Remove unnecessary check for frame parallel decode.

The check will make the code break out from the loop earlier.
But without the check. it works the same in non frame-parallel decode.
In frame parallel decode, this check is wrong as there may be left
frames inside the decoder even if there are no more frames to read.
So the loop should continue until got_data = 0.

Change-Id: I42937cec5d80d1d921a008d78dafce0928c727e0

README

@@ -55,6 +55,7 @@ COMPILING THE APPLICATIONS/LIBRARIES:
     armv6-linux-rvct
     armv6-linux-gcc
     armv6-none-rvct
+    arm64-darwin-gcc
     armv7-android-gcc
     armv7-darwin-gcc
     armv7-linux-rvct
@@ -62,6 +63,7 @@ COMPILING THE APPLICATIONS/LIBRARIES:
     armv7-none-rvct
     armv7-win32-vs11
     armv7-win32-vs12
+    armv7s-darwin-gcc
     mips32-linux-gcc
     ppc32-darwin8-gcc
     ppc32-darwin9-gcc
@@ -79,6 +81,7 @@ COMPILING THE APPLICATIONS/LIBRARIES:
     x86-darwin11-gcc
     x86-darwin12-gcc
     x86-darwin13-gcc
+    x86-iphonesimulator-gcc
     x86-linux-gcc
     x86-linux-icc
     x86-os2-gcc
@@ -95,6 +98,7 @@ COMPILING THE APPLICATIONS/LIBRARIES:
     x86_64-darwin11-gcc
     x86_64-darwin12-gcc
     x86_64-darwin13-gcc
+    x86_64-iphonesimulator-gcc
     x86_64-linux-gcc
     x86_64-linux-icc
     x86_64-solaris-gcc

build/make/Makefile

@@ -330,7 +330,10 @@ endef
 ifneq ($(target),)
 include $(SRC_PATH_BARE)/$(target:-$(TOOLCHAIN)=).mk
 endif
-ifeq ($(filter %clean,$(MAKECMDGOALS)),)
+
+skip_deps := $(filter %clean,$(MAKECMDGOALS))
+skip_deps += $(findstring testdata,$(MAKECMDGOALS))
+ifeq ($(strip $(skip_deps)),)
   # Older versions of make don't like -include directives with no arguments
   ifneq ($(filter %.d,$(OBJS-yes:.o=.d)),)
     -include $(filter %.d,$(OBJS-yes:.o=.d))

build/make/configure.sh

@@ -799,10 +799,10 @@ process_common_toolchain() {
     arm*)
         # on arm, isa versions are supersets
         case ${tgt_isa} in
-        armv8)
+        arm64|armv8)
             soft_enable neon
             ;;
-        armv7)
+        armv7|armv7s)
             soft_enable neon
             soft_enable neon_asm
             soft_enable media
@@ -831,7 +831,7 @@ process_common_toolchain() {
             arch_int=${arch_int%%te}
             check_add_asflags --defsym ARCHITECTURE=${arch_int}
             tune_cflags="-mtune="
-            if [ ${tgt_isa} = "armv7" ]; then
+            if [ ${tgt_isa} = "armv7" ] || [ ${tgt_isa} = "armv7s" ]; then
                 if [ -z "${float_abi}" ]; then
                     check_cpp <<EOF && float_abi=hard || float_abi=softfp
 #ifndef __ARM_PCS_VFP
@@ -1048,14 +1048,6 @@ EOF
         esac
     ;;
     x86*)
-        bits=32
-        enabled x86_64 && bits=64
-        check_cpp <<EOF && bits=x32
-#ifndef __ILP32__
-#error "not x32"
-#endif
-EOF
-
         case  ${tgt_os} in
             win*)
                 enabled gcc && add_cflags -fno-common
@@ -1094,8 +1086,6 @@ EOF
                 esac
             ;;
             gcc*)
-                add_cflags -m${bits}
-                add_ldflags -m${bits}
                 link_with_cc=gcc
                 tune_cflags="-march="
                 setup_gnu_toolchain
@@ -1120,6 +1110,20 @@ EOF
             ;;
         esac
 
+        bits=32
+        enabled x86_64 && bits=64
+        check_cpp <<EOF && bits=x32
+#ifndef __ILP32__
+#error "not x32"
+#endif
+EOF
+        case ${tgt_cc} in
+            gcc*)
+                add_cflags -m${bits}
+                add_ldflags -m${bits}
+            ;;
+        esac
+
         soft_enable runtime_cpu_detect
         # We can't use 'check_cflags' until the compiler is configured and CC is
         # populated.
@@ -1222,9 +1226,11 @@ EOF
         fi
     fi
 
-    # default use_x86inc to yes if pic is no or 64bit or we are not on darwin
+    tgt_os_no_version=$(echo "${tgt_os}" | tr -d "[0-9]")
-    if [ ${tgt_isa} = x86_64 -o ! "$pic" = "yes" -o \
+    # Default use_x86inc to yes when we are 64 bit, non-pic, or on any
-         "${tgt_os#darwin}" = "${tgt_os}"  ]; then
+    # non-Darwin target.
+    if [ "${tgt_isa}" = "x86_64" ] || [ "${pic}" != "yes" ] || \
+            [ "${tgt_os_no_version}" != "darwin" ]; then
         soft_enable use_x86inc
     fi
 

build/make/gen_msvs_proj.sh

@@ -137,7 +137,9 @@ for opt in "$@"; do
         ;;
         --lib) proj_kind="lib"
         ;;
-        --src-path-bare=*) src_path_bare=$(fix_path "$optval")
+        --src-path-bare=*)
+            src_path_bare=$(fix_path "$optval")
+            src_path_bare=${src_path_bare%/}
         ;;
         --static-crt) use_static_runtime=true
         ;;
@@ -151,9 +153,9 @@ for opt in "$@"; do
             esac
         ;;
         -I*)
-            opt="${opt%/}"
             opt=${opt##-I}
             opt=$(fix_path "$opt")
+            opt="${opt%/}"
             incs="${incs}${incs:+;}"${opt}""
             yasmincs="${yasmincs} -I"${opt}""
         ;;
@@ -414,7 +416,7 @@ generate_vcproj() {
                     vpx)
                         tag Tool \
                             Name="VCPreBuildEventTool" \
-                            CommandLine="call obj_int_extract.bat $src_path_bare $plat_no_ws\\\$(ConfigurationName)" \
+                            CommandLine="call obj_int_extract.bat "$src_path_bare" $plat_no_ws\\\$(ConfigurationName)" \
 
                         tag Tool \
                             Name="VCCLCompilerTool" \

build/make/gen_msvs_vcxproj.sh

@@ -157,7 +157,9 @@ for opt in "$@"; do
         ;;
         --lib) proj_kind="lib"
         ;;
-        --src-path-bare=*) src_path_bare=$(fix_path "$optval")
+        --src-path-bare=*)
+            src_path_bare=$(fix_path "$optval")
+            src_path_bare=${src_path_bare%/}
         ;;
         --static-crt) use_static_runtime=true
         ;;
@@ -173,9 +175,9 @@ for opt in "$@"; do
             esac
         ;;
         -I*)
-            opt="${opt%/}"
             opt=${opt##-I}
             opt=$(fix_path "$opt")
+            opt="${opt%/}"
             incs="${incs}${incs:+;}"${opt}""
             yasmincs="${yasmincs} -I"${opt}""
         ;;

build/make/iosbuild.sh

@@ -0,0 +1,248 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE 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.
+##
+##
+## This script generates 'VPX.framework'. An iOS app can encode and decode VPx
+## video by including 'VPX.framework'.
+##
+## Run iosbuild.sh to create 'VPX.framework' in the current directory.
+##
+set -e
+devnull='> /dev/null 2>&1'
+
+BUILD_ROOT="_iosbuild"
+DIST_DIR="_dist"
+FRAMEWORK_DIR="VPX.framework"
+HEADER_DIR="${FRAMEWORK_DIR}/Headers/vpx"
+MAKE_JOBS=1
+LIBVPX_SOURCE_DIR=$(dirname "$0" | sed -e s,/build/make,,)
+LIPO=$(xcrun -sdk iphoneos${SDK} -find lipo)
+ORIG_PWD="$(pwd)"
+TARGETS="arm64-darwin-gcc
+         armv6-darwin-gcc
+         armv7-darwin-gcc
+         armv7s-darwin-gcc
+         x86-iphonesimulator-gcc
+         x86_64-iphonesimulator-gcc"
+
+# Configures for the target specified by $1, and invokes make with the dist
+# target using $DIST_DIR as the distribution output directory.
+build_target() {
+  local target="$1"
+  local old_pwd="$(pwd)"
+
+  vlog "***Building target: ${target}***"
+
+  mkdir "${target}"
+  cd "${target}"
+  eval "../../${LIBVPX_SOURCE_DIR}/configure" --target="${target}" \
+      --disable-docs ${devnull}
+  export DIST_DIR
+  eval make -j ${MAKE_JOBS} dist ${devnull}
+  cd "${old_pwd}"
+
+  vlog "***Done building target: ${target}***"
+}
+
+# Returns the preprocessor symbol for the target specified by $1.
+target_to_preproc_symbol() {
+  target="$1"
+  case "${target}" in
+    arm64-*)
+      echo "__aarch64__"
+      ;;
+    armv6-*)
+      echo "__ARM_ARCH_6__"
+      ;;
+    armv7-*)
+      echo "__ARM_ARCH_7__"
+      ;;
+    armv7s-*)
+      echo "__ARM_ARCH_7S__"
+      ;;
+    x86-*)
+      echo "__i386__"
+      ;;
+    x86_64-*)
+      echo "__x86_64__"
+      ;;
+    *)
+      echo "#error ${target} unknown/unsupported"
+      return 1
+      ;;
+  esac
+}
+
+# Create a vpx_config.h shim that, based on preprocessor settings for the
+# current target CPU, includes the real vpx_config.h for the current target.
+# $1 is the list of targets.
+create_vpx_framework_config_shim() {
+  local targets="$1"
+  local config_file="${HEADER_DIR}/vpx_config.h"
+  local preproc_symbol=""
+  local target=""
+  local include_guard="VPX_FRAMEWORK_HEADERS_VPX_VPX_CONFIG_H_"
+
+  local file_header="/*
+ *  Copyright (c) $(date +%Y) The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+
+/* GENERATED FILE: DO NOT EDIT! */
+
+#ifndef ${include_guard}
+#define ${include_guard}
+
+#if defined"
+
+  printf "%s" "${file_header}" > "${config_file}"
+  for target in ${targets}; do
+    preproc_symbol=$(target_to_preproc_symbol "${target}")
+    printf " ${preproc_symbol}\n" >> "${config_file}"
+    printf "#include \"VPX/vpx/${target}/vpx_config.h\"\n" >> "${config_file}"
+    printf "#elif defined" >> "${config_file}"
+    mkdir "${HEADER_DIR}/${target}"
+    cp -p "${BUILD_ROOT}/${target}/vpx_config.h" "${HEADER_DIR}/${target}"
+  done
+
+  # Consume the last line of output from the loop: We don't want it.
+  sed -i '' -e '$d' "${config_file}"
+
+  printf "#endif\n\n" >> "${config_file}"
+  printf "#endif  // ${include_guard}" >> "${config_file}"
+}
+
+# Configures and builds each target specified by $1, and then builds
+# VPX.framework.
+build_framework() {
+  local lib_list=""
+  local targets="$1"
+  local target=""
+  local target_dist_dir=""
+
+  # Clean up from previous build(s).
+  rm -rf "${BUILD_ROOT}" "${FRAMEWORK_DIR}"
+
+  # Create output dirs.
+  mkdir -p "${BUILD_ROOT}"
+  mkdir -p "${HEADER_DIR}"
+
+  cd "${BUILD_ROOT}"
+
+  for target in ${targets}; do
+    build_target "${target}"
+    target_dist_dir="${BUILD_ROOT}/${target}/${DIST_DIR}"
+    lib_list="${lib_list} ${target_dist_dir}/lib/libvpx.a"
+  done
+
+  cd "${ORIG_PWD}"
+
+  # The basic libvpx API includes are all the same; just grab the most recent
+  # set.
+  cp -p "${target_dist_dir}"/include/vpx/* "${HEADER_DIR}"
+
+  # Build the fat library.
+  ${LIPO} -create ${lib_list} -output ${FRAMEWORK_DIR}/VPX
+
+  # Create the vpx_config.h shim that allows usage of vpx_config.h from
+  # within VPX.framework.
+  create_vpx_framework_config_shim "${targets}"
+
+  # Copy in vpx_version.h.
+  cp -p "${BUILD_ROOT}/${target}/vpx_version.h" "${HEADER_DIR}"
+
+  vlog "Created fat library ${FRAMEWORK_DIR}/VPX containing:"
+  for lib in ${lib_list}; do
+    vlog "  $(echo ${lib} | awk -F / '{print $2, $NF}')"
+  done
+
+  # TODO(tomfinegan): Verify that expected targets are included within
+  # VPX.framework/VPX via lipo -info.
+}
+
+# Trap function. Cleans up the subtree used to build all targets contained in
+# $TARGETS.
+cleanup() {
+  cd "${ORIG_PWD}"
+
+  if [ "${PRESERVE_BUILD_OUTPUT}" != "yes" ]; then
+    rm -rf "${BUILD_ROOT}"
+  fi
+}
+
+iosbuild_usage() {
+cat << EOF
+  Usage: ${0##*/} [arguments]
+    --help: Display this message and exit.
+    --jobs: Number of make jobs.
+    --preserve-build-output: Do not delete the build directory.
+    --show-build-output: Show output from each library build.
+    --verbose: Output information about the environment and each stage of the
+               build.
+EOF
+}
+
+vlog() {
+  if [ "${VERBOSE}" = "yes" ]; then
+    echo "$@"
+  fi
+}
+
+trap cleanup EXIT
+
+# Parse the command line.
+while [ -n "$1" ]; do
+  case "$1" in
+    --help)
+      iosbuild_usage
+      exit
+      ;;
+    --jobs)
+      MAKE_JOBS="$2"
+      shift
+      ;;
+    --preserve-build-output)
+      PRESERVE_BUILD_OUTPUT=yes
+      ;;
+    --show-build-output)
+      devnull=
+      ;;
+    --verbose)
+      VERBOSE=yes
+      ;;
+    *)
+      iosbuild_usage
+      exit 1
+      ;;
+  esac
+  shift
+done
+
+if [ "${VERBOSE}" = "yes" ]; then
+cat << EOF
+  BUILD_ROOT=${BUILD_ROOT}
+  DIST_DIR=${DIST_DIR}
+  FRAMEWORK_DIR=${FRAMEWORK_DIR}
+  HEADER_DIR=${HEADER_DIR}
+  MAKE_JOBS=${MAKE_JOBS}
+  PRESERVE_BUILD_OUTPUT=${PRESERVE_BUILD_OUTPUT}
+  LIBVPX_SOURCE_DIR=${LIBVPX_SOURCE_DIR}
+  LIPO=${LIPO}
+  ORIG_PWD=${ORIG_PWD}
+  TARGETS="${TARGETS}"
+EOF
+fi
+
+build_framework "${TARGETS}"

configure

@@ -96,6 +96,7 @@ all_platforms="${all_platforms} armv6-darwin-gcc"
 all_platforms="${all_platforms} armv6-linux-rvct"
 all_platforms="${all_platforms} armv6-linux-gcc"
 all_platforms="${all_platforms} armv6-none-rvct"
+all_platforms="${all_platforms} arm64-darwin-gcc"
 all_platforms="${all_platforms} armv7-android-gcc"   #neon Cortex-A8
 all_platforms="${all_platforms} armv7-darwin-gcc"    #neon Cortex-A8
 all_platforms="${all_platforms} armv7-linux-rvct"    #neon Cortex-A8
@@ -103,6 +104,7 @@ all_platforms="${all_platforms} armv7-linux-gcc"     #neon Cortex-A8
 all_platforms="${all_platforms} armv7-none-rvct"     #neon Cortex-A8
 all_platforms="${all_platforms} armv7-win32-vs11"
 all_platforms="${all_platforms} armv7-win32-vs12"
+all_platforms="${all_platforms} armv7s-darwin-gcc"
 all_platforms="${all_platforms} mips32-linux-gcc"
 all_platforms="${all_platforms} ppc32-darwin8-gcc"
 all_platforms="${all_platforms} ppc32-darwin9-gcc"
@@ -271,6 +273,8 @@ EXPERIMENT_LIST="
     alpha
     multiple_arf
     spatial_svc
+    denoising
+    fp_mb_stats
 "
 CONFIG_LIST="
     external_build

examples/vp9_spatial_svc_encoder.c

@@ -28,16 +28,6 @@
 #include "vpx/vpx_encoder.h"
 #include "./vpxstats.h"
 
-static const struct arg_enum_list encoding_mode_enum[] = {
-  {"i", INTER_LAYER_PREDICTION_I},
-  {"alt-ip", ALT_INTER_LAYER_PREDICTION_IP},
-  {"ip", INTER_LAYER_PREDICTION_IP},
-  {"gf", USE_GOLDEN_FRAME},
-  {NULL, 0}
-};
-
-static const arg_def_t encoding_mode_arg = ARG_DEF_ENUM(
-    "m", "encoding-mode", 1, "Encoding mode algorithm", encoding_mode_enum);
 static const arg_def_t skip_frames_arg =
     ARG_DEF("s", "skip-frames", 1, "input frames to skip");
 static const arg_def_t frames_arg =
@@ -58,9 +48,6 @@ static const arg_def_t quantizers_arg =
     ARG_DEF("q", "quantizers", 1, "quantizers for non key frames, also will "
             "be applied to key frames if -qn is not specified (lowest to "
             "highest layer)");
-static const arg_def_t quantizers_keyframe_arg =
-    ARG_DEF("qn", "quantizers-keyframe", 1, "quantizers for key frames (lowest "
-        "to highest layer)");
 static const arg_def_t passes_arg =
     ARG_DEF("p", "passes", 1, "Number of passes (1/2)");
 static const arg_def_t pass_arg =
@@ -77,16 +64,13 @@ static const arg_def_t max_bitrate_arg =
     ARG_DEF(NULL, "max-bitrate", 1, "Maximum bitrate");
 
 static const arg_def_t *svc_args[] = {
-  &encoding_mode_arg, &frames_arg,        &width_arg,       &height_arg,
+  &frames_arg,        &width_arg,         &height_arg,
   &timebase_arg,      &bitrate_arg,       &skip_frames_arg, &layers_arg,
-  &kf_dist_arg,       &scale_factors_arg, &quantizers_arg,
+  &kf_dist_arg,       &scale_factors_arg, &quantizers_arg,  &passes_arg,
-  &quantizers_keyframe_arg,               &passes_arg,      &pass_arg,
+  &pass_arg,          &fpf_name_arg,      &min_q_arg,       &max_q_arg,
-  &fpf_name_arg,      &min_q_arg,         &max_q_arg,       &min_bitrate_arg,
+  &min_bitrate_arg,   &max_bitrate_arg,   NULL
-  &max_bitrate_arg,   NULL
 };
 
-static const SVC_ENCODING_MODE default_encoding_mode =
-    INTER_LAYER_PREDICTION_IP;
 static const uint32_t default_frames_to_skip = 0;
 static const uint32_t default_frames_to_code = 60 * 60;
 static const uint32_t default_width = 1920;
@@ -135,7 +119,6 @@ static void parse_command_line(int argc, const char **argv_,
   // initialize SvcContext with parameters that will be passed to vpx_svc_init
   svc_ctx->log_level = SVC_LOG_DEBUG;
   svc_ctx->spatial_layers = default_spatial_layers;
-  svc_ctx->encoding_mode = default_encoding_mode;
 
   // start with default encoder configuration
   res = vpx_codec_enc_config_default(vpx_codec_vp9_cx(), enc_cfg, 0);
@@ -161,9 +144,7 @@ static void parse_command_line(int argc, const char **argv_,
   for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
     arg.argv_step = 1;
 
-    if (arg_match(&arg, &encoding_mode_arg, argi)) {
+    if (arg_match(&arg, &frames_arg, argi)) {
-      svc_ctx->encoding_mode = arg_parse_enum_or_int(&arg);
-    } else if (arg_match(&arg, &frames_arg, argi)) {
       app_input->frames_to_code = arg_parse_uint(&arg);
     } else if (arg_match(&arg, &width_arg, argi)) {
       enc_cfg->g_w = arg_parse_uint(&arg);
@@ -183,9 +164,7 @@ static void parse_command_line(int argc, const char **argv_,
     } else if (arg_match(&arg, &scale_factors_arg, argi)) {
       vpx_svc_set_scale_factors(svc_ctx, arg.val);
     } else if (arg_match(&arg, &quantizers_arg, argi)) {
-      vpx_svc_set_quantizers(svc_ctx, arg.val, 0);
+      vpx_svc_set_quantizers(svc_ctx, arg.val);
-    } else if (arg_match(&arg, &quantizers_keyframe_arg, argi)) {
-      vpx_svc_set_quantizers(svc_ctx, arg.val, 1);
     } else if (arg_match(&arg, &passes_arg, argi)) {
       passes = arg_parse_uint(&arg);
       if (passes < 1 || passes > 2) {
@@ -270,12 +249,12 @@ static void parse_command_line(int argc, const char **argv_,
 
   printf(
       "Codec %s\nframes: %d, skip: %d\n"
-      "mode: %d, layers: %d\n"
+      "layers: %d\n"
       "width %d, height: %d,\n"
       "num: %d, den: %d, bitrate: %d,\n"
       "gop size: %d\n",
       vpx_codec_iface_name(vpx_codec_vp9_cx()), app_input->frames_to_code,
-      app_input->frames_to_skip, svc_ctx->encoding_mode,
+      app_input->frames_to_skip,
       svc_ctx->spatial_layers, enc_cfg->g_w, enc_cfg->g_h,
       enc_cfg->g_timebase.num, enc_cfg->g_timebase.den,
       enc_cfg->rc_target_bitrate, enc_cfg->kf_max_dist);
@@ -296,6 +275,7 @@ int main(int argc, const char **argv) {
   int frame_duration = 1; /* 1 timebase tick per frame */
   FILE *infile = NULL;
   int end_of_stream = 0;
+  int frame_size;
 
   memset(&svc_ctx, 0, sizeof(svc_ctx));
   svc_ctx.log_print = 1;
@@ -351,11 +331,10 @@ int main(int argc, const char **argv) {
       die_codec(&codec, "Failed to encode frame");
     }
     if (!(app_input.passes == 2 && app_input.pass == 1)) {
-      if (vpx_svc_get_frame_size(&svc_ctx) > 0) {
+      while ((frame_size = vpx_svc_get_frame_size(&svc_ctx)) > 0) {
         vpx_video_writer_write_frame(writer,
                                      vpx_svc_get_buffer(&svc_ctx),
-                                     vpx_svc_get_frame_size(&svc_ctx),
+                                     frame_size, pts);
-                                     pts);
       }
     }
     if (vpx_svc_get_rc_stats_buffer_size(&svc_ctx) > 0) {

libs.mk

@@ -170,7 +170,7 @@ CODEC_SRCS-$(BUILD_LIBVPX) += vpx_ports/vpx_once.h
 CODEC_SRCS-$(BUILD_LIBVPX) += $(BUILD_PFX)vpx_config.c
 INSTALL-SRCS-no += $(BUILD_PFX)vpx_config.c
 ifeq ($(ARCH_X86)$(ARCH_X86_64),yes)
-CODEC_SRCS-$(BUILD_LIBVPX) += third_party/x86inc/x86inc.asm
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += third_party/x86inc/x86inc.asm
 endif
 CODEC_EXPORTS-$(BUILD_LIBVPX) += vpx/exports_com
 CODEC_EXPORTS-$(CONFIG_ENCODERS) += vpx/exports_enc

test/convolve_test.cc

@@ -645,6 +645,26 @@ INSTANTIATE_TEST_CASE_P(SSSE3, ConvolveTest, ::testing::Values(
 #endif
 
 #if HAVE_AVX2
+// TODO(jzern): these prototypes can be removed after the avx2 versions are
+// reenabled in vp9_rtcd_defs.pl.
+extern "C" {
+void vp9_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4,
+                             int w, int h);
+void vp9_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h);
+void vp9_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride,
+                        uint8_t *dst, ptrdiff_t dst_stride,
+                        const int16_t *filter_x, int x_step_q4,
+                        const int16_t *filter_y, int y_step_q4,
+                        int w, int h);
+}
+
 const ConvolveFunctions convolve8_avx2(
     vp9_convolve8_horiz_avx2, vp9_convolve8_avg_horiz_ssse3,
     vp9_convolve8_vert_avx2, vp9_convolve8_avg_vert_ssse3,
@@ -655,8 +675,10 @@ INSTANTIATE_TEST_CASE_P(AVX2, ConvolveTest, ::testing::Values(
     make_tuple(8, 4, &convolve8_avx2),
     make_tuple(4, 8, &convolve8_avx2),
     make_tuple(8, 8, &convolve8_avx2),
+    make_tuple(8, 16, &convolve8_avx2)));
+
+INSTANTIATE_TEST_CASE_P(DISABLED_AVX2, ConvolveTest, ::testing::Values(
     make_tuple(16, 8, &convolve8_avx2),
-    make_tuple(8, 16, &convolve8_avx2),
     make_tuple(16, 16, &convolve8_avx2),
     make_tuple(32, 16, &convolve8_avx2),
     make_tuple(16, 32, &convolve8_avx2),

test/cpu_speed_test.cc

@@ -14,38 +14,65 @@
 #include "test/encode_test_driver.h"
 #include "test/i420_video_source.h"
 #include "test/util.h"
+#include "test/y4m_video_source.h"
 
 namespace {
 
+const int kMaxPSNR = 100;
+
 class CpuSpeedTest : public ::libvpx_test::EncoderTest,
     public ::libvpx_test::CodecTestWith2Params<
         libvpx_test::TestMode, int> {
  protected:
-  CpuSpeedTest() : EncoderTest(GET_PARAM(0)) {}
+  CpuSpeedTest()
+      : EncoderTest(GET_PARAM(0)),
+        encoding_mode_(GET_PARAM(1)),
+        set_cpu_used_(GET_PARAM(2)),
+        min_psnr_(kMaxPSNR) {}
   virtual ~CpuSpeedTest() {}
 
   virtual void SetUp() {
     InitializeConfig();
-    SetMode(GET_PARAM(1));
+    SetMode(encoding_mode_);
-    set_cpu_used_ = GET_PARAM(2);
+    if (encoding_mode_ != ::libvpx_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = VPX_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = VPX_CBR;
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    min_psnr_ = kMaxPSNR;
   }
 
   virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
                                   ::libvpx_test::Encoder *encoder) {
     if (video->frame() == 1) {
       encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
+      if (encoding_mode_ != ::libvpx_test::kRealTime) {
         encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
         encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
         encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
         encoder->Control(VP8E_SET_ARNR_TYPE, 3);
       }
     }
+  }
 
   virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
     if (pkt->data.frame.flags & VPX_FRAME_IS_KEY) {
     }
   }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < min_psnr_)
+      min_psnr_ = pkt->data.psnr.psnr[0];
+  }
+
+  ::libvpx_test::TestMode encoding_mode_;
   int set_cpu_used_;
+  double min_psnr_;
 };
 
 TEST_P(CpuSpeedTest, TestQ0) {
@@ -53,7 +80,6 @@ TEST_P(CpuSpeedTest, TestQ0) {
   // without a mismatch when passing in a very low max q.  This pushes
   // the encoder to producing lots of big partitions which will likely
   // extend into the border and test the border condition.
-  cfg_.g_lag_in_frames = 25;
   cfg_.rc_2pass_vbr_minsection_pct = 5;
   cfg_.rc_2pass_vbr_minsection_pct = 2000;
   cfg_.rc_target_bitrate = 400;
@@ -63,16 +89,32 @@ TEST_P(CpuSpeedTest, TestQ0) {
   ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
                                        20);
 
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
   ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(min_psnr_, kMaxPSNR);
 }
 
+TEST_P(CpuSpeedTest, TestScreencastQ0) {
+  ::libvpx_test::Y4mVideoSource video("screendata.y4m", 0, 25);
+  cfg_.g_timebase = video.timebase();
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_minsection_pct = 2000;
+  cfg_.rc_target_bitrate = 400;
+  cfg_.rc_max_quantizer = 0;
+  cfg_.rc_min_quantizer = 0;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(min_psnr_, kMaxPSNR);
+}
 
 TEST_P(CpuSpeedTest, TestEncodeHighBitrate) {
   // Validate that this non multiple of 64 wide clip encodes and decodes
   // without a mismatch when passing in a very low max q.  This pushes
   // the encoder to producing lots of big partitions which will likely
   // extend into the border and test the border condition.
-  cfg_.g_lag_in_frames = 25;
   cfg_.rc_2pass_vbr_minsection_pct = 5;
   cfg_.rc_2pass_vbr_minsection_pct = 2000;
   cfg_.rc_target_bitrate = 12000;
@@ -89,7 +131,6 @@ TEST_P(CpuSpeedTest, TestLowBitrate) {
   // when passing in a very high min q.  This pushes the encoder to producing
   // lots of small partitions which might will test the other condition.
 
-  cfg_.g_lag_in_frames = 25;
   cfg_.rc_2pass_vbr_minsection_pct = 5;
   cfg_.rc_2pass_vbr_minsection_pct = 2000;
   cfg_.rc_target_bitrate = 200;
@@ -108,6 +149,7 @@ using std::tr1::make_tuple;
 
 VP9_INSTANTIATE_TEST_CASE(
     CpuSpeedTest,
-    ::testing::Values(::libvpx_test::kTwoPassGood, ::libvpx_test::kOnePassGood),
+    ::testing::Values(::libvpx_test::kTwoPassGood, ::libvpx_test::kOnePassGood,
+                      ::libvpx_test::kRealTime),
     ::testing::Range(0, 8));
 }  // namespace

test/datarate_test.cc

@@ -576,7 +576,7 @@ TEST_P(DatarateTestVP9Large, BasicRateTargeting3TemporalLayersFrameDropping) {
     // Expect some frame drops in this test: for this 200 frames test,
     // expect at least 10% and not more than 60% drops.
     ASSERT_GE(num_drops_, 20);
-    ASSERT_LE(num_drops_, 120);
+    ASSERT_LE(num_drops_, 130);
   }
 }
 

test/dct16x16_test.cc

@@ -606,4 +606,29 @@ INSTANTIATE_TEST_CASE_P(
     ::testing::Values(
         make_tuple(&vp9_fdct16x16_c, &vp9_idct16x16_256_add_ssse3, 0)));
 #endif
+
+#if HAVE_AVX2
+// TODO(jzern): these prototypes can be removed after the avx2 versions are
+// reenabled in vp9_rtcd_defs.pl.
+extern "C" {
+void vp9_fdct16x16_avx2(const int16_t *input, int16_t *output, int stride);
+void vp9_fht16x16_avx2(const int16_t *input, int16_t *output, int stride,
+                       int tx_type);
+}
+INSTANTIATE_TEST_CASE_P(
+    DISABLED_AVX2, Trans16x16DCT,
+    ::testing::Values(
+        make_tuple(&vp9_fdct16x16_avx2,
+                   &vp9_idct16x16_256_add_c, 0)));
+INSTANTIATE_TEST_CASE_P(
+    AVX2, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht16x16_avx2, &vp9_iht16x16_256_add_c, 3)));
+INSTANTIATE_TEST_CASE_P(
+    DISABLED_AVX2, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht16x16_avx2, &vp9_iht16x16_256_add_c, 0),
+        make_tuple(&vp9_fht16x16_avx2, &vp9_iht16x16_256_add_c, 1),
+        make_tuple(&vp9_fht16x16_avx2, &vp9_iht16x16_256_add_c, 2)));
+#endif
 }  // namespace

test/decode_test_driver.cc

@@ -15,13 +15,27 @@
 
 namespace libvpx_test {
 
+const char kVP8Name[] = "WebM Project VP8";
+
+vpx_codec_err_t Decoder::PeekStream(const uint8_t *cxdata, size_t size,
+                                    vpx_codec_stream_info_t *stream_info) {
+  return vpx_codec_peek_stream_info(CodecInterface(),
+                                    cxdata, static_cast<unsigned int>(size),
+                                    stream_info);
+}
+
 vpx_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size) {
+  return DecodeFrame(cxdata, size, NULL);
+}
+
+vpx_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size,
+                                     void *user_priv) {
   vpx_codec_err_t res_dec;
   InitOnce();
   REGISTER_STATE_CHECK(
       res_dec = vpx_codec_decode(&decoder_,
                                  cxdata, static_cast<unsigned int>(size),
-                                 NULL, 0));
+                                 user_priv, 0));
   return res_dec;
 }
 
@@ -29,13 +43,37 @@ void DecoderTest::RunLoop(CompressedVideoSource *video) {
   vpx_codec_dec_cfg_t dec_cfg = {0};
   Decoder* const decoder = codec_->CreateDecoder(dec_cfg, 0);
   ASSERT_TRUE(decoder != NULL);
+  const char *codec_name = decoder->GetDecoderName();
+  const bool is_vp8 = strncmp(kVP8Name, codec_name, sizeof(kVP8Name) - 1) == 0;
 
   // Decode frames.
-  for (video->Begin(); video->cxdata(); video->Next()) {
+  for (video->Begin(); !::testing::Test::HasFailure() && video->cxdata();
+       video->Next()) {
     PreDecodeFrameHook(*video, decoder);
+
+    vpx_codec_stream_info_t stream_info;
+    stream_info.sz = sizeof(stream_info);
+    const vpx_codec_err_t res_peek = decoder->PeekStream(video->cxdata(),
+                                                         video->frame_size(),
+                                                         &stream_info);
+    if (is_vp8) {
+      /* Vp8's implementation of PeekStream returns an error if the frame you
+       * pass it is not a keyframe, so we only expect VPX_CODEC_OK on the first
+       * frame, which must be a keyframe. */
+      if (video->frame_number() == 0)
+        ASSERT_EQ(VPX_CODEC_OK, res_peek) << "Peek return failed: "
+            << vpx_codec_err_to_string(res_peek);
+    } else {
+      /* The Vp9 implementation of PeekStream returns an error only if the
+       * data passed to it isn't a valid Vp9 chunk. */
+      ASSERT_EQ(VPX_CODEC_OK, res_peek) << "Peek return failed: "
+          << vpx_codec_err_to_string(res_peek);
+    }
+
     vpx_codec_err_t res_dec = decoder->DecodeFrame(video->cxdata(),
                                                    video->frame_size());
-    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+    if (!HandleDecodeResult(res_dec, *video, decoder))
+      break;
 
     DxDataIterator dec_iter = decoder->GetDxData();
     const vpx_image_t *img = NULL;

test/decode_test_driver.h

@@ -49,8 +49,14 @@ class Decoder {
     vpx_codec_destroy(&decoder_);
   }
 
+  vpx_codec_err_t PeekStream(const uint8_t *cxdata, size_t size,
+                             vpx_codec_stream_info_t *stream_info);
+
   vpx_codec_err_t DecodeFrame(const uint8_t *cxdata, size_t size);
 
+  vpx_codec_err_t DecodeFrame(const uint8_t *cxdata, size_t size,
+                              void *user_priv);
+
   DxDataIterator GetDxData() {
     return DxDataIterator(&decoder_);
   }
@@ -85,6 +91,10 @@ class Decoder {
         &decoder_, cb_get, cb_release, user_priv);
   }
 
+  const char* GetDecoderName() {
+    return vpx_codec_iface_name(CodecInterface());
+  }
+
  protected:
   virtual vpx_codec_iface_t* CodecInterface() const = 0;
 
@@ -114,6 +124,14 @@ class DecoderTest {
   virtual void PreDecodeFrameHook(const CompressedVideoSource& video,
                                   Decoder *decoder) {}
 
+  // Hook to be called to handle decode result. Return true to continue.
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res_dec,
+                                  const CompressedVideoSource& /* video */,
+                                  Decoder *decoder) {
+    EXPECT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+    return VPX_CODEC_OK == res_dec;
+  }
+
   // Hook to be called on every decompressed frame.
   virtual void DecompressedFrameHook(const vpx_image_t& img,
                                      const unsigned int frame_number) {}

test/decode_to_md5.sh

@@ -34,7 +34,10 @@ decode_to_md5() {
   local expected_md5="$3"
   local output_file="${VPX_TEST_OUTPUT_DIR}/decode_to_md5_${codec}"
 
-  [ -x "${decoder}" ] || return 1
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${decoder}" "${input_file}" "${output_file}" ${devnull}
 

test/decode_with_drops.sh

@@ -34,7 +34,10 @@ decode_with_drops() {
   local output_file="${VPX_TEST_OUTPUT_DIR}/decode_with_drops_${codec}"
   local drop_mode="$3"
 
-  [ -x "${decoder}" ] || return 1
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${decoder}" "${input_file}" "${output_file}" "${drop_mode}" ${devnull}
 

test/fdct4x4_test.cc

@@ -376,4 +376,19 @@ INSTANTIATE_TEST_CASE_P(
         make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_sse2, 3)));
 #endif
 
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(
+    AVX2, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&vp9_fdct4x4_avx2,
+                   &vp9_idct4x4_16_add_c, 0)));
+INSTANTIATE_TEST_CASE_P(
+    AVX2, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht4x4_avx2, &vp9_iht4x4_16_add_c, 0),
+        make_tuple(&vp9_fht4x4_avx2, &vp9_iht4x4_16_add_c, 1),
+        make_tuple(&vp9_fht4x4_avx2, &vp9_iht4x4_16_add_c, 2),
+        make_tuple(&vp9_fht4x4_avx2, &vp9_iht4x4_16_add_c, 3)));
+#endif
+
 }  // namespace

test/fdct8x8_test.cc

@@ -367,4 +367,18 @@ INSTANTIATE_TEST_CASE_P(
     ::testing::Values(
         make_tuple(&vp9_fdct8x8_ssse3, &vp9_idct8x8_64_add_ssse3, 0)));
 #endif
+
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(
+    AVX2, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vp9_fdct8x8_avx2, &vp9_idct8x8_64_add_c, 0)));
+INSTANTIATE_TEST_CASE_P(
+    AVX2, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht8x8_avx2, &vp9_iht8x8_64_add_c, 0),
+        make_tuple(&vp9_fht8x8_avx2, &vp9_iht8x8_64_add_c, 1),
+        make_tuple(&vp9_fht8x8_avx2, &vp9_iht8x8_64_add_c, 2),
+        make_tuple(&vp9_fht8x8_avx2, &vp9_iht8x8_64_add_c, 3)));
+#endif
 }  // namespace

test/invalid_file_test.cc

@@ -0,0 +1,112 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+class InvalidFileTest
+    : public ::libvpx_test::DecoderTest,
+      public ::libvpx_test::CodecTestWithParam<const char*> {
+ protected:
+  InvalidFileTest() : DecoderTest(GET_PARAM(0)), res_file_(NULL) {}
+
+  virtual ~InvalidFileTest() {
+    if (res_file_ != NULL)
+      fclose(res_file_);
+  }
+
+  void OpenResFile(const std::string &res_file_name_) {
+    res_file_ = libvpx_test::OpenTestDataFile(res_file_name_);
+    ASSERT_TRUE(res_file_ != NULL) << "Result file open failed. Filename: "
+        << res_file_name_;
+  }
+
+  virtual bool HandleDecodeResult(
+      const vpx_codec_err_t res_dec,
+      const libvpx_test::CompressedVideoSource &video,
+      libvpx_test::Decoder *decoder) {
+    EXPECT_TRUE(res_file_ != NULL);
+    int expected_res_dec;
+
+    // Read integer result.
+    const int res = fscanf(res_file_, "%d", &expected_res_dec);
+    EXPECT_NE(res, EOF) << "Read result data failed";
+
+    // Check results match.
+    EXPECT_EQ(expected_res_dec, res_dec)
+        << "Results don't match: frame number = " << video.frame_number();
+
+    return !HasFailure();
+  }
+
+ private:
+  FILE *res_file_;
+};
+
+TEST_P(InvalidFileTest, ReturnCode) {
+  const std::string filename = GET_PARAM(1);
+  libvpx_test::CompressedVideoSource *video = NULL;
+
+  // Open compressed video file.
+  if (filename.substr(filename.length() - 3, 3) == "ivf") {
+    video = new libvpx_test::IVFVideoSource(filename);
+  } else if (filename.substr(filename.length() - 4, 4) == "webm") {
+#if CONFIG_WEBM_IO
+    video = new libvpx_test::WebMVideoSource(filename);
+#else
+    fprintf(stderr, "WebM IO is disabled, skipping test vector %s\n",
+            filename.c_str());
+    return;
+#endif
+  }
+  video->Init();
+
+  // Construct result file name. The file holds a list of expected integer
+  // results, one for each decoded frame.  Any result that doesn't match
+  // the files list will cause a test failure.
+  const std::string res_filename = filename + ".res";
+  OpenResFile(res_filename);
+
+  // Decode frame, and check the md5 matching.
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video));
+  delete video;
+}
+
+const char *const kVP9InvalidFileTests[] = {
+  "invalid-vp90-01.webm",
+  "invalid-vp90-02.webm",
+  "invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.ivf",
+  "invalid-vp90-03.webm",
+  "invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf",
+  "invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf",
+};
+
+#define NELEMENTS(x) static_cast<int>(sizeof(x) / sizeof(x[0]))
+
+VP9_INSTANTIATE_TEST_CASE(InvalidFileTest,
+                          ::testing::ValuesIn(kVP9InvalidFileTests,
+                                              kVP9InvalidFileTests +
+                                              NELEMENTS(kVP9InvalidFileTests)));
+
+}  // namespace

test/md5_helper.h

@@ -28,10 +28,11 @@ class MD5 {
       // plane, we never want to round down and thus skip a pixel so if
       // we are shifting by 1 (chroma_shift) we add 1 before doing the shift.
       // This works only for chroma_shift of 0 and 1.
+      const int bytes_per_sample = (img->fmt & VPX_IMG_FMT_HIGH) ? 2 : 1;
       const int h = plane ? (img->d_h + img->y_chroma_shift) >>
                     img->y_chroma_shift : img->d_h;
-      const int w = plane ? (img->d_w + img->x_chroma_shift) >>
+      const int w = (plane ? (img->d_w + img->x_chroma_shift) >>
-                    img->x_chroma_shift : img->d_w;
+                     img->x_chroma_shift : img->d_w) * bytes_per_sample;
 
       for (int y = 0; y < h; ++y) {
         MD5Update(&md5_, buf, w);

test/postproc.sh

@@ -32,7 +32,10 @@ postproc() {
   local codec="$2"
   local output_file="${VPX_TEST_OUTPUT_DIR}/postproc_${codec}.raw"
 
-  [ -x "${decoder}" ] || return 1
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${decoder}" "${input_file}" "${output_file}" ${devnull}
 

test/resize_util.sh

@@ -33,7 +33,10 @@ resize_util() {
 
   # resize_util is available only when CONFIG_SHARED is disabled.
   if [ -z "$(vpx_config_option_enabled CONFIG_SHARED)" ]; then
-    [ -x "${resizer}" ] || return 1
+    if [ ! -x "${resizer}" ]; then
+      elog "${resizer} does not exist or is not executable."
+      return 1
+    fi
 
     eval "${resizer}" "${YUV_RAW_INPUT}" \
         "${YUV_RAW_INPUT_WIDTH}x${YUV_RAW_INPUT_HEIGHT}" \

test/sad_test.cc

@@ -627,4 +627,24 @@ INSTANTIATE_TEST_CASE_P(SSE3, SADTest, ::testing::Values(
 #endif  // CONFIG_USE_X86INC
 #endif  // HAVE_SSSE3
 
+#if HAVE_AVX2
+#if CONFIG_VP9_ENCODER
+// TODO(jzern): these prototypes can be removed after the avx2 versions are
+// reenabled in vp9_rtcd_defs.pl.
+extern "C" {
+void vp9_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride,
+                          const uint8_t *const ref_ptr[], int ref_stride,
+                          unsigned int *sad_array);
+void vp9_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride,
+                          const uint8_t *const ref_ptr[], int ref_stride,
+                          unsigned int *sad_array);
+}
+const sad_n_by_n_by_4_fn_t sad_64x64x4d_avx2 = vp9_sad64x64x4d_avx2;
+const sad_n_by_n_by_4_fn_t sad_32x32x4d_avx2 = vp9_sad32x32x4d_avx2;
+INSTANTIATE_TEST_CASE_P(DISABLED_AVX2, SADx4Test, ::testing::Values(
+                        make_tuple(32, 32, sad_32x32x4d_avx2),
+                        make_tuple(64, 64, sad_64x64x4d_avx2)));
+#endif  // CONFIG_VP9_ENCODER
+#endif  // HAVE_AVX2
+
 }  // namespace

test/simple_decoder.sh

@@ -32,7 +32,10 @@ simple_decoder() {
   local codec="$2"
   local output_file="${VPX_TEST_OUTPUT_DIR}/simple_decoder_${codec}.raw"
 
-  [ -x "${decoder}" ] || return 1
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${decoder}" "${input_file}" "${output_file}" ${devnull}
 

test/simple_encoder.sh

@@ -29,7 +29,10 @@ simple_encoder() {
   local codec="$1"
   local output_file="${VPX_TEST_OUTPUT_DIR}/simple_encoder_${codec}.ivf"
 
-  [ -x "${encoder}" ] || return 1
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
       "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" 9999 \

test/svc_test.cc

@@ -31,7 +31,6 @@ class SvcTest : public ::testing::Test {
   SvcTest()
       : codec_iface_(0),
         test_file_name_("hantro_collage_w352h288.yuv"),
-        stats_file_name_("hantro_collage_w352h288.stat"),
         codec_initialized_(false),
         decoder_(0) {
     memset(&svc_, 0, sizeof(svc_));
@@ -42,7 +41,6 @@ class SvcTest : public ::testing::Test {
   virtual ~SvcTest() {}
 
   virtual void SetUp() {
-    svc_.encoding_mode = INTER_LAYER_PREDICTION_IP;
     svc_.log_level = SVC_LOG_DEBUG;
     svc_.log_print = 0;
 
@@ -74,7 +72,6 @@ class SvcTest : public ::testing::Test {
   struct vpx_codec_enc_cfg codec_enc_;
   vpx_codec_iface_t *codec_iface_;
   std::string test_file_name_;
-  std::string stats_file_name_;
   bool codec_initialized_;
   Decoder *decoder_;
 };
@@ -133,22 +130,13 @@ TEST_F(SvcTest, SetLayersOption) {
   EXPECT_EQ(3, svc_.spatial_layers);
 }
 
-TEST_F(SvcTest, SetEncodingMode) {
-  vpx_codec_err_t res = vpx_svc_set_options(&svc_, "encoding-mode=alt-ip");
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  codec_initialized_ = true;
-  EXPECT_EQ(ALT_INTER_LAYER_PREDICTION_IP, svc_.encoding_mode);
-}
-
 TEST_F(SvcTest, SetMultipleOptions) {
-  vpx_codec_err_t res = vpx_svc_set_options(&svc_, "layers=2 encoding-mode=ip");
+  vpx_codec_err_t res =
+      vpx_svc_set_options(&svc_, "layers=2 scale-factors=1/3,2/3");
   res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   EXPECT_EQ(VPX_CODEC_OK, res);
   codec_initialized_ = true;
   EXPECT_EQ(2, svc_.spatial_layers);
-  EXPECT_EQ(INTER_LAYER_PREDICTION_IP, svc_.encoding_mode);
 }
 
 TEST_F(SvcTest, SetScaleFactorsOption) {
@@ -179,48 +167,20 @@ TEST_F(SvcTest, SetQuantizersOption) {
   codec_initialized_ = true;
 }
 
-TEST_F(SvcTest, SetKeyFrameQuantizersOption) {
-  svc_.spatial_layers = 2;
-  vpx_codec_err_t res = vpx_svc_set_options(&svc_,
-                                       "quantizers-keyframe=not-quantizers");
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
-  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
-
-  vpx_svc_set_options(&svc_, "quantizers-keyframe=40,45");
-  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  codec_initialized_ = true;
-}
-
 TEST_F(SvcTest, SetQuantizers) {
-  vpx_codec_err_t res = vpx_svc_set_quantizers(NULL, "40,30", 0);
+  vpx_codec_err_t res = vpx_svc_set_quantizers(NULL, "40,30");
   EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
 
-  res = vpx_svc_set_quantizers(&svc_, NULL, 0);
+  res = vpx_svc_set_quantizers(&svc_, NULL);
   EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
 
   svc_.spatial_layers = 2;
-  res = vpx_svc_set_quantizers(&svc_, "40", 0);
+  res = vpx_svc_set_quantizers(&svc_, "40");
   EXPECT_EQ(VPX_CODEC_OK, res);
   res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
 
-  res = vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  res = vpx_svc_set_quantizers(&svc_, "40,30");
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
-  EXPECT_EQ(VPX_CODEC_OK, res);
-  codec_initialized_ = true;
-}
-
-TEST_F(SvcTest, SetKeyFrameQuantizers) {
-  vpx_codec_err_t res = vpx_svc_set_quantizers(NULL, "40,31", 1);
-  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
-
-  res = vpx_svc_set_quantizers(&svc_, NULL, 1);
-  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
-
-  res = vpx_svc_set_quantizers(&svc_, "40,30", 1);
   EXPECT_EQ(VPX_CODEC_OK, res);
   res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   EXPECT_EQ(VPX_CODEC_OK, res);
@@ -251,7 +211,7 @@ TEST_F(SvcTest, SetScaleFactors) {
 TEST_F(SvcTest, FirstFrameHasLayers) {
   svc_.spatial_layers = 2;
   vpx_svc_set_scale_factors(&svc_, "4/16,16/16");
-  vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  vpx_svc_set_quantizers(&svc_, "40,30");
 
   vpx_codec_err_t res =
       vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
@@ -267,9 +227,17 @@ TEST_F(SvcTest, FirstFrameHasLayers) {
                        video.duration(), VPX_DL_GOOD_QUALITY);
   EXPECT_EQ(VPX_CODEC_OK, res);
 
+  if (vpx_svc_get_frame_size(&svc_) == 0) {
+    // Flush encoder
+    res = vpx_svc_encode(&svc_, &codec_, NULL, 0,
+                         video.duration(), VPX_DL_GOOD_QUALITY);
+    EXPECT_EQ(VPX_CODEC_OK, res);
+  }
+
+  int frame_size = vpx_svc_get_frame_size(&svc_);
+  EXPECT_GT(frame_size, 0);
   const vpx_codec_err_t res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
+      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
-      vpx_svc_get_frame_size(&svc_));
 
   // this test fails with a decoder error
   ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
@@ -278,7 +246,10 @@ TEST_F(SvcTest, FirstFrameHasLayers) {
 TEST_F(SvcTest, EncodeThreeFrames) {
   svc_.spatial_layers = 2;
   vpx_svc_set_scale_factors(&svc_, "4/16,16/16");
-  vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  vpx_svc_set_quantizers(&svc_, "40,30");
+  int decoded_frames = 0;
+  vpx_codec_err_t res_dec;
+  int frame_size;
 
   vpx_codec_err_t res =
       vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
@@ -293,13 +264,14 @@ TEST_F(SvcTest, EncodeThreeFrames) {
   // This frame is a keyframe.
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
-  ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(1, vpx_svc_is_keyframe(&svc_));
 
-  vpx_codec_err_t res_dec = decoder_->DecodeFrame(
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
-      vpx_svc_get_frame_size(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
     ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
   // FRAME 1
   video.Next();
@@ -307,12 +279,14 @@ TEST_F(SvcTest, EncodeThreeFrames) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(0, vpx_svc_is_keyframe(&svc_));
 
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
     res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
-      vpx_svc_get_frame_size(&svc_));
     ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
   // FRAME 2
   video.Next();
@@ -320,18 +294,35 @@ TEST_F(SvcTest, EncodeThreeFrames) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(0, vpx_svc_is_keyframe(&svc_));
 
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
     res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
-      vpx_svc_get_frame_size(&svc_));
     ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
+
+  // Flush encoder
+  res = vpx_svc_encode(&svc_, &codec_, NULL, 0,
+                       video.duration(), VPX_DL_GOOD_QUALITY);
+  EXPECT_EQ(VPX_CODEC_OK, res);
+
+  while ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
+
+  EXPECT_EQ(decoded_frames, 3);
 }
 
 TEST_F(SvcTest, GetLayerResolution) {
   svc_.spatial_layers = 2;
   vpx_svc_set_scale_factors(&svc_, "4/16,8/16");
-  vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  vpx_svc_set_quantizers(&svc_, "40,30");
 
   vpx_codec_err_t res =
       vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
@@ -364,11 +355,13 @@ TEST_F(SvcTest, GetLayerResolution) {
   EXPECT_EQ(kHeight * 8 / 16, layer_height);
 }
 
-TEST_F(SvcTest, FirstPassEncode) {
+TEST_F(SvcTest, TwoPassEncode) {
+  // First pass encode
+  std::string stats_buf;
   svc_.spatial_layers = 2;
   codec_enc_.g_pass = VPX_RC_FIRST_PASS;
   vpx_svc_set_scale_factors(&svc_, "4/16,16/16");
-  vpx_svc_set_quantizers(&svc_, "40,30", 0);
+  vpx_svc_set_quantizers(&svc_, "40,30");
 
   vpx_codec_err_t res =
       vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
@@ -383,62 +376,61 @@ TEST_F(SvcTest, FirstPassEncode) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_GT(vpx_svc_get_rc_stats_buffer_size(&svc_), 0U);
+  size_t stats_size = vpx_svc_get_rc_stats_buffer_size(&svc_);
+  EXPECT_GT(stats_size, 0U);
+  const char *stats_data = vpx_svc_get_rc_stats_buffer(&svc_);
+  ASSERT_TRUE(stats_data != NULL);
+  stats_buf.append(stats_data, stats_size);
 
   // FRAME 1
   video.Next();
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
-  ASSERT_EQ(VPX_CODEC_OK, res);
+  stats_size = vpx_svc_get_rc_stats_buffer_size(&svc_);
-  EXPECT_GT(vpx_svc_get_rc_stats_buffer_size(&svc_), 0U);
+  EXPECT_GT(stats_size, 0U);
+  stats_data = vpx_svc_get_rc_stats_buffer(&svc_);
+  ASSERT_TRUE(stats_data != NULL);
+  stats_buf.append(stats_data, stats_size);
 
   // Flush encoder and test EOS packet
   res = vpx_svc_encode(&svc_, &codec_, NULL, video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
-  ASSERT_EQ(VPX_CODEC_OK, res);
+  stats_size = vpx_svc_get_rc_stats_buffer_size(&svc_);
-  EXPECT_GT(vpx_svc_get_rc_stats_buffer_size(&svc_), 0U);
+  EXPECT_GT(stats_size, 0U);
-}
+  stats_data = vpx_svc_get_rc_stats_buffer(&svc_);
+  ASSERT_TRUE(stats_data != NULL);
+  stats_buf.append(stats_data, stats_size);
 
-TEST_F(SvcTest, SecondPassEncode) {
+  // Tear down encoder
-  svc_.spatial_layers = 2;
+  vpx_svc_release(&svc_);
+  vpx_codec_destroy(&codec_);
+
+  // Second pass encode
+  int decoded_frames = 0;
+  vpx_codec_err_t res_dec;
+  int frame_size;
   codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  codec_enc_.rc_twopass_stats_in.buf = &stats_buf[0];
+  codec_enc_.rc_twopass_stats_in.sz = stats_buf.size();
 
-  FILE *const stats_file = libvpx_test::OpenTestDataFile(stats_file_name_);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
-  ASSERT_TRUE(stats_file != NULL) << "Stats file open failed. Filename: "
-      << stats_file;
-
-  struct vpx_fixed_buf stats_buf;
-  fseek(stats_file, 0, SEEK_END);
-  stats_buf.sz = static_cast<size_t>(ftell(stats_file));
-  fseek(stats_file, 0, SEEK_SET);
-
-  stats_buf.buf = malloc(stats_buf.sz);
-  ASSERT_TRUE(stats_buf.buf != NULL);
-  const size_t bytes_read = fread(stats_buf.buf, 1, stats_buf.sz, stats_file);
-  ASSERT_EQ(bytes_read, stats_buf.sz);
-  fclose(stats_file);
-  codec_enc_.rc_twopass_stats_in = stats_buf;
-
-  vpx_codec_err_t res =
-      vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
   ASSERT_EQ(VPX_CODEC_OK, res);
   codec_initialized_ = true;
 
-  libvpx_test::I420VideoSource video(test_file_name_, kWidth, kHeight,
-                                     codec_enc_.g_timebase.den,
-                                     codec_enc_.g_timebase.num, 0, 30);
   // FRAME 0
   video.Begin();
   // This frame is a keyframe.
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(1, vpx_svc_is_keyframe(&svc_));
 
-  vpx_codec_err_t res_dec = decoder_->DecodeFrame(
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
-      vpx_svc_get_frame_size(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
     ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
   // FRAME 1
   video.Next();
@@ -446,12 +438,14 @@ TEST_F(SvcTest, SecondPassEncode) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(0, vpx_svc_is_keyframe(&svc_));
 
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
     res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
-      vpx_svc_get_frame_size(&svc_));
     ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
   // FRAME 2
   video.Next();
@@ -459,14 +453,29 @@ TEST_F(SvcTest, SecondPassEncode) {
   res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
                        video.duration(), VPX_DL_GOOD_QUALITY);
   ASSERT_EQ(VPX_CODEC_OK, res);
-  EXPECT_EQ(0, vpx_svc_is_keyframe(&svc_));
 
+  if ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
     res_dec = decoder_->DecodeFrame(
-      static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)),
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
-      vpx_svc_get_frame_size(&svc_));
     ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
 
-  free(stats_buf.buf);
+  // Flush encoder
+  res = vpx_svc_encode(&svc_, &codec_, NULL, 0,
+                       video.duration(), VPX_DL_GOOD_QUALITY);
+  EXPECT_EQ(VPX_CODEC_OK, res);
+
+  while ((frame_size = vpx_svc_get_frame_size(&svc_)) > 0) {
+    EXPECT_EQ((decoded_frames == 0), vpx_svc_is_keyframe(&svc_));
+    res_dec = decoder_->DecodeFrame(
+        static_cast<const uint8_t *>(vpx_svc_get_buffer(&svc_)), frame_size);
+    ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+    ++decoded_frames;
+  }
+
+  EXPECT_EQ(decoded_frames, 3);
 }
 
 }  // namespace

test/test-data.sha1

@@ -1,6 +1,20 @@
 d5dfb0151c9051f8c85999255645d7a23916d3c0  hantro_collage_w352h288.yuv
-998cec53307c94aa5835aaf8d5731f6a3c7c2e5a  hantro_collage_w352h288.stat
 b87815bf86020c592ccc7a846ba2e28ec8043902  hantro_odd.yuv
+fe346136b9b8c1e6f6084cc106485706915795e4  invalid-vp90-01.webm
+25751f5d3b05ff03f0719ad42cd625348eb8961e  invalid-vp90-01.webm.res
+d78e2fceba5ac942246503ec8366f879c4775ca5  invalid-vp90-02.webm
+2dadee5306245fa5eeb0f99652d0e17afbcba96d  invalid-vp90-02.webm.res
+df1a1453feb3c00d7d89746c7003b4163523bff3  invalid-vp90-03.webm
+8fe6fd82bf537340f586f97a7ae31fb37ccda302  invalid-vp90-03.webm.res
+a432f96ff0a787268e2f94a8092ab161a18d1b06  park_joy_90p_10_420.y4m
+0b194cc312c3a2e84d156a221b0a5eb615dfddc5  park_joy_90p_10_422.y4m
+ff0e0a21dc2adc95b8c1b37902713700655ced17  park_joy_90p_10_444.y4m
+614c32ae1eca391e867c70d19974f0d62664dd99  park_joy_90p_12_420.y4m
+c92825f1ea25c5c37855083a69faac6ac4641a9e  park_joy_90p_12_422.y4m
+b592189b885b6cc85db55cc98512a197d73d3b34  park_joy_90p_12_444.y4m
+4e0eb61e76f0684188d9bc9f3ce61f6b6b77bb2c  park_joy_90p_8_420.y4m
+7a193ff7dfeb96ba5f82b2afd7afa9e1fe83d947  park_joy_90p_8_422.y4m
+bdb7856e6bc93599bdda05c2e773a9f22b6c6d03  park_joy_90p_8_444.y4m
 b1f1c3ec79114b9a0651af24ce634afb44a9a419  rush_hour_444.y4m
 5184c46ddca8b1fadd16742e8500115bc8f749da  vp80-00-comprehensive-001.ivf
 65bf1bbbced81b97bd030f376d1b7f61a224793f  vp80-00-comprehensive-002.ivf
@@ -530,8 +544,6 @@ b6524e4084d15b5d0caaa3d3d1368db30cbee69c  vp90-2-03-deltaq.webm
 7f6d8879336239a43dbb6c9f13178cb11cf7ed09  vp90-2-05-resize.ivf.md5
 bf61ddc1f716eba58d4c9837d4e91031d9ce4ffe  vp90-2-06-bilinear.webm
 f6235f937552e11d8eb331ec55da6b3aa596b9ac  vp90-2-06-bilinear.webm.md5
-495256cfd123fe777b2c0406862ed8468a1f4677  vp91-2-04-yv444.webm
-65e3a7ffef61ab340d9140f335ecc49125970c2c  vp91-2-04-yv444.webm.md5
 0c83a1e414fde3bccd6dc451bbaee68e59974c76  vp90-2-07-frame_parallel.webm
 e5c2c9fb383e5bf3b563480adaeba5b7e3475ecd  vp90-2-07-frame_parallel.webm.md5
 086c7edcffd699ae7d99d710fd7e53b18910ca5b  vp90-2-08-tile_1x2_frame_parallel.webm
@@ -577,6 +589,8 @@ d48c5db1b0f8e60521a7c749696b8067886033a3  vp90-2-09-aq2.webm
 54638c38009198c38c8f3b25c182b709b6c1fd2e  vp90-2-09-lf_deltas.webm.md5
 510d95f3beb3b51c572611fdaeeece12277dac30  vp90-2-10-show-existing-frame.webm
 14d631096f4bfa2d71f7f739aec1448fb3c33bad  vp90-2-10-show-existing-frame.webm.md5
+d2feea7728e8d2c615981d0f47427a4a5a45d881  vp90-2-10-show-existing-frame2.webm
+5f7c7811baa3e4f03be1dd78c33971b727846821  vp90-2-10-show-existing-frame2.webm.md5
 b4318e75f73a6a08992c7326de2fb589c2a794c7  vp90-2-11-size-351x287.webm
 b3c48382cf7d0454e83a02497c229d27720f9e20  vp90-2-11-size-351x287.webm.md5
 8e0096475ea2535bac71d3e2fc09e0c451c444df  vp90-2-11-size-351x288.webm
@@ -639,4 +653,11 @@ e615575ded499ea1d992f3b38e3baa434509cdcd  vp90-2-15-segkey.webm
 e3ab35d4316c5e81325c50f5236ceca4bc0d35df  vp90-2-15-segkey.webm.md5
 9b7ca2cac09d34c4a5d296c1900f93b1e2f69d0d  vp90-2-15-segkey_adpq.webm
 8f46ba5f785d0c2170591a153e0d0d146a7c8090  vp90-2-15-segkey_adpq.webm.md5
-
+0321d507ce62dedc8a51b4e9011f7a19aed9c3dc  vp91-2-04-yuv444.webm
+367e423dd41fdb49aa028574a2cfec5c2f325c5c  vp91-2-04-yuv444.webm.md5
+76024eb753cdac6a5e5703aaea189d35c3c30ac7  invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.ivf
+d3964f9dad9f60363c81b688324d95b4ec7c8038  invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.ivf.res
+83f50908c8dc0ef8760595447a2ff7727489542e  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf
+456d1493e52d32a5c30edf44a27debc1fa6b253a  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf.res
+c123d1f9f02fb4143abb5e271916e3a3080de8f6  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf
+456d1493e52d32a5c30edf44a27debc1fa6b253a  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf.res

test/test.mk

@@ -15,7 +15,7 @@ LIBVPX_TEST_SRCS-yes += video_source.h
 ##
 ## Black box tests only use the public API.
 ##
-LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../md5_utils.h ../md5_utils.c
+LIBVPX_TEST_SRCS-yes                   += ../md5_utils.h ../md5_utils.c
 LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ivf_video_source.h
 LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += ../y4minput.h ../y4minput.c
 LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += aq_segment_test.cc
@@ -30,6 +30,7 @@ LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += cq_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += keyframe_test.cc
 
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += external_frame_buffer_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += user_priv_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += active_map_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += borders_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += cpu_speed_test.cc
@@ -41,6 +42,9 @@ LIBVPX_TEST_SRCS-yes                   += decode_test_driver.h
 LIBVPX_TEST_SRCS-yes                   += encode_test_driver.cc
 LIBVPX_TEST_SRCS-yes                   += encode_test_driver.h
 
+## Y4m parsing.
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += y4m_test.cc ../y4menc.c ../y4menc.h
+
 ## WebM Parsing
 ifeq ($(CONFIG_WEBM_IO), yes)
 LIBWEBM_PARSER_SRCS                    += ../third_party/libwebm/mkvparser.cpp
@@ -54,6 +58,7 @@ LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../webmdec.h
 LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += webm_video_source.h
 endif
 
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += invalid_file_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += test_vector_test.cc
 
 # Currently we only support decoder perf tests for vp9. Also they read from WebM
@@ -131,9 +136,20 @@ endif # CONFIG_SHARED
 ## TEST DATA
 ##
 LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += hantro_collage_w352h288.yuv
-LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += hantro_collage_w352h288.stat
 LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += hantro_odd.yuv
+
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_444.y4m
+
 LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += rush_hour_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += screendata.y4m
 
 LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-001.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-001.ivf.md5
@@ -691,6 +707,8 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-subpixel-00.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-subpixel-00.ivf.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame2.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x287.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x287.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x288.webm
@@ -705,8 +723,6 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_3.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_3.ivf.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-13-largescaling.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-13-largescaling.webm.md5
-LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yv444.webm
-LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yv444.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-4.webm
@@ -755,6 +771,22 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey_adpq.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey_adpq.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv444.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv444.webm.md5
+
+# Invalid files for testing libvpx error checking.
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-01.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-01.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-02.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-02.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf.res
 
 ifeq ($(CONFIG_DECODE_PERF_TESTS),yes)
 # BBB VP9 streams

test/test_vectors.cc

@@ -161,10 +161,11 @@ const char *const kVP9TestVectors[] = {
   "vp90-2-08-tile-4x1.webm", "vp90-2-09-subpixel-00.ivf",
   "vp90-2-02-size-lf-1920x1080.webm", "vp90-2-09-aq2.webm",
   "vp90-2-09-lf_deltas.webm", "vp90-2-10-show-existing-frame.webm",
+  "vp90-2-10-show-existing-frame2.webm",
   "vp90-2-11-size-351x287.webm", "vp90-2-11-size-351x288.webm",
   "vp90-2-11-size-352x287.webm", "vp90-2-12-droppable_1.ivf",
   "vp90-2-12-droppable_2.ivf", "vp90-2-12-droppable_3.ivf",
-  "vp90-2-13-largescaling.webm", "vp91-2-04-yv444.webm",
+  "vp90-2-13-largescaling.webm",
   "vp90-2-14-resize-fp-tiles-1-16.webm",
   "vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm",
   "vp90-2-14-resize-fp-tiles-1-2.webm", "vp90-2-14-resize-fp-tiles-1-4.webm",
@@ -178,7 +179,8 @@ const char *const kVP9TestVectors[] = {
   "vp90-2-14-resize-fp-tiles-4-2.webm", "vp90-2-14-resize-fp-tiles-4-8.webm",
   "vp90-2-14-resize-fp-tiles-8-16.webm", "vp90-2-14-resize-fp-tiles-8-1.webm",
   "vp90-2-14-resize-fp-tiles-8-2.webm", "vp90-2-14-resize-fp-tiles-8-4.webm",
-  "vp90-2-15-segkey.webm", "vp90-2-15-segkey_adpq.webm"
+  "vp90-2-15-segkey.webm", "vp90-2-15-segkey_adpq.webm",
+  "vp91-2-04-yuv444.webm",
 };
 const int kNumVP9TestVectors = NELEMENTS(kVP9TestVectors);
 #endif  // CONFIG_VP9_DECODER

test/tools_common.sh

@@ -17,6 +17,10 @@ VPX_TEST_TOOLS_COMMON_SH=included
 set -e
 devnull='> /dev/null 2>&1'
 
+elog() {
+  echo "$@" 1>&2
+}
+
 vlog() {
   if [ "${VPX_TEST_VERBOSE_OUTPUT}" = "yes" ]; then
     echo "$@"
@@ -456,10 +460,19 @@ vlog "$(basename "${0%.*}") test configuration:
   LIBVPX_BIN_PATH=${LIBVPX_BIN_PATH}
   LIBVPX_CONFIG_PATH=${LIBVPX_CONFIG_PATH}
   LIBVPX_TEST_DATA_PATH=${LIBVPX_TEST_DATA_PATH}
-  VPX_TEST_OUTPUT_DIR=${VPX_TEST_OUTPUT_DIR}
+  VP8_IVF_FILE=${VP8_IVF_FILE}
-  VPX_TEST_VERBOSE_OUTPUT=${VPX_TEST_VERBOSE_OUTPUT}
+  VP9_IVF_FILE=${VP9_IVF_FILE}
+  VP9_WEBM_FILE=${VP9_WEBM_FILE}
+  VPX_TEST_EXE_SUFFIX=${VPX_TEST_EXE_SUFFIX}
   VPX_TEST_FILTER=${VPX_TEST_FILTER}
+  VPX_TEST_OUTPUT_DIR=${VPX_TEST_OUTPUT_DIR}
+  VPX_TEST_RAND=${VPX_TEST_RAND}
   VPX_TEST_RUN_DISABLED_TESTS=${VPX_TEST_RUN_DISABLED_TESTS}
-  VPX_TEST_SHOW_PROGRAM_OUTPUT=${VPX_TEST_SHOW_PROGRAM_OUTPUT}"
+  VPX_TEST_SHOW_PROGRAM_OUTPUT=${VPX_TEST_SHOW_PROGRAM_OUTPUT}
+  VPX_TEST_TEMP_ROOT=${VPX_TEST_TEMP_ROOT}
+  VPX_TEST_VERBOSE_OUTPUT=${VPX_TEST_VERBOSE_OUTPUT}
+  YUV_RAW_INPUT=${YUV_RAW_INPUT}
+  YUV_RAW_INPUT_WIDTH=${YUV_RAW_INPUT_WIDTH}
+  YUV_RAW_INPUT_HEIGHT=${YUV_RAW_INPUT_HEIGHT}"
 
 fi  # End $VPX_TEST_TOOLS_COMMON_SH pseudo include guard.

test/twopass_encoder.sh

@@ -29,7 +29,10 @@ twopass_encoder() {
   local codec="$1"
   local output_file="${VPX_TEST_OUTPUT_DIR}/twopass_encoder_${codec}.ivf"
 
-  [ -x "${encoder}" ] || return 1
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
       "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" \

test/user_priv_test.cc

@@ -0,0 +1,100 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/acm_random.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+#include "vpx/vp8.h"
+
+namespace {
+
+using std::string;
+using libvpx_test::ACMRandom;
+
+#if CONFIG_WEBM_IO
+
+void CheckUserPrivateData(void *user_priv, int *target) {
+  // actual pointer value should be the same as expected.
+  EXPECT_EQ(reinterpret_cast<void *>(target), user_priv) <<
+      "user_priv pointer value does not match.";
+}
+
+// Decodes |filename|. Passes in user_priv data when calling DecodeFrame and
+// compares the user_priv from return img with the original user_priv to see if
+// they match. Both the pointer values and the values inside the addresses
+// should match.
+string DecodeFile(const string &filename) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  libvpx_test::WebMVideoSource video(filename);
+  video.Init();
+
+  vpx_codec_dec_cfg_t cfg = {0};
+  libvpx_test::VP9Decoder decoder(cfg, 0);
+
+  libvpx_test::MD5 md5;
+  int frame_num = 0;
+  for (video.Begin(); !::testing::Test::HasFailure() && video.cxdata();
+       video.Next()) {
+    void *user_priv = reinterpret_cast<void *>(&frame_num);
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size(),
+                            (frame_num == 0) ? NULL : user_priv);
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img = NULL;
+
+    // Get decompressed data.
+    while ((img = dec_iter.Next())) {
+      if (frame_num == 0) {
+        CheckUserPrivateData(img->user_priv, NULL);
+      } else {
+        CheckUserPrivateData(img->user_priv, &frame_num);
+
+        // Also test ctrl_get_reference api.
+        struct vp9_ref_frame ref;
+        // Randomly fetch a reference frame.
+        ref.idx = rnd.Rand8() % 3;
+        decoder.Control(VP9_GET_REFERENCE, &ref);
+
+        CheckUserPrivateData(ref.img.user_priv, NULL);
+      }
+      md5.Add(img);
+    }
+
+    frame_num++;
+  }
+  return string(md5.Get());
+}
+
+TEST(UserPrivTest, VideoDecode) {
+  // no tiles or frame parallel; this exercises the decoding to test the
+  // user_priv.
+  EXPECT_STREQ("b35a1b707b28e82be025d960aba039bc",
+               DecodeFile("vp90-2-03-size-226x226.webm").c_str());
+}
+
+#endif  // CONFIG_WEBM_IO
+
+}  // namespace

test/variance_test.cc

@@ -702,6 +702,57 @@ INSTANTIATE_TEST_CASE_P(
                       make_tuple(6, 6, subpel_avg_variance64x64_ssse3)));
 #endif
 #endif
+
+#if HAVE_AVX2
+// TODO(jzern): these prototypes can be removed after the avx2 versions are
+// reenabled in vp9_rtcd_defs.pl.
+extern "C" {
+unsigned int vp9_sub_pixel_variance32x32_avx2(
+    const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset,
+    const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x64_avx2(
+    const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset,
+    const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_avg_variance32x32_avx2(
+    const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset,
+    const uint8_t *ref_ptr, int ref_stride, unsigned int *sse,
+    const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x64_avx2(
+    const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset,
+    const uint8_t *ref_ptr, int ref_stride, unsigned int *sse,
+    const uint8_t *second_pred);
+}
+const vp9_variance_fn_t variance16x16_avx2 = vp9_variance16x16_avx2;
+const vp9_variance_fn_t variance32x16_avx2 = vp9_variance32x16_avx2;
+const vp9_variance_fn_t variance32x32_avx2 = vp9_variance32x32_avx2;
+const vp9_variance_fn_t variance64x32_avx2 = vp9_variance64x32_avx2;
+const vp9_variance_fn_t variance64x64_avx2 = vp9_variance64x64_avx2;
+INSTANTIATE_TEST_CASE_P(
+    AVX2, VP9VarianceTest,
+    ::testing::Values(make_tuple(4, 4, variance16x16_avx2),
+                      make_tuple(5, 4, variance32x16_avx2),
+                      make_tuple(5, 5, variance32x32_avx2),
+                      make_tuple(6, 5, variance64x32_avx2),
+                      make_tuple(6, 6, variance64x64_avx2)));
+
+const vp9_subpixvariance_fn_t subpel_variance32x32_avx2 =
+    vp9_sub_pixel_variance32x32_avx2;
+const vp9_subpixvariance_fn_t subpel_variance64x64_avx2 =
+    vp9_sub_pixel_variance64x64_avx2;
+INSTANTIATE_TEST_CASE_P(
+    DISABLED_AVX2, VP9SubpelVarianceTest,
+    ::testing::Values(make_tuple(5, 5, subpel_variance32x32_avx2),
+                      make_tuple(6, 6, subpel_variance64x64_avx2)));
+
+const vp9_subp_avg_variance_fn_t subpel_avg_variance32x32_avx2 =
+    vp9_sub_pixel_avg_variance32x32_avx2;
+const vp9_subp_avg_variance_fn_t subpel_avg_variance64x64_avx2 =
+    vp9_sub_pixel_avg_variance64x64_avx2;
+INSTANTIATE_TEST_CASE_P(
+    DISABLED_AVX2, VP9SubpelAvgVarianceTest,
+    ::testing::Values(make_tuple(5, 5, subpel_avg_variance32x32_avx2),
+                      make_tuple(6, 6, subpel_avg_variance64x64_avx2)));
+#endif  // HAVE_AVX2
 #endif  // CONFIG_VP9_ENCODER
 
 }  // namespace vp9

test/video_source.h

@@ -50,6 +50,15 @@ static FILE *OpenTestDataFile(const std::string& file_name) {
   return fopen(path_to_source.c_str(), "rb");
 }
 
+static FILE *OpenTestOutFile(const std::string& file_name) {
+  const std::string path_to_source = GetDataPath() + "/" + file_name;
+  return fopen(path_to_source.c_str(), "wb");
+}
+
+static FILE *OpenTempOutFile() {
+  return tmpfile();
+}
+
 // Abstract base class for test video sources, which provide a stream of
 // vpx_image_t images with associated timestamps and duration.
 class VideoSource {

test/vp8cx_set_ref.sh

@@ -34,7 +34,10 @@ vpx_set_ref() {
   local output_file="${VPX_TEST_OUTPUT_DIR}/vp8cx_set_ref_${codec}.ivf"
   local ref_frame_num=90
 
-  [ -x "${encoder}" ] || return 1
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${encoder}" "${YUV_RAW_INPUT_WIDTH}" "${YUV_RAW_INPUT_HEIGHT}" \
       "${YUV_RAW_INPUT}" "${output_file}" "${ref_frame_num}" \

test/vp9_lossless_test.cc

@@ -36,6 +36,17 @@ class LosslessTestLarge : public ::libvpx_test::EncoderTest,
     SetMode(encoding_mode_);
   }
 
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      // Only call Control if quantizer > 0 to verify that using quantizer
+      // alone will activate lossless
+      if (cfg_.rc_max_quantizer > 0 || cfg_.rc_min_quantizer > 0) {
+        encoder->Control(VP9E_SET_LOSSLESS, 1);
+      }
+    }
+  }
+
   virtual void BeginPassHook(unsigned int /*pass*/) {
     psnr_ = kMaxPsnr;
     nframes_ = 0;
@@ -91,5 +102,24 @@ TEST_P(LosslessTestLarge, TestLossLessEncoding444) {
   EXPECT_GE(psnr_lossless, kMaxPsnr);
 }
 
+TEST_P(LosslessTestLarge, TestLossLessEncodingCtrl) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 25;
+  // Intentionally set Q > 0, to make sure control can be used to activate
+  // lossless
+  cfg_.rc_min_quantizer = 10;
+  cfg_.rc_max_quantizer = 20;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 10);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_lossless = GetMinPsnr();
+  EXPECT_GE(psnr_lossless, kMaxPsnr);
+}
+
 VP9_INSTANTIATE_TEST_CASE(LosslessTestLarge, ALL_TEST_MODES);
 }  // namespace

test/vp9_spatial_svc_encoder.sh

@@ -34,7 +34,10 @@ vp9_spatial_svc_encoder() {
 
   shift
 
-  [ -x "${encoder}" ] || return 1
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${encoder}" -w "${YUV_RAW_INPUT_WIDTH}" -h "${YUV_RAW_INPUT_HEIGHT}" \
       -k "${max_kf}" -f "${frames_to_encode}" "$@" "${YUV_RAW_INPUT}" \

test/vp9_thread_test.cc

@@ -28,11 +28,11 @@ class VP9WorkerThreadTest : public ::testing::TestWithParam<bool> {
  protected:
   virtual ~VP9WorkerThreadTest() {}
   virtual void SetUp() {
-    vp9_worker_init(&worker_);
+    vp9_get_worker_interface()->init(&worker_);
   }
 
   virtual void TearDown() {
-    vp9_worker_end(&worker_);
+    vp9_get_worker_interface()->end(&worker_);
   }
 
   VP9Worker worker_;
@@ -45,10 +45,11 @@ int ThreadHook(void* data, void* return_value) {
 }
 
 TEST_P(VP9WorkerThreadTest, HookSuccess) {
-  EXPECT_NE(vp9_worker_sync(&worker_), 0);  // should be a no-op.
+  // should be a no-op.
+  EXPECT_NE(vp9_get_worker_interface()->sync(&worker_), 0);
 
   for (int i = 0; i < 2; ++i) {
-    EXPECT_NE(vp9_worker_reset(&worker_), 0);
+    EXPECT_NE(vp9_get_worker_interface()->reset(&worker_), 0);
 
     int hook_data = 0;
     int return_value = 1;  // return successfully from the hook
@@ -58,20 +59,21 @@ TEST_P(VP9WorkerThreadTest, HookSuccess) {
 
     const bool synchronous = GetParam();
     if (synchronous) {
-      vp9_worker_execute(&worker_);
+      vp9_get_worker_interface()->execute(&worker_);
     } else {
-      vp9_worker_launch(&worker_);
+      vp9_get_worker_interface()->launch(&worker_);
     }
-    EXPECT_NE(vp9_worker_sync(&worker_), 0);
+    EXPECT_NE(vp9_get_worker_interface()->sync(&worker_), 0);
     EXPECT_FALSE(worker_.had_error);
     EXPECT_EQ(5, hook_data);
 
-    EXPECT_NE(vp9_worker_sync(&worker_), 0);  // should be a no-op.
+    // should be a no-op.
+    EXPECT_NE(vp9_get_worker_interface()->sync(&worker_), 0);
   }
 }
 
 TEST_P(VP9WorkerThreadTest, HookFailure) {
-  EXPECT_NE(vp9_worker_reset(&worker_), 0);
+  EXPECT_NE(vp9_get_worker_interface()->reset(&worker_), 0);
 
   int hook_data = 0;
   int return_value = 0;  // return failure from the hook
@@ -81,26 +83,49 @@ TEST_P(VP9WorkerThreadTest, HookFailure) {
 
   const bool synchronous = GetParam();
   if (synchronous) {
-    vp9_worker_execute(&worker_);
+    vp9_get_worker_interface()->execute(&worker_);
   } else {
-    vp9_worker_launch(&worker_);
+    vp9_get_worker_interface()->launch(&worker_);
   }
-  EXPECT_FALSE(vp9_worker_sync(&worker_));
+  EXPECT_FALSE(vp9_get_worker_interface()->sync(&worker_));
   EXPECT_EQ(1, worker_.had_error);
 
   // Ensure _reset() clears the error and _launch() can be called again.
   return_value = 1;
-  EXPECT_NE(vp9_worker_reset(&worker_), 0);
+  EXPECT_NE(vp9_get_worker_interface()->reset(&worker_), 0);
   EXPECT_FALSE(worker_.had_error);
-  vp9_worker_launch(&worker_);
+  vp9_get_worker_interface()->launch(&worker_);
-  EXPECT_NE(vp9_worker_sync(&worker_), 0);
+  EXPECT_NE(vp9_get_worker_interface()->sync(&worker_), 0);
   EXPECT_FALSE(worker_.had_error);
 }
 
+TEST(VP9WorkerThreadTest, TestInterfaceAPI) {
+  EXPECT_EQ(0, vp9_set_worker_interface(NULL));
+  EXPECT_TRUE(vp9_get_worker_interface() != NULL);
+  for (int i = 0; i < 6; ++i) {
+    VP9WorkerInterface winterface = *vp9_get_worker_interface();
+    switch (i) {
+      default:
+      case 0: winterface.init = NULL; break;
+      case 1: winterface.reset = NULL; break;
+      case 2: winterface.sync = NULL; break;
+      case 3: winterface.launch = NULL; break;
+      case 4: winterface.execute = NULL; break;
+      case 5: winterface.end = NULL; break;
+    }
+    EXPECT_EQ(0, vp9_set_worker_interface(&winterface));
+  }
+}
+
 // -----------------------------------------------------------------------------
 // Multi-threaded decode tests
 
 #if CONFIG_WEBM_IO
+struct FileList {
+  const char *name;
+  const char *expected_md5;
+};
+
 // Decodes |filename| with |num_threads|. Returns the md5 of the decoded frames.
 string DecodeFile(const string& filename, int num_threads) {
   libvpx_test::WebMVideoSource video(filename);
@@ -130,39 +155,77 @@ string DecodeFile(const string& filename, int num_threads) {
   return string(md5.Get());
 }
 
-TEST(VP9DecodeMTTest, MTDecode) {
+void DecodeFiles(const FileList files[]) {
-  // no tiles or frame parallel; this exercises loop filter threading.
+  for (const FileList *iter = files; iter->name != NULL; ++iter) {
-  EXPECT_STREQ("b35a1b707b28e82be025d960aba039bc",
+    SCOPED_TRACE(iter->name);
-               DecodeFile("vp90-2-03-size-226x226.webm", 2).c_str());
+    for (int t = 2; t <= 8; ++t) {
+      EXPECT_EQ(iter->expected_md5, DecodeFile(iter->name, t))
+          << "threads = " << t;
+    }
+  }
 }
 
-TEST(VP9DecodeMTTest, MTDecode2) {
+// Trivial serialized thread worker interface implementation.
-  static const struct {
+// Note any worker that requires synchronization between other workers will
-    const char *name;
+// hang.
-    const char *expected_md5;
+namespace impl {
-  } files[] = {
+
+void Init(VP9Worker *const worker) { memset(worker, 0, sizeof(*worker)); }
+int Reset(VP9Worker *const /*worker*/) { return 1; }
+int Sync(VP9Worker *const worker) { return !worker->had_error; }
+
+void Execute(VP9Worker *const worker) {
+  worker->had_error |= worker->hook(worker->data1, worker->data2);
+}
+
+void Launch(VP9Worker *const worker) { Execute(worker); }
+void End(VP9Worker *const /*worker*/) {}
+
+}  // namespace impl
+
+TEST(VP9WorkerThreadTest, TestSerialInterface) {
+  static const VP9WorkerInterface serial_interface = {
+    impl::Init, impl::Reset, impl::Sync, impl::Launch, impl::Execute, impl::End
+  };
+  // TODO(jzern): Avoid using a file that will use the row-based thread
+  // loopfilter, with the simple serialized implementation it will hang. This is
+  // due to its expectation that rows will be run in parallel as they wait on
+  // progress in the row above before proceeding.
+  static const char expected_md5[] = "b35a1b707b28e82be025d960aba039bc";
+  static const char filename[] = "vp90-2-03-size-226x226.webm";
+  VP9WorkerInterface default_interface = *vp9_get_worker_interface();
+
+  EXPECT_NE(vp9_set_worker_interface(&serial_interface), 0);
+  EXPECT_EQ(expected_md5, DecodeFile(filename, 2));
+
+  // Reset the interface.
+  EXPECT_NE(vp9_set_worker_interface(&default_interface), 0);
+  EXPECT_EQ(expected_md5, DecodeFile(filename, 2));
+}
+
+TEST(VP9DecodeMultiThreadedTest, Decode) {
+  // no tiles or frame parallel; this exercises loop filter threading.
+  EXPECT_EQ("b35a1b707b28e82be025d960aba039bc",
+            DecodeFile("vp90-2-03-size-226x226.webm", 2));
+}
+
+TEST(VP9DecodeMultiThreadedTest, Decode2) {
+  static const FileList files[] = {
     { "vp90-2-08-tile_1x2_frame_parallel.webm",
       "68ede6abd66bae0a2edf2eb9232241b6" },
     { "vp90-2-08-tile_1x4_frame_parallel.webm",
       "368ebc6ebf3a5e478d85b2c3149b2848" },
     { "vp90-2-08-tile_1x8_frame_parallel.webm",
       "17e439da2388aff3a0f69cb22579c6c1" },
+    { NULL, NULL }
   };
 
-  for (int i = 0; i < static_cast<int>(sizeof(files) / sizeof(files[0])); ++i) {
+  DecodeFiles(files);
-    for (int t = 2; t <= 8; ++t) {
-      EXPECT_STREQ(files[i].expected_md5, DecodeFile(files[i].name, t).c_str())
-          << "threads = " << t;
-    }
-  }
 }
 
 // Test tile quantity changes within one file.
-TEST(VP9DecodeMTTest, MTDecode3) {
+TEST(VP9DecodeMultiThreadedTest, Decode3) {
-  static const struct {
+  static const FileList files[] = {
-    const char *name;
-    const char *expected_md5;
-  } files[] = {
     { "vp90-2-14-resize-fp-tiles-1-16.webm",
       "0cd5e632c326297e975f38949c31ea94" },
     { "vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm",
@@ -207,14 +270,10 @@ TEST(VP9DecodeMTTest, MTDecode3) {
       "ae96f21f21b6370cc0125621b441fc52" },
     { "vp90-2-14-resize-fp-tiles-8-4.webm",
       "3eb4f24f10640d42218f7fd7b9fd30d4" },
+    { NULL, NULL }
   };
 
-  for (int i = 0; i < static_cast<int>(sizeof(files) / sizeof(files[0])); ++i) {
+  DecodeFiles(files);
-    for (int t = 2; t <= 8; ++t) {
-      EXPECT_STREQ(files[i].expected_md5, DecodeFile(files[i].name, t).c_str())
-          << "threads = " << t;
-    }
-  }
 }
 #endif  // CONFIG_WEBM_IO
 

test/vpx_temporal_svc_encoder.sh

@@ -39,7 +39,10 @@ vpx_tsvc_encoder() {
 
   shift 2
 
-  [ -x "${encoder}" ] || return 1
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
 
   eval "${encoder}" "${YUV_RAW_INPUT}" "${output_file}" "${codec}" \
       "${YUV_RAW_INPUT_WIDTH}" "${YUV_RAW_INPUT_HEIGHT}" \

test/y4m_test.cc

@@ -0,0 +1,193 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+
+#include <string>
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "./y4menc.h"
+
+namespace {
+
+using std::string;
+using std::tr1::make_tuple;
+
+static const unsigned int kWidth  = 160;
+static const unsigned int kHeight = 90;
+static const unsigned int kFrames = 10;
+
+typedef std::tr1::tuple<const char *, const unsigned int,
+        const vpx_img_fmt, const char *> test_entry_type;
+
+static const test_entry_type kY4mTestVectors[] = {
+  make_tuple("park_joy_90p_8_420.y4m", 8, VPX_IMG_FMT_I420,
+             "e5406275b9fc6bb3436c31d4a05c1cab"),
+  make_tuple("park_joy_90p_8_422.y4m", 8, VPX_IMG_FMT_I422,
+             "284a47a47133b12884ec3a14e959a0b6"),
+  make_tuple("park_joy_90p_8_444.y4m", 8, VPX_IMG_FMT_I444,
+             "90517ff33843d85de712fd4fe60dbed0"),
+  make_tuple("park_joy_90p_10_420.y4m", 10, VPX_IMG_FMT_I42016,
+             "63f21f9f717d8b8631bd2288ee87137b"),
+  make_tuple("park_joy_90p_10_422.y4m", 10, VPX_IMG_FMT_I42216,
+             "48ab51fb540aed07f7ff5af130c9b605"),
+  make_tuple("park_joy_90p_10_444.y4m", 10, VPX_IMG_FMT_I44416,
+             "067bfd75aa85ff9bae91fa3e0edd1e3e"),
+  make_tuple("park_joy_90p_12_420.y4m", 12, VPX_IMG_FMT_I42016,
+             "9e6d8f6508c6e55625f6b697bc461cef"),
+  make_tuple("park_joy_90p_12_422.y4m", 12, VPX_IMG_FMT_I42216,
+             "b239c6b301c0b835485be349ca83a7e3"),
+  make_tuple("park_joy_90p_12_444.y4m", 12, VPX_IMG_FMT_I44416,
+             "5a6481a550821dab6d0192f5c63845e9")
+};
+
+static void write_image_file(const vpx_image_t *img, FILE *file) {
+  int plane, y;
+  for (plane = 0; plane < 3; ++plane) {
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int bytes_per_sample = (img->fmt & VPX_IMG_FMT_HIGH) ? 2 : 1;
+    const int h = (plane ? (img->d_h + img->y_chroma_shift) >>
+                   img->y_chroma_shift : img->d_h);
+    const int w = (plane ? (img->d_w + img->x_chroma_shift) >>
+                   img->x_chroma_shift : img->d_w);
+    for (y = 0; y < h; ++y) {
+      fwrite(buf, bytes_per_sample, w, file);
+      buf += stride;
+    }
+  }
+}
+
+class Y4mVideoSourceTest
+    : public ::testing::TestWithParam<test_entry_type>,
+      public ::libvpx_test::Y4mVideoSource {
+ protected:
+  Y4mVideoSourceTest() : Y4mVideoSource("", 0, 0) {}
+
+  virtual ~Y4mVideoSourceTest() {
+    CloseSource();
+  }
+
+  virtual void Init(const std::string &file_name, int limit) {
+    file_name_ = file_name;
+    start_ = 0;
+    limit_ = limit;
+    frame_ = 0;
+    Begin();
+  }
+
+  // Checks y4m header information
+  void HeaderChecks(unsigned int bit_depth, vpx_img_fmt_t fmt) {
+    ASSERT_TRUE(input_file_ != NULL);
+    ASSERT_EQ(y4m_.pic_w, (int)kWidth);
+    ASSERT_EQ(y4m_.pic_h, (int)kHeight);
+    ASSERT_EQ(img()->d_w, kWidth);
+    ASSERT_EQ(img()->d_h, kHeight);
+    ASSERT_EQ(y4m_.bit_depth, bit_depth);
+    ASSERT_EQ(y4m_.vpx_fmt, fmt);
+    if (fmt == VPX_IMG_FMT_I420 || fmt == VPX_IMG_FMT_I42016) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 3 / 2);
+      ASSERT_EQ(img()->x_chroma_shift, 1U);
+      ASSERT_EQ(img()->y_chroma_shift, 1U);
+    }
+    if (fmt == VPX_IMG_FMT_I422 || fmt == VPX_IMG_FMT_I42216) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 2);
+      ASSERT_EQ(img()->x_chroma_shift, 1U);
+      ASSERT_EQ(img()->y_chroma_shift, 0U);
+    }
+    if (fmt == VPX_IMG_FMT_I444 || fmt == VPX_IMG_FMT_I44416) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 3);
+      ASSERT_EQ(img()->x_chroma_shift, 0U);
+      ASSERT_EQ(img()->y_chroma_shift, 0U);
+    }
+  }
+
+  // Checks MD5 of the raw frame data
+  void Md5Check(const string &expected_md5) {
+    ASSERT_TRUE(input_file_ != NULL);
+    libvpx_test::MD5 md5;
+    for (unsigned int i = start_; i < limit_; i++) {
+      md5.Add(img());
+      Next();
+    }
+    ASSERT_EQ(string(md5.Get()), expected_md5);
+  }
+};
+
+TEST_P(Y4mVideoSourceTest, SourceTest) {
+  const char *filename = GET_PARAM(0);
+  const unsigned int bit_depth = GET_PARAM(1);
+  const vpx_img_fmt format = GET_PARAM(2);
+  const char *md5raw = GET_PARAM(3);
+
+  Init(filename, kFrames);
+  HeaderChecks(bit_depth, format);
+  Md5Check(md5raw);
+}
+
+INSTANTIATE_TEST_CASE_P(C, Y4mVideoSourceTest,
+                        ::testing::ValuesIn(kY4mTestVectors));
+
+class Y4mVideoWriteTest
+    : public Y4mVideoSourceTest {
+ protected:
+  Y4mVideoWriteTest() : Y4mVideoSourceTest() {}
+
+  virtual void ReplaceInputFp(FILE *input_file) {
+    CloseSource();
+    frame_ = 0;
+    input_file_ = input_file;
+    rewind(input_file_);
+    ReadSourceToStart();
+  }
+
+  // Writes out a y4m file and then reads it back
+  void WriteY4mAndReadBack() {
+    ASSERT_TRUE(input_file_ != NULL);
+    char buf[Y4M_BUFFER_SIZE] = {0};
+    const struct VpxRational framerate = {y4m_.fps_n, y4m_.fps_d};
+    FILE *out_file = libvpx_test::OpenTempOutFile();
+    ASSERT_TRUE(out_file != NULL);
+    y4m_write_file_header(buf, sizeof(buf),
+                          kWidth, kHeight,
+                          &framerate, y4m_.vpx_fmt,
+                          y4m_.bit_depth);
+    fputs(buf, out_file);
+    for (unsigned int i = start_; i < limit_; i++) {
+      y4m_write_frame_header(buf, sizeof(buf));
+      fputs(buf, out_file);
+      write_image_file(img(), out_file);
+      Next();
+    }
+    ReplaceInputFp(out_file);
+  }
+
+  virtual void Init(const std::string &file_name, int limit) {
+    Y4mVideoSourceTest::Init(file_name, limit);
+    WriteY4mAndReadBack();
+  }
+};
+
+TEST_P(Y4mVideoWriteTest, WriteTest) {
+  const char *filename = GET_PARAM(0);
+  const unsigned int bit_depth = GET_PARAM(1);
+  const vpx_img_fmt format = GET_PARAM(2);
+  const char *md5raw = GET_PARAM(3);
+
+  Init(filename, kFrames);
+  HeaderChecks(bit_depth, format);
+  Md5Check(md5raw);
+}
+
+INSTANTIATE_TEST_CASE_P(C, Y4mVideoWriteTest,
+                        ::testing::ValuesIn(kY4mTestVectors));
+
+}  // namespace

test/y4m_video_source.h

@@ -38,24 +38,30 @@ class Y4mVideoSource : public VideoSource {
     CloseSource();
   }
 
-  virtual void Begin() {
+  virtual void OpenSource() {
     CloseSource();
     input_file_ = OpenTestDataFile(file_name_);
     ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
                                      << file_name_;
+  }
 
-    y4m_input_open(&y4m_, input_file_, NULL, 0, 0);
+  virtual void ReadSourceToStart() {
+    ASSERT_TRUE(input_file_ != NULL);
+    ASSERT_FALSE(y4m_input_open(&y4m_, input_file_, NULL, 0, 0));
     framerate_numerator_ = y4m_.fps_n;
     framerate_denominator_ = y4m_.fps_d;
-
     frame_ = 0;
     for (unsigned int i = 0; i < start_; i++) {
       Next();
     }
-
     FillFrame();
   }
 
+  virtual void Begin() {
+    OpenSource();
+    ReadSourceToStart();
+  }
+
   virtual void Next() {
     ++frame_;
     FillFrame();

third_party/libmkv/EbmlIDs.h

@@ -0,0 +1,231 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+#ifndef MKV_DEFS_HPP
+#define MKV_DEFS_HPP 1
+
+/* Commenting out values not available in webm, but available in matroska */
+
+enum mkv {
+  EBML = 0x1A45DFA3,
+  EBMLVersion = 0x4286,
+  EBMLReadVersion = 0x42F7,
+  EBMLMaxIDLength = 0x42F2,
+  EBMLMaxSizeLength = 0x42F3,
+  DocType = 0x4282,
+  DocTypeVersion = 0x4287,
+  DocTypeReadVersion = 0x4285,
+/* CRC_32 = 0xBF, */
+  Void = 0xEC,
+  SignatureSlot = 0x1B538667,
+  SignatureAlgo = 0x7E8A,
+  SignatureHash = 0x7E9A,
+  SignaturePublicKey = 0x7EA5,
+  Signature = 0x7EB5,
+  SignatureElements = 0x7E5B,
+  SignatureElementList = 0x7E7B,
+  SignedElement = 0x6532,
+  /* segment */
+  Segment = 0x18538067,
+  /* Meta Seek Information */
+  SeekHead = 0x114D9B74,
+  Seek = 0x4DBB,
+  SeekID = 0x53AB,
+  SeekPosition = 0x53AC,
+  /* Segment Information */
+  Info = 0x1549A966,
+/* SegmentUID = 0x73A4, */
+/* SegmentFilename = 0x7384, */
+/* PrevUID = 0x3CB923, */
+/* PrevFilename = 0x3C83AB, */
+/* NextUID = 0x3EB923, */
+/* NextFilename = 0x3E83BB, */
+/* SegmentFamily = 0x4444, */
+/* ChapterTranslate = 0x6924, */
+/* ChapterTranslateEditionUID = 0x69FC, */
+/* ChapterTranslateCodec = 0x69BF, */
+/* ChapterTranslateID = 0x69A5, */
+  TimecodeScale = 0x2AD7B1,
+  Segment_Duration = 0x4489,
+  DateUTC = 0x4461,
+/* Title = 0x7BA9, */
+  MuxingApp = 0x4D80,
+  WritingApp = 0x5741,
+  /* Cluster */
+  Cluster = 0x1F43B675,
+  Timecode = 0xE7,
+/* SilentTracks = 0x5854, */
+/* SilentTrackNumber = 0x58D7, */
+/* Position = 0xA7, */
+  PrevSize = 0xAB,
+  BlockGroup = 0xA0,
+  Block = 0xA1,
+/* BlockVirtual = 0xA2, */
+  BlockAdditions = 0x75A1,
+  BlockMore = 0xA6,
+  BlockAddID = 0xEE,
+  BlockAdditional = 0xA5,
+  BlockDuration = 0x9B,
+/* ReferencePriority = 0xFA, */
+  ReferenceBlock = 0xFB,
+/* ReferenceVirtual = 0xFD, */
+/* CodecState = 0xA4, */
+/* Slices = 0x8E, */
+/* TimeSlice = 0xE8, */
+  LaceNumber = 0xCC,
+/* FrameNumber = 0xCD, */
+/* BlockAdditionID = 0xCB, */
+/* MkvDelay = 0xCE, */
+/* Cluster_Duration = 0xCF, */
+  SimpleBlock = 0xA3,
+/* EncryptedBlock = 0xAF, */
+  /* Track */
+  Tracks = 0x1654AE6B,
+  TrackEntry = 0xAE,
+  TrackNumber = 0xD7,
+  TrackUID = 0x73C5,
+  TrackType = 0x83,
+  FlagEnabled = 0xB9,
+  FlagDefault = 0x88,
+  FlagForced = 0x55AA,
+  FlagLacing = 0x9C,
+/* MinCache = 0x6DE7, */
+/* MaxCache = 0x6DF8, */
+  DefaultDuration = 0x23E383,
+/* TrackTimecodeScale = 0x23314F, */
+/* TrackOffset = 0x537F, */
+  MaxBlockAdditionID = 0x55EE,
+  Name = 0x536E,
+  Language = 0x22B59C,
+  CodecID = 0x86,
+  CodecPrivate = 0x63A2,
+  CodecName = 0x258688,
+/* AttachmentLink = 0x7446, */
+/* CodecSettings = 0x3A9697, */
+/* CodecInfoURL = 0x3B4040, */
+/* CodecDownloadURL = 0x26B240, */
+/* CodecDecodeAll = 0xAA, */
+/* TrackOverlay = 0x6FAB, */
+/* TrackTranslate = 0x6624, */
+/* TrackTranslateEditionUID = 0x66FC, */
+/* TrackTranslateCodec = 0x66BF, */
+/* TrackTranslateTrackID = 0x66A5, */
+  /* video */
+  Video = 0xE0,
+  FlagInterlaced = 0x9A,
+  StereoMode = 0x53B8,
+  AlphaMode = 0x53C0,
+  PixelWidth = 0xB0,
+  PixelHeight = 0xBA,
+  PixelCropBottom = 0x54AA,
+  PixelCropTop = 0x54BB,
+  PixelCropLeft = 0x54CC,
+  PixelCropRight = 0x54DD,
+  DisplayWidth = 0x54B0,
+  DisplayHeight = 0x54BA,
+  DisplayUnit = 0x54B2,
+  AspectRatioType = 0x54B3,
+/* ColourSpace = 0x2EB524, */
+/* GammaValue = 0x2FB523, */
+  FrameRate = 0x2383E3,
+  /* end video */
+  /* audio */
+  Audio = 0xE1,
+  SamplingFrequency = 0xB5,
+  OutputSamplingFrequency = 0x78B5,
+  Channels = 0x9F,
+/* ChannelPositions = 0x7D7B, */
+  BitDepth = 0x6264,
+  /* end audio */
+  /* content encoding */
+/* ContentEncodings = 0x6d80, */
+/* ContentEncoding = 0x6240, */
+/* ContentEncodingOrder = 0x5031, */
+/* ContentEncodingScope = 0x5032, */
+/* ContentEncodingType = 0x5033, */
+/* ContentCompression = 0x5034, */
+/* ContentCompAlgo = 0x4254, */
+/* ContentCompSettings = 0x4255, */
+/* ContentEncryption = 0x5035, */
+/* ContentEncAlgo = 0x47e1, */
+/* ContentEncKeyID = 0x47e2, */
+/* ContentSignature = 0x47e3, */
+/* ContentSigKeyID = 0x47e4, */
+/* ContentSigAlgo = 0x47e5, */
+/* ContentSigHashAlgo = 0x47e6, */
+  /* end content encoding */
+  /* Cueing Data */
+  Cues = 0x1C53BB6B,
+  CuePoint = 0xBB,
+  CueTime = 0xB3,
+  CueTrackPositions = 0xB7,
+  CueTrack = 0xF7,
+  CueClusterPosition = 0xF1,
+  CueBlockNumber = 0x5378
+/* CueCodecState = 0xEA, */
+/* CueReference = 0xDB, */
+/* CueRefTime = 0x96, */
+/* CueRefCluster = 0x97, */
+/* CueRefNumber = 0x535F, */
+/* CueRefCodecState = 0xEB, */
+  /* Attachment */
+/* Attachments = 0x1941A469, */
+/* AttachedFile = 0x61A7, */
+/* FileDescription = 0x467E, */
+/* FileName = 0x466E, */
+/* FileMimeType = 0x4660, */
+/* FileData = 0x465C, */
+/* FileUID = 0x46AE, */
+/* FileReferral = 0x4675, */
+  /* Chapters */
+/* Chapters = 0x1043A770, */
+/* EditionEntry = 0x45B9, */
+/* EditionUID = 0x45BC, */
+/* EditionFlagHidden = 0x45BD, */
+/* EditionFlagDefault = 0x45DB, */
+/* EditionFlagOrdered = 0x45DD, */
+/* ChapterAtom = 0xB6, */
+/* ChapterUID = 0x73C4, */
+/* ChapterTimeStart = 0x91, */
+/* ChapterTimeEnd = 0x92, */
+/* ChapterFlagHidden = 0x98, */
+/* ChapterFlagEnabled = 0x4598, */
+/* ChapterSegmentUID = 0x6E67, */
+/* ChapterSegmentEditionUID = 0x6EBC, */
+/* ChapterPhysicalEquiv = 0x63C3, */
+/* ChapterTrack = 0x8F, */
+/* ChapterTrackNumber = 0x89, */
+/* ChapterDisplay = 0x80, */
+/* ChapString = 0x85, */
+/* ChapLanguage = 0x437C, */
+/* ChapCountry = 0x437E, */
+/* ChapProcess = 0x6944, */
+/* ChapProcessCodecID = 0x6955, */
+/* ChapProcessPrivate = 0x450D, */
+/* ChapProcessCommand = 0x6911, */
+/* ChapProcessTime = 0x6922, */
+/* ChapProcessData = 0x6933, */
+  /* Tagging */
+/* Tags = 0x1254C367, */
+/* Tag = 0x7373, */
+/* Targets = 0x63C0, */
+/* TargetTypeValue = 0x68CA, */
+/* TargetType = 0x63CA, */
+/* Tagging_TrackUID = 0x63C5, */
+/* Tagging_EditionUID = 0x63C9, */
+/* Tagging_ChapterUID = 0x63C4, */
+/* AttachmentUID = 0x63C6, */
+/* SimpleTag = 0x67C8, */
+/* TagName = 0x45A3, */
+/* TagLanguage = 0x447A, */
+/* TagDefault = 0x4484, */
+/* TagString = 0x4487, */
+/* TagBinary = 0x4485, */
+};
+#endif

third_party/libmkv/EbmlWriter.c

@@ -0,0 +1,157 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+#include "EbmlWriter.h"
+#include <stdlib.h>
+#include <wchar.h>
+#include <string.h>
+#include <limits.h>
+#if defined(_MSC_VER)
+#define LITERALU64(n) n
+#else
+#define LITERALU64(n) n##LLU
+#endif
+
+void Ebml_WriteLen(EbmlGlobal *glob, int64_t val) {
+  /* TODO check and make sure we are not > than 0x0100000000000000LLU */
+  unsigned char size = 8; /* size in bytes to output */
+
+  /* mask to compare for byte size */
+  int64_t minVal = 0xff;
+
+  for (size = 1; size < 8; size ++) {
+    if (val < minVal)
+      break;
+
+    minVal = (minVal << 7);
+  }
+
+  val |= (((uint64_t)0x80) << ((size - 1) * 7));
+
+  Ebml_Serialize(glob, (void *) &val, sizeof(val), size);
+}
+
+void Ebml_WriteString(EbmlGlobal *glob, const char *str) {
+  const size_t size_ = strlen(str);
+  const uint64_t  size = size_;
+  Ebml_WriteLen(glob, size);
+  /* TODO: it's not clear from the spec whether the nul terminator
+   * should be serialized too.  For now we omit the null terminator.
+   */
+  Ebml_Write(glob, str, (unsigned long)size);
+}
+
+void Ebml_WriteUTF8(EbmlGlobal *glob, const wchar_t *wstr) {
+  const size_t strlen = wcslen(wstr);
+
+  /* TODO: it's not clear from the spec whether the nul terminator
+   * should be serialized too.  For now we include it.
+   */
+  const uint64_t  size = strlen;
+
+  Ebml_WriteLen(glob, size);
+  Ebml_Write(glob, wstr, (unsigned long)size);
+}
+
+void Ebml_WriteID(EbmlGlobal *glob, unsigned long class_id) {
+  int len;
+
+  if (class_id >= 0x01000000)
+    len = 4;
+  else if (class_id >= 0x00010000)
+    len = 3;
+  else if (class_id >= 0x00000100)
+    len = 2;
+  else
+    len = 1;
+
+  Ebml_Serialize(glob, (void *)&class_id, sizeof(class_id), len);
+}
+
+void Ebml_SerializeUnsigned64(EbmlGlobal *glob, unsigned long class_id, uint64_t ui) {
+  unsigned char sizeSerialized = 8 | 0x80;
+  Ebml_WriteID(glob, class_id);
+  Ebml_Serialize(glob, &sizeSerialized, sizeof(sizeSerialized), 1);
+  Ebml_Serialize(glob, &ui, sizeof(ui), 8);
+}
+
+void Ebml_SerializeUnsigned(EbmlGlobal *glob, unsigned long class_id, unsigned long ui) {
+  unsigned char size = 8; /* size in bytes to output */
+  unsigned char sizeSerialized = 0;
+  unsigned long minVal;
+
+  Ebml_WriteID(glob, class_id);
+  minVal = 0x7fLU; /* mask to compare for byte size */
+
+  for (size = 1; size < 4; size ++) {
+    if (ui < minVal) {
+      break;
+    }
+
+    minVal <<= 7;
+  }
+
+  sizeSerialized = 0x80 | size;
+  Ebml_Serialize(glob, &sizeSerialized, sizeof(sizeSerialized), 1);
+  Ebml_Serialize(glob, &ui, sizeof(ui), size);
+}
+/* TODO: perhaps this is a poor name for this id serializer helper function */
+void Ebml_SerializeBinary(EbmlGlobal *glob, unsigned long class_id, unsigned long bin) {
+  int size;
+  for (size = 4; size > 1; size--) {
+    if (bin & (unsigned int)0x000000ff << ((size - 1) * 8))
+      break;
+  }
+  Ebml_WriteID(glob, class_id);
+  Ebml_WriteLen(glob, size);
+  Ebml_WriteID(glob, bin);
+}
+
+void Ebml_SerializeFloat(EbmlGlobal *glob, unsigned long class_id, double d) {
+  unsigned char len = 0x88;
+
+  Ebml_WriteID(glob, class_id);
+  Ebml_Serialize(glob, &len, sizeof(len), 1);
+  Ebml_Serialize(glob,  &d, sizeof(d), 8);
+}
+
+void Ebml_WriteSigned16(EbmlGlobal *glob, short val) {
+  signed long out = ((val & 0x003FFFFF) | 0x00200000) << 8;
+  Ebml_Serialize(glob, &out, sizeof(out), 3);
+}
+
+void Ebml_SerializeString(EbmlGlobal *glob, unsigned long class_id, const char *s) {
+  Ebml_WriteID(glob, class_id);
+  Ebml_WriteString(glob, s);
+}
+
+void Ebml_SerializeUTF8(EbmlGlobal *glob, unsigned long class_id, wchar_t *s) {
+  Ebml_WriteID(glob,  class_id);
+  Ebml_WriteUTF8(glob,  s);
+}
+
+void Ebml_SerializeData(EbmlGlobal *glob, unsigned long class_id, unsigned char *data, unsigned long data_length) {
+  Ebml_WriteID(glob, class_id);
+  Ebml_WriteLen(glob, data_length);
+  Ebml_Write(glob,  data, data_length);
+}
+
+void Ebml_WriteVoid(EbmlGlobal *glob, unsigned long vSize) {
+  unsigned char tmp = 0;
+  unsigned long i = 0;
+
+  Ebml_WriteID(glob, 0xEC);
+  Ebml_WriteLen(glob, vSize);
+
+  for (i = 0; i < vSize; i++) {
+    Ebml_Write(glob, &tmp, 1);
+  }
+}
+
+/* TODO Serialize Date */

third_party/libmkv/EbmlWriter.h

@@ -0,0 +1,42 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+#ifndef EBMLWRITER_HPP
+#define EBMLWRITER_HPP
+#include <stddef.h>
+#include "vpx/vpx_integer.h"
+
+/* note: you must define write and serialize functions as well as your own
+ * EBML_GLOBAL
+ *
+ * These functions MUST be implemented
+ */
+
+typedef struct EbmlGlobal EbmlGlobal;
+void  Ebml_Serialize(EbmlGlobal *glob, const void *, int, unsigned long);
+void  Ebml_Write(EbmlGlobal *glob, const void *, unsigned long);
+
+/*****/
+
+void Ebml_WriteLen(EbmlGlobal *glob, int64_t val);
+void Ebml_WriteString(EbmlGlobal *glob, const char *str);
+void Ebml_WriteUTF8(EbmlGlobal *glob, const wchar_t *wstr);
+void Ebml_WriteID(EbmlGlobal *glob, unsigned long class_id);
+void Ebml_SerializeUnsigned64(EbmlGlobal *glob, unsigned long class_id, uint64_t ui);
+void Ebml_SerializeUnsigned(EbmlGlobal *glob, unsigned long class_id, unsigned long ui);
+void Ebml_SerializeBinary(EbmlGlobal *glob, unsigned long class_id, unsigned long ui);
+void Ebml_SerializeFloat(EbmlGlobal *glob, unsigned long class_id, double d);
+/* TODO make this more generic to signed */
+void Ebml_WriteSigned16(EbmlGlobal *glob, short val);
+void Ebml_SerializeString(EbmlGlobal *glob, unsigned long class_id, const char *s);
+void Ebml_SerializeUTF8(EbmlGlobal *glob, unsigned long class_id, wchar_t *s);
+void Ebml_SerializeData(EbmlGlobal *glob, unsigned long class_id, unsigned char *data, unsigned long data_length);
+void Ebml_WriteVoid(EbmlGlobal *glob, unsigned long vSize);
+/* TODO need date function */
+#endif

tools_common.h

@@ -90,6 +90,7 @@ struct VpxInputContext {
   uint32_t width;
   uint32_t height;
   vpx_img_fmt_t fmt;
+  vpx_bit_depth_t bit_depth;
   int only_i420;
   uint32_t fourcc;
   struct VpxRational framerate;

vp8/common/postproc.c

@@ -393,12 +393,12 @@ void vp8_de_noise(VP8_COMMON                 *cm,
                   int                         low_var_thresh,
                   int                         flag)
 {
+    int mbr;
     double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
     int ppl = (int)(level + .5);
-    int mb_rows = source->y_width >> 4;
+    int mb_rows = cm->mb_rows;
-    int mb_cols = source->y_height >> 4;
+    int mb_cols = cm->mb_cols;
     unsigned char *limits = cm->pp_limits_buffer;;
-    int mbr, mbc;
     (void) post;
     (void) low_var_thresh;
     (void) flag;

vp8/common/rtcd_defs.pl

@@ -463,9 +463,7 @@ $vp8_short_walsh4x4_neon_asm=vp8_short_walsh4x4_neon;
 # Quantizer
 #
 add_proto qw/void vp8_regular_quantize_b/, "struct block *, struct blockd *";
-specialize qw/vp8_regular_quantize_b sse2/;
+specialize qw/vp8_regular_quantize_b sse2 sse4_1/;
-# TODO(johann) Update sse4 implementation and re-enable
-#$vp8_regular_quantize_b_sse4_1=vp8_regular_quantize_b_sse4;
 
 add_proto qw/void vp8_fast_quantize_b/, "struct block *, struct blockd *";
 specialize qw/vp8_fast_quantize_b sse2 ssse3 media neon_asm/;
@@ -554,6 +552,9 @@ $vp8_yv12_copy_partial_frame_neon_asm=vp8_yv12_copy_partial_frame_neon;
 if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") {
     add_proto qw/int vp8_denoiser_filter/, "unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
     specialize qw/vp8_denoiser_filter sse2 neon/;
+    add_proto qw/int vp8_denoiser_filter_uv/, "unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
+    specialize qw/vp8_denoiser_filter_uv sse2/;
+
 }
 
 # End of encoder only functions

vp8/common/x86/postproc_mmx.asm

@@ -246,7 +246,6 @@ sym(vp8_mbpost_proc_down_mmx):
 ;                            unsigned char whiteclamp[16],
 ;                            unsigned char bothclamp[16],
 ;                            unsigned int Width, unsigned int Height, int Pitch)
-extern sym(rand)
 global sym(vp8_plane_add_noise_mmx) PRIVATE
 sym(vp8_plane_add_noise_mmx):
     push        rbp
@@ -258,7 +257,7 @@ sym(vp8_plane_add_noise_mmx):
     ; end prolog
 
 .addnoise_loop:
-    call sym(rand) WRT_PLT
+    call sym(LIBVPX_RAND) WRT_PLT
     mov     rcx, arg(1) ;noise
     and     rax, 0xff
     add     rcx, rax

vp8/common/x86/postproc_sse2.asm

@@ -660,7 +660,6 @@ sym(vp8_mbpost_proc_across_ip_xmm):
 ;                            unsigned char whiteclamp[16],
 ;                            unsigned char bothclamp[16],
 ;                            unsigned int Width, unsigned int Height, int Pitch)
-extern sym(rand)
 global sym(vp8_plane_add_noise_wmt) PRIVATE
 sym(vp8_plane_add_noise_wmt):
     push        rbp
@@ -672,7 +671,7 @@ sym(vp8_plane_add_noise_wmt):
     ; end prolog
 
 .addnoise_loop:
-    call sym(rand) WRT_PLT
+    call sym(LIBVPX_RAND) WRT_PLT
     mov     rcx, arg(1) ;noise
     and     rax, 0xff
     add     rcx, rax

vp8/common/x86/postproc_x86.c

@@ -1,24 +0,0 @@
-/*
- *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE 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.
- */
-
-/* On Android NDK, rand is inlined function, but postproc needs rand symbol */
-#if defined(__ANDROID__)
-#define rand __rand
-#include <stdlib.h>
-#undef rand
-
-extern int rand(void)
-{
-  return __rand();
-}
-#else
-/* ISO C forbids an empty translation unit. */
-int vp8_unused;
-#endif

vp8/encoder/bitstream.h

@@ -18,18 +18,18 @@ extern "C" {
 
 #if HAVE_EDSP
 void vp8cx_pack_tokens_armv5(vp8_writer *w, const TOKENEXTRA *p, int xcount,
-                             const vp8_token *,
+                             vp8_token *,
                              const vp8_extra_bit_struct *,
                              const vp8_tree_index *);
 void vp8cx_pack_tokens_into_partitions_armv5(VP8_COMP *,
                                              unsigned char * cx_data,
                                              const unsigned char *cx_data_end,
                                              int num_parts,
-                                             const vp8_token *,
+                                             vp8_token *,
                                              const vp8_extra_bit_struct *,
                                              const vp8_tree_index *);
 void vp8cx_pack_mb_row_tokens_armv5(VP8_COMP *cpi, vp8_writer *w,
-                                    const vp8_token *,
+                                    vp8_token *,
                                     const vp8_extra_bit_struct *,
                                     const vp8_tree_index *);
 # define pack_tokens(a,b,c)                  \

vp8/encoder/denoising.c

@@ -191,10 +191,154 @@ int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride,
     return FILTER_BLOCK;
 }
 
-int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height)
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg_uv,
+                             int mc_avg_uv_stride,
+                             unsigned char *running_avg_uv,
+                             int avg_uv_stride,
+                             unsigned char *sig,
+                             int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    unsigned char *running_avg_uv_start = running_avg_uv;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
+    int r, c;
+    int sum_diff = 0;
+    int sum_block = 0;
+    int adj_val[3] = {3, 4, 6};
+    int shift_inc1 = 0;
+    int shift_inc2 = 1;
+    /* If motion_magnitude is small, making the denoiser more aggressive by
+     * increasing the adjustment for each level. Add another increment for
+     * blocks that are labeled for increase denoising. */
+    if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) {
+      if (increase_denoising) {
+        shift_inc1 = 1;
+        shift_inc2 = 2;
+      }
+      adj_val[0] += shift_inc2;
+      adj_val[1] += shift_inc2;
+      adj_val[2] += shift_inc2;
+    }
+
+    // Avoid denoising color signal if its close to average level.
+    for (r = 0; r < 8; ++r) {
+      for (c = 0; c < 8; ++c) {
+        sum_block += sig[c];
+      }
+      sig += sig_stride;
+    }
+    if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+      return COPY_BLOCK;
+    }
+
+    sig -= sig_stride * 8;
+    for (r = 0; r < 8; ++r) {
+      for (c = 0; c < 8; ++c) {
+        int diff = 0;
+        int adjustment = 0;
+        int absdiff = 0;
+
+        diff = mc_running_avg_uv[c] - sig[c];
+        absdiff = abs(diff);
+
+        // When |diff| <= |3 + shift_inc1|, use pixel value from
+        // last denoised raw.
+        if (absdiff <= 3 + shift_inc1) {
+          running_avg_uv[c] = mc_running_avg_uv[c];
+          sum_diff += diff;
+        } else {
+          if (absdiff >= 4 && absdiff <= 7)
+            adjustment = adj_val[0];
+          else if (absdiff >= 8 && absdiff <= 15)
+            adjustment = adj_val[1];
+          else
+            adjustment = adj_val[2];
+          if (diff > 0) {
+            if ((sig[c] + adjustment) > 255)
+              running_avg_uv[c] = 255;
+            else
+              running_avg_uv[c] = sig[c] + adjustment;
+            sum_diff += adjustment;
+          } else {
+            if ((sig[c] - adjustment) < 0)
+              running_avg_uv[c] = 0;
+            else
+              running_avg_uv[c] = sig[c] - adjustment;
+            sum_diff -= adjustment;
+          }
+        }
+      }
+      /* Update pointers for next iteration. */
+      sig += sig_stride;
+      mc_running_avg_uv += mc_avg_uv_stride;
+      running_avg_uv += avg_uv_stride;
+    }
+
+    sum_diff_thresh= SUM_DIFF_THRESHOLD_UV;
+    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising), check
+      // if we can still apply some (weaker) temporal filtering to this block,
+      // that would otherwise not be denoised at all. Simplest is to apply
+      // an additional adjustment to running_avg_y to bring it closer to sig.
+      // The adjustment is capped by a maximum delta, and chosen such that
+      // in most cases the resulting sum_diff will be within the
+      // accceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over threshold.
+      int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        sig -= sig_stride * 8;
+        mc_running_avg_uv -= mc_avg_uv_stride * 8;
+        running_avg_uv -= avg_uv_stride * 8;
+        for (r = 0; r < 8; ++r) {
+          for (c = 0; c < 8; ++c) {
+            int diff = mc_running_avg_uv[c] - sig[c];
+            int adjustment = abs(diff);
+            if (adjustment > delta)
+              adjustment = delta;
+            if (diff > 0) {
+              // Bring denoised signal down.
+              if (running_avg_uv[c] - adjustment < 0)
+                running_avg_uv[c] = 0;
+              else
+                running_avg_uv[c] = running_avg_uv[c] - adjustment;
+              sum_diff -= adjustment;
+            } else if (diff < 0) {
+              // Bring denoised signal up.
+              if (running_avg_uv[c] + adjustment > 255)
+                running_avg_uv[c] = 255;
+              else
+                running_avg_uv[c] = running_avg_uv[c] + adjustment;
+              sum_diff += adjustment;
+            }
+          }
+          // TODO(marpan): Check here if abs(sum_diff) has gone below the
+          // threshold sum_diff_thresh, and if so, we can exit the row loop.
+          sig += sig_stride;
+          mc_running_avg_uv += mc_avg_uv_stride;
+          running_avg_uv += avg_uv_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh)
+          return COPY_BLOCK;
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+
+    vp8_copy_mem8x8(running_avg_uv_start, avg_uv_stride, sig_start,
+                    sig_stride);
+    return FILTER_BLOCK;
+}
+
+int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
+                          int num_mb_rows, int num_mb_cols)
 {
     int i;
     assert(denoiser);
+    denoiser->num_mb_cols = num_mb_cols;
 
     for (i = 0; i < MAX_REF_FRAMES; i++)
     {
@@ -222,6 +366,10 @@ int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height)
 
     vpx_memset(denoiser->yv12_mc_running_avg.buffer_alloc, 0,
                denoiser->yv12_mc_running_avg.frame_size);
+
+    denoiser->denoise_state = vpx_calloc((num_mb_rows * num_mb_cols), 1);
+    vpx_memset(denoiser->denoise_state, 0, (num_mb_rows * num_mb_cols));
+
     return 0;
 }
 
@@ -235,6 +383,7 @@ void vp8_denoiser_free(VP8_DENOISER *denoiser)
         vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_running_avg[i]);
     }
     vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_mc_running_avg);
+    vpx_free(denoiser->denoise_state);
 }
 
 
@@ -243,17 +392,28 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                              unsigned int best_sse,
                              unsigned int zero_mv_sse,
                              int recon_yoffset,
-                             int recon_uvoffset)
+                             int recon_uvoffset,
+                             loop_filter_info_n *lfi_n,
+                             int mb_row,
+                             int mb_col,
+                             int block_index)
 {
     int mv_row;
     int mv_col;
     unsigned int motion_magnitude2;
     unsigned int sse_thresh;
     int sse_diff_thresh = 0;
+    // Denoise the UV channel.
+    int apply_color_denoise = 0;
+    // Spatial loop filter: only applied selectively based on
+    // temporal filter state of block relative to top/left neighbors.
+    int apply_spatial_loop_filter = 1;
     MV_REFERENCE_FRAME frame = x->best_reference_frame;
     MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame;
 
     enum vp8_denoiser_decision decision = FILTER_BLOCK;
+    enum vp8_denoiser_decision decision_u = FILTER_BLOCK;
+    enum vp8_denoiser_decision decision_v = FILTER_BLOCK;
 
     if (zero_frame)
     {
@@ -263,7 +423,11 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
         MB_MODE_INFO saved_mbmi;
         MACROBLOCKD *filter_xd = &x->e_mbd;
         MB_MODE_INFO *mbmi = &filter_xd->mode_info_context->mbmi;
-        int sse_diff = zero_mv_sse - best_sse;
+        int sse_diff = 0;
+        // Bias on zero motion vector sse.
+        int zero_bias = 95;
+        zero_mv_sse = (unsigned int)((int64_t)zero_mv_sse * zero_bias / 100);
+        sse_diff = zero_mv_sse - best_sse;
 
         saved_mbmi = *mbmi;
 
@@ -362,6 +526,34 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                                        running_avg_y, avg_y_stride,
                                        x->thismb, 16, motion_magnitude2,
                                        x->increase_denoising);
+        denoiser->denoise_state[block_index] = motion_magnitude2 > 0 ?
+            kFilterNonZeroMV : kFilterZeroMV;
+        // Only denoise UV for zero motion, and if y channel was denoised.
+        if (apply_color_denoise &&
+            motion_magnitude2 == 0 &&
+            decision == FILTER_BLOCK) {
+          unsigned char *mc_running_avg_u =
+              denoiser->yv12_mc_running_avg.u_buffer + recon_uvoffset;
+          unsigned char *running_avg_u =
+              denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset;
+          unsigned char *mc_running_avg_v =
+              denoiser->yv12_mc_running_avg.v_buffer + recon_uvoffset;
+          unsigned char *running_avg_v =
+              denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset;
+          int mc_avg_uv_stride = denoiser->yv12_mc_running_avg.uv_stride;
+          int avg_uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
+          int signal_stride = x->block[16].src_stride;
+          decision_u =
+              vp8_denoiser_filter_uv(mc_running_avg_u, mc_avg_uv_stride,
+                                      running_avg_u, avg_uv_stride,
+                                      x->block[16].src + *x->block[16].base_src,
+                                      signal_stride, motion_magnitude2, 0);
+          decision_v =
+              vp8_denoiser_filter_uv(mc_running_avg_v, mc_avg_uv_stride,
+                                      running_avg_v, avg_uv_stride,
+                                      x->block[20].src + *x->block[20].base_src,
+                                      signal_stride, motion_magnitude2, 0);
+        }
     }
     if (decision == COPY_BLOCK)
     {
@@ -372,5 +564,73 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                 x->thismb, 16,
                 denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
                 denoiser->yv12_running_avg[INTRA_FRAME].y_stride);
+        denoiser->denoise_state[block_index] = kNoFilter;
+    }
+    if (apply_color_denoise) {
+      if (decision_u == COPY_BLOCK) {
+        vp8_copy_mem8x8(
+            x->block[16].src + *x->block[16].base_src, x->block[16].src_stride,
+            denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset,
+            denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
+      }
+      if (decision_v == COPY_BLOCK) {
+        vp8_copy_mem8x8(
+            x->block[20].src + *x->block[20].base_src, x->block[16].src_stride,
+            denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset,
+            denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
+      }
+    }
+    // Option to selectively deblock the denoised signal, for y channel only.
+    if (apply_spatial_loop_filter) {
+      loop_filter_info lfi;
+      int apply_filter_col = 0;
+      int apply_filter_row = 0;
+      int apply_filter = 0;
+      int y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
+      int uv_stride =denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
+
+      // Fix filter level to some nominal value for now.
+      int filter_level = 32;
+
+      int hev_index = lfi_n->hev_thr_lut[INTER_FRAME][filter_level];
+      lfi.mblim = lfi_n->mblim[filter_level];
+      lfi.blim = lfi_n->blim[filter_level];
+      lfi.lim = lfi_n->lim[filter_level];
+      lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+      // Apply filter if there is a difference in the denoiser filter state
+      // between the current and left/top block, or if non-zero motion vector
+      // is used for the motion-compensated filtering.
+      if (mb_col > 0) {
+        apply_filter_col = !((denoiser->denoise_state[block_index] ==
+            denoiser->denoise_state[block_index - 1]) &&
+            denoiser->denoise_state[block_index] != kFilterNonZeroMV);
+        if (apply_filter_col) {
+          // Filter left vertical edge.
+          apply_filter = 1;
+          vp8_loop_filter_mbv(
+              denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+              NULL, NULL, y_stride, uv_stride, &lfi);
+        }
+      }
+      if (mb_row > 0) {
+        apply_filter_row = !((denoiser->denoise_state[block_index] ==
+            denoiser->denoise_state[block_index - denoiser->num_mb_cols]) &&
+            denoiser->denoise_state[block_index] != kFilterNonZeroMV);
+        if (apply_filter_row) {
+          // Filter top horizontal edge.
+          apply_filter = 1;
+          vp8_loop_filter_mbh(
+              denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+              NULL, NULL, y_stride, uv_stride, &lfi);
+        }
+      }
+      if (apply_filter) {
+        // Update the signal block |x|. Pixel changes are only to top and/or
+        // left boundary pixels: can we avoid full block copy here.
+        vp8_copy_mem16x16(
+            denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+            y_stride, x->thismb, 16);
+      }
     }
 }

vp8/encoder/denoising.h

@@ -12,6 +12,7 @@
 #define VP8_ENCODER_DENOISING_H_
 
 #include "block.h"
+#include "vp8/common/loopfilter.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -21,19 +22,33 @@ extern "C" {
 #define SUM_DIFF_THRESHOLD_HIGH (16 * 16 * 3)
 #define MOTION_MAGNITUDE_THRESHOLD (8*3)
 
+#define SUM_DIFF_THRESHOLD_UV (96)   // (8 * 8 * 1.5)
+#define SUM_DIFF_THRESHOLD_HIGH_UV (8 * 8 * 2)
+#define SUM_DIFF_FROM_AVG_THRESH_UV (8 * 8 * 4)
+#define MOTION_MAGNITUDE_THRESHOLD_UV (8*3)
+
 enum vp8_denoiser_decision
 {
   COPY_BLOCK,
   FILTER_BLOCK
 };
 
+enum vp8_denoiser_filter_state {
+  kNoFilter,
+  kFilterZeroMV,
+  kFilterNonZeroMV
+};
+
 typedef struct vp8_denoiser
 {
     YV12_BUFFER_CONFIG yv12_running_avg[MAX_REF_FRAMES];
     YV12_BUFFER_CONFIG yv12_mc_running_avg;
+    unsigned char* denoise_state;
+    int num_mb_cols;
 } VP8_DENOISER;
 
-int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height);
+int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
+                          int num_mb_rows, int num_mb_cols);
 
 void vp8_denoiser_free(VP8_DENOISER *denoiser);
 
@@ -42,7 +57,11 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                              unsigned int best_sse,
                              unsigned int zero_mv_sse,
                              int recon_yoffset,
-                             int recon_uvoffset);
+                             int recon_uvoffset,
+                             loop_filter_info_n *lfi_n,
+                             int mb_row,
+                             int mb_col,
+                             int block_index);
 
 #ifdef __cplusplus
 }  // extern "C"

vp8/encoder/encodeframe.c

@@ -1246,7 +1246,7 @@ int vp8cx_encode_inter_macroblock
             x->zbin_mode_boost_enabled = 0;
         }
         vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
-                               &distortion, &intra_error);
+                               &distortion, &intra_error, mb_row, mb_col);
 
         /* switch back to the regular quantizer for the encode */
         if (cpi->sf.improved_quant)

vp8/encoder/onyx_if.c

@@ -98,6 +98,9 @@ extern double vp8_calc_ssimg
 #ifdef OUTPUT_YUV_SRC
 FILE *yuv_file;
 #endif
+#ifdef OUTPUT_YUV_DENOISED
+FILE *yuv_denoised_file;
+#endif
 
 #if 0
 FILE *framepsnr;
@@ -1748,7 +1751,8 @@ void vp8_change_config(VP8_COMP *cpi, VP8_CONFIG *oxcf)
       {
         int width = (cpi->oxcf.Width + 15) & ~15;
         int height = (cpi->oxcf.Height + 15) & ~15;
-        vp8_denoiser_allocate(&cpi->denoiser, width, height);
+        vp8_denoiser_allocate(&cpi->denoiser, width, height,
+                              cpi->common.mb_rows, cpi->common.mb_cols);
       }
     }
 #endif
@@ -1961,6 +1965,9 @@ struct VP8_COMP* vp8_create_compressor(VP8_CONFIG *oxcf)
 #ifdef OUTPUT_YUV_SRC
     yuv_file = fopen("bd.yuv", "ab");
 #endif
+#ifdef OUTPUT_YUV_DENOISED
+    yuv_denoised_file = fopen("denoised.yuv", "ab");
+#endif
 
 #if 0
     framepsnr = fopen("framepsnr.stt", "a");
@@ -2410,6 +2417,9 @@ void vp8_remove_compressor(VP8_COMP **ptr)
 #ifdef OUTPUT_YUV_SRC
     fclose(yuv_file);
 #endif
+#ifdef OUTPUT_YUV_DENOISED
+    fclose(yuv_denoised_file);
+#endif
 
 #if 0
 
@@ -2610,7 +2620,7 @@ int vp8_update_entropy(VP8_COMP *cpi, int update)
 }
 
 
-#if OUTPUT_YUV_SRC
+#if defined(OUTPUT_YUV_SRC) || defined(OUTPUT_YUV_DENOISED)
 void vp8_write_yuv_frame(FILE *yuv_file, YV12_BUFFER_CONFIG *s)
 {
     unsigned char *src = s->y_buffer;
@@ -4430,6 +4440,11 @@ static void encode_frame_to_data_rate
 
     update_reference_frames(cpi);
 
+#ifdef OUTPUT_YUV_DENOISED
+    vp8_write_yuv_frame(yuv_denoised_file,
+                        &cpi->denoiser.yv12_running_avg[INTRA_FRAME]);
+#endif
+
 #if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
     if (cpi->oxcf.error_resilient_mode)
     {

vp8/encoder/pickinter.c

@@ -1168,6 +1168,7 @@ void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
 #if CONFIG_TEMPORAL_DENOISING
     if (cpi->oxcf.noise_sensitivity)
     {
+        int block_index = mb_row * cpi->common.mb_cols + mb_col;
         if (x->best_sse_inter_mode == DC_PRED)
         {
             /* No best MV found. */
@@ -1179,7 +1180,9 @@ void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
         }
         x->increase_denoising = 0;
         vp8_denoiser_denoise_mb(&cpi->denoiser, x, best_sse, zero_mv_sse,
-                                recon_yoffset, recon_uvoffset);
+                                recon_yoffset, recon_uvoffset,
+                                &cpi->common.lf_info, mb_row, mb_col,
+                                block_index);
 
 
         /* Reevaluate ZEROMV after denoising. */

vp8/encoder/rdopt.c

@@ -1935,7 +1935,8 @@ static void update_best_mode(BEST_MODE* best_mode, int this_rd,
 
 void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
                             int recon_uvoffset, int *returnrate,
-                            int *returndistortion, int *returnintra)
+                            int *returndistortion, int *returnintra,
+                            int mb_row, int mb_col)
 {
     BLOCK *b = &x->block[0];
     BLOCKD *d = &x->e_mbd.block[0];
@@ -2510,6 +2511,7 @@ void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
 #if CONFIG_TEMPORAL_DENOISING
     if (cpi->oxcf.noise_sensitivity)
     {
+        int block_index = mb_row * cpi->common.mb_cols + mb_col;
         if (x->best_sse_inter_mode == DC_PRED)
         {
             /* No best MV found. */
@@ -2520,7 +2522,9 @@ void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
             best_sse = best_rd_sse;
         }
         vp8_denoiser_denoise_mb(&cpi->denoiser, x, best_sse, zero_mv_sse,
-                                recon_yoffset, recon_uvoffset);
+                                recon_yoffset, recon_uvoffset,
+                                &cpi->common.lf_info, mb_row, mb_col,
+                                block_index);
 
 
         /* Reevaluate ZEROMV after denoising. */

vp8/encoder/rdopt.h

@@ -70,7 +70,10 @@ static void insertsortsad(int arr[],int idx[], int len)
 }
 
 extern void vp8_initialize_rd_consts(VP8_COMP *cpi, MACROBLOCK *x, int Qvalue);
-extern void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra);
+extern void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x,
+                                   int recon_yoffset, int recon_uvoffset,
+                                   int *returnrate, int *returndistortion,
+                                   int *returnintra, int mb_row, int mb_col);
 extern void vp8_rd_pick_intra_mode(MACROBLOCK *x, int *rate);
 
 

vp8/encoder/x86/denoising_sse2.c

@@ -17,10 +17,23 @@
 #include <emmintrin.h>
 #include "vpx_ports/emmintrin_compat.h"
 
-union sum_union {
+/* Compute the sum of all pixel differences of this MB. */
-    __m128i v;
+static INLINE unsigned int abs_sum_diff_16x1(__m128i acc_diff) {
-    signed char e[16];
+  const __m128i k_1 = _mm_set1_epi16(1);
-};
+  const __m128i acc_diff_lo = _mm_srai_epi16(
+      _mm_unpacklo_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_hi = _mm_srai_epi16(
+      _mm_unpackhi_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
+  const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
+  const __m128i hgfe_dcba = _mm_add_epi32(hg_fe_dc_ba,
+                                          _mm_srli_si128(hg_fe_dc_ba, 8));
+  const __m128i hgfedcba = _mm_add_epi32(hgfe_dcba,
+                                         _mm_srli_si128(hgfe_dcba, 4));
+  unsigned int sum_diff = _mm_cvtsi128_si32(hgfedcba);
+
+  return abs(sum_diff);
+}
 
 int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
                              int mc_avg_y_stride,
@@ -31,7 +44,7 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
 {
     unsigned char *running_avg_y_start = running_avg_y;
     unsigned char *sig_start = sig;
-    int sum_diff_thresh;
+    unsigned int sum_diff_thresh;
     int r;
     int shift_inc  = (increase_denoising &&
         motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
@@ -103,16 +116,10 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
 
     {
         /* Compute the sum of all pixel differences of this MB. */
-        union sum_union s;
+        unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
-        int sum_diff = 0;
-        s.v = acc_diff;
-        sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5]
-                 + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11]
-                 + s.e[12] + s.e[13] + s.e[14] + s.e[15];
-
         sum_diff_thresh = SUM_DIFF_THRESHOLD;
         if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
-        if (abs(sum_diff) > sum_diff_thresh) {
+        if (abs_sum_diff > sum_diff_thresh) {
           // Before returning to copy the block (i.e., apply no denoising),
           // checK if we can still apply some (weaker) temporal filtering to
           // this block, that would otherwise not be denoised at all. Simplest
@@ -123,7 +130,7 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
 
           // The delta is set by the excess of absolute pixel diff over the
           // threshold.
-          int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+          int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
           // Only apply the adjustment for max delta up to 3.
           if (delta < 4) {
             const __m128i k_delta = _mm_set1_epi8(delta);
@@ -162,17 +169,10 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
              mc_running_avg_y += mc_avg_y_stride;
              running_avg_y += avg_y_stride;
             }
-            {
+            abs_sum_diff = abs_sum_diff_16x1(acc_diff);
-              // Update the sum of all pixel differences of this MB.
+            if (abs_sum_diff > sum_diff_thresh) {
-              union sum_union s;
-              s.v = acc_diff;
-              sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5]
-                       + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11]
-                       + s.e[12] + s.e[13] + s.e[14] + s.e[15];
-              if (abs(sum_diff) > sum_diff_thresh) {
               return COPY_BLOCK;
             }
-            }
           } else {
             return COPY_BLOCK;
           }
@@ -182,3 +182,198 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
     vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
     return FILTER_BLOCK;
 }
+
+int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg,
+                             int mc_avg_stride,
+                             unsigned char *running_avg, int avg_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    unsigned char *running_avg_start = running_avg;
+    unsigned char *sig_start = sig;
+    unsigned int sum_diff_thresh;
+    int r;
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 1 : 0;
+    __m128i acc_diff = _mm_setzero_si128();
+    const __m128i k_0 = _mm_setzero_si128();
+    const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+    const __m128i k_8 = _mm_set1_epi8(8);
+    const __m128i k_16 = _mm_set1_epi8(16);
+    /* Modify each level's adjustment according to motion_magnitude. */
+    const __m128i l3 = _mm_set1_epi8(
+                       (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ?
+                        7 + shift_inc : 6);
+    /* Difference between level 3 and level 2 is 2. */
+    const __m128i l32 = _mm_set1_epi8(2);
+    /* Difference between level 2 and level 1 is 1. */
+    const __m128i l21 = _mm_set1_epi8(1);
+
+    {
+      const __m128i k_1 = _mm_set1_epi16(1);
+      __m128i vec_sum_block = _mm_setzero_si128();
+
+      // Avoid denoising color signal if its close to average level.
+      for (r = 0; r < 8; ++r) {
+        const __m128i v_sig = _mm_loadl_epi64((__m128i *)(&sig[0]));
+        const __m128i v_sig_unpack = _mm_unpacklo_epi8(v_sig, k_0);
+        vec_sum_block = _mm_add_epi16(vec_sum_block, v_sig_unpack);
+        sig += sig_stride;
+      }
+      sig -= sig_stride * 8;
+      {
+        const __m128i hg_fe_dc_ba = _mm_madd_epi16(vec_sum_block, k_1);
+        const __m128i hgfe_dcba = _mm_add_epi32(hg_fe_dc_ba,
+                                                _mm_srli_si128(hg_fe_dc_ba, 8));
+        const __m128i hgfedcba = _mm_add_epi32(hgfe_dcba,
+                                               _mm_srli_si128(hgfe_dcba, 4));
+        const int sum_block = _mm_cvtsi128_si32(hgfedcba);
+        if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+          return COPY_BLOCK;
+        }
+      }
+    }
+
+    for (r = 0; r < 4; ++r) {
+        /* Calculate differences */
+        const __m128i v_sig_low = _mm_castpd_si128(
+            _mm_load_sd((double *)(&sig[0])));
+        const __m128i v_sig = _mm_castpd_si128(
+            _mm_loadh_pd(_mm_castsi128_pd(v_sig_low),
+                         (double *)(&sig[sig_stride])));
+        const __m128i v_mc_running_avg_low = _mm_castpd_si128(
+            _mm_load_sd((double *)(&mc_running_avg[0])));
+        const __m128i v_mc_running_avg = _mm_castpd_si128(
+            _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
+                         (double *)(&mc_running_avg[mc_avg_stride])));
+        const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
+        const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
+        /* Obtain the sign. FF if diff is negative. */
+        const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+        /* Clamp absolute difference to 16 to be used to get mask. Doing this
+         * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
+        const __m128i clamped_absdiff = _mm_min_epu8(
+                                        _mm_or_si128(pdiff, ndiff), k_16);
+        /* Get masks for l2 l1 and l0 adjustments */
+        const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
+        const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
+        const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
+        /* Get adjustments for l2, l1, and l0 */
+        __m128i adj2 = _mm_and_si128(mask2, l32);
+        const __m128i adj1 = _mm_and_si128(mask1, l21);
+        const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
+        __m128i adj,  padj, nadj;
+        __m128i v_running_avg;
+
+        /* Combine the adjustments and get absolute adjustments. */
+        adj2 = _mm_add_epi8(adj2, adj1);
+        adj = _mm_sub_epi8(l3, adj2);
+        adj = _mm_andnot_si128(mask0, adj);
+        adj = _mm_or_si128(adj, adj0);
+
+        /* Restore the sign and get positive and negative adjustments. */
+        padj = _mm_andnot_si128(diff_sign, adj);
+        nadj = _mm_and_si128(diff_sign, adj);
+
+        /* Calculate filtered value. */
+        v_running_avg = _mm_adds_epu8(v_sig, padj);
+        v_running_avg = _mm_subs_epu8(v_running_avg, nadj);
+
+        _mm_storel_pd((double *)&running_avg[0],
+                      _mm_castsi128_pd(v_running_avg));
+        _mm_storeh_pd((double *)&running_avg[avg_stride],
+                      _mm_castsi128_pd(v_running_avg));
+
+        /* Adjustments <=7, and each element in acc_diff can fit in signed
+         * char.
+         */
+        acc_diff = _mm_adds_epi8(acc_diff, padj);
+        acc_diff = _mm_subs_epi8(acc_diff, nadj);
+
+        /* Update pointers for next iteration. */
+        sig += sig_stride * 2;
+        mc_running_avg += mc_avg_stride * 2;
+        running_avg += avg_stride * 2;
+    }
+
+    {
+        unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+        sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+        if (abs_sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // checK if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the accceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const __m128i k_delta = _mm_set1_epi8(delta);
+            sig -= sig_stride * 8;
+            mc_running_avg -= mc_avg_stride * 8;
+            running_avg -= avg_stride * 8;
+            for (r = 0; r < 4; ++r) {
+              // Calculate differences.
+              const __m128i v_sig_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&sig[0])));
+              const __m128i v_sig = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_sig_low),
+                               (double *)(&sig[sig_stride])));
+              const __m128i v_mc_running_avg_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&mc_running_avg[0])));
+              const __m128i v_mc_running_avg = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
+                               (double *)(&mc_running_avg[mc_avg_stride])));
+              const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
+              const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
+              // Obtain the sign. FF if diff is negative.
+              const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+              // Clamp absolute difference to delta to get the adjustment.
+              const __m128i adj =
+                  _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+              // Restore the sign and get positive and negative adjustments.
+              __m128i padj, nadj;
+              const __m128i v_running_avg_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&running_avg[0])));
+              __m128i v_running_avg = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_running_avg_low),
+                               (double *)(&running_avg[avg_stride])));
+              padj = _mm_andnot_si128(diff_sign, adj);
+              nadj = _mm_and_si128(diff_sign, adj);
+              // Calculate filtered value.
+              v_running_avg = _mm_subs_epu8(v_running_avg, padj);
+              v_running_avg = _mm_adds_epu8(v_running_avg, nadj);
+
+              _mm_storel_pd((double *)&running_avg[0],
+                            _mm_castsi128_pd(v_running_avg));
+              _mm_storeh_pd((double *)&running_avg[avg_stride],
+                            _mm_castsi128_pd(v_running_avg));
+
+             // Accumulate the adjustments.
+             acc_diff = _mm_subs_epi8(acc_diff, padj);
+             acc_diff = _mm_adds_epi8(acc_diff, nadj);
+
+             // Update pointers for next iteration.
+             sig += sig_stride * 2;
+             mc_running_avg += mc_avg_stride * 2;
+             running_avg += avg_stride * 2;
+            }
+            abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+            if (abs_sum_diff > sum_diff_thresh) {
+              return COPY_BLOCK;
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
+
+    vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride);
+    return FILTER_BLOCK;
+}

vp8/encoder/x86/quantize_sse2.c

@@ -26,11 +26,10 @@
         int cmp = (x[z] < boost) | (y[z] == 0); \
         zbin_boost_ptr++; \
         if (cmp) \
-            goto select_eob_end_##i; \
+            break; \
         qcoeff_ptr[z] = y[z]; \
         eob = i; \
         zbin_boost_ptr = b->zrun_zbin_boost; \
-        select_eob_end_##i:; \
     } while (0)
 
 void vp8_regular_quantize_b_sse2(BLOCK *b, BLOCKD *d)

vp8/encoder/x86/quantize_sse4.asm

@@ -1,256 +0,0 @@
-;
-;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-;  Use of this source code is governed by a BSD-style license and patent
-;  grant that can be found in the LICENSE file in the root of the source
-;  tree. All contributing project authors may be found in the AUTHORS
-;  file in the root of the source tree.
-;
-
-
-%include "vpx_ports/x86_abi_support.asm"
-%include "vp8_asm_enc_offsets.asm"
-
-
-; void vp8_regular_quantize_b_sse4 | arg
-;  (BLOCK  *b,                     |  0
-;   BLOCKD *d)                     |  1
-
-global sym(vp8_regular_quantize_b_sse4) PRIVATE
-sym(vp8_regular_quantize_b_sse4):
-
-%if ABI_IS_32BIT
-    push        rbp
-    mov         rbp, rsp
-    GET_GOT     rbx
-    push        rdi
-    push        rsi
-
-    ALIGN_STACK 16, rax
-    %define qcoeff      0 ; 32
-    %define stack_size 32
-    sub         rsp, stack_size
-%else
-  %if LIBVPX_YASM_WIN64
-    SAVE_XMM 8, u
-    push        rdi
-    push        rsi
-  %endif
-%endif
-    ; end prolog
-
-%if ABI_IS_32BIT
-    mov         rdi, arg(0)                 ; BLOCK *b
-    mov         rsi, arg(1)                 ; BLOCKD *d
-%else
-  %if LIBVPX_YASM_WIN64
-    mov         rdi, rcx                    ; BLOCK *b
-    mov         rsi, rdx                    ; BLOCKD *d
-  %else
-    ;mov         rdi, rdi                    ; BLOCK *b
-    ;mov         rsi, rsi                    ; BLOCKD *d
-  %endif
-%endif
-
-    mov         rax, [rdi + vp8_block_coeff]
-    mov         rcx, [rdi + vp8_block_zbin]
-    mov         rdx, [rdi + vp8_block_round]
-    movd        xmm7, [rdi + vp8_block_zbin_extra]
-
-    ; z
-    movdqa      xmm0, [rax]
-    movdqa      xmm1, [rax + 16]
-
-    ; duplicate zbin_oq_value
-    pshuflw     xmm7, xmm7, 0
-    punpcklwd   xmm7, xmm7
-
-    movdqa      xmm2, xmm0
-    movdqa      xmm3, xmm1
-
-    ; sz
-    psraw       xmm0, 15
-    psraw       xmm1, 15
-
-    ; (z ^ sz)
-    pxor        xmm2, xmm0
-    pxor        xmm3, xmm1
-
-    ; x = abs(z)
-    psubw       xmm2, xmm0
-    psubw       xmm3, xmm1
-
-    ; zbin
-    movdqa      xmm4, [rcx]
-    movdqa      xmm5, [rcx + 16]
-
-    ; *zbin_ptr + zbin_oq_value
-    paddw       xmm4, xmm7
-    paddw       xmm5, xmm7
-
-    movdqa      xmm6, xmm2
-    movdqa      xmm7, xmm3
-
-    ; x - (*zbin_ptr + zbin_oq_value)
-    psubw       xmm6, xmm4
-    psubw       xmm7, xmm5
-
-    ; round
-    movdqa      xmm4, [rdx]
-    movdqa      xmm5, [rdx + 16]
-
-    mov         rax, [rdi + vp8_block_quant_shift]
-    mov         rcx, [rdi + vp8_block_quant]
-    mov         rdx, [rdi + vp8_block_zrun_zbin_boost]
-
-    ; x + round
-    paddw       xmm2, xmm4
-    paddw       xmm3, xmm5
-
-    ; quant
-    movdqa      xmm4, [rcx]
-    movdqa      xmm5, [rcx + 16]
-
-    ; y = x * quant_ptr >> 16
-    pmulhw      xmm4, xmm2
-    pmulhw      xmm5, xmm3
-
-    ; y += x
-    paddw       xmm2, xmm4
-    paddw       xmm3, xmm5
-
-    pxor        xmm4, xmm4
-%if ABI_IS_32BIT
-    movdqa      [rsp + qcoeff], xmm4
-    movdqa      [rsp + qcoeff + 16], xmm4
-%else
-    pxor        xmm8, xmm8
-%endif
-
-    ; quant_shift
-    movdqa      xmm5, [rax]
-
-    ; zrun_zbin_boost
-    mov         rax, rdx
-
-%macro ZIGZAG_LOOP 5
-    ; x
-    pextrw      ecx, %4, %2
-
-    ; if (x >= zbin)
-    sub         cx, WORD PTR[rdx]           ; x - zbin
-    lea         rdx, [rdx + 2]              ; zbin_boost_ptr++
-    jl          .rq_zigzag_loop_%1          ; x < zbin
-
-    pextrw      edi, %3, %2                 ; y
-
-    ; downshift by quant_shift[rc]
-    pextrb      ecx, xmm5, %1               ; quant_shift[rc]
-    sar         edi, cl                     ; also sets Z bit
-    je          .rq_zigzag_loop_%1          ; !y
-%if ABI_IS_32BIT
-    mov         WORD PTR[rsp + qcoeff + %1 *2], di
-%else
-    pinsrw      %5, edi, %2                 ; qcoeff[rc]
-%endif
-    mov         rdx, rax                    ; reset to b->zrun_zbin_boost
-.rq_zigzag_loop_%1:
-%endmacro
-; in vp8_default_zig_zag1d order: see vp8/common/entropy.c
-ZIGZAG_LOOP  0, 0, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  1, 1, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  4, 4, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  8, 0, xmm3, xmm7, xmm8
-ZIGZAG_LOOP  5, 5, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  2, 2, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  3, 3, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  6, 6, xmm2, xmm6, xmm4
-ZIGZAG_LOOP  9, 1, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 12, 4, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 13, 5, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 10, 2, xmm3, xmm7, xmm8
-ZIGZAG_LOOP  7, 7, xmm2, xmm6, xmm4
-ZIGZAG_LOOP 11, 3, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 14, 6, xmm3, xmm7, xmm8
-ZIGZAG_LOOP 15, 7, xmm3, xmm7, xmm8
-
-    mov         rcx, [rsi + vp8_blockd_dequant]
-    mov         rdi, [rsi + vp8_blockd_dqcoeff]
-
-%if ABI_IS_32BIT
-    movdqa      xmm4, [rsp + qcoeff]
-    movdqa      xmm5, [rsp + qcoeff + 16]
-%else
-    %define     xmm5 xmm8
-%endif
-
-    ; y ^ sz
-    pxor        xmm4, xmm0
-    pxor        xmm5, xmm1
-    ; x = (y ^ sz) - sz
-    psubw       xmm4, xmm0
-    psubw       xmm5, xmm1
-
-    ; dequant
-    movdqa      xmm0, [rcx]
-    movdqa      xmm1, [rcx + 16]
-
-    mov         rcx, [rsi + vp8_blockd_qcoeff]
-
-    pmullw      xmm0, xmm4
-    pmullw      xmm1, xmm5
-
-    ; store qcoeff
-    movdqa      [rcx], xmm4
-    movdqa      [rcx + 16], xmm5
-
-    ; store dqcoeff
-    movdqa      [rdi], xmm0
-    movdqa      [rdi + 16], xmm1
-
-    mov         rcx, [rsi + vp8_blockd_eob]
-
-    ; select the last value (in zig_zag order) for EOB
-    pxor        xmm6, xmm6
-    pcmpeqw     xmm4, xmm6
-    pcmpeqw     xmm5, xmm6
-
-    packsswb    xmm4, xmm5
-    pshufb      xmm4, [GLOBAL(zig_zag1d)]
-    pmovmskb    edx, xmm4
-    xor         rdi, rdi
-    mov         eax, -1
-    xor         dx, ax
-    bsr         eax, edx
-    sub         edi, edx
-    sar         edi, 31
-    add         eax, 1
-    and         eax, edi
-
-    mov         BYTE PTR [rcx], al          ; store eob
-
-    ; begin epilog
-%if ABI_IS_32BIT
-    add         rsp, stack_size
-    pop         rsp
-
-    pop         rsi
-    pop         rdi
-    RESTORE_GOT
-    pop         rbp
-%else
-  %undef xmm5
-  %if LIBVPX_YASM_WIN64
-    pop         rsi
-    pop         rdi
-    RESTORE_XMM
-  %endif
-%endif
-
-    ret
-
-SECTION_RODATA
-align 16
-; vp8/common/entropy.c: vp8_default_zig_zag1d
-zig_zag1d:
-    db 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15

vp8/encoder/x86/quantize_sse4.c

@@ -0,0 +1,128 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+
+
+#include <smmintrin.h> /* SSE4.1 */
+
+#include "./vp8_rtcd.h"
+#include "vp8/encoder/block.h"
+#include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */
+
+#define SELECT_EOB(i, z, x, y, q) \
+    do { \
+        short boost = *zbin_boost_ptr; \
+        short x_z = _mm_extract_epi16(x, z); \
+        short y_z = _mm_extract_epi16(y, z); \
+        int cmp = (x_z < boost) | (y_z == 0); \
+        zbin_boost_ptr++; \
+        if (cmp) \
+            break; \
+        q = _mm_insert_epi16(q, y_z, z); \
+        eob = i; \
+        zbin_boost_ptr = b->zrun_zbin_boost; \
+    } while (0)
+
+void vp8_regular_quantize_b_sse4_1(BLOCK *b, BLOCKD *d) {
+    char eob = 0;
+    short *zbin_boost_ptr  = b->zrun_zbin_boost;
+
+    __m128i sz0, x0, sz1, x1, y0, y1, x_minus_zbin0, x_minus_zbin1,
+            dqcoeff0, dqcoeff1;
+    __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
+    __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
+    __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
+    __m128i z1 = _mm_load_si128((__m128i *)(b->coeff+8));
+    __m128i zbin_extra = _mm_cvtsi32_si128(b->zbin_extra);
+    __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
+    __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
+    __m128i round0 = _mm_load_si128((__m128i *)(b->round));
+    __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
+    __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
+    __m128i quant1 = _mm_load_si128((__m128i *)(b->quant + 8));
+    __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
+    __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
+    __m128i qcoeff0 = _mm_setzero_si128();
+    __m128i qcoeff1 = _mm_setzero_si128();
+
+    /* Duplicate to all lanes. */
+    zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0);
+    zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra);
+
+    /* Sign of z: z >> 15 */
+    sz0 = _mm_srai_epi16(z0, 15);
+    sz1 = _mm_srai_epi16(z1, 15);
+
+    /* x = abs(z): (z ^ sz) - sz */
+    x0 = _mm_xor_si128(z0, sz0);
+    x1 = _mm_xor_si128(z1, sz1);
+    x0 = _mm_sub_epi16(x0, sz0);
+    x1 = _mm_sub_epi16(x1, sz1);
+
+    /* zbin[] + zbin_extra */
+    zbin0 = _mm_add_epi16(zbin0, zbin_extra);
+    zbin1 = _mm_add_epi16(zbin1, zbin_extra);
+
+    /* In C x is compared to zbin where zbin = zbin[] + boost + extra. Rebalance
+     * the equation because boost is the only value which can change:
+     * x - (zbin[] + extra) >= boost */
+    x_minus_zbin0 = _mm_sub_epi16(x0, zbin0);
+    x_minus_zbin1 = _mm_sub_epi16(x1, zbin1);
+
+    /* All the remaining calculations are valid whether they are done now with
+     * simd or later inside the loop one at a time. */
+    x0 = _mm_add_epi16(x0, round0);
+    x1 = _mm_add_epi16(x1, round1);
+
+    y0 = _mm_mulhi_epi16(x0, quant0);
+    y1 = _mm_mulhi_epi16(x1, quant1);
+
+    y0 = _mm_add_epi16(y0, x0);
+    y1 = _mm_add_epi16(y1, x1);
+
+    /* Instead of shifting each value independently we convert the scaling
+     * factor with 1 << (16 - shift) so we can use multiply/return high half. */
+    y0 = _mm_mulhi_epi16(y0, quant_shift0);
+    y1 = _mm_mulhi_epi16(y1, quant_shift1);
+
+    /* Return the sign: (y ^ sz) - sz */
+    y0 = _mm_xor_si128(y0, sz0);
+    y1 = _mm_xor_si128(y1, sz1);
+    y0 = _mm_sub_epi16(y0, sz0);
+    y1 = _mm_sub_epi16(y1, sz1);
+
+    /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */
+    SELECT_EOB(1, 0, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(2, 1, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(3, 4, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(4, 0, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(5, 5, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(6, 2, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(7, 3, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(8, 6, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(9, 1, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(10, 4, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(11, 5, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(12, 2, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(13, 7, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(14, 3, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(15, 6, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(16, 7, x_minus_zbin1, y1, qcoeff1);
+
+    _mm_store_si128((__m128i *)(d->qcoeff), qcoeff0);
+    _mm_store_si128((__m128i *)(d->qcoeff + 8), qcoeff1);
+
+    dqcoeff0 = _mm_mullo_epi16(qcoeff0, dequant0);
+    dqcoeff1 = _mm_mullo_epi16(qcoeff1, dequant1);
+
+    _mm_store_si128((__m128i *)(d->dqcoeff), dqcoeff0);
+    _mm_store_si128((__m128i *)(d->dqcoeff + 8), dqcoeff1);
+
+    *d->eob = eob;
+}

vp8/vp8_common.mk

@@ -107,7 +107,6 @@ VP8_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/variance_impl_ssse3.asm
 VP8_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/sad_sse4.asm
 
 ifeq ($(CONFIG_POSTPROC),yes)
-VP8_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/postproc_x86.c
 VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/postproc_mmx.asm
 VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/mfqe_sse2.asm
 VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/postproc_sse2.asm

vp8/vp8cx.mk

@@ -89,6 +89,7 @@ VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/fwalsh_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/quantize_sse2.c
 VP8_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/quantize_ssse3.c
+VP8_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/quantize_sse4.c
 
 ifeq ($(CONFIG_TEMPORAL_DENOISING),yes)
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/denoising_sse2.c
@@ -97,7 +98,6 @@ endif
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/subtract_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/temporal_filter_apply_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp8_enc_stubs_sse2.c
-VP8_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/quantize_sse4.asm
 VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/quantize_mmx.asm
 VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/encodeopt.asm
 VP8_CX_SRCS-$(ARCH_X86_64) += encoder/x86/ssim_opt_x86_64.asm

vp9/common/arm/neon/vp9_convolve_neon.c

@@ -25,12 +25,14 @@ void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
   // Account for the vertical phase needing 3 lines prior and 4 lines post
   int intermediate_height = h + 7;
 
-  if (x_step_q4 != 16 || y_step_q4 != 16)
+  if (x_step_q4 != 16 || y_step_q4 != 16) {
-    return vp9_convolve8_c(src, src_stride,
+    vp9_convolve8_c(src, src_stride,
                     dst, dst_stride,
                     filter_x, x_step_q4,
                     filter_y, y_step_q4,
                     w, h);
+    return;
+  }
 
   /* Filter starting 3 lines back. The neon implementation will ignore the
    * given height and filter a multiple of 4 lines. Since this goes in to
@@ -57,12 +59,14 @@ void vp9_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
   DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
   int intermediate_height = h + 7;
 
-  if (x_step_q4 != 16 || y_step_q4 != 16)
+  if (x_step_q4 != 16 || y_step_q4 != 16) {
-    return vp9_convolve8_avg_c(src, src_stride,
+    vp9_convolve8_avg_c(src, src_stride,
                         dst, dst_stride,
                         filter_x, x_step_q4,
                         filter_y, y_step_q4,
                         w, h);
+    return;
+  }
 
   /* This implementation has the same issues as above. In addition, we only want
    * to average the values after both passes.

vp9/common/arm/neon/vp9_loopfilter_16_neon.c

@@ -9,6 +9,7 @@
  */
 
 #include "./vp9_rtcd.h"
+#include "vpx/vpx_integer.h"
 
 void vp9_lpf_horizontal_8_dual_neon(uint8_t *s, int p /* pitch */,
                                     const uint8_t *blimit0,

vp9/common/vp9_alloccommon.c

@@ -97,19 +97,22 @@ static void free_mi(VP9_COMMON *cm) {
 
 void vp9_free_frame_buffers(VP9_COMMON *cm) {
   int i;
+  BufferPool *const pool = cm->buffer_pool;
 
   for (i = 0; i < FRAME_BUFFERS; ++i) {
-    vp9_free_frame_buffer(&cm->frame_bufs[i].buf);
+    vp9_free_frame_buffer(&pool->frame_bufs[i].buf);
 
-    if (cm->frame_bufs[i].ref_count > 0 &&
+    if (pool->frame_bufs[i].ref_count > 0 &&
-        cm->frame_bufs[i].raw_frame_buffer.data != NULL) {
+        pool->frame_bufs[i].raw_frame_buffer.data != NULL) {
-      cm->release_fb_cb(cm->cb_priv, &cm->frame_bufs[i].raw_frame_buffer);
+      pool->release_fb_cb(pool->cb_priv, &pool->frame_bufs[i].raw_frame_buffer);
-      cm->frame_bufs[i].ref_count = 0;
+      pool->frame_bufs[i].ref_count = 0;
     }
   }
 
   vp9_free_frame_buffer(&cm->post_proc_buffer);
+}
 
+void vp9_free_context_buffers(VP9_COMMON *cm) {
   free_mi(cm);
 
   vpx_free(cm->last_frame_seg_map);
@@ -125,12 +128,15 @@ void vp9_free_frame_buffers(VP9_COMMON *cm) {
 int vp9_resize_frame_buffers(VP9_COMMON *cm, int width, int height) {
   const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
   const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
+#if CONFIG_INTERNAL_STATS || CONFIG_VP9_POSTPROC
   const int ss_x = cm->subsampling_x;
   const int ss_y = cm->subsampling_y;
 
+  // TODO(agrange): this should be conditionally allocated.
   if (vp9_realloc_frame_buffer(&cm->post_proc_buffer, width, height, ss_x, ss_y,
                                VP9_DEC_BORDER_IN_PIXELS, NULL, NULL, NULL) < 0)
     goto fail;
+#endif
 
   set_mb_mi(cm, aligned_width, aligned_height);
 
@@ -165,36 +171,58 @@ int vp9_resize_frame_buffers(VP9_COMMON *cm, int width, int height) {
 
  fail:
   vp9_free_frame_buffers(cm);
+  vp9_free_context_buffers(cm);
   return 1;
 }
 
+static void init_frame_bufs(VP9_COMMON *cm) {
+  int i;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+
+  cm->new_fb_idx = FRAME_BUFFERS - 1;
+  frame_bufs[cm->new_fb_idx].ref_count = 1;
+
+  for (i = 0; i < REF_FRAMES; ++i) {
+    cm->ref_frame_map[i] = i;
+    frame_bufs[i].ref_count = 1;
+  }
+}
+
 int vp9_alloc_frame_buffers(VP9_COMMON *cm, int width, int height) {
-  const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
+  int i;
-  const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
   const int ss_x = cm->subsampling_x;
   const int ss_y = cm->subsampling_y;
-  int i;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   vp9_free_frame_buffers(cm);
 
-  for (i = 0; i < FRAME_BUFFERS; i++) {
+  for (i = 0; i < FRAME_BUFFERS; ++i) {
-    cm->frame_bufs[i].ref_count = 0;
+    frame_bufs[i].ref_count = 0;
-    if (vp9_alloc_frame_buffer(&cm->frame_bufs[i].buf, width, height,
+    if (vp9_alloc_frame_buffer(&frame_bufs[i].buf, width, height,
                                ss_x, ss_y, VP9_ENC_BORDER_IN_PIXELS) < 0)
       goto fail;
   }
 
-  cm->new_fb_idx = FRAME_BUFFERS - 1;
+  init_frame_bufs(cm);
-  cm->frame_bufs[cm->new_fb_idx].ref_count = 1;
-
-  for (i = 0; i < REF_FRAMES; i++) {
-    cm->ref_frame_map[i] = i;
-    cm->frame_bufs[i].ref_count = 1;
-  }
 
+#if CONFIG_INTERNAL_STATS || CONFIG_VP9_POSTPROC
   if (vp9_alloc_frame_buffer(&cm->post_proc_buffer, width, height, ss_x, ss_y,
                              VP9_ENC_BORDER_IN_PIXELS) < 0)
     goto fail;
+#endif
+
+  return 0;
+
+ fail:
+  vp9_free_frame_buffers(cm);
+  return 1;
+}
+
+int vp9_alloc_context_buffers(VP9_COMMON *cm, int width, int height) {
+  const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
+  const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
+
+  vp9_free_context_buffers(cm);
 
   set_mb_mi(cm, aligned_width, aligned_height);
 
@@ -224,13 +252,14 @@ int vp9_alloc_frame_buffers(VP9_COMMON *cm, int width, int height) {
   return 0;
 
  fail:
-  vp9_free_frame_buffers(cm);
+  vp9_free_context_buffers(cm);
   return 1;
 }
 
 void vp9_remove_common(VP9_COMMON *cm) {
   vp9_free_frame_buffers(cm);
-  vp9_free_internal_frame_buffers(&cm->int_frame_buffers);
+  vp9_free_context_buffers(cm);
+  vp9_free_internal_frame_buffers(&cm->buffer_pool->int_frame_buffers);
 }
 
 void vp9_update_frame_size(VP9_COMMON *cm) {

vp9/common/vp9_alloccommon.h

@@ -23,8 +23,12 @@ void vp9_remove_common(struct VP9Common *cm);
 int vp9_resize_frame_buffers(struct VP9Common *cm, int width, int height);
 
 int vp9_alloc_frame_buffers(struct VP9Common *cm, int width, int height);
+int vp9_alloc_state_buffers(struct VP9Common *cm, int width, int height);
+int vp9_alloc_context_buffers(struct VP9Common *cm, int width, int height);
 
 void vp9_free_frame_buffers(struct VP9Common *cm);
+void vp9_free_state_buffers(struct VP9Common *cm);
+void vp9_free_context_buffers(struct VP9Common *cm);
 
 void vp9_update_frame_size(struct VP9Common *cm);
 

vp9/common/vp9_blockd.c

@@ -44,7 +44,7 @@ void vp9_foreach_transformed_block_in_plane(
   // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
   // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
   // transform size varies per plane, look it up in a common way.
-  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi)
+  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi, pd)
                                 : mbmi->tx_size;
   const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
   const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];

vp9/common/vp9_blockd.h

@@ -270,18 +270,20 @@ static INLINE TX_TYPE get_tx_type_4x4(PLANE_TYPE plane_type,
 
 void vp9_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y);
 
-static INLINE TX_SIZE get_uv_tx_size_impl(TX_SIZE y_tx_size, BLOCK_SIZE bsize) {
+static INLINE TX_SIZE get_uv_tx_size_impl(TX_SIZE y_tx_size, BLOCK_SIZE bsize,
+                                          int xss, int yss) {
   if (bsize < BLOCK_8X8) {
     return TX_4X4;
   } else {
-    // TODO(dkovalev): Assuming YUV420 (ss_x == 1, ss_y == 1)
+    const BLOCK_SIZE plane_bsize = ss_size_lookup[bsize][xss][yss];
-    const BLOCK_SIZE plane_bsize = ss_size_lookup[bsize][1][1];
     return MIN(y_tx_size, max_txsize_lookup[plane_bsize]);
   }
 }
 
-static INLINE TX_SIZE get_uv_tx_size(const MB_MODE_INFO *mbmi) {
+static INLINE TX_SIZE get_uv_tx_size(const MB_MODE_INFO *mbmi,
-  return get_uv_tx_size_impl(mbmi->tx_size, mbmi->sb_type);
+                                     const struct macroblockd_plane *pd) {
+  return get_uv_tx_size_impl(mbmi->tx_size, mbmi->sb_type, pd->subsampling_x,
+                             pd->subsampling_y);
 }
 
 static INLINE BLOCK_SIZE get_plane_block_size(BLOCK_SIZE bsize,

vp9/common/vp9_convolve.c

@@ -117,17 +117,25 @@ static void convolve(const uint8_t *src, ptrdiff_t src_stride,
                      const InterpKernel *const y_filters,
                      int y0_q4, int y_step_q4,
                      int w, int h) {
-  // Fixed size intermediate buffer places limits on parameters.
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
-  // Maximum intermediate_height is 324, for y_step_q4 == 80,
+  // 2d filtering proceeds in 2 steps:
-  // h == 64, taps == 8.
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
-  // y_step_q4 of 80 allows for 1/10 scale for 5 layer svc
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
-  uint8_t temp[64 * 324];
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint8_t temp[135 * 64];
   int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + SUBPEL_TAPS;
 
   assert(w <= 64);
   assert(h <= 64);
-  assert(y_step_q4 <= 80);
+  assert(y_step_q4 <= 32);
-  assert(x_step_q4 <= 80);
+  assert(x_step_q4 <= 32);
 
   if (intermediate_height < h)
     intermediate_height = h;

vp9/common/vp9_frame_buffers.c

@@ -76,6 +76,7 @@ int vp9_get_frame_buffer(void *cb_priv, size_t min_size,
 int vp9_release_frame_buffer(void *cb_priv, vpx_codec_frame_buffer_t *fb) {
   InternalFrameBuffer *const int_fb = (InternalFrameBuffer *)fb->priv;
   (void)cb_priv;
+  if (int_fb)
     int_fb->in_use = 0;
   return 0;
 }

vp9/common/vp9_loopfilter.c

@@ -502,7 +502,7 @@ static void build_masks(const loop_filter_info_n *const lfi_n,
   const MB_MODE_INFO *mbmi = &mi->mbmi;
   const BLOCK_SIZE block_size = mbmi->sb_type;
   const TX_SIZE tx_size_y = mbmi->tx_size;
-  const TX_SIZE tx_size_uv = get_uv_tx_size(mbmi);
+  const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1);
   const int filter_level = get_filter_level(lfi_n, mbmi);
   uint64_t *const left_y = &lfm->left_y[tx_size_y];
   uint64_t *const above_y = &lfm->above_y[tx_size_y];
@@ -939,7 +939,7 @@ static void filter_block_plane_non420(VP9_COMMON *cm,
           !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
       const int skip_this_r = skip_this && !block_edge_above;
       const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
-                            ? get_uv_tx_size(&mi[0].mbmi)
+                            ? get_uv_tx_size(&mi[0].mbmi, plane)
                             : mi[0].mbmi.tx_size;
       const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
       const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;

vp9/common/vp9_mvref_common.c

@@ -11,181 +11,6 @@
 
 #include "vp9/common/vp9_mvref_common.h"
 
-#define MVREF_NEIGHBOURS 8
-
-typedef struct position {
-  int row;
-  int col;
-} POSITION;
-
-typedef enum {
-  BOTH_ZERO = 0,
-  ZERO_PLUS_PREDICTED = 1,
-  BOTH_PREDICTED = 2,
-  NEW_PLUS_NON_INTRA = 3,
-  BOTH_NEW = 4,
-  INTRA_PLUS_NON_INTRA = 5,
-  BOTH_INTRA = 6,
-  INVALID_CASE = 9
-} motion_vector_context;
-
-// This is used to figure out a context for the ref blocks. The code flattens
-// an array that would have 3 possible counts (0, 1 & 2) for 3 choices by
-// adding 9 for each intra block, 3 for each zero mv and 1 for each new
-// motion vector. This single number is then converted into a context
-// with a single lookup ( counter_to_context ).
-static const int mode_2_counter[MB_MODE_COUNT] = {
-  9,  // DC_PRED
-  9,  // V_PRED
-  9,  // H_PRED
-  9,  // D45_PRED
-  9,  // D135_PRED
-  9,  // D117_PRED
-  9,  // D153_PRED
-  9,  // D207_PRED
-  9,  // D63_PRED
-  9,  // TM_PRED
-  0,  // NEARESTMV
-  0,  // NEARMV
-  3,  // ZEROMV
-  1,  // NEWMV
-};
-
-// There are 3^3 different combinations of 3 counts that can be either 0,1 or
-// 2. However the actual count can never be greater than 2 so the highest
-// counter we need is 18. 9 is an invalid counter that's never used.
-static const int counter_to_context[19] = {
-  BOTH_PREDICTED,  // 0
-  NEW_PLUS_NON_INTRA,  // 1
-  BOTH_NEW,  // 2
-  ZERO_PLUS_PREDICTED,  // 3
-  NEW_PLUS_NON_INTRA,  // 4
-  INVALID_CASE,  // 5
-  BOTH_ZERO,  // 6
-  INVALID_CASE,  // 7
-  INVALID_CASE,  // 8
-  INTRA_PLUS_NON_INTRA,  // 9
-  INTRA_PLUS_NON_INTRA,  // 10
-  INVALID_CASE,  // 11
-  INTRA_PLUS_NON_INTRA,  // 12
-  INVALID_CASE,  // 13
-  INVALID_CASE,  // 14
-  INVALID_CASE,  // 15
-  INVALID_CASE,  // 16
-  INVALID_CASE,  // 17
-  BOTH_INTRA  // 18
-};
-
-static const POSITION mv_ref_blocks[BLOCK_SIZES][MVREF_NEIGHBOURS] = {
-  // 4X4
-  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
-  // 4X8
-  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
-  // 8X4
-  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
-  // 8X8
-  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
-  // 8X16
-  {{0, -1}, {-1, 0}, {1, -1}, {-1, -1}, {0, -2}, {-2, 0}, {-2, -1}, {-1, -2}},
-  // 16X8
-  {{-1, 0}, {0, -1}, {-1, 1}, {-1, -1}, {-2, 0}, {0, -2}, {-1, -2}, {-2, -1}},
-  // 16X16
-  {{-1, 0}, {0, -1}, {-1, 1}, {1, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
-  // 16X32
-  {{0, -1}, {-1, 0}, {2, -1}, {-1, -1}, {-1, 1}, {0, -3}, {-3, 0}, {-3, -3}},
-  // 32X16
-  {{-1, 0}, {0, -1}, {-1, 2}, {-1, -1}, {1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
-  // 32X32
-  {{-1, 1}, {1, -1}, {-1, 2}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
-  // 32X64
-  {{0, -1}, {-1, 0}, {4, -1}, {-1, 2}, {-1, -1}, {0, -3}, {-3, 0}, {2, -1}},
-  // 64X32
-  {{-1, 0}, {0, -1}, {-1, 4}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-1, 2}},
-  // 64X64
-  {{-1, 3}, {3, -1}, {-1, 4}, {4, -1}, {-1, -1}, {-1, 0}, {0, -1}, {-1, 6}}
-};
-
-static const int idx_n_column_to_subblock[4][2] = {
-  {1, 2},
-  {1, 3},
-  {3, 2},
-  {3, 3}
-};
-
-// clamp_mv_ref
-#define MV_BORDER (16 << 3)  // Allow 16 pels in 1/8th pel units
-
-static void clamp_mv_ref(MV *mv, const MACROBLOCKD *xd) {
-  clamp_mv(mv, xd->mb_to_left_edge - MV_BORDER,
-               xd->mb_to_right_edge + MV_BORDER,
-               xd->mb_to_top_edge - MV_BORDER,
-               xd->mb_to_bottom_edge + MV_BORDER);
-}
-
-// This function returns either the appropriate sub block or block's mv
-// on whether the block_size < 8x8 and we have check_sub_blocks set.
-static INLINE int_mv get_sub_block_mv(const MODE_INFO *candidate, int which_mv,
-                                      int search_col, int block_idx) {
-  return block_idx >= 0 && candidate->mbmi.sb_type < BLOCK_8X8
-          ? candidate->bmi[idx_n_column_to_subblock[block_idx][search_col == 0]]
-              .as_mv[which_mv]
-          : candidate->mbmi.mv[which_mv];
-}
-
-
-// Performs mv sign inversion if indicated by the reference frame combination.
-static INLINE int_mv scale_mv(const MB_MODE_INFO *mbmi, int ref,
-                              const MV_REFERENCE_FRAME this_ref_frame,
-                              const int *ref_sign_bias) {
-  int_mv mv = mbmi->mv[ref];
-  if (ref_sign_bias[mbmi->ref_frame[ref]] != ref_sign_bias[this_ref_frame]) {
-    mv.as_mv.row *= -1;
-    mv.as_mv.col *= -1;
-  }
-  return mv;
-}
-
-// This macro is used to add a motion vector mv_ref list if it isn't
-// already in the list.  If it's the second motion vector it will also
-// skip all additional processing and jump to done!
-#define ADD_MV_REF_LIST(mv) \
-  do { \
-    if (refmv_count) { \
-      if ((mv).as_int != mv_ref_list[0].as_int) { \
-        mv_ref_list[refmv_count] = (mv); \
-        goto Done; \
-      } \
-    } else { \
-      mv_ref_list[refmv_count++] = (mv); \
-    } \
-  } while (0)
-
-// If either reference frame is different, not INTRA, and they
-// are different from each other scale and add the mv to our list.
-#define IF_DIFF_REF_FRAME_ADD_MV(mbmi) \
-  do { \
-    if (is_inter_block(mbmi)) { \
-      if ((mbmi)->ref_frame[0] != ref_frame) \
-        ADD_MV_REF_LIST(scale_mv((mbmi), 0, ref_frame, ref_sign_bias)); \
-      if (has_second_ref(mbmi) && \
-          (mbmi)->ref_frame[1] != ref_frame && \
-          (mbmi)->mv[1].as_int != (mbmi)->mv[0].as_int) \
-        ADD_MV_REF_LIST(scale_mv((mbmi), 1, ref_frame, ref_sign_bias)); \
-    } \
-  } while (0)
-
-
-// Checks that the given mi_row, mi_col and search point
-// are inside the borders of the tile.
-static INLINE int is_inside(const TileInfo *const tile,
-                            int mi_col, int mi_row, int mi_rows,
-                            const POSITION *mi_pos) {
-  return !(mi_row + mi_pos->row < 0 ||
-           mi_col + mi_pos->col < tile->mi_col_start ||
-           mi_row + mi_pos->row >= mi_rows ||
-           mi_col + mi_pos->col >= tile->mi_col_end);
-}
-
 // This function searches the neighbourhood of a given MB/SB
 // to try and find candidate reference vectors.
 static void find_mv_refs_idx(const VP9_COMMON *cm, const MACROBLOCKD *xd,

vp9/common/vp9_mvref_common.h

@@ -21,6 +21,181 @@ extern "C" {
 #define RIGHT_BOTTOM_MARGIN ((VP9_ENC_BORDER_IN_PIXELS -\
                                 VP9_INTERP_EXTEND) << 3)
 
+#define MVREF_NEIGHBOURS 8
+
+typedef struct position {
+  int row;
+  int col;
+} POSITION;
+
+typedef enum {
+  BOTH_ZERO = 0,
+  ZERO_PLUS_PREDICTED = 1,
+  BOTH_PREDICTED = 2,
+  NEW_PLUS_NON_INTRA = 3,
+  BOTH_NEW = 4,
+  INTRA_PLUS_NON_INTRA = 5,
+  BOTH_INTRA = 6,
+  INVALID_CASE = 9
+} motion_vector_context;
+
+// This is used to figure out a context for the ref blocks. The code flattens
+// an array that would have 3 possible counts (0, 1 & 2) for 3 choices by
+// adding 9 for each intra block, 3 for each zero mv and 1 for each new
+// motion vector. This single number is then converted into a context
+// with a single lookup ( counter_to_context ).
+static const int mode_2_counter[MB_MODE_COUNT] = {
+  9,  // DC_PRED
+  9,  // V_PRED
+  9,  // H_PRED
+  9,  // D45_PRED
+  9,  // D135_PRED
+  9,  // D117_PRED
+  9,  // D153_PRED
+  9,  // D207_PRED
+  9,  // D63_PRED
+  9,  // TM_PRED
+  0,  // NEARESTMV
+  0,  // NEARMV
+  3,  // ZEROMV
+  1,  // NEWMV
+};
+
+// There are 3^3 different combinations of 3 counts that can be either 0,1 or
+// 2. However the actual count can never be greater than 2 so the highest
+// counter we need is 18. 9 is an invalid counter that's never used.
+static const int counter_to_context[19] = {
+  BOTH_PREDICTED,  // 0
+  NEW_PLUS_NON_INTRA,  // 1
+  BOTH_NEW,  // 2
+  ZERO_PLUS_PREDICTED,  // 3
+  NEW_PLUS_NON_INTRA,  // 4
+  INVALID_CASE,  // 5
+  BOTH_ZERO,  // 6
+  INVALID_CASE,  // 7
+  INVALID_CASE,  // 8
+  INTRA_PLUS_NON_INTRA,  // 9
+  INTRA_PLUS_NON_INTRA,  // 10
+  INVALID_CASE,  // 11
+  INTRA_PLUS_NON_INTRA,  // 12
+  INVALID_CASE,  // 13
+  INVALID_CASE,  // 14
+  INVALID_CASE,  // 15
+  INVALID_CASE,  // 16
+  INVALID_CASE,  // 17
+  BOTH_INTRA  // 18
+};
+
+static const POSITION mv_ref_blocks[BLOCK_SIZES][MVREF_NEIGHBOURS] = {
+  // 4X4
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 4X8
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X4
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X8
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X16
+  {{0, -1}, {-1, 0}, {1, -1}, {-1, -1}, {0, -2}, {-2, 0}, {-2, -1}, {-1, -2}},
+  // 16X8
+  {{-1, 0}, {0, -1}, {-1, 1}, {-1, -1}, {-2, 0}, {0, -2}, {-1, -2}, {-2, -1}},
+  // 16X16
+  {{-1, 0}, {0, -1}, {-1, 1}, {1, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 16X32
+  {{0, -1}, {-1, 0}, {2, -1}, {-1, -1}, {-1, 1}, {0, -3}, {-3, 0}, {-3, -3}},
+  // 32X16
+  {{-1, 0}, {0, -1}, {-1, 2}, {-1, -1}, {1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 32X32
+  {{-1, 1}, {1, -1}, {-1, 2}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 32X64
+  {{0, -1}, {-1, 0}, {4, -1}, {-1, 2}, {-1, -1}, {0, -3}, {-3, 0}, {2, -1}},
+  // 64X32
+  {{-1, 0}, {0, -1}, {-1, 4}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-1, 2}},
+  // 64X64
+  {{-1, 3}, {3, -1}, {-1, 4}, {4, -1}, {-1, -1}, {-1, 0}, {0, -1}, {-1, 6}}
+};
+
+static const int idx_n_column_to_subblock[4][2] = {
+  {1, 2},
+  {1, 3},
+  {3, 2},
+  {3, 3}
+};
+
+// clamp_mv_ref
+#define MV_BORDER (16 << 3)  // Allow 16 pels in 1/8th pel units
+
+static void clamp_mv_ref(MV *mv, const MACROBLOCKD *xd) {
+  clamp_mv(mv, xd->mb_to_left_edge - MV_BORDER,
+               xd->mb_to_right_edge + MV_BORDER,
+               xd->mb_to_top_edge - MV_BORDER,
+               xd->mb_to_bottom_edge + MV_BORDER);
+}
+
+// This function returns either the appropriate sub block or block's mv
+// on whether the block_size < 8x8 and we have check_sub_blocks set.
+static INLINE int_mv get_sub_block_mv(const MODE_INFO *candidate, int which_mv,
+                                      int search_col, int block_idx) {
+  return block_idx >= 0 && candidate->mbmi.sb_type < BLOCK_8X8
+          ? candidate->bmi[idx_n_column_to_subblock[block_idx][search_col == 0]]
+              .as_mv[which_mv]
+          : candidate->mbmi.mv[which_mv];
+}
+
+
+// Performs mv sign inversion if indicated by the reference frame combination.
+static INLINE int_mv scale_mv(const MB_MODE_INFO *mbmi, int ref,
+                              const MV_REFERENCE_FRAME this_ref_frame,
+                              const int *ref_sign_bias) {
+  int_mv mv = mbmi->mv[ref];
+  if (ref_sign_bias[mbmi->ref_frame[ref]] != ref_sign_bias[this_ref_frame]) {
+    mv.as_mv.row *= -1;
+    mv.as_mv.col *= -1;
+  }
+  return mv;
+}
+
+// This macro is used to add a motion vector mv_ref list if it isn't
+// already in the list.  If it's the second motion vector it will also
+// skip all additional processing and jump to done!
+#define ADD_MV_REF_LIST(mv) \
+  do { \
+    if (refmv_count) { \
+      if ((mv).as_int != mv_ref_list[0].as_int) { \
+        mv_ref_list[refmv_count] = (mv); \
+        goto Done; \
+      } \
+    } else { \
+      mv_ref_list[refmv_count++] = (mv); \
+    } \
+  } while (0)
+
+// If either reference frame is different, not INTRA, and they
+// are different from each other scale and add the mv to our list.
+#define IF_DIFF_REF_FRAME_ADD_MV(mbmi) \
+  do { \
+    if (is_inter_block(mbmi)) { \
+      if ((mbmi)->ref_frame[0] != ref_frame) \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 0, ref_frame, ref_sign_bias)); \
+      if (has_second_ref(mbmi) && \
+          (mbmi)->ref_frame[1] != ref_frame && \
+          (mbmi)->mv[1].as_int != (mbmi)->mv[0].as_int) \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 1, ref_frame, ref_sign_bias)); \
+    } \
+  } while (0)
+
+
+// Checks that the given mi_row, mi_col and search point
+// are inside the borders of the tile.
+static INLINE int is_inside(const TileInfo *const tile,
+                            int mi_col, int mi_row, int mi_rows,
+                            const POSITION *mi_pos) {
+  return !(mi_row + mi_pos->row < 0 ||
+           mi_col + mi_pos->col < tile->mi_col_start ||
+           mi_row + mi_pos->row >= mi_rows ||
+           mi_col + mi_pos->col >= tile->mi_col_end);
+}
+
 // TODO(jingning): this mv clamping function should be block size dependent.
 static INLINE void clamp_mv2(MV *mv, const MACROBLOCKD *xd) {
   clamp_mv(mv, xd->mb_to_left_edge - LEFT_TOP_MARGIN,

vp9/common/vp9_onyxc_int.h

@@ -11,6 +11,7 @@
 #ifndef VP9_COMMON_VP9_ONYXC_INT_H_
 #define VP9_COMMON_VP9_ONYXC_INT_H_
 
+#include <pthread.h>
 #include "./vpx_config.h"
 #include "vpx/internal/vpx_codec_internal.h"
 #include "./vp9_rtcd.h"
@@ -61,8 +62,40 @@ typedef struct {
   int ref_count;
   vpx_codec_frame_buffer_t raw_frame_buffer;
   YV12_BUFFER_CONFIG buf;
+
+  // The Following variables will only be used in frame parallel decode.
+
+  // owner_thread_id indicates which FrameWorker owns this buffer. -1 means
+  // that no FrameWorker owns, or is decoding, this buffer.
+  int owner_worker_id;
+
+  // Buffer has been decoded to (row, col) position. When first start decoding,
+  // they are reset to -1. If a frame has been fully decoded, row and col will
+  // be set to INT_MAX.
+  int row;
+  int col;
 } RefCntBuffer;
 
+typedef struct {
+  // Protect BufferPool from being accessed by several FrameWorkers at
+  // the same time during frame parallel decode.
+  // TODO(hkuang): Try to use atomic variable instead of locking the whole pool.
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t pool_mutex;
+#endif
+
+  // Private data associated with the frame buffer callbacks.
+  void *cb_priv;
+
+  vpx_get_frame_buffer_cb_fn_t get_fb_cb;
+  vpx_release_frame_buffer_cb_fn_t release_fb_cb;
+
+  RefCntBuffer frame_bufs[FRAME_BUFFERS];
+
+  // Handles memory for the codec.
+  InternalFrameBufferList int_frame_buffers;
+} BufferPool;
+
 typedef struct VP9Common {
   struct vpx_internal_error_info  error;
 
@@ -89,10 +122,11 @@ typedef struct VP9Common {
 
   YV12_BUFFER_CONFIG *frame_to_show;
 
-  RefCntBuffer frame_bufs[FRAME_BUFFERS];
-
   int ref_frame_map[REF_FRAMES]; /* maps fb_idx to reference slot */
 
+  // Prepare ref_frame_map for next frame. Only used in frame parallel decode.
+  int next_ref_frame_map[REF_FRAMES];
+
   // TODO(jkoleszar): could expand active_ref_idx to 4, with 0 as intra, and
   // roll new_fb_idx into it.
 
@@ -202,30 +236,33 @@ typedef struct VP9Common {
 
   int log2_tile_cols, log2_tile_rows;
 
-  // Private data associated with the frame buffer callbacks.
+  // External BufferPool passed from outside.
-  void *cb_priv;
+  BufferPool *buffer_pool;
-  vpx_get_frame_buffer_cb_fn_t get_fb_cb;
-  vpx_release_frame_buffer_cb_fn_t release_fb_cb;
-
-  // Handles memory for the codec.
-  InternalFrameBufferList int_frame_buffers;
 
   PARTITION_CONTEXT *above_seg_context;
   ENTROPY_CONTEXT *above_context;
 } VP9_COMMON;
 
 static INLINE YV12_BUFFER_CONFIG *get_frame_new_buffer(VP9_COMMON *cm) {
-  return &cm->frame_bufs[cm->new_fb_idx].buf;
+  return &cm->buffer_pool->frame_bufs[cm->new_fb_idx].buf;
 }
 
 static INLINE int get_free_fb(VP9_COMMON *cm) {
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
   int i;
+
+#if CONFIG_MULTITHREAD
+  pthread_mutex_lock(&cm->buffer_pool->pool_mutex);
+#endif
   for (i = 0; i < FRAME_BUFFERS; i++)
-    if (cm->frame_bufs[i].ref_count == 0)
+    if (frame_bufs[i].ref_count == 0)
       break;
 
   assert(i < FRAME_BUFFERS);
-  cm->frame_bufs[i].ref_count = 1;
+  frame_bufs[i].ref_count = 1;
+#if CONFIG_MULTITHREAD
+  pthread_mutex_unlock(&cm->buffer_pool->pool_mutex);
+#endif
   return i;
 }
 

vp9/common/vp9_quant_common.c

@@ -12,7 +12,6 @@
 #include "vp9/common/vp9_quant_common.h"
 #include "vp9/common/vp9_seg_common.h"
 
-#if 1
 static const int16_t dc_qlookup[QINDEX_RANGE] = {
   4,       8,    8,    9,   10,   11,   12,   12,
   13,     14,   15,   16,   17,   18,   19,   19,
@@ -83,44 +82,6 @@ static const int16_t ac_qlookup[QINDEX_RANGE] = {
   1597, 1628, 1660, 1692, 1725, 1759, 1793, 1828,
 };
 
-void vp9_init_quant_tables(void) { }
-#else
-static int16_t dc_qlookup[QINDEX_RANGE];
-static int16_t ac_qlookup[QINDEX_RANGE];
-
-#define ACDC_MIN 8
-
-// TODO(dkovalev) move to common and reuse
-static double poly3(double a, double b, double c, double d, double x) {
-  return a*x*x*x + b*x*x + c*x + d;
-}
-
-void vp9_init_quant_tables() {
-  int i, val = 4;
-
-  // A "real" q of 1.0 forces lossless mode.
-  // In practice non lossless Q's between 1.0 and 2.0 (represented here by
-  // integer values from 5-7 give poor rd results (lower psnr and often
-  // larger size than the lossless encode. To block out those "not very useful"
-  // values we increment the ac and dc q lookup values by 4 after position 0.
-  ac_qlookup[0] = val;
-  dc_qlookup[0] = val;
-  val += 4;
-
-  for (i = 1; i < QINDEX_RANGE; i++) {
-    const int ac_val = val;
-
-    val = (int)(val * 1.01975);
-    if (val == ac_val)
-      ++val;
-
-    ac_qlookup[i] = (int16_t)ac_val;
-    dc_qlookup[i] = (int16_t)MAX(ACDC_MIN, poly3(0.000000305, -0.00065, 0.9,
-                                                 0.5, ac_val));
-  }
-}
-#endif
-
 int16_t vp9_dc_quant(int qindex, int delta) {
   return dc_qlookup[clamp(qindex + delta, 0, MAXQ)];
 }

vp9/common/vp9_quant_common.h

@@ -22,8 +22,6 @@ extern "C" {
 #define QINDEX_RANGE (MAXQ - MINQ + 1)
 #define QINDEX_BITS 8
 
-void vp9_init_quant_tables();
-
 int16_t vp9_dc_quant(int qindex, int delta);
 int16_t vp9_ac_quant(int qindex, int delta);
 

vp9/common/vp9_reconinter.c

@@ -113,6 +113,18 @@ static MV mi_mv_pred_q4(const MODE_INFO *mi, int idx) {
   return res;
 }
 
+static INLINE int round_mv_comp_q2(int value) {
+  return (value < 0 ? value - 1 : value + 1) / 2;
+}
+
+static MV mi_mv_pred_q2(const MODE_INFO *mi, int idx, int block0, int block1) {
+  MV res = { round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.row +
+                              mi->bmi[block1].as_mv[idx].as_mv.row),
+             round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.col +
+                              mi->bmi[block1].as_mv[idx].as_mv.col) };
+  return res;
+}
+
 // TODO(jkoleszar): yet another mv clamping function :-(
 MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
                              int bw, int bh, int ss_x, int ss_y) {
@@ -139,6 +151,29 @@ MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
   return clamped_mv;
 }
 
+static MV average_split_mvs(const struct macroblockd_plane *pd, int plane,
+                            const MODE_INFO *mi, int ref, int block) {
+  const int ss_idx = ((pd->subsampling_x > 0) << 1) | (pd->subsampling_y > 0);
+  MV res = {0, 0};
+  switch (ss_idx) {
+    case 0:
+      res = mi->bmi[block].as_mv[ref].as_mv;
+      break;
+    case 1:
+      res = mi_mv_pred_q2(mi, ref, block, block + 2);
+      break;
+    case 2:
+      res = mi_mv_pred_q2(mi, ref, block, block + 1);
+      break;
+    case 3:
+      res = mi_mv_pred_q4(mi, ref);
+      break;
+    default:
+      assert(ss_idx <= 3 || ss_idx >= 0);
+  }
+  return res;
+}
+
 static void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
                                    int bw, int bh,
                                    int x, int y, int w, int h,
@@ -154,14 +189,8 @@ static void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
     struct buf_2d *const pre_buf = &pd->pre[ref];
     struct buf_2d *const dst_buf = &pd->dst;
     uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
-
-    // TODO(jkoleszar): All chroma MVs in SPLITMV mode are taken as the
-    // same MV (the average of the 4 luma MVs) but we could do something
-    // smarter for non-4:2:0. Just punt for now, pending the changes to get
-    // rid of SPLITMV mode entirely.
     const MV mv = mi->mbmi.sb_type < BLOCK_8X8
-               ? (plane == 0 ? mi->bmi[block].as_mv[ref].as_mv
+               ? average_split_mvs(pd, plane, mi, ref, block)
-                             : mi_mv_pred_q4(mi, ref))
                : mi->mbmi.mv[ref].as_mv;
 
     // TODO(jkoleszar): This clamping is done in the incorrect place for the
@@ -258,16 +287,11 @@ static void dec_build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
     struct buf_2d *const pre_buf = &pd->pre[ref];
     struct buf_2d *const dst_buf = &pd->dst;
     uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
-
-    // TODO(jkoleszar): All chroma MVs in SPLITMV mode are taken as the
-    // same MV (the average of the 4 luma MVs) but we could do something
-    // smarter for non-4:2:0. Just punt for now, pending the changes to get
-    // rid of SPLITMV mode entirely.
     const MV mv = mi->mbmi.sb_type < BLOCK_8X8
-               ? (plane == 0 ? mi->bmi[block].as_mv[ref].as_mv
+               ? average_split_mvs(pd, plane, mi, ref, block)
-                             : mi_mv_pred_q4(mi, ref))
                : mi->mbmi.mv[ref].as_mv;
 
+
     // TODO(jkoleszar): This clamping is done in the incorrect place for the
     // scaling case. It needs to be done on the scaled MV, not the pre-scaling
     // MV. Note however that it performs the subsampling aware scaling so

vp9/common/vp9_rtcd_defs.pl

@@ -305,15 +305,15 @@ specialize qw/vp9_convolve_avg neon_asm dspr2/, "$sse2_x86inc";
 $vp9_convolve_avg_neon_asm=vp9_convolve_avg_neon;
 
 add_proto qw/void vp9_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8 sse2 ssse3 avx2 neon_asm dspr2/;
+specialize qw/vp9_convolve8 sse2 ssse3 neon_asm dspr2/;
 $vp9_convolve8_neon_asm=vp9_convolve8_neon;
 
 add_proto qw/void vp9_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8_horiz sse2 ssse3 avx2 neon_asm dspr2/;
+specialize qw/vp9_convolve8_horiz sse2 ssse3 neon_asm dspr2/;
 $vp9_convolve8_horiz_neon_asm=vp9_convolve8_horiz_neon;
 
 add_proto qw/void vp9_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vp9_convolve8_vert sse2 ssse3 avx2 neon_asm dspr2/;
+specialize qw/vp9_convolve8_vert sse2 ssse3 neon_asm dspr2/;
 $vp9_convolve8_vert_neon_asm=vp9_convolve8_vert_neon;
 
 add_proto qw/void vp9_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
@@ -402,25 +402,25 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
 
 # variance
 add_proto qw/unsigned int vp9_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance32x16/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance32x16 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance16x32/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance64x32/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance64x32 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance32x64/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance32x32/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance32x32 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance64x64/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance64x64 avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance16x16 mmx/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_variance16x16 mmx avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance16x8 mmx/, "$sse2_x86inc";
@@ -447,10 +447,10 @@ add_proto qw/unsigned int vp9_variance4x4/, "const uint8_t *src_ptr, int source_
 specialize qw/vp9_variance4x4 mmx/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_sub_pixel_variance64x64 avx2/, "$sse2_x86inc", "$ssse3_x86inc";
+specialize qw/vp9_sub_pixel_variance64x64/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
-specialize qw/vp9_sub_pixel_avg_variance64x64 avx2/, "$sse2_x86inc", "$ssse3_x86inc";
+specialize qw/vp9_sub_pixel_avg_variance64x64/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_sub_pixel_variance32x64/, "$sse2_x86inc", "$ssse3_x86inc";
@@ -477,10 +477,10 @@ add_proto qw/unsigned int vp9_sub_pixel_avg_variance16x32/, "const uint8_t *src_
 specialize qw/vp9_sub_pixel_avg_variance16x32/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_sub_pixel_variance32x32 avx2/, "$sse2_x86inc", "$ssse3_x86inc";
+specialize qw/vp9_sub_pixel_variance32x32/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
-specialize qw/vp9_sub_pixel_avg_variance32x32 avx2/, "$sse2_x86inc", "$ssse3_x86inc";
+specialize qw/vp9_sub_pixel_avg_variance32x32/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_sub_pixel_variance16x16/, "$sse2_x86inc", "$ssse3_x86inc";
@@ -653,7 +653,7 @@ add_proto qw/void vp9_sad4x4x8/, "const uint8_t *src_ptr, int  src_stride, const
 specialize qw/vp9_sad4x4x8 sse4/;
 
 add_proto qw/void vp9_sad64x64x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
-specialize qw/vp9_sad64x64x4d sse2 avx2/;
+specialize qw/vp9_sad64x64x4d sse2/;
 
 add_proto qw/void vp9_sad32x64x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad32x64x4d sse2/;
@@ -668,7 +668,7 @@ add_proto qw/void vp9_sad16x32x4d/, "const uint8_t *src_ptr, int  src_stride, co
 specialize qw/vp9_sad16x32x4d sse2/;
 
 add_proto qw/void vp9_sad32x32x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
-specialize qw/vp9_sad32x32x4d sse2 avx2/;
+specialize qw/vp9_sad32x32x4d sse2/;
 
 add_proto qw/void vp9_sad16x16x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad16x16x4d sse2/;
@@ -693,7 +693,7 @@ add_proto qw/void vp9_sad4x4x4d/, "const uint8_t *src_ptr, int  src_stride, cons
 specialize qw/vp9_sad4x4x4d sse/;
 
 add_proto qw/unsigned int vp9_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
-specialize qw/vp9_mse16x16 mmx/, "$sse2_x86inc", "$avx2_x86inc";
+specialize qw/vp9_mse16x16 mmx avx2/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
 specialize qw/vp9_mse8x16/;
@@ -714,6 +714,9 @@ specialize qw/vp9_block_error avx2/, "$sse2_x86inc";
 add_proto qw/void vp9_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
 specialize qw/vp9_subtract_block/, "$sse2_x86inc";
 
+add_proto qw/void vp9_quantize_fp/, "const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+specialize qw/vp9_quantize_fp/, "$ssse3_x86_64";
+
 add_proto qw/void vp9_quantize_b/, "const int16_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
 specialize qw/vp9_quantize_b/, "$ssse3_x86_64";
 
@@ -739,19 +742,31 @@ add_proto qw/void vp9_fht8x8/, "const int16_t *input, int16_t *output, int strid
 specialize qw/vp9_fht8x8 sse2 avx2/;
 
 add_proto qw/void vp9_fht16x16/, "const int16_t *input, int16_t *output, int stride, int tx_type";
-specialize qw/vp9_fht16x16 sse2 avx2/;
+specialize qw/vp9_fht16x16 sse2/;
 
 add_proto qw/void vp9_fwht4x4/, "const int16_t *input, int16_t *output, int stride";
 specialize qw/vp9_fwht4x4/, "$mmx_x86inc";
 
+add_proto qw/void vp9_fdct4x4_1/, "const int16_t *input, int16_t *output, int stride";
+specialize qw/vp9_fdct4x4_1 sse2/;
+
 add_proto qw/void vp9_fdct4x4/, "const int16_t *input, int16_t *output, int stride";
 specialize qw/vp9_fdct4x4 sse2 avx2/;
 
+add_proto qw/void vp9_fdct8x8_1/, "const int16_t *input, int16_t *output, int stride";
+specialize qw/vp9_fdct8x8_1 sse2/;
+
 add_proto qw/void vp9_fdct8x8/, "const int16_t *input, int16_t *output, int stride";
 specialize qw/vp9_fdct8x8 sse2 avx2/, "$ssse3_x86_64";
 
+add_proto qw/void vp9_fdct16x16_1/, "const int16_t *input, int16_t *output, int stride";
+specialize qw/vp9_fdct16x16_1 sse2/;
+
 add_proto qw/void vp9_fdct16x16/, "const int16_t *input, int16_t *output, int stride";
-specialize qw/vp9_fdct16x16 sse2 avx2/;
+specialize qw/vp9_fdct16x16 sse2/;
+
+add_proto qw/void vp9_fdct32x32_1/, "const int16_t *input, int16_t *output, int stride";
+specialize qw/vp9_fdct32x32_1 sse2/;
 
 add_proto qw/void vp9_fdct32x32/, "const int16_t *input, int16_t *output, int stride";
 specialize qw/vp9_fdct32x32 sse2 avx2/;

vp9/common/vp9_scale.c

@@ -33,14 +33,6 @@ static int get_fixed_point_scale_factor(int other_size, int this_size) {
   return (other_size << REF_SCALE_SHIFT) / this_size;
 }
 
-static int check_scale_factors(int other_w, int other_h,
-                               int this_w, int this_h) {
-  return 2 * this_w >= other_w &&
-         2 * this_h >= other_h &&
-         this_w <= 16 * other_w &&
-         this_h <= 16 * other_h;
-}
-
 MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf) {
   const int x_off_q4 = scaled_x(x << SUBPEL_BITS, sf) & SUBPEL_MASK;
   const int y_off_q4 = scaled_y(y << SUBPEL_BITS, sf) & SUBPEL_MASK;
@@ -54,7 +46,7 @@ MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf) {
 void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
                                        int other_w, int other_h,
                                        int this_w, int this_h) {
-  if (!check_scale_factors(other_w, other_h, this_w, this_h)) {
+  if (!valid_ref_frame_size(other_w, other_h, this_w, this_h)) {
     sf->x_scale_fp = REF_INVALID_SCALE;
     sf->y_scale_fp = REF_INVALID_SCALE;
     return;

vp9/common/vp9_scale.h

@@ -46,8 +46,16 @@ static INLINE int vp9_is_valid_scale(const struct scale_factors *sf) {
 }
 
 static INLINE int vp9_is_scaled(const struct scale_factors *sf) {
-  return sf->x_scale_fp != REF_NO_SCALE ||
+  return vp9_is_valid_scale(sf) &&
-         sf->y_scale_fp != REF_NO_SCALE;
+         (sf->x_scale_fp != REF_NO_SCALE || sf->y_scale_fp != REF_NO_SCALE);
+}
+
+static INLINE int valid_ref_frame_size(int ref_width, int ref_height,
+                                      int this_width, int this_height) {
+  return 2 * this_width >= ref_width &&
+         2 * this_height >= ref_height &&
+         this_width <= 16 * ref_width &&
+         this_height <= 16 * ref_height;
 }
 
 #ifdef __cplusplus

vp9/common/x86/vp9_postproc_mmx.asm

@@ -464,7 +464,6 @@ sym(vp9_mbpost_proc_down_mmx):
 ;                            unsigned char whiteclamp[16],
 ;                            unsigned char bothclamp[16],
 ;                            unsigned int width, unsigned int height, int pitch)
-extern sym(rand)
 global sym(vp9_plane_add_noise_mmx) PRIVATE
 sym(vp9_plane_add_noise_mmx):
     push        rbp
@@ -476,7 +475,7 @@ sym(vp9_plane_add_noise_mmx):
     ; end prolog
 
 .addnoise_loop:
-    call sym(rand) WRT_PLT
+    call sym(LIBVPX_RAND) WRT_PLT
     mov     rcx, arg(1) ;noise
     and     rax, 0xff
     add     rcx, rax

vp9/common/x86/vp9_postproc_sse2.asm

@@ -629,7 +629,6 @@ sym(vp9_mbpost_proc_across_ip_xmm):
 ;                            unsigned char whiteclamp[16],
 ;                            unsigned char bothclamp[16],
 ;                            unsigned int width, unsigned int height, int pitch)
-extern sym(rand)
 global sym(vp9_plane_add_noise_wmt) PRIVATE
 sym(vp9_plane_add_noise_wmt):
     push        rbp
@@ -641,7 +640,7 @@ sym(vp9_plane_add_noise_wmt):
     ; end prolog
 
 .addnoise_loop:
-    call sym(rand) WRT_PLT
+    call sym(LIBVPX_RAND) WRT_PLT
     mov     rcx, arg(1) ;noise
     and     rax, 0xff
     add     rcx, rax

vp9/decoder/vp9_decodeframe.c

@@ -410,13 +410,17 @@ static void decode_partition(VP9_COMMON *const cm, MACROBLOCKD *const xd,
                              vp9_reader* r, BLOCK_SIZE bsize) {
   const int hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
   PARTITION_TYPE partition;
-  BLOCK_SIZE subsize;
+  BLOCK_SIZE subsize, uv_subsize;
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
   partition = read_partition(cm, xd, hbs, mi_row, mi_col, bsize, r);
   subsize = get_subsize(bsize, partition);
+  uv_subsize = ss_size_lookup[subsize][cm->subsampling_x][cm->subsampling_y];
+  if (subsize >= BLOCK_8X8 && uv_subsize == BLOCK_INVALID)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid block size.");
   if (subsize < BLOCK_8X8) {
     decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
   } else {
@@ -617,6 +621,7 @@ static void setup_display_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
 }
 
 static void apply_frame_size(VP9_COMMON *cm, int width, int height) {
+  BufferPool *const pool = cm->buffer_pool;
   if (cm->width != width || cm->height != height) {
     // Change in frame size.
     // TODO(agrange) Don't test width/height, check overall size.
@@ -636,8 +641,8 @@ static void apply_frame_size(VP9_COMMON *cm, int width, int height) {
   if (vp9_realloc_frame_buffer(
           get_frame_new_buffer(cm), cm->width, cm->height,
           cm->subsampling_x, cm->subsampling_y, VP9_DEC_BORDER_IN_PIXELS,
-          &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer, cm->get_fb_cb,
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
-          cm->cb_priv)) {
+          pool->cb_priv)) {
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate frame buffer");
   }
@@ -667,9 +672,17 @@ static void setup_frame_size_with_refs(VP9_COMMON *cm,
   if (!found)
     read_frame_size(rb, &width, &height);
 
-  if (width <= 0 || height <= 0)
+  // Check that each of the frames that this frame references has valid
+  // dimensions.
+  for (i = 0; i < REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    const int ref_width = ref_frame->buf->y_width;
+    const int ref_height = ref_frame->buf->y_height;
+
+    if (!valid_ref_frame_size(ref_width, ref_height, width, height))
       vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Referenced frame with invalid size");
+                         "Referenced frame has invalid size");
+  }
 
   apply_frame_size(cm, width, height);
   setup_display_size(cm, rb);
@@ -685,6 +698,10 @@ static void setup_tile_info(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
   while (max_ones-- && vp9_rb_read_bit(rb))
     cm->log2_tile_cols++;
 
+  if (cm->log2_tile_cols > 6)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid number of tile columns");
+
   // rows
   cm->log2_tile_rows = vp9_rb_read_bit(rb);
   if (cm->log2_tile_rows)
@@ -755,6 +772,7 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
                                    const uint8_t *data,
                                    const uint8_t *data_end) {
   VP9_COMMON *const cm = &pbi->common;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
   const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
   const int tile_cols = 1 << cm->log2_tile_cols;
   const int tile_rows = 1 << cm->log2_tile_rows;
@@ -767,7 +785,7 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
     CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
                     vpx_memalign(32, sizeof(LFWorkerData)));
     pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
-    if (pbi->max_threads > 1 && !vp9_worker_reset(&pbi->lf_worker)) {
+    if (pbi->max_threads > 1 && !winterface->reset(&pbi->lf_worker)) {
       vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                          "Loop filter thread creation failed");
     }
@@ -853,13 +871,13 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
         // decoding has completed: finish up the loop filter in this thread.
         if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
 
-        vp9_worker_sync(&pbi->lf_worker);
+        winterface->sync(&pbi->lf_worker);
         lf_data->start = lf_start;
         lf_data->stop = mi_row;
         if (pbi->max_threads > 1) {
-          vp9_worker_launch(&pbi->lf_worker);
+          winterface->launch(&pbi->lf_worker);
         } else {
-          vp9_worker_execute(&pbi->lf_worker);
+          winterface->execute(&pbi->lf_worker);
         }
       }
     }
@@ -868,10 +886,10 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
   // Loopfilter remaining rows in the frame.
   if (cm->lf.filter_level) {
     LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
-    vp9_worker_sync(&pbi->lf_worker);
+    winterface->sync(&pbi->lf_worker);
     lf_data->start = lf_data->stop;
     lf_data->stop = cm->mi_rows;
-    vp9_worker_execute(&pbi->lf_worker);
+    winterface->execute(&pbi->lf_worker);
   }
 
   // Get last tile data.
@@ -915,6 +933,7 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
                                       const uint8_t *data,
                                       const uint8_t *data_end) {
   VP9_COMMON *const cm = &pbi->common;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
   const uint8_t *bit_reader_end = NULL;
   const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
   const int tile_cols = 1 << cm->log2_tile_cols;
@@ -941,11 +960,11 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
       VP9Worker *const worker = &pbi->tile_workers[i];
       ++pbi->num_tile_workers;
 
-      vp9_worker_init(worker);
+      winterface->init(worker);
       CHECK_MEM_ERROR(cm, worker->data1,
                       vpx_memalign(32, sizeof(TileWorkerData)));
       CHECK_MEM_ERROR(cm, worker->data2, vpx_malloc(sizeof(TileInfo)));
-      if (i < num_threads - 1 && !vp9_worker_reset(worker)) {
+      if (i < num_threads - 1 && !winterface->reset(worker)) {
         vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                            "Tile decoder thread creation failed");
       }
@@ -1008,9 +1027,9 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
 
       worker->had_error = 0;
       if (i == num_workers - 1 || n == tile_cols - 1) {
-        vp9_worker_execute(worker);
+        winterface->execute(worker);
       } else {
-        vp9_worker_launch(worker);
+        winterface->launch(worker);
       }
 
       if (buf->col == tile_cols - 1) {
@@ -1022,7 +1041,7 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
 
     for (; i > 0; --i) {
       VP9Worker *const worker = &pbi->tile_workers[i - 1];
-      pbi->mb.corrupted |= !vp9_worker_sync(worker);
+      pbi->mb.corrupted |= !winterface->sync(worker);
     }
     if (final_worker > -1) {
       TileWorkerData *const tile_data =
@@ -1058,8 +1077,9 @@ static BITSTREAM_PROFILE read_profile(struct vp9_read_bit_buffer *rb) {
 static size_t read_uncompressed_header(VP9Decoder *pbi,
                                        struct vp9_read_bit_buffer *rb) {
   VP9_COMMON *const cm = &pbi->common;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+  int mask, i, ref_index = 0;
   size_t sz;
-  int i;
 
   cm->last_frame_type = cm->frame_type;
 
@@ -1077,12 +1097,18 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
     // Show an existing frame directly.
     const int frame_to_show = cm->ref_frame_map[vp9_rb_read_literal(rb, 3)];
 
-    if (cm->frame_bufs[frame_to_show].ref_count < 1)
+#if CONFIG_MULTITHREAD
+    pthread_mutex_lock(&cm->buffer_pool->pool_mutex);
+#endif
+    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1)
       vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                          "Buffer %d does not contain a decoded frame",
                          frame_to_show);
 
-    ref_cnt_fb(cm->frame_bufs, &cm->new_fb_idx, frame_to_show);
+    ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
+#if CONFIG_MULTITHREAD
+    pthread_mutex_unlock(&cm->buffer_pool->pool_mutex);
+#endif
     pbi->refresh_frame_flags = 0;
     cm->lf.filter_level = 0;
     cm->show_frame = 1;
@@ -1138,12 +1164,12 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
       setup_frame_size(cm, rb);
     } else {
       pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
-
       for (i = 0; i < REFS_PER_FRAME; ++i) {
         const int ref = vp9_rb_read_literal(rb, REF_FRAMES_LOG2);
         const int idx = cm->ref_frame_map[ref];
-        cm->frame_refs[i].idx = idx;
+        RefBuffer *const ref_frame = &cm->frame_refs[i];
-        cm->frame_refs[i].buf = &cm->frame_bufs[idx].buf;
+        ref_frame->idx = idx;
+        ref_frame->buf = &frame_bufs[idx].buf;
         cm->ref_frame_sign_bias[LAST_FRAME + i] = vp9_rb_read_bit(rb);
       }
 
@@ -1178,6 +1204,28 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
   // below, forcing the use of context 0 for those frame types.
   cm->frame_context_idx = vp9_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);
 
+  // Update next_ref_frame_map in frame parallel decode.
+  if (pbi->frame_parallel_decode) {
+    for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+      if (mask & 1) {
+        cm->next_ref_frame_map[ref_index] = cm->new_fb_idx;
+#if CONFIG_MULTITHREAD
+        pthread_mutex_lock(&cm->buffer_pool->pool_mutex);
+#endif
+        ++cm->buffer_pool->frame_bufs[cm->new_fb_idx].ref_count;
+#if CONFIG_MULTITHREAD
+        pthread_mutex_unlock(&cm->buffer_pool->pool_mutex);
+#endif
+      } else {
+        cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+      }
+      ++ref_index;
+    }
+
+    for (; ref_index < REF_FRAMES; ++ref_index)
+      cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+  }
+
   if (frame_is_intra_only(cm) || cm->error_resilient_mode)
     vp9_setup_past_independence(cm);
 
@@ -1322,7 +1370,8 @@ void vp9_decode_frame(VP9Decoder *pbi,
                       const uint8_t **p_data_end) {
   VP9_COMMON *const cm = &pbi->common;
   MACROBLOCKD *const xd = &pbi->mb;
-  struct vp9_read_bit_buffer rb = { 0 };
+  struct vp9_read_bit_buffer rb = { NULL, NULL, 0, NULL, 0};
+
   uint8_t clear_data[MAX_VP9_HEADER_SIZE];
   const size_t first_partition_size = read_uncompressed_header(pbi,
       init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
@@ -1373,6 +1422,17 @@ void vp9_decode_frame(VP9Decoder *pbi,
 
   new_fb->corrupted |= xd->corrupted;
 
+  // Update progress in frame parallel decode.
+  if (pbi->frame_parallel_decode) {
+    VP9Worker *worker = pbi->owner_frame_worker;
+    FrameWorkerData *const worker_data = worker->data1;
+    pthread_mutex_lock(&worker_data->stats_mutex);
+    pbi->cur_buf->row = INT_MAX;
+    pbi->cur_buf->col = INT_MAX;
+    pthread_cond_signal(&worker_data->stats_cond);
+    pthread_mutex_unlock(&worker_data->stats_mutex);
+  }
+
   if (!new_fb->corrupted) {
     if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
       vp9_adapt_coef_probs(cm);

vp9/decoder/vp9_decoder.c

@@ -37,12 +37,11 @@ static void initialize_dec() {
 
   if (!init_done) {
     vp9_init_neighbors();
-    vp9_init_quant_tables();
     init_done = 1;
   }
 }
 
-VP9Decoder *vp9_decoder_create() {
+VP9Decoder *vp9_decoder_create(BufferPool *const pool) {
   VP9Decoder *const pbi = vpx_memalign(32, sizeof(*pbi));
   VP9_COMMON *const cm = pbi ? &pbi->common : NULL;
 
@@ -64,9 +63,11 @@ VP9Decoder *vp9_decoder_create() {
 
   // Initialize the references to not point to any frame buffers.
   vpx_memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+  vpx_memset(&cm->next_ref_frame_map, -1, sizeof(cm->next_ref_frame_map));
 
   cm->current_video_frame = 0;
   pbi->ready_for_new_data = 1;
+  pbi->common.buffer_pool = pool;
 
   // vp9_init_dequantizer() is first called here. Add check in
   // frame_init_dequantizer() to avoid unnecessary calling of
@@ -77,7 +78,7 @@ VP9Decoder *vp9_decoder_create() {
 
   cm->error.setjmp = 0;
 
-  vp9_worker_init(&pbi->lf_worker);
+  vp9_get_worker_interface()->init(&pbi->lf_worker);
 
   return pbi;
 }
@@ -87,12 +88,12 @@ void vp9_decoder_remove(VP9Decoder *pbi) {
   int i;
 
   vp9_remove_common(cm);
-  vp9_worker_end(&pbi->lf_worker);
+  vp9_get_worker_interface()->end(&pbi->lf_worker);
   vpx_free(pbi->lf_worker.data1);
   vpx_free(pbi->tile_data);
   for (i = 0; i < pbi->num_tile_workers; ++i) {
     VP9Worker *const worker = &pbi->tile_workers[i];
-    vp9_worker_end(worker);
+    vp9_get_worker_interface()->end(worker);
     vpx_free(worker->data1);
     vpx_free(worker->data2);
   }
@@ -125,7 +126,7 @@ vpx_codec_err_t vp9_copy_reference_dec(VP9Decoder *pbi,
    */
   if (ref_frame_flag == VP9_LAST_FLAG) {
     const YV12_BUFFER_CONFIG *const cfg =
-        &cm->frame_bufs[cm->ref_frame_map[0]].buf;
+        &cm->buffer_pool->frame_bufs[cm->ref_frame_map[0]].buf;
     if (!equal_dimensions(cfg, sd))
       vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                          "Incorrect buffer dimensions");
@@ -144,6 +145,7 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
                                       VP9_REFFRAME ref_frame_flag,
                                       YV12_BUFFER_CONFIG *sd) {
   RefBuffer *ref_buf = NULL;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   // TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
   // encoder is using the frame buffers for. This is just a stub to keep the
@@ -171,11 +173,11 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
     const int free_fb = get_free_fb(cm);
     // Decrease ref_count since it will be increased again in
     // ref_cnt_fb() below.
-    cm->frame_bufs[free_fb].ref_count--;
+    --frame_bufs[free_fb].ref_count;
 
     // Manage the reference counters and copy image.
-    ref_cnt_fb(cm->frame_bufs, ref_fb_ptr, free_fb);
+    ref_cnt_fb(frame_bufs, ref_fb_ptr, free_fb);
-    ref_buf->buf = &cm->frame_bufs[*ref_fb_ptr].buf;
+    ref_buf->buf = &frame_bufs[*ref_fb_ptr].buf;
     vp8_yv12_copy_frame(sd, ref_buf->buf);
   }
 
@@ -185,11 +187,12 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
 
 int vp9_get_reference_dec(VP9Decoder *pbi, int index, YV12_BUFFER_CONFIG **fb) {
   VP9_COMMON *cm = &pbi->common;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   if (index < 0 || index >= REF_FRAMES)
     return -1;
 
-  *fb = &cm->frame_bufs[cm->ref_frame_map[index]].buf;
+  *fb = &frame_bufs[cm->ref_frame_map[index]].buf;
   return 0;
 }
 
@@ -197,21 +200,32 @@ int vp9_get_reference_dec(VP9Decoder *pbi, int index, YV12_BUFFER_CONFIG **fb) {
 static void swap_frame_buffers(VP9Decoder *pbi) {
   int ref_index = 0, mask;
   VP9_COMMON * const cm = &pbi->common;
+  BufferPool * const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
     if (mask & 1) {
       const int old_idx = cm->ref_frame_map[ref_index];
-      ref_cnt_fb(cm->frame_bufs, &cm->ref_frame_map[ref_index],
+#if CONFIG_MULTITHREAD
+      pthread_mutex_lock(&cm->buffer_pool->pool_mutex);
+#endif
+      ref_cnt_fb(frame_bufs, &cm->ref_frame_map[ref_index],
                  cm->new_fb_idx);
-      if (old_idx >= 0 && cm->frame_bufs[old_idx].ref_count == 0)
+      if (old_idx >= 0 && frame_bufs[old_idx].ref_count == 0)
-        cm->release_fb_cb(cm->cb_priv,
+        pool->release_fb_cb(pool->cb_priv,
-                          &cm->frame_bufs[old_idx].raw_frame_buffer);
+                            &frame_bufs[old_idx].raw_frame_buffer);
     }
+#if CONFIG_MULTITHREAD
+      pthread_mutex_unlock(&cm->buffer_pool->pool_mutex);
+#endif
     ++ref_index;
   }
 
   cm->frame_to_show = get_frame_new_buffer(cm);
-  cm->frame_bufs[cm->new_fb_idx].ref_count--;
+
+  if (!pbi->frame_parallel_decode || !cm->show_frame) {
+    --frame_bufs[cm->new_fb_idx].ref_count;
+  }
 
   // Invalidate these references until the next frame starts.
   for (ref_index = 0; ref_index < 3; ref_index++)
@@ -221,6 +235,8 @@ static void swap_frame_buffers(VP9Decoder *pbi) {
 int vp9_receive_compressed_data(VP9Decoder *pbi,
                                 size_t size, const uint8_t **psource) {
   VP9_COMMON *const cm = &pbi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
   const uint8_t *source = *psource;
   int retcode = 0;
 
@@ -240,11 +256,24 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
   }
 
   // Check if the previous frame was a frame without any references to it.
-  if (cm->new_fb_idx >= 0 && cm->frame_bufs[cm->new_fb_idx].ref_count == 0)
+  // Release frame buffer if not decoding in frame parallel mode.
-    cm->release_fb_cb(cm->cb_priv,
+  if (!pbi->frame_parallel_decode && cm->new_fb_idx >= 0
-                      &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer);
+      && frame_bufs[cm->new_fb_idx].ref_count == 0)
+    pool->release_fb_cb(pool->cb_priv,
+                        &frame_bufs[cm->new_fb_idx].raw_frame_buffer);
   cm->new_fb_idx = get_free_fb(cm);
 
+  if (pbi->frame_parallel_decode) {
+    VP9Worker *worker = pbi->owner_frame_worker;
+    FrameWorkerData *const worker_data = worker->data1;
+    pbi->cur_buf = &pool->frame_bufs[cm->new_fb_idx];
+    pool->frame_bufs[cm->new_fb_idx].owner_worker_id = worker_data->worker_id;
+
+    // Reset the decoding progress.
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+  }
+
   if (setjmp(cm->error.jmp)) {
     cm->error.setjmp = 0;
 
@@ -255,11 +284,11 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
     // TODO(jkoleszar): Error concealment is undefined and non-normative
     // at this point, but if it becomes so, [0] may not always be the correct
     // thing to do here.
-    if (cm->frame_refs[0].idx != INT_MAX)
+    if (cm->frame_refs[0].idx != INT_MAX && cm->frame_refs[0].buf != NULL)
       cm->frame_refs[0].buf->corrupted = 1;
 
-    if (cm->frame_bufs[cm->new_fb_idx].ref_count > 0)
+    if (frame_bufs[cm->new_fb_idx].ref_count > 0)
-      cm->frame_bufs[cm->new_fb_idx].ref_count--;
+      --frame_bufs[cm->new_fb_idx].ref_count;
 
     return -1;
   }

vp9/decoder/vp9_decoder.h

@@ -43,8 +43,16 @@ typedef struct VP9Decoder {
 
   int refresh_frame_flags;
 
+  int frame_parallel_decode;  // frame-based threading.
+
+  // TODO(hkuang): Combine this with cur_buf in macroblockd as they are
+  // the same.
+  RefCntBuffer *cur_buf;  //  current decoding reference buffer.
+
   VP9Worker lf_worker;
   VP9Worker *tile_workers;
+  VP9Worker *owner_frame_worker;   // frame_worker that owns this pbi;
+
   int num_tile_workers;
 
   TileData *tile_data;
@@ -76,7 +84,7 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
 int vp9_get_reference_dec(struct VP9Decoder *pbi,
                           int index, YV12_BUFFER_CONFIG **fb);
 
-struct VP9Decoder *vp9_decoder_create();
+struct VP9Decoder *vp9_decoder_create(BufferPool *const pool);
 
 void vp9_decoder_remove(struct VP9Decoder *pbi);
 

vp9/decoder/vp9_dthread.c

@@ -124,7 +124,7 @@ static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
 static int loop_filter_row_worker(void *arg1, void *arg2) {
   TileWorkerData *const tile_data = (TileWorkerData*)arg1;
   LFWorkerData *const lf_data = &tile_data->lfdata;
-
+  (void) arg2;
   loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
                       lf_data->start, lf_data->stop, lf_data->y_only,
                       lf_data->lf_sync, lf_data->num_lf_workers);
@@ -138,6 +138,7 @@ void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
                               int frame_filter_level,
                               int y_only) {
   VP9LfSync *const lf_sync = &pbi->lf_row_sync;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
   // Number of superblock rows and cols
   const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
   const int tile_cols = 1 << cm->log2_tile_cols;
@@ -197,15 +198,15 @@ void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
 
     // Start loopfiltering
     if (i == num_workers - 1) {
-      vp9_worker_execute(worker);
+      winterface->execute(worker);
     } else {
-      vp9_worker_launch(worker);
+      winterface->launch(worker);
     }
   }
 
   // Wait till all rows are finished
   for (i = 0; i < num_workers; ++i) {
-    vp9_worker_sync(&pbi->tile_workers[i]);
+    winterface->sync(&pbi->tile_workers[i]);
   }
 }
 
@@ -278,3 +279,78 @@ void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows) {
     vp9_zero(*lf_sync);
   }
 }
+
+void vp9_frameworker_wait(VP9Worker* const worker, int row, int col,
+                          RefCntBuffer *ref_buf) {
+  FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+  const VP9Decoder *const pbi = worker_data->pbi;
+  const RefCntBuffer *const cur_buf = pbi->cur_buf;
+
+  // Check if worker already release the ref_buf.
+  if (!worker || ref_buf->owner_worker_id == -1) return;
+
+  pthread_mutex_lock(&worker_data->stats_mutex);
+  while (!(cur_buf->row >= row && cur_buf->col >= col)
+         && pbi->cur_buf == ref_buf && ref_buf->owner_worker_id != -1) {
+    pthread_cond_wait(&worker_data->stats_cond, &worker_data->stats_mutex);
+  }
+  pthread_mutex_unlock(&worker_data->stats_mutex);
+}
+
+void vp9_frameworker_broadcast(VP9Worker* const worker, int row, int col) {
+  FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+  const VP9Decoder *const pbi = worker_data->pbi;
+  RefCntBuffer *const cur_buf = pbi->cur_buf;
+
+  pthread_mutex_lock(&worker_data->stats_mutex);
+  cur_buf->row = row;
+  cur_buf->col = col;
+  pthread_cond_signal(&worker_data->stats_cond);
+  pthread_mutex_unlock(&worker_data->stats_mutex);
+}
+
+void vp9_frameworker_copy_context(VP9Worker *const dst_worker,
+                                  const VP9Worker *const src_worker) {
+  FrameWorkerData *const src_worker_data =
+      (FrameWorkerData *)dst_worker->data1;
+  FrameWorkerData *const dst_worker_data =
+      (FrameWorkerData *)src_worker->data1;
+  const VP9_COMMON *const src_cm = &src_worker_data->pbi->common;
+  VP9_COMMON *const dst_cm = &dst_worker_data->pbi->common;
+  int i;
+
+  // Wait until source frame's context is ready.
+  pthread_mutex_lock(&src_worker_data->stats_mutex);
+  while (!src_worker_data->frame_context_ready) {
+    pthread_cond_wait(&src_worker_data->stats_cond,
+                      &src_worker_data->stats_mutex);
+  }
+  pthread_mutex_unlock(&src_worker_data->stats_mutex);
+
+  dst_cm->last_width = src_cm->width;
+  dst_cm->last_height = src_cm->height;
+  dst_cm->subsampling_x = src_cm->subsampling_x;
+  dst_cm->subsampling_y = src_cm->subsampling_y;
+
+  for (i = 0; i < REF_FRAMES; ++i)
+    dst_cm->ref_frame_map[i] = src_cm->next_ref_frame_map[i];
+
+  dst_cm->last_show_frame = src_cm->show_frame;
+
+  dst_cm->prev_mip = src_cm->mip;
+  dst_cm->prev_mi = src_cm->mi;
+  dst_cm->prev_mi_grid_base = src_cm->mi_grid_base;
+  dst_cm->prev_mi_grid_visible = src_cm->mi_grid_visible;
+  dst_cm->lf.last_sharpness_level = src_cm->lf.sharpness_level;
+
+  for (i = 0; i < MAX_REF_LF_DELTAS; ++i) {
+    dst_cm->lf.last_ref_deltas[i] = src_cm->lf.ref_deltas[i];
+    dst_cm->lf.ref_deltas[i] = src_cm->lf.ref_deltas[i];
+  }
+
+  for (i = 0; i < MAX_MODE_LF_DELTAS; ++i)
+    dst_cm-> lf.last_mode_deltas[i] = src_cm->lf.mode_deltas[i];
+
+  for (i = 0; i < FRAME_CONTEXTS; ++i)
+    dst_cm-> frame_contexts[i] = src_cm->frame_contexts[i];
+}

vp9/decoder/vp9_dthread.h

@@ -40,6 +40,28 @@ typedef struct VP9LfSyncData {
   int sync_range;
 } VP9LfSync;
 
+// WorkerData for the FrameWorker thread. It contains all the information of
+// the worker and decode structures for decoding a frame.
+typedef struct FrameWorkerData {
+  struct VP9Decoder *pbi;
+  const uint8_t *data;
+  const uint8_t *data_end;
+  size_t data_size;
+  void *user_priv;
+  int result;
+  int worker_id;
+
+  // scratch_buffer is used in frame parallel mode only.
+  // It is used to make a copy of the compressed data.
+  uint8_t *scratch_buffer;
+  size_t scratch_buffer_size;
+
+  pthread_mutex_t stats_mutex;
+  pthread_cond_t stats_cond;
+
+  int frame_context_ready;  // Current frame's context is ready to read.
+} FrameWorkerData;
+
 // Allocate memory for loopfilter row synchronization.
 void vp9_loop_filter_alloc(struct VP9Common *cm, VP9LfSync *lf_sync,
                            int rows, int width);
@@ -54,4 +76,19 @@ void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
                               int frame_filter_level,
                               int y_only);
 
+// Wait for FrameWorker to finish decoding ref_buf to (r,c) position.
+// Note: worker may already finish decoding ref_buf and release it in order to
+// start decoding next frame. So need to check whether worker is still decoding
+// ref_buf.
+void vp9_frameworker_wait(VP9Worker* const worker, int row, int col,
+                          RefCntBuffer *ref_buf);
+
+// FrameWorker broadcasts its decoding progress so other workers that are
+// waiting it could resume decoding.
+void vp9_frameworker_broadcast(VP9Worker* const worker, int row, int col);
+
+// Copy necessary decoding context from src worker to dst worker.
+void vp9_frameworker_copy_context(VP9Worker *const dst_worker,
+                                 const VP9Worker *const src_worker);
+
 #endif  // VP9_DECODER_VP9_DTHREAD_H_

vp9/decoder/vp9_read_bit_buffer.c

@@ -10,7 +10,7 @@
 #include "vp9/decoder/vp9_read_bit_buffer.h"
 
 size_t vp9_rb_bytes_read(struct vp9_read_bit_buffer *rb) {
-  return rb->bit_offset / CHAR_BIT + (rb->bit_offset % CHAR_BIT > 0);
+  return (rb->bit_offset + CHAR_BIT - 1) / CHAR_BIT;
 }
 
 int vp9_rb_read_bit(struct vp9_read_bit_buffer *rb) {

vp9/decoder/vp9_thread.c

@@ -11,38 +11,42 @@
 //
 // Original source:
 //  http://git.chromium.org/webm/libwebp.git
-//  100644 blob eff8f2a8c20095aade3c292b0e9292dac6cb3587  src/utils/thread.c
+//  100644 blob 08ad4e1fecba302bf1247645e84a7d2779956bc3  src/utils/thread.c
-
 
 #include <assert.h>
 #include <string.h>   // for memset()
 #include "./vp9_thread.h"
-
+#include "vpx_mem/vpx_mem.h"
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
 
 #if CONFIG_MULTITHREAD
 
+struct VP9WorkerImpl {
+  pthread_mutex_t mutex_;
+  pthread_cond_t  condition_;
+  pthread_t       thread_;
+};
+
 //------------------------------------------------------------------------------
 
-static THREADFN thread_loop(void *ptr) {    // thread loop
+static void execute(VP9Worker *const worker);  // Forward declaration.
+
+static THREADFN thread_loop(void *ptr) {
   VP9Worker *const worker = (VP9Worker*)ptr;
   int done = 0;
   while (!done) {
-    pthread_mutex_lock(&worker->mutex_);
+    pthread_mutex_lock(&worker->impl_->mutex_);
     while (worker->status_ == OK) {   // wait in idling mode
-      pthread_cond_wait(&worker->condition_, &worker->mutex_);
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
     }
     if (worker->status_ == WORK) {
-      vp9_worker_execute(worker);
+      execute(worker);
       worker->status_ = OK;
     } else if (worker->status_ == NOT_OK) {   // finish the worker
       done = 1;
     }
-    // signal to the main thread that we're done (for Sync())
+    // signal to the main thread that we're done (for sync())
-    pthread_cond_signal(&worker->condition_);
+    pthread_cond_signal(&worker->impl_->condition_);
-    pthread_mutex_unlock(&worker->mutex_);
+    pthread_mutex_unlock(&worker->impl_->mutex_);
   }
   return THREAD_RETURN(NULL);    // Thread is finished
 }
@@ -50,32 +54,36 @@ static THREADFN thread_loop(void *ptr) {    // thread loop
 // main thread state control
 static void change_state(VP9Worker *const worker,
                          VP9WorkerStatus new_status) {
-  // no-op when attempting to change state on a thread that didn't come up
+  // No-op when attempting to change state on a thread that didn't come up.
-  if (worker->status_ < OK) return;
+  // Checking status_ without acquiring the lock first would result in a data
+  // race.
+  if (worker->impl_ == NULL) return;
 
-  pthread_mutex_lock(&worker->mutex_);
+  pthread_mutex_lock(&worker->impl_->mutex_);
+  if (worker->status_ >= OK) {
     // wait for the worker to finish
     while (worker->status_ != OK) {
-    pthread_cond_wait(&worker->condition_, &worker->mutex_);
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
     }
     // assign new status and release the working thread if needed
     if (new_status != OK) {
       worker->status_ = new_status;
-    pthread_cond_signal(&worker->condition_);
+      pthread_cond_signal(&worker->impl_->condition_);
     }
-  pthread_mutex_unlock(&worker->mutex_);
+  }
+  pthread_mutex_unlock(&worker->impl_->mutex_);
 }
 
 #endif  // CONFIG_MULTITHREAD
 
 //------------------------------------------------------------------------------
 
-void vp9_worker_init(VP9Worker* const worker) {
+static void init(VP9Worker *const worker) {
   memset(worker, 0, sizeof(*worker));
   worker->status_ = NOT_OK;
 }
 
-int vp9_worker_sync(VP9Worker* const worker) {
+static int sync(VP9Worker *const worker) {
 #if CONFIG_MULTITHREAD
   change_state(worker, OK);
 #endif
@@ -83,59 +91,93 @@ int vp9_worker_sync(VP9Worker* const worker) {
   return !worker->had_error;
 }
 
-int vp9_worker_reset(VP9Worker* const worker) {
+static int reset(VP9Worker *const worker) {
   int ok = 1;
   worker->had_error = 0;
   if (worker->status_ < OK) {
 #if CONFIG_MULTITHREAD
-    if (pthread_mutex_init(&worker->mutex_, NULL) ||
+    worker->impl_ = (VP9WorkerImpl*)vpx_calloc(1, sizeof(*worker->impl_));
-        pthread_cond_init(&worker->condition_, NULL)) {
+    if (worker->impl_ == NULL) {
       return 0;
     }
-    pthread_mutex_lock(&worker->mutex_);
+    if (pthread_mutex_init(&worker->impl_->mutex_, NULL)) {
-    ok = !pthread_create(&worker->thread_, NULL, thread_loop, worker);
+      goto Error;
+    }
+    if (pthread_cond_init(&worker->impl_->condition_, NULL)) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      goto Error;
+    }
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    ok = !pthread_create(&worker->impl_->thread_, NULL, thread_loop, worker);
     if (ok) worker->status_ = OK;
-    pthread_mutex_unlock(&worker->mutex_);
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+    if (!ok) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      pthread_cond_destroy(&worker->impl_->condition_);
+ Error:
+      vpx_free(worker->impl_);
+      worker->impl_ = NULL;
+      return 0;
+    }
 #else
     worker->status_ = OK;
 #endif
   } else if (worker->status_ > OK) {
-    ok = vp9_worker_sync(worker);
+    ok = sync(worker);
   }
   assert(!ok || (worker->status_ == OK));
   return ok;
 }
 
-void vp9_worker_execute(VP9Worker* const worker) {
+static void execute(VP9Worker *const worker) {
   if (worker->hook != NULL) {
     worker->had_error |= !worker->hook(worker->data1, worker->data2);
   }
 }
 
-void vp9_worker_launch(VP9Worker* const worker) {
+static void launch(VP9Worker *const worker) {
 #if CONFIG_MULTITHREAD
   change_state(worker, WORK);
 #else
-  vp9_worker_execute(worker);
+  execute(worker);
 #endif
 }
 
-void vp9_worker_end(VP9Worker* const worker) {
+static void end(VP9Worker *const worker) {
   if (worker->status_ >= OK) {
 #if CONFIG_MULTITHREAD
     change_state(worker, NOT_OK);
-    pthread_join(worker->thread_, NULL);
+    pthread_join(worker->impl_->thread_, NULL);
-    pthread_mutex_destroy(&worker->mutex_);
+    pthread_mutex_destroy(&worker->impl_->mutex_);
-    pthread_cond_destroy(&worker->condition_);
+    pthread_cond_destroy(&worker->impl_->condition_);
 #else
     worker->status_ = NOT_OK;
 #endif
   }
+  vpx_free(worker->impl_);
+  worker->impl_ = NULL;
   assert(worker->status_ == NOT_OK);
 }
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+static VP9WorkerInterface g_worker_interface = {
-}    // extern "C"
+  init, reset, sync, launch, execute, end
-#endif
+};
+
+int vp9_set_worker_interface(const VP9WorkerInterface* const winterface) {
+  if (winterface == NULL ||
+      winterface->init == NULL || winterface->reset == NULL ||
+      winterface->sync == NULL || winterface->launch == NULL ||
+      winterface->execute == NULL || winterface->end == NULL) {
+    return 0;
+  }
+  g_worker_interface = *winterface;
+  return 1;
+}
+
+const VP9WorkerInterface *vp9_get_worker_interface(void) {
+  return &g_worker_interface;
+}
+
+//------------------------------------------------------------------------------

vp9/decoder/vp9_thread.h

@@ -11,8 +11,7 @@
 //
 // Original source:
 //  http://git.chromium.org/webm/libwebp.git
-//  100644 blob 13a61a4c84194c3374080cbf03d881d3cd6af40d  src/utils/thread.h
+//  100644 blob 7bd451b124ae3b81596abfbcc823e3cb129d3a38  src/utils/thread.h
-
 
 #ifndef VP9_DECODER_VP9_THREAD_H_
 #define VP9_DECODER_VP9_THREAD_H_
@@ -163,13 +162,12 @@ typedef enum {
 // arguments (data1 and data2), and should return false in case of error.
 typedef int (*VP9WorkerHook)(void*, void*);
 
-// Synchronize object used to launch job in the worker thread
+// Platform-dependent implementation details for the worker.
+typedef struct VP9WorkerImpl VP9WorkerImpl;
+
+// Synchronization object used to launch job in the worker thread
 typedef struct {
-#if CONFIG_MULTITHREAD
+  VP9WorkerImpl *impl_;
-  pthread_mutex_t mutex_;
-  pthread_cond_t  condition_;
-  pthread_t       thread_;
-#endif
   VP9WorkerStatus status_;
   VP9WorkerHook hook;     // hook to call
   void *data1;            // first argument passed to 'hook'
@@ -177,26 +175,40 @@ typedef struct {
   int had_error;          // return value of the last call to 'hook'
 } VP9Worker;
 
+// The interface for all thread-worker related functions. All these functions
+// must be implemented.
+typedef struct {
   // Must be called first, before any other method.
-void vp9_worker_init(VP9Worker* const worker);
+  void (*init)(VP9Worker *const worker);
   // Must be called to initialize the object and spawn the thread. Re-entrant.
   // Will potentially launch the thread. Returns false in case of error.
-int vp9_worker_reset(VP9Worker* const worker);
+  int (*reset)(VP9Worker *const worker);
   // Makes sure the previous work is finished. Returns true if worker->had_error
   // was not set and no error condition was triggered by the working thread.
-int vp9_worker_sync(VP9Worker* const worker);
+  int (*sync)(VP9Worker *const worker);
-// Triggers the thread to call hook() with data1 and data2 argument. These
+  // Triggers the thread to call hook() with data1 and data2 arguments. These
-// hook/data1/data2 can be changed at any time before calling this function,
+  // hook/data1/data2 values can be changed at any time before calling this
-// but not be changed afterward until the next call to vp9_worker_sync().
+  // function, but not be changed afterward until the next call to Sync().
-void vp9_worker_launch(VP9Worker* const worker);
+  void (*launch)(VP9Worker *const worker);
-// This function is similar to vp9_worker_launch() except that it calls the
+  // This function is similar to launch() except that it calls the
   // hook directly instead of using a thread. Convenient to bypass the thread
-// mechanism while still using the VP9Worker structs. vp9_worker_sync() must
+  // mechanism while still using the VP9Worker structs. sync() must
   // still be called afterward (for error reporting).
-void vp9_worker_execute(VP9Worker* const worker);
+  void (*execute)(VP9Worker *const worker);
   // Kill the thread and terminate the object. To use the object again, one
-// must call vp9_worker_reset() again.
+  // must call reset() again.
-void vp9_worker_end(VP9Worker* const worker);
+  void (*end)(VP9Worker *const worker);
+} VP9WorkerInterface;
+
+// Install a new set of threading functions, overriding the defaults. This
+// should be done before any workers are started, i.e., before any encoding or
+// decoding takes place. The contents of the interface struct are copied, it
+// is safe to free the corresponding memory after this call. This function is
+// not thread-safe. Return false in case of invalid pointer or methods.
+int vp9_set_worker_interface(const VP9WorkerInterface *const winterface);
+
+// Retrieve the currently set thread worker interface.
+const VP9WorkerInterface *vp9_get_worker_interface(void);
 
 //------------------------------------------------------------------------------
 

vp9/encoder/vp9_aq_complexity.c

@@ -15,8 +15,19 @@
 
 #include "vp9/encoder/vp9_segmentation.h"
 
-static const double in_frame_q_adj_ratio[MAX_SEGMENTS] =
+#define AQ_C_SEGMENTS  3
-  {1.0, 2.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
+#define AQ_C_STRENGTHS  3
+static const int aq_c_active_segments[AQ_C_STRENGTHS] = {1, 2, 3};
+static const double aq_c_q_adj_factor[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  {{1.0, 1.0, 1.0}, {1.0, 2.0, 1.0}, {1.0, 1.5, 2.5}};
+static const double aq_c_transitions[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  {{1.0, 1.0, 1.0}, {1.0, 0.25, 0.0}, {1.0, 0.5, 0.25}};
+
+static int get_aq_c_strength(int q_index) {
+  // Approximate base quatizer (truncated to int)
+  int base_quant = vp9_ac_quant(q_index, 0) / 4;
+  return (base_quant > 20) + (base_quant > 45);
+}
 
 void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
@@ -29,6 +40,8 @@ void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
       cpi->refresh_alt_ref_frame ||
       (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
     int segment;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex);
+    const int active_segments = aq_c_active_segments[aq_strength];
 
     // Clear down the segment map.
     vpx_memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
@@ -36,9 +49,17 @@ void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
     // Clear down the complexity map used for rd.
     vpx_memset(cpi->complexity_map, 0, cm->mi_rows * cm->mi_cols);
 
-    vp9_enable_segmentation(seg);
     vp9_clearall_segfeatures(seg);
 
+    // Segmentation only makes sense if the target bits per SB is above a
+    // threshold. Below this the overheads will usually outweigh any benefit.
+    if (cpi->rc.sb64_target_rate < 256) {
+      vp9_disable_segmentation(seg);
+      return;
+    }
+
+    vp9_enable_segmentation(seg);
+
     // Select delta coding method.
     seg->abs_delta = SEGMENT_DELTADATA;
 
@@ -46,14 +67,14 @@ void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
     vp9_disable_segfeature(seg, 0, SEG_LVL_ALT_Q);
 
     // Use some of the segments for in frame Q adjustment.
-    for (segment = 1; segment < 2; segment++) {
+    for (segment = 1; segment < active_segments; ++segment) {
       int qindex_delta =
           vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
-                                     in_frame_q_adj_ratio[segment]);
+                                     aq_c_q_adj_factor[aq_strength][segment]);
 
-      // For AQ mode 2, we dont allow Q0 in a segment if the base Q is not 0.
+      // For AQ complexity mode, we dont allow Q0 in a segment if the base
-      // Q0 (lossless) implies 4x4 only and in AQ mode 2 a segment Q delta
+      // Q is not 0. Q0 (lossless) implies 4x4 only and in AQ mode 2 a segment
-      // is sometimes applied without going back around the rd loop.
+      // Q delta is sometimes applied without going back around the rd loop.
       // This could lead to an illegal combination of partition size and q.
       if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
         qindex_delta = -cm->base_qindex + 1;
@@ -66,7 +87,12 @@ void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
   }
 }
 
-// Select a segment for the current SB64
+// Select a segment for the current SB64 block.
+// The choice of segment for a block depends on the ratio of the projected
+// bits for the block vs a target average.
+// An "aq_strength" value determines how many segments are supported,
+// the set of transition points to use and the extent of the quantizer
+// adjustment for each segment (configured in vp9_setup_in_frame_q_adj()).
 void vp9_select_in_frame_q_segment(VP9_COMP *cpi,
                                    int mi_row, int mi_col,
                                    int output_enabled, int projected_rate) {
@@ -89,11 +115,22 @@ void vp9_select_in_frame_q_segment(VP9_COMP *cpi,
     // It is converted to bits * 256 units.
     const int target_rate = (cpi->rc.sb64_target_rate * xmis * ymis * 256) /
                             (bw * bh);
+    const int aq_strength = get_aq_c_strength(cm->base_qindex);
+    const int active_segments = aq_c_active_segments[aq_strength];
 
-    if (projected_rate < (target_rate / 4)) {
+    // The number of segments considered and the transition points used to
-      segment = 1;
+    // select them is determined by the "aq_strength" value.
-    } else {
+    // Currently this loop only supports segments that reduce Q (i.e. where
-      segment = 0;
+    // there is undershoot.
+    // The loop counts down towards segment 0 which is the default segment
+    // with no Q adjustment.
+    segment = active_segments - 1;
+    while (segment > 0) {
+      if (projected_rate <
+          (target_rate * aq_c_transitions[aq_strength][segment])) {
+        break;
+      }
+      --segment;
     }
 
     if (target_rate > 0) {

vp9/encoder/vp9_bitstream.c

@@ -890,14 +890,8 @@ static void write_tile_info(VP9_COMMON *cm, struct vp9_write_bit_buffer *wb) {
 }
 
 static int get_refresh_mask(VP9_COMP *cpi) {
-    // Should the GF or ARF be updated using the transmitted frame or buffer
+    if (!cpi->multi_arf_allowed && cpi->refresh_golden_frame &&
-#if CONFIG_MULTIPLE_ARF
+        cpi->rc.is_src_frame_alt_ref && !cpi->use_svc) {
-    if (!cpi->multi_arf_enabled && cpi->refresh_golden_frame &&
-        !cpi->refresh_alt_ref_frame) {
-#else
-    if (cpi->refresh_golden_frame && !cpi->refresh_alt_ref_frame &&
-        !cpi->use_svc) {
-#endif
       // Preserve the previously existing golden frame and update the frame in
       // the alt ref slot instead. This is highly specific to the use of
       // alt-ref as a forward reference, and this needs to be generalized as
@@ -910,15 +904,10 @@ static int get_refresh_mask(VP9_COMP *cpi) {
              (cpi->refresh_golden_frame << cpi->alt_fb_idx);
     } else {
       int arf_idx = cpi->alt_fb_idx;
-#if CONFIG_MULTIPLE_ARF
+      if ((cpi->pass == 2) && cpi->multi_arf_allowed) {
-      // Determine which ARF buffer to use to encode this ARF frame.
+        const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
-      if (cpi->multi_arf_enabled) {
+        arf_idx = gf_group->arf_update_idx[gf_group->index];
-        int sn = cpi->sequence_number;
-        arf_idx = (cpi->frame_coding_order[sn] < 0) ?
-            cpi->arf_buffer_idx[sn + 1] :
-            cpi->arf_buffer_idx[sn];
       }
-#endif
       return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
              (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
              (cpi->refresh_alt_ref_frame << arf_idx);

vp9/encoder/vp9_block.h

@@ -20,6 +20,12 @@
 extern "C" {
 #endif
 
+typedef struct {
+  unsigned int sse;
+  int sum;
+  unsigned int var;
+} diff;
+
 struct macroblock_plane {
   DECLARE_ALIGNED(16, int16_t, src_diff[64 * 64]);
   int16_t *qcoeff;
@@ -28,6 +34,8 @@ struct macroblock_plane {
   struct buf_2d src;
 
   // Quantizer setings
+  int16_t *quant_fp;
+  int16_t *round_fp;
   int16_t *quant;
   int16_t *quant_shift;
   int16_t *zbin;
@@ -48,7 +56,7 @@ struct macroblock {
 
   MACROBLOCKD e_mbd;
   int skip_block;
-  int select_txfm_size;
+  int select_tx_size;
   int skip_recode;
   int skip_optimize;
   int q_index;
@@ -92,8 +100,6 @@ struct macroblock {
 
   int encode_breakout;
 
-  int in_active_map;
-
   // note that token_costs is the cost when eob node is skipped
   vp9_coeff_cost token_costs[TX_SIZES];
 
@@ -105,6 +111,12 @@ struct macroblock {
   int use_lp32x32fdct;
   int skip_encode;
 
+  // use fast quantization process
+  int quant_fp;
+
+  // skip forward transform and quantization
+  int skip_txfm;
+
   // Used to store sub partition's choices.
   MV pred_mv[MAX_REF_FRAMES];
 

vp9/encoder/vp9_context_tree.h

@@ -33,6 +33,7 @@ typedef struct {
   int is_coded;
   int num_4x4_blk;
   int skip;
+  int skip_txfm;
   int best_mode_index;
   int hybrid_pred_diff;
   int comp_pred_diff;

vp9/encoder/vp9_dct.c

@@ -43,6 +43,17 @@ static void fdct4(const int16_t *input, int16_t *output) {
   output[3] = fdct_round_shift(temp2);
 }
 
+void vp9_fdct4x4_1_c(const int16_t *input, int16_t *output, int stride) {
+  int r, c;
+  int16_t sum = 0;
+  for (r = 0; r < 4; ++r)
+    for (c = 0; c < 4; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum << 1;
+  output[1] = 0;
+}
+
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride) {
   // The 2D transform is done with two passes which are actually pretty
   // similar. In the first one, we transform the columns and transpose
@@ -240,6 +251,17 @@ static void fdct8(const int16_t *input, int16_t *output) {
   output[7] = fdct_round_shift(t3);
 }
 
+void vp9_fdct8x8_1_c(const int16_t *input, int16_t *output, int stride) {
+  int r, c;
+  int16_t sum = 0;
+  for (r = 0; r < 8; ++r)
+    for (c = 0; c < 8; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum;
+  output[1] = 0;
+}
+
 void vp9_fdct8x8_c(const int16_t *input, int16_t *final_output, int stride) {
   int i, j;
   int16_t intermediate[64];
@@ -311,6 +333,17 @@ void vp9_fdct8x8_c(const int16_t *input, int16_t *final_output, int stride) {
   }
 }
 
+void vp9_fdct16x16_1_c(const int16_t *input, int16_t *output, int stride) {
+  int r, c;
+  int16_t sum = 0;
+  for (r = 0; r < 16; ++r)
+    for (c = 0; c < 16; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum >> 1;
+  output[1] = 0;
+}
+
 void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride) {
   // The 2D transform is done with two passes which are actually pretty
   // similar. In the first one, we transform the columns and transpose
@@ -1329,6 +1362,17 @@ static void fdct32(const int *input, int *output, int round) {
   output[31] = dct_32_round(step[31] * cospi_31_64 + step[16] * -cospi_1_64);
 }
 
+void vp9_fdct32x32_1_c(const int16_t *input, int16_t *output, int stride) {
+  int r, c;
+  int16_t sum = 0;
+  for (r = 0; r < 32; ++r)
+    for (c = 0; c < 32; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum >> 3;
+  output[1] = 0;
+}
+
 void vp9_fdct32x32_c(const int16_t *input, int16_t *out, int stride) {
   int i, j;
   int output[32 * 32];

vp9/encoder/vp9_denoiser.c

@@ -0,0 +1,385 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_denoiser.h"
+
+#ifdef OUTPUT_YUV_DENOISED
+static void make_grayscale(YV12_BUFFER_CONFIG *yuv);
+#endif
+
+static const int widths[]  = {4, 4, 8, 8,  8, 16, 16, 16, 32, 32, 32, 64, 64};
+static const int heights[] = {4, 8, 4, 8, 16,  8, 16, 32, 16, 32, 64, 32, 64};
+
+static VP9_DENOISER_DECISION update_running_avg(const uint8_t *mc_avg,
+                                                int mc_avg_stride,
+                                                uint8_t *avg, int avg_stride,
+                                                const uint8_t *sig,
+                                                int sig_stride,
+                                                int increase_denoising,
+                                                BLOCK_SIZE bs) {
+  int r, c;
+  int diff, adj, absdiff;
+  int shift_inc1 = 0, shift_inc2 = 1;
+  int adj_val[] = {3, 4, 6};
+  int total_adj = 0;
+
+  if (increase_denoising) {
+    shift_inc1 = 1;
+    shift_inc2 = 2;
+  }
+
+  for (r = 0; r < heights[bs]; ++r) {
+    for (c = 0; c < widths[bs]; ++c) {
+      diff = mc_avg[c] - sig[c];
+      absdiff = abs(diff);
+
+      if (absdiff <= 3 + shift_inc1) {
+        avg[c] = mc_avg[c];
+        total_adj += diff;
+      } else {
+        switch (absdiff) {
+          case 4: case 5: case 6: case 7:
+            adj = adj_val[0];
+            break;
+          case 8: case 9: case 10: case 11:
+          case 12: case 13: case 14: case 15:
+            adj = adj_val[1];
+            break;
+          default:
+            adj = adj_val[2];
+        }
+        if (diff > 0) {
+          avg[c] = MIN(UINT8_MAX, sig[c] + adj);
+          total_adj += adj;
+        } else {
+          avg[c] = MAX(0, sig[c] - adj);
+          total_adj -= adj;
+        }
+      }
+    }
+    sig += sig_stride;
+    avg += avg_stride;
+    mc_avg += mc_avg_stride;
+  }
+  return total_adj;
+}
+
+static uint8_t *block_start(uint8_t *framebuf, int stride,
+                            int mi_row, int mi_col) {
+  return framebuf + (stride * mi_row * 8) + (mi_col * 8);
+}
+
+void copy_block(uint8_t *dest, int dest_stride,
+                uint8_t *src, int src_stride, BLOCK_SIZE bs) {
+  int r, c;
+  for (r = 0; r < heights[bs]; ++r) {
+    for (c = 0; c < widths[bs]; ++c) {
+      dest[c] = src[c];
+    }
+    dest += dest_stride;
+    src += src_stride;
+  }
+}
+
+static VP9_DENOISER_DECISION perform_motion_compensation(VP9_DENOISER *denoiser,
+                                                         MACROBLOCK *mb,
+                                                         BLOCK_SIZE bs,
+                                                         int increase_denoising,
+                                                         int mi_row,
+                                                         int mi_col) {
+  // constants
+  // TODO(tkopp): empirically determine good constants, or functions of block
+  // size.
+  int NOISE_MOTION_THRESHOLD = 25 * 25;
+  int SSE_DIFF_THRESHOLD = heights[bs] * widths[bs] * 20;
+  unsigned int SSE_THRESH = heights[bs] * widths[bs] * 40;
+  unsigned int SSE_THRESH_HI = heights[bs] * widths[bs] * 60;
+
+  int mv_col, mv_row;
+  int sse_diff = denoiser->zero_mv_sse - denoiser->best_sse;
+  int sse_diff_thresh;
+  int sse_thresh;
+  MV_REFERENCE_FRAME frame;
+  MACROBLOCKD *filter_mbd = &mb->e_mbd;
+  MB_MODE_INFO *mbmi = &filter_mbd->mi[0]->mbmi;
+
+  // We will restore these after motion compensation.
+  MB_MODE_INFO saved_mbmi = *mbmi;
+  struct buf_2d saved_dst = filter_mbd->plane[0].dst;
+  struct buf_2d saved_pre[2];
+  saved_pre[0] = filter_mbd->plane[0].pre[0];
+  saved_pre[1] = filter_mbd->plane[0].pre[1];
+
+  // Decide the threshold for sum squared error.
+  mv_col = denoiser->best_sse_mv.as_mv.col;
+  mv_row = denoiser->best_sse_mv.as_mv.row;
+  if (mv_row * mv_row + mv_col * mv_col > NOISE_MOTION_THRESHOLD) {
+    sse_diff_thresh = 0;
+  } else {
+    sse_diff_thresh = SSE_DIFF_THRESHOLD;
+  }
+
+  frame = denoiser->best_reference_frame;
+
+  // If the best reference frame uses inter-prediction and there is enough of a
+  // difference in sum-squared-error, use it.
+  if (frame != INTRA_FRAME && sse_diff > sse_diff_thresh) {
+    mbmi->ref_frame[0] = denoiser->best_reference_frame;
+    mbmi->mode = denoiser->best_sse_inter_mode;
+    mbmi->mv[0] = denoiser->best_sse_mv;
+  } else {
+    // Otherwise, use the zero reference frame.
+    frame = denoiser->best_zeromv_reference_frame;
+
+    mbmi->ref_frame[0] = denoiser->best_zeromv_reference_frame;
+    mbmi->mode = ZEROMV;
+    mbmi->mv[0].as_int = 0;
+
+    denoiser->best_sse_inter_mode = ZEROMV;
+    denoiser->best_sse_mv.as_int = 0;
+    denoiser->best_sse = denoiser->zero_mv_sse;
+  }
+
+  // Set the pointers in the MACROBLOCKD to point to the buffers in the denoiser
+  // struct.
+  filter_mbd->plane[0].pre[0].buf =
+      block_start(denoiser->running_avg_y[frame].y_buffer,
+                  denoiser->running_avg_y[frame].y_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[0].pre[0].stride = denoiser->running_avg_y[frame].y_stride;
+
+  filter_mbd->plane[1].pre[0].buf =
+      block_start(denoiser->running_avg_y[frame].u_buffer,
+                  denoiser->running_avg_y[frame].uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[1].pre[0].stride = denoiser->running_avg_y[frame].uv_stride;
+
+  filter_mbd->plane[2].pre[0].buf =
+      block_start(denoiser->running_avg_y[frame].v_buffer,
+                  denoiser->running_avg_y[frame].uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[2].pre[0].stride = denoiser->running_avg_y[frame].uv_stride;
+
+  filter_mbd->plane[0].pre[1].buf =
+      block_start(denoiser->running_avg_y[frame].y_buffer,
+                  denoiser->running_avg_y[frame].y_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[0].pre[1].stride = denoiser->running_avg_y[frame].y_stride;
+
+  filter_mbd->plane[1].pre[1].buf =
+      block_start(denoiser->running_avg_y[frame].u_buffer,
+                  denoiser->running_avg_y[frame].uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[1].pre[1].stride = denoiser->running_avg_y[frame].uv_stride;
+
+  filter_mbd->plane[2].pre[1].buf =
+      block_start(denoiser->running_avg_y[frame].v_buffer,
+                  denoiser->running_avg_y[frame].uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[2].pre[1].stride = denoiser->running_avg_y[frame].uv_stride;
+
+  filter_mbd->plane[0].dst.buf =
+      block_start(denoiser->mc_running_avg_y.y_buffer,
+                  denoiser->mc_running_avg_y.y_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[0].dst.stride = denoiser->mc_running_avg_y.y_stride;
+
+  filter_mbd->plane[1].dst.buf =
+      block_start(denoiser->mc_running_avg_y.u_buffer,
+                  denoiser->mc_running_avg_y.uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[1].dst.stride = denoiser->mc_running_avg_y.y_stride;
+
+  filter_mbd->plane[2].dst.buf =
+      block_start(denoiser->mc_running_avg_y.v_buffer,
+                  denoiser->mc_running_avg_y.uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[2].dst.stride = denoiser->mc_running_avg_y.y_stride;
+
+  vp9_build_inter_predictors_sby(filter_mbd, mv_row, mv_col, bs);
+
+  // Restore everything to its original state
+  filter_mbd->plane[0].pre[0] = saved_pre[0];
+  filter_mbd->plane[0].pre[1] = saved_pre[1];
+  filter_mbd->plane[0].dst = saved_dst;
+  *mbmi = saved_mbmi;
+
+  mv_row = denoiser->best_sse_mv.as_mv.row;
+  mv_col = denoiser->best_sse_mv.as_mv.col;
+  sse_thresh = denoiser->increase_denoising ? SSE_THRESH_HI : SSE_THRESH;
+
+  // TODO(tkopp) why 8?
+  if (denoiser->best_sse > sse_thresh ||
+    mv_row * mv_row + mv_col * mv_col > 8 * NOISE_MOTION_THRESHOLD) {
+    return COPY_BLOCK;
+  }
+  return FILTER_BLOCK;
+}
+
+void vp9_denoiser_denoise(VP9_DENOISER *denoiser, MACROBLOCK *mb,
+                          int mi_row, int mi_col, BLOCK_SIZE bs) {
+  VP9_DENOISER_DECISION decision = FILTER_BLOCK;
+
+  YV12_BUFFER_CONFIG avg = denoiser->running_avg_y[INTRA_FRAME];
+  YV12_BUFFER_CONFIG mc_avg = denoiser->mc_running_avg_y;
+  uint8_t *avg_start = block_start(avg.y_buffer, avg.y_stride, mi_row, mi_col);
+  uint8_t *mc_avg_start = block_start(mc_avg.y_buffer, mc_avg.y_stride,
+                                          mi_row, mi_col);
+  struct buf_2d src = mb->plane[0].src;
+
+  decision = perform_motion_compensation(denoiser, mb, bs,
+                                         denoiser->increase_denoising,
+                                         mi_row, mi_col);
+  update_running_avg(mc_avg_start, mc_avg.y_stride, avg_start, avg.y_stride,
+                     mb->plane[0].src.buf, mb->plane[0].src.stride, 0, bs);
+
+  if (decision == FILTER_BLOCK) {
+    // TODO(tkopp)
+  }
+  if (decision == COPY_BLOCK) {
+    copy_block(avg_start, avg.y_stride, src.buf, src.stride, bs);
+  }
+}
+
+static void copy_frame(YV12_BUFFER_CONFIG dest, const YV12_BUFFER_CONFIG src) {
+  int r, c;
+  const uint8_t *srcbuf = src.y_buffer;
+  uint8_t *destbuf = dest.y_buffer;
+  assert(dest.y_width == src.y_width);
+  assert(dest.y_height == src.y_height);
+
+  for (r = 0; r < dest.y_height; ++r) {
+    for (c = 0; c < dest.y_width; ++c) {
+      destbuf[c] = srcbuf[c];
+    }
+    destbuf += dest.y_stride;
+    srcbuf += src.y_stride;
+  }
+}
+
+void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
+                                    YV12_BUFFER_CONFIG src,
+                                    FRAME_TYPE frame_type,
+                                    int refresh_alt_ref_frame,
+                                    int refresh_golden_frame,
+                                    int refresh_last_frame) {
+  if (frame_type == KEY_FRAME) {
+    int i;
+    copy_frame(denoiser->running_avg_y[LAST_FRAME], src);
+    for (i = 2; i < MAX_REF_FRAMES - 1; i++) {
+      copy_frame(denoiser->running_avg_y[i],
+                 denoiser->running_avg_y[LAST_FRAME]);
+    }
+  } else {  /* For non key frames */
+    if (refresh_alt_ref_frame) {
+      copy_frame(denoiser->running_avg_y[ALTREF_FRAME],
+                 denoiser->running_avg_y[INTRA_FRAME]);
+    }
+    if (refresh_golden_frame) {
+      copy_frame(denoiser->running_avg_y[GOLDEN_FRAME],
+                 denoiser->running_avg_y[INTRA_FRAME]);
+    }
+    if (refresh_last_frame) {
+      copy_frame(denoiser->running_avg_y[LAST_FRAME],
+                 denoiser->running_avg_y[INTRA_FRAME]);
+    }
+  }
+}
+
+void vp9_denoiser_reset_frame_stats(VP9_DENOISER *denoiser) {
+  denoiser->zero_mv_sse = UINT_MAX;
+  denoiser->best_sse = UINT_MAX;
+}
+
+void vp9_denoiser_update_frame_stats(VP9_DENOISER *denoiser, MB_MODE_INFO *mbmi,
+                                     unsigned int sse, PREDICTION_MODE mode) {
+  // TODO(tkopp): Use both MVs if possible
+  if (mbmi->mv[0].as_int == 0 && sse < denoiser->zero_mv_sse) {
+    denoiser->zero_mv_sse = sse;
+    denoiser->best_zeromv_reference_frame = mbmi->ref_frame[0];
+  }
+
+  if (mbmi->mv[0].as_int != 0 && sse < denoiser->best_sse) {
+    denoiser->best_sse = sse;
+    denoiser->best_sse_inter_mode = mode;
+    denoiser->best_sse_mv = mbmi->mv[0];
+    denoiser->best_reference_frame = mbmi->ref_frame[0];
+  }
+}
+
+int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
+                       int ssx, int ssy, int border) {
+  int i, fail;
+  assert(denoiser != NULL);
+
+  for (i = 0; i < MAX_REF_FRAMES; ++i) {
+    fail = vp9_alloc_frame_buffer(&denoiser->running_avg_y[i], width, height,
+                                  ssx, ssy, border);
+    if (fail) {
+      vp9_denoiser_free(denoiser);
+      return 1;
+    }
+#ifdef OUTPUT_YUV_DENOISED
+    make_grayscale(&denoiser->running_avg_y[i]);
+#endif
+  }
+
+  fail = vp9_alloc_frame_buffer(&denoiser->mc_running_avg_y, width, height,
+                                ssx, ssy, border);
+  if (fail) {
+    vp9_denoiser_free(denoiser);
+    return 1;
+  }
+#ifdef OUTPUT_YUV_DENOISED
+  make_grayscale(&denoiser->running_avg_y[i]);
+#endif
+  denoiser->increase_denoising = 0;
+
+  return 0;
+}
+
+void vp9_denoiser_free(VP9_DENOISER *denoiser) {
+  int i;
+  if (denoiser == NULL) {
+    return;
+  }
+  for (i = 0; i < MAX_REF_FRAMES; ++i) {
+    if (&denoiser->running_avg_y[i] != NULL) {
+      vp9_free_frame_buffer(&denoiser->running_avg_y[i]);
+    }
+  }
+  if (&denoiser->mc_running_avg_y != NULL) {
+    vp9_free_frame_buffer(&denoiser->mc_running_avg_y);
+  }
+}
+
+#ifdef OUTPUT_YUV_DENOISED
+static void make_grayscale(YV12_BUFFER_CONFIG *yuv) {
+  int r, c;
+  uint8_t *u = yuv->u_buffer;
+  uint8_t *v = yuv->v_buffer;
+
+  // The '/2's are there because we have a 440 buffer, but we want to output
+  // 420.
+  for (r = 0; r < yuv->uv_height / 2; ++r) {
+    for (c = 0; c < yuv->uv_width / 2; ++c) {
+      u[c] = UINT8_MAX / 2;
+      v[c] = UINT8_MAX / 2;
+    }
+    u += yuv->uv_stride + yuv->uv_width / 2;
+    v += yuv->uv_stride + yuv->uv_width / 2;
+  }
+}
+#endif

vp9/encoder/vp9_denoiser.h

@@ -0,0 +1,63 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE 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.
+ */
+
+#ifndef VP9_ENCODER_DENOISER_H_
+#define VP9_ENCODER_DENOISER_H_
+
+#include "vp9/encoder/vp9_block.h"
+#include "vpx_scale/yv12config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum vp9_denoiser_decision {
+  COPY_BLOCK,
+  FILTER_BLOCK
+} VP9_DENOISER_DECISION;
+
+typedef struct vp9_denoiser {
+  YV12_BUFFER_CONFIG running_avg_y[MAX_REF_FRAMES];
+  YV12_BUFFER_CONFIG mc_running_avg_y;
+
+  unsigned int zero_mv_sse;
+  unsigned int best_sse;
+  int increase_denoising;
+  PREDICTION_MODE best_sse_inter_mode;
+  int_mv best_sse_mv;
+  MV_REFERENCE_FRAME best_reference_frame;
+  MV_REFERENCE_FRAME best_zeromv_reference_frame;
+} VP9_DENOISER;
+
+void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
+                                    YV12_BUFFER_CONFIG src,
+                                    FRAME_TYPE frame_type,
+                                    int refresh_alt_ref_frame,
+                                    int refresh_golden_frame,
+                                    int refresh_last_frame);
+
+void vp9_denoiser_denoise(VP9_DENOISER *denoiser, MACROBLOCK *mb,
+                          int mi_row, int mi_col, BLOCK_SIZE bs);
+
+void vp9_denoiser_reset_frame_stats(VP9_DENOISER *denoiser);
+
+void vp9_denoiser_update_frame_stats(VP9_DENOISER *denoiser, MB_MODE_INFO *mbmi,
+                                     unsigned int sse, PREDICTION_MODE mode);
+
+int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
+                       int ssx, int ssy, int border);
+
+void vp9_denoiser_free(VP9_DENOISER *denoiser);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_DENOISER_H_