32_t* const LoopEnd = &X[length];
+
+  __asm__ volatile(
+    "mult   $zero,    $zero                  \n\t"
+    "xor    %[i],     %[i],       %[i]       \n\t"
+    "beq    %[pX],    %[LoopEnd], 2f         \n\t"
+  "1:                                        \n\t"
+    "lw     %[temp0], 0(%[pX])               \n\t"
+    "lw     %[temp1], 0(%[pY])               \n\t"
+    "lw     %[temp2], 4(%[pX])               \n\t"
+    "lw     %[temp3], 4(%[pY])               \n\t"
+    "addiu  %[i],     %[i],       1          \n\t"
+    "addu   %[temp0], %[temp0],   %[temp1]   \n\t"
+    "addu   %[temp2], %[temp2],   %[temp3]   \n\t"
+    "addiu  %[pX],    %[pX],      8          \n\t"
+    "addiu  %[pY],    %[pY],      8          \n\t"
+    "madd   %[i],     %[temp0]               \n\t"
+    "madd   %[i],     %[temp2]               \n\t"
+    "bne    %[pX],    %[LoopEnd], 1b         \n\t"
+  "2:                                        \n\t"
+    "mfhi   %[temp0]                         \n\t"
+    "mflo   %[temp1]                         \n\t"
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
+      [temp2]"=&r"(temp2), [temp3]"=&r"(temp3),
+      [i]"=&r"(i), [pX]"+r"(pX), [pY]"+r"(pY)
+    : [LoopEnd]"r"(LoopEnd)
+    : "memory", "hi", "lo"
+  );
+
+  return (double)((int64_t)temp0 << 32 | temp1);
+}
+
+#define HUFFMAN_COST_PASS                                 \
+  __asm__ volatile(                                       \
+    "sll   %[temp1],  %[temp0],    3           \n\t"      \
+    "addiu %[temp3],  %[streak],   -3          \n\t"      \
+    "addu  %[temp2],  %[pstreaks], %[temp1]    \n\t"      \
+    "blez  %[temp3],  1f                       \n\t"      \
+    "srl   %[temp1],  %[temp1],    1           \n\t"      \
+    "addu  %[temp3],  %[pcnts],    %[temp1]    \n\t"      \
+    "lw    %[temp0],  4(%[temp2])              \n\t"      \
+    "lw    %[temp1],  0(%[temp3])              \n\t"      \
+    "addu  %[temp0],  %[temp0],    %[streak]   \n\t"      \
+    "addiu %[temp1],  %[temp1],    1           \n\t"      \
+    "sw    %[temp0],  4(%[temp2])              \n\t"      \
+    "sw    %[temp1],  0(%[temp3])              \n\t"      \
+    "b     2f                                  \n\t"      \
+  "1:                                          \n\t"      \
+    "lw    %[temp0],  0(%[temp2])              \n\t"      \
+    "addu  %[temp0],  %[temp0],    %[streak]   \n\t"      \
+    "sw    %[temp0],  0(%[temp2])              \n\t"      \
+  "2:                                          \n\t"      \
+    : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),           \
+      [temp3]"=&r"(temp3), [temp0]"+r"(temp0)             \
+    : [pstreaks]"r"(pstreaks), [pcnts]"r"(pcnts),         \
+      [streak]"r"(streak)                                 \
+    : "memory"                                            \
+  );
+
+// Returns the various RLE counts
+static VP8LStreaks HuffmanCostCount(const uint32_t* population, int length) {
+  int i;
+  int streak = 0;
+  VP8LStreaks stats;
+  int* const pstreaks = &stats.streaks[0][0];
+  int* const pcnts = &stats.counts[0];
+  int temp0, temp1, temp2, temp3;
+  memset(&stats, 0, sizeof(stats));
+  for (i = 0; i < length - 1; ++i) {
+    ++streak;
+    if (population[i] == population[i + 1]) {
+      continue;
+    }
+    temp0 = (population[i] != 0);
+    HUFFMAN_COST_PASS
+    streak = 0;
+  }
+  ++streak;
+  temp0 = (population[i] != 0);
+  HUFFMAN_COST_PASS
+
+  return stats;
+}
+
+static VP8LStreaks HuffmanCostCombinedCount(const uint32_t* X,
+                                            const uint32_t* Y, int length) {
+  int i;
+  int streak = 0;
+  VP8LStreaks stats;
+  int* const pstreaks = &stats.streaks[0][0];
+  int* const pcnts = &stats.counts[0];
+  int temp0, temp1, temp2, temp3;
+  memset(&stats, 0, sizeof(stats));
+  for (i = 0; i < length - 1; ++i) {
+    const uint32_t xy = X[i] + Y[i];
+    const uint32_t xy_next = X[i + 1] + Y[i + 1];
+    ++streak;
+    if (xy == xy_next) {
+      continue;
+    }
+    temp0 = (xy != 0);
+    HUFFMAN_COST_PASS
+    streak = 0;
+  }
+  {
+    const uint32_t xy = X[i] + Y[i];
+    ++streak;
+    temp0 = (xy != 0);
+    HUFFMAN_COST_PASS
+  }
+
+  return stats;
+}
+
+#define ASM_START                                       \
+  __asm__ volatile(                                     \
+    ".set   push                            \n\t"       \
+    ".set   at                              \n\t"       \
+    ".set   macro                           \n\t"       \
+  "1:                                       \n\t"
+
+// P2 = P0 + P1
+// A..D - offsets
+// E - temp variable to tell macro
+//     if pointer should be incremented
+// literal_ and successive histograms could be unaligned
+// so we must use ulw and usw
+#define ADD_TO_OUT(A, B, C, D, E, P0, P1, P2)           \
+    "ulw    %[temp0], "#A"(%["#P0"])        \n\t"       \
+    "ulw    %[temp1], "#B"(%["#P0"])        \n\t"       \
+    "ulw    %[temp2], "#C"(%["#P0"])        \n\t"       \
+    "ulw    %[temp3], "#D"(%["#P0"])        \n\t"       \
+    "ulw    %[temp4], "#A"(%["#P1"])        \n\t"       \
+    "ulw    %[temp5], "#B"(%["#P1"])        \n\t"       \
+    "ulw    %[temp6], "#C"(%["#P1"])        \n\t"       \
+    "ulw    %[temp7], "#D"(%["#P1"])        \n\t"       \
+    "addu   %[temp4], %[temp4],   %[temp0]  \n\t"       \
+    "addu   %[temp5], %[temp5],   %[temp1]  \n\t"       \
+    "addu   %[temp6], %[temp6],   %[temp2]  \n\t"       \
+    "addu   %[temp7], %[temp7],   %[temp3]  \n\t"       \
+    "addiu  %["#P0"],  %["#P0"],  16        \n\t"       \
+  ".if "#E" == 1                            \n\t"       \
+    "addiu  %["#P1"],  %["#P1"],  16        \n\t"       \
+  ".endif                                   \n\t"       \
+    "usw    %[temp4], "#A"(%["#P2"])        \n\t"       \
+    "usw    %[temp5], "#B"(%["#P2"])        \n\t"       \
+    "usw    %[temp6], "#C"(%["#P2"])        \n\t"       \
+    "usw    %[temp7], "#D"(%["#P2"])        \n\t"       \
+    "addiu  %["#P2"], %["#P2"],   16        \n\t"       \
+    "bne    %["#P0"], %[LoopEnd], 1b        \n\t"       \
+    ".set   pop                             \n\t"       \
+
+#define ASM_END_COMMON_0                                \
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),         \
+      [temp2]"=&r"(temp2), [temp3]"=&r"(temp3),         \
+      [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),         \
+      [temp6]"=&r"(temp6), [temp7]"=&r"(temp7),         \
+      [pa]"+r"(pa), [pout]"+r"(pout)
+
+#define ASM_END_COMMON_1                                \
+    : [LoopEnd]"r"(LoopEnd)                             \
+    : "memory", "at"                                    \
+  );
+
+#define ASM_END_0                                       \
+    ASM_END_COMMON_0                                    \
+      , [pb]"+r"(pb)                                    \
+    ASM_END_COMMON_1
+
+#define ASM_END_1                                       \
+    ASM_END_COMMON_0                                    \
+    ASM_END_COMMON_1
+
+#define ADD_VECTOR(A, B, OUT, SIZE, EXTRA_SIZE)  do {   \
+  const uint32_t* pa = (const uint32_t*)(A);            \
+  const uint32_t* pb = (const uint32_t*)(B);            \
+  uint32_t* pout = (uint32_t*)(OUT);                    \
+  const uint32_t* const LoopEnd = pa + (SIZE);          \
+  assert((SIZE) % 4 == 0);                              \
+  ASM_START                                             \
+  ADD_TO_OUT(0, 4, 8, 12, 1, pa, pb, pout)              \
+  ASM_END_0                                             \
+  if ((EXTRA_SIZE) > 0) {                               \
+    const int last = (EXTRA_SIZE);                      \
+    int i;                                              \
+    for (i = 0; i < last; ++i) pout[i] = pa[i] + pb[i]; \
+  }                                                     \
+} while (0)
+
+#define ADD_VECTOR_EQ(A, OUT, SIZE, EXTRA_SIZE)  do {   \
+  const uint32_t* pa = (const uint32_t*)(A);            \
+  uint32_t* pout = (uint32_t*)(OUT);                    \
+  const uint32_t* const LoopEnd = pa + (SIZE);          \
+  assert((SIZE) % 4 == 0);                              \
+  ASM_START                                             \
+  ADD_TO_OUT(0, 4, 8, 12, 0, pa, pout, pout)            \
+  ASM_END_1                                             \
+  if ((EXTRA_SIZE) > 0) {                               \
+    const int last = (EXTRA_SIZE);                      \
+    int i;                                              \
+    for (i = 0; i < last; ++i) pout[i] += pa[i];        \
+  }                                                     \
+} while (0)
+
+static void HistogramAdd(const VP8LHistogram* const a,
+                         const VP8LHistogram* const b,
+                         VP8LHistogram* const out) {
+  uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
+  const int extra_cache_size = VP8LHistogramNumCodes(a->palette_code_bits_)
+                             - (NUM_LITERAL_CODES + NUM_LENGTH_CODES);
+  assert(a->palette_code_bits_ == b->palette_code_bits_);
+
+  if (b != out) {
+    ADD_VECTOR(a->literal_, b->literal_, out->literal_,
+               NUM_LITERAL_CODES + NUM_LENGTH_CODES, extra_cache_size);
+    ADD_VECTOR(a->distance_, b->distance_, out->distance_,
+               NUM_DISTANCE_CODES, 0);
+    ADD_VECTOR(a->red_, b->red_, out->red_, NUM_LITERAL_CODES, 0);
+    ADD_VECTOR(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES, 0);
+    ADD_VECTOR(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES, 0);
+  } else {
+    ADD_VECTOR_EQ(a->literal_, out->literal_,
+                  NUM_LITERAL_CODES + NUM_LENGTH_CODES, extra_cache_size);
+    ADD_VECTOR_EQ(a->distance_, out->distance_, NUM_DISTANCE_CODES, 0);
+    ADD_VECTOR_EQ(a->red_, out->red_, NUM_LITERAL_CODES, 0);
+    ADD_VECTOR_EQ(a->blue_, out->blue_, NUM_LITERAL_CODES, 0);
+    ADD_VECTOR_EQ(a->alpha_, out->alpha_, NUM_LITERAL_CODES, 0);
+  }
+}
+
+#undef ADD_VECTOR_EQ
+#undef ADD_VECTOR
+#undef ASM_END_1
+#undef ASM_END_0
+#undef ASM_END_COMMON_1
+#undef ASM_END_COMMON_0
+#undef ADD_TO_OUT
+#undef ASM_START
+
+#endif  // WEBP_USE_MIPS32
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8LDspInitMIPS32(void);
+
+void VP8LDspInitMIPS32(void) {
+#if defined(WEBP_USE_MIPS32)
+  VP8LFastSLog2Slow = FastSLog2Slow;
+  VP8LFastLog2Slow = FastLog2Slow;
+  VP8LExtraCost = ExtraCost;
+  VP8LExtraCostCombined = ExtraCostCombined;
+  VP8LHuffmanCostCount = HuffmanCostCount;
+  VP8LHuffmanCostCombinedCount = HuffmanCostCombinedCount;
+  VP8LHistogramAdd = HistogramAdd;
+#endif  // WEBP_USE_MIPS32
+}

src/dsp/lossless_neon.c

@@ -0,0 +1,357 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// NEON variant of methods for lossless decoder
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_NEON)
+
+#include <arm_neon.h>
+
+#include "./lossless.h"
+#include "./neon.h"
+
+//------------------------------------------------------------------------------
+// Colorspace conversion functions
+
+#if !defined(WORK_AROUND_GCC)
+// gcc 4.6.0 had some trouble (NDK-r9) with this code. We only use it for
+// gcc-4.8.x at least.
+static void ConvertBGRAToRGBA(const uint32_t* src,
+                              int num_pixels, uint8_t* dst) {
+  const uint32_t* const end = src + (num_pixels & ~15);
+  for (; src < end; src += 16) {
+    uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
+    // swap B and R. (VSWP d0,d2 has no intrinsics equivalent!)
+    const uint8x16_t tmp = pixel.val[0];
+    pixel.val[0] = pixel.val[2];
+    pixel.val[2] = tmp;
+    vst4q_u8(dst, pixel);
+    dst += 64;
+  }
+  VP8LConvertBGRAToRGBA_C(src, num_pixels & 15, dst);  // left-overs
+}
+
+static void ConvertBGRAToBGR(const uint32_t* src,
+                             int num_pixels, uint8_t* dst) {
+  const uint32_t* const end = src + (num_pixels & ~15);
+  for (; src < end; src += 16) {
+    const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
+    const uint8x16x3_t tmp = { { pixel.val[0], pixel.val[1], pixel.val[2] } };
+    vst3q_u8(dst, tmp);
+    dst += 48;
+  }
+  VP8LConvertBGRAToBGR_C(src, num_pixels & 15, dst);  // left-overs
+}
+
+static void ConvertBGRAToRGB(const uint32_t* src,
+                             int num_pixels, uint8_t* dst) {
+  const uint32_t* const end = src + (num_pixels & ~15);
+  for (; src < end; src += 16) {
+    const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
+    const uint8x16x3_t tmp = { { pixel.val[2], pixel.val[1], pixel.val[0] } };
+    vst3q_u8(dst, tmp);
+    dst += 48;
+  }
+  VP8LConvertBGRAToRGB_C(src, num_pixels & 15, dst);  // left-overs
+}
+
+#else  // WORK_AROUND_GCC
+
+// gcc-4.6.0 fallback
+
+static const uint8_t kRGBAShuffle[8] = { 2, 1, 0, 3, 6, 5, 4, 7 };
+
+static void ConvertBGRAToRGBA(const uint32_t* src,
+                              int num_pixels, uint8_t* dst) {
+  const uint32_t* const end = src + (num_pixels & ~1);
+  const uint8x8_t shuffle = vld1_u8(kRGBAShuffle);
+  for (; src < end; src += 2) {
+    const uint8x8_t pixels = vld1_u8((uint8_t*)src);
+    vst1_u8(dst, vtbl1_u8(pixels, shuffle));
+    dst += 8;
+  }
+  VP8LConvertBGRAToRGBA_C(src, num_pixels & 1, dst);  // left-overs
+}
+
+static const uint8_t kBGRShuffle[3][8] = {
+  {  0,  1,  2,  4,  5,  6,  8,  9 },
+  { 10, 12, 13, 14, 16, 17, 18, 20 },
+  { 21, 22, 24, 25, 26, 28, 29, 30 }
+};
+
+static void ConvertBGRAToBGR(const uint32_t* src,
+                             int num_pixels, uint8_t* dst) {
+  const uint32_t* const end = src + (num_pixels & ~7);
+  const uint8x8_t shuffle0 = vld1_u8(kBGRShuffle[0]);
+  const uint8x8_t shuffle1 = vld1_u8(kBGRShuffle[1]);
+  const uint8x8_t shuffle2 = vld1_u8(kBGRShuffle[2]);
+  for (; src < end; src += 8) {
+    uint8x8x4_t pixels;
+    INIT_VECTOR4(pixels,
+                 vld1_u8((const uint8_t*)(src + 0)),
+                 vld1_u8((const uint8_t*)(src + 2)),
+                 vld1_u8((const uint8_t*)(src + 4)),
+                 vld1_u8((const uint8_t*)(src + 6)));
+    vst1_u8(dst +  0, vtbl4_u8(pixels, shuffle0));
+    vst1_u8(dst +  8, vtbl4_u8(pixels, shuffle1));
+    vst1_u8(dst + 16, vtbl4_u8(pixels, shuffle2));
+    dst += 8 * 3;
+  }
+  VP8LConvertBGRAToBGR_C(src, num_pixels & 7, dst);  // left-overs
+}
+
+static const uint8_t kRGBShuffle[3][8] = {
+  {  2,  1,  0,  6,  5,  4, 10,  9 },
+  {  8, 14, 13, 12, 18, 17, 16, 22 },
+  { 21, 20, 26, 25, 24, 30, 29, 28 }
+};
+
+static void ConvertBGRAToRGB(const uint32_t* src,
+                             int num_pixels, uint8_t* dst) {
+  const uint32_t* const end = src + (num_pixels & ~7);
+  const uint8x8_t shuffle0 = vld1_u8(kRGBShuffle[0]);
+  const uint8x8_t shuffle1 = vld1_u8(kRGBShuffle[1]);
+  const uint8x8_t shuffle2 = vld1_u8(kRGBShuffle[2]);
+  for (; src < end; src += 8) {
+    uint8x8x4_t pixels;
+    INIT_VECTOR4(pixels,
+                 vld1_u8((const uint8_t*)(src + 0)),
+                 vld1_u8((const uint8_t*)(src + 2)),
+                 vld1_u8((const uint8_t*)(src + 4)),
+                 vld1_u8((const uint8_t*)(src + 6)));
+    vst1_u8(dst +  0, vtbl4_u8(pixels, shuffle0));
+    vst1_u8(dst +  8, vtbl4_u8(pixels, shuffle1));
+    vst1_u8(dst + 16, vtbl4_u8(pixels, shuffle2));
+    dst += 8 * 3;
+  }
+  VP8LConvertBGRAToRGB_C(src, num_pixels & 7, dst);  // left-overs
+}
+
+#endif   // !WORK_AROUND_GCC
+
+//------------------------------------------------------------------------------
+
+#ifdef USE_INTRINSICS
+
+static WEBP_INLINE uint32_t Average2(const uint32_t* const a,
+                                     const uint32_t* const b) {
+  const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
+  const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
+  const uint8x8_t avg = vhadd_u8(a0, b0);
+  return vget_lane_u32(vreinterpret_u32_u8(avg), 0);
+}
+
+static WEBP_INLINE uint32_t Average3(const uint32_t* const a,
+                                     const uint32_t* const b,
+                                     const uint32_t* const c) {
+  const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
+  const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
+  const uint8x8_t c0 = vreinterpret_u8_u64(vcreate_u64(*c));
+  const uint8x8_t avg1 = vhadd_u8(a0, c0);
+  const uint8x8_t avg2 = vhadd_u8(avg1, b0);
+  return vget_lane_u32(vreinterpret_u32_u8(avg2), 0);
+}
+
+static WEBP_INLINE uint32_t Average4(const uint32_t* const a,
+                                     const uint32_t* const b,
+                                     const uint32_t* const c,
+                                     const uint32_t* const d) {
+  const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
+  const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
+  const uint8x8_t c0 = vreinterpret_u8_u64(vcreate_u64(*c));
+  const uint8x8_t d0 = vreinterpret_u8_u64(vcreate_u64(*d));
+  const uint8x8_t avg1 = vhadd_u8(a0, b0);
+  const uint8x8_t avg2 = vhadd_u8(c0, d0);
+  const uint8x8_t avg3 = vhadd_u8(avg1, avg2);
+  return vget_lane_u32(vreinterpret_u32_u8(avg3), 0);
+}
+
+static uint32_t Predictor5(uint32_t left, const uint32_t* const top) {
+  return Average3(&left, top + 0, top + 1);
+}
+
+static uint32_t Predictor6(uint32_t left, const uint32_t* const top) {
+  return Average2(&left, top - 1);
+}
+
+static uint32_t Predictor7(uint32_t left, const uint32_t* const top) {
+  return Average2(&left, top + 0);
+}
+
+static uint32_t Predictor8(uint32_t left, const uint32_t* const top) {
+  (void)left;
+  return Average2(top - 1, top + 0);
+}
+
+static uint32_t Predictor9(uint32_t left, const uint32_t* const top) {
+  (void)left;
+  return Average2(top + 0, top + 1);
+}
+
+static uint32_t Predictor10(uint32_t left, const uint32_t* const top) {
+  return Average4(&left, top - 1, top + 0, top + 1);
+}
+
+//------------------------------------------------------------------------------
+
+static WEBP_INLINE uint32_t Select(const uint32_t* const c0,
+                                   const uint32_t* const c1,
+                                   const uint32_t* const c2) {
+  const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
+  const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
+  const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
+  const uint8x8_t bc = vabd_u8(p1, p2);   // |b-c|
+  const uint8x8_t ac = vabd_u8(p0, p2);   // |a-c|
+  const int16x4_t sum_bc = vreinterpret_s16_u16(vpaddl_u8(bc));
+  const int16x4_t sum_ac = vreinterpret_s16_u16(vpaddl_u8(ac));
+  const int32x2_t diff = vpaddl_s16(vsub_s16(sum_bc, sum_ac));
+  const int32_t pa_minus_pb = vget_lane_s32(diff, 0);
+  return (pa_minus_pb <= 0) ? *c0 : *c1;
+}
+
+static uint32_t Predictor11(uint32_t left, const uint32_t* const top) {
+  return Select(top + 0, &left, top - 1);
+}
+
+static WEBP_INLINE uint32_t ClampedAddSubtractFull(const uint32_t* const c0,
+                                                   const uint32_t* const c1,
+                                                   const uint32_t* const c2) {
+  const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
+  const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
+  const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
+  const uint16x8_t sum0 = vaddl_u8(p0, p1);                // add and widen
+  const uint16x8_t sum1 = vqsubq_u16(sum0, vmovl_u8(p2));  // widen and subtract
+  const uint8x8_t out = vqmovn_u16(sum1);                  // narrow and clamp
+  return vget_lane_u32(vreinterpret_u32_u8(out), 0);
+}
+
+static uint32_t Predictor12(uint32_t left, const uint32_t* const top) {
+  return ClampedAddSubtractFull(&left, top + 0, top - 1);
+}
+
+static WEBP_INLINE uint32_t ClampedAddSubtractHalf(const uint32_t* const c0,
+                                                   const uint32_t* const c1,
+                                                   const uint32_t* const c2) {
+  const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
+  const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
+  const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
+  const uint8x8_t avg = vhadd_u8(p0, p1);                  // Average(c0,c1)
+  const uint8x8_t ab = vshr_n_u8(vqsub_u8(avg, p2), 1);    // (a-b)>>1 saturated
+  const uint8x8_t ba = vshr_n_u8(vqsub_u8(p2, avg), 1);    // (b-a)>>1 saturated
+  const uint8x8_t out = vqsub_u8(vqadd_u8(avg, ab), ba);
+  return vget_lane_u32(vreinterpret_u32_u8(out), 0);
+}
+
+static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
+  return ClampedAddSubtractHalf(&left, top + 0, top - 1);
+}
+
+//------------------------------------------------------------------------------
+// Subtract-Green Transform
+
+// vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
+// non-standard versions there.
+#if defined(__APPLE__) && defined(__aarch64__) && \
+    defined(__apple_build_version__) && (__apple_build_version__< 6020037)
+#define USE_VTBLQ
+#endif
+
+#ifdef USE_VTBLQ
+// 255 = byte will be zeroed
+static const uint8_t kGreenShuffle[16] = {
+  1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255
+};
+
+static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
+                                             const uint8x16_t shuffle) {
+  return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)),
+                     vtbl1q_u8(argb, vget_high_u8(shuffle)));
+}
+#else  // !USE_VTBLQ
+// 255 = byte will be zeroed
+static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255  };
+
+static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
+                                             const uint8x8_t shuffle) {
+  return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
+                     vtbl1_u8(vget_high_u8(argb), shuffle));
+}
+#endif  // USE_VTBLQ
+
+static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
+  const uint32_t* const end = argb_data + (num_pixels & ~3);
+#ifdef USE_VTBLQ
+  const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
+#else
+  const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
+#endif
+  for (; argb_data < end; argb_data += 4) {
+    const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
+    const uint8x16_t greens = DoGreenShuffle(argb, shuffle);
+    vst1q_u8((uint8_t*)argb_data, vsubq_u8(argb, greens));
+  }
+  // fallthrough and finish off with plain-C
+  VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3);
+}
+
+static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
+  const uint32_t* const end = argb_data + (num_pixels & ~3);
+#ifdef USE_VTBLQ
+  const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
+#else
+  const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
+#endif
+  for (; argb_data < end; argb_data += 4) {
+    const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
+    const uint8x16_t greens = DoGreenShuffle(argb, shuffle);
+    vst1q_u8((uint8_t*)argb_data, vaddq_u8(argb, greens));
+  }
+  // fallthrough and finish off with plain-C
+  VP8LAddGreenToBlueAndRed_C(argb_data, num_pixels & 3);
+}
+
+#undef USE_VTBLQ
+
+#endif   // USE_INTRINSICS
+
+#endif   // WEBP_USE_NEON
+
+//------------------------------------------------------------------------------
+
+extern void VP8LDspInitNEON(void);
+
+void VP8LDspInitNEON(void) {
+#if defined(WEBP_USE_NEON)
+  VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;
+  VP8LConvertBGRAToBGR = ConvertBGRAToBGR;
+  VP8LConvertBGRAToRGB = ConvertBGRAToRGB;
+
+#ifdef USE_INTRINSICS
+  VP8LPredictors[5] = Predictor5;
+  VP8LPredictors[6] = Predictor6;
+  VP8LPredictors[7] = Predictor7;
+  VP8LPredictors[8] = Predictor8;
+  VP8LPredictors[9] = Predictor9;
+  VP8LPredictors[10] = Predictor10;
+  VP8LPredictors[11] = Predictor11;
+  VP8LPredictors[12] = Predictor12;
+  VP8LPredictors[13] = Predictor13;
+
+  VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
+  VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
+#endif
+
+#endif   // WEBP_USE_NEON
+}
+
+//------------------------------------------------------------------------------

src/dsp/lossless_sse2.c

@@ -0,0 +1,535 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// SSE2 variant of methods for lossless decoder
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./dsp.h"
+
+#include <assert.h>
+
+#if defined(WEBP_USE_SSE2)
+#include <emmintrin.h>
+#include "./lossless.h"
+
+//------------------------------------------------------------------------------
+// Predictor Transform
+
+static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1,
+                                                   uint32_t c2) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
+  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
+  const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
+  const __m128i V1 = _mm_add_epi16(C0, C1);
+  const __m128i V2 = _mm_sub_epi16(V1, C2);
+  const __m128i b = _mm_packus_epi16(V2, V2);
+  const uint32_t output = _mm_cvtsi128_si32(b);
+  return output;
+}
+
+static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
+                                                   uint32_t c2) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
+  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
+  const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
+  const __m128i avg = _mm_add_epi16(C1, C0);
+  const __m128i A0 = _mm_srli_epi16(avg, 1);
+  const __m128i A1 = _mm_sub_epi16(A0, B0);
+  const __m128i BgtA = _mm_cmpgt_epi16(B0, A0);
+  const __m128i A2 = _mm_sub_epi16(A1, BgtA);
+  const __m128i A3 = _mm_srai_epi16(A2, 1);
+  const __m128i A4 = _mm_add_epi16(A0, A3);
+  const __m128i A5 = _mm_packus_epi16(A4, A4);
+  const uint32_t output = _mm_cvtsi128_si32(A5);
+  return output;
+}
+
+static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
+  int pa_minus_pb;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i A0 = _mm_cvtsi32_si128(a);
+  const __m128i B0 = _mm_cvtsi32_si128(b);
+  const __m128i C0 = _mm_cvtsi32_si128(c);
+  const __m128i AC0 = _mm_subs_epu8(A0, C0);
+  const __m128i CA0 = _mm_subs_epu8(C0, A0);
+  const __m128i BC0 = _mm_subs_epu8(B0, C0);
+  const __m128i CB0 = _mm_subs_epu8(C0, B0);
+  const __m128i AC = _mm_or_si128(AC0, CA0);
+  const __m128i BC = _mm_or_si128(BC0, CB0);
+  const __m128i pa = _mm_unpacklo_epi8(AC, zero);  // |a - c|
+  const __m128i pb = _mm_unpacklo_epi8(BC, zero);  // |b - c|
+  const __m128i diff = _mm_sub_epi16(pb, pa);
+  {
+    int16_t out[8];
+    _mm_storeu_si128((__m128i*)out, diff);
+    pa_minus_pb = out[0] + out[1] + out[2] + out[3];
+  }
+  return (pa_minus_pb <= 0) ? a : b;
+}
+
+static WEBP_INLINE __m128i Average2_128i(uint32_t a0, uint32_t a1) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a0), zero);
+  const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero);
+  const __m128i sum = _mm_add_epi16(A1, A0);
+  const __m128i avg = _mm_srli_epi16(sum, 1);
+  return avg;
+}
+
+static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) {
+  const __m128i avg = Average2_128i(a0, a1);
+  const __m128i A2 = _mm_packus_epi16(avg, avg);
+  const uint32_t output = _mm_cvtsi128_si32(A2);
+  return output;
+}
+
+static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i avg1 = Average2_128i(a0, a2);
+  const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero);
+  const __m128i sum = _mm_add_epi16(avg1, A1);
+  const __m128i avg2 = _mm_srli_epi16(sum, 1);
+  const __m128i A2 = _mm_packus_epi16(avg2, avg2);
+  const uint32_t output = _mm_cvtsi128_si32(A2);
+  return output;
+}
+
+static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1,
+                                     uint32_t a2, uint32_t a3) {
+  const __m128i avg1 = Average2_128i(a0, a1);
+  const __m128i avg2 = Average2_128i(a2, a3);
+  const __m128i sum = _mm_add_epi16(avg2, avg1);
+  const __m128i avg3 = _mm_srli_epi16(sum, 1);
+  const __m128i A0 = _mm_packus_epi16(avg3, avg3);
+  const uint32_t output = _mm_cvtsi128_si32(A0);
+  return output;
+}
+
+static uint32_t Predictor5(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average3(left, top[0], top[1]);
+  return pred;
+}
+static uint32_t Predictor6(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average2(left, top[-1]);
+  return pred;
+}
+static uint32_t Predictor7(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average2(left, top[0]);
+  return pred;
+}
+static uint32_t Predictor8(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average2(top[-1], top[0]);
+  (void)left;
+  return pred;
+}
+static uint32_t Predictor9(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average2(top[0], top[1]);
+  (void)left;
+  return pred;
+}
+static uint32_t Predictor10(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average4(left, top[-1], top[0], top[1]);
+  return pred;
+}
+static uint32_t Predictor11(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Select(top[0], left, top[-1]);
+  return pred;
+}
+static uint32_t Predictor12(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]);
+  return pred;
+}
+static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]);
+  return pred;
+}
+
+//------------------------------------------------------------------------------
+// Subtract-Green Transform
+
+static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
+  const __m128i mask = _mm_set1_epi32(0x0000ff00);
+  int i;
+  for (i = 0; i + 4 <= num_pixels; i += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
+    const __m128i in_00g0 = _mm_and_si128(in, mask);     // 00g0|00g0|...
+    const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8);  // 0g00|0g00|...
+    const __m128i in_000g = _mm_srli_epi32(in_00g0, 8);  // 000g|000g|...
+    const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g);
+    const __m128i out = _mm_sub_epi8(in, in_0g0g);
+    _mm_storeu_si128((__m128i*)&argb_data[i], out);
+  }
+  // fallthrough and finish off with plain-C
+  VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
+}
+
+static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
+  const __m128i mask = _mm_set1_epi32(0x0000ff00);
+  int i;
+  for (i = 0; i + 4 <= num_pixels; i += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
+    const __m128i in_00g0 = _mm_and_si128(in, mask);     // 00g0|00g0|...
+    const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8);  // 0g00|0g00|...
+    const __m128i in_000g = _mm_srli_epi32(in_00g0, 8);  // 000g|000g|...
+    const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g);
+    const __m128i out = _mm_add_epi8(in, in_0g0g);
+    _mm_storeu_si128((__m128i*)&argb_data[i], out);
+  }
+  // fallthrough and finish off with plain-C
+  VP8LAddGreenToBlueAndRed_C(argb_data + i, num_pixels - i);
+}
+
+//------------------------------------------------------------------------------
+// Color Transform
+
+static WEBP_INLINE __m128i ColorTransformDelta(__m128i color_pred,
+                                               __m128i color) {
+  // We simulate signed 8-bit multiplication as:
+  // * Left shift the two (8-bit) numbers by 8 bits,
+  // * Perform a 16-bit signed multiplication and retain the higher 16-bits.
+  const __m128i color_pred_shifted = _mm_slli_epi32(color_pred, 8);
+  const __m128i color_shifted = _mm_slli_epi32(color, 8);
+  // Note: This performs multiplication on 8 packed 16-bit numbers, 4 of which
+  // happen to be zeroes.
+  const __m128i signed_mult =
+      _mm_mulhi_epi16(color_pred_shifted, color_shifted);
+  return _mm_srli_epi32(signed_mult, 5);
+}
+
+static WEBP_INLINE void TransformColor(const VP8LMultipliers* const m,
+                                       uint32_t* argb_data,
+                                       int num_pixels) {
+  const __m128i g_to_r = _mm_set1_epi32(m->green_to_red_);       // multipliers
+  const __m128i g_to_b = _mm_set1_epi32(m->green_to_blue_);
+  const __m128i r_to_b = _mm_set1_epi32(m->red_to_blue_);
+
+  int i;
+
+  for (i = 0; i + 4 <= num_pixels; i += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
+    const __m128i alpha_green_mask = _mm_set1_epi32(0xff00ff00);  // masks
+    const __m128i red_mask = _mm_set1_epi32(0x00ff0000);
+    const __m128i green_mask = _mm_set1_epi32(0x0000ff00);
+    const __m128i lower_8bit_mask  = _mm_set1_epi32(0x000000ff);
+    const __m128i ag = _mm_and_si128(in, alpha_green_mask);      // alpha, green
+    const __m128i r = _mm_srli_epi32(_mm_and_si128(in, red_mask), 16);
+    const __m128i g = _mm_srli_epi32(_mm_and_si128(in, green_mask), 8);
+    const __m128i b = in;
+
+    const __m128i r_delta = ColorTransformDelta(g_to_r, g);      // red
+    const __m128i r_new =
+        _mm_and_si128(_mm_sub_epi32(r, r_delta), lower_8bit_mask);
+    const __m128i r_new_shifted = _mm_slli_epi32(r_new, 16);
+
+    const __m128i b_delta_1 = ColorTransformDelta(g_to_b, g);    // blue
+    const __m128i b_delta_2 = ColorTransformDelta(r_to_b, r);
+    const __m128i b_delta = _mm_add_epi32(b_delta_1, b_delta_2);
+    const __m128i b_new =
+        _mm_and_si128(_mm_sub_epi32(b, b_delta), lower_8bit_mask);
+
+    const __m128i out = _mm_or_si128(_mm_or_si128(ag, r_new_shifted), b_new);
+    _mm_storeu_si128((__m128i*)&argb_data[i], out);
+  }
+
+  // Fall-back to C-version for left-overs.
+  VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
+}
+
+static WEBP_INLINE void TransformColorInverse(const VP8LMultipliers* const m,
+                                              uint32_t* argb_data,
+                                              int num_pixels) {
+  const __m128i g_to_r = _mm_set1_epi32(m->green_to_red_);       // multipliers
+  const __m128i g_to_b = _mm_set1_epi32(m->green_to_blue_);
+  const __m128i r_to_b = _mm_set1_epi32(m->red_to_blue_);
+
+  int i;
+
+  for (i = 0; i + 4 <= num_pixels; i += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
+    const __m128i alpha_green_mask = _mm_set1_epi32(0xff00ff00);  // masks
+    const __m128i red_mask = _mm_set1_epi32(0x00ff0000);
+    const __m128i green_mask = _mm_set1_epi32(0x0000ff00);
+    const __m128i lower_8bit_mask  = _mm_set1_epi32(0x000000ff);
+    const __m128i ag = _mm_and_si128(in, alpha_green_mask);      // alpha, green
+    const __m128i r = _mm_srli_epi32(_mm_and_si128(in, red_mask), 16);
+    const __m128i g = _mm_srli_epi32(_mm_and_si128(in, green_mask), 8);
+    const __m128i b = in;
+
+    const __m128i r_delta = ColorTransformDelta(g_to_r, g);      // red
+    const __m128i r_new =
+        _mm_and_si128(_mm_add_epi32(r, r_delta), lower_8bit_mask);
+    const __m128i r_new_shifted = _mm_slli_epi32(r_new, 16);
+
+    const __m128i b_delta_1 = ColorTransformDelta(g_to_b, g);    // blue
+    const __m128i b_delta_2 = ColorTransformDelta(r_to_b, r_new);
+    const __m128i b_delta = _mm_add_epi32(b_delta_1, b_delta_2);
+    const __m128i b_new =
+        _mm_and_si128(_mm_add_epi32(b, b_delta), lower_8bit_mask);
+
+    const __m128i out = _mm_or_si128(_mm_or_si128(ag, r_new_shifted), b_new);
+    _mm_storeu_si128((__m128i*)&argb_data[i], out);
+  }
+
+  // Fall-back to C-version for left-overs.
+  VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i);
+}
+
+//------------------------------------------------------------------------------
+// Color-space conversion functions
+
+static void ConvertBGRAToRGBA(const uint32_t* src,
+                              int num_pixels, uint8_t* dst) {
+  const __m128i* in = (const __m128i*)src;
+  __m128i* out = (__m128i*)dst;
+  while (num_pixels >= 8) {
+    const __m128i bgra0 = _mm_loadu_si128(in++);     // bgra0|bgra1|bgra2|bgra3
+    const __m128i bgra4 = _mm_loadu_si128(in++);     // bgra4|bgra5|bgra6|bgra7
+    const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4);  // b0b4g0g4r0r4a0a4...
+    const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4);  // b2b6g2g6r2r6a2a6...
+    const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h);   // b0b2b4b6g0g2g4g6...
+    const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h);   // b1b3b5b7g1g3g5g7...
+    const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h);   // b0...b7 | g0...g7
+    const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h);   // r0...r7 | a0...a7
+    const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h);  // g0...g7 | a0...a7
+    const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l);  // r0...r7 | b0...b7
+    const __m128i rg0 = _mm_unpacklo_epi8(rb0, ga0);   // r0g0r1g1 ... r6g6r7g7
+    const __m128i ba0 = _mm_unpackhi_epi8(rb0, ga0);   // b0a0b1a1 ... b6a6b7a7
+    const __m128i rgba0 = _mm_unpacklo_epi16(rg0, ba0);  // rgba0|rgba1...
+    const __m128i rgba4 = _mm_unpackhi_epi16(rg0, ba0);  // rgba4|rgba5...
+    _mm_storeu_si128(out++, rgba0);
+    _mm_storeu_si128(out++, rgba4);
+    num_pixels -= 8;
+  }
+  // left-overs
+  VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
+}
+
+static void ConvertBGRAToRGBA4444(const uint32_t* src,
+                                  int num_pixels, uint8_t* dst) {
+  const __m128i mask_0x0f = _mm_set1_epi8(0x0f);
+  const __m128i mask_0xf0 = _mm_set1_epi8(0xf0);
+  const __m128i* in = (const __m128i*)src;
+  __m128i* out = (__m128i*)dst;
+  while (num_pixels >= 8) {
+    const __m128i bgra0 = _mm_loadu_si128(in++);     // bgra0|bgra1|bgra2|bgra3
+    const __m128i bgra4 = _mm_loadu_si128(in++);     // bgra4|bgra5|bgra6|bgra7
+    const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4);  // b0b4g0g4r0r4a0a4...
+    const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4);  // b2b6g2g6r2r6a2a6...
+    const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h);    // b0b2b4b6g0g2g4g6...
+    const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h);    // b1b3b5b7g1g3g5g7...
+    const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h);    // b0...b7 | g0...g7
+    const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h);    // r0...r7 | a0...a7
+    const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h);   // g0...g7 | a0...a7
+    const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l);   // r0...r7 | b0...b7
+    const __m128i ga1 = _mm_srli_epi16(ga0, 4);         // g0-|g1-|...|a6-|a7-
+    const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0);  // -r0|-r1|...|-b6|-a7
+    const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f);  // g0-|g1-|...|a6-|a7-
+    const __m128i rgba0 = _mm_or_si128(ga2, rb1);       // rg0..rg7 | ba0..ba7
+    const __m128i rgba1 = _mm_srli_si128(rgba0, 8);     // ba0..ba7 | 0
+#ifdef WEBP_SWAP_16BIT_CSP
+    const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0);  // barg0...barg7
+#else
+    const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1);  // rgba0...rgba7
+#endif
+    _mm_storeu_si128(out++, rgba);
+    num_pixels -= 8;
+  }
+  // left-overs
+  VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
+}
+
+static void ConvertBGRAToRGB565(const uint32_t* src,
+                                int num_pixels, uint8_t* dst) {
+  const __m128i mask_0xe0 = _mm_set1_epi8(0xe0);
+  const __m128i mask_0xf8 = _mm_set1_epi8(0xf8);
+  const __m128i mask_0x07 = _mm_set1_epi8(0x07);
+  const __m128i* in = (const __m128i*)src;
+  __m128i* out = (__m128i*)dst;
+  while (num_pixels >= 8) {
+    const __m128i bgra0 = _mm_loadu_si128(in++);     // bgra0|bgra1|bgra2|bgra3
+    const __m128i bgra4 = _mm_loadu_si128(in++);     // bgra4|bgra5|bgra6|bgra7
+    const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4);  // b0b4g0g4r0r4a0a4...
+    const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4);  // b2b6g2g6r2r6a2a6...
+    const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h);      // b0b2b4b6g0g2g4g6...
+    const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h);      // b1b3b5b7g1g3g5g7...
+    const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h);      // b0...b7 | g0...g7
+    const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h);      // r0...r7 | a0...a7
+    const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h);     // g0...g7 | a0...a7
+    const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l);     // r0...r7 | b0...b7
+    const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8);    // -r0..-r7|-b0..-b7
+    const __m128i g_lo1 = _mm_srli_epi16(ga0, 5);
+    const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07);  // g0-...g7-|xx (3b)
+    const __m128i g_hi1 = _mm_slli_epi16(ga0, 3);
+    const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0);  // -g0...-g7|xx (3b)
+    const __m128i b0 = _mm_srli_si128(rb1, 8);              // -b0...-b7|0
+    const __m128i rg1 = _mm_or_si128(rb1, g_lo2);           // gr0...gr7|xx
+    const __m128i b1 = _mm_srli_epi16(b0, 3);
+    const __m128i gb1 = _mm_or_si128(b1, g_hi2);            // bg0...bg7|xx
+#ifdef WEBP_SWAP_16BIT_CSP
+    const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1);     // rggb0...rggb7
+#else
+    const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1);     // bgrb0...bgrb7
+#endif
+    _mm_storeu_si128(out++, rgba);
+    num_pixels -= 8;
+  }
+  // left-overs
+  VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
+}
+
+static void ConvertBGRAToBGR(const uint32_t* src,
+                             int num_pixels, uint8_t* dst) {
+  const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff);
+  const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0);
+  const __m128i* in = (const __m128i*)src;
+  const uint8_t* const end = dst + num_pixels * 3;
+  // the last storel_epi64 below writes 8 bytes starting at offset 18
+  while (dst + 26 <= end) {
+    const __m128i bgra0 = _mm_loadu_si128(in++);     // bgra0|bgra1|bgra2|bgra3
+    const __m128i bgra4 = _mm_loadu_si128(in++);     // bgra4|bgra5|bgra6|bgra7
+    const __m128i a0l = _mm_and_si128(bgra0, mask_l);   // bgr0|0|bgr0|0
+    const __m128i a4l = _mm_and_si128(bgra4, mask_l);   // bgr0|0|bgr0|0
+    const __m128i a0h = _mm_and_si128(bgra0, mask_h);   // 0|bgr0|0|bgr0
+    const __m128i a4h = _mm_and_si128(bgra4, mask_h);   // 0|bgr0|0|bgr0
+    const __m128i b0h = _mm_srli_epi64(a0h, 8);         // 000b|gr00|000b|gr00
+    const __m128i b4h = _mm_srli_epi64(a4h, 8);         // 000b|gr00|000b|gr00
+    const __m128i c0 = _mm_or_si128(a0l, b0h);          // rgbrgb00|rgbrgb00
+    const __m128i c4 = _mm_or_si128(a4l, b4h);          // rgbrgb00|rgbrgb00
+    const __m128i c2 = _mm_srli_si128(c0, 8);
+    const __m128i c6 = _mm_srli_si128(c4, 8);
+    _mm_storel_epi64((__m128i*)(dst +   0), c0);
+    _mm_storel_epi64((__m128i*)(dst +   6), c2);
+    _mm_storel_epi64((__m128i*)(dst +  12), c4);
+    _mm_storel_epi64((__m128i*)(dst +  18), c6);
+    dst += 24;
+    num_pixels -= 8;
+  }
+  // left-overs
+  VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst);
+}
+
+//------------------------------------------------------------------------------
+
+#define LINE_SIZE 16    // 8 or 16
+static void AddVector(const uint32_t* a, const uint32_t* b, uint32_t* out,
+                      int size) {
+  int i;
+  assert(size % LINE_SIZE == 0);
+  for (i = 0; i < size; i += LINE_SIZE) {
+    const __m128i a0 = _mm_loadu_si128((__m128i*)&a[i +  0]);
+    const __m128i a1 = _mm_loadu_si128((__m128i*)&a[i +  4]);
+#if (LINE_SIZE == 16)
+    const __m128i a2 = _mm_loadu_si128((__m128i*)&a[i +  8]);
+    const __m128i a3 = _mm_loadu_si128((__m128i*)&a[i + 12]);
+#endif
+    const __m128i b0 = _mm_loadu_si128((__m128i*)&b[i +  0]);
+    const __m128i b1 = _mm_loadu_si128((__m128i*)&b[i +  4]);
+#if (LINE_SIZE == 16)
+    const __m128i b2 = _mm_loadu_si128((__m128i*)&b[i +  8]);
+    const __m128i b3 = _mm_loadu_si128((__m128i*)&b[i + 12]);
+#endif
+    _mm_storeu_si128((__m128i*)&out[i +  0], _mm_add_epi32(a0, b0));
+    _mm_storeu_si128((__m128i*)&out[i +  4], _mm_add_epi32(a1, b1));
+#if (LINE_SIZE == 16)
+    _mm_storeu_si128((__m128i*)&out[i +  8], _mm_add_epi32(a2, b2));
+    _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
+#endif
+  }
+}
+
+static void AddVectorEq(const uint32_t* a, uint32_t* out, int size) {
+  int i;
+  assert(size % LINE_SIZE == 0);
+  for (i = 0; i < size; i += LINE_SIZE) {
+    const __m128i a0 = _mm_loadu_si128((__m128i*)&a[i +  0]);
+    const __m128i a1 = _mm_loadu_si128((__m128i*)&a[i +  4]);
+#if (LINE_SIZE == 16)
+    const __m128i a2 = _mm_loadu_si128((__m128i*)&a[i +  8]);
+    const __m128i a3 = _mm_loadu_si128((__m128i*)&a[i + 12]);
+#endif
+    const __m128i b0 = _mm_loadu_si128((__m128i*)&out[i +  0]);
+    const __m128i b1 = _mm_loadu_si128((__m128i*)&out[i +  4]);
+#if (LINE_SIZE == 16)
+    const __m128i b2 = _mm_loadu_si128((__m128i*)&out[i +  8]);
+    const __m128i b3 = _mm_loadu_si128((__m128i*)&out[i + 12]);
+#endif
+    _mm_storeu_si128((__m128i*)&out[i +  0], _mm_add_epi32(a0, b0));
+    _mm_storeu_si128((__m128i*)&out[i +  4], _mm_add_epi32(a1, b1));
+#if (LINE_SIZE == 16)
+    _mm_storeu_si128((__m128i*)&out[i +  8], _mm_add_epi32(a2, b2));
+    _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
+#endif
+  }
+}
+#undef LINE_SIZE
+
+// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
+// that's ok since the histogram values are less than 1<<28 (max picture size).
+static void HistogramAdd(const VP8LHistogram* const a,
+                         const VP8LHistogram* const b,
+                         VP8LHistogram* const out) {
+  int i;
+  const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_);
+  assert(a->palette_code_bits_ == b->palette_code_bits_);
+  if (b != out) {
+    AddVector(a->literal_, b->literal_, out->literal_, NUM_LITERAL_CODES);
+    AddVector(a->red_, b->red_, out->red_, NUM_LITERAL_CODES);
+    AddVector(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES);
+    AddVector(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES);
+  } else {
+    AddVectorEq(a->literal_, out->literal_, NUM_LITERAL_CODES);
+    AddVectorEq(a->red_, out->red_, NUM_LITERAL_CODES);
+    AddVectorEq(a->blue_, out->blue_, NUM_LITERAL_CODES);
+    AddVectorEq(a->alpha_, out->alpha_, NUM_LITERAL_CODES);
+  }
+  for (i = NUM_LITERAL_CODES; i < literal_size; ++i) {
+    out->literal_[i] = a->literal_[i] + b->literal_[i];
+  }
+  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
+    out->distance_[i] = a->distance_[i] + b->distance_[i];
+  }
+}
+
+#endif   // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+
+extern void VP8LDspInitSSE2(void);
+
+void VP8LDspInitSSE2(void) {
+#if defined(WEBP_USE_SSE2)
+  VP8LPredictors[5] = Predictor5;
+  VP8LPredictors[6] = Predictor6;
+  VP8LPredictors[7] = Predictor7;
+  VP8LPredictors[8] = Predictor8;
+  VP8LPredictors[9] = Predictor9;
+  VP8LPredictors[10] = Predictor10;
+  VP8LPredictors[11] = Predictor11;
+  VP8LPredictors[12] = Predictor12;
+  VP8LPredictors[13] = Predictor13;
+
+  VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
+  VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
+
+  VP8LTransformColor = TransformColor;
+  VP8LTransformColorInverse = TransformColorInverse;
+
+  VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;
+  VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444;
+  VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565;
+  VP8LConvertBGRAToBGR = ConvertBGRAToBGR;
+
+  VP8LHistogramAdd = HistogramAdd;
+#endif   // WEBP_USE_SSE2
+}
+
+//------------------------------------------------------------------------------

src/dsp/neon.h

@@ -0,0 +1,82 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//  NEON common code.
+
+#ifndef WEBP_DSP_NEON_H_
+#define WEBP_DSP_NEON_H_
+
+#include <arm_neon.h>
+
+#include "./dsp.h"
+
+// Right now, some intrinsics functions seem slower, so we disable them
+// everywhere except aarch64 where the inline assembly is incompatible.
+#if defined(__aarch64__)
+#define USE_INTRINSICS   // use intrinsics when possible
+#endif
+
+#define INIT_VECTOR2(v, a, b) do {  \
+  v.val[0] = a;                     \
+  v.val[1] = b;                     \
+} while (0)
+
+#define INIT_VECTOR3(v, a, b, c) do {  \
+  v.val[0] = a;                        \
+  v.val[1] = b;                        \
+  v.val[2] = c;                        \
+} while (0)
+
+#define INIT_VECTOR4(v, a, b, c, d) do {  \
+  v.val[0] = a;                           \
+  v.val[1] = b;                           \
+  v.val[2] = c;                           \
+  v.val[3] = d;                           \
+} while (0)
+
+// if using intrinsics, this flag avoids some functions that make gcc-4.6.3
+// crash ("internal compiler error: in immed_double_const, at emit-rtl.").
+// (probably similar to gcc.gnu.org/bugzilla/show_bug.cgi?id=48183)
+#if !(LOCAL_GCC_PREREQ(4,8) || defined(__aarch64__))
+#define WORK_AROUND_GCC
+#endif
+
+static WEBP_INLINE int32x4x4_t Transpose4x4(const int32x4x4_t rows) {
+  uint64x2x2_t row01, row23;
+
+  row01.val[0] = vreinterpretq_u64_s32(rows.val[0]);
+  row01.val[1] = vreinterpretq_u64_s32(rows.val[1]);
+  row23.val[0] = vreinterpretq_u64_s32(rows.val[2]);
+  row23.val[1] = vreinterpretq_u64_s32(rows.val[3]);
+  // Transpose 64-bit values (there's no vswp equivalent)
+  {
+    const uint64x1_t row0h = vget_high_u64(row01.val[0]);
+    const uint64x1_t row2l = vget_low_u64(row23.val[0]);
+    const uint64x1_t row1h = vget_high_u64(row01.val[1]);
+    const uint64x1_t row3l = vget_low_u64(row23.val[1]);
+    row01.val[0] = vcombine_u64(vget_low_u64(row01.val[0]), row2l);
+    row23.val[0] = vcombine_u64(row0h, vget_high_u64(row23.val[0]));
+    row01.val[1] = vcombine_u64(vget_low_u64(row01.val[1]), row3l);
+    row23.val[1] = vcombine_u64(row1h, vget_high_u64(row23.val[1]));
+  }
+  {
+    const int32x4x2_t out01 = vtrnq_s32(vreinterpretq_s32_u64(row01.val[0]),
+                                        vreinterpretq_s32_u64(row01.val[1]));
+    const int32x4x2_t out23 = vtrnq_s32(vreinterpretq_s32_u64(row23.val[0]),
+                                        vreinterpretq_s32_u64(row23.val[1]));
+    int32x4x4_t out;
+    out.val[0] = out01.val[0];
+    out.val[1] = out01.val[1];
+    out.val[2] = out23.val[0];
+    out.val[3] = out23.val[1];
+    return out;
+  }
+}
+
+#endif  // WEBP_DSP_NEON_H_

src/dsp/upsampling.c

@@ -106,57 +106,6 @@ UPSAMPLE_FUNC(UpsampleRgb565LinePair,  VP8YuvToRgb565,  2)
 
 #endif  // FANCY_UPSAMPLING
 
-//------------------------------------------------------------------------------
-// simple point-sampling
-
-#define SAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                                    \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
-                      const uint8_t* u, const uint8_t* v,                      \
-                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
-  int i;                                                                       \
-  for (i = 0; i < len - 1; i += 2) {                                           \
-    FUNC(top_y[0], u[0], v[0], top_dst);                                       \
-    FUNC(top_y[1], u[0], v[0], top_dst + XSTEP);                               \
-    FUNC(bottom_y[0], u[0], v[0], bottom_dst);                                 \
-    FUNC(bottom_y[1], u[0], v[0], bottom_dst + XSTEP);                         \
-    top_y += 2;                                                                \
-    bottom_y += 2;                                                             \
-    u++;                                                                       \
-    v++;                                                                       \
-    top_dst += 2 * XSTEP;                                                      \
-    bottom_dst += 2 * XSTEP;                                                   \
-  }                                                                            \
-  if (i == len - 1) {    /* last one */                                        \
-    FUNC(top_y[0], u[0], v[0], top_dst);                                       \
-    FUNC(bottom_y[0], u[0], v[0], bottom_dst);                                 \
-  }                                                                            \
-}
-
-// All variants implemented.
-SAMPLE_FUNC(SampleRgbLinePair,      VP8YuvToRgb,  3)
-SAMPLE_FUNC(SampleBgrLinePair,      VP8YuvToBgr,  3)
-SAMPLE_FUNC(SampleRgbaLinePair,     VP8YuvToRgba, 4)
-SAMPLE_FUNC(SampleBgraLinePair,     VP8YuvToBgra, 4)
-SAMPLE_FUNC(SampleArgbLinePair,     VP8YuvToArgb, 4)
-SAMPLE_FUNC(SampleRgba4444LinePair, VP8YuvToRgba4444, 2)
-SAMPLE_FUNC(SampleRgb565LinePair,   VP8YuvToRgb565, 2)
-
-#undef SAMPLE_FUNC
-
-const WebPSampleLinePairFunc WebPSamplers[MODE_LAST] = {
-  SampleRgbLinePair,       // MODE_RGB
-  SampleRgbaLinePair,      // MODE_RGBA
-  SampleBgrLinePair,       // MODE_BGR
-  SampleBgraLinePair,      // MODE_BGRA
-  SampleArgbLinePair,      // MODE_ARGB
-  SampleRgba4444LinePair,  // MODE_RGBA_4444
-  SampleRgb565LinePair,    // MODE_RGB_565
-  SampleRgbaLinePair,      // MODE_rgbA
-  SampleBgraLinePair,      // MODE_bgrA
-  SampleArgbLinePair,      // MODE_Argb
-  SampleRgba4444LinePair   // MODE_rgbA_4444
-};
-
 //------------------------------------------------------------------------------
 
 #if !defined(FANCY_UPSAMPLING)
@@ -235,85 +184,17 @@ const WebPYUV444Converter WebPYUV444Converters[MODE_LAST] = {
 };
 
 //------------------------------------------------------------------------------
-// Premultiplied modes
+// Main calls
 
-// non dithered-modes
+extern void WebPInitUpsamplersSSE2(void);
+extern void WebPInitUpsamplersNEON(void);
 
-// (x * a * 32897) >> 23 is bit-wise equivalent to (int)(x * a / 255.)
-// for all 8bit x or a. For bit-wise equivalence to (int)(x * a / 255. + .5),
-// one can use instead: (x * a * 65793 + (1 << 23)) >> 24
-#if 1     // (int)(x * a / 255.)
-#define MULTIPLIER(a)   ((a) * 32897UL)
-#define PREMULTIPLY(x, m) (((x) * (m)) >> 23)
-#else     // (int)(x * a / 255. + .5)
-#define MULTIPLIER(a) ((a) * 65793UL)
-#define PREMULTIPLY(x, m) (((x) * (m) + (1UL << 23)) >> 24)
-#endif
-
-static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_first,
-                               int w, int h, int stride) {
-  while (h-- > 0) {
-    uint8_t* const rgb = rgba + (alpha_first ? 1 : 0);
-    const uint8_t* const alpha = rgba + (alpha_first ? 0 : 3);
-    int i;
-    for (i = 0; i < w; ++i) {
-      const uint32_t a = alpha[4 * i];
-      if (a != 0xff) {
-        const uint32_t mult = MULTIPLIER(a);
-        rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult);
-        rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult);
-        rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult);
-      }
-    }
-    rgba += stride;
-  }
-}
-#undef MULTIPLIER
-#undef PREMULTIPLY
-
-// rgbA4444
-
-#define MULTIPLIER(a)  ((a) * 0x1111)    // 0x1111 ~= (1 << 16) / 15
-
-static WEBP_INLINE uint8_t dither_hi(uint8_t x) {
-  return (x & 0xf0) | (x >> 4);
-}
-
-static WEBP_INLINE uint8_t dither_lo(uint8_t x) {
-  return (x & 0x0f) | (x << 4);
-}
-
-static WEBP_INLINE uint8_t multiply(uint8_t x, uint32_t m) {
-  return (x * m) >> 16;
-}
-
-static void ApplyAlphaMultiply4444(uint8_t* rgba4444,
-                                   int w, int h, int stride) {
-  while (h-- > 0) {
-    int i;
-    for (i = 0; i < w; ++i) {
-      const uint8_t a = (rgba4444[2 * i + 1] & 0x0f);
-      const uint32_t mult = MULTIPLIER(a);
-      const uint8_t r = multiply(dither_hi(rgba4444[2 * i + 0]), mult);
-      const uint8_t g = multiply(dither_lo(rgba4444[2 * i + 0]), mult);
-      const uint8_t b = multiply(dither_hi(rgba4444[2 * i + 1]), mult);
-      rgba4444[2 * i + 0] = (r & 0xf0) | ((g >> 4) & 0x0f);
-      rgba4444[2 * i + 1] = (b & 0xf0) | a;
-    }
-    rgba4444 += stride;
-  }
-}
-#undef MULTIPLIER
-
-void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int)
-    = ApplyAlphaMultiply;
-void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int)
-    = ApplyAlphaMultiply4444;
-
-//------------------------------------------------------------------------------
-// Main call
+static volatile VP8CPUInfo upsampling_last_cpuinfo_used2 =
+    (VP8CPUInfo)&upsampling_last_cpuinfo_used2;
 
 void WebPInitUpsamplers(void) {
+  if (upsampling_last_cpuinfo_used2 == VP8GetCPUInfo) return;
+
 #ifdef FANCY_UPSAMPLING
   WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair;
   WebPUpsamplers[MODE_RGBA]      = UpsampleRgbaLinePair;
@@ -322,6 +203,10 @@ void WebPInitUpsamplers(void) {
   WebPUpsamplers[MODE_ARGB]      = UpsampleArgbLinePair;
   WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
   WebPUpsamplers[MODE_RGB_565]   = UpsampleRgb565LinePair;
+  WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_Argb]      = UpsampleArgbLinePair;
+  WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
 
   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
   if (VP8GetCPUInfo != NULL) {
@@ -337,30 +222,7 @@ void WebPInitUpsamplers(void) {
 #endif
   }
 #endif  // FANCY_UPSAMPLING
+  upsampling_last_cpuinfo_used2 = VP8GetCPUInfo;
 }
 
-void WebPInitPremultiply(void) {
-  WebPApplyAlphaMultiply = ApplyAlphaMultiply;
-  WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply4444;
-
-#ifdef FANCY_UPSAMPLING
-  WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair;
-  WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair;
-  WebPUpsamplers[MODE_Argb]      = UpsampleArgbLinePair;
-  WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
-
-  if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
-    if (VP8GetCPUInfo(kSSE2)) {
-      WebPInitPremultiplySSE2();
-    }
-#endif
-#if defined(WEBP_USE_NEON)
-    if (VP8GetCPUInfo(kNEON)) {
-      WebPInitPremultiplyNEON();
-    }
-#endif
-  }
-#endif  // FANCY_UPSAMPLING
-}
-
+//------------------------------------------------------------------------------

src/dsp/upsampling_neon.c

@@ -19,6 +19,7 @@
 #include <assert.h>
 #include <arm_neon.h>
 #include <string.h>
+#include "./neon.h"
 #include "./yuv.h"
 
 #ifdef FANCY_UPSAMPLING
@@ -61,8 +62,9 @@
   d = vrhadd_u8(d, diag1);                                              \
                                                                         \
   {                                                                     \
-    const uint8x8x2_t a_b = {{ a, b }};                                 \
-    const uint8x8x2_t c_d = {{ c, d }};                                 \
+    uint8x8x2_t a_b, c_d;                                               \
+    INIT_VECTOR2(a_b, a, b);                                            \
+    INIT_VECTOR2(c_d, c, d);                                            \
     vst2_u8(out,      a_b);                                             \
     vst2_u8(out + 32, c_d);                                             \
   }                                                                     \
@@ -89,25 +91,29 @@ static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,
 
 static const int16_t kCoeffs[4] = { kYScale, kVToR, kUToG, kVToG };
 
-#define v255 vmov_n_u8(255)
+#define v255 vdup_n_u8(255)
 
 #define STORE_Rgb(out, r, g, b) do {                                    \
-  const uint8x8x3_t r_g_b = {{ r, g, b }};                              \
+  uint8x8x3_t r_g_b;                                                    \
+  INIT_VECTOR3(r_g_b, r, g, b);                                         \
   vst3_u8(out, r_g_b);                                                  \
 } while (0)
 
 #define STORE_Bgr(out, r, g, b) do {                                    \
-  const uint8x8x3_t b_g_r = {{ b, g, r }};                              \
+  uint8x8x3_t b_g_r;                                                    \
+  INIT_VECTOR3(b_g_r, b, g, r);                                         \
   vst3_u8(out, b_g_r);                                                  \
 } while (0)
 
 #define STORE_Rgba(out, r, g, b) do {                                   \
-  const uint8x8x4_t r_g_b_v255 = {{ r, g, b, v255 }};                   \
+  uint8x8x4_t r_g_b_v255;                                               \
+  INIT_VECTOR4(r_g_b_v255, r, g, b, v255);                              \
   vst4_u8(out, r_g_b_v255);                                             \
 } while (0)
 
 #define STORE_Bgra(out, r, g, b) do {                                   \
-  const uint8x8x4_t b_g_r_v255 = {{ b, g, r, v255 }};                   \
+  uint8x8x4_t b_g_r_v255;                                               \
+  INIT_VECTOR4(b_g_r_v255, b, g, r, v255);                              \
   vst4_u8(out, b_g_r_v255);                                             \
 } while (0)
 
@@ -190,9 +196,9 @@ static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y,    \
   const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;                  \
                                                                         \
   const int16x4_t cf16 = vld1_s16(kCoeffs);                             \
-  const int32x2_t cf32 = vmov_n_s32(kUToB);                             \
-  const uint8x8_t u16  = vmov_n_u8(16);                                 \
-  const uint8x8_t u128 = vmov_n_u8(128);                                \
+  const int32x2_t cf32 = vdup_n_s32(kUToB);                             \
+  const uint8x8_t u16  = vdup_n_u8(16);                                 \
+  const uint8x8_t u128 = vdup_n_u8(128);                                \
                                                                         \
   /* Treat the first pixel in regular way */                            \
   assert(top_y != NULL);                                                \
@@ -225,10 +231,10 @@ static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y,    \
 }
 
 // NEON variants of the fancy upsampler.
-NEON_UPSAMPLE_FUNC(UpsampleRgbLinePairNEON,  Rgb,  3)
-NEON_UPSAMPLE_FUNC(UpsampleBgrLinePairNEON,  Bgr,  3)
-NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePairNEON, Rgba, 4)
-NEON_UPSAMPLE_FUNC(UpsampleBgraLinePairNEON, Bgra, 4)
+NEON_UPSAMPLE_FUNC(UpsampleRgbLinePair,  Rgb,  3)
+NEON_UPSAMPLE_FUNC(UpsampleBgrLinePair,  Bgr,  3)
+NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePair, Rgba, 4)
+NEON_UPSAMPLE_FUNC(UpsampleBgraLinePair, Bgra, 4)
 
 #endif  // FANCY_UPSAMPLING
 
@@ -236,30 +242,26 @@ NEON_UPSAMPLE_FUNC(UpsampleBgraLinePairNEON, Bgra, 4)
 
 //------------------------------------------------------------------------------
 
+extern void WebPInitUpsamplersNEON(void);
+
 #ifdef FANCY_UPSAMPLING
 
 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 
 void WebPInitUpsamplersNEON(void) {
 #if defined(WEBP_USE_NEON)
-  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePairNEON;
-  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairNEON;
-  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePairNEON;
-  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairNEON;
-#endif   // WEBP_USE_NEON
-}
-
-void WebPInitPremultiplyNEON(void) {
-#if defined(WEBP_USE_NEON)
-  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairNEON;
-  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairNEON;
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair;
+  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair;
+  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
 #endif   // WEBP_USE_NEON
 }
 
 #else
 
 // this empty function is to avoid an empty .o
-void WebPInitPremultiplyNEON(void) {}
+void WebPInitUpsamplersNEON(void) {}
 
 #endif  // FANCY_UPSAMPLING
-

src/dsp/upsampling_sse2.c

@@ -169,10 +169,10 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
 }
 
 // SSE2 variants of the fancy upsampler.
-SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePairSSE2,  VP8YuvToRgb,  3)
-SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePairSSE2,  VP8YuvToBgr,  3)
-SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePairSSE2, VP8YuvToRgba, 4)
-SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4)
+SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePair,  VP8YuvToRgb,  3)
+SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePair,  VP8YuvToBgr,  3)
+SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePair, VP8YuvToRgba, 4)
+SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePair, VP8YuvToBgra, 4)
 
 #undef GET_M
 #undef PACK_AND_STORE
@@ -188,6 +188,8 @@ SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4)
 
 //------------------------------------------------------------------------------
 
+extern void WebPInitUpsamplersSSE2(void);
+
 #ifdef FANCY_UPSAMPLING
 
 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
@@ -195,24 +197,18 @@ extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 void WebPInitUpsamplersSSE2(void) {
 #if defined(WEBP_USE_SSE2)
   VP8YUVInitSSE2();
-  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePairSSE2;
-  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2;
-  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePairSSE2;
-  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2;
-#endif   // WEBP_USE_SSE2
-}
-
-void WebPInitPremultiplySSE2(void) {
-#if defined(WEBP_USE_SSE2)
-  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2;
-  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2;
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair;
+  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair;
+  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
 #endif   // WEBP_USE_SSE2
 }
 
 #else
 
 // this empty function is to avoid an empty .o
-void WebPInitPremultiplySSE2(void) {}
+void WebPInitUpsamplersSSE2(void) {}
 
 #endif  // FANCY_UPSAMPLING
-

src/dsp/yuv.c

@@ -7,13 +7,12 @@
 // be found in the AUTHORS file in the root of the source tree.
 // -----------------------------------------------------------------------------
 //
-// YUV->RGB conversion function
+// YUV->RGB conversion functions
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include "./yuv.h"
 
-
 #if defined(WEBP_YUV_USE_TABLE)
 
 static int done = 0;
@@ -68,140 +67,94 @@ void VP8YUVInit(void) {}
 #endif  // WEBP_YUV_USE_TABLE
 
 //-----------------------------------------------------------------------------
-// SSE2 extras
+// Plain-C version
 
-#if defined(WEBP_USE_SSE2)
+#define ROW_FUNC(FUNC_NAME, FUNC, XSTEP)                                       \
+static void FUNC_NAME(const uint8_t* y,                                        \
+                      const uint8_t* u, const uint8_t* v,                      \
+                      uint8_t* dst, int len) {                                 \
+  const uint8_t* const end = dst + (len & ~1) * XSTEP;                         \
+  while (dst != end) {                                                         \
+    FUNC(y[0], u[0], v[0], dst);                                               \
+    FUNC(y[1], u[0], v[0], dst + XSTEP);                                       \
+    y += 2;                                                                    \
+    ++u;                                                                       \
+    ++v;                                                                       \
+    dst += 2 * XSTEP;                                                          \
+  }                                                                            \
+  if (len & 1) {                                                               \
+    FUNC(y[0], u[0], v[0], dst);                                               \
+  }                                                                            \
+}                                                                              \
 
-#ifdef FANCY_UPSAMPLING
+// All variants implemented.
+ROW_FUNC(YuvToRgbRow,      VP8YuvToRgb,  3)
+ROW_FUNC(YuvToBgrRow,      VP8YuvToBgr,  3)
+ROW_FUNC(YuvToRgbaRow,     VP8YuvToRgba, 4)
+ROW_FUNC(YuvToBgraRow,     VP8YuvToBgra, 4)
+ROW_FUNC(YuvToArgbRow,     VP8YuvToArgb, 4)
+ROW_FUNC(YuvToRgba4444Row, VP8YuvToRgba4444, 2)
+ROW_FUNC(YuvToRgb565Row,   VP8YuvToRgb565, 2)
 
-#include <emmintrin.h>
-#include <string.h>   // for memcpy
+#undef ROW_FUNC
 
-typedef union {   // handy struct for converting SSE2 registers
-  int32_t i32[4];
-  uint8_t u8[16];
-  __m128i m;
-} VP8kCstSSE2;
-
-static int done_sse2 = 0;
-static VP8kCstSSE2 VP8kUtoRGBA[256], VP8kVtoRGBA[256], VP8kYtoRGBA[256];
-
-void VP8YUVInitSSE2(void) {
-  if (!done_sse2) {
-    int i;
-    for (i = 0; i < 256; ++i) {
-      VP8kYtoRGBA[i].i32[0] =
-        VP8kYtoRGBA[i].i32[1] =
-        VP8kYtoRGBA[i].i32[2] = (i - 16) * kYScale + YUV_HALF2;
-      VP8kYtoRGBA[i].i32[3] = 0xff << YUV_FIX2;
-
-      VP8kUtoRGBA[i].i32[0] = 0;
-      VP8kUtoRGBA[i].i32[1] = -kUToG * (i - 128);
-      VP8kUtoRGBA[i].i32[2] =  kUToB * (i - 128);
-      VP8kUtoRGBA[i].i32[3] = 0;
-
-      VP8kVtoRGBA[i].i32[0] =  kVToR * (i - 128);
-      VP8kVtoRGBA[i].i32[1] = -kVToG * (i - 128);
-      VP8kVtoRGBA[i].i32[2] = 0;
-      VP8kVtoRGBA[i].i32[3] = 0;
+// Main call for processing a plane with a WebPSamplerRowFunc function:
+void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
+                             const uint8_t* u, const uint8_t* v, int uv_stride,
+                             uint8_t* dst, int dst_stride,
+                             int width, int height, WebPSamplerRowFunc func) {
+  int j;
+  for (j = 0; j < height; ++j) {
+    func(y, u, v, dst, width);
+    y += y_stride;
+    if (j & 1) {
+      u += uv_stride;
+      v += uv_stride;
     }
-    done_sse2 = 1;
+    dst += dst_stride;
   }
 }
 
-static WEBP_INLINE __m128i VP8GetRGBA32b(int y, int u, int v) {
-  const __m128i u_part = _mm_loadu_si128(&VP8kUtoRGBA[u].m);
-  const __m128i v_part = _mm_loadu_si128(&VP8kVtoRGBA[v].m);
-  const __m128i y_part = _mm_loadu_si128(&VP8kYtoRGBA[y].m);
-  const __m128i uv_part = _mm_add_epi32(u_part, v_part);
-  const __m128i rgba1 = _mm_add_epi32(y_part, uv_part);
-  const __m128i rgba2 = _mm_srai_epi32(rgba1, YUV_FIX2);
-  return rgba2;
-}
+//-----------------------------------------------------------------------------
+// Main call
 
-static WEBP_INLINE void VP8YuvToRgbSSE2(uint8_t y, uint8_t u, uint8_t v,
-                                        uint8_t* const rgb) {
-  const __m128i tmp0 = VP8GetRGBA32b(y, u, v);
-  const __m128i tmp1 = _mm_packs_epi32(tmp0, tmp0);
-  const __m128i tmp2 = _mm_packus_epi16(tmp1, tmp1);
-  // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
-  _mm_storel_epi64((__m128i*)rgb, tmp2);
-}
+WebPSamplerRowFunc WebPSamplers[MODE_LAST];
 
-static WEBP_INLINE void VP8YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
-                                        uint8_t* const bgr) {
-  const __m128i tmp0 = VP8GetRGBA32b(y, u, v);
-  const __m128i tmp1 = _mm_shuffle_epi32(tmp0, _MM_SHUFFLE(3, 0, 1, 2));
-  const __m128i tmp2 = _mm_packs_epi32(tmp1, tmp1);
-  const __m128i tmp3 = _mm_packus_epi16(tmp2, tmp2);
-  // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
-  _mm_storel_epi64((__m128i*)bgr, tmp3);
-}
+extern void WebPInitSamplersSSE2(void);
+extern void WebPInitSamplersMIPS32(void);
 
-void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                    uint8_t* dst) {
-  int n;
-  for (n = 0; n < 32; n += 4) {
-    const __m128i tmp0_1 = VP8GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
-    const __m128i tmp0_2 = VP8GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
-    const __m128i tmp0_3 = VP8GetRGBA32b(y[n + 2], u[n + 2], v[n + 2]);
-    const __m128i tmp0_4 = VP8GetRGBA32b(y[n + 3], u[n + 3], v[n + 3]);
-    const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
-    const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
-    const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
-    _mm_storeu_si128((__m128i*)dst, tmp2);
-    dst += 4 * 4;
-  }
-}
+static volatile VP8CPUInfo yuv_last_cpuinfo_used =
+    (VP8CPUInfo)&yuv_last_cpuinfo_used;
 
-void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                    uint8_t* dst) {
-  int n;
-  for (n = 0; n < 32; n += 2) {
-    const __m128i tmp0_1 = VP8GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
-    const __m128i tmp0_2 = VP8GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
-    const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
-    const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
-    const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
-    const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
-    _mm_storel_epi64((__m128i*)dst, tmp3);
-    dst += 4 * 2;
-  }
-}
+void WebPInitSamplers(void) {
+  if (yuv_last_cpuinfo_used == VP8GetCPUInfo) return;
 
-void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                   uint8_t* dst) {
-  int n;
-  uint8_t tmp0[2 * 3 + 5 + 15];
-  uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15);  // align
-  for (n = 0; n < 30; ++n) {   // we directly stomp the *dst memory
-    VP8YuvToRgbSSE2(y[n], u[n], v[n], dst + n * 3);
-  }
-  // Last two pixels are special: we write in a tmp buffer before sending
-  // to dst.
-  VP8YuvToRgbSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
-  VP8YuvToRgbSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
-  memcpy(dst + n * 3, tmp, 2 * 3);
-}
-
-void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
-                   uint8_t* dst) {
-  int n;
-  uint8_t tmp0[2 * 3 + 5 + 15];
-  uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15);  // align
-  for (n = 0; n < 30; ++n) {
-    VP8YuvToBgrSSE2(y[n], u[n], v[n], dst + n * 3);
-  }
-  VP8YuvToBgrSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
-  VP8YuvToBgrSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
-  memcpy(dst + n * 3, tmp, 2 * 3);
-}
-
-#else
-
-void VP8YUVInitSSE2(void) {}
-
-#endif  // FANCY_UPSAMPLING
+  WebPSamplers[MODE_RGB]       = YuvToRgbRow;
+  WebPSamplers[MODE_RGBA]      = YuvToRgbaRow;
+  WebPSamplers[MODE_BGR]       = YuvToBgrRow;
+  WebPSamplers[MODE_BGRA]      = YuvToBgraRow;
+  WebPSamplers[MODE_ARGB]      = YuvToArgbRow;
+  WebPSamplers[MODE_RGBA_4444] = YuvToRgba4444Row;
+  WebPSamplers[MODE_RGB_565]   = YuvToRgb565Row;
+  WebPSamplers[MODE_rgbA]      = YuvToRgbaRow;
+  WebPSamplers[MODE_bgrA]      = YuvToBgraRow;
+  WebPSamplers[MODE_Argb]      = YuvToArgbRow;
+  WebPSamplers[MODE_rgbA_4444] = YuvToRgba4444Row;
 
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitSamplersSSE2();
+    }
 #endif  // WEBP_USE_SSE2
+#if defined(WEBP_USE_MIPS32)
+    if (VP8GetCPUInfo(kMIPS32)) {
+      WebPInitSamplersMIPS32();
+    }
+#endif  // WEBP_USE_MIPS32
+  }
+  yuv_last_cpuinfo_used = VP8GetCPUInfo;
+}
 
+//-----------------------------------------------------------------------------

src/dsp/yuv.h

@@ -245,6 +245,10 @@ void VP8YUVInit(void);
 
 #if defined(WEBP_USE_SSE2)
 
+// When the following is defined, tables are initialized statically, adding ~12k
+// to the binary size. Otherwise, they are initialized at run-time (small cost).
+#define WEBP_YUV_USE_SSE2_TABLES
+
 #if defined(FANCY_UPSAMPLING)
 // Process 32 pixels and store the result (24b or 32b per pixel) in *dst.
 void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
@@ -298,12 +302,12 @@ static WEBP_INLINE int VP8RGBToY(int r, int g, int b, int rounding) {
   return (luma + rounding) >> YUV_FIX;  // no need to clip
 }
 
-static WEBP_INLINE int VP8_RGB_TO_U(int r, int g, int b, int rounding) {
+static WEBP_INLINE int VP8RGBToU(int r, int g, int b, int rounding) {
   const int u = -11058 * r - 21710 * g + 32768 * b;
   return VP8ClipUV(u, rounding);
 }
 
-static WEBP_INLINE int VP8_RGB_TO_V(int r, int g, int b, int rounding) {
+static WEBP_INLINE int VP8RGBToV(int r, int g, int b, int rounding) {
   const int v = 32768 * r - 27439 * g - 5329 * b;
   return VP8ClipUV(v, rounding);
 }

src/dsp/yuv_mips32.c

@@ -0,0 +1,100 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// MIPS version of YUV to RGB upsampling functions.
+//
+// Author(s):  Djordje Pesut    (djordje.pesut@imgtec.com)
+//             Jovan Zelincevic (jovan.zelincevic@imgtec.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_MIPS32)
+
+#include "./yuv.h"
+
+//------------------------------------------------------------------------------
+// simple point-sampling
+
+#define ROW_FUNC(FUNC_NAME, XSTEP, R, G, B, A)                                 \
+static void FUNC_NAME(const uint8_t* y,                                        \
+                      const uint8_t* u, const uint8_t* v,                      \
+                      uint8_t* dst, int len) {                                 \
+  int i, r, g, b;                                                              \
+  int temp0, temp1, temp2, temp3, temp4;                                       \
+  for (i = 0; i < (len >> 1); i++) {                                           \
+    temp1 = kVToR * v[0];                                                      \
+    temp3 = kVToG * v[0];                                                      \
+    temp2 = kUToG * u[0];                                                      \
+    temp4 = kUToB * u[0];                                                      \
+    temp0 = kYScale * y[0];                                                    \
+    temp1 += kRCst;                                                            \
+    temp3 -= kGCst;                                                            \
+    temp2 += temp3;                                                            \
+    temp4 += kBCst;                                                            \
+    r = VP8Clip8(temp0 + temp1);                                               \
+    g = VP8Clip8(temp0 - temp2);                                               \
+    b = VP8Clip8(temp0 + temp4);                                               \
+    temp0 = kYScale * y[1];                                                    \
+    dst[R] = r;                                                                \
+    dst[G] = g;                                                                \
+    dst[B] = b;                                                                \
+    if (A) dst[A] = 0xff;                                                      \
+    r = VP8Clip8(temp0 + temp1);                                               \
+    g = VP8Clip8(temp0 - temp2);                                               \
+    b = VP8Clip8(temp0 + temp4);                                               \
+    dst[R + XSTEP] = r;                                                        \
+    dst[G + XSTEP] = g;                                                        \
+    dst[B + XSTEP] = b;                                                        \
+    if (A) dst[A + XSTEP] = 0xff;                                              \
+    y += 2;                                                                    \
+    ++u;                                                                       \
+    ++v;                                                                       \
+    dst += 2 * XSTEP;                                                          \
+  }                                                                            \
+  if (len & 1) {                                                               \
+    temp1 = kVToR * v[0];                                                      \
+    temp3 = kVToG * v[0];                                                      \
+    temp2 = kUToG * u[0];                                                      \
+    temp4 = kUToB * u[0];                                                      \
+    temp0 = kYScale * y[0];                                                    \
+    temp1 += kRCst;                                                            \
+    temp3 -= kGCst;                                                            \
+    temp2 += temp3;                                                            \
+    temp4 += kBCst;                                                            \
+    r = VP8Clip8(temp0 + temp1);                                               \
+    g = VP8Clip8(temp0 - temp2);                                               \
+    b = VP8Clip8(temp0 + temp4);                                               \
+    dst[R] = r;                                                                \
+    dst[G] = g;                                                                \
+    dst[B] = b;                                                                \
+    if (A) dst[A] = 0xff;                                                      \
+  }                                                                            \
+}
+
+ROW_FUNC(YuvToRgbRow,      3, 0, 1, 2, 0)
+ROW_FUNC(YuvToRgbaRow,     4, 0, 1, 2, 3)
+ROW_FUNC(YuvToBgrRow,      3, 2, 1, 0, 0)
+ROW_FUNC(YuvToBgraRow,     4, 2, 1, 0, 3)
+
+#undef ROW_FUNC
+
+#endif   // WEBP_USE_MIPS32
+
+//------------------------------------------------------------------------------
+
+extern void WebPInitSamplersMIPS32(void);
+
+void WebPInitSamplersMIPS32(void) {
+#if defined(WEBP_USE_MIPS32)
+  WebPSamplers[MODE_RGB]  = YuvToRgbRow;
+  WebPSamplers[MODE_RGBA] = YuvToRgbaRow;
+  WebPSamplers[MODE_BGR]  = YuvToBgrRow;
+  WebPSamplers[MODE_BGRA] = YuvToBgraRow;
+#endif  // WEBP_USE_MIPS32
+}

src/dsp/yuv_sse2.c

@@ -0,0 +1,322 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// YUV->RGB conversion functions
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./yuv.h"
+
+#if defined(WEBP_USE_SSE2)
+
+#include <emmintrin.h>
+#include <string.h>   // for memcpy
+
+typedef union {   // handy struct for converting SSE2 registers
+  int32_t i32[4];
+  uint8_t u8[16];
+  __m128i m;
+} VP8kCstSSE2;
+
+#if defined(WEBP_YUV_USE_SSE2_TABLES)
+
+#include "./yuv_tables_sse2.h"
+
+void VP8YUVInitSSE2(void) {}
+
+#else
+
+static int done_sse2 = 0;
+static VP8kCstSSE2 VP8kUtoRGBA[256], VP8kVtoRGBA[256], VP8kYtoRGBA[256];
+
+void VP8YUVInitSSE2(void) {
+  if (!done_sse2) {
+    int i;
+    for (i = 0; i < 256; ++i) {
+      VP8kYtoRGBA[i].i32[0] =
+        VP8kYtoRGBA[i].i32[1] =
+        VP8kYtoRGBA[i].i32[2] = (i - 16) * kYScale + YUV_HALF2;
+      VP8kYtoRGBA[i].i32[3] = 0xff << YUV_FIX2;
+
+      VP8kUtoRGBA[i].i32[0] = 0;
+      VP8kUtoRGBA[i].i32[1] = -kUToG * (i - 128);
+      VP8kUtoRGBA[i].i32[2] =  kUToB * (i - 128);
+      VP8kUtoRGBA[i].i32[3] = 0;
+
+      VP8kVtoRGBA[i].i32[0] =  kVToR * (i - 128);
+      VP8kVtoRGBA[i].i32[1] = -kVToG * (i - 128);
+      VP8kVtoRGBA[i].i32[2] = 0;
+      VP8kVtoRGBA[i].i32[3] = 0;
+    }
+    done_sse2 = 1;
+
+#if 0   // code used to generate 'yuv_tables_sse2.h'
+    printf("static const VP8kCstSSE2 VP8kYtoRGBA[256] = {\n");
+    for (i = 0; i < 256; ++i) {
+      printf("  {{0x%.8x, 0x%.8x, 0x%.8x, 0x%.8x}},\n",
+             VP8kYtoRGBA[i].i32[0], VP8kYtoRGBA[i].i32[1],
+             VP8kYtoRGBA[i].i32[2], VP8kYtoRGBA[i].i32[3]);
+    }
+    printf("};\n\n");
+    printf("static const VP8kCstSSE2 VP8kUtoRGBA[256] = {\n");
+    for (i = 0; i < 256; ++i) {
+      printf("  {{0, 0x%.8x, 0x%.8x, 0}},\n",
+             VP8kUtoRGBA[i].i32[1], VP8kUtoRGBA[i].i32[2]);
+    }
+    printf("};\n\n");
+    printf("static VP8kCstSSE2 VP8kVtoRGBA[256] = {\n");
+    for (i = 0; i < 256; ++i) {
+      printf("  {{0x%.8x, 0x%.8x, 0, 0}},\n",
+             VP8kVtoRGBA[i].i32[0], VP8kVtoRGBA[i].i32[1]);
+    }
+    printf("};\n\n");
+#endif
+  }
+}
+
+#endif  // WEBP_YUV_USE_SSE2_TABLES
+
+//-----------------------------------------------------------------------------
+
+static WEBP_INLINE __m128i LoadUVPart(int u, int v) {
+  const __m128i u_part = _mm_loadu_si128(&VP8kUtoRGBA[u].m);
+  const __m128i v_part = _mm_loadu_si128(&VP8kVtoRGBA[v].m);
+  const __m128i uv_part = _mm_add_epi32(u_part, v_part);
+  return uv_part;
+}
+
+static WEBP_INLINE __m128i GetRGBA32bWithUV(int y, const __m128i uv_part) {
+  const __m128i y_part = _mm_loadu_si128(&VP8kYtoRGBA[y].m);
+  const __m128i rgba1 = _mm_add_epi32(y_part, uv_part);
+  const __m128i rgba2 = _mm_srai_epi32(rgba1, YUV_FIX2);
+  return rgba2;
+}
+
+static WEBP_INLINE __m128i GetRGBA32b(int y, int u, int v) {
+  const __m128i uv_part = LoadUVPart(u, v);
+  return GetRGBA32bWithUV(y, uv_part);
+}
+
+static WEBP_INLINE void YuvToRgbSSE2(uint8_t y, uint8_t u, uint8_t v,
+                                     uint8_t* const rgb) {
+  const __m128i tmp0 = GetRGBA32b(y, u, v);
+  const __m128i tmp1 = _mm_packs_epi32(tmp0, tmp0);
+  const __m128i tmp2 = _mm_packus_epi16(tmp1, tmp1);
+  // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
+  _mm_storel_epi64((__m128i*)rgb, tmp2);
+}
+
+static WEBP_INLINE void YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
+                                     uint8_t* const bgr) {
+  const __m128i tmp0 = GetRGBA32b(y, u, v);
+  const __m128i tmp1 = _mm_shuffle_epi32(tmp0, _MM_SHUFFLE(3, 0, 1, 2));
+  const __m128i tmp2 = _mm_packs_epi32(tmp1, tmp1);
+  const __m128i tmp3 = _mm_packus_epi16(tmp2, tmp2);
+  // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
+  _mm_storel_epi64((__m128i*)bgr, tmp3);
+}
+
+//-----------------------------------------------------------------------------
+// Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
+
+#ifdef FANCY_UPSAMPLING
+
+void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                    uint8_t* dst) {
+  int n;
+  for (n = 0; n < 32; n += 4) {
+    const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
+    const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
+    const __m128i tmp0_3 = GetRGBA32b(y[n + 2], u[n + 2], v[n + 2]);
+    const __m128i tmp0_4 = GetRGBA32b(y[n + 3], u[n + 3], v[n + 3]);
+    const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
+    const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
+    const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
+    _mm_storeu_si128((__m128i*)dst, tmp2);
+    dst += 4 * 4;
+  }
+}
+
+void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                    uint8_t* dst) {
+  int n;
+  for (n = 0; n < 32; n += 2) {
+    const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
+    const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
+    const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
+    const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
+    const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
+    const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
+    _mm_storel_epi64((__m128i*)dst, tmp3);
+    dst += 4 * 2;
+  }
+}
+
+void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                   uint8_t* dst) {
+  int n;
+  uint8_t tmp0[2 * 3 + 5 + 15];
+  uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15);  // align
+  for (n = 0; n < 30; ++n) {   // we directly stomp the *dst memory
+    YuvToRgbSSE2(y[n], u[n], v[n], dst + n * 3);
+  }
+  // Last two pixels are special: we write in a tmp buffer before sending
+  // to dst.
+  YuvToRgbSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
+  YuvToRgbSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
+  memcpy(dst + n * 3, tmp, 2 * 3);
+}
+
+void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                   uint8_t* dst) {
+  int n;
+  uint8_t tmp0[2 * 3 + 5 + 15];
+  uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15);  // align
+  for (n = 0; n < 30; ++n) {
+    YuvToBgrSSE2(y[n], u[n], v[n], dst + n * 3);
+  }
+  YuvToBgrSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
+  YuvToBgrSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
+  memcpy(dst + n * 3, tmp, 2 * 3);
+}
+
+#endif  // FANCY_UPSAMPLING
+
+//-----------------------------------------------------------------------------
+// Arbitrary-length row conversion functions
+
+static void YuvToRgbaRowSSE2(const uint8_t* y,
+                             const uint8_t* u, const uint8_t* v,
+                             uint8_t* dst, int len) {
+  int n;
+  for (n = 0; n + 4 <= len; n += 4) {
+    const __m128i uv_0 = LoadUVPart(u[0], v[0]);
+    const __m128i uv_1 = LoadUVPart(u[1], v[1]);
+    const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
+    const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
+    const __m128i tmp0_3 = GetRGBA32bWithUV(y[2], uv_1);
+    const __m128i tmp0_4 = GetRGBA32bWithUV(y[3], uv_1);
+    const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
+    const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
+    const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
+    _mm_storeu_si128((__m128i*)dst, tmp2);
+    dst += 4 * 4;
+    y += 4;
+    u += 2;
+    v += 2;
+  }
+  // Finish off
+  while (n < len) {
+    VP8YuvToRgba(y[0], u[0], v[0], dst);
+    dst += 4;
+    ++y;
+    u += (n & 1);
+    v += (n & 1);
+    ++n;
+  }
+}
+
+static void YuvToBgraRowSSE2(const uint8_t* y,
+                             const uint8_t* u, const uint8_t* v,
+                             uint8_t* dst, int len) {
+  int n;
+  for (n = 0; n + 2 <= len; n += 2) {
+    const __m128i uv_0 = LoadUVPart(u[0], v[0]);
+    const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
+    const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
+    const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
+    const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
+    const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
+    const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
+    _mm_storel_epi64((__m128i*)dst, tmp3);
+    dst += 4 * 2;
+    y += 2;
+    ++u;
+    ++v;
+  }
+  // Finish off
+  if (len & 1) {
+    VP8YuvToBgra(y[0], u[0], v[0], dst);
+  }
+}
+
+static void YuvToArgbRowSSE2(const uint8_t* y,
+                             const uint8_t* u, const uint8_t* v,
+                             uint8_t* dst, int len) {
+  int n;
+  for (n = 0; n + 2 <= len; n += 2) {
+    const __m128i uv_0 = LoadUVPart(u[0], v[0]);
+    const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
+    const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
+    const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(2, 1, 0, 3));
+    const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(2, 1, 0, 3));
+    const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
+    const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
+    _mm_storel_epi64((__m128i*)dst, tmp3);
+    dst += 4 * 2;
+    y += 2;
+    ++u;
+    ++v;
+  }
+  // Finish off
+  if (len & 1) {
+    VP8YuvToArgb(y[0], u[0], v[0], dst);
+  }
+}
+
+static void YuvToRgbRowSSE2(const uint8_t* y,
+                            const uint8_t* u, const uint8_t* v,
+                            uint8_t* dst, int len) {
+  int n;
+  for (n = 0; n + 2 < len; ++n) {   // we directly stomp the *dst memory
+    YuvToRgbSSE2(y[0], u[0], v[0], dst);  // stomps 8 bytes
+    dst += 3;
+    ++y;
+    u += (n & 1);
+    v += (n & 1);
+  }
+  VP8YuvToRgb(y[0], u[0], v[0], dst);
+  if (len > 1) {
+    VP8YuvToRgb(y[1], u[n & 1], v[n & 1], dst + 3);
+  }
+}
+
+static void YuvToBgrRowSSE2(const uint8_t* y,
+                            const uint8_t* u, const uint8_t* v,
+                            uint8_t* dst, int len) {
+  int n;
+  for (n = 0; n + 2 < len; ++n) {   // we directly stomp the *dst memory
+    YuvToBgrSSE2(y[0], u[0], v[0], dst);  // stomps 8 bytes
+    dst += 3;
+    ++y;
+    u += (n & 1);
+    v += (n & 1);
+  }
+  VP8YuvToBgr(y[0], u[0], v[0], dst + 0);
+  if (len > 1) {
+    VP8YuvToBgr(y[1], u[n & 1], v[n & 1], dst + 3);
+  }
+}
+
+#endif  // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void WebPInitSamplersSSE2(void);
+
+void WebPInitSamplersSSE2(void) {
+#if defined(WEBP_USE_SSE2)
+  WebPSamplers[MODE_RGB]  = YuvToRgbRowSSE2;
+  WebPSamplers[MODE_RGBA] = YuvToRgbaRowSSE2;
+  WebPSamplers[MODE_BGR]  = YuvToBgrRowSSE2;
+  WebPSamplers[MODE_BGRA] = YuvToBgraRowSSE2;
+  WebPSamplers[MODE_ARGB] = YuvToArgbRowSSE2;
+#endif  // WEBP_USE_SSE2
+}

src/dsp/yuv_tables_sse2.h

@@ -0,0 +1,536 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// SSE2 tables for YUV->RGB conversion (12kB overall)
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+// This file is not compiled, but #include'd directly from yuv.c
+// Only used if WEBP_YUV_USE_SSE2_TABLES is defined.
+
+static const VP8kCstSSE2 VP8kYtoRGBA[256] = {
+  {{0xfffb77b0, 0xfffb77b0, 0xfffb77b0, 0x003fc000}},
+  {{0xfffbc235, 0xfffbc235, 0xfffbc235, 0x003fc000}},
+  {{0xfffc0cba, 0xfffc0cba, 0xfffc0cba, 0x003fc000}},
+  {{0xfffc573f, 0xfffc573f, 0xfffc573f, 0x003fc000}},
+  {{0xfffca1c4, 0xfffca1c4, 0xfffca1c4, 0x003fc000}},
+  {{0xfffcec49, 0xfffcec49, 0xfffcec49, 0x003fc000}},
+  {{0xfffd36ce, 0xfffd36ce, 0xfffd36ce, 0x003fc000}},
+  {{0xfffd8153, 0xfffd8153, 0xfffd8153, 0x003fc000}},
+  {{0xfffdcbd8, 0xfffdcbd8, 0xfffdcbd8, 0x003fc000}},
+  {{0xfffe165d, 0xfffe165d, 0xfffe165d, 0x003fc000}},
+  {{0xfffe60e2, 0xfffe60e2, 0xfffe60e2, 0x003fc000}},
+  {{0xfffeab67, 0xfffeab67, 0xfffeab67, 0x003fc000}},
+  {{0xfffef5ec, 0xfffef5ec, 0xfffef5ec, 0x003fc000}},
+  {{0xffff4071, 0xffff4071, 0xffff4071, 0x003fc000}},
+  {{0xffff8af6, 0xffff8af6, 0xffff8af6, 0x003fc000}},
+  {{0xffffd57b, 0xffffd57b, 0xffffd57b, 0x003fc000}},
+  {{0x00002000, 0x00002000, 0x00002000, 0x003fc000}},
+  {{0x00006a85, 0x00006a85, 0x00006a85, 0x003fc000}},
+  {{0x0000b50a, 0x0000b50a, 0x0000b50a, 0x003fc000}},
+  {{0x0000ff8f, 0x0000ff8f, 0x0000ff8f, 0x003fc000}},
+  {{0x00014a14, 0x00014a14, 0x00014a14, 0x003fc000}},
+  {{0x00019499, 0x00019499, 0x00019499, 0x003fc000}},
+  {{0x0001df1e, 0x0001df1e, 0x0001df1e, 0x003fc000}},
+  {{0x000229a3, 0x000229a3, 0x000229a3, 0x003fc000}},
+  {{0x00027428, 0x00027428, 0x00027428, 0x003fc000}},
+  {{0x0002bead, 0x0002bead, 0x0002bead, 0x003fc000}},
+  {{0x00030932, 0x00030932, 0x00030932, 0x003fc000}},
+  {{0x000353b7, 0x000353b7, 0x000353b7, 0x003fc000}},
+  {{0x00039e3c, 0x00039e3c, 0x00039e3c, 0x003fc000}},
+  {{0x0003e8c1, 0x0003e8c1, 0x0003e8c1, 0x003fc000}},
+  {{0x00043346, 0x00043346, 0x00043346, 0x003fc000}},
+  {{0x00047dcb, 0x00047dcb, 0x00047dcb, 0x003fc000}},
+  {{0x0004c850, 0x0004c850, 0x0004c850, 0x003fc000}},
+  {{0x000512d5, 0x000512d5, 0x000512d5, 0x003fc000}},
+  {{0x00055d5a, 0x00055d5a, 0x00055d5a, 0x003fc000}},
+  {{0x0005a7df, 0x0005a7df, 0x0005a7df, 0x003fc000}},
+  {{0x0005f264, 0x0005f264, 0x0005f264, 0x003fc000}},
+  {{0x00063ce9, 0x00063ce9, 0x00063ce9, 0x003fc000}},
+  {{0x0006876e, 0x0006876e, 0x0006876e, 0x003fc000}},
+  {{0x0006d1f3, 0x0006d1f3, 0x0006d1f3, 0x003fc000}},
+  {{0x00071c78, 0x00071c78, 0x00071c78, 0x003fc000}},
+  {{0x000766fd, 0x000766fd, 0x000766fd, 0x003fc000}},
+  {{0x0007b182, 0x0007b182, 0x0007b182, 0x003fc000}},
+  {{0x0007fc07, 0x0007fc07, 0x0007fc07, 0x003fc000}},
+  {{0x0008468c, 0x0008468c, 0x0008468c, 0x003fc000}},
+  {{0x00089111, 0x00089111, 0x00089111, 0x003fc000}},
+  {{0x0008db96, 0x0008db96, 0x0008db96, 0x003fc000}},
+  {{0x0009261b, 0x0009261b, 0x0009261b, 0x003fc000}},
+  {{0x000970a0, 0x000970a0, 0x000970a0, 0x003fc000}},
+  {{0x0009bb25, 0x0009bb25, 0x0009bb25, 0x003fc000}},
+  {{0x000a05aa, 0x000a05aa, 0x000a05aa, 0x003fc000}},
+  {{0x000a502f, 0x000a502f, 0x000a502f, 0x003fc000}},
+  {{0x000a9ab4, 0x000a9ab4, 0x000a9ab4, 0x003fc000}},
+  {{0x000ae539, 0x000ae539, 0x000ae539, 0x003fc000}},
+  {{0x000b2fbe, 0x000b2fbe, 0x000b2fbe, 0x003fc000}},
+  {{0x000b7a43, 0x000b7a43, 0x000b7a43, 0x003fc000}},
+  {{0x000bc4c8, 0x000bc4c8, 0x000bc4c8, 0x003fc000}},
+  {{0x000c0f4d, 0x000c0f4d, 0x000c0f4d, 0x003fc000}},
+  {{0x000c59d2, 0x000c59d2, 0x000c59d2, 0x003fc000}},
+  {{0x000ca457, 0x000ca457, 0x000ca457, 0x003fc000}},
+  {{0x000ceedc, 0x000ceedc, 0x000ceedc, 0x003fc000}},
+  {{0x000d3961, 0x000d3961, 0x000d3961, 0x003fc000}},
+  {{0x000d83e6, 0x000d83e6, 0x000d83e6, 0x003fc000}},
+  {{0x000dce6b, 0x000dce6b, 0x000dce6b, 0x003fc000}},
+  {{0x000e18f0, 0x000e18f0, 0x000e18f0, 0x003fc000}},
+  {{0x000e6375, 0x000e6375, 0x000e6375, 0x003fc000}},
+  {{0x000eadfa, 0x000eadfa, 0x000eadfa, 0x003fc000}},
+  {{0x000ef87f, 0x000ef87f, 0x000ef87f, 0x003fc000}},
+  {{0x000f4304, 0x000f4304, 0x000f4304, 0x003fc000}},
+  {{0x000f8d89, 0x000f8d89, 0x000f8d89, 0x003fc000}},
+  {{0x000fd80e, 0x000fd80e, 0x000fd80e, 0x003fc000}},
+  {{0x00102293, 0x00102293, 0x00102293, 0x003fc000}},
+  {{0x00106d18, 0x00106d18, 0x00106d18, 0x003fc000}},
+  {{0x0010b79d, 0x0010b79d, 0x0010b79d, 0x003fc000}},
+  {{0x00110222, 0x00110222, 0x00110222, 0x003fc000}},
+  {{0x00114ca7, 0x00114ca7, 0x00114ca7, 0x003fc000}},
+  {{0x0011972c, 0x0011972c, 0x0011972c, 0x003fc000}},
+  {{0x0011e1b1, 0x0011e1b1, 0x0011e1b1, 0x003fc000}},
+  {{0x00122c36, 0x00122c36, 0x00122c36, 0x003fc000}},
+  {{0x001276bb, 0x001276bb, 0x001276bb, 0x003fc000}},
+  {{0x0012c140, 0x0012c140, 0x0012c140, 0x003fc000}},
+  {{0x00130bc5, 0x00130bc5, 0x00130bc5, 0x003fc000}},
+  {{0x0013564a, 0x0013564a, 0x0013564a, 0x003fc000}},
+  {{0x0013a0cf, 0x0013a0cf, 0x0013a0cf, 0x003fc000}},
+  {{0x0013eb54, 0x0013eb54, 0x0013eb54, 0x003fc000}},
+  {{0x001435d9, 0x001435d9, 0x001435d9, 0x003fc000}},
+  {{0x0014805e, 0x0014805e, 0x0014805e, 0x003fc000}},
+  {{0x0014cae3, 0x0014cae3, 0x0014cae3, 0x003fc000}},
+  {{0x00151568, 0x00151568, 0x00151568, 0x003fc000}},
+  {{0x00155fed, 0x00155fed, 0x00155fed, 0x003fc000}},
+  {{0x0015aa72, 0x0015aa72, 0x0015aa72, 0x003fc000}},
+  {{0x0015f4f7, 0x0015f4f7, 0x0015f4f7, 0x003fc000}},
+  {{0x00163f7c, 0x00163f7c, 0x00163f7c, 0x003fc000}},
+  {{0x00168a01, 0x00168a01, 0x00168a01, 0x003fc000}},
+  {{0x0016d486, 0x0016d486, 0x0016d486, 0x003fc000}},
+  {{0x00171f0b, 0x00171f0b, 0x00171f0b, 0x003fc000}},
+  {{0x00176990, 0x00176990, 0x00176990, 0x003fc000}},
+  {{0x0017b415, 0x0017b415, 0x0017b415, 0x003fc000}},
+  {{0x0017fe9a, 0x0017fe9a, 0x0017fe9a, 0x003fc000}},
+  {{0x0018491f, 0x0018491f, 0x0018491f, 0x003fc000}},
+  {{0x001893a4, 0x001893a4, 0x001893a4, 0x003fc000}},
+  {{0x0018de29, 0x0018de29, 0x0018de29, 0x003fc000}},
+  {{0x001928ae, 0x001928ae, 0x001928ae, 0x003fc000}},
+  {{0x00197333, 0x00197333, 0x00197333, 0x003fc000}},
+  {{0x0019bdb8, 0x0019bdb8, 0x0019bdb8, 0x003fc000}},
+  {{0x001a083d, 0x001a083d, 0x001a083d, 0x003fc000}},
+  {{0x001a52c2, 0x001a52c2, 0x001a52c2, 0x003fc000}},
+  {{0x001a9d47, 0x001a9d47, 0x001a9d47, 0x003fc000}},
+  {{0x001ae7cc, 0x001ae7cc, 0x001ae7cc, 0x003fc000}},
+  {{0x001b3251, 0x001b3251, 0x001b3251, 0x003fc000}},
+  {{0x001b7cd6, 0x001b7cd6, 0x001b7cd6, 0x003fc000}},
+  {{0x001bc75b, 0x001bc75b, 0x001bc75b, 0x003fc000}},
+  {{0x001c11e0, 0x001c11e0, 0x001c11e0, 0x003fc000}},
+  {{0x001c5c65, 0x001c5c65, 0x001c5c65, 0x003fc000}},
+  {{0x001ca6ea, 0x001ca6ea, 0x001ca6ea, 0x003fc000}},
+  {{0x001cf16f, 0x001cf16f, 0x001cf16f, 0x003fc000}},
+  {{0x001d3bf4, 0x001d3bf4, 0x001d3bf4, 0x003fc000}},
+  {{0x001d8679, 0x001d8679, 0x001d8679, 0x003fc000}},
+  {{0x001dd0fe, 0x001dd0fe, 0x001dd0fe, 0x003fc000}},
+  {{0x001e1b83, 0x001e1b83, 0x001e1b83, 0x003fc000}},
+  {{0x001e6608, 0x001e6608, 0x001e6608, 0x003fc000}},
+  {{0x001eb08d, 0x001eb08d, 0x001eb08d, 0x003fc000}},
+  {{0x001efb12, 0x001efb12, 0x001efb12, 0x003fc000}},
+  {{0x001f4597, 0x001f4597, 0x001f4597, 0x003fc000}},
+  {{0x001f901c, 0x001f901c, 0x001f901c, 0x003fc000}},
+  {{0x001fdaa1, 0x001fdaa1, 0x001fdaa1, 0x003fc000}},
+  {{0x00202526, 0x00202526, 0x00202526, 0x003fc000}},
+  {{0x00206fab, 0x00206fab, 0x00206fab, 0x003fc000}},
+  {{0x0020ba30, 0x0020ba30, 0x0020ba30, 0x003fc000}},
+  {{0x002104b5, 0x002104b5, 0x002104b5, 0x003fc000}},
+  {{0x00214f3a, 0x00214f3a, 0x00214f3a, 0x003fc000}},
+  {{0x002199bf, 0x002199bf, 0x002199bf, 0x003fc000}},
+  {{0x0021e444, 0x0021e444, 0x0021e444, 0x003fc000}},
+  {{0x00222ec9, 0x00222ec9, 0x00222ec9, 0x003fc000}},
+  {{0x0022794e, 0x0022794e, 0x0022794e, 0x003fc000}},
+  {{0x0022c3d3, 0x0022c3d3, 0x0022c3d3, 0x003fc000}},
+  {{0x00230e58, 0x00230e58, 0x00230e58, 0x003fc000}},
+  {{0x002358dd, 0x002358dd, 0x002358dd, 0x003fc000}},
+  {{0x0023a362, 0x0023a362, 0x0023a362, 0x003fc000}},
+  {{0x0023ede7, 0x0023ede7, 0x0023ede7, 0x003fc000}},
+  {{0x0024386c, 0x0024386c, 0x0024386c, 0x003fc000}},
+  {{0x002482f1, 0x002482f1, 0x002482f1, 0x003fc000}},
+  {{0x0024cd76, 0x0024cd76, 0x0024cd76, 0x003fc000}},
+  {{0x002517fb, 0x002517fb, 0x002517fb, 0x003fc000}},
+  {{0x00256280, 0x00256280, 0x00256280, 0x003fc000}},
+  {{0x0025ad05, 0x0025ad05, 0x0025ad05, 0x003fc000}},
+  {{0x0025f78a, 0x0025f78a, 0x0025f78a, 0x003fc000}},
+  {{0x0026420f, 0x0026420f, 0x0026420f, 0x003fc000}},
+  {{0x00268c94, 0x00268c94, 0x00268c94, 0x003fc000}},
+  {{0x0026d719, 0x0026d719, 0x0026d719, 0x003fc000}},
+  {{0x0027219e, 0x0027219e, 0x0027219e, 0x003fc000}},
+  {{0x00276c23, 0x00276c23, 0x00276c23, 0x003fc000}},
+  {{0x0027b6a8, 0x0027b6a8, 0x0027b6a8, 0x003fc000}},
+  {{0x0028012d, 0x0028012d, 0x0028012d, 0x003fc000}},
+  {{0x00284bb2, 0x00284bb2, 0x00284bb2, 0x003fc000}},
+  {{0x00289637, 0x00289637, 0x00289637, 0x003fc000}},
+  {{0x0028e0bc, 0x0028e0bc, 0x0028e0bc, 0x003fc000}},
+  {{0x00292b41, 0x00292b41, 0x00292b41, 0x003fc000}},
+  {{0x002975c6, 0x002975c6, 0x002975c6, 0x003fc000}},
+  {{0x0029c04b, 0x0029c04b, 0x0029c04b, 0x003fc000}},
+  {{0x002a0ad0, 0x002a0ad0, 0x002a0ad0, 0x003fc000}},
+  {{0x002a5555, 0x002a5555, 0x002a5555, 0x003fc000}},
+  {{0x002a9fda, 0x002a9fda, 0x002a9fda, 0x003fc000}},
+  {{0x002aea5f, 0x002aea5f, 0x002aea5f, 0x003fc000}},
+  {{0x002b34e4, 0x002b34e4, 0x002b34e4, 0x003fc000}},
+  {{0x002b7f69, 0x002b7f69, 0x002b7f69, 0x003fc000}},
+  {{0x002bc9ee, 0x002bc9ee, 0x002bc9ee, 0x003fc000}},
+  {{0x002c1473, 0x002c1473, 0x002c1473, 0x003fc000}},
+  {{0x002c5ef8, 0x002c5ef8, 0x002c5ef8, 0x003fc000}},
+  {{0x002ca97d, 0x002ca97d, 0x002ca97d, 0x003fc000}},
+  {{0x002cf402, 0x002cf402, 0x002cf402, 0x003fc000}},
+  {{0x002d3e87, 0x002d3e87, 0x002d3e87, 0x003fc000}},
+  {{0x002d890c, 0x002d890c, 0x002d890c, 0x003fc000}},
+  {{0x002dd391, 0x002dd391, 0x002dd391, 0x003fc000}},
+  {{0x002e1e16, 0x002e1e16, 0x002e1e16, 0x003fc000}},
+  {{0x002e689b, 0x002e689b, 0x002e689b, 0x003fc000}},
+  {{0x002eb320, 0x002eb320, 0x002eb320, 0x003fc000}},
+  {{0x002efda5, 0x002efda5, 0x002efda5, 0x003fc000}},
+  {{0x002f482a, 0x002f482a, 0x002f482a, 0x003fc000}},
+  {{0x002f92af, 0x002f92af, 0x002f92af, 0x003fc000}},
+  {{0x002fdd34, 0x002fdd34, 0x002fdd34, 0x003fc000}},
+  {{0x003027b9, 0x003027b9, 0x003027b9, 0x003fc000}},
+  {{0x0030723e, 0x0030723e, 0x0030723e, 0x003fc000}},
+  {{0x0030bcc3, 0x0030bcc3, 0x0030bcc3, 0x003fc000}},
+  {{0x00310748, 0x00310748, 0x00310748, 0x003fc000}},
+  {{0x003151cd, 0x003151cd, 0x003151cd, 0x003fc000}},
+  {{0x00319c52, 0x00319c52, 0x00319c52, 0x003fc000}},
+  {{0x0031e6d7, 0x0031e6d7, 0x0031e6d7, 0x003fc000}},
+  {{0x0032315c, 0x0032315c, 0x0032315c, 0x003fc000}},
+  {{0x00327be1, 0x00327be1, 0x00327be1, 0x003fc000}},
+  {{0x0032c666, 0x0032c666, 0x0032c666, 0x003fc000}},
+  {{0x003310eb, 0x003310eb, 0x003310eb, 0x003fc000}},
+  {{0x00335b70, 0x00335b70, 0x00335b70, 0x003fc000}},
+  {{0x0033a5f5, 0x0033a5f5, 0x0033a5f5, 0x003fc000}},
+  {{0x0033f07a, 0x0033f07a, 0x0033f07a, 0x003fc000}},
+  {{0x00343aff, 0x00343aff, 0x00343aff, 0x003fc000}},
+  {{0x00348584, 0x00348584, 0x00348584, 0x003fc000}},
+  {{0x0034d009, 0x0034d009, 0x0034d009, 0x003fc000}},
+  {{0x00351a8e, 0x00351a8e, 0x00351a8e, 0x003fc000}},
+  {{0x00356513, 0x00356513, 0x00356513, 0x003fc000}},
+  {{0x0035af98, 0x0035af98, 0x0035af98, 0x003fc000}},
+  {{0x0035fa1d, 0x0035fa1d, 0x0035fa1d, 0x003fc000}},
+  {{0x003644a2, 0x003644a2, 0x003644a2, 0x003fc000}},
+  {{0x00368f27, 0x00368f27, 0x00368f27, 0x003fc000}},
+  {{0x0036d9ac, 0x0036d9ac, 0x0036d9ac, 0x003fc000}},
+  {{0x00372431, 0x00372431, 0x00372431, 0x003fc000}},
+  {{0x00376eb6, 0x00376eb6, 0x00376eb6, 0x003fc000}},
+  {{0x0037b93b, 0x0037b93b, 0x0037b93b, 0x003fc000}},
+  {{0x003803c0, 0x003803c0, 0x003803c0, 0x003fc000}},
+  {{0x00384e45, 0x00384e45, 0x00384e45, 0x003fc000}},
+  {{0x003898ca, 0x003898ca, 0x003898ca, 0x003fc000}},
+  {{0x0038e34f, 0x0038e34f, 0x0038e34f, 0x003fc000}},
+  {{0x00392dd4, 0x00392dd4, 0x00392dd4, 0x003fc000}},
+  {{0x00397859, 0x00397859, 0x00397859, 0x003fc000}},
+  {{0x0039c2de, 0x0039c2de, 0x0039c2de, 0x003fc000}},
+  {{0x003a0d63, 0x003a0d63, 0x003a0d63, 0x003fc000}},
+  {{0x003a57e8, 0x003a57e8, 0x003a57e8, 0x003fc000}},
+  {{0x003aa26d, 0x003aa26d, 0x003aa26d, 0x003fc000}},
+  {{0x003aecf2, 0x003aecf2, 0x003aecf2, 0x003fc000}},
+  {{0x003b3777, 0x003b3777, 0x003b3777, 0x003fc000}},
+  {{0x003b81fc, 0x003b81fc, 0x003b81fc, 0x003fc000}},
+  {{0x003bcc81, 0x003bcc81, 0x003bcc81, 0x003fc000}},
+  {{0x003c1706, 0x003c1706, 0x003c1706, 0x003fc000}},
+  {{0x003c618b, 0x003c618b, 0x003c618b, 0x003fc000}},
+  {{0x003cac10, 0x003cac10, 0x003cac10, 0x003fc000}},
+  {{0x003cf695, 0x003cf695, 0x003cf695, 0x003fc000}},
+  {{0x003d411a, 0x003d411a, 0x003d411a, 0x003fc000}},
+  {{0x003d8b9f, 0x003d8b9f, 0x003d8b9f, 0x003fc000}},
+  {{0x003dd624, 0x003dd624, 0x003dd624, 0x003fc000}},
+  {{0x003e20a9, 0x003e20a9, 0x003e20a9, 0x003fc000}},
+  {{0x003e6b2e, 0x003e6b2e, 0x003e6b2e, 0x003fc000}},
+  {{0x003eb5b3, 0x003eb5b3, 0x003eb5b3, 0x003fc000}},
+  {{0x003f0038, 0x003f0038, 0x003f0038, 0x003fc000}},
+  {{0x003f4abd, 0x003f4abd, 0x003f4abd, 0x003fc000}},
+  {{0x003f9542, 0x003f9542, 0x003f9542, 0x003fc000}},
+  {{0x003fdfc7, 0x003fdfc7, 0x003fdfc7, 0x003fc000}},
+  {{0x00402a4c, 0x00402a4c, 0x00402a4c, 0x003fc000}},
+  {{0x004074d1, 0x004074d1, 0x004074d1, 0x003fc000}},
+  {{0x0040bf56, 0x0040bf56, 0x0040bf56, 0x003fc000}},
+  {{0x004109db, 0x004109db, 0x004109db, 0x003fc000}},
+  {{0x00415460, 0x00415460, 0x00415460, 0x003fc000}},
+  {{0x00419ee5, 0x00419ee5, 0x00419ee5, 0x003fc000}},
+  {{0x0041e96a, 0x0041e96a, 0x0041e96a, 0x003fc000}},
+  {{0x004233ef, 0x004233ef, 0x004233ef, 0x003fc000}},
+  {{0x00427e74, 0x00427e74, 0x00427e74, 0x003fc000}},
+  {{0x0042c8f9, 0x0042c8f9, 0x0042c8f9, 0x003fc000}},
+  {{0x0043137e, 0x0043137e, 0x0043137e, 0x003fc000}},
+  {{0x00435e03, 0x00435e03, 0x00435e03, 0x003fc000}},
+  {{0x0043a888, 0x0043a888, 0x0043a888, 0x003fc000}},
+  {{0x0043f30d, 0x0043f30d, 0x0043f30d, 0x003fc000}},
+  {{0x00443d92, 0x00443d92, 0x00443d92, 0x003fc000}},
+  {{0x00448817, 0x00448817, 0x00448817, 0x003fc000}},
+  {{0x0044d29c, 0x0044d29c, 0x0044d29c, 0x003fc000}},
+  {{0x00451d21, 0x00451d21, 0x00451d21, 0x003fc000}},
+  {{0x004567a6, 0x004567a6, 0x004567a6, 0x003fc000}},
+  {{0x0045b22b, 0x0045b22b, 0x0045b22b, 0x003fc000}}
+};
+
+static const VP8kCstSSE2 VP8kUtoRGBA[256] = {
+  {{0, 0x000c8980, 0xffbf7300, 0}}, {{0, 0x000c706d, 0xffbff41a, 0}},
+  {{0, 0x000c575a, 0xffc07534, 0}}, {{0, 0x000c3e47, 0xffc0f64e, 0}},
+  {{0, 0x000c2534, 0xffc17768, 0}}, {{0, 0x000c0c21, 0xffc1f882, 0}},
+  {{0, 0x000bf30e, 0xffc2799c, 0}}, {{0, 0x000bd9fb, 0xffc2fab6, 0}},
+  {{0, 0x000bc0e8, 0xffc37bd0, 0}}, {{0, 0x000ba7d5, 0xffc3fcea, 0}},
+  {{0, 0x000b8ec2, 0xffc47e04, 0}}, {{0, 0x000b75af, 0xffc4ff1e, 0}},
+  {{0, 0x000b5c9c, 0xffc58038, 0}}, {{0, 0x000b4389, 0xffc60152, 0}},
+  {{0, 0x000b2a76, 0xffc6826c, 0}}, {{0, 0x000b1163, 0xffc70386, 0}},
+  {{0, 0x000af850, 0xffc784a0, 0}}, {{0, 0x000adf3d, 0xffc805ba, 0}},
+  {{0, 0x000ac62a, 0xffc886d4, 0}}, {{0, 0x000aad17, 0xffc907ee, 0}},
+  {{0, 0x000a9404, 0xffc98908, 0}}, {{0, 0x000a7af1, 0xffca0a22, 0}},
+  {{0, 0x000a61de, 0xffca8b3c, 0}}, {{0, 0x000a48cb, 0xffcb0c56, 0}},
+  {{0, 0x000a2fb8, 0xffcb8d70, 0}}, {{0, 0x000a16a5, 0xffcc0e8a, 0}},
+  {{0, 0x0009fd92, 0xffcc8fa4, 0}}, {{0, 0x0009e47f, 0xffcd10be, 0}},
+  {{0, 0x0009cb6c, 0xffcd91d8, 0}}, {{0, 0x0009b259, 0xffce12f2, 0}},
+  {{0, 0x00099946, 0xffce940c, 0}}, {{0, 0x00098033, 0xffcf1526, 0}},
+  {{0, 0x00096720, 0xffcf9640, 0}}, {{0, 0x00094e0d, 0xffd0175a, 0}},
+  {{0, 0x000934fa, 0xffd09874, 0}}, {{0, 0x00091be7, 0xffd1198e, 0}},
+  {{0, 0x000902d4, 0xffd19aa8, 0}}, {{0, 0x0008e9c1, 0xffd21bc2, 0}},
+  {{0, 0x0008d0ae, 0xffd29cdc, 0}}, {{0, 0x0008b79b, 0xffd31df6, 0}},
+  {{0, 0x00089e88, 0xffd39f10, 0}}, {{0, 0x00088575, 0xffd4202a, 0}},
+  {{0, 0x00086c62, 0xffd4a144, 0}}, {{0, 0x0008534f, 0xffd5225e, 0}},
+  {{0, 0x00083a3c, 0xffd5a378, 0}}, {{0, 0x00082129, 0xffd62492, 0}},
+  {{0, 0x00080816, 0xffd6a5ac, 0}}, {{0, 0x0007ef03, 0xffd726c6, 0}},
+  {{0, 0x0007d5f0, 0xffd7a7e0, 0}}, {{0, 0x0007bcdd, 0xffd828fa, 0}},
+  {{0, 0x0007a3ca, 0xffd8aa14, 0}}, {{0, 0x00078ab7, 0xffd92b2e, 0}},
+  {{0, 0x000771a4, 0xffd9ac48, 0}}, {{0, 0x00075891, 0xffda2d62, 0}},
+  {{0, 0x00073f7e, 0xffdaae7c, 0}}, {{0, 0x0007266b, 0xffdb2f96, 0}},
+  {{0, 0x00070d58, 0xffdbb0b0, 0}}, {{0, 0x0006f445, 0xffdc31ca, 0}},
+  {{0, 0x0006db32, 0xffdcb2e4, 0}}, {{0, 0x0006c21f, 0xffdd33fe, 0}},
+  {{0, 0x0006a90c, 0xffddb518, 0}}, {{0, 0x00068ff9, 0xffde3632, 0}},
+  {{0, 0x000676e6, 0xffdeb74c, 0}}, {{0, 0x00065dd3, 0xffdf3866, 0}},
+  {{0, 0x000644c0, 0xffdfb980, 0}}, {{0, 0x00062bad, 0xffe03a9a, 0}},
+  {{0, 0x0006129a, 0xffe0bbb4, 0}}, {{0, 0x0005f987, 0xffe13cce, 0}},
+  {{0, 0x0005e074, 0xffe1bde8, 0}}, {{0, 0x0005c761, 0xffe23f02, 0}},
+  {{0, 0x0005ae4e, 0xffe2c01c, 0}}, {{0, 0x0005953b, 0xffe34136, 0}},
+  {{0, 0x00057c28, 0xffe3c250, 0}}, {{0, 0x00056315, 0xffe4436a, 0}},
+  {{0, 0x00054a02, 0xffe4c484, 0}}, {{0, 0x000530ef, 0xffe5459e, 0}},
+  {{0, 0x000517dc, 0xffe5c6b8, 0}}, {{0, 0x0004fec9, 0xffe647d2, 0}},
+  {{0, 0x0004e5b6, 0xffe6c8ec, 0}}, {{0, 0x0004cca3, 0xffe74a06, 0}},
+  {{0, 0x0004b390, 0xffe7cb20, 0}}, {{0, 0x00049a7d, 0xffe84c3a, 0}},
+  {{0, 0x0004816a, 0xffe8cd54, 0}}, {{0, 0x00046857, 0xffe94e6e, 0}},
+  {{0, 0x00044f44, 0xffe9cf88, 0}}, {{0, 0x00043631, 0xffea50a2, 0}},
+  {{0, 0x00041d1e, 0xffead1bc, 0}}, {{0, 0x0004040b, 0xffeb52d6, 0}},
+  {{0, 0x0003eaf8, 0xffebd3f0, 0}}, {{0, 0x0003d1e5, 0xffec550a, 0}},
+  {{0, 0x0003b8d2, 0xffecd624, 0}}, {{0, 0x00039fbf, 0xffed573e, 0}},
+  {{0, 0x000386ac, 0xffedd858, 0}}, {{0, 0x00036d99, 0xffee5972, 0}},
+  {{0, 0x00035486, 0xffeeda8c, 0}}, {{0, 0x00033b73, 0xffef5ba6, 0}},
+  {{0, 0x00032260, 0xffefdcc0, 0}}, {{0, 0x0003094d, 0xfff05dda, 0}},
+  {{0, 0x0002f03a, 0xfff0def4, 0}}, {{0, 0x0002d727, 0xfff1600e, 0}},
+  {{0, 0x0002be14, 0xfff1e128, 0}}, {{0, 0x0002a501, 0xfff26242, 0}},
+  {{0, 0x00028bee, 0xfff2e35c, 0}}, {{0, 0x000272db, 0xfff36476, 0}},
+  {{0, 0x000259c8, 0xfff3e590, 0}}, {{0, 0x000240b5, 0xfff466aa, 0}},
+  {{0, 0x000227a2, 0xfff4e7c4, 0}}, {{0, 0x00020e8f, 0xfff568de, 0}},
+  {{0, 0x0001f57c, 0xfff5e9f8, 0}}, {{0, 0x0001dc69, 0xfff66b12, 0}},
+  {{0, 0x0001c356, 0xfff6ec2c, 0}}, {{0, 0x0001aa43, 0xfff76d46, 0}},
+  {{0, 0x00019130, 0xfff7ee60, 0}}, {{0, 0x0001781d, 0xfff86f7a, 0}},
+  {{0, 0x00015f0a, 0xfff8f094, 0}}, {{0, 0x000145f7, 0xfff971ae, 0}},
+  {{0, 0x00012ce4, 0xfff9f2c8, 0}}, {{0, 0x000113d1, 0xfffa73e2, 0}},
+  {{0, 0x0000fabe, 0xfffaf4fc, 0}}, {{0, 0x0000e1ab, 0xfffb7616, 0}},
+  {{0, 0x0000c898, 0xfffbf730, 0}}, {{0, 0x0000af85, 0xfffc784a, 0}},
+  {{0, 0x00009672, 0xfffcf964, 0}}, {{0, 0x00007d5f, 0xfffd7a7e, 0}},
+  {{0, 0x0000644c, 0xfffdfb98, 0}}, {{0, 0x00004b39, 0xfffe7cb2, 0}},
+  {{0, 0x00003226, 0xfffefdcc, 0}}, {{0, 0x00001913, 0xffff7ee6, 0}},
+  {{0, 0x00000000, 0x00000000, 0}}, {{0, 0xffffe6ed, 0x0000811a, 0}},
+  {{0, 0xffffcdda, 0x00010234, 0}}, {{0, 0xffffb4c7, 0x0001834e, 0}},
+  {{0, 0xffff9bb4, 0x00020468, 0}}, {{0, 0xffff82a1, 0x00028582, 0}},
+  {{0, 0xffff698e, 0x0003069c, 0}}, {{0, 0xffff507b, 0x000387b6, 0}},
+  {{0, 0xffff3768, 0x000408d0, 0}}, {{0, 0xffff1e55, 0x000489ea, 0}},
+  {{0, 0xffff0542, 0x00050b04, 0}}, {{0, 0xfffeec2f, 0x00058c1e, 0}},
+  {{0, 0xfffed31c, 0x00060d38, 0}}, {{0, 0xfffeba09, 0x00068e52, 0}},
+  {{0, 0xfffea0f6, 0x00070f6c, 0}}, {{0, 0xfffe87e3, 0x00079086, 0}},
+  {{0, 0xfffe6ed0, 0x000811a0, 0}}, {{0, 0xfffe55bd, 0x000892ba, 0}},
+  {{0, 0xfffe3caa, 0x000913d4, 0}}, {{0, 0xfffe2397, 0x000994ee, 0}},
+  {{0, 0xfffe0a84, 0x000a1608, 0}}, {{0, 0xfffdf171, 0x000a9722, 0}},
+  {{0, 0xfffdd85e, 0x000b183c, 0}}, {{0, 0xfffdbf4b, 0x000b9956, 0}},
+  {{0, 0xfffda638, 0x000c1a70, 0}}, {{0, 0xfffd8d25, 0x000c9b8a, 0}},
+  {{0, 0xfffd7412, 0x000d1ca4, 0}}, {{0, 0xfffd5aff, 0x000d9dbe, 0}},
+  {{0, 0xfffd41ec, 0x000e1ed8, 0}}, {{0, 0xfffd28d9, 0x000e9ff2, 0}},
+  {{0, 0xfffd0fc6, 0x000f210c, 0}}, {{0, 0xfffcf6b3, 0x000fa226, 0}},
+  {{0, 0xfffcdda0, 0x00102340, 0}}, {{0, 0xfffcc48d, 0x0010a45a, 0}},
+  {{0, 0xfffcab7a, 0x00112574, 0}}, {{0, 0xfffc9267, 0x0011a68e, 0}},
+  {{0, 0xfffc7954, 0x001227a8, 0}}, {{0, 0xfffc6041, 0x0012a8c2, 0}},
+  {{0, 0xfffc472e, 0x001329dc, 0}}, {{0, 0xfffc2e1b, 0x0013aaf6, 0}},
+  {{0, 0xfffc1508, 0x00142c10, 0}}, {{0, 0xfffbfbf5, 0x0014ad2a, 0}},
+  {{0, 0xfffbe2e2, 0x00152e44, 0}}, {{0, 0xfffbc9cf, 0x0015af5e, 0}},
+  {{0, 0xfffbb0bc, 0x00163078, 0}}, {{0, 0xfffb97a9, 0x0016b192, 0}},
+  {{0, 0xfffb7e96, 0x001732ac, 0}}, {{0, 0xfffb6583, 0x0017b3c6, 0}},
+  {{0, 0xfffb4c70, 0x001834e0, 0}}, {{0, 0xfffb335d, 0x0018b5fa, 0}},
+  {{0, 0xfffb1a4a, 0x00193714, 0}}, {{0, 0xfffb0137, 0x0019b82e, 0}},
+  {{0, 0xfffae824, 0x001a3948, 0}}, {{0, 0xfffacf11, 0x001aba62, 0}},
+  {{0, 0xfffab5fe, 0x001b3b7c, 0}}, {{0, 0xfffa9ceb, 0x001bbc96, 0}},
+  {{0, 0xfffa83d8, 0x001c3db0, 0}}, {{0, 0xfffa6ac5, 0x001cbeca, 0}},
+  {{0, 0xfffa51b2, 0x001d3fe4, 0}}, {{0, 0xfffa389f, 0x001dc0fe, 0}},
+  {{0, 0xfffa1f8c, 0x001e4218, 0}}, {{0, 0xfffa0679, 0x001ec332, 0}},
+  {{0, 0xfff9ed66, 0x001f444c, 0}}, {{0, 0xfff9d453, 0x001fc566, 0}},
+  {{0, 0xfff9bb40, 0x00204680, 0}}, {{0, 0xfff9a22d, 0x0020c79a, 0}},
+  {{0, 0xfff9891a, 0x002148b4, 0}}, {{0, 0xfff97007, 0x0021c9ce, 0}},
+  {{0, 0xfff956f4, 0x00224ae8, 0}}, {{0, 0xfff93de1, 0x0022cc02, 0}},
+  {{0, 0xfff924ce, 0x00234d1c, 0}}, {{0, 0xfff90bbb, 0x0023ce36, 0}},
+  {{0, 0xfff8f2a8, 0x00244f50, 0}}, {{0, 0xfff8d995, 0x0024d06a, 0}},
+  {{0, 0xfff8c082, 0x00255184, 0}}, {{0, 0xfff8a76f, 0x0025d29e, 0}},
+  {{0, 0xfff88e5c, 0x002653b8, 0}}, {{0, 0xfff87549, 0x0026d4d2, 0}},
+  {{0, 0xfff85c36, 0x002755ec, 0}}, {{0, 0xfff84323, 0x0027d706, 0}},
+  {{0, 0xfff82a10, 0x00285820, 0}}, {{0, 0xfff810fd, 0x0028d93a, 0}},
+  {{0, 0xfff7f7ea, 0x00295a54, 0}}, {{0, 0xfff7ded7, 0x0029db6e, 0}},
+  {{0, 0xfff7c5c4, 0x002a5c88, 0}}, {{0, 0xfff7acb1, 0x002adda2, 0}},
+  {{0, 0xfff7939e, 0x002b5ebc, 0}}, {{0, 0xfff77a8b, 0x002bdfd6, 0}},
+  {{0, 0xfff76178, 0x002c60f0, 0}}, {{0, 0xfff74865, 0x002ce20a, 0}},
+  {{0, 0xfff72f52, 0x002d6324, 0}}, {{0, 0xfff7163f, 0x002de43e, 0}},
+  {{0, 0xfff6fd2c, 0x002e6558, 0}}, {{0, 0xfff6e419, 0x002ee672, 0}},
+  {{0, 0xfff6cb06, 0x002f678c, 0}}, {{0, 0xfff6b1f3, 0x002fe8a6, 0}},
+  {{0, 0xfff698e0, 0x003069c0, 0}}, {{0, 0xfff67fcd, 0x0030eada, 0}},
+  {{0, 0xfff666ba, 0x00316bf4, 0}}, {{0, 0xfff64da7, 0x0031ed0e, 0}},
+  {{0, 0xfff63494, 0x00326e28, 0}}, {{0, 0xfff61b81, 0x0032ef42, 0}},
+  {{0, 0xfff6026e, 0x0033705c, 0}}, {{0, 0xfff5e95b, 0x0033f176, 0}},
+  {{0, 0xfff5d048, 0x00347290, 0}}, {{0, 0xfff5b735, 0x0034f3aa, 0}},
+  {{0, 0xfff59e22, 0x003574c4, 0}}, {{0, 0xfff5850f, 0x0035f5de, 0}},
+  {{0, 0xfff56bfc, 0x003676f8, 0}}, {{0, 0xfff552e9, 0x0036f812, 0}},
+  {{0, 0xfff539d6, 0x0037792c, 0}}, {{0, 0xfff520c3, 0x0037fa46, 0}},
+  {{0, 0xfff507b0, 0x00387b60, 0}}, {{0, 0xfff4ee9d, 0x0038fc7a, 0}},
+  {{0, 0xfff4d58a, 0x00397d94, 0}}, {{0, 0xfff4bc77, 0x0039feae, 0}},
+  {{0, 0xfff4a364, 0x003a7fc8, 0}}, {{0, 0xfff48a51, 0x003b00e2, 0}},
+  {{0, 0xfff4713e, 0x003b81fc, 0}}, {{0, 0xfff4582b, 0x003c0316, 0}},
+  {{0, 0xfff43f18, 0x003c8430, 0}}, {{0, 0xfff42605, 0x003d054a, 0}},
+  {{0, 0xfff40cf2, 0x003d8664, 0}}, {{0, 0xfff3f3df, 0x003e077e, 0}},
+  {{0, 0xfff3dacc, 0x003e8898, 0}}, {{0, 0xfff3c1b9, 0x003f09b2, 0}},
+  {{0, 0xfff3a8a6, 0x003f8acc, 0}}, {{0, 0xfff38f93, 0x00400be6, 0}}
+};
+
+static VP8kCstSSE2 VP8kVtoRGBA[256] = {
+  {{0xffcced80, 0x001a0400, 0, 0}}, {{0xffcd53a5, 0x0019cff8, 0, 0}},
+  {{0xffcdb9ca, 0x00199bf0, 0, 0}}, {{0xffce1fef, 0x001967e8, 0, 0}},
+  {{0xffce8614, 0x001933e0, 0, 0}}, {{0xffceec39, 0x0018ffd8, 0, 0}},
+  {{0xffcf525e, 0x0018cbd0, 0, 0}}, {{0xffcfb883, 0x001897c8, 0, 0}},
+  {{0xffd01ea8, 0x001863c0, 0, 0}}, {{0xffd084cd, 0x00182fb8, 0, 0}},
+  {{0xffd0eaf2, 0x0017fbb0, 0, 0}}, {{0xffd15117, 0x0017c7a8, 0, 0}},
+  {{0xffd1b73c, 0x001793a0, 0, 0}}, {{0xffd21d61, 0x00175f98, 0, 0}},
+  {{0xffd28386, 0x00172b90, 0, 0}}, {{0xffd2e9ab, 0x0016f788, 0, 0}},
+  {{0xffd34fd0, 0x0016c380, 0, 0}}, {{0xffd3b5f5, 0x00168f78, 0, 0}},
+  {{0xffd41c1a, 0x00165b70, 0, 0}}, {{0xffd4823f, 0x00162768, 0, 0}},
+  {{0xffd4e864, 0x0015f360, 0, 0}}, {{0xffd54e89, 0x0015bf58, 0, 0}},
+  {{0xffd5b4ae, 0x00158b50, 0, 0}}, {{0xffd61ad3, 0x00155748, 0, 0}},
+  {{0xffd680f8, 0x00152340, 0, 0}}, {{0xffd6e71d, 0x0014ef38, 0, 0}},
+  {{0xffd74d42, 0x0014bb30, 0, 0}}, {{0xffd7b367, 0x00148728, 0, 0}},
+  {{0xffd8198c, 0x00145320, 0, 0}}, {{0xffd87fb1, 0x00141f18, 0, 0}},
+  {{0xffd8e5d6, 0x0013eb10, 0, 0}}, {{0xffd94bfb, 0x0013b708, 0, 0}},
+  {{0xffd9b220, 0x00138300, 0, 0}}, {{0xffda1845, 0x00134ef8, 0, 0}},
+  {{0xffda7e6a, 0x00131af0, 0, 0}}, {{0xffdae48f, 0x0012e6e8, 0, 0}},
+  {{0xffdb4ab4, 0x0012b2e0, 0, 0}}, {{0xffdbb0d9, 0x00127ed8, 0, 0}},
+  {{0xffdc16fe, 0x00124ad0, 0, 0}}, {{0xffdc7d23, 0x001216c8, 0, 0}},
+  {{0xffdce348, 0x0011e2c0, 0, 0}}, {{0xffdd496d, 0x0011aeb8, 0, 0}},
+  {{0xffddaf92, 0x00117ab0, 0, 0}}, {{0xffde15b7, 0x001146a8, 0, 0}},
+  {{0xffde7bdc, 0x001112a0, 0, 0}}, {{0xffdee201, 0x0010de98, 0, 0}},
+  {{0xffdf4826, 0x0010aa90, 0, 0}}, {{0xffdfae4b, 0x00107688, 0, 0}},
+  {{0xffe01470, 0x00104280, 0, 0}}, {{0xffe07a95, 0x00100e78, 0, 0}},
+  {{0xffe0e0ba, 0x000fda70, 0, 0}}, {{0xffe146df, 0x000fa668, 0, 0}},
+  {{0xffe1ad04, 0x000f7260, 0, 0}}, {{0xffe21329, 0x000f3e58, 0, 0}},
+  {{0xffe2794e, 0x000f0a50, 0, 0}}, {{0xffe2df73, 0x000ed648, 0, 0}},
+  {{0xffe34598, 0x000ea240, 0, 0}}, {{0xffe3abbd, 0x000e6e38, 0, 0}},
+  {{0xffe411e2, 0x000e3a30, 0, 0}}, {{0xffe47807, 0x000e0628, 0, 0}},
+  {{0xffe4de2c, 0x000dd220, 0, 0}}, {{0xffe54451, 0x000d9e18, 0, 0}},
+  {{0xffe5aa76, 0x000d6a10, 0, 0}}, {{0xffe6109b, 0x000d3608, 0, 0}},
+  {{0xffe676c0, 0x000d0200, 0, 0}}, {{0xffe6dce5, 0x000ccdf8, 0, 0}},
+  {{0xffe7430a, 0x000c99f0, 0, 0}}, {{0xffe7a92f, 0x000c65e8, 0, 0}},
+  {{0xffe80f54, 0x000c31e0, 0, 0}}, {{0xffe87579, 0x000bfdd8, 0, 0}},
+  {{0xffe8db9e, 0x000bc9d0, 0, 0}}, {{0xffe941c3, 0x000b95c8, 0, 0}},
+  {{0xffe9a7e8, 0x000b61c0, 0, 0}}, {{0xffea0e0d, 0x000b2db8, 0, 0}},
+  {{0xffea7432, 0x000af9b0, 0, 0}}, {{0xffeada57, 0x000ac5a8, 0, 0}},
+  {{0xffeb407c, 0x000a91a0, 0, 0}}, {{0xffeba6a1, 0x000a5d98, 0, 0}},
+  {{0xffec0cc6, 0x000a2990, 0, 0}}, {{0xffec72eb, 0x0009f588, 0, 0}},
+  {{0xffecd910, 0x0009c180, 0, 0}}, {{0xffed3f35, 0x00098d78, 0, 0}},
+  {{0xffeda55a, 0x00095970, 0, 0}}, {{0xffee0b7f, 0x00092568, 0, 0}},
+  {{0xffee71a4, 0x0008f160, 0, 0}}, {{0xffeed7c9, 0x0008bd58, 0, 0}},
+  {{0xffef3dee, 0x00088950, 0, 0}}, {{0xffefa413, 0x00085548, 0, 0}},
+  {{0xfff00a38, 0x00082140, 0, 0}}, {{0xfff0705d, 0x0007ed38, 0, 0}},
+  {{0xfff0d682, 0x0007b930, 0, 0}}, {{0xfff13ca7, 0x00078528, 0, 0}},
+  {{0xfff1a2cc, 0x00075120, 0, 0}}, {{0xfff208f1, 0x00071d18, 0, 0}},
+  {{0xfff26f16, 0x0006e910, 0, 0}}, {{0xfff2d53b, 0x0006b508, 0, 0}},
+  {{0xfff33b60, 0x00068100, 0, 0}}, {{0xfff3a185, 0x00064cf8, 0, 0}},
+  {{0xfff407aa, 0x000618f0, 0, 0}}, {{0xfff46dcf, 0x0005e4e8, 0, 0}},
+  {{0xfff4d3f4, 0x0005b0e0, 0, 0}}, {{0xfff53a19, 0x00057cd8, 0, 0}},
+  {{0xfff5a03e, 0x000548d0, 0, 0}}, {{0xfff60663, 0x000514c8, 0, 0}},
+  {{0xfff66c88, 0x0004e0c0, 0, 0}}, {{0xfff6d2ad, 0x0004acb8, 0, 0}},
+  {{0xfff738d2, 0x000478b0, 0, 0}}, {{0xfff79ef7, 0x000444a8, 0, 0}},
+  {{0xfff8051c, 0x000410a0, 0, 0}}, {{0xfff86b41, 0x0003dc98, 0, 0}},
+  {{0xfff8d166, 0x0003a890, 0, 0}}, {{0xfff9378b, 0x00037488, 0, 0}},
+  {{0xfff99db0, 0x00034080, 0, 0}}, {{0xfffa03d5, 0x00030c78, 0, 0}},
+  {{0xfffa69fa, 0x0002d870, 0, 0}}, {{0xfffad01f, 0x0002a468, 0, 0}},
+  {{0xfffb3644, 0x00027060, 0, 0}}, {{0xfffb9c69, 0x00023c58, 0, 0}},
+  {{0xfffc028e, 0x00020850, 0, 0}}, {{0xfffc68b3, 0x0001d448, 0, 0}},
+  {{0xfffcced8, 0x0001a040, 0, 0}}, {{0xfffd34fd, 0x00016c38, 0, 0}},
+  {{0xfffd9b22, 0x00013830, 0, 0}}, {{0xfffe0147, 0x00010428, 0, 0}},
+  {{0xfffe676c, 0x0000d020, 0, 0}}, {{0xfffecd91, 0x00009c18, 0, 0}},
+  {{0xffff33b6, 0x00006810, 0, 0}}, {{0xffff99db, 0x00003408, 0, 0}},
+  {{0x00000000, 0x00000000, 0, 0}}, {{0x00006625, 0xffffcbf8, 0, 0}},
+  {{0x0000cc4a, 0xffff97f0, 0, 0}}, {{0x0001326f, 0xffff63e8, 0, 0}},
+  {{0x00019894, 0xffff2fe0, 0, 0}}, {{0x0001feb9, 0xfffefbd8, 0, 0}},
+  {{0x000264de, 0xfffec7d0, 0, 0}}, {{0x0002cb03, 0xfffe93c8, 0, 0}},
+  {{0x00033128, 0xfffe5fc0, 0, 0}}, {{0x0003974d, 0xfffe2bb8, 0, 0}},
+  {{0x0003fd72, 0xfffdf7b0, 0, 0}}, {{0x00046397, 0xfffdc3a8, 0, 0}},
+  {{0x0004c9bc, 0xfffd8fa0, 0, 0}}, {{0x00052fe1, 0xfffd5b98, 0, 0}},
+  {{0x00059606, 0xfffd2790, 0, 0}}, {{0x0005fc2b, 0xfffcf388, 0, 0}},
+  {{0x00066250, 0xfffcbf80, 0, 0}}, {{0x0006c875, 0xfffc8b78, 0, 0}},
+  {{0x00072e9a, 0xfffc5770, 0, 0}}, {{0x000794bf, 0xfffc2368, 0, 0}},
+  {{0x0007fae4, 0xfffbef60, 0, 0}}, {{0x00086109, 0xfffbbb58, 0, 0}},
+  {{0x0008c72e, 0xfffb8750, 0, 0}}, {{0x00092d53, 0xfffb5348, 0, 0}},
+  {{0x00099378, 0xfffb1f40, 0, 0}}, {{0x0009f99d, 0xfffaeb38, 0, 0}},
+  {{0x000a5fc2, 0xfffab730, 0, 0}}, {{0x000ac5e7, 0xfffa8328, 0, 0}},
+  {{0x000b2c0c, 0xfffa4f20, 0, 0}}, {{0x000b9231, 0xfffa1b18, 0, 0}},
+  {{0x000bf856, 0xfff9e710, 0, 0}}, {{0x000c5e7b, 0xfff9b308, 0, 0}},
+  {{0x000cc4a0, 0xfff97f00, 0, 0}}, {{0x000d2ac5, 0xfff94af8, 0, 0}},
+  {{0x000d90ea, 0xfff916f0, 0, 0}}, {{0x000df70f, 0xfff8e2e8, 0, 0}},
+  {{0x000e5d34, 0xfff8aee0, 0, 0}}, {{0x000ec359, 0xfff87ad8, 0, 0}},
+  {{0x000f297e, 0xfff846d0, 0, 0}}, {{0x000f8fa3, 0xfff812c8, 0, 0}},
+  {{0x000ff5c8, 0xfff7dec0, 0, 0}}, {{0x00105bed, 0xfff7aab8, 0, 0}},
+  {{0x0010c212, 0xfff776b0, 0, 0}}, {{0x00112837, 0xfff742a8, 0, 0}},
+  {{0x00118e5c, 0xfff70ea0, 0, 0}}, {{0x0011f481, 0xfff6da98, 0, 0}},
+  {{0x00125aa6, 0xfff6a690, 0, 0}}, {{0x0012c0cb, 0xfff67288, 0, 0}},
+  {{0x001326f0, 0xfff63e80, 0, 0}}, {{0x00138d15, 0xfff60a78, 0, 0}},
+  {{0x0013f33a, 0xfff5d670, 0, 0}}, {{0x0014595f, 0xfff5a268, 0, 0}},
+  {{0x0014bf84, 0xfff56e60, 0, 0}}, {{0x001525a9, 0xfff53a58, 0, 0}},
+  {{0x00158bce, 0xfff50650, 0, 0}}, {{0x0015f1f3, 0xfff4d248, 0, 0}},
+  {{0x00165818, 0xfff49e40, 0, 0}}, {{0x0016be3d, 0xfff46a38, 0, 0}},
+  {{0x00172462, 0xfff43630, 0, 0}}, {{0x00178a87, 0xfff40228, 0, 0}},
+  {{0x0017f0ac, 0xfff3ce20, 0, 0}}, {{0x001856d1, 0xfff39a18, 0, 0}},
+  {{0x0018bcf6, 0xfff36610, 0, 0}}, {{0x0019231b, 0xfff33208, 0, 0}},
+  {{0x00198940, 0xfff2fe00, 0, 0}}, {{0x0019ef65, 0xfff2c9f8, 0, 0}},
+  {{0x001a558a, 0xfff295f0, 0, 0}}, {{0x001abbaf, 0xfff261e8, 0, 0}},
+  {{0x001b21d4, 0xfff22de0, 0, 0}}, {{0x001b87f9, 0xfff1f9d8, 0, 0}},
+  {{0x001bee1e, 0xfff1c5d0, 0, 0}}, {{0x001c5443, 0xfff191c8, 0, 0}},
+  {{0x001cba68, 0xfff15dc0, 0, 0}}, {{0x001d208d, 0xfff129b8, 0, 0}},
+  {{0x001d86b2, 0xfff0f5b0, 0, 0}}, {{0x001decd7, 0xfff0c1a8, 0, 0}},
+  {{0x001e52fc, 0xfff08da0, 0, 0}}, {{0x001eb921, 0xfff05998, 0, 0}},
+  {{0x001f1f46, 0xfff02590, 0, 0}}, {{0x001f856b, 0xffeff188, 0, 0}},
+  {{0x001feb90, 0xffefbd80, 0, 0}}, {{0x002051b5, 0xffef8978, 0, 0}},
+  {{0x0020b7da, 0xffef5570, 0, 0}}, {{0x00211dff, 0xffef2168, 0, 0}},
+  {{0x00218424, 0xffeeed60, 0, 0}}, {{0x0021ea49, 0xffeeb958, 0, 0}},
+  {{0x0022506e, 0xffee8550, 0, 0}}, {{0x0022b693, 0xffee5148, 0, 0}},
+  {{0x00231cb8, 0xffee1d40, 0, 0}}, {{0x002382dd, 0xffede938, 0, 0}},
+  {{0x0023e902, 0xffedb530, 0, 0}}, {{0x00244f27, 0xffed8128, 0, 0}},
+  {{0x0024b54c, 0xffed4d20, 0, 0}}, {{0x00251b71, 0xffed1918, 0, 0}},
+  {{0x00258196, 0xffece510, 0, 0}}, {{0x0025e7bb, 0xffecb108, 0, 0}},
+  {{0x00264de0, 0xffec7d00, 0, 0}}, {{0x0026b405, 0xffec48f8, 0, 0}},
+  {{0x00271a2a, 0xffec14f0, 0, 0}}, {{0x0027804f, 0xffebe0e8, 0, 0}},
+  {{0x0027e674, 0xffebace0, 0, 0}}, {{0x00284c99, 0xffeb78d8, 0, 0}},
+  {{0x0028b2be, 0xffeb44d0, 0, 0}}, {{0x002918e3, 0xffeb10c8, 0, 0}},
+  {{0x00297f08, 0xffeadcc0, 0, 0}}, {{0x0029e52d, 0xffeaa8b8, 0, 0}},
+  {{0x002a4b52, 0xffea74b0, 0, 0}}, {{0x002ab177, 0xffea40a8, 0, 0}},
+  {{0x002b179c, 0xffea0ca0, 0, 0}}, {{0x002b7dc1, 0xffe9d898, 0, 0}},
+  {{0x002be3e6, 0xffe9a490, 0, 0}}, {{0x002c4a0b, 0xffe97088, 0, 0}},
+  {{0x002cb030, 0xffe93c80, 0, 0}}, {{0x002d1655, 0xffe90878, 0, 0}},
+  {{0x002d7c7a, 0xffe8d470, 0, 0}}, {{0x002de29f, 0xffe8a068, 0, 0}},
+  {{0x002e48c4, 0xffe86c60, 0, 0}}, {{0x002eaee9, 0xffe83858, 0, 0}},
+  {{0x002f150e, 0xffe80450, 0, 0}}, {{0x002f7b33, 0xffe7d048, 0, 0}},
+  {{0x002fe158, 0xffe79c40, 0, 0}}, {{0x0030477d, 0xffe76838, 0, 0}},
+  {{0x0030ada2, 0xffe73430, 0, 0}}, {{0x003113c7, 0xffe70028, 0, 0}},
+  {{0x003179ec, 0xffe6cc20, 0, 0}}, {{0x0031e011, 0xffe69818, 0, 0}},
+  {{0x00324636, 0xffe66410, 0, 0}}, {{0x0032ac5b, 0xffe63008, 0, 0}}
+};

src/enc/Makefile.am

@@ -1,4 +1,3 @@
-AM_CPPFLAGS = -I$(top_srcdir)/src
 noinst_LTLIBRARIES = libwebpencode.la
 
 libwebpencode_la_SOURCES =
@@ -12,8 +11,11 @@ libwebpencode_la_SOURCES += filter.c
 libwebpencode_la_SOURCES += frame.c
 libwebpencode_la_SOURCES += histogram.c
 libwebpencode_la_SOURCES += iterator.c
-libwebpencode_la_SOURCES += layer.c
 libwebpencode_la_SOURCES += picture.c
+libwebpencode_la_SOURCES += picture_csp.c
+libwebpencode_la_SOURCES += picture_psnr.c
+libwebpencode_la_SOURCES += picture_rescale.c
+libwebpencode_la_SOURCES += picture_tools.c
 libwebpencode_la_SOURCES += quant.c
 libwebpencode_la_SOURCES += syntax.c
 libwebpencode_la_SOURCES += token.c

src/enc/alpha.c

@@ -17,6 +17,7 @@
 #include "./vp8enci.h"
 #include "../utils/filters.h"
 #include "../utils/quant_levels.h"
+#include "../utils/utils.h"
 #include "../webp/format_constants.h"
 
 // -----------------------------------------------------------------------------
@@ -34,7 +35,7 @@
 //
 // 'output' corresponds to the buffer containing compressed alpha data.
 //          This buffer is allocated by this method and caller should call
-//          free(*output) when done.
+//          WebPSafeFree(*output) when done.
 // 'output_size' corresponds to size of this compressed alpha buffer.
 //
 // Returns 1 on successfully encoding the alpha and
@@ -46,12 +47,11 @@
 
 static int EncodeLossless(const uint8_t* const data, int width, int height,
                           int effort_level,  // in [0..6] range
-                          VP8BitWriter* const bw,
+                          VP8LBitWriter* const bw,
                           WebPAuxStats* const stats) {
   int ok = 0;
   WebPConfig config;
   WebPPicture picture;
-  VP8LBitWriter tmp_bw;
 
   WebPPictureInit(&picture);
   picture.width = width;
@@ -83,16 +83,15 @@ static int EncodeLossless(const uint8_t* const data, int width, int height,
   config.quality = 8.f * effort_level;
   assert(config.quality >= 0 && config.quality <= 100.f);
 
-  ok = VP8LBitWriterInit(&tmp_bw, (width * height) >> 3);
-  ok = ok && (VP8LEncodeStream(&config, &picture, &tmp_bw) == VP8_ENC_OK);
+  ok = (VP8LEncodeStream(&config, &picture, bw) == VP8_ENC_OK);
   WebPPictureFree(&picture);
-  if (ok) {
-    const uint8_t* const buffer = VP8LBitWriterFinish(&tmp_bw);
-    const size_t buffer_size = VP8LBitWriterNumBytes(&tmp_bw);
-    VP8BitWriterAppend(bw, buffer, buffer_size);
+  ok = ok && !bw->error_;
+  if (!ok) {
+    VP8LBitWriterDestroy(bw);
+    return 0;
   }
-  VP8LBitWriterDestroy(&tmp_bw);
-  return ok && !bw->error_;
+  return 1;
+
 }
 
 // -----------------------------------------------------------------------------
@@ -114,8 +113,10 @@ static int EncodeAlphaInternal(const uint8_t* const data, int width, int height,
   const uint8_t* alpha_src;
   WebPFilterFunc filter_func;
   uint8_t header;
-  size_t expected_size;
   const size_t data_size = width * height;
+  const uint8_t* output = NULL;
+  size_t output_size = 0;
+  VP8LBitWriter tmp_bw;
 
   assert((uint64_t)data_size == (uint64_t)width * height);  // as per spec
   assert(filter >= 0 && filter < WEBP_FILTER_LAST);
@@ -124,15 +125,6 @@ static int EncodeAlphaInternal(const uint8_t* const data, int width, int height,
   assert(sizeof(header) == ALPHA_HEADER_LEN);
   // TODO(skal): have a common function and #define's to validate alpha params.
 
-  expected_size =
-      (method == ALPHA_NO_COMPRESSION) ? (ALPHA_HEADER_LEN + data_size)
-                                       : (data_size >> 5);
-  header = method | (filter << 2);
-  if (reduce_levels) header |= ALPHA_PREPROCESSED_LEVELS << 4;
-
-  VP8BitWriterInit(&result->bw, expected_size);
-  VP8BitWriterAppend(&result->bw, &header, ALPHA_HEADER_LEN);
-
   filter_func = WebPFilters[filter];
   if (filter_func != NULL) {
     filter_func(data, width, height, width, tmp_alpha);
@@ -141,14 +133,42 @@ static int EncodeAlphaInternal(const uint8_t* const data, int width, int height,
     alpha_src = data;
   }
 
-  if (method == ALPHA_NO_COMPRESSION) {
-    ok = VP8BitWriterAppend(&result->bw, alpha_src, width * height);
-    ok = ok && !result->bw.error_;
-  } else {
-    ok = EncodeLossless(alpha_src, width, height, effort_level,
-                        &result->bw, &result->stats);
-    VP8BitWriterFinish(&result->bw);
+  if (method != ALPHA_NO_COMPRESSION) {
+    ok = VP8LBitWriterInit(&tmp_bw, data_size >> 3);
+    ok = ok && EncodeLossless(alpha_src, width, height, effort_level,
+                              &tmp_bw, &result->stats);
+    if (ok) {
+      output = VP8LBitWriterFinish(&tmp_bw);
+      output_size = VP8LBitWriterNumBytes(&tmp_bw);
+      if (output_size > data_size) {
+        // compressed size is larger than source! Revert to uncompressed mode.
+        method = ALPHA_NO_COMPRESSION;
+        VP8LBitWriterDestroy(&tmp_bw);
+      }
+    } else {
+      VP8LBitWriterDestroy(&tmp_bw);
+      return 0;
+    }
   }
+
+  if (method == ALPHA_NO_COMPRESSION) {
+    output = alpha_src;
+    output_size = data_size;
+    ok = 1;
+  }
+
+  // Emit final result.
+  header = method | (filter << 2);
+  if (reduce_levels) header |= ALPHA_PREPROCESSED_LEVELS << 4;
+
+  VP8BitWriterInit(&result->bw, ALPHA_HEADER_LEN + output_size);
+  ok = ok && VP8BitWriterAppend(&result->bw, &header, ALPHA_HEADER_LEN);
+  ok = ok && VP8BitWriterAppend(&result->bw, output, output_size);
+
+  if (method != ALPHA_NO_COMPRESSION) {
+    VP8LBitWriterDestroy(&tmp_bw);
+  }
+  ok = ok && !result->bw.error_;
   result->score = VP8BitWriterSize(&result->bw);
   return ok;
 }
@@ -231,7 +251,7 @@ static int ApplyFiltersAndEncode(const uint8_t* alpha, int width, int height,
       GetFilterMap(alpha, width, height, filter, effort_level);
   InitFilterTrial(&best);
   if (try_map != FILTER_TRY_NONE) {
-    uint8_t* filtered_alpha =  (uint8_t*)malloc(data_size);
+    uint8_t* filtered_alpha =  (uint8_t*)WebPSafeMalloc(1ULL, data_size);
     if (filtered_alpha == NULL) return 0;
 
     for (filter = WEBP_FILTER_NONE; ok && try_map; ++filter, try_map >>= 1) {
@@ -248,7 +268,7 @@ static int ApplyFiltersAndEncode(const uint8_t* alpha, int width, int height,
         }
       }
     }
-    free(filtered_alpha);
+    WebPSafeFree(filtered_alpha);
   } else {
     ok = EncodeAlphaInternal(alpha, width, height, method, WEBP_FILTER_NONE,
                              reduce_levels, effort_level, NULL, &best);
@@ -298,7 +318,7 @@ static int EncodeAlpha(VP8Encoder* const enc,
     filter = WEBP_FILTER_NONE;
   }
 
-  quant_alpha = (uint8_t*)malloc(data_size);
+  quant_alpha = (uint8_t*)WebPSafeMalloc(1ULL, data_size);
   if (quant_alpha == NULL) {
     return 0;
   }
@@ -325,7 +345,7 @@ static int EncodeAlpha(VP8Encoder* const enc,
     }
   }
 
-  free(quant_alpha);
+  WebPSafeFree(quant_alpha);
   return ok;
 }
 
@@ -346,7 +366,7 @@ static int CompressAlphaJob(VP8Encoder* const enc, void* dummy) {
     return 0;
   }
   if (alpha_size != (uint32_t)alpha_size) {  // Sanity check.
-    free(alpha_data);
+    WebPSafeFree(alpha_data);
     return 0;
   }
   enc->alpha_data_size_ = (uint32_t)alpha_size;
@@ -361,7 +381,7 @@ void VP8EncInitAlpha(VP8Encoder* const enc) {
   enc->alpha_data_size_ = 0;
   if (enc->thread_level_ > 0) {
     WebPWorker* const worker = &enc->alpha_worker_;
-    WebPWorkerInit(worker);
+    WebPGetWorkerInterface()->Init(worker);
     worker->data1 = enc;
     worker->data2 = NULL;
     worker->hook = (WebPWorkerHook)CompressAlphaJob;
@@ -372,10 +392,11 @@ int VP8EncStartAlpha(VP8Encoder* const enc) {
   if (enc->has_alpha_) {
     if (enc->thread_level_ > 0) {
       WebPWorker* const worker = &enc->alpha_worker_;
-      if (!WebPWorkerReset(worker)) {    // Makes sure worker is good to go.
+      // Makes sure worker is good to go.
+      if (!WebPGetWorkerInterface()->Reset(worker)) {
         return 0;
       }
-      WebPWorkerLaunch(worker);
+      WebPGetWorkerInterface()->Launch(worker);
       return 1;
     } else {
       return CompressAlphaJob(enc, NULL);   // just do the job right away
@@ -388,7 +409,7 @@ int VP8EncFinishAlpha(VP8Encoder* const enc) {
   if (enc->has_alpha_) {
     if (enc->thread_level_ > 0) {
       WebPWorker* const worker = &enc->alpha_worker_;
-      if (!WebPWorkerSync(worker)) return 0;  // error
+      if (!WebPGetWorkerInterface()->Sync(worker)) return 0;  // error
     }
   }
   return WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
@@ -398,10 +419,12 @@ int VP8EncDeleteAlpha(VP8Encoder* const enc) {
   int ok = 1;
   if (enc->thread_level_ > 0) {
     WebPWorker* const worker = &enc->alpha_worker_;
-    ok = WebPWorkerSync(worker);  // finish anything left in flight
-    WebPWorkerEnd(worker);  // still need to end the worker, even if !ok
+    // finish anything left in flight
+    ok = WebPGetWorkerInterface()->Sync(worker);
+    // still need to end the worker, even if !ok
+    WebPGetWorkerInterface()->End(worker);
   }
-  free(enc->alpha_data_);
+  WebPSafeFree(enc->alpha_data_);
   enc->alpha_data_ = NULL;
   enc->alpha_data_size_ = 0;
   enc->has_alpha_ = 0;

src/enc/analysis.c

@@ -30,7 +30,7 @@ static void SmoothSegmentMap(VP8Encoder* const enc) {
   const int w = enc->mb_w_;
   const int h = enc->mb_h_;
   const int majority_cnt_3_x_3_grid = 5;
-  uint8_t* const tmp = (uint8_t*)WebPSafeMalloc((uint64_t)w * h, sizeof(*tmp));
+  uint8_t* const tmp = (uint8_t*)WebPSafeMalloc(w * h, sizeof(*tmp));
   assert((uint64_t)(w * h) == (uint64_t)w * h);   // no overflow, as per spec
 
   if (tmp == NULL) return;
@@ -63,7 +63,7 @@ static void SmoothSegmentMap(VP8Encoder* const enc) {
       mb->segment_ = tmp[x + y * w];
     }
   }
-  free(tmp);
+  WebPSafeFree(tmp);
 }
 
 //------------------------------------------------------------------------------
@@ -141,7 +141,11 @@ static void MergeHistograms(const VP8Histogram* const in,
 
 static void AssignSegments(VP8Encoder* const enc,
                            const int alphas[MAX_ALPHA + 1]) {
-  const int nb = enc->segment_hdr_.num_segments_;
+  // 'num_segments_' is previously validated and <= NUM_MB_SEGMENTS, but an
+  // explicit check is needed to avoid spurious warning about 'n + 1' exceeding
+  // array bounds of 'centers' with some compilers (noticed with gcc-4.9).
+  const int nb = (enc->segment_hdr_.num_segments_ < NUM_MB_SEGMENTS) ?
+                 enc->segment_hdr_.num_segments_ : NUM_MB_SEGMENTS;
   int centers[NUM_MB_SEGMENTS];
   int weighted_average = 0;
   int map[MAX_ALPHA + 1];
@@ -151,6 +155,7 @@ static void AssignSegments(VP8Encoder* const enc,
   int accum[NUM_MB_SEGMENTS], dist_accum[NUM_MB_SEGMENTS];
 
   assert(nb >= 1);
+  assert(nb <= NUM_MB_SEGMENTS);
 
   // bracket the input
   for (n = 0; n <= MAX_ALPHA && alphas[n] == 0; ++n) {}
@@ -225,18 +230,15 @@ static void AssignSegments(VP8Encoder* const enc,
 // susceptibility and set best modes for this macroblock.
 // Segment assignment is done later.
 
-// Number of modes to inspect for alpha_ evaluation. For high-quality settings
-// (method >= FAST_ANALYSIS_METHOD) we don't need to test all the possible modes
-// during the analysis phase.
-#define FAST_ANALYSIS_METHOD 4  // method above which we do partial analysis
+// Number of modes to inspect for alpha_ evaluation. We don't need to test all
+// the possible modes during the analysis phase: we risk falling into a local
+// optimum, or be subject to boundary effect
 #define MAX_INTRA16_MODE 2
 #define MAX_INTRA4_MODE  2
 #define MAX_UV_MODE      2
 
 static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
-  const int max_mode =
-      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA16_MODE
-                                                  : NUM_PRED_MODES;
+  const int max_mode = MAX_INTRA16_MODE;
   int mode;
   int best_alpha = DEFAULT_ALPHA;
   int best_mode = 0;
@@ -262,9 +264,7 @@ static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
 static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
                                    int best_alpha) {
   uint8_t modes[16];
-  const int max_mode =
-      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA4_MODE
-                                                  : NUM_BMODES;
+  const int max_mode = MAX_INTRA4_MODE;
   int i4_alpha;
   VP8Histogram total_histo = { { 0 } };
   int cur_histo = 0;
@@ -306,10 +306,9 @@ static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
 static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
   int best_alpha = DEFAULT_ALPHA;
   int best_mode = 0;
-  const int max_mode =
-      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_UV_MODE
-                                                  : NUM_PRED_MODES;
+  const int max_mode = MAX_UV_MODE;
   int mode;
+
   VP8MakeChroma8Preds(it);
   for (mode = 0; mode < max_mode; ++mode) {
     VP8Histogram histo = { { 0 } };
@@ -425,7 +424,7 @@ static void MergeJobs(const SegmentJob* const src, SegmentJob* const dst) {
 // initialize the job struct with some TODOs
 static void InitSegmentJob(VP8Encoder* const enc, SegmentJob* const job,
                            int start_row, int end_row) {
-  WebPWorkerInit(&job->worker);
+  WebPGetWorkerInterface()->Init(&job->worker);
   job->worker.data1 = job;
   job->worker.data2 = &job->it;
   job->worker.hook = (WebPWorkerHook)DoSegmentsJob;
@@ -458,6 +457,8 @@ int VP8EncAnalyze(VP8Encoder* const enc) {
 #else
     const int do_mt = 0;
 #endif
+    const WebPWorkerInterface* const worker_interface =
+        WebPGetWorkerInterface();
     SegmentJob main_job;
     if (do_mt) {
       SegmentJob side_job;
@@ -467,23 +468,23 @@ int VP8EncAnalyze(VP8Encoder* const enc) {
       InitSegmentJob(enc, &side_job, split_row, last_row);
       // we don't need to call Reset() on main_job.worker, since we're calling
       // WebPWorkerExecute() on it
-      ok &= WebPWorkerReset(&side_job.worker);
+      ok &= worker_interface->Reset(&side_job.worker);
       // launch the two jobs in parallel
       if (ok) {
-        WebPWorkerLaunch(&side_job.worker);
-        WebPWorkerExecute(&main_job.worker);
-        ok &= WebPWorkerSync(&side_job.worker);
-        ok &= WebPWorkerSync(&main_job.worker);
+        worker_interface->Launch(&side_job.worker);
+        worker_interface->Execute(&main_job.worker);
+        ok &= worker_interface->Sync(&side_job.worker);
+        ok &= worker_interface->Sync(&main_job.worker);
       }
-      WebPWorkerEnd(&side_job.worker);
+      worker_interface->End(&side_job.worker);
       if (ok) MergeJobs(&side_job, &main_job);  // merge results together
     } else {
       // Even for single-thread case, we use the generic Worker tools.
       InitSegmentJob(enc, &main_job, 0, last_row);
-      WebPWorkerExecute(&main_job.worker);
-      ok &= WebPWorkerSync(&main_job.worker);
+      worker_interface->Execute(&main_job.worker);
+      ok &= worker_interface->Sync(&main_job.worker);
     }
-    WebPWorkerEnd(&main_job.worker);
+    worker_interface->End(&main_job.worker);
     if (ok) {
       enc->alpha_ = main_job.alpha / total_mb;
       enc->uv_alpha_ = main_job.uv_alpha / total_mb;

src/enc/backward_references.c

@@ -12,7 +12,6 @@
 
 #include <assert.h>
 #include <math.h>
-#include <stdio.h>
 
 #include "./backward_references.h"
 #include "./histogram.h"
@@ -22,10 +21,12 @@
 
 #define VALUES_IN_BYTE 256
 
-#define HASH_BITS 18
-#define HASH_SIZE (1 << HASH_BITS)
 #define HASH_MULTIPLIER (0xc6a4a7935bd1e995ULL)
 
+#define MIN_BLOCK_SIZE 256  // minimum block size for backward references
+
+#define MAX_ENTROPY    (1e30f)
+
 // 1M window (4M bytes) minus 120 special codes for short distances.
 #define WINDOW_SIZE ((1 << 20) - 120)
 
@@ -33,14 +34,6 @@
 #define MIN_LENGTH 2
 #define MAX_LENGTH 4096
 
-typedef struct {
-  // Stores the most recently added position with the given hash value.
-  int32_t hash_to_first_index_[HASH_SIZE];
-  // chain_[pos] stores the previous position with the same hash value
-  // for every pixel in the image.
-  int32_t* chain_;
-} HashChain;
-
 // -----------------------------------------------------------------------------
 
 static const uint8_t plane_to_code_lut[128] = {
@@ -78,65 +71,152 @@ static WEBP_INLINE int FindMatchLength(const uint32_t* const array1,
 // -----------------------------------------------------------------------------
 //  VP8LBackwardRefs
 
-void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs) {
-  if (refs != NULL) {
-    refs->refs = NULL;
-    refs->size = 0;
-    refs->max_size = 0;
-  }
-}
+struct PixOrCopyBlock {
+  PixOrCopyBlock* next_;   // next block (or NULL)
+  PixOrCopy* start_;       // data start
+  int size_;               // currently used size
+};
 
-void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs) {
-  if (refs != NULL) {
-    free(refs->refs);
-    VP8LInitBackwardRefs(refs);
-  }
-}
-
-int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size) {
+static void ClearBackwardRefs(VP8LBackwardRefs* const refs) {
   assert(refs != NULL);
-  refs->size = 0;
-  refs->max_size = 0;
-  refs->refs = (PixOrCopy*)WebPSafeMalloc((uint64_t)max_size,
-                                          sizeof(*refs->refs));
-  if (refs->refs == NULL) return 0;
-  refs->max_size = max_size;
+  if (refs->tail_ != NULL) {
+    *refs->tail_ = refs->free_blocks_;  // recycle all blocks at once
+  }
+  refs->free_blocks_ = refs->refs_;
+  refs->tail_ = &refs->refs_;
+  refs->last_block_ = NULL;
+  refs->refs_ = NULL;
+}
+
+void VP8LBackwardRefsClear(VP8LBackwardRefs* const refs) {
+  assert(refs != NULL);
+  ClearBackwardRefs(refs);
+  while (refs->free_blocks_ != NULL) {
+    PixOrCopyBlock* const next = refs->free_blocks_->next_;
+    WebPSafeFree(refs->free_blocks_);
+    refs->free_blocks_ = next;
+  }
+}
+
+void VP8LBackwardRefsInit(VP8LBackwardRefs* const refs, int block_size) {
+  assert(refs != NULL);
+  memset(refs, 0, sizeof(*refs));
+  refs->tail_ = &refs->refs_;
+  refs->block_size_ =
+      (block_size < MIN_BLOCK_SIZE) ? MIN_BLOCK_SIZE : block_size;
+}
+
+VP8LRefsCursor VP8LRefsCursorInit(const VP8LBackwardRefs* const refs) {
+  VP8LRefsCursor c;
+  c.cur_block_ = refs->refs_;
+  if (refs->refs_ != NULL) {
+    c.cur_pos = c.cur_block_->start_;
+    c.last_pos_ = c.cur_pos + c.cur_block_->size_;
+  } else {
+    c.cur_pos = NULL;
+    c.last_pos_ = NULL;
+  }
+  return c;
+}
+
+void VP8LRefsCursorNextBlock(VP8LRefsCursor* const c) {
+  PixOrCopyBlock* const b = c->cur_block_->next_;
+  c->cur_pos = (b == NULL) ? NULL : b->start_;
+  c->last_pos_ = (b == NULL) ? NULL : b->start_ + b->size_;
+  c->cur_block_ = b;
+}
+
+// Create a new block, either from the free list or allocated
+static PixOrCopyBlock* BackwardRefsNewBlock(VP8LBackwardRefs* const refs) {
+  PixOrCopyBlock* b = refs->free_blocks_;
+  if (b == NULL) {   // allocate new memory chunk
+    const size_t total_size =
+        sizeof(*b) + refs->block_size_ * sizeof(*b->start_);
+    b = (PixOrCopyBlock*)WebPSafeMalloc(1ULL, total_size);
+    if (b == NULL) {
+      refs->error_ |= 1;
+      return NULL;
+    }
+    b->start_ = (PixOrCopy*)((uint8_t*)b + sizeof(*b));  // not always aligned
+  } else {  // recycle from free-list
+    refs->free_blocks_ = b->next_;
+  }
+  *refs->tail_ = b;
+  refs->tail_ = &b->next_;
+  refs->last_block_ = b;
+  b->next_ = NULL;
+  b->size_ = 0;
+  return b;
+}
+
+static WEBP_INLINE void BackwardRefsCursorAdd(VP8LBackwardRefs* const refs,
+                                              const PixOrCopy v) {
+  PixOrCopyBlock* b = refs->last_block_;
+  if (b == NULL || b->size_ == refs->block_size_) {
+    b = BackwardRefsNewBlock(refs);
+    if (b == NULL) return;   // refs->error_ is set
+  }
+  b->start_[b->size_++] = v;
+}
+
+int VP8LBackwardRefsCopy(const VP8LBackwardRefs* const src,
+                         VP8LBackwardRefs* const dst) {
+  const PixOrCopyBlock* b = src->refs_;
+  ClearBackwardRefs(dst);
+  assert(src->block_size_ == dst->block_size_);
+  while (b != NULL) {
+    PixOrCopyBlock* const new_b = BackwardRefsNewBlock(dst);
+    if (new_b == NULL) return 0;   // dst->error_ is set
+    memcpy(new_b->start_, b->start_, b->size_ * sizeof(*b->start_));
+    new_b->size_ = b->size_;
+    b = b->next_;
+  }
   return 1;
 }
 
 // -----------------------------------------------------------------------------
 // Hash chains
 
-static WEBP_INLINE uint64_t GetPixPairHash64(const uint32_t* const argb) {
-  uint64_t key = ((uint64_t)(argb[1]) << 32) | argb[0];
-  key = (key * HASH_MULTIPLIER) >> (64 - HASH_BITS);
-  return key;
-}
-
-static int HashChainInit(HashChain* const p, int size) {
+// initialize as empty
+static void HashChainInit(VP8LHashChain* const p) {
   int i;
-  p->chain_ = (int*)WebPSafeMalloc((uint64_t)size, sizeof(*p->chain_));
-  if (p->chain_ == NULL) {
-    return 0;
-  }
-  for (i = 0; i < size; ++i) {
+  assert(p != NULL);
+  for (i = 0; i < p->size_; ++i) {
     p->chain_[i] = -1;
   }
   for (i = 0; i < HASH_SIZE; ++i) {
     p->hash_to_first_index_[i] = -1;
   }
+}
+
+int VP8LHashChainInit(VP8LHashChain* const p, int size) {
+  assert(p->size_ == 0);
+  assert(p->chain_ == NULL);
+  assert(size > 0);
+  p->chain_ = (int*)WebPSafeMalloc(size, sizeof(*p->chain_));
+  if (p->chain_ == NULL) return 0;
+  p->size_ = size;
+  HashChainInit(p);
   return 1;
 }
 
-static void HashChainDelete(HashChain* const p) {
-  if (p != NULL) {
-    free(p->chain_);
-    free(p);
-  }
+void VP8LHashChainClear(VP8LHashChain* const p) {
+  assert(p != NULL);
+  WebPSafeFree(p->chain_);
+  p->size_ = 0;
+  p->chain_ = NULL;
+}
+
+// -----------------------------------------------------------------------------
+
+static WEBP_INLINE uint64_t GetPixPairHash64(const uint32_t* const argb) {
+  uint64_t key = ((uint64_t)argb[1] << 32) | argb[0];
+  key = (key * HASH_MULTIPLIER) >> (64 - HASH_BITS);
+  return key;
 }
 
 // Insertion of two pixels at a time.
-static void HashChainInsert(HashChain* const p,
+static void HashChainInsert(VP8LHashChain* const p,
                             const uint32_t* const argb, int pos) {
   const uint64_t hash_code = GetPixPairHash64(argb);
   p->chain_[pos] = p->hash_to_first_index_[hash_code];
@@ -161,7 +241,7 @@ static void GetParamsForHashChainFindCopy(int quality, int xsize,
   *iter_limit = (cache_bits > 0) ? iter_neg : iter_neg / 2;
 }
 
-static int HashChainFindCopy(const HashChain* const p,
+static int HashChainFindCopy(const VP8LHashChain* const p,
                              int base_position, int xsize_signed,
                              const uint32_t* const argb, int max_len,
                              int window_size, int iter_pos, int iter_limit,
@@ -185,10 +265,8 @@ static int HashChainFindCopy(const HashChain* const p,
     uint64_t val;
     uint32_t curr_length;
     uint32_t distance;
-    const uint64_t* const ptr1 =
-        (const uint64_t*)(argb + pos + best_length - 1);
-    const uint64_t* const ptr2 =
-        (const uint64_t*)(argb_start + best_length - 1);
+    const uint32_t* const ptr1 = (argb + pos + best_length - 1);
+    const uint32_t* const ptr2 = (argb_start + best_length - 1);
 
     if (iter_pos < 0) {
       if (iter_pos < iter_limit || best_val >= 0xff0000) {
@@ -199,7 +277,7 @@ static int HashChainFindCopy(const HashChain* const p,
 
     // Before 'expensive' linear match, check if the two arrays match at the
     // current best length index and also for the succeeding elements.
-    if (*ptr1 != *ptr2) continue;
+    if (ptr1[0] != ptr2[0] || ptr1[1] != ptr2[1]) continue;
 
     curr_length = FindMatchLength(argb + pos, argb_start, max_len);
     if (curr_length < best_length) continue;
@@ -237,64 +315,61 @@ static int HashChainFindCopy(const HashChain* const p,
 }
 
 static WEBP_INLINE void PushBackCopy(VP8LBackwardRefs* const refs, int length) {
-  int size = refs->size;
   while (length >= MAX_LENGTH) {
-    refs->refs[size++] = PixOrCopyCreateCopy(1, MAX_LENGTH);
+    BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(1, MAX_LENGTH));
     length -= MAX_LENGTH;
   }
   if (length > 0) {
-    refs->refs[size++] = PixOrCopyCreateCopy(1, length);
+    BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(1, length));
   }
-  refs->size = size;
 }
 
-static void BackwardReferencesRle(int xsize, int ysize,
-                                  const uint32_t* const argb,
-                                  VP8LBackwardRefs* const refs) {
+static int BackwardReferencesRle(int xsize, int ysize,
+                                 const uint32_t* const argb,
+                                 VP8LBackwardRefs* const refs) {
   const int pix_count = xsize * ysize;
   int match_len = 0;
   int i;
-  refs->size = 0;
+  ClearBackwardRefs(refs);
   PushBackCopy(refs, match_len);    // i=0 case
-  refs->refs[refs->size++] = PixOrCopyCreateLiteral(argb[0]);
+  BackwardRefsCursorAdd(refs, PixOrCopyCreateLiteral(argb[0]));
   for (i = 1; i < pix_count; ++i) {
     if (argb[i] == argb[i - 1]) {
       ++match_len;
     } else {
       PushBackCopy(refs, match_len);
       match_len = 0;
-      refs->refs[refs->size++] = PixOrCopyCreateLiteral(argb[i]);
+      BackwardRefsCursorAdd(refs, PixOrCopyCreateLiteral(argb[i]));
     }
   }
   PushBackCopy(refs, match_len);
+  return !refs->error_;
 }
 
 static int BackwardReferencesHashChain(int xsize, int ysize,
                                        const uint32_t* const argb,
                                        int cache_bits, int quality,
+                                       VP8LHashChain* const hash_chain,
                                        VP8LBackwardRefs* const refs) {
   int i;
   int ok = 0;
   int cc_init = 0;
   const int use_color_cache = (cache_bits > 0);
   const int pix_count = xsize * ysize;
-  HashChain* const hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
   VP8LColorCache hashers;
   int window_size = WINDOW_SIZE;
   int iter_pos = 1;
   int iter_limit = -1;
 
-  if (hash_chain == NULL) return 0;
   if (use_color_cache) {
     cc_init = VP8LColorCacheInit(&hashers, cache_bits);
     if (!cc_init) goto Error;
   }
 
-  if (!HashChainInit(hash_chain, pix_count)) goto Error;
-
-  refs->size = 0;
+  ClearBackwardRefs(refs);
   GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                 &window_size, &iter_pos, &iter_limit);
+  HashChainInit(hash_chain);
   for (i = 0; i < pix_count; ) {
     // Alternative#1: Code the pixels starting at 'i' using backward reference.
     int offset = 0;
@@ -320,14 +395,15 @@ static int BackwardReferencesHashChain(int xsize, int ysize,
         if (len2 > len + 1) {
           const uint32_t pixel = argb[i];
           // Alternative#2 is a better match. So push pixel at 'i' as literal.
+          PixOrCopy v;
           if (use_color_cache && VP8LColorCacheContains(&hashers, pixel)) {
             const int ix = VP8LColorCacheGetIndex(&hashers, pixel);
-            refs->refs[refs->size] = PixOrCopyCreateCacheIdx(ix);
+            v = PixOrCopyCreateCacheIdx(ix);
           } else {
             if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
-            refs->refs[refs->size] = PixOrCopyCreateLiteral(pixel);
+            v = PixOrCopyCreateLiteral(pixel);
           }
-          ++refs->size;
+          BackwardRefsCursorAdd(refs, v);
           i++;  // Backward reference to be done for next pixel.
           len = len2;
           offset = offset2;
@@ -336,7 +412,7 @@ static int BackwardReferencesHashChain(int xsize, int ysize,
       if (len >= MAX_LENGTH) {
         len = MAX_LENGTH - 1;
       }
-      refs->refs[refs->size++] = PixOrCopyCreateCopy(offset, len);
+      BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(offset, len));
       if (use_color_cache) {
         for (k = 0; k < len; ++k) {
           VP8LColorCacheInsert(&hashers, argb[i + k]);
@@ -352,25 +428,25 @@ static int BackwardReferencesHashChain(int xsize, int ysize,
       i += len;
     } else {
       const uint32_t pixel = argb[i];
+      PixOrCopy v;
       if (use_color_cache && VP8LColorCacheContains(&hashers, pixel)) {
         // push pixel as a PixOrCopyCreateCacheIdx pixel
         const int ix = VP8LColorCacheGetIndex(&hashers, pixel);
-        refs->refs[refs->size] = PixOrCopyCreateCacheIdx(ix);
+        v = PixOrCopyCreateCacheIdx(ix);
       } else {
         if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
-        refs->refs[refs->size] = PixOrCopyCreateLiteral(pixel);
+        v = PixOrCopyCreateLiteral(pixel);
       }
-      ++refs->size;
+      BackwardRefsCursorAdd(refs, v);
       if (i + 1 < pix_count) {
         HashChainInsert(hash_chain, &argb[i], i);
       }
       ++i;
     }
   }
-  ok = 1;
+  ok = !refs->error_;
 Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  HashChainDelete(hash_chain);
   return ok;
 }
 
@@ -387,11 +463,12 @@ typedef struct {
 static int BackwardReferencesTraceBackwards(
     int xsize, int ysize, int recursive_cost_model,
     const uint32_t* const argb, int quality, int cache_bits,
+    VP8LHashChain* const hash_chain,
     VP8LBackwardRefs* const refs);
 
 static void ConvertPopulationCountTableToBitEstimates(
-    int num_symbols, const int population_counts[], double output[]) {
-  int sum = 0;
+    int num_symbols, const uint32_t population_counts[], double output[]) {
+  uint32_t sum = 0;
   int nonzeros = 0;
   int i;
   for (i = 0; i < num_symbols; ++i) {
@@ -412,39 +489,45 @@ static void ConvertPopulationCountTableToBitEstimates(
 
 static int CostModelBuild(CostModel* const m, int xsize, int ysize,
                           int recursion_level, const uint32_t* const argb,
-                          int quality, int cache_bits) {
+                          int quality, int cache_bits,
+                          VP8LHashChain* const hash_chain,
+                          VP8LBackwardRefs* const refs) {
   int ok = 0;
-  VP8LHistogram histo;
-  VP8LBackwardRefs refs;
-
-  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize)) goto Error;
+  VP8LHistogram* histo = NULL;
 
+  ClearBackwardRefs(refs);
   if (recursion_level > 0) {
     if (!BackwardReferencesTraceBackwards(xsize, ysize, recursion_level - 1,
-                                          argb, quality, cache_bits, &refs)) {
+                                          argb, quality, cache_bits, hash_chain,
+                                          refs)) {
       goto Error;
     }
   } else {
     if (!BackwardReferencesHashChain(xsize, ysize, argb, cache_bits, quality,
-                                     &refs)) {
+                                     hash_chain, refs)) {
       goto Error;
     }
   }
-  VP8LHistogramCreate(&histo, &refs, cache_bits);
+  histo = VP8LAllocateHistogram(cache_bits);
+  if (histo == NULL) goto Error;
+
+  VP8LHistogramCreate(histo, refs, cache_bits);
+
   ConvertPopulationCountTableToBitEstimates(
-      VP8LHistogramNumCodes(&histo), histo.literal_, m->literal_);
+      VP8LHistogramNumCodes(histo->palette_code_bits_),
+      histo->literal_, m->literal_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.red_, m->red_);
+      VALUES_IN_BYTE, histo->red_, m->red_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.blue_, m->blue_);
+      VALUES_IN_BYTE, histo->blue_, m->blue_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.alpha_, m->alpha_);
+      VALUES_IN_BYTE, histo->alpha_, m->alpha_);
   ConvertPopulationCountTableToBitEstimates(
-      NUM_DISTANCE_CODES, histo.distance_, m->distance_);
+      NUM_DISTANCE_CODES, histo->distance_, m->distance_);
   ok = 1;
 
  Error:
-  VP8LClearBackwardRefs(&refs);
+  VP8LFreeHistogram(histo);
   return ok;
 }
 
@@ -476,16 +559,16 @@ static WEBP_INLINE double GetDistanceCost(const CostModel* const m,
 
 static int BackwardReferencesHashChainDistanceOnly(
     int xsize, int ysize, int recursive_cost_model, const uint32_t* const argb,
-    int quality, int cache_bits, uint32_t* const dist_array) {
+    int quality, int cache_bits, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs* const refs, uint32_t* const dist_array) {
   int i;
   int ok = 0;
   int cc_init = 0;
   const int pix_count = xsize * ysize;
   const int use_color_cache = (cache_bits > 0);
   float* const cost =
-      (float*)WebPSafeMalloc((uint64_t)pix_count, sizeof(*cost));
-  CostModel* cost_model = (CostModel*)malloc(sizeof(*cost_model));
-  HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+      (float*)WebPSafeMalloc(pix_count, sizeof(*cost));
+  CostModel* cost_model = (CostModel*)WebPSafeMalloc(1ULL, sizeof(*cost_model));
   VP8LColorCache hashers;
   const double mul0 = (recursive_cost_model != 0) ? 1.0 : 0.68;
   const double mul1 = (recursive_cost_model != 0) ? 1.0 : 0.82;
@@ -494,9 +577,7 @@ static int BackwardReferencesHashChainDistanceOnly(
   int iter_pos = 1;
   int iter_limit = -1;
 
-  if (cost == NULL || cost_model == NULL || hash_chain == NULL) goto Error;
-
-  if (!HashChainInit(hash_chain, pix_count)) goto Error;
+  if (cost == NULL || cost_model == NULL) goto Error;
 
   if (use_color_cache) {
     cc_init = VP8LColorCacheInit(&hashers, cache_bits);
@@ -504,7 +585,7 @@ static int BackwardReferencesHashChainDistanceOnly(
   }
 
   if (!CostModelBuild(cost_model, xsize, ysize, recursive_cost_model, argb,
-                      quality, cache_bits)) {
+                      quality, cache_bits, hash_chain, refs)) {
     goto Error;
   }
 
@@ -515,6 +596,7 @@ static int BackwardReferencesHashChainDistanceOnly(
   dist_array[0] = 0;
   GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                 &window_size, &iter_pos, &iter_limit);
+  HashChainInit(hash_chain);
   for (i = 0; i < pix_count; ++i) {
     double prev_cost = 0.0;
     int shortmax;
@@ -589,12 +671,11 @@ static int BackwardReferencesHashChainDistanceOnly(
   }
   // Last pixel still to do, it can only be a single step if not reached
   // through cheaper means already.
-  ok = 1;
+  ok = !refs->error_;
 Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  HashChainDelete(hash_chain);
-  free(cost_model);
-  free(cost);
+  WebPSafeFree(cost_model);
+  WebPSafeFree(cost);
   return ok;
 }
 
@@ -621,6 +702,7 @@ static int BackwardReferencesHashChainFollowChosenPath(
     int xsize, int ysize, const uint32_t* const argb,
     int quality, int cache_bits,
     const uint32_t* const chosen_path, int chosen_path_size,
+    VP8LHashChain* const hash_chain,
     VP8LBackwardRefs* const refs) {
   const int pix_count = xsize * ysize;
   const int use_color_cache = (cache_bits > 0);
@@ -633,20 +715,17 @@ static int BackwardReferencesHashChainFollowChosenPath(
   int window_size = WINDOW_SIZE;
   int iter_pos = 1;
   int iter_limit = -1;
-  HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
   VP8LColorCache hashers;
 
-  if (hash_chain == NULL || !HashChainInit(hash_chain, pix_count)) {
-    goto Error;
-  }
   if (use_color_cache) {
     cc_init = VP8LColorCacheInit(&hashers, cache_bits);
     if (!cc_init) goto Error;
   }
 
-  refs->size = 0;
+  ClearBackwardRefs(refs);
   GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                 &window_size, &iter_pos, &iter_limit);
+  HashChainInit(hash_chain);
   for (ix = 0; ix < chosen_path_size; ++ix, ++size) {
     int offset = 0;
     int len = 0;
@@ -656,7 +735,7 @@ static int BackwardReferencesHashChainFollowChosenPath(
                         window_size, iter_pos, iter_limit,
                         &offset, &len);
       assert(len == max_len);
-      refs->refs[size] = PixOrCopyCreateCopy(offset, len);
+      BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(offset, len));
       if (use_color_cache) {
         for (k = 0; k < len; ++k) {
           VP8LColorCacheInsert(&hashers, argb[i + k]);
@@ -670,26 +749,25 @@ static int BackwardReferencesHashChainFollowChosenPath(
       }
       i += len;
     } else {
+      PixOrCopy v;
       if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
         // push pixel as a color cache index
         const int idx = VP8LColorCacheGetIndex(&hashers, argb[i]);
-        refs->refs[size] = PixOrCopyCreateCacheIdx(idx);
+        v = PixOrCopyCreateCacheIdx(idx);
       } else {
         if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
-        refs->refs[size] = PixOrCopyCreateLiteral(argb[i]);
+        v = PixOrCopyCreateLiteral(argb[i]);
       }
+      BackwardRefsCursorAdd(refs, v);
       if (i + 1 < pix_count) {
         HashChainInsert(hash_chain, &argb[i], i);
       }
       ++i;
     }
   }
-  assert(size <= refs->max_size);
-  refs->size = size;
-  ok = 1;
+  ok = !refs->error_;
 Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  HashChainDelete(hash_chain);
   return ok;
 }
 
@@ -698,142 +776,129 @@ static int BackwardReferencesTraceBackwards(int xsize, int ysize,
                                             int recursive_cost_model,
                                             const uint32_t* const argb,
                                             int quality, int cache_bits,
+                                            VP8LHashChain* const hash_chain,
                                             VP8LBackwardRefs* const refs) {
   int ok = 0;
   const int dist_array_size = xsize * ysize;
   uint32_t* chosen_path = NULL;
   int chosen_path_size = 0;
   uint32_t* dist_array =
-      (uint32_t*)WebPSafeMalloc((uint64_t)dist_array_size, sizeof(*dist_array));
+      (uint32_t*)WebPSafeMalloc(dist_array_size, sizeof(*dist_array));
 
   if (dist_array == NULL) goto Error;
 
   if (!BackwardReferencesHashChainDistanceOnly(
-      xsize, ysize, recursive_cost_model, argb, quality, cache_bits,
-      dist_array)) {
+      xsize, ysize, recursive_cost_model, argb, quality, cache_bits, hash_chain,
+      refs, dist_array)) {
     goto Error;
   }
   TraceBackwards(dist_array, dist_array_size, &chosen_path, &chosen_path_size);
   if (!BackwardReferencesHashChainFollowChosenPath(
       xsize, ysize, argb, quality, cache_bits, chosen_path, chosen_path_size,
-      refs)) {
+      hash_chain, refs)) {
     goto Error;
   }
   ok = 1;
  Error:
-  free(dist_array);
+  WebPSafeFree(dist_array);
   return ok;
 }
 
 static void BackwardReferences2DLocality(int xsize,
-                                         VP8LBackwardRefs* const refs) {
-  int i;
-  for (i = 0; i < refs->size; ++i) {
-    if (PixOrCopyIsCopy(&refs->refs[i])) {
-      const int dist = refs->refs[i].argb_or_distance;
+                                         const VP8LBackwardRefs* const refs) {
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  while (VP8LRefsCursorOk(&c)) {
+    if (PixOrCopyIsCopy(c.cur_pos)) {
+      const int dist = c.cur_pos->argb_or_distance;
       const int transformed_dist = DistanceToPlaneCode(xsize, dist);
-      refs->refs[i].argb_or_distance = transformed_dist;
+      c.cur_pos->argb_or_distance = transformed_dist;
     }
+    VP8LRefsCursorNext(&c);
   }
 }
 
-int VP8LGetBackwardReferences(int width, int height,
-                              const uint32_t* const argb,
-                              int quality, int cache_bits, int use_2d_locality,
-                              VP8LBackwardRefs* const best) {
-  int ok = 0;
+VP8LBackwardRefs* VP8LGetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int cache_bits, int use_2d_locality, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs refs_array[2]) {
   int lz77_is_useful;
-  VP8LBackwardRefs refs_rle, refs_lz77;
   const int num_pix = width * height;
-
-  VP8LBackwardRefsAlloc(&refs_rle, num_pix);
-  VP8LBackwardRefsAlloc(&refs_lz77, num_pix);
-  VP8LInitBackwardRefs(best);
-  if (refs_rle.refs == NULL || refs_lz77.refs == NULL) {
- Error1:
-    VP8LClearBackwardRefs(&refs_rle);
-    VP8LClearBackwardRefs(&refs_lz77);
-    goto End;
-  }
+  VP8LBackwardRefs* best = NULL;
+  VP8LBackwardRefs* const refs_lz77 = &refs_array[0];
+  VP8LBackwardRefs* const refs_rle = &refs_array[1];
 
   if (!BackwardReferencesHashChain(width, height, argb, cache_bits, quality,
-                                   &refs_lz77)) {
-    goto End;
+                                   hash_chain, refs_lz77)) {
+    return NULL;
+  }
+  if (!BackwardReferencesRle(width, height, argb, refs_rle)) {
+    return NULL;
   }
-  // Backward Reference using RLE only.
-  BackwardReferencesRle(width, height, argb, &refs_rle);
 
   {
     double bit_cost_lz77, bit_cost_rle;
-    VP8LHistogram* const histo = (VP8LHistogram*)malloc(sizeof(*histo));
-    if (histo == NULL) goto Error1;
-    // Evaluate lz77 coding
-    VP8LHistogramCreate(histo, &refs_lz77, cache_bits);
+    VP8LHistogram* const histo = VP8LAllocateHistogram(cache_bits);
+    if (histo == NULL) return NULL;
+    // Evaluate LZ77 coding.
+    VP8LHistogramCreate(histo, refs_lz77, cache_bits);
     bit_cost_lz77 = VP8LHistogramEstimateBits(histo);
-    // Evaluate RLE coding
-    VP8LHistogramCreate(histo, &refs_rle, cache_bits);
+    // Evaluate RLE coding.
+    VP8LHistogramCreate(histo, refs_rle, cache_bits);
     bit_cost_rle = VP8LHistogramEstimateBits(histo);
     // Decide if LZ77 is useful.
     lz77_is_useful = (bit_cost_lz77 < bit_cost_rle);
-    free(histo);
+    VP8LFreeHistogram(histo);
   }
 
   // Choose appropriate backward reference.
   if (lz77_is_useful) {
     // TraceBackwards is costly. Don't execute it at lower quality.
     const int try_lz77_trace_backwards = (quality >= 25);
-    *best = refs_lz77;   // default guess: lz77 is better
-    VP8LClearBackwardRefs(&refs_rle);
+    best = refs_lz77;   // default guess: lz77 is better
     if (try_lz77_trace_backwards) {
       // Set recursion level for large images using a color cache.
       const int recursion_level =
           (num_pix < 320 * 200) && (cache_bits > 0) ? 1 : 0;
-      VP8LBackwardRefs refs_trace;
-      if (!VP8LBackwardRefsAlloc(&refs_trace, num_pix)) {
-        goto End;
-      }
+      VP8LBackwardRefs* const refs_trace = &refs_array[1];
+      ClearBackwardRefs(refs_trace);
       if (BackwardReferencesTraceBackwards(width, height, recursion_level, argb,
-                                           quality, cache_bits, &refs_trace)) {
-        VP8LClearBackwardRefs(&refs_lz77);
-        *best = refs_trace;
+                                           quality, cache_bits, hash_chain,
+                                           refs_trace)) {
+        best = refs_trace;
       }
     }
   } else {
-    VP8LClearBackwardRefs(&refs_lz77);
-    *best = refs_rle;
+    best = refs_rle;
   }
 
   if (use_2d_locality) BackwardReferences2DLocality(width, best);
 
-  ok = 1;
-
- End:
-  if (!ok) {
-    VP8LClearBackwardRefs(best);
-  }
-  return ok;
+  return best;
 }
 
-// Returns 1 on success.
-static int ComputeCacheHistogram(const uint32_t* const argb,
-                                 int xsize, int ysize,
-                                 const VP8LBackwardRefs* const refs,
-                                 int cache_bits,
-                                 VP8LHistogram* const histo) {
+// Returns entropy for the given cache bits.
+static double ComputeCacheEntropy(const uint32_t* const argb,
+                                  int xsize, int ysize,
+                                  const VP8LBackwardRefs* const refs,
+                                  int cache_bits) {
   int pixel_index = 0;
-  int i;
   uint32_t k;
-  VP8LColorCache hashers;
   const int use_color_cache = (cache_bits > 0);
   int cc_init = 0;
+  double entropy = MAX_ENTROPY;
+  const double kSmallPenaltyForLargeCache = 4.0;
+  VP8LColorCache hashers;
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  VP8LHistogram* histo = VP8LAllocateHistogram(cache_bits);
+  if (histo == NULL) goto Error;
 
   if (use_color_cache) {
     cc_init = VP8LColorCacheInit(&hashers, cache_bits);
-    if (!cc_init) return 0;
+    if (!cc_init) goto Error;
   }
 
-  for (i = 0; i < refs->size; ++i) {
-    const PixOrCopy* const v = &refs->refs[i];
+  while (VP8LRefsCursorOk(&c)) {
+    const PixOrCopy* const v = c.cur_pos;
     if (PixOrCopyIsLiteral(v)) {
       if (use_color_cache &&
           VP8LColorCacheContains(&hashers, argb[pixel_index])) {
@@ -853,42 +918,58 @@ static int ComputeCacheHistogram(const uint32_t* const argb,
       }
     }
     pixel_index += PixOrCopyLength(v);
+    VP8LRefsCursorNext(&c);
   }
   assert(pixel_index == xsize * ysize);
   (void)xsize;  // xsize is not used in non-debug compilations otherwise.
   (void)ysize;  // ysize is not used in non-debug compilations otherwise.
+  entropy = VP8LHistogramEstimateBits(histo) +
+      kSmallPenaltyForLargeCache * cache_bits;
+ Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  return 1;
+  VP8LFreeHistogram(histo);
+  return entropy;
 }
 
-// Returns how many bits are to be used for a color cache.
+// *best_cache_bits will contain how many bits are to be used for a color cache.
+// Returns 0 in case of memory error.
 int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
-                                      int xsize, int ysize,
+                                      int xsize, int ysize, int quality,
+                                      VP8LHashChain* const hash_chain,
+                                      VP8LBackwardRefs* const refs,
                                       int* const best_cache_bits) {
-  int ok = 0;
-  int cache_bits;
-  double lowest_entropy = 1e99;
-  VP8LBackwardRefs refs;
-  static const double kSmallPenaltyForLargeCache = 4.0;
-  static const int quality = 30;
-  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize) ||
-      !BackwardReferencesHashChain(xsize, ysize, argb, 0, quality, &refs)) {
-    goto Error;
+  int eval_low = 1;
+  int eval_high = 1;
+  double entropy_low = MAX_ENTROPY;
+  double entropy_high = MAX_ENTROPY;
+  int cache_bits_low = 0;
+  int cache_bits_high = MAX_COLOR_CACHE_BITS;
+
+  if (!BackwardReferencesHashChain(xsize, ysize, argb, 0, quality, hash_chain,
+                                   refs)) {
+    return 0;
   }
-  for (cache_bits = 0; cache_bits <= MAX_COLOR_CACHE_BITS; ++cache_bits) {
-    double cur_entropy;
-    VP8LHistogram histo;
-    VP8LHistogramInit(&histo, cache_bits);
-    ComputeCacheHistogram(argb, xsize, ysize, &refs, cache_bits, &histo);
-    cur_entropy = VP8LHistogramEstimateBits(&histo) +
-        kSmallPenaltyForLargeCache * cache_bits;
-    if (cache_bits == 0 || cur_entropy < lowest_entropy) {
-      *best_cache_bits = cache_bits;
-      lowest_entropy = cur_entropy;
+  // Do a binary search to find the optimal entropy for cache_bits.
+  while (cache_bits_high - cache_bits_low > 1) {
+    if (eval_low) {
+      entropy_low =
+          ComputeCacheEntropy(argb, xsize, ysize, refs, cache_bits_low);
+      eval_low = 0;
+    }
+    if (eval_high) {
+      entropy_high =
+          ComputeCacheEntropy(argb, xsize, ysize, refs, cache_bits_high);
+      eval_high = 0;
+    }
+    if (entropy_high < entropy_low) {
+      *best_cache_bits = cache_bits_high;
+      cache_bits_low = (cache_bits_low + cache_bits_high) / 2;
+      eval_low = 1;
+    } else {
+      *best_cache_bits = cache_bits_low;
+      cache_bits_high = (cache_bits_low + cache_bits_high) / 2;
+      eval_high = 1;
     }
   }
-  ok = 1;
- Error:
-  VP8LClearBackwardRefs(&refs);
-  return ok;
+  return 1;
 }

src/enc/backward_references.h

@@ -113,36 +113,96 @@ static WEBP_INLINE uint32_t PixOrCopyDistance(const PixOrCopy* const p) {
 }
 
 // -----------------------------------------------------------------------------
-// VP8LBackwardRefs
+// VP8LHashChain
 
+#define HASH_BITS 18
+#define HASH_SIZE (1 << HASH_BITS)
+
+typedef struct VP8LHashChain VP8LHashChain;
+struct VP8LHashChain {
+  // Stores the most recently added position with the given hash value.
+  int32_t hash_to_first_index_[HASH_SIZE];
+  // chain_[pos] stores the previous position with the same hash value
+  // for every pixel in the image.
+  int32_t* chain_;
+  // This is the maximum size of the hash_chain that can be constructed.
+  // Typically this is the pixel count (width x height) for a given image.
+  int size_;
+};
+
+// Must be called first, to set size.
+int VP8LHashChainInit(VP8LHashChain* const p, int size);
+void VP8LHashChainClear(VP8LHashChain* const p);  // release memory
+
+// -----------------------------------------------------------------------------
+// VP8LBackwardRefs (block-based backward-references storage)
+
+// maximum number of reference blocks the image will be segmented into
+#define MAX_REFS_BLOCK_PER_IMAGE 16
+
+typedef struct PixOrCopyBlock PixOrCopyBlock;   // forward declaration
+typedef struct VP8LBackwardRefs VP8LBackwardRefs;
+
+// Container for blocks chain
+struct VP8LBackwardRefs {
+  int block_size_;               // common block-size
+  int error_;                    // set to true if some memory error occurred
+  PixOrCopyBlock* refs_;         // list of currently used blocks
+  PixOrCopyBlock** tail_;        // for list recycling
+  PixOrCopyBlock* free_blocks_;  // free-list
+  PixOrCopyBlock* last_block_;   // used for adding new refs (internal)
+};
+
+// Initialize the object. 'block_size' is the common block size to store
+// references (typically, width * height / MAX_REFS_BLOCK_PER_IMAGE).
+void VP8LBackwardRefsInit(VP8LBackwardRefs* const refs, int block_size);
+// Release memory for backward references.
+void VP8LBackwardRefsClear(VP8LBackwardRefs* const refs);
+// Copies the 'src' backward refs to the 'dst'. Returns 0 in case of error.
+int VP8LBackwardRefsCopy(const VP8LBackwardRefs* const src,
+                         VP8LBackwardRefs* const dst);
+
+// Cursor for iterating on references content
 typedef struct {
-  PixOrCopy* refs;
-  int size;      // currently used
-  int max_size;  // maximum capacity
-} VP8LBackwardRefs;
+  // public:
+  PixOrCopy* cur_pos;           // current position
+  // private:
+  PixOrCopyBlock* cur_block_;   // current block in the refs list
+  const PixOrCopy* last_pos_;   // sentinel for switching to next block
+} VP8LRefsCursor;
 
-// Initialize the object. Must be called first. 'refs' can be NULL.
-void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs);
-
-// Release memory and re-initialize the object. 'refs' can be NULL.
-void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs);
-
-// Allocate 'max_size' references. Returns false in case of memory error.
-int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size);
+// Returns a cursor positioned at the beginning of the references list.
+VP8LRefsCursor VP8LRefsCursorInit(const VP8LBackwardRefs* const refs);
+// Returns true if cursor is pointing at a valid position.
+static WEBP_INLINE int VP8LRefsCursorOk(const VP8LRefsCursor* const c) {
+  return (c->cur_pos != NULL);
+}
+// Move to next block of references. Internal, not to be called directly.
+void VP8LRefsCursorNextBlock(VP8LRefsCursor* const c);
+// Move to next position, or NULL. Should not be called if !VP8LRefsCursorOk().
+static WEBP_INLINE void VP8LRefsCursorNext(VP8LRefsCursor* const c) {
+  assert(c != NULL);
+  assert(VP8LRefsCursorOk(c));
+  if (++c->cur_pos == c->last_pos_) VP8LRefsCursorNextBlock(c);
+}
 
 // -----------------------------------------------------------------------------
 // Main entry points
 
 // Evaluates best possible backward references for specified quality.
 // Further optimize for 2D locality if use_2d_locality flag is set.
-int VP8LGetBackwardReferences(int width, int height,
-                              const uint32_t* const argb,
-                              int quality, int cache_bits, int use_2d_locality,
-                              VP8LBackwardRefs* const best);
+// The return value is the pointer to the best of the two backward refs viz,
+// refs[0] or refs[1].
+VP8LBackwardRefs* VP8LGetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int cache_bits, int use_2d_locality, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs refs[2]);
 
 // Produce an estimate for a good color cache size for the image.
 int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
-                                      int xsize, int ysize,
+                                      int xsize, int ysize, int quality,
+                                      VP8LHashChain* const hash_chain,
+                                      VP8LBackwardRefs* const ref,
                                       int* const best_cache_bits);
 
 #ifdef __cplusplus

src/enc/config.c

@@ -111,7 +111,11 @@ int WebPValidateConfig(const WebPConfig* config) {
     return 0;
   if (config->show_compressed < 0 || config->show_compressed > 1)
     return 0;
+#if WEBP_ENCODER_ABI_VERSION > 0x0204
+  if (config->preprocessing < 0 || config->preprocessing > 7)
+#else
   if (config->preprocessing < 0 || config->preprocessing > 3)
+#endif
     return 0;
   if (config->partitions < 0 || config->partitions > 3)
     return 0;
@@ -138,3 +142,25 @@ int WebPValidateConfig(const WebPConfig* config) {
 
 //------------------------------------------------------------------------------
 
+#if WEBP_ENCODER_ABI_VERSION > 0x0202
+#define MAX_LEVEL 9
+
+// Mapping between -z level and -m / -q parameter settings.
+static const struct {
+  uint8_t method_;
+  uint8_t quality_;
+} kLosslessPresets[MAX_LEVEL + 1] = {
+  { 0,  0 }, { 1, 20 }, { 2, 25 }, { 3, 30 }, { 3, 50 },
+  { 4, 50 }, { 4, 75 }, { 4, 90 }, { 5, 90 }, { 6, 100 }
+};
+
+int WebPConfigLosslessPreset(WebPConfig* config, int level) {
+  if (config == NULL || level < 0 || level > MAX_LEVEL) return 0;
+  config->lossless = 1;
+  config->method = kLosslessPresets[level].method_;
+  config->quality = kLosslessPresets[level].quality_;
+  return 1;
+}
+#endif
+
+//------------------------------------------------------------------------------

src/enc/cost.c

@@ -360,9 +360,10 @@ void VP8CalculateLevelCosts(VP8Proba* const proba) {
       for (ctx = 0; ctx < NUM_CTX; ++ctx) {
         const uint8_t* const p = proba->coeffs_[ctype][band][ctx];
         uint16_t* const table = proba->level_cost_[ctype][band][ctx];
-        const int cost_base = VP8BitCost(1, p[1]);
+        const int cost0 = (ctx > 0) ? VP8BitCost(1, p[0]) : 0;
+        const int cost_base = VP8BitCost(1, p[1]) + cost0;
         int v;
-        table[0] = VP8BitCost(0, p[1]);
+        table[0] = VP8BitCost(0, p[1]) + cost0;
         for (v = 1; v <= MAX_VARIABLE_LEVEL; ++v) {
           table[v] = cost_base + VariableLevelCost(v, p);
         }
@@ -486,4 +487,249 @@ const uint16_t VP8FixedCostsI4[NUM_BMODES][NUM_BMODES][NUM_BMODES] = {
 };
 
 //------------------------------------------------------------------------------
+// Mode costs
 
+static int GetResidualCost(int ctx0, const VP8Residual* const res) {
+  int n = res->first;
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  const int p0 = res->prob[n][ctx0][0];
+  const uint16_t* t = res->cost[n][ctx0];
+  // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0
+  // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll
+  // be missing during the loop.
+  int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0;
+
+  if (res->last < 0) {
+    return VP8BitCost(0, p0);
+  }
+  for (; n < res->last; ++n) {
+    const int v = abs(res->coeffs[n]);
+    const int b = VP8EncBands[n + 1];
+    const int ctx = (v >= 2) ? 2 : v;
+    cost += VP8LevelCost(t, v);
+    t = res->cost[b][ctx];
+  }
+  // Last coefficient is always non-zero
+  {
+    const int v = abs(res->coeffs[n]);
+    assert(v != 0);
+    cost += VP8LevelCost(t, v);
+    if (n < 15) {
+      const int b = VP8EncBands[n + 1];
+      const int ctx = (v == 1) ? 1 : 2;
+      const int last_p0 = res->prob[b][ctx][0];
+      cost += VP8BitCost(0, last_p0);
+    }
+  }
+  return cost;
+}
+
+//------------------------------------------------------------------------------
+// init function
+
+#if defined(WEBP_USE_MIPS32)
+extern int VP8GetResidualCostMIPS32(int ctx0, const VP8Residual* const res);
+#endif  // WEBP_USE_MIPS32
+
+// TODO(skal): this, and GetResidualCost(), should probably go somewhere
+// under src/dsp/ at some point.
+VP8GetResidualCostFunc VP8GetResidualCost;
+
+void VP8GetResidualCostInit(void) {
+  VP8GetResidualCost = GetResidualCost;
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_MIPS32)
+    if (VP8GetCPUInfo(kMIPS32)) {
+      VP8GetResidualCost = VP8GetResidualCostMIPS32;
+    }
+#endif
+  }
+}
+
+//------------------------------------------------------------------------------
+// helper functions for residuals struct VP8Residual.
+
+void VP8InitResidual(int first, int coeff_type,
+                     VP8Encoder* const enc, VP8Residual* const res) {
+  res->coeff_type = coeff_type;
+  res->prob  = enc->proba_.coeffs_[coeff_type];
+  res->stats = enc->proba_.stats_[coeff_type];
+  res->cost  = enc->proba_.level_cost_[coeff_type];
+  res->first = first;
+}
+
+static void SetResidualCoeffs(const int16_t* const coeffs,
+                              VP8Residual* const res) {
+  int n;
+  res->last = -1;
+  assert(res->first == 0 || coeffs[0] == 0);
+  for (n = 15; n >= 0; --n) {
+    if (coeffs[n]) {
+      res->last = n;
+      break;
+    }
+  }
+  res->coeffs = coeffs;
+}
+
+//------------------------------------------------------------------------------
+// init function
+
+#if defined(WEBP_USE_SSE2)
+extern void VP8SetResidualCoeffsSSE2(const int16_t* const coeffs,
+                                     VP8Residual* const res);
+#endif  // WEBP_USE_SSE2
+
+VP8SetResidualCoeffsFunc VP8SetResidualCoeffs;
+
+void VP8SetResidualCoeffsInit(void) {
+  VP8SetResidualCoeffs = SetResidualCoeffs;
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      VP8SetResidualCoeffs = VP8SetResidualCoeffsSSE2;
+    }
+#endif
+  }
+}
+
+//------------------------------------------------------------------------------
+// Mode costs
+
+int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]) {
+  const int x = (it->i4_ & 3), y = (it->i4_ >> 2);
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int R = 0;
+  int ctx;
+
+  VP8InitResidual(0, 3, enc, &res);
+  ctx = it->top_nz_[x] + it->left_nz_[y];
+  VP8SetResidualCoeffs(levels, &res);
+  R += VP8GetResidualCost(ctx, &res);
+  return R;
+}
+
+int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd) {
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int x, y;
+  int R = 0;
+
+  VP8IteratorNzToBytes(it);   // re-import the non-zero context
+
+  // DC
+  VP8InitResidual(0, 1, enc, &res);
+  VP8SetResidualCoeffs(rd->y_dc_levels, &res);
+  R += VP8GetResidualCost(it->top_nz_[8] + it->left_nz_[8], &res);
+
+  // AC
+  VP8InitResidual(1, 0, enc, &res);
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      R += VP8GetResidualCost(ctx, &res);
+      it->top_nz_[x] = it->left_nz_[y] = (res.last >= 0);
+    }
+  }
+  return R;
+}
+
+int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int ch, x, y;
+  int R = 0;
+
+  VP8IteratorNzToBytes(it);  // re-import the non-zero context
+
+  VP8InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        R += VP8GetResidualCost(ctx, &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = (res.last >= 0);
+      }
+    }
+  }
+  return R;
+}
+
+
+//------------------------------------------------------------------------------
+// Recording of token probabilities.
+
+// Record proba context used
+static int Record(int bit, proba_t* const stats) {
+  proba_t p = *stats;
+  if (p >= 0xffff0000u) {               // an overflow is inbound.
+    p = ((p + 1u) >> 1) & 0x7fff7fffu;  // -> divide the stats by 2.
+  }
+  // record bit count (lower 16 bits) and increment total count (upper 16 bits).
+  p += 0x00010000u + bit;
+  *stats = p;
+  return bit;
+}
+
+// We keep the table-free variant around for reference, in case.
+#define USE_LEVEL_CODE_TABLE
+
+// Simulate block coding, but only record statistics.
+// Note: no need to record the fixed probas.
+int VP8RecordCoeffs(int ctx, const VP8Residual* const res) {
+  int n = res->first;
+  // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  proba_t* s = res->stats[n][ctx];
+  if (res->last  < 0) {
+    Record(0, s + 0);
+    return 0;
+  }
+  while (n <= res->last) {
+    int v;
+    Record(1, s + 0);  // order of record doesn't matter
+    while ((v = res->coeffs[n++]) == 0) {
+      Record(0, s + 1);
+      s = res->stats[VP8EncBands[n]][0];
+    }
+    Record(1, s + 1);
+    if (!Record(2u < (unsigned int)(v + 1), s + 2)) {  // v = -1 or 1
+      s = res->stats[VP8EncBands[n]][1];
+    } else {
+      v = abs(v);
+#if !defined(USE_LEVEL_CODE_TABLE)
+      if (!Record(v > 4, s + 3)) {
+        if (Record(v != 2, s + 4))
+          Record(v == 4, s + 5);
+      } else if (!Record(v > 10, s + 6)) {
+        Record(v > 6, s + 7);
+      } else if (!Record((v >= 3 + (8 << 2)), s + 8)) {
+        Record((v >= 3 + (8 << 1)), s + 9);
+      } else {
+        Record((v >= 3 + (8 << 3)), s + 10);
+      }
+#else
+      if (v > MAX_VARIABLE_LEVEL) {
+        v = MAX_VARIABLE_LEVEL;
+      }
+
+      {
+        const int bits = VP8LevelCodes[v - 1][1];
+        int pattern = VP8LevelCodes[v - 1][0];
+        int i;
+        for (i = 0; (pattern >>= 1) != 0; ++i) {
+          const int mask = 2 << i;
+          if (pattern & 1) Record(!!(bits & mask), s + 3 + i);
+        }
+      }
+#endif
+      s = res->stats[VP8EncBands[n]][2];
+    }
+  }
+  if (n < 16) Record(0, s + 0);
+  return 1;
+}
+
+//------------------------------------------------------------------------------

src/enc/cost.h

@@ -14,12 +14,38 @@
 #ifndef WEBP_ENC_COST_H_
 #define WEBP_ENC_COST_H_
 
+#include <assert.h>
+#include <stdlib.h>
 #include "./vp8enci.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+// On-the-fly info about the current set of residuals. Handy to avoid
+// passing zillions of params.
+typedef struct {
+  int first;
+  int last;
+  const int16_t* coeffs;
+
+  int coeff_type;
+  ProbaArray* prob;
+  StatsArray* stats;
+  CostArray*  cost;
+} VP8Residual;
+
+void VP8InitResidual(int first, int coeff_type,
+                     VP8Encoder* const enc, VP8Residual* const res);
+
+typedef void (*VP8SetResidualCoeffsFunc)(const int16_t* const coeffs,
+                                         VP8Residual* const res);
+extern VP8SetResidualCoeffsFunc VP8SetResidualCoeffs;
+
+void VP8SetResidualCoeffsInit(void);  // must be called first
+
+int VP8RecordCoeffs(int ctx, const VP8Residual* const res);
+
 // approximate cost per level:
 extern const uint16_t VP8LevelFixedCosts[MAX_LEVEL + 1];
 extern const uint16_t VP8EntropyCost[256];        // 8bit fixed-point log(p)
@@ -29,6 +55,12 @@ static WEBP_INLINE int VP8BitCost(int bit, uint8_t proba) {
   return !bit ? VP8EntropyCost[proba] : VP8EntropyCost[255 - proba];
 }
 
+// Cost calculation function.
+typedef int (*VP8GetResidualCostFunc)(int ctx0, const VP8Residual* const res);
+extern VP8GetResidualCostFunc VP8GetResidualCost;
+
+void VP8GetResidualCostInit(void);  // must be called first
+
 // Level cost calculations
 extern const uint16_t VP8LevelCodes[MAX_VARIABLE_LEVEL][2];
 void VP8CalculateLevelCosts(VP8Proba* const proba);

src/enc/filter.c

@@ -13,6 +13,7 @@
 
 #include <assert.h>
 #include "./vp8enci.h"
+#include "../dsp/dsp.h"
 
 // This table gives, for a given sharpness, the filtering strength to be
 // used (at least) in order to filter a given edge step delta.
@@ -61,180 +62,6 @@ int VP8FilterStrengthFromDelta(int sharpness, int delta) {
   return kLevelsFromDelta[sharpness][pos];
 }
 
-// -----------------------------------------------------------------------------
-// NOTE: clip1, tables and InitTables are repeated entries of dsp.c
-static uint8_t abs0[255 + 255 + 1];     // abs(i)
-static uint8_t abs1[255 + 255 + 1];     // abs(i)>>1
-static int8_t sclip1[1020 + 1020 + 1];  // clips [-1020, 1020] to [-128, 127]
-static int8_t sclip2[112 + 112 + 1];    // clips [-112, 112] to [-16, 15]
-static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
-
-static int tables_ok = 0;
-
-static void InitTables(void) {
-  if (!tables_ok) {
-    int i;
-    for (i = -255; i <= 255; ++i) {
-      abs0[255 + i] = (i < 0) ? -i : i;
-      abs1[255 + i] = abs0[255 + i] >> 1;
-    }
-    for (i = -1020; i <= 1020; ++i) {
-      sclip1[1020 + i] = (i < -128) ? -128 : (i > 127) ? 127 : i;
-    }
-    for (i = -112; i <= 112; ++i) {
-      sclip2[112 + i] = (i < -16) ? -16 : (i > 15) ? 15 : i;
-    }
-    for (i = -255; i <= 255 + 255; ++i) {
-      clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i;
-    }
-    tables_ok = 1;
-  }
-}
-
-//------------------------------------------------------------------------------
-// Edge filtering functions
-
-// 4 pixels in, 2 pixels out
-static WEBP_INLINE void do_filter2(uint8_t* p, int step) {
-  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
-  const int a = 3 * (q0 - p0) + sclip1[1020 + p1 - q1];
-  const int a1 = sclip2[112 + ((a + 4) >> 3)];
-  const int a2 = sclip2[112 + ((a + 3) >> 3)];
-  p[-step] = clip1[255 + p0 + a2];
-  p[    0] = clip1[255 + q0 - a1];
-}
-
-// 4 pixels in, 4 pixels out
-static WEBP_INLINE void do_filter4(uint8_t* p, int step) {
-  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
-  const int a = 3 * (q0 - p0);
-  const int a1 = sclip2[112 + ((a + 4) >> 3)];
-  const int a2 = sclip2[112 + ((a + 3) >> 3)];
-  const int a3 = (a1 + 1) >> 1;
-  p[-2*step] = clip1[255 + p1 + a3];
-  p[-  step] = clip1[255 + p0 + a2];
-  p[      0] = clip1[255 + q0 - a1];
-  p[   step] = clip1[255 + q1 - a3];
-}
-
-// high edge-variance
-static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) {
-  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
-  return (abs0[255 + p1 - p0] > thresh) || (abs0[255 + q1 - q0] > thresh);
-}
-
-static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int thresh) {
-  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
-  return (2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) <= thresh;
-}
-
-static WEBP_INLINE int needs_filter2(const uint8_t* p,
-                                     int step, int t, int it) {
-  const int p3 = p[-4*step], p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step];
-  const int q0 = p[0], q1 = p[step], q2 = p[2*step], q3 = p[3*step];
-  if ((2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) > t)
-    return 0;
-  return abs0[255 + p3 - p2] <= it && abs0[255 + p2 - p1] <= it &&
-         abs0[255 + p1 - p0] <= it && abs0[255 + q3 - q2] <= it &&
-         abs0[255 + q2 - q1] <= it && abs0[255 + q1 - q0] <= it;
-}
-
-//------------------------------------------------------------------------------
-// Simple In-loop filtering (Paragraph 15.2)
-
-static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
-  int i;
-  for (i = 0; i < 16; ++i) {
-    if (needs_filter(p + i, stride, thresh)) {
-      do_filter2(p + i, stride);
-    }
-  }
-}
-
-static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
-  int i;
-  for (i = 0; i < 16; ++i) {
-    if (needs_filter(p + i * stride, 1, thresh)) {
-      do_filter2(p + i * stride, 1);
-    }
-  }
-}
-
-static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
-  int k;
-  for (k = 3; k > 0; --k) {
-    p += 4 * stride;
-    SimpleVFilter16(p, stride, thresh);
-  }
-}
-
-static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
-  int k;
-  for (k = 3; k > 0; --k) {
-    p += 4;
-    SimpleHFilter16(p, stride, thresh);
-  }
-}
-
-//------------------------------------------------------------------------------
-// Complex In-loop filtering (Paragraph 15.3)
-
-static WEBP_INLINE void FilterLoop24(uint8_t* p,
-                                     int hstride, int vstride, int size,
-                                     int thresh, int ithresh, int hev_thresh) {
-  while (size-- > 0) {
-    if (needs_filter2(p, hstride, thresh, ithresh)) {
-      if (hev(p, hstride, hev_thresh)) {
-        do_filter2(p, hstride);
-      } else {
-        do_filter4(p, hstride);
-      }
-    }
-    p += vstride;
-  }
-}
-
-// on three inner edges
-static void VFilter16i(uint8_t* p, int stride,
-                       int thresh, int ithresh, int hev_thresh) {
-  int k;
-  for (k = 3; k > 0; --k) {
-    p += 4 * stride;
-    FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh);
-  }
-}
-
-static void HFilter16i(uint8_t* p, int stride,
-                       int thresh, int ithresh, int hev_thresh) {
-  int k;
-  for (k = 3; k > 0; --k) {
-    p += 4;
-    FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh);
-  }
-}
-
-static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
-                      int thresh, int ithresh, int hev_thresh) {
-  FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
-  FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
-}
-
-static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
-                      int thresh, int ithresh, int hev_thresh) {
-  FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
-  FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
-}
-
-//------------------------------------------------------------------------------
-
-void (*VP8EncVFilter16i)(uint8_t*, int, int, int, int) = VFilter16i;
-void (*VP8EncHFilter16i)(uint8_t*, int, int, int, int) = HFilter16i;
-void (*VP8EncVFilter8i)(uint8_t*, uint8_t*, int, int, int, int) = VFilter8i;
-void (*VP8EncHFilter8i)(uint8_t*, uint8_t*, int, int, int, int) = HFilter8i;
-
-void (*VP8EncSimpleVFilter16i)(uint8_t*, int, int) = SimpleVFilter16i;
-void (*VP8EncSimpleHFilter16i)(uint8_t*, int, int) = SimpleHFilter16i;
-
 //------------------------------------------------------------------------------
 // Paragraph 15.4: compute the inner-edge filtering strength
 
@@ -266,14 +93,14 @@ static void DoFilter(const VP8EncIterator* const it, int level) {
   memcpy(y_dst, it->yuv_out_, YUV_SIZE * sizeof(uint8_t));
 
   if (enc->filter_hdr_.simple_ == 1) {   // simple
-    VP8EncSimpleHFilter16i(y_dst, BPS, limit);
-    VP8EncSimpleVFilter16i(y_dst, BPS, limit);
+    VP8SimpleHFilter16i(y_dst, BPS, limit);
+    VP8SimpleVFilter16i(y_dst, BPS, limit);
   } else {    // complex
     const int hev_thresh = (level >= 40) ? 2 : (level >= 15) ? 1 : 0;
-    VP8EncHFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
-    VP8EncHFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
-    VP8EncVFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
-    VP8EncVFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+    VP8HFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8HFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+    VP8VFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8VFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
   }
 }
 
@@ -387,7 +214,6 @@ static double GetMBSSIM(const uint8_t* yuv1, const uint8_t* yuv2) {
 void VP8InitFilter(VP8EncIterator* const it) {
   if (it->lf_stats_ != NULL) {
     int s, i;
-    InitTables();
     for (s = 0; s < NUM_MB_SEGMENTS; s++) {
       for (i = 0; i < MAX_LF_LEVELS; i++) {
         (*it->lf_stats_)[s][i] = 0;
@@ -468,4 +294,3 @@ void VP8AdjustFilterStrength(VP8EncIterator* const it) {
 }
 
 // -----------------------------------------------------------------------------
-

src/enc/frame.c

@@ -11,8 +11,6 @@
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
-#include <assert.h>
-#include <stdlib.h>
 #include <string.h>
 #include <math.h>
 
@@ -23,19 +21,6 @@
 #define SEGMENT_VISU 0
 #define DEBUG_SEARCH 0    // useful to track search convergence
 
-// On-the-fly info about the current set of residuals. Handy to avoid
-// passing zillions of params.
-typedef struct {
-  int first;
-  int last;
-  const int16_t* coeffs;
-
-  int coeff_type;
-  ProbaArray* prob;
-  StatsArray* stats;
-  CostArray*  cost;
-} VP8Residual;
-
 //------------------------------------------------------------------------------
 // multi-pass convergence
 
@@ -142,83 +127,6 @@ static int FinalizeSkipProba(VP8Encoder* const enc) {
   return size;
 }
 
-//------------------------------------------------------------------------------
-// Recording of token probabilities.
-
-static void ResetTokenStats(VP8Encoder* const enc) {
-  VP8Proba* const proba = &enc->proba_;
-  memset(proba->stats_, 0, sizeof(proba->stats_));
-}
-
-// Record proba context used
-static int Record(int bit, proba_t* const stats) {
-  proba_t p = *stats;
-  if (p >= 0xffff0000u) {               // an overflow is inbound.
-    p = ((p + 1u) >> 1) & 0x7fff7fffu;  // -> divide the stats by 2.
-  }
-  // record bit count (lower 16 bits) and increment total count (upper 16 bits).
-  p += 0x00010000u + bit;
-  *stats = p;
-  return bit;
-}
-
-// We keep the table free variant around for reference, in case.
-#define USE_LEVEL_CODE_TABLE
-
-// Simulate block coding, but only record statistics.
-// Note: no need to record the fixed probas.
-static int RecordCoeffs(int ctx, const VP8Residual* const res) {
-  int n = res->first;
-  // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
-  proba_t* s = res->stats[n][ctx];
-  if (res->last  < 0) {
-    Record(0, s + 0);
-    return 0;
-  }
-  while (n <= res->last) {
-    int v;
-    Record(1, s + 0);  // order of record doesn't matter
-    while ((v = res->coeffs[n++]) == 0) {
-      Record(0, s + 1);
-      s = res->stats[VP8EncBands[n]][0];
-    }
-    Record(1, s + 1);
-    if (!Record(2u < (unsigned int)(v + 1), s + 2)) {  // v = -1 or 1
-      s = res->stats[VP8EncBands[n]][1];
-    } else {
-      v = abs(v);
-#if !defined(USE_LEVEL_CODE_TABLE)
-      if (!Record(v > 4, s + 3)) {
-        if (Record(v != 2, s + 4))
-          Record(v == 4, s + 5);
-      } else if (!Record(v > 10, s + 6)) {
-        Record(v > 6, s + 7);
-      } else if (!Record((v >= 3 + (8 << 2)), s + 8)) {
-        Record((v >= 3 + (8 << 1)), s + 9);
-      } else {
-        Record((v >= 3 + (8 << 3)), s + 10);
-      }
-#else
-      if (v > MAX_VARIABLE_LEVEL)
-        v = MAX_VARIABLE_LEVEL;
-
-      {
-        const int bits = VP8LevelCodes[v - 1][1];
-        int pattern = VP8LevelCodes[v - 1][0];
-        int i;
-        for (i = 0; (pattern >>= 1) != 0; ++i) {
-          const int mask = 2 << i;
-          if (pattern & 1) Record(!!(bits & mask), s + 3 + i);
-        }
-      }
-#endif
-      s = res->stats[VP8EncBands[n]][2];
-    }
-  }
-  if (n < 16) Record(0, s + 0);
-  return 1;
-}
-
 // Collect statistics and deduce probabilities for next coding pass.
 // Return the total bit-cost for coding the probability updates.
 static int CalcTokenProba(int nb, int total) {
@@ -231,6 +139,11 @@ static int BranchCost(int nb, int total, int proba) {
   return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba);
 }
 
+static void ResetTokenStats(VP8Encoder* const enc) {
+  VP8Proba* const proba = &enc->proba_;
+  memset(proba->stats_, 0, sizeof(proba->stats_));
+}
+
 static int FinalizeTokenProbas(VP8Proba* const proba) {
   int has_changed = 0;
   int size = 0;
@@ -308,131 +221,6 @@ static void SetSegmentProbas(VP8Encoder* const enc) {
   }
 }
 
-//------------------------------------------------------------------------------
-// helper functions for residuals struct VP8Residual.
-
-static void InitResidual(int first, int coeff_type,
-                         VP8Encoder* const enc, VP8Residual* const res) {
-  res->coeff_type = coeff_type;
-  res->prob  = enc->proba_.coeffs_[coeff_type];
-  res->stats = enc->proba_.stats_[coeff_type];
-  res->cost  = enc->proba_.level_cost_[coeff_type];
-  res->first = first;
-}
-
-static void SetResidualCoeffs(const int16_t* const coeffs,
-                              VP8Residual* const res) {
-  int n;
-  res->last = -1;
-  for (n = 15; n >= res->first; --n) {
-    if (coeffs[n]) {
-      res->last = n;
-      break;
-    }
-  }
-  res->coeffs = coeffs;
-}
-
-//------------------------------------------------------------------------------
-// Mode costs
-
-static int GetResidualCost(int ctx0, const VP8Residual* const res) {
-  int n = res->first;
-  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
-  int p0 = res->prob[n][ctx0][0];
-  const uint16_t* t = res->cost[n][ctx0];
-  int cost;
-
-  if (res->last < 0) {
-    return VP8BitCost(0, p0);
-  }
-  cost = VP8BitCost(1, p0);
-  for (; n < res->last; ++n) {
-    const int v = abs(res->coeffs[n]);
-    const int b = VP8EncBands[n + 1];
-    const int ctx = (v >= 2) ? 2 : v;
-    cost += VP8LevelCost(t, v);
-    t = res->cost[b][ctx];
-    // the masking trick is faster than "if (v) cost += ..." with clang
-    cost += (v ? ~0U : 0) & VP8BitCost(1, res->prob[b][ctx][0]);
-  }
-  // Last coefficient is always non-zero
-  {
-    const int v = abs(res->coeffs[n]);
-    assert(v != 0);
-    cost += VP8LevelCost(t, v);
-    if (n < 15) {
-      const int b = VP8EncBands[n + 1];
-      const int ctx = (v == 1) ? 1 : 2;
-      const int last_p0 = res->prob[b][ctx][0];
-      cost += VP8BitCost(0, last_p0);
-    }
-  }
-  return cost;
-}
-
-int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]) {
-  const int x = (it->i4_ & 3), y = (it->i4_ >> 2);
-  VP8Residual res;
-  VP8Encoder* const enc = it->enc_;
-  int R = 0;
-  int ctx;
-
-  InitResidual(0, 3, enc, &res);
-  ctx = it->top_nz_[x] + it->left_nz_[y];
-  SetResidualCoeffs(levels, &res);
-  R += GetResidualCost(ctx, &res);
-  return R;
-}
-
-int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd) {
-  VP8Residual res;
-  VP8Encoder* const enc = it->enc_;
-  int x, y;
-  int R = 0;
-
-  VP8IteratorNzToBytes(it);   // re-import the non-zero context
-
-  // DC
-  InitResidual(0, 1, enc, &res);
-  SetResidualCoeffs(rd->y_dc_levels, &res);
-  R += GetResidualCost(it->top_nz_[8] + it->left_nz_[8], &res);
-
-  // AC
-  InitResidual(1, 0, enc, &res);
-  for (y = 0; y < 4; ++y) {
-    for (x = 0; x < 4; ++x) {
-      const int ctx = it->top_nz_[x] + it->left_nz_[y];
-      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
-      R += GetResidualCost(ctx, &res);
-      it->top_nz_[x] = it->left_nz_[y] = (res.last >= 0);
-    }
-  }
-  return R;
-}
-
-int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
-  VP8Residual res;
-  VP8Encoder* const enc = it->enc_;
-  int ch, x, y;
-  int R = 0;
-
-  VP8IteratorNzToBytes(it);  // re-import the non-zero context
-
-  InitResidual(0, 2, enc, &res);
-  for (ch = 0; ch <= 2; ch += 2) {
-    for (y = 0; y < 2; ++y) {
-      for (x = 0; x < 2; ++x) {
-        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
-        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
-        R += GetResidualCost(ctx, &res);
-        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = (res.last >= 0);
-      }
-    }
-  }
-  return R;
-}
-
 //------------------------------------------------------------------------------
 // Coefficient coding
 
@@ -521,32 +309,32 @@ static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it,
 
   pos1 = VP8BitWriterPos(bw);
   if (i16) {
-    InitResidual(0, 1, enc, &res);
-    SetResidualCoeffs(rd->y_dc_levels, &res);
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(rd->y_dc_levels, &res);
     it->top_nz_[8] = it->left_nz_[8] =
       PutCoeffs(bw, it->top_nz_[8] + it->left_nz_[8], &res);
-    InitResidual(1, 0, enc, &res);
+    VP8InitResidual(1, 0, enc, &res);
   } else {
-    InitResidual(0, 3, enc, &res);
+    VP8InitResidual(0, 3, enc, &res);
   }
 
   // luma-AC
   for (y = 0; y < 4; ++y) {
     for (x = 0; x < 4; ++x) {
       const int ctx = it->top_nz_[x] + it->left_nz_[y];
-      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
       it->top_nz_[x] = it->left_nz_[y] = PutCoeffs(bw, ctx, &res);
     }
   }
   pos2 = VP8BitWriterPos(bw);
 
   // U/V
-  InitResidual(0, 2, enc, &res);
+  VP8InitResidual(0, 2, enc, &res);
   for (ch = 0; ch <= 2; ch += 2) {
     for (y = 0; y < 2; ++y) {
       for (x = 0; x < 2; ++x) {
         const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
-        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
         it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
             PutCoeffs(bw, ctx, &res);
       }
@@ -571,33 +359,33 @@ static void RecordResiduals(VP8EncIterator* const it,
   VP8IteratorNzToBytes(it);
 
   if (it->mb_->type_ == 1) {   // i16x16
-    InitResidual(0, 1, enc, &res);
-    SetResidualCoeffs(rd->y_dc_levels, &res);
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(rd->y_dc_levels, &res);
     it->top_nz_[8] = it->left_nz_[8] =
-      RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res);
-    InitResidual(1, 0, enc, &res);
+      VP8RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res);
+    VP8InitResidual(1, 0, enc, &res);
   } else {
-    InitResidual(0, 3, enc, &res);
+    VP8InitResidual(0, 3, enc, &res);
   }
 
   // luma-AC
   for (y = 0; y < 4; ++y) {
     for (x = 0; x < 4; ++x) {
       const int ctx = it->top_nz_[x] + it->left_nz_[y];
-      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
-      it->top_nz_[x] = it->left_nz_[y] = RecordCoeffs(ctx, &res);
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] = VP8RecordCoeffs(ctx, &res);
     }
   }
 
   // U/V
-  InitResidual(0, 2, enc, &res);
+  VP8InitResidual(0, 2, enc, &res);
   for (ch = 0; ch <= 2; ch += 2) {
     for (y = 0; y < 2; ++y) {
       for (x = 0; x < 2; ++x) {
         const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
-        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
         it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
-            RecordCoeffs(ctx, &res);
+            VP8RecordCoeffs(ctx, &res);
       }
     }
   }
@@ -610,8 +398,8 @@ static void RecordResiduals(VP8EncIterator* const it,
 
 #if !defined(DISABLE_TOKEN_BUFFER)
 
-static void RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
-                         VP8TBuffer* const tokens) {
+static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
+                        VP8TBuffer* const tokens) {
   int x, y, ch;
   VP8Residual res;
   VP8Encoder* const enc = it->enc_;
@@ -619,44 +407,45 @@ static void RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
   VP8IteratorNzToBytes(it);
   if (it->mb_->type_ == 1) {   // i16x16
     const int ctx = it->top_nz_[8] + it->left_nz_[8];
-    InitResidual(0, 1, enc, &res);
-    SetResidualCoeffs(rd->y_dc_levels, &res);
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(rd->y_dc_levels, &res);
     it->top_nz_[8] = it->left_nz_[8] =
         VP8RecordCoeffTokens(ctx, 1,
                              res.first, res.last, res.coeffs, tokens);
-    RecordCoeffs(ctx, &res);
-    InitResidual(1, 0, enc, &res);
+    VP8RecordCoeffs(ctx, &res);
+    VP8InitResidual(1, 0, enc, &res);
   } else {
-    InitResidual(0, 3, enc, &res);
+    VP8InitResidual(0, 3, enc, &res);
   }
 
   // luma-AC
   for (y = 0; y < 4; ++y) {
     for (x = 0; x < 4; ++x) {
       const int ctx = it->top_nz_[x] + it->left_nz_[y];
-      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
       it->top_nz_[x] = it->left_nz_[y] =
           VP8RecordCoeffTokens(ctx, res.coeff_type,
                                res.first, res.last, res.coeffs, tokens);
-      RecordCoeffs(ctx, &res);
+      VP8RecordCoeffs(ctx, &res);
     }
   }
 
   // U/V
-  InitResidual(0, 2, enc, &res);
+  VP8InitResidual(0, 2, enc, &res);
   for (ch = 0; ch <= 2; ch += 2) {
     for (y = 0; y < 2; ++y) {
       for (x = 0; x < 2; ++x) {
         const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
-        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
         it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
             VP8RecordCoeffTokens(ctx, 2,
                                  res.first, res.last, res.coeffs, tokens);
-        RecordCoeffs(ctx, &res);
+        VP8RecordCoeffs(ctx, &res);
       }
     }
   }
   VP8IteratorBytesToNz(it);
+  return !tokens->error_;
 }
 
 #endif    // !DISABLE_TOKEN_BUFFER
@@ -719,7 +508,7 @@ static void StoreSideInfo(const VP8EncIterator* const it) {
       }
       case 7: *info = mb->alpha_; break;
       default: *info = 0; break;
-    };
+    }
   }
 #if SEGMENT_VISU  // visualize segments and prediction modes
   SetBlock(it->yuv_out_ + Y_OFF, mb->segment_ * 64, 16);
@@ -863,7 +652,10 @@ static int PreLoopInitialize(VP8Encoder* const enc) {
   for (p = 0; ok && p < enc->num_parts_; ++p) {
     ok = VP8BitWriterInit(enc->parts_ + p, bytes_per_parts);
   }
-  if (!ok) VP8EncFreeBitWriters(enc);  // malloc error occurred
+  if (!ok) {
+    VP8EncFreeBitWriters(enc);  // malloc error occurred
+    WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
   return ok;
 }
 
@@ -928,11 +720,6 @@ int VP8EncLoop(VP8Encoder* const enc) {
     } else {   // reset predictors after a skip
       ResetAfterSkip(&it);
     }
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-    if (enc->use_layer_) {
-      VP8EncCodeLayerBlock(&it);
-    }
-#endif
     StoreSideInfo(&it);
     VP8StoreFilterStats(&it);
     VP8IteratorExport(&it);
@@ -997,14 +784,13 @@ int VP8EncTokenLoop(VP8Encoder* const enc) {
         cnt = max_count;
       }
       VP8Decimate(&it, &info, rd_opt);
-      RecordTokens(&it, &info, &enc->tokens_);
+      ok = RecordTokens(&it, &info, &enc->tokens_);
+      if (!ok) {
+        WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+        break;
+      }
       size_p0 += info.H;
       distortion += info.D;
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-      if (enc->use_layer_) {
-        VP8EncCodeLayerBlock(&it);
-      }
-#endif
       if (is_last_pass) {
         StoreSideInfo(&it);
         VP8StoreFilterStats(&it);

src/enc/histogram.c

@@ -10,31 +10,64 @@
 // Author: Jyrki Alakuijala (jyrki@google.com)
 //
 #ifdef HAVE_CONFIG_H
-#include "config.h"
+#include "../webp/config.h"
 #endif
 
 #include <math.h>
-#include <stdio.h>
 
 #include "./backward_references.h"
 #include "./histogram.h"
 #include "../dsp/lossless.h"
 #include "../utils/utils.h"
 
+#define MAX_COST 1.e38
+
+// Number of partitions for the three dominant (literal, red and blue) symbol
+// costs.
+#define NUM_PARTITIONS 4
+// The size of the bin-hash corresponding to the three dominant costs.
+#define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS)
+
 static void HistogramClear(VP8LHistogram* const p) {
-  memset(p->literal_, 0, sizeof(p->literal_));
-  memset(p->red_, 0, sizeof(p->red_));
-  memset(p->blue_, 0, sizeof(p->blue_));
-  memset(p->alpha_, 0, sizeof(p->alpha_));
-  memset(p->distance_, 0, sizeof(p->distance_));
-  p->bit_cost_ = 0;
+  uint32_t* const literal = p->literal_;
+  const int cache_bits = p->palette_code_bits_;
+  const int histo_size = VP8LGetHistogramSize(cache_bits);
+  memset(p, 0, histo_size);
+  p->palette_code_bits_ = cache_bits;
+  p->literal_ = literal;
+}
+
+static void HistogramCopy(const VP8LHistogram* const src,
+                          VP8LHistogram* const dst) {
+  uint32_t* const dst_literal = dst->literal_;
+  const int dst_cache_bits = dst->palette_code_bits_;
+  const int histo_size = VP8LGetHistogramSize(dst_cache_bits);
+  assert(src->palette_code_bits_ == dst_cache_bits);
+  memcpy(dst, src, histo_size);
+  dst->literal_ = dst_literal;
+}
+
+int VP8LGetHistogramSize(int cache_bits) {
+  const int literal_size = VP8LHistogramNumCodes(cache_bits);
+  const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size;
+  assert(total_size <= (size_t)0x7fffffff);
+  return (int)total_size;
+}
+
+void VP8LFreeHistogram(VP8LHistogram* const histo) {
+  WebPSafeFree(histo);
+}
+
+void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) {
+  WebPSafeFree(histo);
 }
 
 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
                             VP8LHistogram* const histo) {
-  int i;
-  for (i = 0; i < refs->size; ++i) {
-    VP8LHistogramAddSinglePixOrCopy(histo, &refs->refs[i]);
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  while (VP8LRefsCursorOk(&c)) {
+    VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos);
+    VP8LRefsCursorNext(&c);
   }
 }
 
@@ -53,13 +86,24 @@ void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) {
   HistogramClear(p);
 }
 
+VP8LHistogram* VP8LAllocateHistogram(int cache_bits) {
+  VP8LHistogram* histo = NULL;
+  const int total_size = VP8LGetHistogramSize(cache_bits);
+  uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
+  if (memory == NULL) return NULL;
+  histo = (VP8LHistogram*)memory;
+  // literal_ won't necessary be aligned.
+  histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
+  VP8LHistogramInit(histo, cache_bits);
+  return histo;
+}
+
 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
   int i;
   VP8LHistogramSet* set;
-  VP8LHistogram* bulk;
-  const uint64_t total_size = sizeof(*set)
-                            + (uint64_t)size * sizeof(*set->histograms)
-                            + (uint64_t)size * sizeof(**set->histograms);
+  const size_t total_size = sizeof(*set)
+                            + sizeof(*set->histograms) * size
+                            + (size_t)VP8LGetHistogramSize(cache_bits) * size;
   uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
   if (memory == NULL) return NULL;
 
@@ -67,12 +111,15 @@ VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
   memory += sizeof(*set);
   set->histograms = (VP8LHistogram**)memory;
   memory += size * sizeof(*set->histograms);
-  bulk = (VP8LHistogram*)memory;
   set->max_size = size;
   set->size = size;
   for (i = 0; i < size; ++i) {
-    set->histograms[i] = bulk + i;
+    set->histograms[i] = (VP8LHistogram*)memory;
+    // literal_ won't necessary be aligned.
+    set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
     VP8LHistogramInit(set->histograms[i], cache_bits);
+    // There's no padding/alignment between successive histograms.
+    memory += VP8LGetHistogramSize(cache_bits);
   }
   return set;
 }
@@ -87,36 +134,21 @@ void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
     ++histo->literal_[PixOrCopyLiteral(v, 1)];
     ++histo->blue_[PixOrCopyLiteral(v, 0)];
   } else if (PixOrCopyIsCacheIdx(v)) {
-    int literal_ix = 256 + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
+    const int literal_ix =
+        NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
     ++histo->literal_[literal_ix];
   } else {
     int code, extra_bits;
     VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
-    ++histo->literal_[256 + code];
+    ++histo->literal_[NUM_LITERAL_CODES + code];
     VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
     ++histo->distance_[code];
   }
 }
 
-static double BitsEntropy(const int* const array, int n) {
-  double retval = 0.;
-  int sum = 0;
-  int nonzeros = 0;
-  int max_val = 0;
-  int i;
+static WEBP_INLINE double BitsEntropyRefine(int nonzeros, int sum, int max_val,
+                                            double retval) {
   double mix;
-  for (i = 0; i < n; ++i) {
-    if (array[i] != 0) {
-      sum += array[i];
-      ++nonzeros;
-      retval -= VP8LFastSLog2(array[i]);
-      if (max_val < array[i]) {
-        max_val = array[i];
-      }
-    }
-  }
-  retval += VP8LFastSLog2(sum);
-
   if (nonzeros < 5) {
     if (nonzeros <= 1) {
       return 0;
@@ -147,95 +179,142 @@ static double BitsEntropy(const int* const array, int n) {
   }
 }
 
-// Returns the cost encode the rle-encoded entropy code.
-// The constants in this function are experimental.
-static double HuffmanCost(const int* const population, int length) {
+static double BitsEntropy(const uint32_t* const array, int n) {
+  double retval = 0.;
+  uint32_t sum = 0;
+  int nonzeros = 0;
+  uint32_t max_val = 0;
+  int i;
+  for (i = 0; i < n; ++i) {
+    if (array[i] != 0) {
+      sum += array[i];
+      ++nonzeros;
+      retval -= VP8LFastSLog2(array[i]);
+      if (max_val < array[i]) {
+        max_val = array[i];
+      }
+    }
+  }
+  retval += VP8LFastSLog2(sum);
+  return BitsEntropyRefine(nonzeros, sum, max_val, retval);
+}
+
+static double BitsEntropyCombined(const uint32_t* const X,
+                                  const uint32_t* const Y, int n) {
+  double retval = 0.;
+  int sum = 0;
+  int nonzeros = 0;
+  int max_val = 0;
+  int i;
+  for (i = 0; i < n; ++i) {
+    const int xy = X[i] + Y[i];
+    if (xy != 0) {
+      sum += xy;
+      ++nonzeros;
+      retval -= VP8LFastSLog2(xy);
+      if (max_val < xy) {
+        max_val = xy;
+      }
+    }
+  }
+  retval += VP8LFastSLog2(sum);
+  return BitsEntropyRefine(nonzeros, sum, max_val, retval);
+}
+
+static double InitialHuffmanCost(void) {
   // Small bias because Huffman code length is typically not stored in
   // full length.
   static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3;
   static const double kSmallBias = 9.1;
-  double retval = kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
-  int streak = 0;
-  int i = 0;
-  for (; i < length - 1; ++i) {
-    ++streak;
-    if (population[i] == population[i + 1]) {
-      continue;
-    }
- last_streak_hack:
-    // population[i] points now to the symbol in the streak of same values.
-    if (streak > 3) {
-      if (population[i] == 0) {
-        retval += 1.5625 + 0.234375 * streak;
-      } else {
-        retval += 2.578125 + 0.703125 * streak;
-      }
-    } else {
-      if (population[i] == 0) {
-        retval += 1.796875 * streak;
-      } else {
-        retval += 3.28125 * streak;
-      }
-    }
-    streak = 0;
-  }
-  if (i == length - 1) {
-    ++streak;
-    goto last_streak_hack;
-  }
+  return kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
+}
+
+// Finalize the Huffman cost based on streak numbers and length type (<3 or >=3)
+static double FinalHuffmanCost(const VP8LStreaks* const stats) {
+  double retval = InitialHuffmanCost();
+  retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1];
+  retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1];
+  retval += 1.796875 * stats->streaks[0][0];
+  retval += 3.28125 * stats->streaks[1][0];
   return retval;
 }
 
-static double PopulationCost(const int* const population, int length) {
+// Trampolines
+static double HuffmanCost(const uint32_t* const population, int length) {
+  const VP8LStreaks stats = VP8LHuffmanCostCount(population, length);
+  return FinalHuffmanCost(&stats);
+}
+
+static double HuffmanCostCombined(const uint32_t* const X,
+                                  const uint32_t* const Y, int length) {
+  const VP8LStreaks stats = VP8LHuffmanCostCombinedCount(X, Y, length);
+  return FinalHuffmanCost(&stats);
+}
+
+// Aggregated costs
+static double PopulationCost(const uint32_t* const population, int length) {
   return BitsEntropy(population, length) + HuffmanCost(population, length);
 }
 
-static double ExtraCost(const int* const population, int length) {
-  int i;
-  double cost = 0.;
-  for (i = 2; i < length - 2; ++i) cost += (i >> 1) * population[i + 2];
-  return cost;
+static double GetCombinedEntropy(const uint32_t* const X,
+                                 const uint32_t* const Y, int length) {
+  return BitsEntropyCombined(X, Y, length) + HuffmanCostCombined(X, Y, length);
 }
 
 // Estimates the Entropy + Huffman + other block overhead size cost.
 double VP8LHistogramEstimateBits(const VP8LHistogram* const p) {
-  return PopulationCost(p->literal_, VP8LHistogramNumCodes(p))
-       + PopulationCost(p->red_, 256)
-       + PopulationCost(p->blue_, 256)
-       + PopulationCost(p->alpha_, 256)
-       + PopulationCost(p->distance_, NUM_DISTANCE_CODES)
-       + ExtraCost(p->literal_ + 256, NUM_LENGTH_CODES)
-       + ExtraCost(p->distance_, NUM_DISTANCE_CODES);
+  return
+      PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_))
+      + PopulationCost(p->red_, NUM_LITERAL_CODES)
+      + PopulationCost(p->blue_, NUM_LITERAL_CODES)
+      + PopulationCost(p->alpha_, NUM_LITERAL_CODES)
+      + PopulationCost(p->distance_, NUM_DISTANCE_CODES)
+      + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
+      + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
 }
 
 double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p) {
-  return BitsEntropy(p->literal_, VP8LHistogramNumCodes(p))
-       + BitsEntropy(p->red_, 256)
-       + BitsEntropy(p->blue_, 256)
-       + BitsEntropy(p->alpha_, 256)
-       + BitsEntropy(p->distance_, NUM_DISTANCE_CODES)
-       + ExtraCost(p->literal_ + 256, NUM_LENGTH_CODES)
-       + ExtraCost(p->distance_, NUM_DISTANCE_CODES);
+  return
+      BitsEntropy(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_))
+      + BitsEntropy(p->red_, NUM_LITERAL_CODES)
+      + BitsEntropy(p->blue_, NUM_LITERAL_CODES)
+      + BitsEntropy(p->alpha_, NUM_LITERAL_CODES)
+      + BitsEntropy(p->distance_, NUM_DISTANCE_CODES)
+      + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
+      + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
 }
 
 // -----------------------------------------------------------------------------
 // Various histogram combine/cost-eval functions
 
-// Adds 'in' histogram to 'out'
-static void HistogramAdd(const VP8LHistogram* const in,
-                         VP8LHistogram* const out) {
-  int i;
-  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
-    out->literal_[i] += in->literal_[i];
-  }
-  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
-    out->distance_[i] += in->distance_[i];
-  }
-  for (i = 0; i < 256; ++i) {
-    out->red_[i] += in->red_[i];
-    out->blue_[i] += in->blue_[i];
-    out->alpha_[i] += in->alpha_[i];
-  }
+static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
+                                       const VP8LHistogram* const b,
+                                       double cost_threshold,
+                                       double* cost) {
+  const int palette_code_bits = a->palette_code_bits_;
+  assert(a->palette_code_bits_ == b->palette_code_bits_);
+  *cost += GetCombinedEntropy(a->literal_, b->literal_,
+                              VP8LHistogramNumCodes(palette_code_bits));
+  *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
+                                 b->literal_ + NUM_LITERAL_CODES,
+                                 NUM_LENGTH_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  *cost += GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  *cost += GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  *cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  *cost += GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES);
+  *cost += VP8LExtraCostCombined(a->distance_, b->distance_,
+                                 NUM_DISTANCE_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  return 1;
 }
 
 // Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing
@@ -250,41 +329,14 @@ static double HistogramAddEval(const VP8LHistogram* const a,
                                double cost_threshold) {
   double cost = 0;
   const double sum_cost = a->bit_cost_ + b->bit_cost_;
-  int i;
-
   cost_threshold += sum_cost;
 
-  // palette_code_bits_ is part of the cost evaluation for literal_.
-  // TODO(skal): remove/simplify this palette_code_bits_?
-  out->palette_code_bits_ =
-      (a->palette_code_bits_ > b->palette_code_bits_) ? a->palette_code_bits_ :
-                                                        b->palette_code_bits_;
-  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
-    out->literal_[i] = a->literal_[i] + b->literal_[i];
+  if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
+    VP8LHistogramAdd(a, b, out);
+    out->bit_cost_ = cost;
+    out->palette_code_bits_ = a->palette_code_bits_;
   }
-  cost += PopulationCost(out->literal_, VP8LHistogramNumCodes(out));
-  cost += ExtraCost(out->literal_ + 256, NUM_LENGTH_CODES);
-  if (cost > cost_threshold) return cost;
 
-  for (i = 0; i < 256; ++i) out->red_[i] = a->red_[i] + b->red_[i];
-  cost += PopulationCost(out->red_, 256);
-  if (cost > cost_threshold) return cost;
-
-  for (i = 0; i < 256; ++i) out->blue_[i] = a->blue_[i] + b->blue_[i];
-  cost += PopulationCost(out->blue_, 256);
-  if (cost > cost_threshold) return cost;
-
-  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
-    out->distance_[i] = a->distance_[i] + b->distance_[i];
-  }
-  cost += PopulationCost(out->distance_, NUM_DISTANCE_CODES);
-  cost += ExtraCost(out->distance_, NUM_DISTANCE_CODES);
-  if (cost > cost_threshold) return cost;
-
-  for (i = 0; i < 256; ++i) out->alpha_[i] = a->alpha_[i] + b->alpha_[i];
-  cost += PopulationCost(out->alpha_, 256);
-
-  out->bit_cost_ = cost;
   return cost - sum_cost;
 }
 
@@ -294,52 +346,92 @@ static double HistogramAddEval(const VP8LHistogram* const a,
 static double HistogramAddThresh(const VP8LHistogram* const a,
                                  const VP8LHistogram* const b,
                                  double cost_threshold) {
-  int tmp[PIX_OR_COPY_CODES_MAX];  // <= max storage we'll need
-  int i;
   double cost = -a->bit_cost_;
-
-  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
-    tmp[i] = a->literal_[i] + b->literal_[i];
-  }
-  // note that the tests are ordered so that the usually largest
-  // cost shares come first.
-  cost += PopulationCost(tmp, VP8LHistogramNumCodes(a));
-  cost += ExtraCost(tmp + 256, NUM_LENGTH_CODES);
-  if (cost > cost_threshold) return cost;
-
-  for (i = 0; i < 256; ++i) tmp[i] = a->red_[i] + b->red_[i];
-  cost += PopulationCost(tmp, 256);
-  if (cost > cost_threshold) return cost;
-
-  for (i = 0; i < 256; ++i) tmp[i] = a->blue_[i] + b->blue_[i];
-  cost += PopulationCost(tmp, 256);
-  if (cost > cost_threshold) return cost;
-
-  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
-    tmp[i] = a->distance_[i] + b->distance_[i];
-  }
-  cost += PopulationCost(tmp, NUM_DISTANCE_CODES);
-  cost += ExtraCost(tmp, NUM_DISTANCE_CODES);
-  if (cost > cost_threshold) return cost;
-
-  for (i = 0; i < 256; ++i) tmp[i] = a->alpha_[i] + b->alpha_[i];
-  cost += PopulationCost(tmp, 256);
-
+  GetCombinedHistogramEntropy(a, b, cost_threshold, &cost);
   return cost;
 }
 
 // -----------------------------------------------------------------------------
 
-static void HistogramBuildImage(int xsize, int histo_bits,
-                                const VP8LBackwardRefs* const backward_refs,
-                                VP8LHistogramSet* const image) {
-  int i;
+// The structure to keep track of cost range for the three dominant entropy
+// symbols.
+// TODO(skal): Evaluate if float can be used here instead of double for
+// representing the entropy costs.
+typedef struct {
+  double literal_max_;
+  double literal_min_;
+  double red_max_;
+  double red_min_;
+  double blue_max_;
+  double blue_min_;
+} DominantCostRange;
+
+static void DominantCostRangeInit(DominantCostRange* const c) {
+  c->literal_max_ = 0.;
+  c->literal_min_ = MAX_COST;
+  c->red_max_ = 0.;
+  c->red_min_ = MAX_COST;
+  c->blue_max_ = 0.;
+  c->blue_min_ = MAX_COST;
+}
+
+static void UpdateDominantCostRange(
+    const VP8LHistogram* const h, DominantCostRange* const c) {
+  if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_;
+  if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_;
+  if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_;
+  if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_;
+  if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_;
+  if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_;
+}
+
+static void UpdateHistogramCost(VP8LHistogram* const h) {
+  const double alpha_cost = PopulationCost(h->alpha_, NUM_LITERAL_CODES);
+  const double distance_cost =
+      PopulationCost(h->distance_, NUM_DISTANCE_CODES) +
+      VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
+  const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
+  h->literal_cost_ = PopulationCost(h->literal_, num_codes) +
+                     VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES,
+                                   NUM_LENGTH_CODES);
+  h->red_cost_ = PopulationCost(h->red_, NUM_LITERAL_CODES);
+  h->blue_cost_ = PopulationCost(h->blue_, NUM_LITERAL_CODES);
+  h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ +
+                 alpha_cost + distance_cost;
+}
+
+static int GetBinIdForEntropy(double min, double max, double val) {
+  const double range = max - min + 1e-6;
+  const double delta = val - min;
+  return (int)(NUM_PARTITIONS * delta / range);
+}
+
+// TODO(vikasa): Evaluate, if there's any correlation between red & blue.
+static int GetHistoBinIndex(
+    const VP8LHistogram* const h, const DominantCostRange* const c) {
+  const int bin_id =
+      GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_) +
+      NUM_PARTITIONS * GetBinIdForEntropy(c->red_min_, c->red_max_,
+                                          h->red_cost_) +
+      NUM_PARTITIONS * NUM_PARTITIONS * GetBinIdForEntropy(c->literal_min_,
+                                                           c->literal_max_,
+                                                           h->literal_cost_);
+  assert(bin_id < BIN_SIZE);
+  return bin_id;
+}
+
+// Construct the histograms from backward references.
+static void HistogramBuild(
+    int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs,
+    VP8LHistogramSet* const image_histo) {
   int x = 0, y = 0;
   const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
-  VP8LHistogram** const histograms = image->histograms;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs);
   assert(histo_bits > 0);
-  for (i = 0; i < backward_refs->size; ++i) {
-    const PixOrCopy* const v = &backward_refs->refs[i];
+  // Construct the Histo from a given backward references.
+  while (VP8LRefsCursorOk(&c)) {
+    const PixOrCopy* const v = c.cur_pos;
     const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
     VP8LHistogramAddSinglePixOrCopy(histograms[ix], v);
     x += PixOrCopyLength(v);
@@ -347,9 +439,119 @@ static void HistogramBuildImage(int xsize, int histo_bits,
       x -= xsize;
       ++y;
     }
+    VP8LRefsCursorNext(&c);
   }
 }
 
+// Copies the histograms and computes its bit_cost.
+static void HistogramCopyAndAnalyze(
+    VP8LHistogramSet* const orig_histo, VP8LHistogramSet* const image_histo) {
+  int i;
+  const int histo_size = orig_histo->size;
+  VP8LHistogram** const orig_histograms = orig_histo->histograms;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  for (i = 0; i < histo_size; ++i) {
+    VP8LHistogram* const histo = orig_histograms[i];
+    UpdateHistogramCost(histo);
+    // Copy histograms from orig_histo[] to image_histo[].
+    HistogramCopy(histo, histograms[i]);
+  }
+}
+
+// Partition histograms to different entropy bins for three dominant (literal,
+// red and blue) symbol costs and compute the histogram aggregate bit_cost.
+static void HistogramAnalyzeEntropyBin(
+    VP8LHistogramSet* const image_histo, int16_t* const bin_map) {
+  int i;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  const int histo_size = image_histo->size;
+  const int bin_depth = histo_size + 1;
+  DominantCostRange cost_range;
+  DominantCostRangeInit(&cost_range);
+
+  // Analyze the dominant (literal, red and blue) entropy costs.
+  for (i = 0; i < histo_size; ++i) {
+    VP8LHistogram* const histo = histograms[i];
+    UpdateDominantCostRange(histo, &cost_range);
+  }
+
+  // bin-hash histograms on three of the dominant (literal, red and blue)
+  // symbol costs.
+  for (i = 0; i < histo_size; ++i) {
+    int num_histos;
+    VP8LHistogram* const histo = histograms[i];
+    const int16_t bin_id = (int16_t)GetHistoBinIndex(histo, &cost_range);
+    const int bin_offset = bin_id * bin_depth;
+    // bin_map[n][0] for every bin 'n' maintains the counter for the number of
+    // histograms in that bin.
+    // Get and increment the num_histos in that bin.
+    num_histos = ++bin_map[bin_offset];
+    assert(bin_offset + num_histos < bin_depth * BIN_SIZE);
+    // Add histogram i'th index at num_histos (last) position in the bin_map.
+    bin_map[bin_offset + num_histos] = i;
+  }
+}
+
+// Compact the histogram set by moving the valid one left in the set to the
+// head and moving the ones that have been merged to other histograms towards
+// the end.
+// TODO(vikasa): Evaluate if this method can be avoided by altering the code
+// logic of HistogramCombineEntropyBin main loop.
+static void HistogramCompactBins(VP8LHistogramSet* const image_histo) {
+  int start = 0;
+  int end = image_histo->size - 1;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  while (start < end) {
+    while (start <= end && histograms[start] != NULL &&
+           histograms[start]->bit_cost_ != 0.) {
+      ++start;
+    }
+    while (start <= end && histograms[end]->bit_cost_ == 0.) {
+      histograms[end] = NULL;
+      --end;
+    }
+    if (start < end) {
+      assert(histograms[start] != NULL);
+      assert(histograms[end] != NULL);
+      HistogramCopy(histograms[end], histograms[start]);
+      histograms[end] = NULL;
+      --end;
+    }
+  }
+  image_histo->size = end + 1;
+}
+
+static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo,
+                                       VP8LHistogram* const histos,
+                                       int16_t* const bin_map, int bin_depth,
+                                       double combine_cost_factor) {
+  int bin_id;
+  VP8LHistogram* cur_combo = histos;
+  VP8LHistogram** const histograms = image_histo->histograms;
+
+  for (bin_id = 0; bin_id < BIN_SIZE; ++bin_id) {
+    const int bin_offset = bin_id * bin_depth;
+    const int num_histos = bin_map[bin_offset];
+    const int idx1 = bin_map[bin_offset + 1];
+    int n;
+    for (n = 2; n <= num_histos; ++n) {
+      const int idx2 = bin_map[bin_offset + n];
+      const double bit_cost_idx2 = histograms[idx2]->bit_cost_;
+      if (bit_cost_idx2 > 0.) {
+        const double bit_cost_thresh = -bit_cost_idx2 * combine_cost_factor;
+        const double curr_cost_diff =
+            HistogramAddEval(histograms[idx1], histograms[idx2],
+                             cur_combo, bit_cost_thresh);
+        if (curr_cost_diff < bit_cost_thresh) {
+          HistogramCopy(cur_combo, histograms[idx1]);
+          histograms[idx2]->bit_cost_ = 0.;
+        }
+      }
+    }
+  }
+  HistogramCompactBins(image_histo);
+}
+
 static uint32_t MyRand(uint32_t *seed) {
   *seed *= 16807U;
   if (*seed == 0) {
@@ -358,48 +560,45 @@ static uint32_t MyRand(uint32_t *seed) {
   return *seed;
 }
 
-static int HistogramCombine(const VP8LHistogramSet* const in,
-                            VP8LHistogramSet* const out, int iter_mult,
-                            int num_pairs, int num_tries_no_success) {
-  int ok = 0;
-  int i, iter;
+static void HistogramCombine(VP8LHistogramSet* const image_histo,
+                             VP8LHistogramSet* const histos, int quality) {
+  int iter;
   uint32_t seed = 0;
   int tries_with_no_success = 0;
-  int out_size = in->size;
-  const int outer_iters = in->size * iter_mult;
+  int image_histo_size = image_histo->size;
+  const int iter_mult = (quality < 25) ? 2 : 2 + (quality - 25) / 8;
+  const int outer_iters = image_histo_size * iter_mult;
+  const int num_pairs = image_histo_size / 2;
+  const int num_tries_no_success = outer_iters / 2;
   const int min_cluster_size = 2;
-  VP8LHistogram* const histos = (VP8LHistogram*)malloc(2 * sizeof(*histos));
-  VP8LHistogram* cur_combo = histos + 0;    // trial merged histogram
-  VP8LHistogram* best_combo = histos + 1;   // best merged histogram so far
-  if (histos == NULL) goto End;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  VP8LHistogram* cur_combo = histos->histograms[0];   // trial histogram
+  VP8LHistogram* best_combo = histos->histograms[1];  // best histogram so far
 
-  // Copy histograms from in[] to out[].
-  assert(in->size <= out->size);
-  for (i = 0; i < in->size; ++i) {
-    in->histograms[i]->bit_cost_ = VP8LHistogramEstimateBits(in->histograms[i]);
-    *out->histograms[i] = *in->histograms[i];
-  }
-
-  // Collapse similar histograms in 'out'.
-  for (iter = 0; iter < outer_iters && out_size >= min_cluster_size; ++iter) {
+  // Collapse similar histograms in 'image_histo'.
+  for (iter = 0;
+       iter < outer_iters && image_histo_size >= min_cluster_size;
+       ++iter) {
     double best_cost_diff = 0.;
     int best_idx1 = -1, best_idx2 = 1;
     int j;
-    const int num_tries = (num_pairs < out_size) ? num_pairs : out_size;
+    const int num_tries =
+        (num_pairs < image_histo_size) ? num_pairs : image_histo_size;
     seed += iter;
     for (j = 0; j < num_tries; ++j) {
       double curr_cost_diff;
       // Choose two histograms at random and try to combine them.
-      const uint32_t idx1 = MyRand(&seed) % out_size;
+      const uint32_t idx1 = MyRand(&seed) % image_histo_size;
       const uint32_t tmp = (j & 7) + 1;
-      const uint32_t diff = (tmp < 3) ? tmp : MyRand(&seed) % (out_size - 1);
-      const uint32_t idx2 = (idx1 + diff + 1) % out_size;
+      const uint32_t diff =
+          (tmp < 3) ? tmp : MyRand(&seed) % (image_histo_size - 1);
+      const uint32_t idx2 = (idx1 + diff + 1) % image_histo_size;
       if (idx1 == idx2) {
         continue;
       }
+
       // Calculate cost reduction on combining.
-      curr_cost_diff = HistogramAddEval(out->histograms[idx1],
-                                        out->histograms[idx2],
+      curr_cost_diff = HistogramAddEval(histograms[idx1], histograms[idx2],
                                         cur_combo, best_cost_diff);
       if (curr_cost_diff < best_cost_diff) {    // found a better pair?
         {     // swap cur/best combo histograms
@@ -414,12 +613,12 @@ static int HistogramCombine(const VP8LHistogramSet* const in,
     }
 
     if (best_idx1 >= 0) {
-      *out->histograms[best_idx1] = *best_combo;
+      HistogramCopy(best_combo, histograms[best_idx1]);
       // swap best_idx2 slot with last one (which is now unused)
-      --out_size;
-      if (best_idx2 != out_size) {
-        out->histograms[best_idx2] = out->histograms[out_size];
-        out->histograms[out_size] = NULL;   // just for sanity check.
+      --image_histo_size;
+      if (best_idx2 != image_histo_size) {
+        HistogramCopy(histograms[image_histo_size], histograms[best_idx2]);
+        histograms[image_histo_size] = NULL;
       }
       tries_with_no_success = 0;
     }
@@ -427,38 +626,28 @@ static int HistogramCombine(const VP8LHistogramSet* const in,
       break;
     }
   }
-  out->size = out_size;
-  ok = 1;
-
- End:
-  free(histos);
-  return ok;
+  image_histo->size = image_histo_size;
 }
 
 // -----------------------------------------------------------------------------
 // Histogram refinement
 
-// What is the bit cost of moving square_histogram from cur_symbol to candidate.
-static double HistogramDistance(const VP8LHistogram* const square_histogram,
-                                const VP8LHistogram* const candidate,
-                                double cost_threshold) {
-  return HistogramAddThresh(candidate, square_histogram, cost_threshold);
-}
-
 // Find the best 'out' histogram for each of the 'in' histograms.
 // Note: we assume that out[]->bit_cost_ is already up-to-date.
-static void HistogramRemap(const VP8LHistogramSet* const in,
-                           const VP8LHistogramSet* const out,
+static void HistogramRemap(const VP8LHistogramSet* const orig_histo,
+                           const VP8LHistogramSet* const image_histo,
                            uint16_t* const symbols) {
   int i;
-  for (i = 0; i < in->size; ++i) {
+  VP8LHistogram** const orig_histograms = orig_histo->histograms;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  for (i = 0; i < orig_histo->size; ++i) {
     int best_out = 0;
     double best_bits =
-        HistogramDistance(in->histograms[i], out->histograms[0], 1.e38);
+        HistogramAddThresh(histograms[0], orig_histograms[i], MAX_COST);
     int k;
-    for (k = 1; k < out->size; ++k) {
+    for (k = 1; k < image_histo->size; ++k) {
       const double cur_bits =
-          HistogramDistance(in->histograms[i], out->histograms[k], best_bits);
+          HistogramAddThresh(histograms[k], orig_histograms[i], best_bits);
       if (cur_bits < best_bits) {
         best_bits = cur_bits;
         best_out = k;
@@ -468,45 +657,85 @@ static void HistogramRemap(const VP8LHistogramSet* const in,
   }
 
   // Recompute each out based on raw and symbols.
-  for (i = 0; i < out->size; ++i) {
-    HistogramClear(out->histograms[i]);
+  for (i = 0; i < image_histo->size; ++i) {
+    HistogramClear(histograms[i]);
   }
-  for (i = 0; i < in->size; ++i) {
-    HistogramAdd(in->histograms[i], out->histograms[symbols[i]]);
+
+  for (i = 0; i < orig_histo->size; ++i) {
+    const int idx = symbols[i];
+    VP8LHistogramAdd(orig_histograms[i], histograms[idx], histograms[idx]);
   }
 }
 
+static double GetCombineCostFactor(int histo_size, int quality) {
+  double combine_cost_factor = 0.16;
+  if (histo_size > 256) combine_cost_factor /= 2.;
+  if (histo_size > 512) combine_cost_factor /= 2.;
+  if (histo_size > 1024) combine_cost_factor /= 2.;
+  if (quality <= 50) combine_cost_factor /= 2.;
+  return combine_cost_factor;
+}
+
 int VP8LGetHistoImageSymbols(int xsize, int ysize,
                              const VP8LBackwardRefs* const refs,
                              int quality, int histo_bits, int cache_bits,
-                             VP8LHistogramSet* const image_in,
+                             VP8LHistogramSet* const image_histo,
                              uint16_t* const histogram_symbols) {
   int ok = 0;
   const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
   const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
-  const int histo_image_raw_size = histo_xsize * histo_ysize;
+  const int image_histo_raw_size = histo_xsize * histo_ysize;
 
-  // Heuristic params for HistogramCombine().
-  const int num_tries_no_success = 8 + (quality >> 1);
-  const int iter_mult = (quality < 27) ? 1 : 1 + ((quality - 27) >> 4);
-  const int num_pairs = (quality < 25) ? 10 : (5 * quality) >> 3;
+  // The bin_map for every bin follows following semantics:
+  // bin_map[n][0] = num_histo; // The number of histograms in that bin.
+  // bin_map[n][1] = index of first histogram in that bin;
+  // bin_map[n][num_histo] = index of last histogram in that bin;
+  // bin_map[n][num_histo + 1] ... bin_map[n][bin_depth - 1] = un-used indices.
+  const int bin_depth = image_histo_raw_size + 1;
+  int16_t* bin_map = NULL;
+  VP8LHistogramSet* const histos = VP8LAllocateHistogramSet(2, cache_bits);
+  VP8LHistogramSet* const orig_histo =
+      VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
 
-  VP8LHistogramSet* const image_out =
-      VP8LAllocateHistogramSet(histo_image_raw_size, cache_bits);
-  if (image_out == NULL) return 0;
-
-  // Build histogram image.
-  HistogramBuildImage(xsize, histo_bits, refs, image_out);
-  // Collapse similar histograms.
-  if (!HistogramCombine(image_out, image_in, iter_mult, num_pairs,
-                        num_tries_no_success)) {
+  if (orig_histo == NULL || histos == NULL) {
     goto Error;
   }
+
+  // Don't attempt linear bin-partition heuristic for:
+  // histograms of small sizes, as bin_map will be very sparse and;
+  // Higher qualities (> 90), to preserve the compression gains at those
+  // quality settings.
+  if (orig_histo->size > 2 * BIN_SIZE && quality < 90) {
+    const int bin_map_size = bin_depth * BIN_SIZE;
+    bin_map = (int16_t*)WebPSafeCalloc(bin_map_size, sizeof(*bin_map));
+    if (bin_map == NULL) goto Error;
+  }
+
+  // Construct the histograms from backward references.
+  HistogramBuild(xsize, histo_bits, refs, orig_histo);
+  // Copies the histograms and computes its bit_cost.
+  HistogramCopyAndAnalyze(orig_histo, image_histo);
+
+  if (bin_map != NULL) {
+    const double combine_cost_factor =
+        GetCombineCostFactor(image_histo_raw_size, quality);
+    HistogramAnalyzeEntropyBin(orig_histo, bin_map);
+    // Collapse histograms with similar entropy.
+    HistogramCombineEntropyBin(image_histo, histos->histograms[0],
+                               bin_map, bin_depth, combine_cost_factor);
+  }
+
+  // Collapse similar histograms by random histogram-pair compares.
+  HistogramCombine(image_histo, histos, quality);
+
   // Find the optimal map from original histograms to the final ones.
-  HistogramRemap(image_out, image_in, histogram_symbols);
+  HistogramRemap(orig_histo, image_histo, histogram_symbols);
+
   ok = 1;
 
-Error:
-  free(image_out);
+ Error:
+  WebPSafeFree(bin_map);
+  VP8LFreeHistogramSet(orig_histo);
+  VP8LFreeHistogramSet(histos);
   return ok;
 }

src/enc/histogram.h

@@ -32,18 +32,21 @@ extern "C" {
 typedef struct {
   // literal_ contains green literal, palette-code and
   // copy-length-prefix histogram
-  int literal_[PIX_OR_COPY_CODES_MAX];
-  int red_[256];
-  int blue_[256];
-  int alpha_[256];
+  uint32_t* literal_;         // Pointer to the allocated buffer for literal.
+  uint32_t red_[NUM_LITERAL_CODES];
+  uint32_t blue_[NUM_LITERAL_CODES];
+  uint32_t alpha_[NUM_LITERAL_CODES];
   // Backward reference prefix-code histogram.
-  int distance_[NUM_DISTANCE_CODES];
+  uint32_t distance_[NUM_DISTANCE_CODES];
   int palette_code_bits_;
-  double bit_cost_;   // cached value of VP8LHistogramEstimateBits(this)
+  double bit_cost_;      // cached value of VP8LHistogramEstimateBits(this)
+  double literal_cost_;  // Cached values of dominant entropy costs:
+  double red_cost_;      //   literal, red & blue.
+  double blue_cost_;
 } VP8LHistogram;
 
 // Collection of histograms with fixed capacity, allocated as one
-// big memory chunk. Can be destroyed by simply calling 'free()'.
+// big memory chunk. Can be destroyed by calling WebPSafeFree().
 typedef struct {
   int size;         // number of slots currently in use
   int max_size;     // maximum capacity
@@ -59,6 +62,9 @@ void VP8LHistogramCreate(VP8LHistogram* const p,
                          const VP8LBackwardRefs* const refs,
                          int palette_code_bits);
 
+// Return the size of the histogram for a given palette_code_bits.
+int VP8LGetHistogramSize(int palette_code_bits);
+
 // Set the palette_code_bits and reset the stats.
 void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits);
 
@@ -66,10 +72,21 @@ void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits);
 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
                             VP8LHistogram* const histo);
 
+// Free the memory allocated for the histogram.
+void VP8LFreeHistogram(VP8LHistogram* const histo);
+
+// Free the memory allocated for the histogram set.
+void VP8LFreeHistogramSet(VP8LHistogramSet* const histo);
+
 // Allocate an array of pointer to histograms, allocated and initialized
 // using 'cache_bits'. Return NULL in case of memory error.
 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits);
 
+// Allocate and initialize histogram object with specified 'cache_bits'.
+// Returns NULL in case of memory error.
+// Special case of VP8LAllocateHistogramSet, with size equals 1.
+VP8LHistogram* VP8LAllocateHistogram(int cache_bits);
+
 // Accumulate a token 'v' into a histogram.
 void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
                                      const PixOrCopy* const v);
@@ -82,9 +99,9 @@ double VP8LHistogramEstimateBits(const VP8LHistogram* const p);
 // represent the entropy code itself.
 double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p);
 
-static WEBP_INLINE int VP8LHistogramNumCodes(const VP8LHistogram* const p) {
-  return 256 + NUM_LENGTH_CODES +
-      ((p->palette_code_bits_ > 0) ? (1 << p->palette_code_bits_) : 0);
+static WEBP_INLINE int VP8LHistogramNumCodes(int palette_code_bits) {
+  return NUM_LITERAL_CODES + NUM_LENGTH_CODES +
+      ((palette_code_bits > 0) ? (1 << palette_code_bits) : 0);
 }
 
 // Builds the histogram image.

src/enc/layer.c

@@ -1,44 +0,0 @@
-// Copyright 2011 Google Inc. All Rights Reserved.
-//
-// Use of this source code is governed by a BSD-style license
-// that can be found in the COPYING file in the root of the source
-// tree. An additional intellectual property rights grant can be found
-// in the file PATENTS. All contributing project authors may
-// be found in the AUTHORS file in the root of the source tree.
-// -----------------------------------------------------------------------------
-//
-// Enhancement layer (for YUV444/422)
-//
-// Author: Skal (pascal.massimino@gmail.com)
-
-#include <stdlib.h>
-
-#include "./vp8enci.h"
-
-//------------------------------------------------------------------------------
-
-void VP8EncInitLayer(VP8Encoder* const enc) {
-  enc->use_layer_ = (enc->pic_->u0 != NULL);
-  enc->layer_data_size_ = 0;
-  enc->layer_data_ = NULL;
-  if (enc->use_layer_) {
-    VP8BitWriterInit(&enc->layer_bw_, enc->mb_w_ * enc->mb_h_ * 3);
-  }
-}
-
-void VP8EncCodeLayerBlock(VP8EncIterator* it) {
-  (void)it;   // remove a warning
-}
-
-int VP8EncFinishLayer(VP8Encoder* const enc) {
-  if (enc->use_layer_) {
-    enc->layer_data_ = VP8BitWriterFinish(&enc->layer_bw_);
-    enc->layer_data_size_ = VP8BitWriterSize(&enc->layer_bw_);
-  }
-  return 1;
-}
-
-void VP8EncDeleteLayer(VP8Encoder* enc) {
-  free(enc->layer_data_);
-}
-