blob: 5500d7f5a64e0008dcf7c61e70d345d6779c6502 [file] [log] [blame]
/*
* Copyright 2011 The LibYuv Project Authors. All rights reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "libyuv/row.h"
// This module is for Visual C 32/64 bit and clangcl 32 bit
#if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
(defined(_M_IX86) || (defined(_M_X64) && !defined(__clang__)))
#if defined(_M_X64)
#include <emmintrin.h>
#include <tmmintrin.h> // For _mm_maddubs_epi16
#endif
#ifdef __cplusplus
namespace libyuv {
extern "C" {
#endif
// 64 bit
#if defined(_M_X64)
// Read 4 UV from 422, upsample to 8 UV.
#define READYUV422 \
xmm0 = _mm_cvtsi32_si128(*(uint32_t*)u_buf); \
xmm1 = _mm_cvtsi32_si128(*(uint32_t*)(u_buf + offset)); \
xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); \
xmm0 = _mm_unpacklo_epi16(xmm0, xmm0); \
u_buf += 4; \
xmm4 = _mm_loadl_epi64((__m128i*)y_buf); \
xmm4 = _mm_unpacklo_epi8(xmm4, xmm4); \
y_buf += 8;
// Read 4 UV from 422, upsample to 8 UV. With 8 Alpha.
#define READYUVA422 \
xmm0 = _mm_cvtsi32_si128(*(uint32_t*)u_buf); \
xmm1 = _mm_cvtsi32_si128(*(uint32_t*)(u_buf + offset)); \
xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); \
xmm0 = _mm_unpacklo_epi16(xmm0, xmm0); \
u_buf += 4; \
xmm4 = _mm_loadl_epi64((__m128i*)y_buf); \
xmm4 = _mm_unpacklo_epi8(xmm4, xmm4); \
y_buf += 8; \
xmm5 = _mm_loadl_epi64((__m128i*)a_buf); \
a_buf += 8;
// Convert 8 pixels: 8 UV and 8 Y.
#define YUVTORGB(yuvconstants) \
xmm1 = _mm_loadu_si128(&xmm0); \
xmm2 = _mm_loadu_si128(&xmm0); \
xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)yuvconstants->kUVToB); \
xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)yuvconstants->kUVToG); \
xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)yuvconstants->kUVToR); \
xmm0 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasB, xmm0); \
xmm1 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasG, xmm1); \
xmm2 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasR, xmm2); \
xmm4 = _mm_mulhi_epu16(xmm4, *(__m128i*)yuvconstants->kYToRgb); \
xmm0 = _mm_adds_epi16(xmm0, xmm4); \
xmm1 = _mm_adds_epi16(xmm1, xmm4); \
xmm2 = _mm_adds_epi16(xmm2, xmm4); \
xmm0 = _mm_srai_epi16(xmm0, 6); \
xmm1 = _mm_srai_epi16(xmm1, 6); \
xmm2 = _mm_srai_epi16(xmm2, 6); \
xmm0 = _mm_packus_epi16(xmm0, xmm0); \
xmm1 = _mm_packus_epi16(xmm1, xmm1); \
xmm2 = _mm_packus_epi16(xmm2, xmm2);
// Store 8 ARGB values.
#define STOREARGB \
xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); \
xmm2 = _mm_unpacklo_epi8(xmm2, xmm5); \
xmm1 = _mm_loadu_si128(&xmm0); \
xmm0 = _mm_unpacklo_epi16(xmm0, xmm2); \
xmm1 = _mm_unpackhi_epi16(xmm1, xmm2); \
_mm_storeu_si128((__m128i*)dst_argb, xmm0); \
_mm_storeu_si128((__m128i*)(dst_argb + 16), xmm1); \
dst_argb += 32;
#if defined(HAS_I422TOARGBROW_SSSE3)
void I422ToARGBRow_SSSE3(const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__m128i xmm0, xmm1, xmm2, xmm4;
const __m128i xmm5 = _mm_set1_epi8(-1);
const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
while (width > 0) {
READYUV422
YUVTORGB(yuvconstants)
STOREARGB
width -= 8;
}
}
#endif
#if defined(HAS_I422ALPHATOARGBROW_SSSE3)
void I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
const uint8_t* a_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__m128i xmm0, xmm1, xmm2, xmm4, xmm5;
const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
while (width > 0) {
READYUVA422
YUVTORGB(yuvconstants)
STOREARGB
width -= 8;
}
}
#endif
// 32 bit
#else // defined(_M_X64)
#ifdef HAS_ARGBTOYROW_SSSE3
// Constants for ARGB.
static const vec8 kARGBToY = {13, 65, 33, 0, 13, 65, 33, 0,
13, 65, 33, 0, 13, 65, 33, 0};
// JPeg full range.
static const vec8 kARGBToYJ = {15, 75, 38, 0, 15, 75, 38, 0,
15, 75, 38, 0, 15, 75, 38, 0};
static const vec8 kARGBToU = {112, -74, -38, 0, 112, -74, -38, 0,
112, -74, -38, 0, 112, -74, -38, 0};
static const vec8 kARGBToUJ = {127, -84, -43, 0, 127, -84, -43, 0,
127, -84, -43, 0, 127, -84, -43, 0};
static const vec8 kARGBToV = {
-18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
};
static const vec8 kARGBToVJ = {-20, -107, 127, 0, -20, -107, 127, 0,
-20, -107, 127, 0, -20, -107, 127, 0};
// vpshufb for vphaddw + vpackuswb packed to shorts.
static const lvec8 kShufARGBToUV_AVX = {
0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
// Constants for BGRA.
static const vec8 kBGRAToY = {0, 33, 65, 13, 0, 33, 65, 13,
0, 33, 65, 13, 0, 33, 65, 13};
static const vec8 kBGRAToU = {0, -38, -74, 112, 0, -38, -74, 112,
0, -38, -74, 112, 0, -38, -74, 112};
static const vec8 kBGRAToV = {0, 112, -94, -18, 0, 112, -94, -18,
0, 112, -94, -18, 0, 112, -94, -18};
// Constants for ABGR.
static const vec8 kABGRToY = {33, 65, 13, 0, 33, 65, 13, 0,
33, 65, 13, 0, 33, 65, 13, 0};
static const vec8 kABGRToU = {-38, -74, 112, 0, -38, -74, 112, 0,
-38, -74, 112, 0, -38, -74, 112, 0};
static const vec8 kABGRToV = {112, -94, -18, 0, 112, -94, -18, 0,
112, -94, -18, 0, 112, -94, -18, 0};
// Constants for RGBA.
static const vec8 kRGBAToY = {0, 13, 65, 33, 0, 13, 65, 33,
0, 13, 65, 33, 0, 13, 65, 33};
static const vec8 kRGBAToU = {0, 112, -74, -38, 0, 112, -74, -38,
0, 112, -74, -38, 0, 112, -74, -38};
static const vec8 kRGBAToV = {0, -18, -94, 112, 0, -18, -94, 112,
0, -18, -94, 112, 0, -18, -94, 112};
static const uvec8 kAddY16 = {16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u,
16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u};
// 7 bit fixed point 0.5.
static const vec16 kAddYJ64 = {64, 64, 64, 64, 64, 64, 64, 64};
static const uvec8 kAddUV128 = {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
static const uvec16 kAddUVJ128 = {0x8080u, 0x8080u, 0x8080u, 0x8080u,
0x8080u, 0x8080u, 0x8080u, 0x8080u};
// Shuffle table for converting RGB24 to ARGB.
static const uvec8 kShuffleMaskRGB24ToARGB = {
0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u};
// Shuffle table for converting RAW to ARGB.
static const uvec8 kShuffleMaskRAWToARGB = {2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u,
8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u};
// Shuffle table for converting RAW to RGB24. First 8.
static const uvec8 kShuffleMaskRAWToRGB24_0 = {
2u, 1u, 0u, 5u, 4u, 3u, 8u, 7u,
128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
// Shuffle table for converting RAW to RGB24. Middle 8.
static const uvec8 kShuffleMaskRAWToRGB24_1 = {
2u, 7u, 6u, 5u, 10u, 9u, 8u, 13u,
128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
// Shuffle table for converting RAW to RGB24. Last 8.
static const uvec8 kShuffleMaskRAWToRGB24_2 = {
8u, 7u, 12u, 11u, 10u, 15u, 14u, 13u,
128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
// Shuffle table for converting ARGB to RGB24.
static const uvec8 kShuffleMaskARGBToRGB24 = {
0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u};
// Shuffle table for converting ARGB to RAW.
static const uvec8 kShuffleMaskARGBToRAW = {
2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u};
// Shuffle table for converting ARGBToRGB24 for I422ToRGB24. First 8 + next 4
static const uvec8 kShuffleMaskARGBToRGB24_0 = {
0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u};
// YUY2 shuf 16 Y to 32 Y.
static const lvec8 kShuffleYUY2Y = {0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10,
10, 12, 12, 14, 14, 0, 0, 2, 2, 4, 4,
6, 6, 8, 8, 10, 10, 12, 12, 14, 14};
// YUY2 shuf 8 UV to 16 UV.
static const lvec8 kShuffleYUY2UV = {1, 3, 1, 3, 5, 7, 5, 7, 9, 11, 9,
11, 13, 15, 13, 15, 1, 3, 1, 3, 5, 7,
5, 7, 9, 11, 9, 11, 13, 15, 13, 15};
// UYVY shuf 16 Y to 32 Y.
static const lvec8 kShuffleUYVYY = {1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11,
11, 13, 13, 15, 15, 1, 1, 3, 3, 5, 5,
7, 7, 9, 9, 11, 11, 13, 13, 15, 15};
// UYVY shuf 8 UV to 16 UV.
static const lvec8 kShuffleUYVYUV = {0, 2, 0, 2, 4, 6, 4, 6, 8, 10, 8,
10, 12, 14, 12, 14, 0, 2, 0, 2, 4, 6,
4, 6, 8, 10, 8, 10, 12, 14, 12, 14};
// NV21 shuf 8 VU to 16 UV.
static const lvec8 kShuffleNV21 = {
1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6,
};
// Duplicates gray value 3 times and fills in alpha opaque.
__declspec(naked) void J400ToARGBRow_SSE2(const uint8_t* src_y,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, [esp + 4] // src_y
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
pcmpeqb xmm5, xmm5 // generate mask 0xff000000
pslld xmm5, 24
convertloop:
movq xmm0, qword ptr [eax]
lea eax, [eax + 8]
punpcklbw xmm0, xmm0
movdqa xmm1, xmm0
punpcklwd xmm0, xmm0
punpckhwd xmm1, xmm1
por xmm0, xmm5
por xmm1, xmm5
movdqu [edx], xmm0
movdqu [edx + 16], xmm1
lea edx, [edx + 32]
sub ecx, 8
jg convertloop
ret
}
}
#ifdef HAS_J400TOARGBROW_AVX2
// Duplicates gray value 3 times and fills in alpha opaque.
__declspec(naked) void J400ToARGBRow_AVX2(const uint8_t* src_y,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, [esp + 4] // src_y
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0xff000000
vpslld ymm5, ymm5, 24
convertloop:
vmovdqu xmm0, [eax]
lea eax, [eax + 16]
vpermq ymm0, ymm0, 0xd8
vpunpcklbw ymm0, ymm0, ymm0
vpermq ymm0, ymm0, 0xd8
vpunpckhwd ymm1, ymm0, ymm0
vpunpcklwd ymm0, ymm0, ymm0
vpor ymm0, ymm0, ymm5
vpor ymm1, ymm1, ymm5
vmovdqu [edx], ymm0
vmovdqu [edx + 32], ymm1
lea edx, [edx + 64]
sub ecx, 16
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_J400TOARGBROW_AVX2
__declspec(naked) void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, [esp + 4] // src_rgb24
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
pcmpeqb xmm5, xmm5 // generate mask 0xff000000
pslld xmm5, 24
movdqa xmm4, xmmword ptr kShuffleMaskRGB24ToARGB
convertloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 16]
movdqu xmm3, [eax + 32]
lea eax, [eax + 48]
movdqa xmm2, xmm3
palignr xmm2, xmm1, 8 // xmm2 = { xmm3[0:3] xmm1[8:15]}
pshufb xmm2, xmm4
por xmm2, xmm5
palignr xmm1, xmm0, 12 // xmm1 = { xmm3[0:7] xmm0[12:15]}
pshufb xmm0, xmm4
movdqu [edx + 32], xmm2
por xmm0, xmm5
pshufb xmm1, xmm4
movdqu [edx], xmm0
por xmm1, xmm5
palignr xmm3, xmm3, 4 // xmm3 = { xmm3[4:15]}
pshufb xmm3, xmm4
movdqu [edx + 16], xmm1
por xmm3, xmm5
movdqu [edx + 48], xmm3
lea edx, [edx + 64]
sub ecx, 16
jg convertloop
ret
}
}
__declspec(naked) void RAWToARGBRow_SSSE3(const uint8_t* src_raw,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, [esp + 4] // src_raw
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
pcmpeqb xmm5, xmm5 // generate mask 0xff000000
pslld xmm5, 24
movdqa xmm4, xmmword ptr kShuffleMaskRAWToARGB
convertloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 16]
movdqu xmm3, [eax + 32]
lea eax, [eax + 48]
movdqa xmm2, xmm3
palignr xmm2, xmm1, 8 // xmm2 = { xmm3[0:3] xmm1[8:15]}
pshufb xmm2, xmm4
por xmm2, xmm5
palignr xmm1, xmm0, 12 // xmm1 = { xmm3[0:7] xmm0[12:15]}
pshufb xmm0, xmm4
movdqu [edx + 32], xmm2
por xmm0, xmm5
pshufb xmm1, xmm4
movdqu [edx], xmm0
por xmm1, xmm5
palignr xmm3, xmm3, 4 // xmm3 = { xmm3[4:15]}
pshufb xmm3, xmm4
movdqu [edx + 16], xmm1
por xmm3, xmm5
movdqu [edx + 48], xmm3
lea edx, [edx + 64]
sub ecx, 16
jg convertloop
ret
}
}
__declspec(naked) void RAWToRGB24Row_SSSE3(const uint8_t* src_raw,
uint8_t* dst_rgb24,
int width) {
__asm {
mov eax, [esp + 4] // src_raw
mov edx, [esp + 8] // dst_rgb24
mov ecx, [esp + 12] // width
movdqa xmm3, xmmword ptr kShuffleMaskRAWToRGB24_0
movdqa xmm4, xmmword ptr kShuffleMaskRAWToRGB24_1
movdqa xmm5, xmmword ptr kShuffleMaskRAWToRGB24_2
convertloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 4]
movdqu xmm2, [eax + 8]
lea eax, [eax + 24]
pshufb xmm0, xmm3
pshufb xmm1, xmm4
pshufb xmm2, xmm5
movq qword ptr [edx], xmm0
movq qword ptr [edx + 8], xmm1
movq qword ptr [edx + 16], xmm2
lea edx, [edx + 24]
sub ecx, 8
jg convertloop
ret
}
}
// pmul method to replicate bits.
// Math to replicate bits:
// (v << 8) | (v << 3)
// v * 256 + v * 8
// v * (256 + 8)
// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
// 20 instructions.
__declspec(naked) void RGB565ToARGBRow_SSE2(const uint8_t* src_rgb565,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, 0x01080108 // generate multiplier to repeat 5 bits
movd xmm5, eax
pshufd xmm5, xmm5, 0
mov eax, 0x20802080 // multiplier shift by 5 and then repeat 6 bits
movd xmm6, eax
pshufd xmm6, xmm6, 0
pcmpeqb xmm3, xmm3 // generate mask 0xf800f800 for Red
psllw xmm3, 11
pcmpeqb xmm4, xmm4 // generate mask 0x07e007e0 for Green
psllw xmm4, 10
psrlw xmm4, 5
pcmpeqb xmm7, xmm7 // generate mask 0xff00ff00 for Alpha
psllw xmm7, 8
mov eax, [esp + 4] // src_rgb565
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
sub edx, eax
sub edx, eax
convertloop:
movdqu xmm0, [eax] // fetch 8 pixels of bgr565
movdqa xmm1, xmm0
movdqa xmm2, xmm0
pand xmm1, xmm3 // R in upper 5 bits
psllw xmm2, 11 // B in upper 5 bits
pmulhuw xmm1, xmm5 // * (256 + 8)
pmulhuw xmm2, xmm5 // * (256 + 8)
psllw xmm1, 8
por xmm1, xmm2 // RB
pand xmm0, xmm4 // G in middle 6 bits
pmulhuw xmm0, xmm6 // << 5 * (256 + 4)
por xmm0, xmm7 // AG
movdqa xmm2, xmm1
punpcklbw xmm1, xmm0
punpckhbw xmm2, xmm0
movdqu [eax * 2 + edx], xmm1 // store 4 pixels of ARGB
movdqu [eax * 2 + edx + 16], xmm2 // store next 4 pixels of ARGB
lea eax, [eax + 16]
sub ecx, 8
jg convertloop
ret
}
}
#ifdef HAS_RGB565TOARGBROW_AVX2
// pmul method to replicate bits.
// Math to replicate bits:
// (v << 8) | (v << 3)
// v * 256 + v * 8
// v * (256 + 8)
// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
__declspec(naked) void RGB565ToARGBRow_AVX2(const uint8_t* src_rgb565,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, 0x01080108 // generate multiplier to repeat 5 bits
vmovd xmm5, eax
vbroadcastss ymm5, xmm5
mov eax, 0x20802080 // multiplier shift by 5 and then repeat 6 bits
vmovd xmm6, eax
vbroadcastss ymm6, xmm6
vpcmpeqb ymm3, ymm3, ymm3 // generate mask 0xf800f800 for Red
vpsllw ymm3, ymm3, 11
vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0x07e007e0 for Green
vpsllw ymm4, ymm4, 10
vpsrlw ymm4, ymm4, 5
vpcmpeqb ymm7, ymm7, ymm7 // generate mask 0xff00ff00 for Alpha
vpsllw ymm7, ymm7, 8
mov eax, [esp + 4] // src_rgb565
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
sub edx, eax
sub edx, eax
convertloop:
vmovdqu ymm0, [eax] // fetch 16 pixels of bgr565
vpand ymm1, ymm0, ymm3 // R in upper 5 bits
vpsllw ymm2, ymm0, 11 // B in upper 5 bits
vpmulhuw ymm1, ymm1, ymm5 // * (256 + 8)
vpmulhuw ymm2, ymm2, ymm5 // * (256 + 8)
vpsllw ymm1, ymm1, 8
vpor ymm1, ymm1, ymm2 // RB
vpand ymm0, ymm0, ymm4 // G in middle 6 bits
vpmulhuw ymm0, ymm0, ymm6 // << 5 * (256 + 4)
vpor ymm0, ymm0, ymm7 // AG
vpermq ymm0, ymm0, 0xd8 // mutate for unpack
vpermq ymm1, ymm1, 0xd8
vpunpckhbw ymm2, ymm1, ymm0
vpunpcklbw ymm1, ymm1, ymm0
vmovdqu [eax * 2 + edx], ymm1 // store 4 pixels of ARGB
vmovdqu [eax * 2 + edx + 32], ymm2 // store next 4 pixels of ARGB
lea eax, [eax + 32]
sub ecx, 16
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_RGB565TOARGBROW_AVX2
#ifdef HAS_ARGB1555TOARGBROW_AVX2
__declspec(naked) void ARGB1555ToARGBRow_AVX2(const uint8_t* src_argb1555,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, 0x01080108 // generate multiplier to repeat 5 bits
vmovd xmm5, eax
vbroadcastss ymm5, xmm5
mov eax, 0x42004200 // multiplier shift by 6 and then repeat 5 bits
vmovd xmm6, eax
vbroadcastss ymm6, xmm6
vpcmpeqb ymm3, ymm3, ymm3 // generate mask 0xf800f800 for Red
vpsllw ymm3, ymm3, 11
vpsrlw ymm4, ymm3, 6 // generate mask 0x03e003e0 for Green
vpcmpeqb ymm7, ymm7, ymm7 // generate mask 0xff00ff00 for Alpha
vpsllw ymm7, ymm7, 8
mov eax, [esp + 4] // src_argb1555
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
sub edx, eax
sub edx, eax
convertloop:
vmovdqu ymm0, [eax] // fetch 16 pixels of 1555
vpsllw ymm1, ymm0, 1 // R in upper 5 bits
vpsllw ymm2, ymm0, 11 // B in upper 5 bits
vpand ymm1, ymm1, ymm3
vpmulhuw ymm2, ymm2, ymm5 // * (256 + 8)
vpmulhuw ymm1, ymm1, ymm5 // * (256 + 8)
vpsllw ymm1, ymm1, 8
vpor ymm1, ymm1, ymm2 // RB
vpsraw ymm2, ymm0, 8 // A
vpand ymm0, ymm0, ymm4 // G in middle 5 bits
vpmulhuw ymm0, ymm0, ymm6 // << 6 * (256 + 8)
vpand ymm2, ymm2, ymm7
vpor ymm0, ymm0, ymm2 // AG
vpermq ymm0, ymm0, 0xd8 // mutate for unpack
vpermq ymm1, ymm1, 0xd8
vpunpckhbw ymm2, ymm1, ymm0
vpunpcklbw ymm1, ymm1, ymm0
vmovdqu [eax * 2 + edx], ymm1 // store 8 pixels of ARGB
vmovdqu [eax * 2 + edx + 32], ymm2 // store next 8 pixels of ARGB
lea eax, [eax + 32]
sub ecx, 16
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_ARGB1555TOARGBROW_AVX2
#ifdef HAS_ARGB4444TOARGBROW_AVX2
__declspec(naked) void ARGB4444ToARGBRow_AVX2(const uint8_t* src_argb4444,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, 0x0f0f0f0f // generate mask 0x0f0f0f0f
vmovd xmm4, eax
vbroadcastss ymm4, xmm4
vpslld ymm5, ymm4, 4 // 0xf0f0f0f0 for high nibbles
mov eax, [esp + 4] // src_argb4444
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
sub edx, eax
sub edx, eax
convertloop:
vmovdqu ymm0, [eax] // fetch 16 pixels of bgra4444
vpand ymm2, ymm0, ymm5 // mask high nibbles
vpand ymm0, ymm0, ymm4 // mask low nibbles
vpsrlw ymm3, ymm2, 4
vpsllw ymm1, ymm0, 4
vpor ymm2, ymm2, ymm3
vpor ymm0, ymm0, ymm1
vpermq ymm0, ymm0, 0xd8 // mutate for unpack
vpermq ymm2, ymm2, 0xd8
vpunpckhbw ymm1, ymm0, ymm2
vpunpcklbw ymm0, ymm0, ymm2
vmovdqu [eax * 2 + edx], ymm0 // store 8 pixels of ARGB
vmovdqu [eax * 2 + edx + 32], ymm1 // store next 8 pixels of ARGB
lea eax, [eax + 32]
sub ecx, 16
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_ARGB4444TOARGBROW_AVX2
// 24 instructions
__declspec(naked) void ARGB1555ToARGBRow_SSE2(const uint8_t* src_argb1555,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, 0x01080108 // generate multiplier to repeat 5 bits
movd xmm5, eax
pshufd xmm5, xmm5, 0
mov eax, 0x42004200 // multiplier shift by 6 and then repeat 5 bits
movd xmm6, eax
pshufd xmm6, xmm6, 0
pcmpeqb xmm3, xmm3 // generate mask 0xf800f800 for Red
psllw xmm3, 11
movdqa xmm4, xmm3 // generate mask 0x03e003e0 for Green
psrlw xmm4, 6
pcmpeqb xmm7, xmm7 // generate mask 0xff00ff00 for Alpha
psllw xmm7, 8
mov eax, [esp + 4] // src_argb1555
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
sub edx, eax
sub edx, eax
convertloop:
movdqu xmm0, [eax] // fetch 8 pixels of 1555
movdqa xmm1, xmm0
movdqa xmm2, xmm0
psllw xmm1, 1 // R in upper 5 bits
psllw xmm2, 11 // B in upper 5 bits
pand xmm1, xmm3
pmulhuw xmm2, xmm5 // * (256 + 8)
pmulhuw xmm1, xmm5 // * (256 + 8)
psllw xmm1, 8
por xmm1, xmm2 // RB
movdqa xmm2, xmm0
pand xmm0, xmm4 // G in middle 5 bits
psraw xmm2, 8 // A
pmulhuw xmm0, xmm6 // << 6 * (256 + 8)
pand xmm2, xmm7
por xmm0, xmm2 // AG
movdqa xmm2, xmm1
punpcklbw xmm1, xmm0
punpckhbw xmm2, xmm0
movdqu [eax * 2 + edx], xmm1 // store 4 pixels of ARGB
movdqu [eax * 2 + edx + 16], xmm2 // store next 4 pixels of ARGB
lea eax, [eax + 16]
sub ecx, 8
jg convertloop
ret
}
}
// 18 instructions.
__declspec(naked) void ARGB4444ToARGBRow_SSE2(const uint8_t* src_argb4444,
uint8_t* dst_argb,
int width) {
__asm {
mov eax, 0x0f0f0f0f // generate mask 0x0f0f0f0f
movd xmm4, eax
pshufd xmm4, xmm4, 0
movdqa xmm5, xmm4 // 0xf0f0f0f0 for high nibbles
pslld xmm5, 4
mov eax, [esp + 4] // src_argb4444
mov edx, [esp + 8] // dst_argb
mov ecx, [esp + 12] // width
sub edx, eax
sub edx, eax
convertloop:
movdqu xmm0, [eax] // fetch 8 pixels of bgra4444
movdqa xmm2, xmm0
pand xmm0, xmm4 // mask low nibbles
pand xmm2, xmm5 // mask high nibbles
movdqa xmm1, xmm0
movdqa xmm3, xmm2
psllw xmm1, 4
psrlw xmm3, 4
por xmm0, xmm1
por xmm2, xmm3
movdqa xmm1, xmm0
punpcklbw xmm0, xmm2
punpckhbw xmm1, xmm2
movdqu [eax * 2 + edx], xmm0 // store 4 pixels of ARGB
movdqu [eax * 2 + edx + 16], xmm1 // store next 4 pixels of ARGB
lea eax, [eax + 16]
sub ecx, 8
jg convertloop
ret
}
}
__declspec(naked) void ARGBToRGB24Row_SSSE3(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
mov ecx, [esp + 12] // width
movdqa xmm6, xmmword ptr kShuffleMaskARGBToRGB24
convertloop:
movdqu xmm0, [eax] // fetch 16 pixels of argb
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
lea eax, [eax + 64]
pshufb xmm0, xmm6 // pack 16 bytes of ARGB to 12 bytes of RGB
pshufb xmm1, xmm6
pshufb xmm2, xmm6
pshufb xmm3, xmm6
movdqa xmm4, xmm1 // 4 bytes from 1 for 0
psrldq xmm1, 4 // 8 bytes from 1
pslldq xmm4, 12 // 4 bytes from 1 for 0
movdqa xmm5, xmm2 // 8 bytes from 2 for 1
por xmm0, xmm4 // 4 bytes from 1 for 0
pslldq xmm5, 8 // 8 bytes from 2 for 1
movdqu [edx], xmm0 // store 0
por xmm1, xmm5 // 8 bytes from 2 for 1
psrldq xmm2, 8 // 4 bytes from 2
pslldq xmm3, 4 // 12 bytes from 3 for 2
por xmm2, xmm3 // 12 bytes from 3 for 2
movdqu [edx + 16], xmm1 // store 1
movdqu [edx + 32], xmm2 // store 2
lea edx, [edx + 48]
sub ecx, 16
jg convertloop
ret
}
}
__declspec(naked) void ARGBToRAWRow_SSSE3(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
mov ecx, [esp + 12] // width
movdqa xmm6, xmmword ptr kShuffleMaskARGBToRAW
convertloop:
movdqu xmm0, [eax] // fetch 16 pixels of argb
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
lea eax, [eax + 64]
pshufb xmm0, xmm6 // pack 16 bytes of ARGB to 12 bytes of RGB
pshufb xmm1, xmm6
pshufb xmm2, xmm6
pshufb xmm3, xmm6
movdqa xmm4, xmm1 // 4 bytes from 1 for 0
psrldq xmm1, 4 // 8 bytes from 1
pslldq xmm4, 12 // 4 bytes from 1 for 0
movdqa xmm5, xmm2 // 8 bytes from 2 for 1
por xmm0, xmm4 // 4 bytes from 1 for 0
pslldq xmm5, 8 // 8 bytes from 2 for 1
movdqu [edx], xmm0 // store 0
por xmm1, xmm5 // 8 bytes from 2 for 1
psrldq xmm2, 8 // 4 bytes from 2
pslldq xmm3, 4 // 12 bytes from 3 for 2
por xmm2, xmm3 // 12 bytes from 3 for 2
movdqu [edx + 16], xmm1 // store 1
movdqu [edx + 32], xmm2 // store 2
lea edx, [edx + 48]
sub ecx, 16
jg convertloop
ret
}
}
__declspec(naked) void ARGBToRGB565Row_SSE2(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
mov ecx, [esp + 12] // width
pcmpeqb xmm3, xmm3 // generate mask 0x0000001f
psrld xmm3, 27
pcmpeqb xmm4, xmm4 // generate mask 0x000007e0
psrld xmm4, 26
pslld xmm4, 5
pcmpeqb xmm5, xmm5 // generate mask 0xfffff800
pslld xmm5, 11
convertloop:
movdqu xmm0, [eax] // fetch 4 pixels of argb
movdqa xmm1, xmm0 // B
movdqa xmm2, xmm0 // G
pslld xmm0, 8 // R
psrld xmm1, 3 // B
psrld xmm2, 5 // G
psrad xmm0, 16 // R
pand xmm1, xmm3 // B
pand xmm2, xmm4 // G
pand xmm0, xmm5 // R
por xmm1, xmm2 // BG
por xmm0, xmm1 // BGR
packssdw xmm0, xmm0
lea eax, [eax + 16]
movq qword ptr [edx], xmm0 // store 4 pixels of RGB565
lea edx, [edx + 8]
sub ecx, 4
jg convertloop
ret
}
}
__declspec(naked) void ARGBToRGB565DitherRow_SSE2(const uint8_t* src_argb,
uint8_t* dst_rgb,
const uint32_t dither4,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
movd xmm6, [esp + 12] // dither4
mov ecx, [esp + 16] // width
punpcklbw xmm6, xmm6 // make dither 16 bytes
movdqa xmm7, xmm6
punpcklwd xmm6, xmm6
punpckhwd xmm7, xmm7
pcmpeqb xmm3, xmm3 // generate mask 0x0000001f
psrld xmm3, 27
pcmpeqb xmm4, xmm4 // generate mask 0x000007e0
psrld xmm4, 26
pslld xmm4, 5
pcmpeqb xmm5, xmm5 // generate mask 0xfffff800
pslld xmm5, 11
convertloop:
movdqu xmm0, [eax] // fetch 4 pixels of argb
paddusb xmm0, xmm6 // add dither
movdqa xmm1, xmm0 // B
movdqa xmm2, xmm0 // G
pslld xmm0, 8 // R
psrld xmm1, 3 // B
psrld xmm2, 5 // G
psrad xmm0, 16 // R
pand xmm1, xmm3 // B
pand xmm2, xmm4 // G
pand xmm0, xmm5 // R
por xmm1, xmm2 // BG
por xmm0, xmm1 // BGR
packssdw xmm0, xmm0
lea eax, [eax + 16]
movq qword ptr [edx], xmm0 // store 4 pixels of RGB565
lea edx, [edx + 8]
sub ecx, 4
jg convertloop
ret
}
}
#ifdef HAS_ARGBTORGB565DITHERROW_AVX2
__declspec(naked) void ARGBToRGB565DitherRow_AVX2(const uint8_t* src_argb,
uint8_t* dst_rgb,
const uint32_t dither4,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
vbroadcastss xmm6, [esp + 12] // dither4
mov ecx, [esp + 16] // width
vpunpcklbw xmm6, xmm6, xmm6 // make dither 32 bytes
vpermq ymm6, ymm6, 0xd8
vpunpcklwd ymm6, ymm6, ymm6
vpcmpeqb ymm3, ymm3, ymm3 // generate mask 0x0000001f
vpsrld ymm3, ymm3, 27
vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0x000007e0
vpsrld ymm4, ymm4, 26
vpslld ymm4, ymm4, 5
vpslld ymm5, ymm3, 11 // generate mask 0x0000f800
convertloop:
vmovdqu ymm0, [eax] // fetch 8 pixels of argb
vpaddusb ymm0, ymm0, ymm6 // add dither
vpsrld ymm2, ymm0, 5 // G
vpsrld ymm1, ymm0, 3 // B
vpsrld ymm0, ymm0, 8 // R
vpand ymm2, ymm2, ymm4 // G
vpand ymm1, ymm1, ymm3 // B
vpand ymm0, ymm0, ymm5 // R
vpor ymm1, ymm1, ymm2 // BG
vpor ymm0, ymm0, ymm1 // BGR
vpackusdw ymm0, ymm0, ymm0
vpermq ymm0, ymm0, 0xd8
lea eax, [eax + 32]
vmovdqu [edx], xmm0 // store 8 pixels of RGB565
lea edx, [edx + 16]
sub ecx, 8
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_ARGBTORGB565DITHERROW_AVX2
// TODO(fbarchard): Improve sign extension/packing.
__declspec(naked) void ARGBToARGB1555Row_SSE2(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
mov ecx, [esp + 12] // width
pcmpeqb xmm4, xmm4 // generate mask 0x0000001f
psrld xmm4, 27
movdqa xmm5, xmm4 // generate mask 0x000003e0
pslld xmm5, 5
movdqa xmm6, xmm4 // generate mask 0x00007c00
pslld xmm6, 10
pcmpeqb xmm7, xmm7 // generate mask 0xffff8000
pslld xmm7, 15
convertloop:
movdqu xmm0, [eax] // fetch 4 pixels of argb
movdqa xmm1, xmm0 // B
movdqa xmm2, xmm0 // G
movdqa xmm3, xmm0 // R
psrad xmm0, 16 // A
psrld xmm1, 3 // B
psrld xmm2, 6 // G
psrld xmm3, 9 // R
pand xmm0, xmm7 // A
pand xmm1, xmm4 // B
pand xmm2, xmm5 // G
pand xmm3, xmm6 // R
por xmm0, xmm1 // BA
por xmm2, xmm3 // GR
por xmm0, xmm2 // BGRA
packssdw xmm0, xmm0
lea eax, [eax + 16]
movq qword ptr [edx], xmm0 // store 4 pixels of ARGB1555
lea edx, [edx + 8]
sub ecx, 4
jg convertloop
ret
}
}
__declspec(naked) void ARGBToARGB4444Row_SSE2(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
mov ecx, [esp + 12] // width
pcmpeqb xmm4, xmm4 // generate mask 0xf000f000
psllw xmm4, 12
movdqa xmm3, xmm4 // generate mask 0x00f000f0
psrlw xmm3, 8
convertloop:
movdqu xmm0, [eax] // fetch 4 pixels of argb
movdqa xmm1, xmm0
pand xmm0, xmm3 // low nibble
pand xmm1, xmm4 // high nibble
psrld xmm0, 4
psrld xmm1, 8
por xmm0, xmm1
packuswb xmm0, xmm0
lea eax, [eax + 16]
movq qword ptr [edx], xmm0 // store 4 pixels of ARGB4444
lea edx, [edx + 8]
sub ecx, 4
jg convertloop
ret
}
}
#ifdef HAS_ARGBTORGB565ROW_AVX2
__declspec(naked) void ARGBToRGB565Row_AVX2(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
mov ecx, [esp + 12] // width
vpcmpeqb ymm3, ymm3, ymm3 // generate mask 0x0000001f
vpsrld ymm3, ymm3, 27
vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0x000007e0
vpsrld ymm4, ymm4, 26
vpslld ymm4, ymm4, 5
vpslld ymm5, ymm3, 11 // generate mask 0x0000f800
convertloop:
vmovdqu ymm0, [eax] // fetch 8 pixels of argb
vpsrld ymm2, ymm0, 5 // G
vpsrld ymm1, ymm0, 3 // B
vpsrld ymm0, ymm0, 8 // R
vpand ymm2, ymm2, ymm4 // G
vpand ymm1, ymm1, ymm3 // B
vpand ymm0, ymm0, ymm5 // R
vpor ymm1, ymm1, ymm2 // BG
vpor ymm0, ymm0, ymm1 // BGR
vpackusdw ymm0, ymm0, ymm0
vpermq ymm0, ymm0, 0xd8
lea eax, [eax + 32]
vmovdqu [edx], xmm0 // store 8 pixels of RGB565
lea edx, [edx + 16]
sub ecx, 8
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_ARGBTORGB565ROW_AVX2
#ifdef HAS_ARGBTOARGB1555ROW_AVX2
__declspec(naked) void ARGBToARGB1555Row_AVX2(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
mov ecx, [esp + 12] // width
vpcmpeqb ymm4, ymm4, ymm4
vpsrld ymm4, ymm4, 27 // generate mask 0x0000001f
vpslld ymm5, ymm4, 5 // generate mask 0x000003e0
vpslld ymm6, ymm4, 10 // generate mask 0x00007c00
vpcmpeqb ymm7, ymm7, ymm7 // generate mask 0xffff8000
vpslld ymm7, ymm7, 15
convertloop:
vmovdqu ymm0, [eax] // fetch 8 pixels of argb
vpsrld ymm3, ymm0, 9 // R
vpsrld ymm2, ymm0, 6 // G
vpsrld ymm1, ymm0, 3 // B
vpsrad ymm0, ymm0, 16 // A
vpand ymm3, ymm3, ymm6 // R
vpand ymm2, ymm2, ymm5 // G
vpand ymm1, ymm1, ymm4 // B
vpand ymm0, ymm0, ymm7 // A
vpor ymm0, ymm0, ymm1 // BA
vpor ymm2, ymm2, ymm3 // GR
vpor ymm0, ymm0, ymm2 // BGRA
vpackssdw ymm0, ymm0, ymm0
vpermq ymm0, ymm0, 0xd8
lea eax, [eax + 32]
vmovdqu [edx], xmm0 // store 8 pixels of ARGB1555
lea edx, [edx + 16]
sub ecx, 8
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_ARGBTOARGB1555ROW_AVX2
#ifdef HAS_ARGBTOARGB4444ROW_AVX2
__declspec(naked) void ARGBToARGB4444Row_AVX2(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_rgb
mov ecx, [esp + 12] // width
vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0xf000f000
vpsllw ymm4, ymm4, 12
vpsrlw ymm3, ymm4, 8 // generate mask 0x00f000f0
convertloop:
vmovdqu ymm0, [eax] // fetch 8 pixels of argb
vpand ymm1, ymm0, ymm4 // high nibble
vpand ymm0, ymm0, ymm3 // low nibble
vpsrld ymm1, ymm1, 8
vpsrld ymm0, ymm0, 4
vpor ymm0, ymm0, ymm1
vpackuswb ymm0, ymm0, ymm0
vpermq ymm0, ymm0, 0xd8
lea eax, [eax + 32]
vmovdqu [edx], xmm0 // store 8 pixels of ARGB4444
lea edx, [edx + 16]
sub ecx, 8
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_ARGBTOARGB4444ROW_AVX2
// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
__declspec(naked) void ARGBToYRow_SSSE3(const uint8_t* src_argb,
uint8_t* dst_y,
int width) {
__asm {
mov eax, [esp + 4] /* src_argb */
mov edx, [esp + 8] /* dst_y */
mov ecx, [esp + 12] /* width */
movdqa xmm4, xmmword ptr kARGBToY
movdqa xmm5, xmmword ptr kAddY16
convertloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
pmaddubsw xmm0, xmm4
pmaddubsw xmm1, xmm4
pmaddubsw xmm2, xmm4
pmaddubsw xmm3, xmm4
lea eax, [eax + 64]
phaddw xmm0, xmm1
phaddw xmm2, xmm3
psrlw xmm0, 7
psrlw xmm2, 7
packuswb xmm0, xmm2
paddb xmm0, xmm5
movdqu [edx], xmm0
lea edx, [edx + 16]
sub ecx, 16
jg convertloop
ret
}
}
// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
__declspec(naked) void ARGBToYJRow_SSSE3(const uint8_t* src_argb,
uint8_t* dst_y,
int width) {
__asm {
mov eax, [esp + 4] /* src_argb */
mov edx, [esp + 8] /* dst_y */
mov ecx, [esp + 12] /* width */
movdqa xmm4, xmmword ptr kARGBToYJ
movdqa xmm5, xmmword ptr kAddYJ64
convertloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
pmaddubsw xmm0, xmm4
pmaddubsw xmm1, xmm4
pmaddubsw xmm2, xmm4
pmaddubsw xmm3, xmm4
lea eax, [eax + 64]
phaddw xmm0, xmm1
phaddw xmm2, xmm3
paddw xmm0, xmm5 // Add .5 for rounding.
paddw xmm2, xmm5
psrlw xmm0, 7
psrlw xmm2, 7
packuswb xmm0, xmm2
movdqu [edx], xmm0
lea edx, [edx + 16]
sub ecx, 16
jg convertloop
ret
}
}
#ifdef HAS_ARGBTOYROW_AVX2
// vpermd for vphaddw + vpackuswb vpermd.
static const lvec32 kPermdARGBToY_AVX = {0, 4, 1, 5, 2, 6, 3, 7};
// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
__declspec(naked) void ARGBToYRow_AVX2(const uint8_t* src_argb,
uint8_t* dst_y,
int width) {
__asm {
mov eax, [esp + 4] /* src_argb */
mov edx, [esp + 8] /* dst_y */
mov ecx, [esp + 12] /* width */
vbroadcastf128 ymm4, xmmword ptr kARGBToY
vbroadcastf128 ymm5, xmmword ptr kAddY16
vmovdqu ymm6, ymmword ptr kPermdARGBToY_AVX
convertloop:
vmovdqu ymm0, [eax]
vmovdqu ymm1, [eax + 32]
vmovdqu ymm2, [eax + 64]
vmovdqu ymm3, [eax + 96]
vpmaddubsw ymm0, ymm0, ymm4
vpmaddubsw ymm1, ymm1, ymm4
vpmaddubsw ymm2, ymm2, ymm4
vpmaddubsw ymm3, ymm3, ymm4
lea eax, [eax + 128]
vphaddw ymm0, ymm0, ymm1 // mutates.
vphaddw ymm2, ymm2, ymm3
vpsrlw ymm0, ymm0, 7
vpsrlw ymm2, ymm2, 7
vpackuswb ymm0, ymm0, ymm2 // mutates.
vpermd ymm0, ymm6, ymm0 // For vphaddw + vpackuswb mutation.
vpaddb ymm0, ymm0, ymm5 // add 16 for Y
vmovdqu [edx], ymm0
lea edx, [edx + 32]
sub ecx, 32
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_ARGBTOYROW_AVX2
#ifdef HAS_ARGBTOYJROW_AVX2
// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
__declspec(naked) void ARGBToYJRow_AVX2(const uint8_t* src_argb,
uint8_t* dst_y,
int width) {
__asm {
mov eax, [esp + 4] /* src_argb */
mov edx, [esp + 8] /* dst_y */
mov ecx, [esp + 12] /* width */
vbroadcastf128 ymm4, xmmword ptr kARGBToYJ
vbroadcastf128 ymm5, xmmword ptr kAddYJ64
vmovdqu ymm6, ymmword ptr kPermdARGBToY_AVX
convertloop:
vmovdqu ymm0, [eax]
vmovdqu ymm1, [eax + 32]
vmovdqu ymm2, [eax + 64]
vmovdqu ymm3, [eax + 96]
vpmaddubsw ymm0, ymm0, ymm4
vpmaddubsw ymm1, ymm1, ymm4
vpmaddubsw ymm2, ymm2, ymm4
vpmaddubsw ymm3, ymm3, ymm4
lea eax, [eax + 128]
vphaddw ymm0, ymm0, ymm1 // mutates.
vphaddw ymm2, ymm2, ymm3
vpaddw ymm0, ymm0, ymm5 // Add .5 for rounding.
vpaddw ymm2, ymm2, ymm5
vpsrlw ymm0, ymm0, 7
vpsrlw ymm2, ymm2, 7
vpackuswb ymm0, ymm0, ymm2 // mutates.
vpermd ymm0, ymm6, ymm0 // For vphaddw + vpackuswb mutation.
vmovdqu [edx], ymm0
lea edx, [edx + 32]
sub ecx, 32
jg convertloop
vzeroupper
ret
}
}
#endif // HAS_ARGBTOYJROW_AVX2
__declspec(naked) void BGRAToYRow_SSSE3(const uint8_t* src_argb,
uint8_t* dst_y,
int width) {
__asm {
mov eax, [esp + 4] /* src_argb */
mov edx, [esp + 8] /* dst_y */
mov ecx, [esp + 12] /* width */
movdqa xmm4, xmmword ptr kBGRAToY
movdqa xmm5, xmmword ptr kAddY16
convertloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
pmaddubsw xmm0, xmm4
pmaddubsw xmm1, xmm4
pmaddubsw xmm2, xmm4
pmaddubsw xmm3, xmm4
lea eax, [eax + 64]
phaddw xmm0, xmm1
phaddw xmm2, xmm3
psrlw xmm0, 7
psrlw xmm2, 7
packuswb xmm0, xmm2
paddb xmm0, xmm5
movdqu [edx], xmm0
lea edx, [edx + 16]
sub ecx, 16
jg convertloop
ret
}
}
__declspec(naked) void ABGRToYRow_SSSE3(const uint8_t* src_argb,
uint8_t* dst_y,
int width) {
__asm {
mov eax, [esp + 4] /* src_argb */
mov edx, [esp + 8] /* dst_y */
mov ecx, [esp + 12] /* width */
movdqa xmm4, xmmword ptr kABGRToY
movdqa xmm5, xmmword ptr kAddY16
convertloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
pmaddubsw xmm0, xmm4
pmaddubsw xmm1, xmm4
pmaddubsw xmm2, xmm4
pmaddubsw xmm3, xmm4
lea eax, [eax + 64]
phaddw xmm0, xmm1
phaddw xmm2, xmm3
psrlw xmm0, 7
psrlw xmm2, 7
packuswb xmm0, xmm2
paddb xmm0, xmm5
movdqu [edx], xmm0
lea edx, [edx + 16]
sub ecx, 16
jg convertloop
ret
}
}
__declspec(naked) void RGBAToYRow_SSSE3(const uint8_t* src_argb,
uint8_t* dst_y,
int width) {
__asm {
mov eax, [esp + 4] /* src_argb */
mov edx, [esp + 8] /* dst_y */
mov ecx, [esp + 12] /* width */
movdqa xmm4, xmmword ptr kRGBAToY
movdqa xmm5, xmmword ptr kAddY16
convertloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
pmaddubsw xmm0, xmm4
pmaddubsw xmm1, xmm4
pmaddubsw xmm2, xmm4
pmaddubsw xmm3, xmm4
lea eax, [eax + 64]
phaddw xmm0, xmm1
phaddw xmm2, xmm3
psrlw xmm0, 7
psrlw xmm2, 7
packuswb xmm0, xmm2
paddb xmm0, xmm5
movdqu [edx], xmm0
lea edx, [edx + 16]
sub ecx, 16
jg convertloop
ret
}
}
__declspec(naked) void ARGBToUVRow_SSSE3(const uint8_t* src_argb0,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
__asm {
push esi
push edi
mov eax, [esp + 8 + 4] // src_argb
mov esi, [esp + 8 + 8] // src_stride_argb
mov edx, [esp + 8 + 12] // dst_u
mov edi, [esp + 8 + 16] // dst_v
mov ecx, [esp + 8 + 20] // width
movdqa xmm5, xmmword ptr kAddUV128
movdqa xmm6, xmmword ptr kARGBToV
movdqa xmm7, xmmword ptr kARGBToU
sub edi, edx // stride from u to v
convertloop:
/* step 1 - subsample 16x2 argb pixels to 8x1 */
movdqu xmm0, [eax]
movdqu xmm4, [eax + esi]
pavgb xmm0, xmm4
movdqu xmm1, [eax + 16]
movdqu xmm4, [eax + esi + 16]
pavgb xmm1, xmm4
movdqu xmm2, [eax + 32]
movdqu xmm4, [eax + esi + 32]
pavgb xmm2, xmm4
movdqu xmm3, [eax + 48]
movdqu xmm4, [eax + esi + 48]
pavgb xmm3, xmm4
lea eax, [eax + 64]
movdqa xmm4, xmm0
shufps xmm0, xmm1, 0x88
shufps xmm4, xmm1, 0xdd
pavgb xmm0, xmm4
movdqa xmm4, xmm2
shufps xmm2, xmm3, 0x88
shufps xmm4, xmm3, 0xdd
pavgb xmm2, xmm4
// step 2 - convert to U and V
// from here down is very similar to Y code except
// instead of 16 different pixels, its 8 pixels of U and 8 of V
movdqa xmm1, xmm0
movdqa xmm3, xmm2
pmaddubsw xmm0, xmm7 // U
pmaddubsw xmm2, xmm7
pmaddubsw xmm1, xmm6 // V
pmaddubsw xmm3, xmm6
phaddw xmm0, xmm2
phaddw xmm1, xmm3
psraw xmm0, 8
psraw xmm1, 8
packsswb xmm0, xmm1
paddb xmm0, xmm5 // -> unsigned
// step 3 - store 8 U and 8 V values
movlps qword ptr [edx], xmm0 // U
movhps qword ptr [edx + edi], xmm0 // V
lea edx, [edx + 8]
sub ecx, 16
jg convertloop
pop edi
pop esi
ret
}
}
__declspec(naked) void ARGBToUVJRow_SSSE3(const uint8_t* src_argb0,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
__asm {
push esi
push edi
mov eax, [esp + 8 + 4] // src_argb
mov esi, [esp + 8 + 8] // src_stride_argb
mov edx, [esp + 8 + 12] // dst_u
mov edi, [esp + 8 + 16] // dst_v
mov ecx, [esp + 8 + 20] // width
movdqa xmm5, xmmword ptr kAddUVJ128
movdqa xmm6, xmmword ptr kARGBToVJ
movdqa xmm7, xmmword ptr kARGBToUJ
sub edi, edx // stride from u to v
convertloop:
/* step 1 - subsample 16x2 argb pixels to 8x1 */
movdqu xmm0, [eax]
movdqu xmm4, [eax + esi]
pavgb xmm0, xmm4
movdqu xmm1, [eax + 16]
movdqu xmm4, [eax + esi + 16]
pavgb xmm1, xmm4
movdqu xmm2, [eax + 32]
movdqu xmm4, [eax + esi + 32]
pavgb xmm2, xmm4
movdqu xmm3, [eax + 48]
movdqu xmm4, [eax + esi + 48]
pavgb xmm3, xmm4
lea eax, [eax + 64]
movdqa xmm4, xmm0
shufps xmm0, xmm1, 0x88
shufps xmm4, xmm1, 0xdd
pavgb xmm0, xmm4
movdqa xmm4, xmm2
shufps xmm2, xmm3, 0x88
shufps xmm4, xmm3, 0xdd
pavgb xmm2, xmm4
// step 2 - convert to U and V
// from here down is very similar to Y code except
// instead of 16 different pixels, its 8 pixels of U and 8 of V
movdqa xmm1, xmm0
movdqa xmm3, xmm2
pmaddubsw xmm0, xmm7 // U
pmaddubsw xmm2, xmm7
pmaddubsw xmm1, xmm6 // V
pmaddubsw xmm3, xmm6
phaddw xmm0, xmm2
phaddw xmm1, xmm3
paddw xmm0, xmm5 // +.5 rounding -> unsigned
paddw xmm1, xmm5
psraw xmm0, 8
psraw xmm1, 8
packsswb xmm0, xmm1
// step 3 - store 8 U and 8 V values
movlps qword ptr [edx], xmm0 // U
movhps qword ptr [edx + edi], xmm0 // V
lea edx, [edx + 8]
sub ecx, 16
jg convertloop
pop edi
pop esi
ret
}
}
#ifdef HAS_ARGBTOUVROW_AVX2
__declspec(naked) void ARGBToUVRow_AVX2(const uint8_t* src_argb0,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
__asm {
push esi
push edi
mov eax, [esp + 8 + 4] // src_argb
mov esi, [esp + 8 + 8] // src_stride_argb
mov edx, [esp + 8 + 12] // dst_u
mov edi, [esp + 8 + 16] // dst_v
mov ecx, [esp + 8 + 20] // width
vbroadcastf128 ymm5, xmmword ptr kAddUV128
vbroadcastf128 ymm6, xmmword ptr kARGBToV
vbroadcastf128 ymm7, xmmword ptr kARGBToU
sub edi, edx // stride from u to v
convertloop:
/* step 1 - subsample 32x2 argb pixels to 16x1 */
vmovdqu ymm0, [eax]
vmovdqu ymm1, [eax + 32]
vmovdqu ymm2, [eax + 64]
vmovdqu ymm3, [eax + 96]
vpavgb ymm0, ymm0, [eax + esi]
vpavgb ymm1, ymm1, [eax + esi + 32]
vpavgb ymm2, ymm2, [eax + esi + 64]
vpavgb ymm3, ymm3, [eax + esi + 96]
lea eax, [eax + 128]
vshufps ymm4, ymm0, ymm1, 0x88
vshufps ymm0, ymm0, ymm1, 0xdd
vpavgb ymm0, ymm0, ymm4 // mutated by vshufps
vshufps ymm4, ymm2, ymm3, 0x88
vshufps ymm2, ymm2, ymm3, 0xdd
vpavgb ymm2, ymm2, ymm4 // mutated by vshufps
// step 2 - convert to U and V
// from here down is very similar to Y code except
// instead of 32 different pixels, its 16 pixels of U and 16 of V
vpmaddubsw ymm1, ymm0, ymm7 // U
vpmaddubsw ymm3, ymm2, ymm7
vpmaddubsw ymm0, ymm0, ymm6 // V
vpmaddubsw ymm2, ymm2, ymm6
vphaddw ymm1, ymm1, ymm3 // mutates
vphaddw ymm0, ymm0, ymm2
vpsraw ymm1, ymm1, 8
vpsraw ymm0, ymm0, 8
vpacksswb ymm0, ymm1, ymm0 // mutates
vpermq ymm0, ymm0, 0xd8 // For vpacksswb
vpshufb ymm0, ymm0, ymmword ptr kShufARGBToUV_AVX // for vshufps/vphaddw
vpaddb ymm0, ymm0, ymm5 // -> unsigned
// step 3 - store 16 U and 16 V values
vextractf128 [edx], ymm0, 0 // U
vextractf128 [edx + edi], ymm0, 1 // V
lea edx, [edx + 16]
sub ecx, 32
jg convertloop
pop edi
pop esi
vzeroupper
ret
}
}
#endif // HAS_ARGBTOUVROW_AVX2
#ifdef HAS_ARGBTOUVJROW_AVX2
__declspec(naked) void ARGBToUVJRow_AVX2(const uint8_t* src_argb0,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
__asm {
push esi
push edi
mov eax, [esp + 8 + 4] // src_argb
mov esi, [esp + 8 + 8] // src_stride_argb
mov edx, [esp + 8 + 12] // dst_u
mov edi, [esp + 8 + 16] // dst_v
mov ecx, [esp + 8 + 20] // width
vbroadcastf128 ymm5, xmmword ptr kAddUV128
vbroadcastf128 ymm6, xmmword ptr kARGBToV
vbroadcastf128 ymm7, xmmword ptr kARGBToU
sub edi, edx // stride from u to v
convertloop:
/* step 1 - subsample 32x2 argb pixels to 16x1 */
vmovdqu ymm0, [eax]
vmovdqu ymm1, [eax + 32]
vmovdqu ymm2, [eax + 64]
vmovdqu ymm3, [eax + 96]
vpavgb ymm0, ymm0, [eax + esi]
vpavgb ymm1, ymm1, [eax + esi + 32]
vpavgb ymm2, ymm2, [eax + esi + 64]
vpavgb ymm3, ymm3, [eax + esi + 96]
lea eax, [eax + 128]
vshufps ymm4, ymm0, ymm1, 0x88
vshufps ymm0, ymm0, ymm1, 0xdd
vpavgb ymm0, ymm0, ymm4 // mutated by vshufps
vshufps ymm4, ymm2, ymm3, 0x88
vshufps ymm2, ymm2, ymm3, 0xdd
vpavgb ymm2, ymm2, ymm4 // mutated by vshufps
// step 2 - convert to U and V
// from here down is very similar to Y code except
// instead of 32 different pixels, its 16 pixels of U and 16 of V
vpmaddubsw ymm1, ymm0, ymm7 // U
vpmaddubsw ymm3, ymm2, ymm7
vpmaddubsw ymm0, ymm0, ymm6 // V
vpmaddubsw ymm2, ymm2, ymm6
vphaddw ymm1, ymm1, ymm3 // mutates
vphaddw ymm0, ymm0, ymm2
vpaddw ymm1, ymm1, ymm5 // +.5 rounding -> unsigned
vpaddw ymm0, ymm0, ymm5
vpsraw ymm1, ymm1, 8
vpsraw ymm0, ymm0, 8
vpacksswb ymm0, ymm1, ymm0 // mutates
vpermq ymm0, ymm0, 0xd8 // For vpacksswb
vpshufb ymm0, ymm0, ymmword ptr kShufARGBToUV_AVX // for vshufps/vphaddw
// step 3 - store 16 U and 16 V values
vextractf128 [edx], ymm0, 0 // U
vextractf128 [edx + edi], ymm0, 1 // V
lea edx, [edx + 16]
sub ecx, 32
jg convertloop
pop edi
pop esi
vzeroupper
ret
}
}
#endif // HAS_ARGBTOUVJROW_AVX2
__declspec(naked) void ARGBToUV444Row_SSSE3(const uint8_t* src_argb0,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
__asm {
push edi
mov eax, [esp + 4 + 4] // src_argb
mov edx, [esp + 4 + 8] // dst_u
mov edi, [esp + 4 + 12] // dst_v
mov ecx, [esp + 4 + 16] // width
movdqa xmm5, xmmword ptr kAddUV128
movdqa xmm6, xmmword ptr kARGBToV
movdqa xmm7, xmmword ptr kARGBToU
sub edi, edx // stride from u to v
convertloop:
/* convert to U and V */
movdqu xmm0, [eax] // U
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
pmaddubsw xmm0, xmm7
pmaddubsw xmm1, xmm7
pmaddubsw xmm2, xmm7
pmaddubsw xmm3, xmm7
phaddw xmm0, xmm1
phaddw xmm2, xmm3
psraw xmm0, 8
psraw xmm2, 8
packsswb xmm0, xmm2
paddb xmm0, xmm5
movdqu [edx], xmm0
movdqu xmm0, [eax] // V
movdqu xmm1, [eax + 16]
movdqu xmm2, [eax + 32]
movdqu xmm3, [eax + 48]
pmaddubsw xmm0, xmm6
pmaddubsw xmm1, xmm6
pmaddubsw xmm2, xmm6
pmaddubsw xmm3, xmm6
phaddw xmm0, xmm1
phaddw xmm2, xmm3
psraw xmm0, 8
psraw xmm2, 8
packsswb xmm0, xmm2
paddb xmm0, xmm5
lea eax, [eax + 64]
movdqu [edx + edi], xmm0
lea edx, [edx + 16]
sub ecx, 16
jg convertloop
pop edi
ret
}
}
__declspec(naked) void BGRAToUVRow_SSSE3(const uint8_t* src_argb0,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
__asm {
push esi
push edi
mov eax, [esp + 8 + 4] // src_argb
mov esi, [esp + 8 + 8] // src_stride_argb
mov edx, [esp + 8 + 12] // dst_u
mov edi, [esp + 8 + 16] // dst_v
mov ecx, [esp + 8 + 20] // width
movdqa xmm5, xmmword ptr kAddUV128
movdqa xmm6, xmmword ptr kBGRAToV
movdqa xmm7, xmmword ptr kBGRAToU
sub edi, edx // stride from u to v
convertloop:
/* step 1 - subsample 16x2 argb pixels to 8x1 */
movdqu xmm0, [eax]
movdqu xmm4, [eax + esi]
pavgb xmm0, xmm4
movdqu xmm1, [eax + 16]
movdqu xmm4, [eax + esi + 16]
pavgb xmm1, xmm4
movdqu xmm2, [eax + 32]
movdqu xmm4, [eax + esi + 32]
pavgb xmm2, xmm4
movdqu xmm3, [eax + 48]
movdqu xmm4, [eax + esi + 48]
pavgb xmm3, xmm4
lea eax, [eax + 64]
movdqa xmm4, xmm0
shufps xmm0, xmm1, 0x88
shufps xmm4, xmm1, 0xdd
pavgb xmm0, xmm4
movdqa xmm4, xmm2
shufps xmm2, xmm3, 0x88
shufps xmm4, xmm3, 0xdd
pavgb xmm2, xmm4
// step 2 - convert to U and V
// from here down is very similar to Y code except
// instead of 16 different pixels, its 8 pixels of U and 8 of V
movdqa xmm1, xmm0
movdqa xmm3, xmm2
pmaddubsw xmm0, xmm7 // U
pmaddubsw xmm2, xmm7
pmaddubsw xmm1, xmm6 // V
pmaddubsw xmm3, xmm6
phaddw xmm0, xmm2
phaddw xmm1, xmm3
psraw xmm0, 8
psraw xmm1, 8
packsswb xmm0, xmm1
paddb xmm0, xmm5 // -> unsigned
// step 3 - store 8 U and 8 V values
movlps qword ptr [edx], xmm0 // U
movhps qword ptr [edx + edi], xmm0 // V
lea edx, [edx + 8]
sub ecx, 16
jg convertloop
pop edi
pop esi
ret
}
}
__declspec(naked) void ABGRToUVRow_SSSE3(const uint8_t* src_argb0,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
__asm {
push esi
push edi
mov eax, [esp + 8 + 4] // src_argb
mov esi, [esp + 8 + 8] // src_stride_argb
mov edx, [esp + 8 + 12] // dst_u
mov edi, [esp + 8 + 16] // dst_v
mov ecx, [esp + 8 + 20] // width
movdqa xmm5, xmmword ptr kAddUV128
movdqa xmm6, xmmword ptr kABGRToV
movdqa xmm7, xmmword ptr kABGRToU
sub edi, edx // stride from u to v
convertloop:
/* step 1 - subsample 16x2 argb pixels to 8x1 */
movdqu xmm0, [eax]
movdqu xmm4, [eax + esi]
pavgb xmm0, xmm4
movdqu xmm1, [eax + 16]
movdqu xmm4, [eax + esi + 16]
pavgb xmm1, xmm4
movdqu xmm2, [eax + 32]
movdqu xmm4, [eax + esi + 32]
pavgb xmm2, xmm4
movdqu xmm3, [eax + 48]
movdqu xmm4, [eax + esi + 48]
pavgb xmm3, xmm4
lea eax, [eax + 64]
movdqa xmm4, xmm0
shufps xmm0, xmm1, 0x88
shufps xmm4, xmm1, 0xdd
pavgb xmm0, xmm4
movdqa xmm4, xmm2
shufps xmm2, xmm3, 0x88
shufps xmm4, xmm3, 0xdd
pavgb xmm2, xmm4
// step 2 - convert to U and V
// from here down is very similar to Y code except
// instead of 16 different pixels, its 8 pixels of U and 8 of V
movdqa xmm1, xmm0
movdqa xmm3, xmm2
pmaddubsw xmm0, xmm7 // U
pmaddubsw xmm2, xmm7
pmaddubsw xmm1, xmm6 // V
pmaddubsw xmm3, xmm6
phaddw xmm0, xmm2
phaddw xmm1, xmm3
psraw xmm0, 8
psraw xmm1, 8
packsswb xmm0, xmm1
paddb xmm0, xmm5 // -> unsigned
// step 3 - store 8 U and 8 V values
movlps qword ptr [edx], xmm0 // U
movhps qword ptr [edx + edi], xmm0 // V
lea edx, [edx + 8]
sub ecx, 16
jg convertloop
pop edi
pop esi
ret
}
}
__declspec(naked) void RGBAToUVRow_SSSE3(const uint8_t* src_argb0,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
__asm {
push esi
push edi
mov eax, [esp + 8 + 4] // src_argb
mov esi, [esp + 8 + 8] // src_stride_argb
mov edx, [esp + 8 + 12] // dst_u
mov edi, [esp + 8 + 16] // dst_v
mov ecx, [esp + 8 + 20] // width
movdqa xmm5, xmmword ptr kAddUV128
movdqa xmm6, xmmword ptr kRGBAToV
movdqa xmm7, xmmword ptr kRGBAToU
sub edi, edx // stride from u to v
convertloop:
/* step 1 - subsample 16x2 argb pixels to 8x1 */
movdqu xmm0, [eax]
movdqu xmm4, [eax + esi]
pavgb xmm0, xmm4
movdqu xmm1, [eax + 16]
movdqu xmm4, [eax + esi + 16]
pavgb xmm1, xmm4
movdqu xmm2, [eax + 32]
movdqu xmm4, [eax + esi + 32]
pavgb xmm2, xmm4
movdqu xmm3, [eax + 48]
movdqu xmm4, [eax + esi + 48]
pavgb xmm3, xmm4
lea eax, [eax + 64]
movdqa xmm4, xmm0
shufps xmm0, xmm1, 0x88
shufps xmm4, xmm1, 0xdd
pavgb xmm0, xmm4
movdqa xmm4, xmm2
shufps xmm2, xmm3, 0x88
shufps xmm4, xmm3, 0xdd
pavgb xmm2, xmm4
// step 2 - convert to U and V
// from here down is very similar to Y code except
// instead of 16 different pixels, its 8 pixels of U and 8 of V
movdqa xmm1, xmm0
movdqa xmm3, xmm2
pmaddubsw xmm0, xmm7 // U
pmaddubsw xmm2, xmm7
pmaddubsw xmm1, xmm6 // V
pmaddubsw xmm3, xmm6
phaddw xmm0, xmm2
phaddw xmm1, xmm3
psraw xmm0, 8
psraw xmm1, 8
packsswb xmm0, xmm1
paddb xmm0, xmm5 // -> unsigned
// step 3 - store 8 U and 8 V values
movlps qword ptr [edx], xmm0 // U
movhps qword ptr [edx + edi], xmm0 // V
lea edx, [edx + 8]
sub ecx, 16
jg convertloop
pop edi
pop esi
ret
}
}
#endif // HAS_ARGBTOYROW_SSSE3
// Read 16 UV from 444
#define READYUV444_AVX2 \
__asm { \
__asm vmovdqu xmm0, [esi] /* U */ \
__asm vmovdqu xmm1, [esi + edi] /* V */ \
__asm lea esi, [esi + 16] \
__asm vpermq ymm0, ymm0, 0xd8 \
__asm vpermq ymm1, ymm1, 0xd8 \
__asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */ \
__asm vmovdqu xmm4, [eax] /* Y */ \
__asm vpermq ymm4, ymm4, 0xd8 \
__asm vpunpcklbw ymm4, ymm4, ymm4 \
__asm lea eax, [eax + 16]}
// Read 8 UV from 422, upsample to 16 UV.
#define READYUV422_AVX2 \
__asm { \
__asm vmovq xmm0, qword ptr [esi] /* U */ \
__asm vmovq xmm1, qword ptr [esi + edi] /* V */ \
__asm lea esi, [esi + 8] \
__asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */ \
__asm vpermq ymm0, ymm0, 0xd8 \
__asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */ \
__asm vmovdqu xmm4, [eax] /* Y */ \
__asm vpermq ymm4, ymm4, 0xd8 \
__asm vpunpcklbw ymm4, ymm4, ymm4 \
__asm lea eax, [eax + 16]}
// Read 8 UV from 422, upsample to 16 UV. With 16 Alpha.
#define READYUVA422_AVX2 \
__asm { \
__asm vmovq xmm0, qword ptr [esi] /* U */ \
__asm vmovq xmm1, qword ptr [esi + edi] /* V */ \
__asm lea esi, [esi + 8] \
__asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */ \
__asm vpermq ymm0, ymm0, 0xd8 \
__asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */ \
__asm vmovdqu xmm4, [eax] /* Y */ \
__asm vpermq ymm4, ymm4, 0xd8 \
__asm vpunpcklbw ymm4, ymm4, ymm4 \
__asm lea eax, [eax + 16] \
__asm vmovdqu xmm5, [ebp] /* A */ \
__asm vpermq ymm5, ymm5, 0xd8 \
__asm lea ebp, [ebp + 16]}
// Read 8 UV from NV12, upsample to 16 UV.
#define READNV12_AVX2 \
__asm { \
__asm vmovdqu xmm0, [esi] /* UV */ \
__asm lea esi, [esi + 16] \
__asm vpermq ymm0, ymm0, 0xd8 \
__asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */ \
__asm vmovdqu xmm4, [eax] /* Y */ \
__asm vpermq ymm4, ymm4, 0xd8 \
__asm vpunpcklbw ymm4, ymm4, ymm4 \
__asm lea eax, [eax + 16]}
// Read 8 UV from NV21, upsample to 16 UV.
#define READNV21_AVX2 \
__asm { \
__asm vmovdqu xmm0, [esi] /* UV */ \
__asm lea esi, [esi + 16] \
__asm vpermq ymm0, ymm0, 0xd8 \
__asm vpshufb ymm0, ymm0, ymmword ptr kShuffleNV21 \
__asm vmovdqu xmm4, [eax] /* Y */ \
__asm vpermq ymm4, ymm4, 0xd8 \
__asm vpunpcklbw ymm4, ymm4, ymm4 \
__asm lea eax, [eax + 16]}
// Read 8 YUY2 with 16 Y and upsample 8 UV to 16 UV.
#define READYUY2_AVX2 \
__asm { \
__asm vmovdqu ymm4, [eax] /* YUY2 */ \
__asm vpshufb ymm4, ymm4, ymmword ptr kShuffleYUY2Y \
__asm vmovdqu ymm0, [eax] /* UV */ \
__asm vpshufb ymm0, ymm0, ymmword ptr kShuffleYUY2UV \
__asm lea eax, [eax + 32]}
// Read 8 UYVY with 16 Y and upsample 8 UV to 16 UV.
#define READUYVY_AVX2 \
__asm { \
__asm vmovdqu ymm4, [eax] /* UYVY */ \
__asm vpshufb ymm4, ymm4, ymmword ptr kShuffleUYVYY \
__asm vmovdqu ymm0, [eax] /* UV */ \
__asm vpshufb ymm0, ymm0, ymmword ptr kShuffleUYVYUV \
__asm lea eax, [eax + 32]}
// Convert 16 pixels: 16 UV and 16 Y.
#define YUVTORGB_AVX2(YuvConstants) \
__asm { \
__asm vpmaddubsw ymm2, ymm0, ymmword ptr [YuvConstants + KUVTOR] /* R UV */\
__asm vpmaddubsw ymm1, ymm0, ymmword ptr [YuvConstants + KUVTOG] /* G UV */\
__asm vpmaddubsw ymm0, ymm0, ymmword ptr [YuvConstants + KUVTOB] /* B UV */\
__asm vmovdqu ymm3, ymmword ptr [YuvConstants + KUVBIASR] \
__asm vpsubw ymm2, ymm3, ymm2 \
__asm vmovdqu ymm3, ymmword ptr [YuvConstants + KUVBIASG] \
__asm vpsubw ymm1, ymm3, ymm1 \
__asm vmovdqu ymm3, ymmword ptr [YuvConstants + KUVBIASB] \
__asm vpsubw ymm0, ymm3, ymm0 /* Step 2: Find Y contribution to 16 R,G,B values */ \
__asm vpmulhuw ymm4, ymm4, ymmword ptr [YuvConstants + KYTORGB] \
__asm vpaddsw ymm0, ymm0, ymm4 /* B += Y */ \
__asm vpaddsw ymm1, ymm1, ymm4 /* G += Y */ \
__asm vpaddsw ymm2, ymm2, ymm4 /* R += Y */ \
__asm vpsraw ymm0, ymm0, 6 \
__asm vpsraw ymm1, ymm1, 6 \
__asm vpsraw ymm2, ymm2, 6 \
__asm vpackuswb ymm0, ymm0, ymm0 /* B */ \
__asm vpackuswb ymm1, ymm1, ymm1 /* G */ \
__asm vpackuswb ymm2, ymm2, ymm2 /* R */ \
}
// Store 16 ARGB values.
#define STOREARGB_AVX2 \
__asm { \
__asm vpunpcklbw ymm0, ymm0, ymm1 /* BG */ \
__asm vpermq ymm0, ymm0, 0xd8 \
__asm vpunpcklbw ymm2, ymm2, ymm5 /* RA */ \
__asm vpermq ymm2, ymm2, 0xd8 \
__asm vpunpcklwd ymm1, ymm0, ymm2 /* BGRA first 8 pixels */ \
__asm vpunpckhwd ymm0, ymm0, ymm2 /* BGRA next 8 pixels */ \
__asm vmovdqu 0[edx], ymm1 \
__asm vmovdqu 32[edx], ymm0 \
__asm lea edx, [edx + 64]}
// Store 16 RGBA values.
#define STORERGBA_AVX2 \
__asm { \
__asm vpunpcklbw ymm1, ymm1, ymm2 /* GR */ \
__asm vpermq ymm1, ymm1, 0xd8 \
__asm vpunpcklbw ymm2, ymm5, ymm0 /* AB */ \
__asm vpermq ymm2, ymm2, 0xd8 \
__asm vpunpcklwd ymm0, ymm2, ymm1 /* ABGR first 8 pixels */ \
__asm vpunpckhwd ymm1, ymm2, ymm1 /* ABGR next 8 pixels */ \
__asm vmovdqu [edx], ymm0 \
__asm vmovdqu [edx + 32], ymm1 \
__asm lea edx, [edx + 64]}
#ifdef HAS_I422TOARGBROW_AVX2
// 16 pixels
// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
__declspec(naked) void I422ToARGBRow_AVX2(
const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push esi
push edi
push ebx
mov eax, [esp + 12 + 4] // Y
mov esi, [esp + 12 + 8] // U
mov edi, [esp + 12 + 12] // V
mov edx, [esp + 12 + 16] // argb
mov ebx, [esp + 12 + 20] // yuvconstants
mov ecx, [esp + 12 + 24] // width
sub edi, esi
vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha
convertloop:
READYUV422_AVX2
YUVTORGB_AVX2(ebx)
STOREARGB_AVX2
sub ecx, 16
jg convertloop
pop ebx
pop edi
pop esi
vzeroupper
ret
}
}
#endif // HAS_I422TOARGBROW_AVX2
#ifdef HAS_I422ALPHATOARGBROW_AVX2
// 16 pixels
// 8 UV values upsampled to 16 UV, mixed with 16 Y and 16 A producing 16 ARGB.
__declspec(naked) void I422AlphaToARGBRow_AVX2(
const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
const uint8_t* a_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push esi
push edi
push ebx
push ebp
mov eax, [esp + 16 + 4] // Y
mov esi, [esp + 16 + 8] // U
mov edi, [esp + 16 + 12] // V
mov ebp, [esp + 16 + 16] // A
mov edx, [esp + 16 + 20] // argb
mov ebx, [esp + 16 + 24] // yuvconstants
mov ecx, [esp + 16 + 28] // width
sub edi, esi
convertloop:
READYUVA422_AVX2
YUVTORGB_AVX2(ebx)
STOREARGB_AVX2
sub ecx, 16
jg convertloop
pop ebp
pop ebx
pop edi
pop esi
vzeroupper
ret
}
}
#endif // HAS_I422ALPHATOARGBROW_AVX2
#ifdef HAS_I444TOARGBROW_AVX2
// 16 pixels
// 16 UV values with 16 Y producing 16 ARGB (64 bytes).
__declspec(naked) void I444ToARGBRow_AVX2(
const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push esi
push edi
push ebx
mov eax, [esp + 12 + 4] // Y
mov esi, [esp + 12 + 8] // U
mov edi, [esp + 12 + 12] // V
mov edx, [esp + 12 + 16] // argb
mov ebx, [esp + 12 + 20] // yuvconstants
mov ecx, [esp + 12 + 24] // width
sub edi, esi
vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha
convertloop:
READYUV444_AVX2
YUVTORGB_AVX2(ebx)
STOREARGB_AVX2
sub ecx, 16
jg convertloop
pop ebx
pop edi
pop esi
vzeroupper
ret
}
}
#endif // HAS_I444TOARGBROW_AVX2
#ifdef HAS_NV12TOARGBROW_AVX2
// 16 pixels.
// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
__declspec(naked) void NV12ToARGBRow_AVX2(
const uint8_t* y_buf,
const uint8_t* uv_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push esi
push ebx
mov eax, [esp + 8 + 4] // Y
mov esi, [esp + 8 + 8] // UV
mov edx, [esp + 8 + 12] // argb
mov ebx, [esp + 8 + 16] // yuvconstants
mov ecx, [esp + 8 + 20] // width
vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha
convertloop:
READNV12_AVX2
YUVTORGB_AVX2(ebx)
STOREARGB_AVX2
sub ecx, 16
jg convertloop
pop ebx
pop esi
vzeroupper
ret
}
}
#endif // HAS_NV12TOARGBROW_AVX2
#ifdef HAS_NV21TOARGBROW_AVX2
// 16 pixels.
// 8 VU values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
__declspec(naked) void NV21ToARGBRow_AVX2(
const uint8_t* y_buf,
const uint8_t* vu_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push esi
push ebx
mov eax, [esp + 8 + 4] // Y
mov esi, [esp + 8 + 8] // VU
mov edx, [esp + 8 + 12] // argb
mov ebx, [esp + 8 + 16] // yuvconstants
mov ecx, [esp + 8 + 20] // width
vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha
convertloop:
READNV21_AVX2
YUVTORGB_AVX2(ebx)
STOREARGB_AVX2
sub ecx, 16
jg convertloop
pop ebx
pop esi
vzeroupper
ret
}
}
#endif // HAS_NV21TOARGBROW_AVX2
#ifdef HAS_YUY2TOARGBROW_AVX2
// 16 pixels.
// 8 YUY2 values with 16 Y and 8 UV producing 16 ARGB (64 bytes).
__declspec(naked) void YUY2ToARGBRow_AVX2(
const uint8_t* src_yuy2,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push ebx
mov eax, [esp + 4 + 4] // yuy2
mov edx, [esp + 4 + 8] // argb
mov ebx, [esp + 4 + 12] // yuvconstants
mov ecx, [esp + 4 + 16] // width
vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha
convertloop:
READYUY2_AVX2
YUVTORGB_AVX2(ebx)
STOREARGB_AVX2
sub ecx, 16
jg convertloop
pop ebx
vzeroupper
ret
}
}
#endif // HAS_YUY2TOARGBROW_AVX2
#ifdef HAS_UYVYTOARGBROW_AVX2
// 16 pixels.
// 8 UYVY values with 16 Y and 8 UV producing 16 ARGB (64 bytes).
__declspec(naked) void UYVYToARGBRow_AVX2(
const uint8_t* src_uyvy,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push ebx
mov eax, [esp + 4 + 4] // uyvy
mov edx, [esp + 4 + 8] // argb
mov ebx, [esp + 4 + 12] // yuvconstants
mov ecx, [esp + 4 + 16] // width
vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha
convertloop:
READUYVY_AVX2
YUVTORGB_AVX2(ebx)
STOREARGB_AVX2
sub ecx, 16
jg convertloop
pop ebx
vzeroupper
ret
}
}
#endif // HAS_UYVYTOARGBROW_AVX2
#ifdef HAS_I422TORGBAROW_AVX2
// 16 pixels
// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
__declspec(naked) void I422ToRGBARow_AVX2(
const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push esi
push edi
push ebx
mov eax, [esp + 12 + 4] // Y
mov esi, [esp + 12 + 8] // U
mov edi, [esp + 12 + 12] // V
mov edx, [esp + 12 + 16] // abgr
mov ebx, [esp + 12 + 20] // yuvconstants
mov ecx, [esp + 12 + 24] // width
sub edi, esi
vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha
convertloop:
READYUV422_AVX2
YUVTORGB_AVX2(ebx)
STORERGBA_AVX2
sub ecx, 16
jg convertloop
pop ebx
pop edi
pop esi
vzeroupper
ret
}
}
#endif // HAS_I422TORGBAROW_AVX2
#if defined(HAS_I422TOARGBROW_SSSE3)
// TODO(fbarchard): Read that does half size on Y and treats 420 as 444.
// Allows a conversion with half size scaling.
// Read 8 UV from 444.
#define READYUV444 \
__asm { \
__asm movq xmm0, qword ptr [esi] /* U */ \
__asm movq xmm1, qword ptr [esi + edi] /* V */ \
__asm lea esi, [esi + 8] \
__asm punpcklbw xmm0, xmm1 /* UV */ \
__asm movq xmm4, qword ptr [eax] \
__asm punpcklbw xmm4, xmm4 \
__asm lea eax, [eax + 8]}
// Read 4 UV from 422, upsample to 8 UV.
#define READYUV422 \
__asm { \
__asm movd xmm0, [esi] /* U */ \
__asm movd xmm1, [esi + edi] /* V */ \
__asm lea esi, [esi + 4] \
__asm punpcklbw xmm0, xmm1 /* UV */ \
__asm punpcklwd xmm0, xmm0 /* UVUV (upsample) */ \
__asm movq xmm4, qword ptr [eax] \
__asm punpcklbw xmm4, xmm4 \
__asm lea eax, [eax + 8]}
// Read 4 UV from 422, upsample to 8 UV. With 8 Alpha.
#define READYUVA422 \
__asm { \
__asm movd xmm0, [esi] /* U */ \
__asm movd xmm1, [esi + edi] /* V */ \
__asm lea esi, [esi + 4] \
__asm punpcklbw xmm0, xmm1 /* UV */ \
__asm punpcklwd xmm0, xmm0 /* UVUV (upsample) */ \
__asm movq xmm4, qword ptr [eax] /* Y */ \
__asm punpcklbw xmm4, xmm4 \
__asm lea eax, [eax + 8] \
__asm movq xmm5, qword ptr [ebp] /* A */ \
__asm lea ebp, [ebp + 8]}
// Read 4 UV from NV12, upsample to 8 UV.
#define READNV12 \
__asm { \
__asm movq xmm0, qword ptr [esi] /* UV */ \
__asm lea esi, [esi + 8] \
__asm punpcklwd xmm0, xmm0 /* UVUV (upsample) */ \
__asm movq xmm4, qword ptr [eax] \
__asm punpcklbw xmm4, xmm4 \
__asm lea eax, [eax + 8]}
// Read 4 VU from NV21, upsample to 8 UV.
#define READNV21 \
__asm { \
__asm movq xmm0, qword ptr [esi] /* UV */ \
__asm lea esi, [esi + 8] \
__asm pshufb xmm0, xmmword ptr kShuffleNV21 \
__asm movq xmm4, qword ptr [eax] \
__asm punpcklbw xmm4, xmm4 \
__asm lea eax, [eax + 8]}
// Read 4 YUY2 with 8 Y and upsample 4 UV to 8 UV.
#define READYUY2 \
__asm { \
__asm movdqu xmm4, [eax] /* YUY2 */ \
__asm pshufb xmm4, xmmword ptr kShuffleYUY2Y \
__asm movdqu xmm0, [eax] /* UV */ \
__asm pshufb xmm0, xmmword ptr kShuffleYUY2UV \
__asm lea eax, [eax + 16]}
// Read 4 UYVY with 8 Y and upsample 4 UV to 8 UV.
#define READUYVY \
__asm { \
__asm movdqu xmm4, [eax] /* UYVY */ \
__asm pshufb xmm4, xmmword ptr kShuffleUYVYY \
__asm movdqu xmm0, [eax] /* UV */ \
__asm pshufb xmm0, xmmword ptr kShuffleUYVYUV \
__asm lea eax, [eax + 16]}
// Convert 8 pixels: 8 UV and 8 Y.
#define YUVTORGB(YuvConstants) \
__asm { \
__asm movdqa xmm1, xmm0 \
__asm movdqa xmm2, xmm0 \
__asm movdqa xmm3, xmm0 \
__asm movdqa xmm0, xmmword ptr [YuvConstants + KUVBIASB] \
__asm pmaddubsw xmm1, xmmword ptr [YuvConstants + KUVTOB] \
__asm psubw xmm0, xmm1 \
__asm movdqa xmm1, xmmword ptr [YuvConstants + KUVBIASG] \
__asm pmaddubsw xmm2, xmmword ptr [YuvConstants + KUVTOG] \
__asm psubw xmm1, xmm2 \
__asm movdqa xmm2, xmmword ptr [YuvConstants + KUVBIASR] \
__asm pmaddubsw xmm3, xmmword ptr [YuvConstants + KUVTOR] \
__asm psubw xmm2, xmm3 \
__asm pmulhuw xmm4, xmmword ptr [YuvConstants + KYTORGB] \
__asm paddsw xmm0, xmm4 /* B += Y */ \
__asm paddsw xmm1, xmm4 /* G += Y */ \
__asm paddsw xmm2, xmm4 /* R += Y */ \
__asm psraw xmm0, 6 \
__asm psraw xmm1, 6 \
__asm psraw xmm2, 6 \
__asm packuswb xmm0, xmm0 /* B */ \
__asm packuswb xmm1, xmm1 /* G */ \
__asm packuswb xmm2, xmm2 /* R */ \
}
// Store 8 ARGB values.
#define STOREARGB \
__asm { \
__asm punpcklbw xmm0, xmm1 /* BG */ \
__asm punpcklbw xmm2, xmm5 /* RA */ \
__asm movdqa xmm1, xmm0 \
__asm punpcklwd xmm0, xmm2 /* BGRA first 4 pixels */ \
__asm punpckhwd xmm1, xmm2 /* BGRA next 4 pixels */ \
__asm movdqu 0[edx], xmm0 \
__asm movdqu 16[edx], xmm1 \
__asm lea edx, [edx + 32]}
// Store 8 BGRA values.
#define STOREBGRA \
__asm { \
__asm pcmpeqb xmm5, xmm5 /* generate 0xffffffff for alpha */ \
__asm punpcklbw xmm1, xmm0 /* GB */ \
__asm punpcklbw xmm5, xmm2 /* AR */ \
__asm movdqa xmm0, xmm5 \
__asm punpcklwd xmm5, xmm1 /* BGRA first 4 pixels */ \
__asm punpckhwd xmm0, xmm1 /* BGRA next 4 pixels */ \
__asm movdqu 0[edx], xmm5 \
__asm movdqu 16[edx], xmm0 \
__asm lea edx, [edx + 32]}
// Store 8 RGBA values.
#define STORERGBA \
__asm { \
__asm pcmpeqb xmm5, xmm5 /* generate 0xffffffff for alpha */ \
__asm punpcklbw xmm1, xmm2 /* GR */ \
__asm punpcklbw xmm5, xmm0 /* AB */ \
__asm movdqa xmm0, xmm5 \
__asm punpcklwd xmm5, xmm1 /* RGBA first 4 pixels */ \
__asm punpckhwd xmm0, xmm1 /* RGBA next 4 pixels */ \
__asm movdqu 0[edx], xmm5 \
__asm movdqu 16[edx], xmm0 \
__asm lea edx, [edx + 32]}
// Store 8 RGB24 values.
#define STORERGB24 \
__asm {/* Weave into RRGB */ \
__asm punpcklbw xmm0, xmm1 /* BG */ \
__asm punpcklbw xmm2, xmm2 /* RR */ \
__asm movdqa xmm1, xmm0 \
__asm punpcklwd xmm0, xmm2 /* BGRR first 4 pixels */ \
__asm punpckhwd xmm1, xmm2 /* BGRR next 4 pixels */ /* RRGB -> RGB24 */ \
__asm pshufb xmm0, xmm5 /* Pack first 8 and last 4 bytes. */ \
__asm pshufb xmm1, xmm6 /* Pack first 12 bytes. */ \
__asm palignr xmm1, xmm0, 12 /* last 4 bytes of xmm0 + 12 xmm1 */ \
__asm movq qword ptr 0[edx], xmm0 /* First 8 bytes */ \
__asm movdqu 8[edx], xmm1 /* Last 16 bytes */ \
__asm lea edx, [edx + 24]}
// Store 8 RGB565 values.
#define STORERGB565 \
__asm {/* Weave into RRGB */ \
__asm punpcklbw xmm0, xmm1 /* BG */ \
__asm punpcklbw xmm2, xmm2 /* RR */ \
__asm movdqa xmm1, xmm0 \
__asm punpcklwd xmm0, xmm2 /* BGRR first 4 pixels */ \
__asm punpckhwd xmm1, xmm2 /* BGRR next 4 pixels */ /* RRGB -> RGB565 */ \
__asm movdqa xmm3, xmm0 /* B first 4 pixels of argb */ \
__asm movdqa xmm2, xmm0 /* G */ \
__asm pslld xmm0, 8 /* R */ \
__asm psrld xmm3, 3 /* B */ \
__asm psrld xmm2, 5 /* G */ \
__asm psrad xmm0, 16 /* R */ \
__asm pand xmm3, xmm5 /* B */ \
__asm pand xmm2, xmm6 /* G */ \
__asm pand xmm0, xmm7 /* R */ \
__asm por xmm3, xmm2 /* BG */ \
__asm por xmm0, xmm3 /* BGR */ \
__asm movdqa xmm3, xmm1 /* B next 4 pixels of argb */ \
__asm movdqa xmm2, xmm1 /* G */ \
__asm pslld xmm1, 8 /* R */ \
__asm psrld xmm3, 3 /* B */ \
__asm psrld xmm2, 5 /* G */ \
__asm psrad xmm1, 16 /* R */ \
__asm pand xmm3, xmm5 /* B */ \
__asm pand xmm2, xmm6 /* G */ \
__asm pand xmm1, xmm7 /* R */ \
__asm por xmm3, xmm2 /* BG */ \
__asm por xmm1, xmm3 /* BGR */ \
__asm packssdw xmm0, xmm1 \
__asm movdqu 0[edx], xmm0 /* store 8 pixels of RGB565 */ \
__asm lea edx, [edx + 16]}
// 8 pixels.
// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
__declspec(naked) void I444ToARGBRow_SSSE3(
const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push esi
push edi
push ebx
mov eax, [esp + 12 + 4] // Y
mov esi, [esp + 12 + 8] // U
mov edi, [esp + 12 + 12] // V
mov edx, [esp + 12 + 16] // argb
mov ebx, [esp + 12 + 20] // yuvconstants
mov ecx, [esp + 12 + 24] // width
sub edi, esi
pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha
convertloop:
READYUV444
YUVTORGB(ebx)
STOREARGB
sub ecx, 8
jg convertloop
pop ebx
pop edi
pop esi
ret
}
}
// 8 pixels.
// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB24 (24 bytes).
__declspec(naked) void I422ToRGB24Row_SSSE3(
const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
uint8_t* dst_rgb24,
const struct YuvConstants* yuvconstants,
int width) {
__asm {
push esi
push edi
push ebx
mov eax, [esp + 12 + 4] // Y