dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2025-07-16 13:59:51 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity

SSE2: yuv->rgb speed-up for point-sampling

Browse Source 
- use statically initialized tables (if WEBP_YUV_USE_SSE2_TABLES is defined)
 - use SSE2 row conversion for yuv->ARGB / RGBA / ABGR / RGB / BGR
 - clean-up and harmonize the WebpUpsamplers[] usage.

Change-Id: Ic5f3659a995927bd7363defac99c1fc03a85a47d
This commit is contained in:
skal
2014-05-22 09:49:42 +02:00
parent 1b99c09cdc
commit a05dc1402c
11 changed files with 863 additions and 161 deletions
Download Patch File Download Diff File Expand all files Collapse all files

306
src/dsp/yuv.c
View File

@ -13,7 +13,6 @@
#include "./yuv.h"
#if defined(WEBP_YUV_USE_TABLE)
static int done = 0;
@ -72,8 +71,6 @@ void VP8YUVInit(void) {}
#if defined(WEBP_USE_SSE2)
#ifdef FANCY_UPSAMPLING
#include <emmintrin.h>
#include <string.h> // for memcpy
@ -83,6 +80,14 @@ typedef union { // handy struct for converting SSE2 registers
__m128i m;
} VP8kCstSSE2;
#if defined(WEBP_YUV_USE_SSE2_TABLES)
#include "./yuv_tables_sse2.h"
void VP8YUVInitSSE2(void) {}
#else
static int done_sse2 = 0;
static VP8kCstSSE2 VP8kUtoRGBA[256], VP8kVtoRGBA[256], VP8kYtoRGBA[256];
@ -106,31 +111,66 @@ void VP8YUVInitSSE2(void) {
VP8kVtoRGBA[i].i32[3] = 0;
}
done_sse2 = 1;
#if 0 // code used to generate 'yuv_tables_sse2.h'
printf("static const VP8kCstSSE2 VP8kYtoRGBA[256] = {\n");
for (i = 0; i < 256; ++i) {
printf(" {{0x%.8x, 0x%.8x, 0x%.8x, 0x%.8x}},\n",
VP8kYtoRGBA[i].i32[0], VP8kYtoRGBA[i].i32[1],
VP8kYtoRGBA[i].i32[2], VP8kYtoRGBA[i].i32[3]);
}
printf("};\n\n");
printf("static const VP8kCstSSE2 VP8kUtoRGBA[256] = {\n");
for (i = 0; i < 256; ++i) {
printf(" {{0, 0x%.8x, 0x%.8x, 0}},\n",
VP8kUtoRGBA[i].i32[1], VP8kUtoRGBA[i].i32[2]);
}
printf("};\n\n");
printf("static VP8kCstSSE2 VP8kVtoRGBA[256] = {\n");
for (i = 0; i < 256; ++i) {
printf(" {{0x%.8x, 0x%.8x, 0, 0}},\n",
VP8kVtoRGBA[i].i32[0], VP8kVtoRGBA[i].i32[1]);
}
printf("};\n\n");
#endif
}
}
static WEBP_INLINE __m128i VP8GetRGBA32b(int y, int u, int v) {
#endif // WEBP_YUV_USE_SSE2_TABLES
//-----------------------------------------------------------------------------
static WEBP_INLINE __m128i LoadUVPart(int u, int v) {
const __m128i u_part = _mm_loadu_si128(&VP8kUtoRGBA[u].m);
const __m128i v_part = _mm_loadu_si128(&VP8kVtoRGBA[v].m);
const __m128i y_part = _mm_loadu_si128(&VP8kYtoRGBA[y].m);
const __m128i uv_part = _mm_add_epi32(u_part, v_part);
return uv_part;
}
static WEBP_INLINE __m128i GetRGBA32bWithUV(int y, const __m128i uv_part) {
const __m128i y_part = _mm_loadu_si128(&VP8kYtoRGBA[y].m);
const __m128i rgba1 = _mm_add_epi32(y_part, uv_part);
const __m128i rgba2 = _mm_srai_epi32(rgba1, YUV_FIX2);
return rgba2;
}
static WEBP_INLINE void VP8YuvToRgbSSE2(uint8_t y, uint8_t u, uint8_t v,
uint8_t* const rgb) {
const __m128i tmp0 = VP8GetRGBA32b(y, u, v);
static WEBP_INLINE __m128i GetRGBA32b(int y, int u, int v) {
const __m128i uv_part = LoadUVPart(u, v);
return GetRGBA32bWithUV(y, uv_part);
}
static WEBP_INLINE void YuvToRgbSSE2(uint8_t y, uint8_t u, uint8_t v,
uint8_t* const rgb) {
const __m128i tmp0 = GetRGBA32b(y, u, v);
const __m128i tmp1 = _mm_packs_epi32(tmp0, tmp0);
const __m128i tmp2 = _mm_packus_epi16(tmp1, tmp1);
// Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
_mm_storel_epi64((__m128i*)rgb, tmp2);
}
static WEBP_INLINE void VP8YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
uint8_t* const bgr) {
const __m128i tmp0 = VP8GetRGBA32b(y, u, v);
static WEBP_INLINE void YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
uint8_t* const bgr) {
const __m128i tmp0 = GetRGBA32b(y, u, v);
const __m128i tmp1 = _mm_shuffle_epi32(tmp0, _MM_SHUFFLE(3, 0, 1, 2));
const __m128i tmp2 = _mm_packs_epi32(tmp1, tmp1);
const __m128i tmp3 = _mm_packus_epi16(tmp2, tmp2);
@ -138,14 +178,19 @@ static WEBP_INLINE void VP8YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
_mm_storel_epi64((__m128i*)bgr, tmp3);
}
//-----------------------------------------------------------------------------
// Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
#ifdef FANCY_UPSAMPLING
void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst) {
int n;
for (n = 0; n < 32; n += 4) {
const __m128i tmp0_1 = VP8GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
const __m128i tmp0_2 = VP8GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
const __m128i tmp0_3 = VP8GetRGBA32b(y[n + 2], u[n + 2], v[n + 2]);
const __m128i tmp0_4 = VP8GetRGBA32b(y[n + 3], u[n + 3], v[n + 3]);
const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
const __m128i tmp0_3 = GetRGBA32b(y[n + 2], u[n + 2], v[n + 2]);
const __m128i tmp0_4 = GetRGBA32b(y[n + 3], u[n + 3], v[n + 3]);
const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
@ -158,8 +203,8 @@ void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst) {
int n;
for (n = 0; n < 32; n += 2) {
const __m128i tmp0_1 = VP8GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
const __m128i tmp0_2 = VP8GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
@ -175,12 +220,12 @@ void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t tmp0[2 * 3 + 5 + 15];
uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15); // align
for (n = 0; n < 30; ++n) { // we directly stomp the *dst memory
VP8YuvToRgbSSE2(y[n], u[n], v[n], dst + n * 3);
YuvToRgbSSE2(y[n], u[n], v[n], dst + n * 3);
}
// Last two pixels are special: we write in a tmp buffer before sending
// to dst.
VP8YuvToRgbSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
VP8YuvToRgbSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
YuvToRgbSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
YuvToRgbSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
memcpy(dst + n * 3, tmp, 2 * 3);
}
@ -190,18 +235,225 @@ void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t tmp0[2 * 3 + 5 + 15];
uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15); // align
for (n = 0; n < 30; ++n) {
VP8YuvToBgrSSE2(y[n], u[n], v[n], dst + n * 3);
YuvToBgrSSE2(y[n], u[n], v[n], dst + n * 3);
}
VP8YuvToBgrSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
VP8YuvToBgrSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
YuvToBgrSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
YuvToBgrSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
memcpy(dst + n * 3, tmp, 2 * 3);
}
#else
void VP8YUVInitSSE2(void) {}
#endif // FANCY_UPSAMPLING
//-----------------------------------------------------------------------------
// Arbitrary-length row conversion functions
static void YuvToRgbaRowSSE2(const uint8_t* y,
const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
int n;
for (n = 0; n + 4 <= len; n += 4) {
const __m128i uv_0 = LoadUVPart(u[0], v[0]);
const __m128i uv_1 = LoadUVPart(u[1], v[1]);
const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
const __m128i tmp0_3 = GetRGBA32bWithUV(y[2], uv_1);
const __m128i tmp0_4 = GetRGBA32bWithUV(y[3], uv_1);
const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
_mm_storeu_si128((__m128i*)dst, tmp2);
dst += 4 * 4;
y += 4;
u += 2;
v += 2;
}
// Finish off
while (n < len) {
VP8YuvToRgba(y[0], u[0], v[0], dst);
dst += 4;
++y;
u += (n & 1);
v += (n & 1);
++n;
}
}
static void YuvToBgraRowSSE2(const uint8_t* y,
const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
int n;
for (n = 0; n + 2 <= len; n += 2) {
const __m128i uv_0 = LoadUVPart(u[0], v[0]);
const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
_mm_storel_epi64((__m128i*)dst, tmp3);
dst += 4 * 2;
y += 2;
++u;
++v;
}
// Finish off
if (len & 1) {
VP8YuvToBgra(y[0], u[0], v[0], dst);
}
}
static void YuvToArgbRowSSE2(const uint8_t* y,
const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
int n;
for (n = 0; n + 2 <= len; n += 2) {
const __m128i uv_0 = LoadUVPart(u[0], v[0]);
const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(2, 1, 0, 3));
const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(2, 1, 0, 3));
const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
_mm_storel_epi64((__m128i*)dst, tmp3);
dst += 4 * 2;
y += 2;
++u;
++v;
}
// Finish off
if (len & 1) {
VP8YuvToArgb(y[0], u[0], v[0], dst);
}
}
static void YuvToRgbRowSSE2(const uint8_t* y,
const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
int n;
for (n = 0; n + 2 < len; ++n) { // we directly stomp the *dst memory
YuvToRgbSSE2(y[0], u[0], v[0], dst); // stomps 8 bytes
dst += 3;
++y;
u += (n & 1);
v += (n & 1);
}
VP8YuvToRgb(y[0], u[0], v[0], dst);
if (len > 1) {
VP8YuvToRgb(y[1], u[n & 1], v[n & 1], dst + 3);
}
}
static void YuvToBgrRowSSE2(const uint8_t* y,
const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
int n;
for (n = 0; n + 2 < len; ++n) { // we directly stomp the *dst memory
YuvToBgrSSE2(y[0], u[0], v[0], dst); // stomps 8 bytes
dst += 3;
++y;
u += (n & 1);
v += (n & 1);
}
VP8YuvToBgr(y[0], u[0], v[0], dst + 0);
if (len > 1) {
VP8YuvToBgr(y[1], u[n & 1], v[n & 1], dst + 3);
}
}
extern void WebPInitSamplersSSE2(void);
void WebPInitSamplersSSE2(void) {
WebPSamplers[MODE_RGB] = YuvToRgbRowSSE2;
WebPSamplers[MODE_RGBA] = YuvToRgbaRowSSE2;
WebPSamplers[MODE_BGR] = YuvToBgrRowSSE2;
WebPSamplers[MODE_BGRA] = YuvToBgraRowSSE2;
WebPSamplers[MODE_ARGB] = YuvToArgbRowSSE2;
}
#endif // WEBP_USE_SSE2
//-----------------------------------------------------------------------------
// Plain-C version
#define ROW_FUNC(FUNC_NAME, FUNC, XSTEP) \
static void FUNC_NAME(const uint8_t* y, \
const uint8_t* u, const uint8_t* v, \
uint8_t* dst, int len) { \
const uint8_t* const end = dst + (len & ~1) * XSTEP; \
while (dst != end) { \
FUNC(y[0], u[0], v[0], dst); \
FUNC(y[1], u[0], v[0], dst + XSTEP); \
y += 2; \
++u; \
++v; \
dst += 2 * XSTEP; \
} \
if (len & 1) { \
FUNC(y[0], u[0], v[0], dst); \
} \
} \
// All variants implemented.
ROW_FUNC(YuvToRgbRow, VP8YuvToRgb, 3)
ROW_FUNC(YuvToBgrRow, VP8YuvToBgr, 3)
ROW_FUNC(YuvToRgbaRow, VP8YuvToRgba, 4)
ROW_FUNC(YuvToBgraRow, VP8YuvToBgra, 4)
ROW_FUNC(YuvToArgbRow, VP8YuvToArgb, 4)
ROW_FUNC(YuvToRgba4444Row, VP8YuvToRgba4444, 2)
ROW_FUNC(YuvToRgb565Row, VP8YuvToRgb565, 2)
#undef ROW_FUNC
// Main call for processing a plane with a WebPSamplerRowFunc function:
void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
const uint8_t* u, const uint8_t* v, int uv_stride,
uint8_t* dst, int dst_stride,
int width, int height, WebPSamplerRowFunc func) {
int j;
for (j = 0; j < height; ++j) {
func(y, u, v, dst, width);
y += y_stride;
if (j & 1) {
u += uv_stride;
v += uv_stride;
}
dst += dst_stride;
}
}
//-----------------------------------------------------------------------------
// Main call
WebPSamplerRowFunc WebPSamplers[MODE_LAST];
extern void WebPInitSamplersSSE2(void);
extern void WebPInitSamplersMIPS32(void);
void WebPInitSamplers(void) {
WebPSamplers[MODE_RGB] = YuvToRgbRow;
WebPSamplers[MODE_RGBA] = YuvToRgbaRow;
WebPSamplers[MODE_BGR] = YuvToBgrRow;
WebPSamplers[MODE_BGRA] = YuvToBgraRow;
WebPSamplers[MODE_ARGB] = YuvToArgbRow;
WebPSamplers[MODE_RGBA_4444] = YuvToRgba4444Row;
WebPSamplers[MODE_RGB_565] = YuvToRgb565Row;
WebPSamplers[MODE_rgbA] = YuvToRgbaRow;
WebPSamplers[MODE_bgrA] = YuvToBgraRow;
WebPSamplers[MODE_Argb] = YuvToArgbRow;
WebPSamplers[MODE_rgbA_4444] = YuvToRgba4444Row;
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
#if defined(WEBP_USE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
WebPInitSamplersSSE2();
}
#endif // WEBP_USE_SSE2
#if defined(WEBP_USE_MIPS32)
if (VP8GetCPUInfo(kMIPS32)) {
WebPInitSamplersMIPS32();
}
#endif // WEBP_USE_MIPS32
}
}
//-----------------------------------------------------------------------------