generic-library/webp
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

dsp/lossless: split enc/dec functions

Browse Source 
adds lossless_enc*.c; reduces the size of the decode-only so: ~78K
w/gcc-4.8.2 on x86_64.

Change-Id: If5e4610b67d05eba5896bc64bab79e9df92b2092
This commit is contained in:
James Zern 2015-03-20 19:09:49 -07:00
parent 9064adc8a8
commit 553051f741
15 changed files with 1884 additions and 1597 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
Android.mk
View File

@ -53,7 +53,6 @@ dsp_dec_srcs := \
src/dsp/filters_mips_dsp_r2.c \
src/dsp/filters_sse2.c \
src/dsp/lossless.c \
src/dsp/lossless_mips32.c \
src/dsp/lossless_mips_dsp_r2.c \
src/dsp/lossless_neon.$(NEON) \
src/dsp/lossless_sse2.c \
@ -80,6 +79,11 @@ dsp_enc_srcs := \
src/dsp/enc_mips_dsp_r2.c \
src/dsp/enc_neon.$(NEON) \
src/dsp/enc_sse2.c \
src/dsp/lossless_enc.c \
src/dsp/lossless_enc_mips32.c \
src/dsp/lossless_enc_mips_dsp_r2.c \
src/dsp/lossless_enc_neon.$(NEON) \
src/dsp/lossless_enc_sse2.c \
enc_srcs := \
src/enc/alpha.c \

6
Makefile.vc
View File

@ -200,7 +200,6 @@ DSP_DEC_OBJS = \
$(DIROBJ)\dsp\filters_mips_dsp_r2.obj \
$(DIROBJ)\dsp\filters_sse2.obj \
$(DIROBJ)\dsp\lossless.obj \
$(DIROBJ)\dsp\lossless_mips32.obj \
$(DIROBJ)\dsp\lossless_mips_dsp_r2.obj \
$(DIROBJ)\dsp\lossless_neon.obj \
$(DIROBJ)\dsp\lossless_sse2.obj \
@ -230,6 +229,11 @@ DSP_ENC_OBJS = \
$(DIROBJ)\dsp\enc_mips_dsp_r2.obj \
$(DIROBJ)\dsp\enc_neon.obj \
$(DIROBJ)\dsp\enc_sse2.obj \
$(DIROBJ)\dsp\lossless_enc.obj \
$(DIROBJ)\dsp\lossless_enc_mips32.obj \
$(DIROBJ)\dsp\lossless_enc_mips_dsp_r2.obj \
$(DIROBJ)\dsp\lossless_enc_neon.obj \
$(DIROBJ)\dsp\lossless_enc_sse2.obj \
EX_FORMAT_DEC_OBJS = \
$(DIROBJ)\examples\jpegdec.obj \

6
makefile.unix
View File

@ -130,7 +130,6 @@ DSP_DEC_OBJS = \
src/dsp/filters_mips_dsp_r2.o \
src/dsp/filters_sse2.o \
src/dsp/lossless.o \
src/dsp/lossless_mips32.o \
src/dsp/lossless_mips_dsp_r2.o \
src/dsp/lossless_neon.o \
src/dsp/lossless_sse2.o \
@ -160,6 +159,11 @@ DSP_ENC_OBJS = \
src/dsp/enc_mips_dsp_r2.o \
src/dsp/enc_neon.o \
src/dsp/enc_sse2.o \
src/dsp/lossless_enc.o \
src/dsp/lossless_enc_mips32.o \
src/dsp/lossless_enc_mips_dsp_r2.o \
src/dsp/lossless_enc_neon.o \
src/dsp/lossless_enc_sse2.o \
ENC_OBJS = \
src/enc/alpha.o \

6
src/dsp/Makefile.am
View File

@ -23,7 +23,6 @@ COMMON_SOURCES += filters.c
COMMON_SOURCES += filters_mips_dsp_r2.c
COMMON_SOURCES += lossless.c
COMMON_SOURCES += lossless.h
COMMON_SOURCES += lossless_mips32.c
COMMON_SOURCES += lossless_mips_dsp_r2.c
COMMON_SOURCES += lossless_neon.c
COMMON_SOURCES += mips_macro.h
@ -49,6 +48,10 @@ ENC_SOURCES += enc.c
ENC_SOURCES += enc_mips32.c
ENC_SOURCES += enc_mips_dsp_r2.c
ENC_SOURCES += enc_neon.c
ENC_SOURCES += lossless_enc.c
ENC_SOURCES += lossless_enc_mips32.c
ENC_SOURCES += lossless_enc_mips_dsp_r2.c
ENC_SOURCES += lossless_enc_neon.c
libwebpdsp_avx2_la_SOURCES =
libwebpdsp_avx2_la_SOURCES += enc_avx2.c
@ -75,6 +78,7 @@ libwebpdsp_sse2_la_SOURCES =
libwebpdsp_sse2_la_SOURCES += argb_sse2.c
libwebpdsp_sse2_la_SOURCES += cost_sse2.c
libwebpdsp_sse2_la_SOURCES += enc_sse2.c
libwebpdsp_sse2_la_SOURCES += lossless_enc_sse2.c
libwebpdsp_sse2_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
libwebpdsp_sse2_la_CFLAGS = $(AM_CFLAGS) $(SSE2_FLAGS)
libwebpdsp_sse2_la_LIBADD = libwebpdspdecode_sse2.la

1186
src/dsp/lossless.c
View File

File diff suppressed because it is too large Load Diff

1
src/dsp/lossless.h
View File

@ -89,6 +89,7 @@ void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels);
// Must be called before calling any of the above methods.
void VP8LDspInit(void);
void VP8LEncDspInit(void);
//------------------------------------------------------------------------------
// Image transforms.

1294
src/dsp/lossless_enc.c Normal file
View File

File diff suppressed because it is too large Load Diff

10
src/dsp/lossless_mips32.c → src/dsp/lossless_enc_mips32.c
View File

@ -1,4 +1,4 @@
// Copyright 2014 Google Inc. All Rights Reserved.
// Copyright 2015 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
@ -399,9 +399,9 @@ static void HistogramAdd(const VP8LHistogram* const a,
//------------------------------------------------------------------------------
// Entry point
extern void VP8LDspInitMIPS32(void);
extern void VP8LEncDspInitMIPS32(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMIPS32(void) {
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPS32(void) {
VP8LFastSLog2Slow = FastSLog2Slow;
VP8LFastLog2Slow = FastLog2Slow;
VP8LExtraCost = ExtraCost;
@ -413,7 +413,7 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMIPS32(void) {
#else // !WEBP_USE_MIPS32
extern void VP8LDspInitMIPS32(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMIPS32(void) {}
extern void VP8LEncDspInitMIPS32(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPS32(void) {}
#endif // WEBP_USE_MIPS32

276
src/dsp/lossless_enc_mips_dsp_r2.c Normal file
View File

@ -0,0 +1,276 @@
// Copyright 2015 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// Image transform methods for lossless encoder.
//
// Author(s): Djordje Pesut (djordje.pesut@imgtec.com)
// Jovan Zelincevic (jovan.zelincevic@imgtec.com)
#include "./dsp.h"
#if defined(WEBP_USE_MIPS_DSP_R2)
#include "./lossless.h"
static void SubtractGreenFromBlueAndRed(uint32_t* argb_data,
int num_pixels) {
uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
uint32_t* const p_loop1_end = argb_data + (num_pixels & ~3);
uint32_t* const p_loop2_end = p_loop1_end + (num_pixels & 3);
__asm__ volatile (
".set push \n\t"
".set noreorder \n\t"
"beq %[argb_data], %[p_loop1_end], 3f \n\t"
" nop \n\t"
"0: \n\t"
"lw %[temp0], 0(%[argb_data]) \n\t"
"lw %[temp1], 4(%[argb_data]) \n\t"
"lw %[temp2], 8(%[argb_data]) \n\t"
"lw %[temp3], 12(%[argb_data]) \n\t"
"ext %[temp4], %[temp0], 8, 8 \n\t"
"ext %[temp5], %[temp1], 8, 8 \n\t"
"ext %[temp6], %[temp2], 8, 8 \n\t"
"ext %[temp7], %[temp3], 8, 8 \n\t"
"addiu %[argb_data], %[argb_data], 16 \n\t"
"replv.ph %[temp4], %[temp4] \n\t"
"replv.ph %[temp5], %[temp5] \n\t"
"replv.ph %[temp6], %[temp6] \n\t"
"replv.ph %[temp7], %[temp7] \n\t"
"subu.qb %[temp0], %[temp0], %[temp4] \n\t"
"subu.qb %[temp1], %[temp1], %[temp5] \n\t"
"subu.qb %[temp2], %[temp2], %[temp6] \n\t"
"subu.qb %[temp3], %[temp3], %[temp7] \n\t"
"sw %[temp0], -16(%[argb_data]) \n\t"
"sw %[temp1], -12(%[argb_data]) \n\t"
"sw %[temp2], -8(%[argb_data]) \n\t"
"bne %[argb_data], %[p_loop1_end], 0b \n\t"
" sw %[temp3], -4(%[argb_data]) \n\t"
"3: \n\t"
"beq %[argb_data], %[p_loop2_end], 2f \n\t"
" nop \n\t"
"1: \n\t"
"lw %[temp0], 0(%[argb_data]) \n\t"
"addiu %[argb_data], %[argb_data], 4 \n\t"
"ext %[temp4], %[temp0], 8, 8 \n\t"
"replv.ph %[temp4], %[temp4] \n\t"
"subu.qb %[temp0], %[temp0], %[temp4] \n\t"
"bne %[argb_data], %[p_loop2_end], 1b \n\t"
" sw %[temp0], -4(%[argb_data]) \n\t"
"2: \n\t"
".set pop \n\t"
: [argb_data]"+&r"(argb_data), [temp0]"=&r"(temp0),
[temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3),
[temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6),
[temp7]"=&r"(temp7)
: [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end)
: "memory"
);
}
static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred,
int8_t color) {
return (uint32_t)((int)(color_pred) * color) >> 5;
}
static void TransformColor(const VP8LMultipliers* const m, uint32_t* data,
int num_pixels) {
int temp0, temp1, temp2, temp3, temp4, temp5;
uint32_t argb, argb1, new_red, new_red1;
const uint32_t G_to_R = m->green_to_red_;
const uint32_t G_to_B = m->green_to_blue_;
const uint32_t R_to_B = m->red_to_blue_;
uint32_t* const p_loop_end = data + (num_pixels & ~1);
__asm__ volatile (
".set push \n\t"
".set noreorder \n\t"
"beq %[data], %[p_loop_end], 1f \n\t"
" nop \n\t"
"replv.ph %[temp0], %[G_to_R] \n\t"
"replv.ph %[temp1], %[G_to_B] \n\t"
"replv.ph %[temp2], %[R_to_B] \n\t"
"shll.ph %[temp0], %[temp0], 8 \n\t"
"shll.ph %[temp1], %[temp1], 8 \n\t"
"shll.ph %[temp2], %[temp2], 8 \n\t"
"shra.ph %[temp0], %[temp0], 8 \n\t"
"shra.ph %[temp1], %[temp1], 8 \n\t"
"shra.ph %[temp2], %[temp2], 8 \n\t"
"0: \n\t"
"lw %[argb], 0(%[data]) \n\t"
"lw %[argb1], 4(%[data]) \n\t"
"lhu %[new_red], 2(%[data]) \n\t"
"lhu %[new_red1], 6(%[data]) \n\t"
"precrq.qb.ph %[temp3], %[argb], %[argb1] \n\t"
"precr.qb.ph %[temp4], %[argb], %[argb1] \n\t"
"preceu.ph.qbra %[temp3], %[temp3] \n\t"
"preceu.ph.qbla %[temp4], %[temp4] \n\t"
"shll.ph %[temp3], %[temp3], 8 \n\t"
"shll.ph %[temp4], %[temp4], 8 \n\t"
"shra.ph %[temp3], %[temp3], 8 \n\t"
"shra.ph %[temp4], %[temp4], 8 \n\t"
"mul.ph %[temp5], %[temp3], %[temp0] \n\t"
"mul.ph %[temp3], %[temp3], %[temp1] \n\t"
"mul.ph %[temp4], %[temp4], %[temp2] \n\t"
"addiu %[data], %[data], 8 \n\t"
"ins %[new_red1], %[new_red], 16, 16 \n\t"
"ins %[argb1], %[argb], 16, 16 \n\t"
"shra.ph %[temp5], %[temp5], 5 \n\t"
"shra.ph %[temp3], %[temp3], 5 \n\t"
"shra.ph %[temp4], %[temp4], 5 \n\t"
"subu.ph %[new_red1], %[new_red1], %[temp5] \n\t"
"subu.ph %[argb1], %[argb1], %[temp3] \n\t"
"preceu.ph.qbra %[temp5], %[new_red1] \n\t"
"subu.ph %[argb1], %[argb1], %[temp4] \n\t"
"preceu.ph.qbra %[temp3], %[argb1] \n\t"
"sb %[temp5], -2(%[data]) \n\t"
"sb %[temp3], -4(%[data]) \n\t"
"sra %[temp5], %[temp5], 16 \n\t"
"sra %[temp3], %[temp3], 16 \n\t"
"sb %[temp5], -6(%[data]) \n\t"
"bne %[data], %[p_loop_end], 0b \n\t"
" sb %[temp3], -8(%[data]) \n\t"
"1: \n\t"
".set pop \n\t"
: [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
[temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
[new_red1]"=&r"(new_red1), [new_red]"=&r"(new_red),
[argb]"=&r"(argb), [argb1]"=&r"(argb1), [data]"+&r"(data)
: [G_to_R]"r"(G_to_R), [R_to_B]"r"(R_to_B),
[G_to_B]"r"(G_to_B), [p_loop_end]"r"(p_loop_end)
: "memory", "hi", "lo"
);
if (num_pixels & 1) {
const uint32_t argb_ = data[0];
const uint32_t green = argb_ >> 8;
const uint32_t red = argb_ >> 16;
uint32_t new_blue = argb_;
new_red = red;
new_red -= ColorTransformDelta(m->green_to_red_, green);
new_red &= 0xff;
new_blue -= ColorTransformDelta(m->green_to_blue_, green);
new_blue -= ColorTransformDelta(m->red_to_blue_, red);
new_blue &= 0xff;
data[0] = (argb_ & 0xff00ff00u) | (new_red << 16) | (new_blue);
}
}
static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue,
uint8_t red_to_blue,
uint32_t argb) {
const uint32_t green = argb >> 8;
const uint32_t red = argb >> 16;
uint8_t new_blue = argb;
new_blue -= ColorTransformDelta(green_to_blue, green);
new_blue -= ColorTransformDelta(red_to_blue, red);
return (new_blue & 0xff);
}
static void CollectColorBlueTransforms(const uint32_t* argb, int stride,
int tile_width, int tile_height,
int green_to_blue, int red_to_blue,
int histo[]) {
const int rtb = (red_to_blue << 16) | (red_to_blue & 0xffff);
const int gtb = (green_to_blue << 16) | (green_to_blue & 0xffff);
const uint32_t mask = 0xff00ffu;
while (tile_height-- > 0) {
int x;
const uint32_t* p_argb = argb;
argb += stride;
for (x = 0; x < (tile_width >> 1); ++x) {
int temp0, temp1, temp2, temp3, temp4, temp5, temp6;
__asm__ volatile (
"lw %[temp0], 0(%[p_argb]) \n\t"
"lw %[temp1], 4(%[p_argb]) \n\t"
"precr.qb.ph %[temp2], %[temp0], %[temp1] \n\t"
"ins %[temp1], %[temp0], 16, 16 \n\t"
"shra.ph %[temp2], %[temp2], 8 \n\t"
"shra.ph %[temp3], %[temp1], 8 \n\t"
"mul.ph %[temp5], %[temp2], %[rtb] \n\t"
"mul.ph %[temp6], %[temp3], %[gtb] \n\t"
"and %[temp4], %[temp1], %[mask] \n\t"
"addiu %[p_argb], %[p_argb], 8 \n\t"
"shra.ph %[temp5], %[temp5], 5 \n\t"
"shra.ph %[temp6], %[temp6], 5 \n\t"
"subu.qb %[temp2], %[temp4], %[temp5] \n\t"
"subu.qb %[temp2], %[temp2], %[temp6] \n\t"
: [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
[temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4),
[temp5]"=&r"(temp5), [temp6]"=&r"(temp6)
: [rtb]"r"(rtb), [gtb]"r"(gtb), [mask]"r"(mask)
: "memory", "hi", "lo"
);
++histo[(uint8_t)(temp2 >> 16)];
++histo[(uint8_t)temp2];
}
if (tile_width & 1) {
++histo[TransformColorBlue(green_to_blue, red_to_blue, *p_argb)];
}
}
}
static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red,
uint32_t argb) {
const uint32_t green = argb >> 8;
uint32_t new_red = argb >> 16;
new_red -= ColorTransformDelta(green_to_red, green);
return (new_red & 0xff);
}
static void CollectColorRedTransforms(const uint32_t* argb, int stride,
int tile_width, int tile_height,
int green_to_red, int histo[]) {
const int gtr = (green_to_red << 16) | (green_to_red & 0xffff);
while (tile_height-- > 0) {
int x;
const uint32_t* p_argb = argb;
argb += stride;
for (x = 0; x < (tile_width >> 1); ++x) {
int temp0, temp1, temp2, temp3, temp4;
__asm__ volatile (
"lw %[temp0], 0(%[p_argb]) \n\t"
"lw %[temp1], 4(%[p_argb]) \n\t"
"precrq.ph.w %[temp4], %[temp0], %[temp1] \n\t"
"ins %[temp1], %[temp0], 16, 16 \n\t"
"shra.ph %[temp3], %[temp1], 8 \n\t"
"mul.ph %[temp2], %[temp3], %[gtr] \n\t"
"addiu %[p_argb], %[p_argb], 8 \n\t"
"shra.ph %[temp2], %[temp2], 5 \n\t"
"subu.qb %[temp2], %[temp4], %[temp2] \n\t"
: [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
[temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4)
: [gtr]"r"(gtr)
: "memory", "hi", "lo"
);
++histo[(uint8_t)(temp2 >> 16)];
++histo[(uint8_t)temp2];
}
if (tile_width & 1) {
++histo[TransformColorRed(green_to_red, *p_argb)];
}
}
}
//------------------------------------------------------------------------------
// Entry point
extern void VP8LEncDspInitMIPSdspR2(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPSdspR2(void) {
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
VP8LTransformColor = TransformColor;
VP8LCollectColorBlueTransforms = CollectColorBlueTransforms;
VP8LCollectColorRedTransforms = CollectColorRedTransforms;
}
#else // !WEBP_USE_MIPS_DSP_R2
extern void VP8LEncDspInitMIPSdspR2(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPSdspR2(void) {}
#endif // WEBP_USE_MIPS_DSP_R2

93
src/dsp/lossless_enc_neon.c Normal file
View File

@ -0,0 +1,93 @@
// Copyright 2015 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// NEON variant of methods for lossless encoder
//
// Author: Skal (pascal.massimino@gmail.com)
#include "./dsp.h"
#if defined(WEBP_USE_NEON)
#include <arm_neon.h>
#include "./lossless.h"
#include "./neon.h"
#ifdef WEBP_USE_INTRINSICS
//------------------------------------------------------------------------------
// Subtract-Green Transform
// vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
// non-standard versions there.
#if defined(__APPLE__) && defined(__aarch64__) && \
defined(__apple_build_version__) && (__apple_build_version__< 6020037)
#define USE_VTBLQ
#endif
#ifdef USE_VTBLQ
// 255 = byte will be zeroed
static const uint8_t kGreenShuffle[16] = {
1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255
};
static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
const uint8x16_t shuffle) {
return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)),
vtbl1q_u8(argb, vget_high_u8(shuffle)));
}
#else // !USE_VTBLQ
// 255 = byte will be zeroed
static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255 };
static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
const uint8x8_t shuffle) {
return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
vtbl1_u8(vget_high_u8(argb), shuffle));
}
#endif // USE_VTBLQ
static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
const uint32_t* const end = argb_data + (num_pixels & ~3);
#ifdef USE_VTBLQ
const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
#else
const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
#endif
for (; argb_data < end; argb_data += 4) {
const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
const uint8x16_t greens = DoGreenShuffle(argb, shuffle);
vst1q_u8((uint8_t*)argb_data, vsubq_u8(argb, greens));
}
// fallthrough and finish off with plain-C
VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3);
}
#undef USE_VTBLQ
#endif // WEBP_USE_INTRINSICS
//------------------------------------------------------------------------------
// Entry point
extern void VP8LEncDspInitNEON(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitNEON(void) {
#ifdef WEBP_USE_INTRINSICS
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
#endif
}
#else // !WEBP_USE_NEON
extern void VP8LEncDspInitNEON(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitNEON(void) {}
#endif // WEBP_USE_NEON

194
src/dsp/lossless_enc_sse2.c Normal file
View File

@ -0,0 +1,194 @@
// Copyright 2015 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
// SSE2 variant of methods for lossless encoder
//
// Author: Skal (pascal.massimino@gmail.com)
#include "./dsp.h"
#if defined(WEBP_USE_SSE2)
#include <assert.h>
#include <emmintrin.h>
#include "./lossless.h"
//------------------------------------------------------------------------------
// Subtract-Green Transform
static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
const __m128i mask = _mm_set1_epi32(0x0000ff00);
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
const __m128i in_00g0 = _mm_and_si128(in, mask); // 00g0|00g0|...
const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8); // 0g00|0g00|...
const __m128i in_000g = _mm_srli_epi32(in_00g0, 8); // 000g|000g|...
const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g);
const __m128i out = _mm_sub_epi8(in, in_0g0g);
_mm_storeu_si128((__m128i*)&argb_data[i], out);
}
// fallthrough and finish off with plain-C
VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
}
//------------------------------------------------------------------------------
// Color Transform
static WEBP_INLINE __m128i ColorTransformDelta(__m128i color_pred,
__m128i color) {
// We simulate signed 8-bit multiplication as:
// * Left shift the two (8-bit) numbers by 8 bits,
// * Perform a 16-bit signed multiplication and retain the higher 16-bits.
const __m128i color_pred_shifted = _mm_slli_epi32(color_pred, 8);
const __m128i color_shifted = _mm_slli_epi32(color, 8);
// Note: This performs multiplication on 8 packed 16-bit numbers, 4 of which
// happen to be zeroes.
const __m128i signed_mult =
_mm_mulhi_epi16(color_pred_shifted, color_shifted);
return _mm_srli_epi32(signed_mult, 5);
}
static WEBP_INLINE void TransformColor(const VP8LMultipliers* const m,
uint32_t* argb_data,
int num_pixels) {
const __m128i g_to_r = _mm_set1_epi32(m->green_to_red_); // multipliers
const __m128i g_to_b = _mm_set1_epi32(m->green_to_blue_);
const __m128i r_to_b = _mm_set1_epi32(m->red_to_blue_);
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
const __m128i alpha_green_mask = _mm_set1_epi32(0xff00ff00); // masks
const __m128i red_mask = _mm_set1_epi32(0x00ff0000);
const __m128i green_mask = _mm_set1_epi32(0x0000ff00);
const __m128i lower_8bit_mask = _mm_set1_epi32(0x000000ff);
const __m128i ag = _mm_and_si128(in, alpha_green_mask); // alpha, green
const __m128i r = _mm_srli_epi32(_mm_and_si128(in, red_mask), 16);
const __m128i g = _mm_srli_epi32(_mm_and_si128(in, green_mask), 8);
const __m128i b = in;
const __m128i r_delta = ColorTransformDelta(g_to_r, g); // red
const __m128i r_new =
_mm_and_si128(_mm_sub_epi32(r, r_delta), lower_8bit_mask);
const __m128i r_new_shifted = _mm_slli_epi32(r_new, 16);
const __m128i b_delta_1 = ColorTransformDelta(g_to_b, g); // blue
const __m128i b_delta_2 = ColorTransformDelta(r_to_b, r);
const __m128i b_delta = _mm_add_epi32(b_delta_1, b_delta_2);
const __m128i b_new =
_mm_and_si128(_mm_sub_epi32(b, b_delta), lower_8bit_mask);
const __m128i out = _mm_or_si128(_mm_or_si128(ag, r_new_shifted), b_new);
_mm_storeu_si128((__m128i*)&argb_data[i], out);
}
// Fall-back to C-version for left-overs.
VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
}
//------------------------------------------------------------------------------
#define LINE_SIZE 16 // 8 or 16
static void AddVector(const uint32_t* a, const uint32_t* b, uint32_t* out,
int size) {
int i;
assert(size % LINE_SIZE == 0);
for (i = 0; i < size; i += LINE_SIZE) {
const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
#if (LINE_SIZE == 16)
const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
#endif
const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]);
const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
#if (LINE_SIZE == 16)
const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i + 8]);
const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]);
#endif
_mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
_mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
#if (LINE_SIZE == 16)
_mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
_mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
#endif
}
}
static void AddVectorEq(const uint32_t* a, uint32_t* out, int size) {
int i;
assert(size % LINE_SIZE == 0);
for (i = 0; i < size; i += LINE_SIZE) {
const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
#if (LINE_SIZE == 16)
const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
#endif
const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i + 0]);
const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i + 4]);
#if (LINE_SIZE == 16)
const __m128i b2 = _mm_loadu_si128((const __m128i*)&out[i + 8]);
const __m128i b3 = _mm_loadu_si128((const __m128i*)&out[i + 12]);
#endif
_mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
_mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
#if (LINE_SIZE == 16)
_mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
_mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
#endif
}
}
#undef LINE_SIZE
// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
// that's ok since the histogram values are less than 1<<28 (max picture size).
static void HistogramAdd(const VP8LHistogram* const a,
const VP8LHistogram* const b,
VP8LHistogram* const out) {
int i;
const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_);
assert(a->palette_code_bits_ == b->palette_code_bits_);
if (b != out) {
AddVector(a->literal_, b->literal_, out->literal_, NUM_LITERAL_CODES);
AddVector(a->red_, b->red_, out->red_, NUM_LITERAL_CODES);
AddVector(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES);
AddVector(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES);
} else {
AddVectorEq(a->literal_, out->literal_, NUM_LITERAL_CODES);
AddVectorEq(a->red_, out->red_, NUM_LITERAL_CODES);
AddVectorEq(a->blue_, out->blue_, NUM_LITERAL_CODES);
AddVectorEq(a->alpha_, out->alpha_, NUM_LITERAL_CODES);
}
for (i = NUM_LITERAL_CODES; i < literal_size; ++i) {
out->literal_[i] = a->literal_[i] + b->literal_[i];
}
for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
out->distance_[i] = a->distance_[i] + b->distance_[i];
}
}
//------------------------------------------------------------------------------
// Entry point
extern void VP8LEncDspInitSSE2(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
VP8LTransformColor = TransformColor;
VP8LHistogramAdd = HistogramAdd;
}
#else // !WEBP_USE_SSE2
extern void VP8LEncDspInitSSE2(void);
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {}
#endif // WEBP_USE_SSE2

242
src/dsp/lossless_mips_dsp_r2.c
View File

@ -145,61 +145,6 @@ static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
return temp1;
}
static void SubtractGreenFromBlueAndRed(uint32_t* argb_data,
int num_pixels) {
uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
uint32_t* const p_loop1_end = argb_data + (num_pixels & ~3);
uint32_t* const p_loop2_end = p_loop1_end + (num_pixels & 3);
__asm__ volatile (
".set push \n\t"
".set noreorder \n\t"
"beq %[argb_data], %[p_loop1_end], 3f \n\t"
" nop \n\t"
"0: \n\t"
"lw %[temp0], 0(%[argb_data]) \n\t"
"lw %[temp1], 4(%[argb_data]) \n\t"
"lw %[temp2], 8(%[argb_data]) \n\t"
"lw %[temp3], 12(%[argb_data]) \n\t"
"ext %[temp4], %[temp0], 8, 8 \n\t"
"ext %[temp5], %[temp1], 8, 8 \n\t"
"ext %[temp6], %[temp2], 8, 8 \n\t"
"ext %[temp7], %[temp3], 8, 8 \n\t"
"addiu %[argb_data], %[argb_data], 16 \n\t"
"replv.ph %[temp4], %[temp4] \n\t"
"replv.ph %[temp5], %[temp5] \n\t"
"replv.ph %[temp6], %[temp6] \n\t"
"replv.ph %[temp7], %[temp7] \n\t"
"subu.qb %[temp0], %[temp0], %[temp4] \n\t"
"subu.qb %[temp1], %[temp1], %[temp5] \n\t"
"subu.qb %[temp2], %[temp2], %[temp6] \n\t"
"subu.qb %[temp3], %[temp3], %[temp7] \n\t"
"sw %[temp0], -16(%[argb_data]) \n\t"
"sw %[temp1], -12(%[argb_data]) \n\t"
"sw %[temp2], -8(%[argb_data]) \n\t"
"bne %[argb_data], %[p_loop1_end], 0b \n\t"
" sw %[temp3], -4(%[argb_data]) \n\t"
"3: \n\t"
"beq %[argb_data], %[p_loop2_end], 2f \n\t"
" nop \n\t"
"1: \n\t"
"lw %[temp0], 0(%[argb_data]) \n\t"
"addiu %[argb_data], %[argb_data], 4 \n\t"
"ext %[temp4], %[temp0], 8, 8 \n\t"
"replv.ph %[temp4], %[temp4] \n\t"
"subu.qb %[temp0], %[temp0], %[temp4] \n\t"
"bne %[argb_data], %[p_loop2_end], 1b \n\t"
" sw %[temp0], -4(%[argb_data]) \n\t"
"2: \n\t"
".set pop \n\t"
: [argb_data]"+&r"(argb_data), [temp0]"=&r"(temp0),
[temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3),
[temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6),
[temp7]"=&r"(temp7)
: [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end)
: "memory"
);
}
static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
int temp0, temp1, temp2, temp3, temp4, temp5;
__asm__ volatile (
@ -280,189 +225,6 @@ static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
return ClampedAddSubtractHalf(left, top[0], top[-1]);
}
static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred,
int8_t color) {
return (uint32_t)((int)(color_pred) * color) >> 5;
}
static void TransformColor(const VP8LMultipliers* const m, uint32_t* data,
int num_pixels) {
int temp0, temp1, temp2, temp3, temp4, temp5;
uint32_t argb, argb1, new_red, new_red1;
const uint32_t G_to_R = m->green_to_red_;
const uint32_t G_to_B = m->green_to_blue_;
const uint32_t R_to_B = m->red_to_blue_;
uint32_t* const p_loop_end = data + (num_pixels & ~1);
__asm__ volatile (
".set push \n\t"
".set noreorder \n\t"
"beq %[data], %[p_loop_end], 1f \n\t"
" nop \n\t"
"replv.ph %[temp0], %[G_to_R] \n\t"
"replv.ph %[temp1], %[G_to_B] \n\t"
"replv.ph %[temp2], %[R_to_B] \n\t"
"shll.ph %[temp0], %[temp0], 8 \n\t"
"shll.ph %[temp1], %[temp1], 8 \n\t"
"shll.ph %[temp2], %[temp2], 8 \n\t"
"shra.ph %[temp0], %[temp0], 8 \n\t"
"shra.ph %[temp1], %[temp1], 8 \n\t"
"shra.ph %[temp2], %[temp2], 8 \n\t"
"0: \n\t"
"lw %[argb], 0(%[data]) \n\t"
"lw %[argb1], 4(%[data]) \n\t"
"lhu %[new_red], 2(%[data]) \n\t"
"lhu %[new_red1], 6(%[data]) \n\t"
"precrq.qb.ph %[temp3], %[argb], %[argb1] \n\t"
"precr.qb.ph %[temp4], %[argb], %[argb1] \n\t"
"preceu.ph.qbra %[temp3], %[temp3] \n\t"
"preceu.ph.qbla %[temp4], %[temp4] \n\t"
"shll.ph %[temp3], %[temp3], 8 \n\t"
"shll.ph %[temp4], %[temp4], 8 \n\t"
"shra.ph %[temp3], %[temp3], 8 \n\t"
"shra.ph %[temp4], %[temp4], 8 \n\t"
"mul.ph %[temp5], %[temp3], %[temp0] \n\t"
"mul.ph %[temp3], %[temp3], %[temp1] \n\t"
"mul.ph %[temp4], %[temp4], %[temp2] \n\t"
"addiu %[data], %[data], 8 \n\t"
"ins %[new_red1], %[new_red], 16, 16 \n\t"
"ins %[argb1], %[argb], 16, 16 \n\t"
"shra.ph %[temp5], %[temp5], 5 \n\t"
"shra.ph %[temp3], %[temp3], 5 \n\t"
"shra.ph %[temp4], %[temp4], 5 \n\t"
"subu.ph %[new_red1], %[new_red1], %[temp5] \n\t"
"subu.ph %[argb1], %[argb1], %[temp3] \n\t"
"preceu.ph.qbra %[temp5], %[new_red1] \n\t"
"subu.ph %[argb1], %[argb1], %[temp4] \n\t"
"preceu.ph.qbra %[temp3], %[argb1] \n\t"
"sb %[temp5], -2(%[data]) \n\t"
"sb %[temp3], -4(%[data]) \n\t"
"sra %[temp5], %[temp5], 16 \n\t"
"sra %[temp3], %[temp3], 16 \n\t"
"sb %[temp5], -6(%[data]) \n\t"
"bne %[data], %[p_loop_end], 0b \n\t"
" sb %[temp3], -8(%[data]) \n\t"
"1: \n\t"
".set pop \n\t"
: [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
[temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
[new_red1]"=&r"(new_red1), [new_red]"=&r"(new_red),
[argb]"=&r"(argb), [argb1]"=&r"(argb1), [data]"+&r"(data)
: [G_to_R]"r"(G_to_R), [R_to_B]"r"(R_to_B),
[G_to_B]"r"(G_to_B), [p_loop_end]"r"(p_loop_end)
: "memory", "hi", "lo"
);
if (num_pixels & 1) {
const uint32_t argb_ = data[0];
const uint32_t green = argb_ >> 8;
const uint32_t red = argb_ >> 16;
uint32_t new_blue = argb_;
new_red = red;
new_red -= ColorTransformDelta(m->green_to_red_, green);
new_red &= 0xff;
new_blue -= ColorTransformDelta(m->green_to_blue_, green);
new_blue -= ColorTransformDelta(m->red_to_blue_, red);
new_blue &= 0xff;
data[0] = (argb_ & 0xff00ff00u) | (new_red << 16) | (new_blue);
}
}
static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue,
uint8_t red_to_blue,
uint32_t argb) {
const uint32_t green = argb >> 8;
const uint32_t red = argb >> 16;
uint8_t new_blue = argb;
new_blue -= ColorTransformDelta(green_to_blue, green);
new_blue -= ColorTransformDelta(red_to_blue, red);
return (new_blue & 0xff);
}
static void CollectColorBlueTransforms(const uint32_t* argb, int stride,
int tile_width, int tile_height,
int green_to_blue, int red_to_blue,
int histo[]) {
const int rtb = (red_to_blue << 16) | (red_to_blue & 0xffff);
const int gtb = (green_to_blue << 16) | (green_to_blue & 0xffff);
const uint32_t mask = 0xff00ffu;
while (tile_height-- > 0) {
int x;
const uint32_t* p_argb = argb;
argb += stride;
for (x = 0; x < (tile_width >> 1); ++x) {
int temp0, temp1, temp2, temp3, temp4, temp5, temp6;
__asm__ volatile (
"lw %[temp0], 0(%[p_argb]) \n\t"
"lw %[temp1], 4(%[p_argb]) \n\t"
"precr.qb.ph %[temp2], %[temp0], %[temp1] \n\t"
"ins %[temp1], %[temp0], 16, 16 \n\t"
"shra.ph %[temp2], %[temp2], 8 \n\t"
"shra.ph %[temp3], %[temp1], 8 \n\t"
"mul.ph %[temp5], %[temp2], %[rtb] \n\t"
"mul.ph %[temp6], %[temp3], %[gtb] \n\t"
"and %[temp4], %[temp1], %[mask] \n\t"
"addiu %[p_argb], %[p_argb], 8 \n\t"
"shra.ph %[temp5], %[temp5], 5 \n\t"
"shra.ph %[temp6], %[temp6], 5 \n\t"
"subu.qb %[temp2], %[temp4], %[temp5] \n\t"
"subu.qb %[temp2], %[temp2], %[temp6] \n\t"
: [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
[temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4),
[temp5]"=&r"(temp5), [temp6]"=&r"(temp6)
: [rtb]"r"(rtb), [gtb]"r"(gtb), [mask]"r"(mask)
: "memory", "hi", "lo"
);
++histo[(uint8_t)(temp2 >> 16)];
++histo[(uint8_t)temp2];
}
if (tile_width & 1) {
++histo[TransformColorBlue(green_to_blue, red_to_blue, *p_argb)];
}
}
}
static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red,
uint32_t argb) {
const uint32_t green = argb >> 8;
uint32_t new_red = argb >> 16;
new_red -= ColorTransformDelta(green_to_red, green);
return (new_red & 0xff);
}
static void CollectColorRedTransforms(const uint32_t* argb, int stride,
int tile_width, int tile_height,
int green_to_red, int histo[]) {
const int gtr = (green_to_red << 16) | (green_to_red & 0xffff);
while (tile_height-- > 0) {
int x;
const uint32_t* p_argb = argb;
argb += stride;
for (x = 0; x < (tile_width >> 1); ++x) {
int temp0, temp1, temp2, temp3, temp4;
__asm__ volatile (
"lw %[temp0], 0(%[p_argb]) \n\t"
"lw %[temp1], 4(%[p_argb]) \n\t"
"precrq.ph.w %[temp4], %[temp0], %[temp1] \n\t"
"ins %[temp1], %[temp0], 16, 16 \n\t"
"shra.ph %[temp3], %[temp1], 8 \n\t"
"mul.ph %[temp2], %[temp3], %[gtr] \n\t"
"addiu %[p_argb], %[p_argb], 8 \n\t"
"shra.ph %[temp2], %[temp2], 5 \n\t"
"subu.qb %[temp2], %[temp4], %[temp2] \n\t"
: [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
[temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4)
: [gtr]"r"(gtr)
: "memory", "hi", "lo"
);
++histo[(uint8_t)(temp2 >> 16)];
++histo[(uint8_t)temp2];
}
if (tile_width & 1) {
++histo[TransformColorRed(green_to_red, *p_argb)];
}
}
}
// Add green to blue and red channels (i.e. perform the inverse transform of
// 'subtract green').
static void AddGreenToBlueAndRed(uint32_t* data, int num_pixels) {
@ -902,10 +664,6 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMIPSdspR2(void) {
VP8LPredictors[11] = Predictor11;
VP8LPredictors[12] = Predictor12;
VP8LPredictors[13] = Predictor13;
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
VP8LTransformColor = TransformColor;
VP8LCollectColorBlueTransforms = CollectColorBlueTransforms;
VP8LCollectColorRedTransforms = CollectColorRedTransforms;
VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
VP8LTransformColorInverse = TransformColorInverse;
VP8LConvertBGRAToRGB = ConvertBGRAToRGB;

17
src/dsp/lossless_neon.c
View File

@ -288,22 +288,6 @@ static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
}
#endif // USE_VTBLQ
static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
const uint32_t* const end = argb_data + (num_pixels & ~3);
#ifdef USE_VTBLQ
const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
#else
const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
#endif
for (; argb_data < end; argb_data += 4) {
const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
const uint8x16_t greens = DoGreenShuffle(argb, shuffle);
vst1q_u8((uint8_t*)argb_data, vsubq_u8(argb, greens));
}
// fallthrough and finish off with plain-C
VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3);
}
static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
const uint32_t* const end = argb_data + (num_pixels & ~3);
#ifdef USE_VTBLQ
@ -345,7 +329,6 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitNEON(void) {
VP8LPredictors[12] = Predictor12;
VP8LPredictors[13] = Predictor13;
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
#endif
}

142
src/dsp/lossless_sse2.c
View File

@ -155,22 +155,6 @@ static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
//------------------------------------------------------------------------------
// Subtract-Green Transform
static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
const __m128i mask = _mm_set1_epi32(0x0000ff00);
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
const __m128i in_00g0 = _mm_and_si128(in, mask); // 00g0|00g0|...
const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8); // 0g00|0g00|...
const __m128i in_000g = _mm_srli_epi32(in_00g0, 8); // 000g|000g|...
const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g);
const __m128i out = _mm_sub_epi8(in, in_0g0g);
_mm_storeu_si128((__m128i*)&argb_data[i], out);
}
// fallthrough and finish off with plain-C
VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
}
static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
const __m128i mask = _mm_set1_epi32(0x0000ff00);
int i;
@ -204,45 +188,6 @@ static WEBP_INLINE __m128i ColorTransformDelta(__m128i color_pred,
return _mm_srli_epi32(signed_mult, 5);
}
static WEBP_INLINE void TransformColor(const VP8LMultipliers* const m,
uint32_t* argb_data,
int num_pixels) {
const __m128i g_to_r = _mm_set1_epi32(m->green_to_red_); // multipliers
const __m128i g_to_b = _mm_set1_epi32(m->green_to_blue_);
const __m128i r_to_b = _mm_set1_epi32(m->red_to_blue_);
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
const __m128i alpha_green_mask = _mm_set1_epi32(0xff00ff00); // masks
const __m128i red_mask = _mm_set1_epi32(0x00ff0000);
const __m128i green_mask = _mm_set1_epi32(0x0000ff00);
const __m128i lower_8bit_mask = _mm_set1_epi32(0x000000ff);
const __m128i ag = _mm_and_si128(in, alpha_green_mask); // alpha, green
const __m128i r = _mm_srli_epi32(_mm_and_si128(in, red_mask), 16);
const __m128i g = _mm_srli_epi32(_mm_and_si128(in, green_mask), 8);
const __m128i b = in;
const __m128i r_delta = ColorTransformDelta(g_to_r, g); // red
const __m128i r_new =
_mm_and_si128(_mm_sub_epi32(r, r_delta), lower_8bit_mask);
const __m128i r_new_shifted = _mm_slli_epi32(r_new, 16);
const __m128i b_delta_1 = ColorTransformDelta(g_to_b, g); // blue
const __m128i b_delta_2 = ColorTransformDelta(r_to_b, r);
const __m128i b_delta = _mm_add_epi32(b_delta_1, b_delta_2);
const __m128i b_new =
_mm_and_si128(_mm_sub_epi32(b, b_delta), lower_8bit_mask);
const __m128i out = _mm_or_si128(_mm_or_si128(ag, r_new_shifted), b_new);
_mm_storeu_si128((__m128i*)&argb_data[i], out);
}
// Fall-back to C-version for left-overs.
VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
}
static WEBP_INLINE void TransformColorInverse(const VP8LMultipliers* const m,
uint32_t* argb_data,
int num_pixels) {
@ -416,88 +361,6 @@ static void ConvertBGRAToBGR(const uint32_t* src,
VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst);
}
//------------------------------------------------------------------------------
#define LINE_SIZE 16 // 8 or 16
static void AddVector(const uint32_t* a, const uint32_t* b, uint32_t* out,
int size) {
int i;
assert(size % LINE_SIZE == 0);
for (i = 0; i < size; i += LINE_SIZE) {
const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
#if (LINE_SIZE == 16)
const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
#endif
const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]);
const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
#if (LINE_SIZE == 16)
const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i + 8]);
const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]);
#endif
_mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
_mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
#if (LINE_SIZE == 16)
_mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
_mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
#endif
}
}
static void AddVectorEq(const uint32_t* a, uint32_t* out, int size) {
int i;
assert(size % LINE_SIZE == 0);
for (i = 0; i < size; i += LINE_SIZE) {
const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
#if (LINE_SIZE == 16)
const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
#endif
const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i + 0]);
const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i + 4]);
#if (LINE_SIZE == 16)
const __m128i b2 = _mm_loadu_si128((const __m128i*)&out[i + 8]);
const __m128i b3 = _mm_loadu_si128((const __m128i*)&out[i + 12]);
#endif
_mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
_mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
#if (LINE_SIZE == 16)
_mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
_mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
#endif
}
}
#undef LINE_SIZE
// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
// that's ok since the histogram values are less than 1<<28 (max picture size).
static void HistogramAdd(const VP8LHistogram* const a,
const VP8LHistogram* const b,
VP8LHistogram* const out) {
int i;
const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_);
assert(a->palette_code_bits_ == b->palette_code_bits_);
if (b != out) {
AddVector(a->literal_, b->literal_, out->literal_, NUM_LITERAL_CODES);
AddVector(a->red_, b->red_, out->red_, NUM_LITERAL_CODES);
AddVector(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES);
AddVector(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES);
} else {
AddVectorEq(a->literal_, out->literal_, NUM_LITERAL_CODES);
AddVectorEq(a->red_, out->red_, NUM_LITERAL_CODES);
AddVectorEq(a->blue_, out->blue_, NUM_LITERAL_CODES);
AddVectorEq(a->alpha_, out->alpha_, NUM_LITERAL_CODES);
}
for (i = NUM_LITERAL_CODES; i < literal_size; ++i) {
out->literal_[i] = a->literal_[i] + b->literal_[i];
}
for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
out->distance_[i] = a->distance_[i] + b->distance_[i];
}
}
//------------------------------------------------------------------------------
// Entry point
@ -514,18 +377,13 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE2(void) {
VP8LPredictors[12] = Predictor12;
VP8LPredictors[13] = Predictor13;
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
VP8LTransformColor = TransformColor;
VP8LTransformColorInverse = TransformColorInverse;
VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;
VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444;
VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565;
VP8LConvertBGRAToBGR = ConvertBGRAToBGR;
VP8LHistogramAdd = HistogramAdd;
}
#else // !WEBP_USE_SSE2

2
src/enc/vp8l.c
View File

@ -1216,7 +1216,7 @@ static VP8LEncoder* VP8LEncoderNew(const WebPConfig* const config,
enc->config_ = config;
enc->pic_ = picture;
VP8LDspInit();
VP8LEncDspInit();
return enc;
}