mirror of https://github.com/opencv/opencv.git
Commit: 698b8844e3
Tag: https://github.com/webmproject/libwebp/releases/tag/v1.0.0
pull/11690/head
Peter Rekdal Sunde
7 years ago
committed by
Alexander Alekhin
parent
b2cff44084
commit
9492f46d77
51 changed files with 1572 additions and 1093 deletions
@ -1,68 +0,0 @@ |
||||
// Copyright 2014 Google Inc. All Rights Reserved.
|
||||
//
|
||||
// Use of this source code is governed by a BSD-style license
|
||||
// that can be found in the COPYING file in the root of the source
|
||||
// tree. An additional intellectual property rights grant can be found
|
||||
// in the file PATENTS. All contributing project authors may
|
||||
// be found in the AUTHORS file in the root of the source tree.
|
||||
// -----------------------------------------------------------------------------
|
||||
//
|
||||
// ARGB making functions.
|
||||
//
|
||||
// Author: Djordje Pesut (djordje.pesut@imgtec.com)
|
||||
|
||||
#include "./dsp.h" |
||||
|
||||
static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) { |
||||
return (((uint32_t)a << 24) | (r << 16) | (g << 8) | b); |
||||
} |
||||
|
||||
static void PackARGB(const uint8_t* a, const uint8_t* r, const uint8_t* g, |
||||
const uint8_t* b, int len, uint32_t* out) { |
||||
int i; |
||||
for (i = 0; i < len; ++i) { |
||||
out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]); |
||||
} |
||||
} |
||||
|
||||
static void PackRGB(const uint8_t* r, const uint8_t* g, const uint8_t* b, |
||||
int len, int step, uint32_t* out) { |
||||
int i, offset = 0; |
||||
for (i = 0; i < len; ++i) { |
||||
out[i] = MakeARGB32(0xff, r[offset], g[offset], b[offset]); |
||||
offset += step; |
||||
} |
||||
} |
||||
|
||||
void (*VP8PackARGB)(const uint8_t*, const uint8_t*, const uint8_t*, |
||||
const uint8_t*, int, uint32_t*); |
||||
void (*VP8PackRGB)(const uint8_t*, const uint8_t*, const uint8_t*, |
||||
int, int, uint32_t*); |
||||
|
||||
extern void VP8EncDspARGBInitMIPSdspR2(void); |
||||
extern void VP8EncDspARGBInitSSE2(void); |
||||
|
||||
static volatile VP8CPUInfo argb_last_cpuinfo_used = |
||||
(VP8CPUInfo)&argb_last_cpuinfo_used; |
||||
|
||||
WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspARGBInit(void) { |
||||
if (argb_last_cpuinfo_used == VP8GetCPUInfo) return; |
||||
|
||||
VP8PackARGB = PackARGB; |
||||
VP8PackRGB = PackRGB; |
||||
|
||||
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
|
||||
if (VP8GetCPUInfo != NULL) { |
||||
#if defined(WEBP_USE_SSE2) |
||||
if (VP8GetCPUInfo(kSSE2)) { |
||||
VP8EncDspARGBInitSSE2(); |
||||
} |
||||
#endif |
||||
#if defined(WEBP_USE_MIPS_DSP_R2) |
||||
if (VP8GetCPUInfo(kMIPSdspR2)) { |
||||
VP8EncDspARGBInitMIPSdspR2(); |
||||
} |
||||
#endif |
||||
} |
||||
argb_last_cpuinfo_used = VP8GetCPUInfo; |
||||
} |
||||
@ -1,110 +0,0 @@ |
||||
// Copyright 2014 Google Inc. All Rights Reserved.
|
||||
//
|
||||
// Use of this source code is governed by a BSD-style license
|
||||
// that can be found in the COPYING file in the root of the source
|
||||
// tree. An additional intellectual property rights grant can be found
|
||||
// in the file PATENTS. All contributing project authors may
|
||||
// be found in the AUTHORS file in the root of the source tree.
|
||||
// -----------------------------------------------------------------------------
|
||||
//
|
||||
// ARGB making functions (mips version).
|
||||
//
|
||||
// Author: Djordje Pesut (djordje.pesut@imgtec.com)
|
||||
|
||||
#include "./dsp.h" |
||||
|
||||
#if defined(WEBP_USE_MIPS_DSP_R2) |
||||
|
||||
static void PackARGB(const uint8_t* a, const uint8_t* r, const uint8_t* g, |
||||
const uint8_t* b, int len, uint32_t* out) { |
||||
int temp0, temp1, temp2, temp3, offset; |
||||
const int rest = len & 1; |
||||
const uint32_t* const loop_end = out + len - rest; |
||||
const int step = 4; |
||||
__asm__ volatile ( |
||||
"xor %[offset], %[offset], %[offset] \n\t" |
||||
"beq %[loop_end], %[out], 0f \n\t" |
||||
"2: \n\t" |
||||
"lbux %[temp0], %[offset](%[a]) \n\t" |
||||
"lbux %[temp1], %[offset](%[r]) \n\t" |
||||
"lbux %[temp2], %[offset](%[g]) \n\t" |
||||
"lbux %[temp3], %[offset](%[b]) \n\t" |
||||
"ins %[temp1], %[temp0], 16, 16 \n\t" |
||||
"ins %[temp3], %[temp2], 16, 16 \n\t" |
||||
"addiu %[out], %[out], 4 \n\t" |
||||
"precr.qb.ph %[temp0], %[temp1], %[temp3] \n\t" |
||||
"sw %[temp0], -4(%[out]) \n\t" |
||||
"addu %[offset], %[offset], %[step] \n\t" |
||||
"bne %[loop_end], %[out], 2b \n\t" |
||||
"0: \n\t" |
||||
"beq %[rest], $zero, 1f \n\t" |
||||
"lbux %[temp0], %[offset](%[a]) \n\t" |
||||
"lbux %[temp1], %[offset](%[r]) \n\t" |
||||
"lbux %[temp2], %[offset](%[g]) \n\t" |
||||
"lbux %[temp3], %[offset](%[b]) \n\t" |
||||
"ins %[temp1], %[temp0], 16, 16 \n\t" |
||||
"ins %[temp3], %[temp2], 16, 16 \n\t" |
||||
"precr.qb.ph %[temp0], %[temp1], %[temp3] \n\t" |
||||
"sw %[temp0], 0(%[out]) \n\t" |
||||
"1: \n\t" |
||||
: [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), |
||||
[temp3]"=&r"(temp3), [offset]"=&r"(offset), [out]"+&r"(out) |
||||
: [a]"r"(a), [r]"r"(r), [g]"r"(g), [b]"r"(b), [step]"r"(step), |
||||
[loop_end]"r"(loop_end), [rest]"r"(rest) |
||||
: "memory" |
||||
); |
||||
} |
||||
|
||||
static void PackRGB(const uint8_t* r, const uint8_t* g, const uint8_t* b, |
||||
int len, int step, uint32_t* out) { |
||||
int temp0, temp1, temp2, offset; |
||||
const int rest = len & 1; |
||||
const int a = 0xff; |
||||
const uint32_t* const loop_end = out + len - rest; |
||||
__asm__ volatile ( |
||||
"xor %[offset], %[offset], %[offset] \n\t" |
||||
"beq %[loop_end], %[out], 0f \n\t" |
||||
"2: \n\t" |
||||
"lbux %[temp0], %[offset](%[r]) \n\t" |
||||
"lbux %[temp1], %[offset](%[g]) \n\t" |
||||
"lbux %[temp2], %[offset](%[b]) \n\t" |
||||
"ins %[temp0], %[a], 16, 16 \n\t" |
||||
"ins %[temp2], %[temp1], 16, 16 \n\t" |
||||
"addiu %[out], %[out], 4 \n\t" |
||||
"precr.qb.ph %[temp0], %[temp0], %[temp2] \n\t" |
||||
"sw %[temp0], -4(%[out]) \n\t" |
||||
"addu %[offset], %[offset], %[step] \n\t" |
||||
"bne %[loop_end], %[out], 2b \n\t" |
||||
"0: \n\t" |
||||
"beq %[rest], $zero, 1f \n\t" |
||||
"lbux %[temp0], %[offset](%[r]) \n\t" |
||||
"lbux %[temp1], %[offset](%[g]) \n\t" |
||||
"lbux %[temp2], %[offset](%[b]) \n\t" |
||||
"ins %[temp0], %[a], 16, 16 \n\t" |
||||
"ins %[temp2], %[temp1], 16, 16 \n\t" |
||||
"precr.qb.ph %[temp0], %[temp0], %[temp2] \n\t" |
||||
"sw %[temp0], 0(%[out]) \n\t" |
||||
"1: \n\t" |
||||
: [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), |
||||
[offset]"=&r"(offset), [out]"+&r"(out) |
||||
: [a]"r"(a), [r]"r"(r), [g]"r"(g), [b]"r"(b), [step]"r"(step), |
||||
[loop_end]"r"(loop_end), [rest]"r"(rest) |
||||
: "memory" |
||||
); |
||||
} |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Entry point
|
||||
|
||||
extern void VP8EncDspARGBInitMIPSdspR2(void); |
||||
|
||||
WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspARGBInitMIPSdspR2(void) { |
||||
VP8PackARGB = PackARGB; |
||||
VP8PackRGB = PackRGB; |
||||
} |
||||
|
||||
#else // !WEBP_USE_MIPS_DSP_R2
|
||||
|
||||
WEBP_DSP_INIT_STUB(VP8EncDspARGBInitMIPSdspR2) |
||||
|
||||
#endif // WEBP_USE_MIPS_DSP_R2
|
||||
@ -1,70 +0,0 @@ |
||||
// Copyright 2014 Google Inc. All Rights Reserved.
|
||||
//
|
||||
// Use of this source code is governed by a BSD-style license
|
||||
// that can be found in the COPYING file in the root of the source
|
||||
// tree. An additional intellectual property rights grant can be found
|
||||
// in the file PATENTS. All contributing project authors may
|
||||
// be found in the AUTHORS file in the root of the source tree.
|
||||
// -----------------------------------------------------------------------------
|
||||
//
|
||||
// ARGB making functions (SSE2 version).
|
||||
//
|
||||
// Author: Skal (pascal.massimino@gmail.com)
|
||||
|
||||
#include "./dsp.h" |
||||
|
||||
#if defined(WEBP_USE_SSE2) |
||||
|
||||
#include <assert.h> |
||||
#include <emmintrin.h> |
||||
#include <string.h> |
||||
|
||||
static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) { |
||||
return (((uint32_t)a << 24) | (r << 16) | (g << 8) | b); |
||||
} |
||||
|
||||
static void PackARGB(const uint8_t* a, const uint8_t* r, const uint8_t* g, |
||||
const uint8_t* b, int len, uint32_t* out) { |
||||
if (g == r + 1) { // RGBA input order. Need to swap R and B.
|
||||
int i = 0; |
||||
const int len_max = len & ~3; // max length processed in main loop
|
||||
const __m128i red_blue_mask = _mm_set1_epi32(0x00ff00ffu); |
||||
assert(b == r + 2); |
||||
assert(a == r + 3); |
||||
for (; i < len_max; i += 4) { |
||||
const __m128i A = _mm_loadu_si128((const __m128i*)(r + 4 * i)); |
||||
const __m128i B = _mm_and_si128(A, red_blue_mask); // R 0 B 0
|
||||
const __m128i C = _mm_andnot_si128(red_blue_mask, A); // 0 G 0 A
|
||||
const __m128i D = _mm_shufflelo_epi16(B, _MM_SHUFFLE(2, 3, 0, 1)); |
||||
const __m128i E = _mm_shufflehi_epi16(D, _MM_SHUFFLE(2, 3, 0, 1)); |
||||
const __m128i F = _mm_or_si128(E, C); |
||||
_mm_storeu_si128((__m128i*)(out + i), F); |
||||
} |
||||
for (; i < len; ++i) { |
||||
out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]); |
||||
} |
||||
} else { |
||||
assert(g == b + 1); |
||||
assert(r == b + 2); |
||||
assert(a == b + 3); |
||||
memcpy(out, b, len * 4); |
||||
} |
||||
} |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Entry point
|
||||
|
||||
extern void VP8EncDspARGBInitSSE2(void); |
||||
|
||||
WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspARGBInitSSE2(void) { |
||||
extern void (*VP8PackARGB)(const uint8_t*, const uint8_t*, const uint8_t*, |
||||
const uint8_t*, int, uint32_t*); |
||||
|
||||
VP8PackARGB = PackARGB; |
||||
} |
||||
|
||||
#else // !WEBP_USE_SSE2
|
||||
|
||||
WEBP_DSP_INIT_STUB(VP8EncDspARGBInitSSE2) |
||||
|
||||
#endif // WEBP_USE_SSE2
|
||||
@ -0,0 +1,132 @@ |
||||
// Copyright 2016 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.
|
||||
// -----------------------------------------------------------------------------
|
||||
//
|
||||
// SSE4 code common to several files.
|
||||
//
|
||||
// Author: Vincent Rabaud (vrabaud@google.com)
|
||||
|
||||
#ifndef WEBP_DSP_COMMON_SSE41_H_ |
||||
#define WEBP_DSP_COMMON_SSE41_H_ |
||||
|
||||
#ifdef __cplusplus |
||||
extern "C" { |
||||
#endif |
||||
|
||||
#if defined(WEBP_USE_SSE41) |
||||
#include <smmintrin.h> |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Channel mixing.
|
||||
// Shuffles the input buffer as A0 0 0 A1 0 0 A2 ...
|
||||
#define WEBP_SSE41_SHUFF(OUT, IN0, IN1) \ |
||||
OUT##0 = _mm_shuffle_epi8(*IN0, shuff0); \
|
||||
OUT##1 = _mm_shuffle_epi8(*IN0, shuff1); \
|
||||
OUT##2 = _mm_shuffle_epi8(*IN0, shuff2); \
|
||||
OUT##3 = _mm_shuffle_epi8(*IN1, shuff0); \
|
||||
OUT##4 = _mm_shuffle_epi8(*IN1, shuff1); \
|
||||
OUT##5 = _mm_shuffle_epi8(*IN1, shuff2); |
||||
|
||||
// Pack the planar buffers
|
||||
// rrrr... rrrr... gggg... gggg... bbbb... bbbb....
|
||||
// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ...
|
||||
static WEBP_INLINE void VP8PlanarTo24b_SSE41( |
||||
__m128i* const in0, __m128i* const in1, __m128i* const in2, |
||||
__m128i* const in3, __m128i* const in4, __m128i* const in5) { |
||||
__m128i R0, R1, R2, R3, R4, R5; |
||||
__m128i G0, G1, G2, G3, G4, G5; |
||||
__m128i B0, B1, B2, B3, B4, B5; |
||||
|
||||
// Process R.
|
||||
{ |
||||
const __m128i shuff0 = _mm_set_epi8( |
||||
5, -1, -1, 4, -1, -1, 3, -1, -1, 2, -1, -1, 1, -1, -1, 0); |
||||
const __m128i shuff1 = _mm_set_epi8( |
||||
-1, 10, -1, -1, 9, -1, -1, 8, -1, -1, 7, -1, -1, 6, -1, -1); |
||||
const __m128i shuff2 = _mm_set_epi8( |
||||
-1, -1, 15, -1, -1, 14, -1, -1, 13, -1, -1, 12, -1, -1, 11, -1); |
||||
WEBP_SSE41_SHUFF(R, in0, in1) |
||||
} |
||||
|
||||
// Process G.
|
||||
{ |
||||
// Same as before, just shifted to the left by one and including the right
|
||||
// padding.
|
||||
const __m128i shuff0 = _mm_set_epi8( |
||||
-1, -1, 4, -1, -1, 3, -1, -1, 2, -1, -1, 1, -1, -1, 0, -1); |
||||
const __m128i shuff1 = _mm_set_epi8( |
||||
10, -1, -1, 9, -1, -1, 8, -1, -1, 7, -1, -1, 6, -1, -1, 5); |
||||
const __m128i shuff2 = _mm_set_epi8( |
||||
-1, 15, -1, -1, 14, -1, -1, 13, -1, -1, 12, -1, -1, 11, -1, -1); |
||||
WEBP_SSE41_SHUFF(G, in2, in3) |
||||
} |
||||
|
||||
// Process B.
|
||||
{ |
||||
const __m128i shuff0 = _mm_set_epi8( |
||||
-1, 4, -1, -1, 3, -1, -1, 2, -1, -1, 1, -1, -1, 0, -1, -1); |
||||
const __m128i shuff1 = _mm_set_epi8( |
||||
-1, -1, 9, -1, -1, 8, -1, -1, 7, -1, -1, 6, -1, -1, 5, -1); |
||||
const __m128i shuff2 = _mm_set_epi8( |
||||
15, -1, -1, 14, -1, -1, 13, -1, -1, 12, -1, -1, 11, -1, -1, 10); |
||||
WEBP_SSE41_SHUFF(B, in4, in5) |
||||
} |
||||
|
||||
// OR the different channels.
|
||||
{ |
||||
const __m128i RG0 = _mm_or_si128(R0, G0); |
||||
const __m128i RG1 = _mm_or_si128(R1, G1); |
||||
const __m128i RG2 = _mm_or_si128(R2, G2); |
||||
const __m128i RG3 = _mm_or_si128(R3, G3); |
||||
const __m128i RG4 = _mm_or_si128(R4, G4); |
||||
const __m128i RG5 = _mm_or_si128(R5, G5); |
||||
*in0 = _mm_or_si128(RG0, B0); |
||||
*in1 = _mm_or_si128(RG1, B1); |
||||
*in2 = _mm_or_si128(RG2, B2); |
||||
*in3 = _mm_or_si128(RG3, B3); |
||||
*in4 = _mm_or_si128(RG4, B4); |
||||
*in5 = _mm_or_si128(RG5, B5); |
||||
} |
||||
} |
||||
|
||||
#undef WEBP_SSE41_SHUFF |
||||
|
||||
// Convert four packed four-channel buffers like argbargbargbargb... into the
|
||||
// split channels aaaaa ... rrrr ... gggg .... bbbbb ......
|
||||
static WEBP_INLINE void VP8L32bToPlanar_SSE41(__m128i* const in0, |
||||
__m128i* const in1, |
||||
__m128i* const in2, |
||||
__m128i* const in3) { |
||||
// aaaarrrrggggbbbb
|
||||
const __m128i shuff0 = |
||||
_mm_set_epi8(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0); |
||||
const __m128i A0 = _mm_shuffle_epi8(*in0, shuff0); |
||||
const __m128i A1 = _mm_shuffle_epi8(*in1, shuff0); |
||||
const __m128i A2 = _mm_shuffle_epi8(*in2, shuff0); |
||||
const __m128i A3 = _mm_shuffle_epi8(*in3, shuff0); |
||||
// A0A1R0R1
|
||||
// G0G1B0B1
|
||||
// A2A3R2R3
|
||||
// G0G1B0B1
|
||||
const __m128i B0 = _mm_unpacklo_epi32(A0, A1); |
||||
const __m128i B1 = _mm_unpackhi_epi32(A0, A1); |
||||
const __m128i B2 = _mm_unpacklo_epi32(A2, A3); |
||||
const __m128i B3 = _mm_unpackhi_epi32(A2, A3); |
||||
*in3 = _mm_unpacklo_epi64(B0, B2); |
||||
*in2 = _mm_unpackhi_epi64(B0, B2); |
||||
*in1 = _mm_unpacklo_epi64(B1, B3); |
||||
*in0 = _mm_unpackhi_epi64(B1, B3); |
||||
} |
||||
|
||||
#endif // WEBP_USE_SSE41
|
||||
|
||||
#ifdef __cplusplus |
||||
} // extern "C"
|
||||
#endif |
||||
|
||||
#endif // WEBP_DSP_COMMON_SSE41_H_
|
||||
@ -0,0 +1,239 @@ |
||||
// Copyright 2011 Google Inc. All Rights Reserved.
|
||||
//
|
||||
// Use of this source code is governed by a BSD-style license
|
||||
// that can be found in the COPYING file in the root of the source
|
||||
// tree. An additional intellectual property rights grant can be found
|
||||
// in the file PATENTS. All contributing project authors may
|
||||
// be found in the AUTHORS file in the root of the source tree.
|
||||
// -----------------------------------------------------------------------------
|
||||
//
|
||||
// SSE41 version of YUV to RGB upsampling functions.
|
||||
//
|
||||
// Author: somnath@google.com (Somnath Banerjee)
|
||||
|
||||
#include "src/dsp/dsp.h" |
||||
|
||||
#if defined(WEBP_USE_SSE41) |
||||
|
||||
#include <assert.h> |
||||
#include <smmintrin.h> |
||||
#include <string.h> |
||||
#include "src/dsp/yuv.h" |
||||
|
||||
#ifdef FANCY_UPSAMPLING |
||||
|
||||
#if !defined(WEBP_REDUCE_CSP) |
||||
|
||||
// We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows
|
||||
// u = (9*a + 3*b + 3*c + d + 8) / 16
|
||||
// = (a + (a + 3*b + 3*c + d) / 8 + 1) / 2
|
||||
// = (a + m + 1) / 2
|
||||
// where m = (a + 3*b + 3*c + d) / 8
|
||||
// = ((a + b + c + d) / 2 + b + c) / 4
|
||||
//
|
||||
// Let's say k = (a + b + c + d) / 4.
|
||||
// We can compute k as
|
||||
// k = (s + t + 1) / 2 - ((a^d) | (b^c) | (s^t)) & 1
|
||||
// where s = (a + d + 1) / 2 and t = (b + c + 1) / 2
|
||||
//
|
||||
// Then m can be written as
|
||||
// m = (k + t + 1) / 2 - (((b^c) & (s^t)) | (k^t)) & 1
|
||||
|
||||
// Computes out = (k + in + 1) / 2 - ((ij & (s^t)) | (k^in)) & 1
|
||||
#define GET_M(ij, in, out) do { \ |
||||
const __m128i tmp0 = _mm_avg_epu8(k, (in)); /* (k + in + 1) / 2 */ \
|
||||
const __m128i tmp1 = _mm_and_si128((ij), st); /* (ij) & (s^t) */ \
|
||||
const __m128i tmp2 = _mm_xor_si128(k, (in)); /* (k^in) */ \
|
||||
const __m128i tmp3 = _mm_or_si128(tmp1, tmp2); /* ((ij) & (s^t)) | (k^in) */\
|
||||
const __m128i tmp4 = _mm_and_si128(tmp3, one); /* & 1 -> lsb_correction */ \
|
||||
(out) = _mm_sub_epi8(tmp0, tmp4); /* (k + in + 1) / 2 - lsb_correction */ \
|
||||
} while (0) |
||||
|
||||
// pack and store two alternating pixel rows
|
||||
#define PACK_AND_STORE(a, b, da, db, out) do { \ |
||||
const __m128i t_a = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \
|
||||
const __m128i t_b = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \
|
||||
const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b); \
|
||||
const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b); \
|
||||
_mm_store_si128(((__m128i*)(out)) + 0, t_1); \
|
||||
_mm_store_si128(((__m128i*)(out)) + 1, t_2); \
|
||||
} while (0) |
||||
|
||||
// Loads 17 pixels each from rows r1 and r2 and generates 32 pixels.
|
||||
#define UPSAMPLE_32PIXELS(r1, r2, out) { \ |
||||
const __m128i one = _mm_set1_epi8(1); \
|
||||
const __m128i a = _mm_loadu_si128((const __m128i*)&(r1)[0]); \
|
||||
const __m128i b = _mm_loadu_si128((const __m128i*)&(r1)[1]); \
|
||||
const __m128i c = _mm_loadu_si128((const __m128i*)&(r2)[0]); \
|
||||
const __m128i d = _mm_loadu_si128((const __m128i*)&(r2)[1]); \
|
||||
\
|
||||
const __m128i s = _mm_avg_epu8(a, d); /* s = (a + d + 1) / 2 */ \
|
||||
const __m128i t = _mm_avg_epu8(b, c); /* t = (b + c + 1) / 2 */ \
|
||||
const __m128i st = _mm_xor_si128(s, t); /* st = s^t */ \
|
||||
\
|
||||
const __m128i ad = _mm_xor_si128(a, d); /* ad = a^d */ \
|
||||
const __m128i bc = _mm_xor_si128(b, c); /* bc = b^c */ \
|
||||
\
|
||||
const __m128i t1 = _mm_or_si128(ad, bc); /* (a^d) | (b^c) */ \
|
||||
const __m128i t2 = _mm_or_si128(t1, st); /* (a^d) | (b^c) | (s^t) */ \
|
||||
const __m128i t3 = _mm_and_si128(t2, one); /* (a^d) | (b^c) | (s^t) & 1 */ \
|
||||
const __m128i t4 = _mm_avg_epu8(s, t); \
|
||||
const __m128i k = _mm_sub_epi8(t4, t3); /* k = (a + b + c + d) / 4 */ \
|
||||
__m128i diag1, diag2; \
|
||||
\
|
||||
GET_M(bc, t, diag1); /* diag1 = (a + 3b + 3c + d) / 8 */ \
|
||||
GET_M(ad, s, diag2); /* diag2 = (3a + b + c + 3d) / 8 */ \
|
||||
\
|
||||
/* pack the alternate pixels */ \
|
||||
PACK_AND_STORE(a, b, diag1, diag2, (out) + 0); /* store top */ \
|
||||
PACK_AND_STORE(c, d, diag2, diag1, (out) + 2 * 32); /* store bottom */ \
|
||||
} |
||||
|
||||
// Turn the macro into a function for reducing code-size when non-critical
|
||||
static void Upsample32Pixels_SSE41(const uint8_t r1[], const uint8_t r2[], |
||||
uint8_t* const out) { |
||||
UPSAMPLE_32PIXELS(r1, r2, out); |
||||
} |
||||
|
||||
#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) { \ |
||||
uint8_t r1[17], r2[17]; \
|
||||
memcpy(r1, (tb), (num_pixels)); \
|
||||
memcpy(r2, (bb), (num_pixels)); \
|
||||
/* replicate last byte */ \
|
||||
memset(r1 + (num_pixels), r1[(num_pixels) - 1], 17 - (num_pixels)); \
|
||||
memset(r2 + (num_pixels), r2[(num_pixels) - 1], 17 - (num_pixels)); \
|
||||
/* using the shared function instead of the macro saves ~3k code size */ \
|
||||
Upsample32Pixels_SSE41(r1, r2, out); \
|
||||
} |
||||
|
||||
#define CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, \ |
||||
top_dst, bottom_dst, cur_x) do { \
|
||||
FUNC##32_SSE41((top_y) + (cur_x), r_u, r_v, (top_dst) + (cur_x) * (XSTEP)); \
|
||||
if ((bottom_y) != NULL) { \
|
||||
FUNC##32_SSE41((bottom_y) + (cur_x), r_u + 64, r_v + 64, \
|
||||
(bottom_dst) + (cur_x) * (XSTEP)); \
|
||||
} \
|
||||
} while (0) |
||||
|
||||
#define SSE4_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ |
||||
static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \
|
||||
const uint8_t* top_u, const uint8_t* top_v, \
|
||||
const uint8_t* cur_u, const uint8_t* cur_v, \
|
||||
uint8_t* top_dst, uint8_t* bottom_dst, int len) { \
|
||||
int uv_pos, pos; \
|
||||
/* 16byte-aligned array to cache reconstructed u and v */ \
|
||||
uint8_t uv_buf[14 * 32 + 15] = { 0 }; \
|
||||
uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \
|
||||
uint8_t* const r_v = r_u + 32; \
|
||||
\
|
||||
assert(top_y != NULL); \
|
||||
{ /* Treat the first pixel in regular way */ \
|
||||
const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \
|
||||
const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \
|
||||
const int u0_t = (top_u[0] + u_diag) >> 1; \
|
||||
const int v0_t = (top_v[0] + v_diag) >> 1; \
|
||||
FUNC(top_y[0], u0_t, v0_t, top_dst); \
|
||||
if (bottom_y != NULL) { \
|
||||
const int u0_b = (cur_u[0] + u_diag) >> 1; \
|
||||
const int v0_b = (cur_v[0] + v_diag) >> 1; \
|
||||
FUNC(bottom_y[0], u0_b, v0_b, bottom_dst); \
|
||||
} \
|
||||
} \
|
||||
/* For UPSAMPLE_32PIXELS, 17 u/v values must be read-able for each block */ \
|
||||
for (pos = 1, uv_pos = 0; pos + 32 + 1 <= len; pos += 32, uv_pos += 16) { \
|
||||
UPSAMPLE_32PIXELS(top_u + uv_pos, cur_u + uv_pos, r_u); \
|
||||
UPSAMPLE_32PIXELS(top_v + uv_pos, cur_v + uv_pos, r_v); \
|
||||
CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst, pos); \
|
||||
} \
|
||||
if (len > 1) { \
|
||||
const int left_over = ((len + 1) >> 1) - (pos >> 1); \
|
||||
uint8_t* const tmp_top_dst = r_u + 4 * 32; \
|
||||
uint8_t* const tmp_bottom_dst = tmp_top_dst + 4 * 32; \
|
||||
uint8_t* const tmp_top = tmp_bottom_dst + 4 * 32; \
|
||||
uint8_t* const tmp_bottom = (bottom_y == NULL) ? NULL : tmp_top + 32; \
|
||||
assert(left_over > 0); \
|
||||
UPSAMPLE_LAST_BLOCK(top_u + uv_pos, cur_u + uv_pos, left_over, r_u); \
|
||||
UPSAMPLE_LAST_BLOCK(top_v + uv_pos, cur_v + uv_pos, left_over, r_v); \
|
||||
memcpy(tmp_top, top_y + pos, len - pos); \
|
||||
if (bottom_y != NULL) memcpy(tmp_bottom, bottom_y + pos, len - pos); \
|
||||
CONVERT2RGB_32(FUNC, XSTEP, tmp_top, tmp_bottom, tmp_top_dst, \
|
||||
tmp_bottom_dst, 0); \
|
||||
memcpy(top_dst + pos * (XSTEP), tmp_top_dst, (len - pos) * (XSTEP)); \
|
||||
if (bottom_y != NULL) { \
|
||||
memcpy(bottom_dst + pos * (XSTEP), tmp_bottom_dst, \
|
||||
(len - pos) * (XSTEP)); \
|
||||
} \
|
||||
} \
|
||||
} |
||||
|
||||
// SSE4 variants of the fancy upsampler.
|
||||
SSE4_UPSAMPLE_FUNC(UpsampleRgbLinePair_SSE41, VP8YuvToRgb, 3) |
||||
SSE4_UPSAMPLE_FUNC(UpsampleBgrLinePair_SSE41, VP8YuvToBgr, 3) |
||||
|
||||
#undef GET_M |
||||
#undef PACK_AND_STORE |
||||
#undef UPSAMPLE_32PIXELS |
||||
#undef UPSAMPLE_LAST_BLOCK |
||||
#undef CONVERT2RGB |
||||
#undef CONVERT2RGB_32 |
||||
#undef SSE4_UPSAMPLE_FUNC |
||||
|
||||
#endif // WEBP_REDUCE_CSP
|
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Entry point
|
||||
|
||||
extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; |
||||
|
||||
extern void WebPInitUpsamplersSSE41(void); |
||||
|
||||
WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersSSE41(void) { |
||||
#if !defined(WEBP_REDUCE_CSP) |
||||
WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair_SSE41; |
||||
WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair_SSE41; |
||||
#endif // WEBP_REDUCE_CSP
|
||||
} |
||||
|
||||
#endif // FANCY_UPSAMPLING
|
||||
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */]; |
||||
extern void WebPInitYUV444ConvertersSSE41(void); |
||||
|
||||
#define YUV444_FUNC(FUNC_NAME, CALL, CALL_C, XSTEP) \ |
||||
extern void CALL_C(const uint8_t* y, const uint8_t* u, const uint8_t* v, \
|
||||
uint8_t* dst, int len); \
|
||||
static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \
|
||||
uint8_t* dst, int len) { \
|
||||
int i; \
|
||||
const int max_len = len & ~31; \
|
||||
for (i = 0; i < max_len; i += 32) { \
|
||||
CALL(y + i, u + i, v + i, dst + i * (XSTEP)); \
|
||||
} \
|
||||
if (i < len) { /* C-fallback */ \
|
||||
CALL_C(y + i, u + i, v + i, dst + i * (XSTEP), len - i); \
|
||||
} \
|
||||
} |
||||
|
||||
#if !defined(WEBP_REDUCE_CSP) |
||||
YUV444_FUNC(Yuv444ToRgb_SSE41, VP8YuvToRgb32_SSE41, WebPYuv444ToRgb_C, 3); |
||||
YUV444_FUNC(Yuv444ToBgr_SSE41, VP8YuvToBgr32_SSE41, WebPYuv444ToBgr_C, 3); |
||||
#endif // WEBP_REDUCE_CSP
|
||||
|
||||
WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersSSE41(void) { |
||||
#if !defined(WEBP_REDUCE_CSP) |
||||
WebPYUV444Converters[MODE_RGB] = Yuv444ToRgb_SSE41; |
||||
WebPYUV444Converters[MODE_BGR] = Yuv444ToBgr_SSE41; |
||||
#endif // WEBP_REDUCE_CSP
|
||||
} |
||||
|
||||
#else |
||||
|
||||
WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersSSE41) |
||||
|
||||
#endif // WEBP_USE_SSE41
|
||||
|
||||
#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_SSE41)) |
||||
WEBP_DSP_INIT_STUB(WebPInitUpsamplersSSE41) |
||||
#endif |
||||
@ -0,0 +1,613 @@ |
||||
// Copyright 2014 Google Inc. All Rights Reserved.
|
||||
//
|
||||
// Use of this source code is governed by a BSD-style license
|
||||
// that can be found in the COPYING file in the root of the source
|
||||
// tree. An additional intellectual property rights grant can be found
|
||||
// in the file PATENTS. All contributing project authors may
|
||||
// be found in the AUTHORS file in the root of the source tree.
|
||||
// -----------------------------------------------------------------------------
|
||||
//
|
||||
// YUV->RGB conversion functions
|
||||
//
|
||||
// Author: Skal (pascal.massimino@gmail.com)
|
||||
|
||||
#include "src/dsp/yuv.h" |
||||
|
||||
#if defined(WEBP_USE_SSE41) |
||||
|
||||
#include "src/dsp/common_sse41.h" |
||||
#include <stdlib.h> |
||||
#include <smmintrin.h> |
||||
|
||||
//-----------------------------------------------------------------------------
|
||||
// Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
|
||||
|
||||
// These constants are 14b fixed-point version of ITU-R BT.601 constants.
|
||||
// R = (19077 * y + 26149 * v - 14234) >> 6
|
||||
// G = (19077 * y - 6419 * u - 13320 * v + 8708) >> 6
|
||||
// B = (19077 * y + 33050 * u - 17685) >> 6
|
||||
static void ConvertYUV444ToRGB_SSE41(const __m128i* const Y0, |
||||
const __m128i* const U0, |
||||
const __m128i* const V0, |
||||
__m128i* const R, |
||||
__m128i* const G, |
||||
__m128i* const B) { |
||||
const __m128i k19077 = _mm_set1_epi16(19077); |
||||
const __m128i k26149 = _mm_set1_epi16(26149); |
||||
const __m128i k14234 = _mm_set1_epi16(14234); |
||||
// 33050 doesn't fit in a signed short: only use this with unsigned arithmetic
|
||||
const __m128i k33050 = _mm_set1_epi16((short)33050); |
||||
const __m128i k17685 = _mm_set1_epi16(17685); |
||||
const __m128i k6419 = _mm_set1_epi16(6419); |
||||
const __m128i k13320 = _mm_set1_epi16(13320); |
||||
const __m128i k8708 = _mm_set1_epi16(8708); |
||||
|
||||
const __m128i Y1 = _mm_mulhi_epu16(*Y0, k19077); |
||||
|
||||
const __m128i R0 = _mm_mulhi_epu16(*V0, k26149); |
||||
const __m128i R1 = _mm_sub_epi16(Y1, k14234); |
||||
const __m128i R2 = _mm_add_epi16(R1, R0); |
||||
|
||||
const __m128i G0 = _mm_mulhi_epu16(*U0, k6419); |
||||
const __m128i G1 = _mm_mulhi_epu16(*V0, k13320); |
||||
const __m128i G2 = _mm_add_epi16(Y1, k8708); |
||||
const __m128i G3 = _mm_add_epi16(G0, G1); |
||||
const __m128i G4 = _mm_sub_epi16(G2, G3); |
||||
|
||||
// be careful with the saturated *unsigned* arithmetic here!
|
||||
const __m128i B0 = _mm_mulhi_epu16(*U0, k33050); |
||||
const __m128i B1 = _mm_adds_epu16(B0, Y1); |
||||
const __m128i B2 = _mm_subs_epu16(B1, k17685); |
||||
|
||||
// use logical shift for B2, which can be larger than 32767
|
||||
*R = _mm_srai_epi16(R2, 6); // range: [-14234, 30815]
|
||||
*G = _mm_srai_epi16(G4, 6); // range: [-10953, 27710]
|
||||
*B = _mm_srli_epi16(B2, 6); // range: [0, 34238]
|
||||
} |
||||
|
||||
// Load the bytes into the *upper* part of 16b words. That's "<< 8", basically.
|
||||
static WEBP_INLINE __m128i Load_HI_16_SSE41(const uint8_t* src) { |
||||
const __m128i zero = _mm_setzero_si128(); |
||||
return _mm_unpacklo_epi8(zero, _mm_loadl_epi64((const __m128i*)src)); |
||||
} |
||||
|
||||
// Load and replicate the U/V samples
|
||||
static WEBP_INLINE __m128i Load_UV_HI_8_SSE41(const uint8_t* src) { |
||||
const __m128i zero = _mm_setzero_si128(); |
||||
const __m128i tmp0 = _mm_cvtsi32_si128(*(const uint32_t*)src); |
||||
const __m128i tmp1 = _mm_unpacklo_epi8(zero, tmp0); |
||||
return _mm_unpacklo_epi16(tmp1, tmp1); // replicate samples
|
||||
} |
||||
|
||||
// Convert 32 samples of YUV444 to R/G/B
|
||||
static void YUV444ToRGB_SSE41(const uint8_t* const y, |
||||
const uint8_t* const u, |
||||
const uint8_t* const v, |
||||
__m128i* const R, __m128i* const G, |
||||
__m128i* const B) { |
||||
const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_HI_16_SSE41(u), |
||||
V0 = Load_HI_16_SSE41(v); |
||||
ConvertYUV444ToRGB_SSE41(&Y0, &U0, &V0, R, G, B); |
||||
} |
||||
|
||||
// Convert 32 samples of YUV420 to R/G/B
|
||||
static void YUV420ToRGB_SSE41(const uint8_t* const y, |
||||
const uint8_t* const u, |
||||
const uint8_t* const v, |
||||
__m128i* const R, __m128i* const G, |
||||
__m128i* const B) { |
||||
const __m128i Y0 = Load_HI_16_SSE41(y), U0 = Load_UV_HI_8_SSE41(u), |
||||
V0 = Load_UV_HI_8_SSE41(v); |
||||
ConvertYUV444ToRGB_SSE41(&Y0, &U0, &V0, R, G, B); |
||||
} |
||||
|
||||
// Pack the planar buffers
|
||||
// rrrr... rrrr... gggg... gggg... bbbb... bbbb....
|
||||
// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ...
|
||||
static WEBP_INLINE void PlanarTo24b_SSE41( |
||||
__m128i* const in0, __m128i* const in1, __m128i* const in2, |
||||
__m128i* const in3, __m128i* const in4, __m128i* const in5, |
||||
uint8_t* const rgb) { |
||||
// The input is 6 registers of sixteen 8b but for the sake of explanation,
|
||||
// let's take 6 registers of four 8b values.
|
||||
// To pack, we will keep taking one every two 8b integer and move it
|
||||
// around as follows:
|
||||
// Input:
|
||||
// r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7
|
||||
// Split the 6 registers in two sets of 3 registers: the first set as the even
|
||||
// 8b bytes, the second the odd ones:
|
||||
// r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7
|
||||
// Repeat the same permutations twice more:
|
||||
// r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7
|
||||
// r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7
|
||||
VP8PlanarTo24b_SSE41(in0, in1, in2, in3, in4, in5); |
||||
|
||||
_mm_storeu_si128((__m128i*)(rgb + 0), *in0); |
||||
_mm_storeu_si128((__m128i*)(rgb + 16), *in1); |
||||
_mm_storeu_si128((__m128i*)(rgb + 32), *in2); |
||||
_mm_storeu_si128((__m128i*)(rgb + 48), *in3); |
||||
_mm_storeu_si128((__m128i*)(rgb + 64), *in4); |
||||
_mm_storeu_si128((__m128i*)(rgb + 80), *in5); |
||||
} |
||||
|
||||
void VP8YuvToRgb32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v, |
||||
uint8_t* dst) { |
||||
__m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; |
||||
__m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5; |
||||
|
||||
YUV444ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0); |
||||
YUV444ToRGB_SSE41(y + 8, u + 8, v + 8, &R1, &G1, &B1); |
||||
YUV444ToRGB_SSE41(y + 16, u + 16, v + 16, &R2, &G2, &B2); |
||||
YUV444ToRGB_SSE41(y + 24, u + 24, v + 24, &R3, &G3, &B3); |
||||
|
||||
// Cast to 8b and store as RRRRGGGGBBBB.
|
||||
rgb0 = _mm_packus_epi16(R0, R1); |
||||
rgb1 = _mm_packus_epi16(R2, R3); |
||||
rgb2 = _mm_packus_epi16(G0, G1); |
||||
rgb3 = _mm_packus_epi16(G2, G3); |
||||
rgb4 = _mm_packus_epi16(B0, B1); |
||||
rgb5 = _mm_packus_epi16(B2, B3); |
||||
|
||||
// Pack as RGBRGBRGBRGB.
|
||||
PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst); |
||||
} |
||||
|
||||
void VP8YuvToBgr32_SSE41(const uint8_t* y, const uint8_t* u, const uint8_t* v, |
||||
uint8_t* dst) { |
||||
__m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; |
||||
__m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5; |
||||
|
||||
YUV444ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0); |
||||
YUV444ToRGB_SSE41(y + 8, u + 8, v + 8, &R1, &G1, &B1); |
||||
YUV444ToRGB_SSE41(y + 16, u + 16, v + 16, &R2, &G2, &B2); |
||||
YUV444ToRGB_SSE41(y + 24, u + 24, v + 24, &R3, &G3, &B3); |
||||
|
||||
// Cast to 8b and store as BBBBGGGGRRRR.
|
||||
bgr0 = _mm_packus_epi16(B0, B1); |
||||
bgr1 = _mm_packus_epi16(B2, B3); |
||||
bgr2 = _mm_packus_epi16(G0, G1); |
||||
bgr3 = _mm_packus_epi16(G2, G3); |
||||
bgr4 = _mm_packus_epi16(R0, R1); |
||||
bgr5= _mm_packus_epi16(R2, R3); |
||||
|
||||
// Pack as BGRBGRBGRBGR.
|
||||
PlanarTo24b_SSE41(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst); |
||||
} |
||||
|
||||
//-----------------------------------------------------------------------------
|
||||
// Arbitrary-length row conversion functions
|
||||
|
||||
static void YuvToRgbRow_SSE41(const uint8_t* y, |
||||
const uint8_t* u, const uint8_t* v, |
||||
uint8_t* dst, int len) { |
||||
int n; |
||||
for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) { |
||||
__m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; |
||||
__m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5; |
||||
|
||||
YUV420ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0); |
||||
YUV420ToRGB_SSE41(y + 8, u + 4, v + 4, &R1, &G1, &B1); |
||||
YUV420ToRGB_SSE41(y + 16, u + 8, v + 8, &R2, &G2, &B2); |
||||
YUV420ToRGB_SSE41(y + 24, u + 12, v + 12, &R3, &G3, &B3); |
||||
|
||||
// Cast to 8b and store as RRRRGGGGBBBB.
|
||||
rgb0 = _mm_packus_epi16(R0, R1); |
||||
rgb1 = _mm_packus_epi16(R2, R3); |
||||
rgb2 = _mm_packus_epi16(G0, G1); |
||||
rgb3 = _mm_packus_epi16(G2, G3); |
||||
rgb4 = _mm_packus_epi16(B0, B1); |
||||
rgb5 = _mm_packus_epi16(B2, B3); |
||||
|
||||
// Pack as RGBRGBRGBRGB.
|
||||
PlanarTo24b_SSE41(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst); |
||||
|
||||
y += 32; |
||||
u += 16; |
||||
v += 16; |
||||
} |
||||
for (; n < len; ++n) { // Finish off
|
||||
VP8YuvToRgb(y[0], u[0], v[0], dst); |
||||
dst += 3; |
||||
y += 1; |
||||
u += (n & 1); |
||||
v += (n & 1); |
||||
} |
||||
} |
||||
|
||||
static void YuvToBgrRow_SSE41(const uint8_t* y, |
||||
const uint8_t* u, const uint8_t* v, |
||||
uint8_t* dst, int len) { |
||||
int n; |
||||
for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) { |
||||
__m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; |
||||
__m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5; |
||||
|
||||
YUV420ToRGB_SSE41(y + 0, u + 0, v + 0, &R0, &G0, &B0); |
||||
YUV420ToRGB_SSE41(y + 8, u + 4, v + 4, &R1, &G1, &B1); |
||||
YUV420ToRGB_SSE41(y + 16, u + 8, v + 8, &R2, &G2, &B2); |
||||
YUV420ToRGB_SSE41(y + 24, u + 12, v + 12, &R3, &G3, &B3); |
||||
|
||||
// Cast to 8b and store as BBBBGGGGRRRR.
|
||||
bgr0 = _mm_packus_epi16(B0, B1); |
||||
bgr1 = _mm_packus_epi16(B2, B3); |
||||
bgr2 = _mm_packus_epi16(G0, G1); |
||||
bgr3 = _mm_packus_epi16(G2, G3); |
||||
bgr4 = _mm_packus_epi16(R0, R1); |
||||
bgr5 = _mm_packus_epi16(R2, R3); |
||||
|
||||
// Pack as BGRBGRBGRBGR.
|
||||
PlanarTo24b_SSE41(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst); |
||||
|
||||
y += 32; |
||||
u += 16; |
||||
v += 16; |
||||
} |
||||
for (; n < len; ++n) { // Finish off
|
||||
VP8YuvToBgr(y[0], u[0], v[0], dst); |
||||
dst += 3; |
||||
y += 1; |
||||
u += (n & 1); |
||||
v += (n & 1); |
||||
} |
||||
} |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Entry point
|
||||
|
||||
extern void WebPInitSamplersSSE41(void); |
||||
|
||||
WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersSSE41(void) { |
||||
WebPSamplers[MODE_RGB] = YuvToRgbRow_SSE41; |
||||
WebPSamplers[MODE_BGR] = YuvToBgrRow_SSE41; |
||||
} |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// RGB24/32 -> YUV converters
|
||||
|
||||
// Load eight 16b-words from *src.
|
||||
#define LOAD_16(src) _mm_loadu_si128((const __m128i*)(src)) |
||||
// Store either 16b-words into *dst
|
||||
#define STORE_16(V, dst) _mm_storeu_si128((__m128i*)(dst), (V)) |
||||
|
||||
#define WEBP_SSE41_SHUFF(OUT) do { \ |
||||
const __m128i tmp0 = _mm_shuffle_epi8(A0, shuff0); \
|
||||
const __m128i tmp1 = _mm_shuffle_epi8(A1, shuff1); \
|
||||
const __m128i tmp2 = _mm_shuffle_epi8(A2, shuff2); \
|
||||
const __m128i tmp3 = _mm_shuffle_epi8(A3, shuff0); \
|
||||
const __m128i tmp4 = _mm_shuffle_epi8(A4, shuff1); \
|
||||
const __m128i tmp5 = _mm_shuffle_epi8(A5, shuff2); \
|
||||
\
|
||||
/* OR everything to get one channel */ \
|
||||
const __m128i tmp6 = _mm_or_si128(tmp0, tmp1); \
|
||||
const __m128i tmp7 = _mm_or_si128(tmp3, tmp4); \
|
||||
out[OUT + 0] = _mm_or_si128(tmp6, tmp2); \
|
||||
out[OUT + 1] = _mm_or_si128(tmp7, tmp5); \
|
||||
} while (0); |
||||
|
||||
// Unpack the 8b input rgbrgbrgbrgb ... as contiguous registers:
|
||||
// rrrr... rrrr... gggg... gggg... bbbb... bbbb....
|
||||
// Similar to PlanarTo24bHelper(), but in reverse order.
|
||||
static WEBP_INLINE void RGB24PackedToPlanar_SSE41( |
||||
const uint8_t* const rgb, __m128i* const out /*out[6]*/) { |
||||
const __m128i A0 = _mm_loadu_si128((const __m128i*)(rgb + 0)); |
||||
const __m128i A1 = _mm_loadu_si128((const __m128i*)(rgb + 16)); |
||||
const __m128i A2 = _mm_loadu_si128((const __m128i*)(rgb + 32)); |
||||
const __m128i A3 = _mm_loadu_si128((const __m128i*)(rgb + 48)); |
||||
const __m128i A4 = _mm_loadu_si128((const __m128i*)(rgb + 64)); |
||||
const __m128i A5 = _mm_loadu_si128((const __m128i*)(rgb + 80)); |
||||
|
||||
// Compute RR.
|
||||
{ |
||||
const __m128i shuff0 = _mm_set_epi8( |
||||
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 15, 12, 9, 6, 3, 0); |
||||
const __m128i shuff1 = _mm_set_epi8( |
||||
-1, -1, -1, -1, -1, 14, 11, 8, 5, 2, -1, -1, -1, -1, -1, -1); |
||||
const __m128i shuff2 = _mm_set_epi8( |
||||
13, 10, 7, 4, 1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1); |
||||
WEBP_SSE41_SHUFF(0) |
||||
} |
||||
// Compute GG.
|
||||
{ |
||||
const __m128i shuff0 = _mm_set_epi8( |
||||
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 13, 10, 7, 4, 1); |
||||
const __m128i shuff1 = _mm_set_epi8( |
||||
-1, -1, -1, -1, -1, 15, 12, 9, 6, 3, 0, -1, -1, -1, -1, -1); |
||||
const __m128i shuff2 = _mm_set_epi8( |
||||
14, 11, 8, 5, 2, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1); |
||||
WEBP_SSE41_SHUFF(2) |
||||
} |
||||
// Compute BB.
|
||||
{ |
||||
const __m128i shuff0 = _mm_set_epi8( |
||||
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 14, 11, 8, 5, 2); |
||||
const __m128i shuff1 = _mm_set_epi8( |
||||
-1, -1, -1, -1, -1, -1, 13, 10, 7, 4, 1, -1, -1, -1, -1, -1); |
||||
const __m128i shuff2 = _mm_set_epi8( |
||||
15, 12, 9, 6, 3, 0, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1); |
||||
WEBP_SSE41_SHUFF(4) |
||||
} |
||||
} |
||||
|
||||
#undef WEBP_SSE41_SHUFF |
||||
|
||||
// Convert 8 packed ARGB to r[], g[], b[]
|
||||
static WEBP_INLINE void RGB32PackedToPlanar_SSE41( |
||||
const uint32_t* const argb, __m128i* const rgb /*in[6]*/) { |
||||
const __m128i zero = _mm_setzero_si128(); |
||||
__m128i a0 = LOAD_16(argb + 0); |
||||
__m128i a1 = LOAD_16(argb + 4); |
||||
__m128i a2 = LOAD_16(argb + 8); |
||||
__m128i a3 = LOAD_16(argb + 12); |
||||
VP8L32bToPlanar_SSE41(&a0, &a1, &a2, &a3); |
||||
rgb[0] = _mm_unpacklo_epi8(a1, zero); |
||||
rgb[1] = _mm_unpackhi_epi8(a1, zero); |
||||
rgb[2] = _mm_unpacklo_epi8(a2, zero); |
||||
rgb[3] = _mm_unpackhi_epi8(a2, zero); |
||||
rgb[4] = _mm_unpacklo_epi8(a3, zero); |
||||
rgb[5] = _mm_unpackhi_epi8(a3, zero); |
||||
} |
||||
|
||||
// This macro computes (RG * MULT_RG + GB * MULT_GB + ROUNDER) >> DESCALE_FIX
|
||||
// It's a macro and not a function because we need to use immediate values with
|
||||
// srai_epi32, e.g.
|
||||
#define TRANSFORM(RG_LO, RG_HI, GB_LO, GB_HI, MULT_RG, MULT_GB, \ |
||||
ROUNDER, DESCALE_FIX, OUT) do { \
|
||||
const __m128i V0_lo = _mm_madd_epi16(RG_LO, MULT_RG); \
|
||||
const __m128i V0_hi = _mm_madd_epi16(RG_HI, MULT_RG); \
|
||||
const __m128i V1_lo = _mm_madd_epi16(GB_LO, MULT_GB); \
|
||||
const __m128i V1_hi = _mm_madd_epi16(GB_HI, MULT_GB); \
|
||||
const __m128i V2_lo = _mm_add_epi32(V0_lo, V1_lo); \
|
||||
const __m128i V2_hi = _mm_add_epi32(V0_hi, V1_hi); \
|
||||
const __m128i V3_lo = _mm_add_epi32(V2_lo, ROUNDER); \
|
||||
const __m128i V3_hi = _mm_add_epi32(V2_hi, ROUNDER); \
|
||||
const __m128i V5_lo = _mm_srai_epi32(V3_lo, DESCALE_FIX); \
|
||||
const __m128i V5_hi = _mm_srai_epi32(V3_hi, DESCALE_FIX); \
|
||||
(OUT) = _mm_packs_epi32(V5_lo, V5_hi); \
|
||||
} while (0) |
||||
|
||||
#define MK_CST_16(A, B) _mm_set_epi16((B), (A), (B), (A), (B), (A), (B), (A)) |
||||
static WEBP_INLINE void ConvertRGBToY_SSE41(const __m128i* const R, |
||||
const __m128i* const G, |
||||
const __m128i* const B, |
||||
__m128i* const Y) { |
||||
const __m128i kRG_y = MK_CST_16(16839, 33059 - 16384); |
||||
const __m128i kGB_y = MK_CST_16(16384, 6420); |
||||
const __m128i kHALF_Y = _mm_set1_epi32((16 << YUV_FIX) + YUV_HALF); |
||||
|
||||
const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G); |
||||
const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G); |
||||
const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B); |
||||
const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B); |
||||
TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_y, kGB_y, kHALF_Y, YUV_FIX, *Y); |
||||
} |
||||
|
||||
static WEBP_INLINE void ConvertRGBToUV_SSE41(const __m128i* const R, |
||||
const __m128i* const G, |
||||
const __m128i* const B, |
||||
__m128i* const U, |
||||
__m128i* const V) { |
||||
const __m128i kRG_u = MK_CST_16(-9719, -19081); |
||||
const __m128i kGB_u = MK_CST_16(0, 28800); |
||||
const __m128i kRG_v = MK_CST_16(28800, 0); |
||||
const __m128i kGB_v = MK_CST_16(-24116, -4684); |
||||
const __m128i kHALF_UV = _mm_set1_epi32(((128 << YUV_FIX) + YUV_HALF) << 2); |
||||
|
||||
const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G); |
||||
const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G); |
||||
const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B); |
||||
const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B); |
||||
TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_u, kGB_u, |
||||
kHALF_UV, YUV_FIX + 2, *U); |
||||
TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_v, kGB_v, |
||||
kHALF_UV, YUV_FIX + 2, *V); |
||||
} |
||||
|
||||
#undef MK_CST_16 |
||||
#undef TRANSFORM |
||||
|
||||
static void ConvertRGB24ToY_SSE41(const uint8_t* rgb, uint8_t* y, int width) { |
||||
const int max_width = width & ~31; |
||||
int i; |
||||
for (i = 0; i < max_width; rgb += 3 * 16 * 2) { |
||||
__m128i rgb_plane[6]; |
||||
int j; |
||||
|
||||
RGB24PackedToPlanar_SSE41(rgb, rgb_plane); |
||||
|
||||
for (j = 0; j < 2; ++j, i += 16) { |
||||
const __m128i zero = _mm_setzero_si128(); |
||||
__m128i r, g, b, Y0, Y1; |
||||
|
||||
// Convert to 16-bit Y.
|
||||
r = _mm_unpacklo_epi8(rgb_plane[0 + j], zero); |
||||
g = _mm_unpacklo_epi8(rgb_plane[2 + j], zero); |
||||
b = _mm_unpacklo_epi8(rgb_plane[4 + j], zero); |
||||
ConvertRGBToY_SSE41(&r, &g, &b, &Y0); |
||||
|
||||
// Convert to 16-bit Y.
|
||||
r = _mm_unpackhi_epi8(rgb_plane[0 + j], zero); |
||||
g = _mm_unpackhi_epi8(rgb_plane[2 + j], zero); |
||||
b = _mm_unpackhi_epi8(rgb_plane[4 + j], zero); |
||||
ConvertRGBToY_SSE41(&r, &g, &b, &Y1); |
||||
|
||||
// Cast to 8-bit and store.
|
||||
STORE_16(_mm_packus_epi16(Y0, Y1), y + i); |
||||
} |
||||
} |
||||
for (; i < width; ++i, rgb += 3) { // left-over
|
||||
y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF); |
||||
} |
||||
} |
||||
|
||||
static void ConvertBGR24ToY_SSE41(const uint8_t* bgr, uint8_t* y, int width) { |
||||
const int max_width = width & ~31; |
||||
int i; |
||||
for (i = 0; i < max_width; bgr += 3 * 16 * 2) { |
||||
__m128i bgr_plane[6]; |
||||
int j; |
||||
|
||||
RGB24PackedToPlanar_SSE41(bgr, bgr_plane); |
||||
|
||||
for (j = 0; j < 2; ++j, i += 16) { |
||||
const __m128i zero = _mm_setzero_si128(); |
||||
__m128i r, g, b, Y0, Y1; |
||||
|
||||
// Convert to 16-bit Y.
|
||||
b = _mm_unpacklo_epi8(bgr_plane[0 + j], zero); |
||||
g = _mm_unpacklo_epi8(bgr_plane[2 + j], zero); |
||||
r = _mm_unpacklo_epi8(bgr_plane[4 + j], zero); |
||||
ConvertRGBToY_SSE41(&r, &g, &b, &Y0); |
||||
|
||||
// Convert to 16-bit Y.
|
||||
b = _mm_unpackhi_epi8(bgr_plane[0 + j], zero); |
||||
g = _mm_unpackhi_epi8(bgr_plane[2 + j], zero); |
||||
r = _mm_unpackhi_epi8(bgr_plane[4 + j], zero); |
||||
ConvertRGBToY_SSE41(&r, &g, &b, &Y1); |
||||
|
||||
// Cast to 8-bit and store.
|
||||
STORE_16(_mm_packus_epi16(Y0, Y1), y + i); |
||||
} |
||||
} |
||||
for (; i < width; ++i, bgr += 3) { // left-over
|
||||
y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF); |
||||
} |
||||
} |
||||
|
||||
static void ConvertARGBToY_SSE41(const uint32_t* argb, uint8_t* y, int width) { |
||||
const int max_width = width & ~15; |
||||
int i; |
||||
for (i = 0; i < max_width; i += 16) { |
||||
__m128i Y0, Y1, rgb[6]; |
||||
RGB32PackedToPlanar_SSE41(&argb[i], rgb); |
||||
ConvertRGBToY_SSE41(&rgb[0], &rgb[2], &rgb[4], &Y0); |
||||
ConvertRGBToY_SSE41(&rgb[1], &rgb[3], &rgb[5], &Y1); |
||||
STORE_16(_mm_packus_epi16(Y0, Y1), y + i); |
||||
} |
||||
for (; i < width; ++i) { // left-over
|
||||
const uint32_t p = argb[i]; |
||||
y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff, |
||||
YUV_HALF); |
||||
} |
||||
} |
||||
|
||||
// Horizontal add (doubled) of two 16b values, result is 16b.
|
||||
// in: A | B | C | D | ... -> out: 2*(A+B) | 2*(C+D) | ...
|
||||
static void HorizontalAddPack_SSE41(const __m128i* const A, |
||||
const __m128i* const B, |
||||
__m128i* const out) { |
||||
const __m128i k2 = _mm_set1_epi16(2); |
||||
const __m128i C = _mm_madd_epi16(*A, k2); |
||||
const __m128i D = _mm_madd_epi16(*B, k2); |
||||
*out = _mm_packs_epi32(C, D); |
||||
} |
||||
|
||||
static void ConvertARGBToUV_SSE41(const uint32_t* argb, |
||||
uint8_t* u, uint8_t* v, |
||||
int src_width, int do_store) { |
||||
const int max_width = src_width & ~31; |
||||
int i; |
||||
for (i = 0; i < max_width; i += 32, u += 16, v += 16) { |
||||
__m128i rgb[6], U0, V0, U1, V1; |
||||
RGB32PackedToPlanar_SSE41(&argb[i], rgb); |
||||
HorizontalAddPack_SSE41(&rgb[0], &rgb[1], &rgb[0]); |
||||
HorizontalAddPack_SSE41(&rgb[2], &rgb[3], &rgb[2]); |
||||
HorizontalAddPack_SSE41(&rgb[4], &rgb[5], &rgb[4]); |
||||
ConvertRGBToUV_SSE41(&rgb[0], &rgb[2], &rgb[4], &U0, &V0); |
||||
|
||||
RGB32PackedToPlanar_SSE41(&argb[i + 16], rgb); |
||||
HorizontalAddPack_SSE41(&rgb[0], &rgb[1], &rgb[0]); |
||||
HorizontalAddPack_SSE41(&rgb[2], &rgb[3], &rgb[2]); |
||||
HorizontalAddPack_SSE41(&rgb[4], &rgb[5], &rgb[4]); |
||||
ConvertRGBToUV_SSE41(&rgb[0], &rgb[2], &rgb[4], &U1, &V1); |
||||
|
||||
U0 = _mm_packus_epi16(U0, U1); |
||||
V0 = _mm_packus_epi16(V0, V1); |
||||
if (!do_store) { |
||||
const __m128i prev_u = LOAD_16(u); |
||||
const __m128i prev_v = LOAD_16(v); |
||||
U0 = _mm_avg_epu8(U0, prev_u); |
||||
V0 = _mm_avg_epu8(V0, prev_v); |
||||
} |
||||
STORE_16(U0, u); |
||||
STORE_16(V0, v); |
||||
} |
||||
if (i < src_width) { // left-over
|
||||
WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store); |
||||
} |
||||
} |
||||
|
||||
// Convert 16 packed ARGB 16b-values to r[], g[], b[]
|
||||
static WEBP_INLINE void RGBA32PackedToPlanar_16b_SSE41( |
||||
const uint16_t* const rgbx, |
||||
__m128i* const r, __m128i* const g, __m128i* const b) { |
||||
const __m128i in0 = LOAD_16(rgbx + 0); // r0 | g0 | b0 |x| r1 | g1 | b1 |x
|
||||
const __m128i in1 = LOAD_16(rgbx + 8); // r2 | g2 | b2 |x| r3 | g3 | b3 |x
|
||||
const __m128i in2 = LOAD_16(rgbx + 16); // r4 | ...
|
||||
const __m128i in3 = LOAD_16(rgbx + 24); // r6 | ...
|
||||
// aarrggbb as 16-bit.
|
||||
const __m128i shuff0 = |
||||
_mm_set_epi8(-1, -1, -1, -1, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0); |
||||
const __m128i shuff1 = |
||||
_mm_set_epi8(13, 12, 5, 4, -1, -1, -1, -1, 11, 10, 3, 2, 9, 8, 1, 0); |
||||
const __m128i A0 = _mm_shuffle_epi8(in0, shuff0); |
||||
const __m128i A1 = _mm_shuffle_epi8(in1, shuff1); |
||||
const __m128i A2 = _mm_shuffle_epi8(in2, shuff0); |
||||
const __m128i A3 = _mm_shuffle_epi8(in3, shuff1); |
||||
// R0R1G0G1
|
||||
// B0B1****
|
||||
// R2R3G2G3
|
||||
// B2B3****
|
||||
// (OR is used to free port 5 for the unpack)
|
||||
const __m128i B0 = _mm_unpacklo_epi32(A0, A1); |
||||
const __m128i B1 = _mm_or_si128(A0, A1); |
||||
const __m128i B2 = _mm_unpacklo_epi32(A2, A3); |
||||
const __m128i B3 = _mm_or_si128(A2, A3); |
||||
// Gather the channels.
|
||||
*r = _mm_unpacklo_epi64(B0, B2); |
||||
*g = _mm_unpackhi_epi64(B0, B2); |
||||
*b = _mm_unpackhi_epi64(B1, B3); |
||||
} |
||||
|
||||
static void ConvertRGBA32ToUV_SSE41(const uint16_t* rgb, |
||||
uint8_t* u, uint8_t* v, int width) { |
||||
const int max_width = width & ~15; |
||||
const uint16_t* const last_rgb = rgb + 4 * max_width; |
||||
while (rgb < last_rgb) { |
||||
__m128i r, g, b, U0, V0, U1, V1; |
||||
RGBA32PackedToPlanar_16b_SSE41(rgb + 0, &r, &g, &b); |
||||
ConvertRGBToUV_SSE41(&r, &g, &b, &U0, &V0); |
||||
RGBA32PackedToPlanar_16b_SSE41(rgb + 32, &r, &g, &b); |
||||
ConvertRGBToUV_SSE41(&r, &g, &b, &U1, &V1); |
||||
STORE_16(_mm_packus_epi16(U0, U1), u); |
||||
STORE_16(_mm_packus_epi16(V0, V1), v); |
||||
u += 16; |
||||
v += 16; |
||||
rgb += 2 * 32; |
||||
} |
||||
if (max_width < width) { // left-over
|
||||
WebPConvertRGBA32ToUV_C(rgb, u, v, width - max_width); |
||||
} |
||||
} |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
extern void WebPInitConvertARGBToYUVSSE41(void); |
||||
|
||||
WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVSSE41(void) { |
||||
WebPConvertARGBToY = ConvertARGBToY_SSE41; |
||||
WebPConvertARGBToUV = ConvertARGBToUV_SSE41; |
||||
|
||||
WebPConvertRGB24ToY = ConvertRGB24ToY_SSE41; |
||||
WebPConvertBGR24ToY = ConvertBGR24ToY_SSE41; |
||||
|
||||
WebPConvertRGBA32ToUV = ConvertRGBA32ToUV_SSE41; |
||||
} |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
#else // !WEBP_USE_SSE41
|
||||
|
||||
WEBP_DSP_INIT_STUB(WebPInitSamplersSSE41) |
||||
WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVSSE41) |
||||
|
||||
#endif // WEBP_USE_SSE41
|
||||
@ -1,455 +0,0 @@ |
||||
// 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.
|
||||
// -----------------------------------------------------------------------------
|
||||
//
|
||||
// Author: Mislav Bradac (mislavm@google.com)
|
||||
//
|
||||
|
||||
#include "src/enc/delta_palettization_enc.h" |
||||
|
||||
#ifdef WEBP_EXPERIMENTAL_FEATURES |
||||
#include "src/webp/types.h" |
||||
#include "src/dsp/lossless.h" |
||||
|
||||
#define MK_COL(r, g, b) (((r) << 16) + ((g) << 8) + (b)) |
||||
|
||||
// Format allows palette up to 256 entries, but more palette entries produce
|
||||
// bigger entropy. In the future it will probably be useful to add more entries
|
||||
// that are far from the origin of the palette or choose remaining entries
|
||||
// dynamically.
|
||||
#define DELTA_PALETTE_SIZE 226 |
||||
|
||||
// Palette used for delta_palettization. Entries are roughly sorted by distance
|
||||
// of their signed equivalents from the origin.
|
||||
static const uint32_t kDeltaPalette[DELTA_PALETTE_SIZE] = { |
||||
MK_COL(0u, 0u, 0u), |
||||
MK_COL(255u, 255u, 255u), |
||||
MK_COL(1u, 1u, 1u), |
||||
MK_COL(254u, 254u, 254u), |
||||
MK_COL(2u, 2u, 2u), |
||||
MK_COL(4u, 4u, 4u), |
||||
MK_COL(252u, 252u, 252u), |
||||
MK_COL(250u, 0u, 0u), |
||||
MK_COL(0u, 250u, 0u), |
||||
MK_COL(0u, 0u, 250u), |
||||
MK_COL(6u, 0u, 0u), |
||||
MK_COL(0u, 6u, 0u), |
||||
MK_COL(0u, 0u, 6u), |
||||
MK_COL(0u, 0u, 248u), |
||||
MK_COL(0u, 0u, 8u), |
||||
MK_COL(0u, 248u, 0u), |
||||
MK_COL(0u, 248u, 248u), |
||||
MK_COL(0u, 248u, 8u), |
||||
MK_COL(0u, 8u, 0u), |
||||
MK_COL(0u, 8u, 248u), |
||||
MK_COL(0u, 8u, 8u), |
||||
MK_COL(8u, 8u, 8u), |
||||
MK_COL(248u, 0u, 0u), |
||||
MK_COL(248u, 0u, 248u), |
||||
MK_COL(248u, 0u, 8u), |
||||
MK_COL(248u, 248u, 0u), |
||||
MK_COL(248u, 8u, 0u), |
||||
MK_COL(8u, 0u, 0u), |
||||
MK_COL(8u, 0u, 248u), |
||||
MK_COL(8u, 0u, 8u), |
||||
MK_COL(8u, 248u, 0u), |
||||
MK_COL(8u, 8u, 0u), |
||||
MK_COL(23u, 23u, 23u), |
||||
MK_COL(13u, 13u, 13u), |
||||
MK_COL(232u, 232u, 232u), |
||||
MK_COL(244u, 244u, 244u), |
||||
MK_COL(245u, 245u, 250u), |
||||
MK_COL(50u, 50u, 50u), |
||||
MK_COL(204u, 204u, 204u), |
||||
MK_COL(236u, 236u, 236u), |
||||
MK_COL(16u, 16u, 16u), |
||||
MK_COL(240u, 16u, 16u), |
||||
MK_COL(16u, 240u, 16u), |
||||
MK_COL(240u, 240u, 16u), |
||||
MK_COL(16u, 16u, 240u), |
||||
MK_COL(240u, 16u, 240u), |
||||
MK_COL(16u, 240u, 240u), |
||||
MK_COL(240u, 240u, 240u), |
||||
MK_COL(0u, 0u, 232u), |
||||
MK_COL(0u, 232u, 0u), |
||||
MK_COL(232u, 0u, 0u), |
||||
MK_COL(0u, 0u, 24u), |
||||
MK_COL(0u, 24u, 0u), |
||||
MK_COL(24u, 0u, 0u), |
||||
MK_COL(32u, 32u, 32u), |
||||
MK_COL(224u, 32u, 32u), |
||||
MK_COL(32u, 224u, 32u), |
||||
MK_COL(224u, 224u, 32u), |
||||
MK_COL(32u, 32u, 224u), |
||||
MK_COL(224u, 32u, 224u), |
||||
MK_COL(32u, 224u, 224u), |
||||
MK_COL(224u, 224u, 224u), |
||||
MK_COL(0u, 0u, 176u), |
||||
MK_COL(0u, 0u, 80u), |
||||
MK_COL(0u, 176u, 0u), |
||||
MK_COL(0u, 176u, 176u), |
||||
MK_COL(0u, 176u, 80u), |
||||
MK_COL(0u, 80u, 0u), |
||||
MK_COL(0u, 80u, 176u), |
||||
MK_COL(0u, 80u, 80u), |
||||
MK_COL(176u, 0u, 0u), |
||||
MK_COL(176u, 0u, 176u), |
||||
MK_COL(176u, 0u, 80u), |
||||
MK_COL(176u, 176u, 0u), |
||||
MK_COL(176u, 80u, 0u), |
||||
MK_COL(80u, 0u, 0u), |
||||
MK_COL(80u, 0u, 176u), |
||||
MK_COL(80u, 0u, 80u), |
||||
MK_COL(80u, 176u, 0u), |
||||
MK_COL(80u, 80u, 0u), |
||||
MK_COL(0u, 0u, 152u), |
||||
MK_COL(0u, 0u, 104u), |
||||
MK_COL(0u, 152u, 0u), |
||||
MK_COL(0u, 152u, 152u), |
||||
MK_COL(0u, 152u, 104u), |
||||
MK_COL(0u, 104u, 0u), |
||||
MK_COL(0u, 104u, 152u), |
||||
MK_COL(0u, 104u, 104u), |
||||
MK_COL(152u, 0u, 0u), |
||||
MK_COL(152u, 0u, 152u), |
||||
MK_COL(152u, 0u, 104u), |
||||
MK_COL(152u, 152u, 0u), |
||||
MK_COL(152u, 104u, 0u), |
||||
MK_COL(104u, 0u, 0u), |
||||
MK_COL(104u, 0u, 152u), |
||||
MK_COL(104u, 0u, 104u), |
||||
MK_COL(104u, 152u, 0u), |
||||
MK_COL(104u, 104u, 0u), |
||||
MK_COL(216u, 216u, 216u), |
||||
MK_COL(216u, 216u, 40u), |
||||
MK_COL(216u, 216u, 176u), |
||||
MK_COL(216u, 216u, 80u), |
||||
MK_COL(216u, 40u, 216u), |
||||
MK_COL(216u, 40u, 40u), |
||||
MK_COL(216u, 40u, 176u), |
||||
MK_COL(216u, 40u, 80u), |
||||
MK_COL(216u, 176u, 216u), |
||||
MK_COL(216u, 176u, 40u), |
||||
MK_COL(216u, 176u, 176u), |
||||
MK_COL(216u, 176u, 80u), |
||||
MK_COL(216u, 80u, 216u), |
||||
MK_COL(216u, 80u, 40u), |
||||
MK_COL(216u, 80u, 176u), |
||||
MK_COL(216u, 80u, 80u), |
||||
MK_COL(40u, 216u, 216u), |
||||
MK_COL(40u, 216u, 40u), |
||||
MK_COL(40u, 216u, 176u), |
||||
MK_COL(40u, 216u, 80u), |
||||
MK_COL(40u, 40u, 216u), |
||||
MK_COL(40u, 40u, 40u), |
||||
MK_COL(40u, 40u, 176u), |
||||
MK_COL(40u, 40u, 80u), |
||||
MK_COL(40u, 176u, 216u), |
||||
MK_COL(40u, 176u, 40u), |
||||
MK_COL(40u, 176u, 176u), |
||||
MK_COL(40u, 176u, 80u), |
||||
MK_COL(40u, 80u, 216u), |
||||
MK_COL(40u, 80u, 40u), |
||||
MK_COL(40u, 80u, 176u), |
||||
MK_COL(40u, 80u, 80u), |
||||
MK_COL(80u, 216u, 216u), |
||||
MK_COL(80u, 216u, 40u), |
||||
MK_COL(80u, 216u, 176u), |
||||
MK_COL(80u, 216u, 80u), |
||||
MK_COL(80u, 40u, 216u), |
||||
MK_COL(80u, 40u, 40u), |
||||
MK_COL(80u, 40u, 176u), |
||||
MK_COL(80u, 40u, 80u), |
||||
MK_COL(80u, 176u, 216u), |
||||
MK_COL(80u, 176u, 40u), |
||||
MK_COL(80u, 176u, 176u), |
||||
MK_COL(80u, 176u, 80u), |
||||
MK_COL(80u, 80u, 216u), |
||||
MK_COL(80u, 80u, 40u), |
||||
MK_COL(80u, 80u, 176u), |
||||
MK_COL(80u, 80u, 80u), |
||||
MK_COL(0u, 0u, 192u), |
||||
MK_COL(0u, 0u, 64u), |
||||
MK_COL(0u, 0u, 128u), |
||||
MK_COL(0u, 192u, 0u), |
||||
MK_COL(0u, 192u, 192u), |
||||
MK_COL(0u, 192u, 64u), |
||||
MK_COL(0u, 192u, 128u), |
||||
MK_COL(0u, 64u, 0u), |
||||
MK_COL(0u, 64u, 192u), |
||||
MK_COL(0u, 64u, 64u), |
||||
MK_COL(0u, 64u, 128u), |
||||
MK_COL(0u, 128u, 0u), |
||||
MK_COL(0u, 128u, 192u), |
||||
MK_COL(0u, 128u, 64u), |
||||
MK_COL(0u, 128u, 128u), |
||||
MK_COL(176u, 216u, 216u), |
||||
MK_COL(176u, 216u, 40u), |
||||
MK_COL(176u, 216u, 176u), |
||||
MK_COL(176u, 216u, 80u), |
||||
MK_COL(176u, 40u, 216u), |
||||
MK_COL(176u, 40u, 40u), |
||||
MK_COL(176u, 40u, 176u), |
||||
MK_COL(176u, 40u, 80u), |
||||
MK_COL(176u, 176u, 216u), |
||||
MK_COL(176u, 176u, 40u), |
||||
MK_COL(176u, 176u, 176u), |
||||
MK_COL(176u, 176u, 80u), |
||||
MK_COL(176u, 80u, 216u), |
||||
MK_COL(176u, 80u, 40u), |
||||
MK_COL(176u, 80u, 176u), |
||||
MK_COL(176u, 80u, 80u), |
||||
MK_COL(192u, 0u, 0u), |
||||
MK_COL(192u, 0u, 192u), |
||||
MK_COL(192u, 0u, 64u), |
||||
MK_COL(192u, 0u, 128u), |
||||
MK_COL(192u, 192u, 0u), |
||||
MK_COL(192u, 192u, 192u), |
||||
MK_COL(192u, 192u, 64u), |
||||
MK_COL(192u, 192u, 128u), |
||||
MK_COL(192u, 64u, 0u), |
||||
MK_COL(192u, 64u, 192u), |
||||
MK_COL(192u, 64u, 64u), |
||||
MK_COL(192u, 64u, 128u), |
||||
MK_COL(192u, 128u, 0u), |
||||
MK_COL(192u, 128u, 192u), |
||||
MK_COL(192u, 128u, 64u), |
||||
MK_COL(192u, 128u, 128u), |
||||
MK_COL(64u, 0u, 0u), |
||||
MK_COL(64u, 0u, 192u), |
||||
MK_COL(64u, 0u, 64u), |
||||
MK_COL(64u, 0u, 128u), |
||||
MK_COL(64u, 192u, 0u), |
||||
MK_COL(64u, 192u, 192u), |
||||
MK_COL(64u, 192u, 64u), |
||||
MK_COL(64u, 192u, 128u), |
||||
MK_COL(64u, 64u, 0u), |
||||
MK_COL(64u, 64u, 192u), |
||||
MK_COL(64u, 64u, 64u), |
||||
MK_COL(64u, 64u, 128u), |
||||
MK_COL(64u, 128u, 0u), |
||||
MK_COL(64u, 128u, 192u), |
||||
MK_COL(64u, 128u, 64u), |
||||
MK_COL(64u, 128u, 128u), |
||||
MK_COL(128u, 0u, 0u), |
||||
MK_COL(128u, 0u, 192u), |
||||
MK_COL(128u, 0u, 64u), |
||||
MK_COL(128u, 0u, 128u), |
||||
MK_COL(128u, 192u, 0u), |
||||
MK_COL(128u, 192u, 192u), |
||||
MK_COL(128u, 192u, 64u), |
||||
MK_COL(128u, 192u, 128u), |
||||
MK_COL(128u, 64u, 0u), |
||||
MK_COL(128u, 64u, 192u), |
||||
MK_COL(128u, 64u, 64u), |
||||
MK_COL(128u, 64u, 128u), |
||||
MK_COL(128u, 128u, 0u), |
||||
MK_COL(128u, 128u, 192u), |
||||
MK_COL(128u, 128u, 64u), |
||||
MK_COL(128u, 128u, 128u), |
||||
}; |
||||
|
||||
#undef MK_COL |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// TODO(skal): move the functions to dsp/lossless.c when the correct
|
||||
// granularity is found. For now, we'll just copy-paste some useful bits
|
||||
// here instead.
|
||||
|
||||
// In-place sum of each component with mod 256.
|
||||
static WEBP_INLINE void AddPixelsEq(uint32_t* a, uint32_t b) { |
||||
const uint32_t alpha_and_green = (*a & 0xff00ff00u) + (b & 0xff00ff00u); |
||||
const uint32_t red_and_blue = (*a & 0x00ff00ffu) + (b & 0x00ff00ffu); |
||||
*a = (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu); |
||||
} |
||||
|
||||
static WEBP_INLINE uint32_t Clip255(uint32_t a) { |
||||
if (a < 256) { |
||||
return a; |
||||
} |
||||
// return 0, when a is a negative integer.
|
||||
// return 255, when a is positive.
|
||||
return ~a >> 24; |
||||
} |
||||
|
||||
// Delta palettization functions.
|
||||
static WEBP_INLINE int Square(int x) { |
||||
return x * x; |
||||
} |
||||
|
||||
static WEBP_INLINE uint32_t Intensity(uint32_t a) { |
||||
return |
||||
30 * ((a >> 16) & 0xff) + |
||||
59 * ((a >> 8) & 0xff) + |
||||
11 * ((a >> 0) & 0xff); |
||||
} |
||||
|
||||
static uint32_t CalcDist(uint32_t predicted_value, uint32_t actual_value, |
||||
uint32_t palette_entry) { |
||||
int i; |
||||
uint32_t distance = 0; |
||||
AddPixelsEq(&predicted_value, palette_entry); |
||||
for (i = 0; i < 32; i += 8) { |
||||
const int32_t av = (actual_value >> i) & 0xff; |
||||
const int32_t pv = (predicted_value >> i) & 0xff; |
||||
distance += Square(pv - av); |
||||
} |
||||
// We sum square of intensity difference with factor 10, but because Intensity
|
||||
// returns 100 times real intensity we need to multiply differences of colors
|
||||
// by 1000.
|
||||
distance *= 1000u; |
||||
distance += Square(Intensity(predicted_value) |
||||
- Intensity(actual_value)); |
||||
return distance; |
||||
} |
||||
|
||||
static uint32_t Predict(int x, int y, uint32_t* image) { |
||||
const uint32_t t = (y == 0) ? ARGB_BLACK : image[x]; |
||||
const uint32_t l = (x == 0) ? ARGB_BLACK : image[x - 1]; |
||||
const uint32_t p = |
||||
(((((t >> 24) & 0xff) + ((l >> 24) & 0xff)) / 2) << 24) + |
||||
(((((t >> 16) & 0xff) + ((l >> 16) & 0xff)) / 2) << 16) + |
||||
(((((t >> 8) & 0xff) + ((l >> 8) & 0xff)) / 2) << 8) + |
||||
(((((t >> 0) & 0xff) + ((l >> 0) & 0xff)) / 2) << 0); |
||||
if (x == 0 && y == 0) return ARGB_BLACK; |
||||
if (x == 0) return t; |
||||
if (y == 0) return l; |
||||
return p; |
||||
} |
||||
|
||||
static WEBP_INLINE int AddSubtractComponentFullWithCoefficient( |
||||
int a, int b, int c) { |
||||
return Clip255(a + ((b - c) >> 2)); |
||||
} |
||||
|
||||
static WEBP_INLINE uint32_t ClampedAddSubtractFullWithCoefficient( |
||||
uint32_t c0, uint32_t c1, uint32_t c2) { |
||||
const int a = AddSubtractComponentFullWithCoefficient( |
||||
c0 >> 24, c1 >> 24, c2 >> 24); |
||||
const int r = AddSubtractComponentFullWithCoefficient((c0 >> 16) & 0xff, |
||||
(c1 >> 16) & 0xff, |
||||
(c2 >> 16) & 0xff); |
||||
const int g = AddSubtractComponentFullWithCoefficient((c0 >> 8) & 0xff, |
||||
(c1 >> 8) & 0xff, |
||||
(c2 >> 8) & 0xff); |
||||
const int b = AddSubtractComponentFullWithCoefficient( |
||||
c0 & 0xff, c1 & 0xff, c2 & 0xff); |
||||
return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b; |
||||
} |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
// Find palette entry with minimum error from difference of actual pixel value
|
||||
// and predicted pixel value. Propagate error of pixel to its top and left pixel
|
||||
// in src array. Write predicted_value + palette_entry to new_image. Return
|
||||
// index of best palette entry.
|
||||
static int FindBestPaletteEntry(uint32_t src, uint32_t predicted_value, |
||||
const uint32_t palette[], int palette_size) { |
||||
int i; |
||||
int idx = 0; |
||||
uint32_t best_distance = CalcDist(predicted_value, src, palette[0]); |
||||
for (i = 1; i < palette_size; ++i) { |
||||
const uint32_t distance = CalcDist(predicted_value, src, palette[i]); |
||||
if (distance < best_distance) { |
||||
best_distance = distance; |
||||
idx = i; |
||||
} |
||||
} |
||||
return idx; |
||||
} |
||||
|
||||
static void ApplyBestPaletteEntry(int x, int y, |
||||
uint32_t new_value, uint32_t palette_value, |
||||
uint32_t* src, int src_stride, |
||||
uint32_t* new_image) { |
||||
AddPixelsEq(&new_value, palette_value); |
||||
if (x > 0) { |
||||
src[x - 1] = ClampedAddSubtractFullWithCoefficient(src[x - 1], |
||||
new_value, src[x]); |
||||
} |
||||
if (y > 0) { |
||||
src[x - src_stride] = |
||||
ClampedAddSubtractFullWithCoefficient(src[x - src_stride], |
||||
new_value, src[x]); |
||||
} |
||||
new_image[x] = new_value; |
||||
} |
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Main entry point
|
||||
|
||||
static WebPEncodingError ApplyDeltaPalette(uint32_t* src, uint32_t* dst, |
||||
uint32_t src_stride, |
||||
uint32_t dst_stride, |
||||
const uint32_t* palette, |
||||
int palette_size, |
||||
int width, int height, |
||||
int num_passes) { |
||||
int x, y; |
||||
WebPEncodingError err = VP8_ENC_OK; |
||||
uint32_t* new_image = (uint32_t*)WebPSafeMalloc(width, sizeof(*new_image)); |
||||
uint8_t* const tmp_row = (uint8_t*)WebPSafeMalloc(width, sizeof(*tmp_row)); |
||||
if (new_image == NULL || tmp_row == NULL) { |
||||
err = VP8_ENC_ERROR_OUT_OF_MEMORY; |
||||
goto Error; |
||||
} |
||||
|
||||
while (num_passes--) { |
||||
uint32_t* cur_src = src; |
||||
uint32_t* cur_dst = dst; |
||||
for (y = 0; y < height; ++y) { |
||||
for (x = 0; x < width; ++x) { |
||||
const uint32_t predicted_value = Predict(x, y, new_image); |
||||
tmp_row[x] = FindBestPaletteEntry(cur_src[x], predicted_value, |
||||
palette, palette_size); |
||||
ApplyBestPaletteEntry(x, y, predicted_value, palette[tmp_row[x]], |
||||
cur_src, src_stride, new_image); |
||||
} |
||||
for (x = 0; x < width; ++x) { |
||||
cur_dst[x] = palette[tmp_row[x]]; |
||||
} |
||||
cur_src += src_stride; |
||||
cur_dst += dst_stride; |
||||
} |
||||
} |
||||
Error: |
||||
WebPSafeFree(new_image); |
||||
WebPSafeFree(tmp_row); |
||||
return err; |
||||
} |
||||
|
||||
// replaces enc->argb_ by a palettizable approximation of it,
|
||||
// and generates optimal enc->palette_[]
|
||||
WebPEncodingError WebPSearchOptimalDeltaPalette(VP8LEncoder* const enc) { |
||||
const WebPPicture* const pic = enc->pic_; |
||||
uint32_t* src = pic->argb; |
||||
uint32_t* dst = enc->argb_; |
||||
const int width = pic->width; |
||||
const int height = pic->height; |
||||
|
||||
WebPEncodingError err = VP8_ENC_OK; |
||||
memcpy(enc->palette_, kDeltaPalette, sizeof(kDeltaPalette)); |
||||
enc->palette_[DELTA_PALETTE_SIZE - 1] = src[0] - 0xff000000u; |
||||
enc->palette_size_ = DELTA_PALETTE_SIZE; |
||||
err = ApplyDeltaPalette(src, dst, pic->argb_stride, enc->current_width_, |
||||
enc->palette_, enc->palette_size_, |
||||
width, height, 2); |
||||
if (err != VP8_ENC_OK) goto Error; |
||||
|
||||
Error: |
||||
return err; |
||||
} |
||||
|
||||
#else // !WEBP_EXPERIMENTAL_FEATURES
|
||||
|
||||
WebPEncodingError WebPSearchOptimalDeltaPalette(VP8LEncoder* const enc) { |
||||
(void)enc; |
||||
return VP8_ENC_ERROR_INVALID_CONFIGURATION; |
||||
} |
||||
|
||||
#endif // WEBP_EXPERIMENTAL_FEATURES
|
||||
@ -1,25 +0,0 @@ |
||||
// 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.
|
||||
// -----------------------------------------------------------------------------
|
||||
//
|
||||
// Author: Mislav Bradac (mislavm@google.com)
|
||||
//
|
||||
|
||||
#ifndef WEBP_ENC_DELTA_PALETTIZATION_ENC_H_ |
||||
#define WEBP_ENC_DELTA_PALETTIZATION_ENC_H_ |
||||
|
||||
#include "src/webp/encode.h" |
||||
#include "src/enc/vp8li_enc.h" |
||||
|
||||
// Replaces enc->argb_[] input by a palettizable approximation of it,
|
||||
// and generates optimal enc->palette_[].
|
||||
// This function can revert enc->use_palette_ / enc->use_predict_ flag
|
||||
// if delta-palettization is not producing expected saving.
|
||||
WebPEncodingError WebPSearchOptimalDeltaPalette(VP8LEncoder* const enc); |
||||
|
||||
#endif // WEBP_ENC_DELTA_PALETTIZATION_ENC_H_
|
||||
Loading…
Reference in new issue