Skip to content

L
libyuv
Commit 585a1261 authored by fbarchard@google.com's avatar fbarchard@google.com
Browse files
Options

rewrite ARGBToI420 with SSSE3 

TEST=talk unittests
BUG=none
Review URL: http://webrtc-codereview.appspot.com/251003

git-svn-id: http://libyuv.googlecode.com/svn/trunk@46 16f28f9a-4ce2-e073-06de-1de4eb20be90
parent 8cfa3073
Expand all Show whitespace changes
Inline Side-by-side
Showing with 396 additions and 242 deletions
include/libyuv/convert.h
View file @ 585a1261
...@@ -16,93 +16,81 @@ ...@@ -16,93 +16,81 @@
namespace libyuv { namespace libyuv {
int int I420ToRGB24(const uint8* src_y, int src_stride_y,
I420ToRGB24(const uint8* src_yplane, int src_ystride, const uint8* src_u, int src_stride_u,
const uint8* src_uplane, int src_ustride, const uint8* src_v, int src_stride_v,
const uint8* src_vplane, int src_vstride, uint8* dst_frame, int dst_stride_frame,
uint8* dst_frame, int dst_stride, int width, int height);
int src_width, int src_height);
int I420ToARGB4444(const uint8* src_y, int src_stride_y,
int const uint8* src_u, int src_stride_u,
I420ToARGB4444(const uint8* src_yplane, int src_ystride, const uint8* src_v, int src_stride_v,
const uint8* src_uplane, int src_ustride, uint8* dst_frame, int dst_stride_frame,
const uint8* src_vplane, int src_vstride, int width, int height);
uint8* dst_frame, int dst_stride,
int src_width, int src_height); int I420ToRGB565(const uint8* src_y, int src_stride_y,
const uint8* src_u, int src_stride_u,
int const uint8* src_v, int src_stride_v,
I420ToRGB565(const uint8* src_yplane, int src_ystride, uint8* dst_frame, int dst_stride_frame,
const uint8* src_uplane, int src_ustride, int width, int height);
const uint8* src_vplane, int src_vstride,
uint8* dst_frame, int dst_stride, int I420ToARGB1555(const uint8* src_y, int src_stride_y,
int src_width, int src_height); const uint8* src_u, int src_stride_u,
const uint8* src_v, int src_stride_v,
int uint8* dst_frame, int dst_stride_frame,
I420ToARGB1555(const uint8* src_yplane, int src_ystride, int width, int height);
const uint8* src_uplane, int src_ustride,
const uint8* src_vplane, int src_vstride, int I420ToYUY2(const uint8* src_y, int src_stride_y,
uint8* dst_frame, int dst_stride, const uint8* src_u, int src_stride_u,
int src_width, int src_height); const uint8* src_v, int src_stride_v,
uint8* dst_frame, int dst_stride_frame,
int int width, int height);
I420ToYUY2(const uint8* src_yplane, int src_ystride,
const uint8* src_uplane, int src_ustride, int I420ToUYVY(const uint8* src_y, int src_stride_y,
const uint8* src_vplane, int src_vstride, const uint8* src_u, int src_stride_u,
uint8* dst_frame, int dst_stride, const uint8* src_v, int src_stride_v,
int src_width, int src_height); uint8* dst_frame, int dst_stride_frame,
int width, int height);
int
I420ToUYVY(const uint8* src_yplane, int src_ystride, // TODO(fbarchard): Deprecated - this is same as BG24ToARGB with -height
const uint8* src_uplane, int src_ustride, int RGB24ToARGB(const uint8* src_frame, int src_stride_frame,
const uint8* src_vplane, int src_vstride, uint8* dst_frame, int dst_stride_frame,
uint8* dst_frame, int dst_stride, int width, int height);
int src_width, int src_height);
int RGB24ToI420(const uint8* src_frame, int src_stride_frame,
int uint8* dst_y, int dst_stride_y,
RGB24ToARGB(const uint8* src_frame, int src_stride, uint8* dst_u, int dst_stride_u,
uint8* dst_frame, int dst_stride, uint8* dst_v, int dst_stride_v,
int src_width, int src_height); int width, int height);
int int RAWToI420(const uint8* src_frame, int src_stride_frame,
RGB24ToI420(const uint8* src_frame, int src_stride, uint8* dst_y, int dst_stride_y,
uint8* dst_yplane, int dst_ystride, uint8* dst_u, int dst_stride_u,
uint8* dst_uplane, int dst_ustride, uint8* dst_v, int dst_stride_v,
uint8* dst_vplane, int dst_vstride, int width, int height);
int src_width, int src_height);
int ABGRToI420(const uint8* src_frame, int src_stride_frame,
int uint8* dst_y, int dst_stride_y,
RAWToI420(const uint8* src_frame, int src_stride, uint8* dst_u, int dst_stride_u,
uint8* dst_yplane, int dst_ystride, uint8* dst_v, int dst_stride_v,
uint8* dst_uplane, int dst_ustride, int width, int height);
uint8* dst_vplane, int dst_vstride,
int src_width, int src_height); int BGRAToI420(const uint8* src_frame, int src_stride_frame,
uint8* dst_y, int dst_stride_y,
int uint8* dst_u, int dst_stride_u,
ABGRToI420(const uint8* src_frame, int src_stride, uint8* dst_v, int dst_stride_v,
uint8* dst_yplane, int dst_ystride, int width, int height);
uint8* dst_uplane, int dst_ustride,
uint8* dst_vplane, int dst_vstride, int ARGBToI420(const uint8* src_frame, int src_stride_frame,
int src_width, int src_height); uint8* dst_y, int dst_stride_y,
uint8* dst_u, int dst_stride_u,
int uint8* dst_v, int dst_stride_v,
BGRAToI420(const uint8* src_frame, int src_stride, int width, int height);
uint8* dst_yplane, int dst_ystride,
uint8* dst_uplane, int dst_ustride, int NV12ToRGB565(const uint8* src_y, int src_stride_y,
uint8* dst_vplane, int dst_vstride, const uint8* src_uv, int src_stride_uv,
int src_width, int src_height); uint8* dst_frame, int dst_stride_frame,
int width, int height);
int
ARGBToI420(const uint8* src_frame, int src_stride,
uint8* dst_yplane, int dst_ystride,
uint8* dst_uplane, int dst_ustride,
uint8* dst_vplane, int dst_vstride,
int src_width, int src_height);
int
NV12ToRGB565(const uint8* src_yplane, int src_ystride,
const uint8* src_uvplane, int src_uvstride,
uint8* dst_frame, int dst_stride,
int src_width, int src_height);
} // namespace libyuv } // namespace libyuv
......
source/convert.cc
View file @ 585a1261
This diff is collapsed.
source/format_conversion.cc
View file @ 585a1261
...@@ -12,21 +12,10 @@ ...@@ -12,21 +12,10 @@
#include "libyuv/cpu_id.h" #include "libyuv/cpu_id.h"
#include "video_common.h" #include "video_common.h"
#include "row.h"
namespace libyuv { namespace libyuv {
// Most code in here is inspired by the material at
// http://www.siliconimaging.com/RGB%20Bayer.htm
// Forces compiler to inline, even against its better judgement. Use wisely.
#if defined(__GNUC__)
#define FORCE_INLINE __attribute__((always_inline))
#elif defined(WIN32)
#define FORCE_INLINE __forceinline
#else
#define FORCE_INLINE
#endif
// Note: to do this with Neon vld4.8 would load ARGB values into 4 registers // Note: to do this with Neon vld4.8 would load ARGB values into 4 registers
// and vst would select which 2 components to write. The low level would need // and vst would select which 2 components to write. The low level would need
// to be ARGBToBG, ARGBToGB, ARGBToRG, ARGBToGR // to be ARGBToBG, ARGBToGB, ARGBToRG, ARGBToGR
...@@ -333,46 +322,6 @@ int BayerRGBToARGB(const uint8* src_bayer, int src_stride_bayer, ...@@ -333,46 +322,6 @@ int BayerRGBToARGB(const uint8* src_bayer, int src_stride_bayer,
return 0; return 0;
} }
// Taken from http://en.wikipedia.org/wiki/YUV
static FORCE_INLINE int RGBToY(uint8 r, uint8 g, uint8 b) {
return (( 66 * r + 129 * g + 25 * b + 128) >> 8) + 16;
}
static FORCE_INLINE int RGBToU(uint8 r, uint8 g, uint8 b) {
return ((-38 * r - 74 * g + 112 * b + 128) >> 8) + 128;
}
static FORCE_INLINE int RGBToV(uint8 r, uint8 g, uint8 b) {
return ((112 * r - 94 * g - 18 * b + 128) >> 8) + 128;
}
static void ARGBtoYRow(const uint8* src_argb0,
uint8* dst_y, int width) {
for (int x = 0; x < width; ++x) {
dst_y[0] = RGBToY(src_argb0[2], src_argb0[1], src_argb0[0]);
src_argb0 += 4;
dst_y += 1;
}
}
static void ARGBtoUVRow(const uint8* src_argb0, int src_stride_argb,
uint8* dst_u,
uint8* dst_v,
int width) {
const uint8* src_argb1 = src_argb0 + src_stride_argb;
for (int x = 0; x < width - 1; x += 2) {
uint8 ab = (src_argb0[0] + src_argb0[4] + src_argb1[0] + src_argb1[4]) >> 2;
uint8 ag = (src_argb0[1] + src_argb0[5] + src_argb1[1] + src_argb1[5]) >> 2;
uint8 ar = (src_argb0[2] + src_argb0[6] + src_argb1[2] + src_argb1[6]) >> 2;
dst_u[0] = RGBToU(ar, ag, ab);
dst_v[0] = RGBToV(ar, ag, ab);
src_argb0 += 8;
src_argb1 += 8;
dst_u += 1;
dst_v += 1;
}
}
// Converts any Bayer RGB format to ARGB. // Converts any Bayer RGB format to ARGB.
int BayerRGBToI420(const uint8* src_bayer, int src_stride_bayer, int BayerRGBToI420(const uint8* src_bayer, int src_stride_bayer,
uint32 src_fourcc_bayer, uint32 src_fourcc_bayer,
...@@ -395,6 +344,28 @@ int BayerRGBToI420(const uint8* src_bayer, int src_stride_bayer, ...@@ -395,6 +344,28 @@ int BayerRGBToI420(const uint8* src_bayer, int src_stride_bayer,
uint8* dst_rgb, int pix); uint8* dst_rgb, int pix);
void (*BayerRow1)(const uint8* src_bayer, int src_stride_bayer, void (*BayerRow1)(const uint8* src_bayer, int src_stride_bayer,
uint8* dst_rgb, int pix); uint8* dst_rgb, int pix);
void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix);
void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
uint8* dst_u, uint8* dst_v, int width);
#define kMaxStride (2048 * 4)
SIMD_ALIGNED(uint8 row[kMaxStride * 2]);
#if defined(HAS_ARGBTOYROW_SSSE3)
if (libyuv::TestCpuFlag(libyuv::kCpuHasSSSE3) &&
(width % 8 == 0) &&
IS_ALIGNED(row, 16) && (kMaxStride % 16 == 0) &&
IS_ALIGNED(dst_y, 8) && (dst_stride_y % 8 == 0)) {
ARGBToYRow = ARGBToYRow_SSSE3;
#if defined(HAS_ARGBTOUVROW_SSSE3)
ARGBToUVRow = ARGBToUVRow_SSSE3;
#else
ARGBToUVRow = ARGBToUVRow_C;
#endif
} else
#endif
{
ARGBToYRow = ARGBToYRow_C;
ARGBToUVRow = ARGBToUVRow_C;
}
switch (src_fourcc_bayer) { switch (src_fourcc_bayer) {
default: default:
...@@ -417,24 +388,23 @@ int BayerRGBToI420(const uint8* src_bayer, int src_stride_bayer, ...@@ -417,24 +388,23 @@ int BayerRGBToI420(const uint8* src_bayer, int src_stride_bayer,
break; break;
} }
#define kMaxStride 2048 * 4
uint8 row[kMaxStride * 2];
for (int y = 0; y < (height - 1); y += 2) { for (int y = 0; y < (height - 1); y += 2) {
BayerRow0(src_bayer, src_stride_bayer, row, width); BayerRow0(src_bayer, src_stride_bayer, row, width);
BayerRow1(src_bayer + src_stride_bayer, -src_stride_bayer, BayerRow1(src_bayer + src_stride_bayer, -src_stride_bayer,
row + kMaxStride, width); row + kMaxStride, width);
ARGBtoYRow(row, dst_y, width); ARGBToYRow(row, dst_y, width);
ARGBtoYRow(row + kMaxStride, dst_y + dst_stride_y, width); ARGBToYRow(row + kMaxStride, dst_y + dst_stride_y, width);
ARGBtoUVRow(row, kMaxStride, dst_u, dst_v, width); ARGBToUVRow(row, kMaxStride, dst_u, dst_v, width);
src_bayer += src_stride_bayer * 2; src_bayer += src_stride_bayer * 2;
dst_y += dst_stride_y * 2; dst_y += dst_stride_y * 2;
dst_u += dst_stride_u; dst_u += dst_stride_u;
dst_v += dst_stride_v; dst_v += dst_stride_v;
} }
// TODO(fbarchard): Make sure this filters properly
if (height & 1) { if (height & 1) {
BayerRow0(src_bayer, src_stride_bayer, row, width); BayerRow0(src_bayer, src_stride_bayer, row, width);
ARGBtoYRow(row, dst_y, width); ARGBToYRow(row, dst_y, width);
ARGBtoUVRow(row, 0, dst_u, dst_v, width); ARGBToUVRow(row, 0, dst_u, dst_v, width);
} }
return 0; return 0;
} }
......
source/planar_functions.cc
View file @ 585a1261
...@@ -68,15 +68,6 @@ extern "C" TALIGN16(const uint8, kShuffleMaskRAWToARGB[16]) = { ...@@ -68,15 +68,6 @@ extern "C" TALIGN16(const uint8, kShuffleMaskRAWToARGB[16]) = {
2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u 2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
}; };
// Constant multiplication table for converting ARGB to I400.
extern "C" TALIGN16(const uint8, kMultiplyMaskARGBToI400[16]) = {
13u, 64u, 33u, 0u, 13u, 64u, 33u, 0u, 13u, 64u, 33u, 0u, 13u, 64u, 33u, 0u
};
extern "C" TALIGN16(const uint8, kMultiplyMaskARGBToI400_2[16]) = {
1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u
};
#if defined(WIN32) && !defined(COVERAGE_ENABLED) #if defined(WIN32) && !defined(COVERAGE_ENABLED)
#define HAS_SPLITUV_SSE2 #define HAS_SPLITUV_SSE2
__declspec(naked) __declspec(naked)
...@@ -215,7 +206,7 @@ int I420Copy(const uint8* src_y, int src_stride_y, ...@@ -215,7 +206,7 @@ int I420Copy(const uint8* src_y, int src_stride_y,
static void SetRow32_NEON(uint8* dst, uint32 v32, int count) { static void SetRow32_NEON(uint8* dst, uint32 v32, int count) {
__asm__ volatile __asm__ volatile
( (
"vdup.u32 q0, %2 \n" // duplicate 4 ints "vdup.u32 {q0}, %2 \n" // duplicate 4 ints
"1:\n" "1:\n"
"vst1.u32 {q0}, [%0]! \n" // store "vst1.u32 {q0}, [%0]! \n" // store
"subs %1, %1, #16 \n" // 16 processed per loop "subs %1, %1, #16 \n" // 16 processed per loop
...@@ -393,16 +384,16 @@ int I422ToI420(const uint8* src_y, int src_stride_y, ...@@ -393,16 +384,16 @@ int I422ToI420(const uint8* src_y, int src_stride_y,
} }
static void I420CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1, static void I420CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
uint8* dst, int dst_stride, uint8* dst, int dst_stride_frame,
int width, int height) { int width, int height) {
// Copy plane // Copy plane
for (int y = 0; y < height; y += 2) { for (int y = 0; y < height; y += 2) {
memcpy(dst, src, width); memcpy(dst, src, width);
src += src_stride_0; src += src_stride_0;
dst += dst_stride; dst += dst_stride_frame;
memcpy(dst, src, width); memcpy(dst, src, width);
src += src_stride_1; src += src_stride_1;
dst += dst_stride; dst += dst_stride_frame;
} }
} }
...@@ -503,13 +494,13 @@ int NV12ToI420(const uint8* src_y, int src_stride_y, ...@@ -503,13 +494,13 @@ int NV12ToI420(const uint8* src_y, int src_stride_y,
// Convert NV12 to I420. Deprecated. // Convert NV12 to I420. Deprecated.
int NV12ToI420(const uint8* src_y, int NV12ToI420(const uint8* src_y,
const uint8* src_uv, const uint8* src_uv,
int src_stride, int src_stride_frame,
uint8* dst_y, int dst_stride_y, uint8* dst_y, int dst_stride_y,
uint8* dst_u, int dst_stride_u, uint8* dst_u, int dst_stride_u,
uint8* dst_v, int dst_stride_v, uint8* dst_v, int dst_stride_v,
int width, int height) { int width, int height) {
return X420ToI420(src_y, src_stride, src_stride, return X420ToI420(src_y, src_stride_frame, src_stride_frame,
src_uv, src_stride, src_uv, src_stride_frame,
dst_y, dst_stride_y, dst_y, dst_stride_y,
dst_u, dst_stride_u, dst_u, dst_stride_u,
dst_v, dst_stride_v, dst_v, dst_stride_v,
...@@ -1371,38 +1362,6 @@ __asm { ...@@ -1371,38 +1362,6 @@ __asm {
} }
} }
#define HAS_ARGBTOI400ROW_SSSE3
__declspec(naked)
static void ARGBToI400Row_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_y
mov ecx, [esp + 12] // pix
movdqa xmm7, _kMultiplyMaskARGBToI400
movdqa xmm6, _kMultiplyMaskARGBToI400_2
movdqa xmm5, xmm6
psllw xmm5, 4 // Generate a mask of 0x10 on each byte.
convertloop :
movdqa xmm0, [eax]
pmaddubsw xmm0, xmm7
movdqa xmm1, [eax + 16]
psrlw xmm0, 7
pmaddubsw xmm1, xmm7
lea eax, [eax + 32]
psrlw xmm1, 7
packuswb xmm0, xmm1
pmaddubsw xmm0, xmm6
packuswb xmm0, xmm0
paddb xmm0, xmm5
movq qword ptr [edx], xmm0
lea edx, [edx + 8]
sub ecx, 8
ja convertloop
ret
}
}
#elif (defined(__x86_64__) || defined(__i386__)) && \ #elif (defined(__x86_64__) || defined(__i386__)) && \
!defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR) !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
...@@ -1554,39 +1513,6 @@ static void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, ...@@ -1554,39 +1513,6 @@ static void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
); );
} }
#define HAS_ARGBTOI400ROW_SSSE3
static void ARGBToI400Row_SSSE3(const uint8* src_argb, uint8* dst_y,
int pix) {
asm volatile(
"movdqa (%3),%%xmm7\n"
"movdqa (%4),%%xmm6\n"
"movdqa %%xmm6,%%xmm5\n"
"psllw $0x4,%%xmm5\n" // Generate a mask of 0x10 on each byte.
"1:"
"movdqa (%0),%%xmm0\n"
"pmaddubsw %%xmm7,%%xmm0\n"
"movdqa 0x10(%0),%%xmm1\n"
"psrlw $0x7,%%xmm0\n"
"pmaddubsw %%xmm7,%%xmm1\n"
"lea 0x20(%0),%0\n"
"psrlw $0x7,%%xmm1\n"
"packuswb %%xmm1,%%xmm0\n"
"pmaddubsw %%xmm6,%%xmm0\n"
"packuswb %%xmm0,%%xmm0\n"
"paddb %%xmm5,%%xmm0\n"
"movq %%xmm0,(%1)\n"
"lea 0x8(%1),%1\n"
"sub $0x8,%2\n"
"ja 1b\n"
: "+r"(src_argb), // %0
"+r"(dst_y), // %1
"+r"(pix) // %2
: "r"(kMultiplyMaskARGBToI400), // %3
"r"(kMultiplyMaskARGBToI400_2) // %4
: "memory"
);
}
#endif #endif
static void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix) { static void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix) {
...@@ -1812,16 +1738,6 @@ int BGRAToARGB(const uint8* src_bgra, int src_stride_bgra, ...@@ -1812,16 +1738,6 @@ int BGRAToARGB(const uint8* src_bgra, int src_stride_bgra,
return 0; return 0;
} }
static void ARGBToI400Row_C(const uint8* src_argb, uint8* dst_y, int pix) {
for (int x = 0; x < pix; ++x) {
uint32 b = static_cast<uint32>(src_argb[0] * 13u);
uint32 g = static_cast<uint32>(src_argb[1] * 64u);
uint32 r = static_cast<uint32>(src_argb[2] * 33u);
*(dst_y++) = static_cast<uint8>(((b + g + r) >> 7) + 16u);
src_argb += 4;
}
}
// Convert ARGB to I400. // Convert ARGB to I400.
int ARGBToI400(const uint8* src_argb, int src_stride_argb, int ARGBToI400(const uint8* src_argb, int src_stride_argb,
uint8* dst_y, int dst_stride_y, uint8* dst_y, int dst_stride_y,
...@@ -1831,21 +1747,21 @@ int ARGBToI400(const uint8* src_argb, int src_stride_argb, ...@@ -1831,21 +1747,21 @@ int ARGBToI400(const uint8* src_argb, int src_stride_argb,
src_argb = src_argb + (height - 1) * src_stride_argb; src_argb = src_argb + (height - 1) * src_stride_argb;
src_stride_argb = -src_stride_argb; src_stride_argb = -src_stride_argb;
} }
void (*ARGBToI400Row)(const uint8* src_argb, uint8* dst_y, int pix); void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix);
#if defined(HAS_ARGBTOI400ROW_SSSE3) #if defined(HAS_ARGBTOYROW_SSSE3)
if (libyuv::TestCpuFlag(libyuv::kCpuHasSSSE3) && if (libyuv::TestCpuFlag(libyuv::kCpuHasSSSE3) &&
(width % 4 == 0) && (width % 4 == 0) &&
IS_ALIGNED(src_argb, 16) && (src_stride_argb % 16 == 0) && IS_ALIGNED(src_argb, 16) && (src_stride_argb % 16 == 0) &&
IS_ALIGNED(dst_y, 16) && (dst_stride_y % 16 == 0)) { IS_ALIGNED(dst_y, 16) && (dst_stride_y % 16 == 0)) {
ARGBToI400Row = ARGBToI400Row_SSSE3; ARGBToYRow = ARGBToYRow_SSSE3;
} else } else
#endif #endif
{ {
ARGBToI400Row = ARGBToI400Row_C; ARGBToYRow = ARGBToYRow_C;
} }
for (int y = 0; y < height; ++y) { for (int y = 0; y < height; ++y) {
ARGBToI400Row(src_argb, dst_y, width); ARGBToYRow(src_argb, dst_y, width);
src_argb += src_stride_argb; src_argb += src_stride_argb;
dst_y += dst_stride_y; dst_y += dst_stride_y;
} }
......
source/row.h
View file @ 585a1261
...@@ -13,6 +13,16 @@ ...@@ -13,6 +13,16 @@
#include "libyuv/basic_types.h" #include "libyuv/basic_types.h"
#if (defined(WIN32) || defined(__x86_64__) || defined(__i386__)) \
&& !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
#define HAS_ARGBTOYROW_SSSE3
#endif
#if defined(WIN32) \
&& !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
#define HAS_ARGBTOUVROW_SSSE3
#endif
extern "C" { extern "C" {
void FastConvertYUVToRGB32Row(const uint8* y_buf, void FastConvertYUVToRGB32Row(const uint8* y_buf,
const uint8* u_buf, const uint8* u_buf,
...@@ -42,11 +52,24 @@ void FastConvertYToRGB32Row(const uint8* y_buf, ...@@ -42,11 +52,24 @@ void FastConvertYToRGB32Row(const uint8* y_buf,
uint8* rgb_buf, uint8* rgb_buf,
int width); int width);
#ifdef HAS_ARGBTOYROW_SSSE3
void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
uint8* dst_u, uint8* dst_v, int width);
#endif
void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
void ARGBToUVRow_C(const uint8* src_argb0, int src_stride_argb,
uint8* dst_u, uint8* dst_v, int width);
#if defined(_MSC_VER) #if defined(_MSC_VER)
#define SIMD_ALIGNED(var) __declspec(align(16)) var #define SIMD_ALIGNED(var) __declspec(align(16)) var
#define TALIGN16(t, var) static __declspec(align(16)) t _ ## var
#else #else
#define SIMD_ALIGNED(var) var __attribute__((aligned(16))) #define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
#define TALIGN16(t, var) t var __attribute__((aligned(16)))
#endif #endif
#ifdef OSX #ifdef OSX
extern SIMD_ALIGNED(const int16 kCoefficientsRgbY[768][4]); extern SIMD_ALIGNED(const int16 kCoefficientsRgbY[768][4]);
extern SIMD_ALIGNED(const int16 kCoefficientsBgraY[768][4]); extern SIMD_ALIGNED(const int16 kCoefficientsBgraY[768][4]);
......
source/row_posix.cc
View file @ 585a1261
...@@ -12,6 +12,91 @@ ...@@ -12,6 +12,91 @@
extern "C" { extern "C" {
#ifdef HAS_ARGBTOYROW_SSSE3
// Constant multiplication table for converting ARGB to I400.
extern "C" TALIGN16(const uint8, kMultiplyMaskARGBToI400[16]) = {
13u, 64u, 33u, 0u, 13u, 64u, 33u, 0u, 13u, 64u, 33u, 0u, 13u, 64u, 33u, 0u
};
extern "C" TALIGN16(const uint8, kAdd16[16]) = {
1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u, 1u
};
void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
asm volatile(
"movdqa (%3),%%xmm7\n"
"movdqa (%4),%%xmm6\n"
"movdqa %%xmm6,%%xmm5\n"
"psllw $0x4,%%xmm5\n" // Generate a mask of 0x10 on each byte.
"1:"
"movdqa (%0),%%xmm0\n"
"pmaddubsw %%xmm7,%%xmm0\n"
"movdqa 0x10(%0),%%xmm1\n"
"psrlw $0x7,%%xmm0\n"
"pmaddubsw %%xmm7,%%xmm1\n"
"lea 0x20(%0),%0\n"
"psrlw $0x7,%%xmm1\n"
"packuswb %%xmm1,%%xmm0\n"
"pmaddubsw %%xmm6,%%xmm0\n"
"packuswb %%xmm0,%%xmm0\n"
"paddb %%xmm5,%%xmm0\n"
"movq %%xmm0,(%1)\n"
"lea 0x8(%1),%1\n"
"sub $0x8,%2\n"
"ja 1b\n"
: "+r"(src_argb), // %0
"+r"(dst_y), // %1
"+r"(pix) // %2
: "r"(kMultiplyMaskARGBToI400), // %3
"r"(kAdd16) // %4
: "memory"
);
}
#endif
static inline int RGBToY(uint8 r, uint8 g, uint8 b) {
return (( 66 * r + 129 * g + 25 * b + 128) >> 8) + 16;
}
static inline int RGBToU(uint8 r, uint8 g, uint8 b) {
return ((-38 * r - 74 * g + 112 * b + 128) >> 8) + 128;
}
static inline int RGBToV(uint8 r, uint8 g, uint8 b) {
return ((112 * r - 94 * g - 18 * b + 128) >> 8) + 128;
}
void ARGBToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) {
for (int x = 0; x < width; ++x) {
dst_y[0] = RGBToY(src_argb0[2], src_argb0[1], src_argb0[0]);
src_argb0 += 4;
dst_y += 1;
}
}
void ARGBToUVRow_C(const uint8* src_argb0, int src_stride_argb,
uint8* dst_u, uint8* dst_v, int width) {
const uint8* src_argb1 = src_argb0 + src_stride_argb;
for (int x = 0; x < width - 1; x += 2) {
uint8 ab = (src_argb0[0] + src_argb0[4] + src_argb1[0] + src_argb1[4]) >> 2;
uint8 ag = (src_argb0[1] + src_argb0[5] + src_argb1[1] + src_argb1[5]) >> 2;
uint8 ar = (src_argb0[2] + src_argb0[6] + src_argb1[2] + src_argb1[6]) >> 2;
dst_u[0] = RGBToU(ar, ag, ab);
dst_v[0] = RGBToV(ar, ag, ab);
src_argb0 += 8;
src_argb1 += 8;
dst_u += 1;
dst_v += 1;
}
if (width & 1) {
uint8 ab = (src_argb0[0] + src_argb1[0]) >> 1;
uint8 ag = (src_argb0[1] + src_argb1[1]) >> 1;
uint8 ar = (src_argb0[2] + src_argb1[2]) >> 1;
dst_u[0] = RGBToU(ar, ag, ab);
dst_v[0] = RGBToV(ar, ag, ab);
}
}
#if defined(__x86_64__) #if defined(__x86_64__)
// 64 bit linux gcc version // 64 bit linux gcc version
......
source/row_win.cc
View file @ 585a1261
...@@ -12,6 +12,176 @@ ...@@ -12,6 +12,176 @@
extern "C" { extern "C" {
#ifdef HAS_ARGBTOYROW_SSSE3
#define TALIGN16(t, var) static __declspec(align(16)) t _ ## var
// Constant multiplication table for converting ARGB to I400.
extern "C" TALIGN16(const int8, kRGBToY[16]) = {
13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
};
extern "C" TALIGN16(const int8, kRGBToU[16]) = {
112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
};
extern "C" TALIGN16(const int8, kRGBToV[16]) = {
-18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
};
extern "C" TALIGN16(const uint8, kAddY16[16]) = {
16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u,
16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u,
};
extern "C" TALIGN16(const uint8, kAddUV128[16]) = {
128u, 0u, 128u, 0u, 128u, 0u, 128u, 0u,
128u, 0u, 128u, 0u, 128u, 0u, 128u, 0u
};
__declspec(naked)
void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
__asm {
mov eax, [esp + 4] // src_argb
mov edx, [esp + 8] // dst_y
mov ecx, [esp + 12] // pix
movdqa xmm7, _kRGBToY
movdqa xmm6, _kAddY16
pcmpeqb xmm5, xmm5 // Generate mask 0x0000ffff
psrld xmm5, 16
convertloop :
movdqa xmm0, [eax]
movdqa xmm1, [eax + 16]
pmaddubsw xmm0, xmm7
lea eax, [eax + 32]
pmaddubsw xmm1, xmm7 // BG ra BG ra BG ra BG ra
palignr xmm2, xmm0, 2 // AR xx AR xx AR xx AR xx
paddw xmm2, xmm0 // BGRA xx BGRA xx BGRA xx BGRA xx
pand xmm2, xmm5 // BGRA 00 BGRA 00 BGRA 00 BGRA 00
palignr xmm3, xmm1, 2
paddw xmm3, xmm1
pand xmm3, xmm5 // BGRA 00 BGRA 00 BGRA 00 BGRA 00
packssdw xmm2, xmm3 // BGRA BGRA BGRA BGRA BGRA BGRA BGRA BGRA
psrlw xmm2, 7 // 0B xx 0B xx 0B xx 0B xx
packuswb xmm2, xmm2
paddb xmm2, xmm6
movq qword ptr [edx], xmm2
lea edx, [edx + 8]
sub ecx, 8
ja convertloop
ret
}
}
__declspec(naked)
void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
uint8* dst_u, uint8* 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] // pix
movdqa xmm7, _kRGBToU
movdqa xmm6, _kRGBToV
movdqa xmm5, _kAddUV128
pcmpeqb xmm4, xmm4 // Generate mask 0x0000ffff
psrld xmm4, 16
convertloop :
// step 1 - subsample 8x2 argb pixels to 4x1
movdqa xmm0, [eax] // 32x2 -> 32x1
movdqa xmm1, [eax + 16]
movdqa xmm2, [eax + esi]
movdqa xmm3, [eax + esi + 16]
lea eax, [eax + 32]
pavgb xmm0, xmm2
pavgb xmm1, xmm3
movdqa xmm2, xmm0 // 32x1 -> 16x1
shufps xmm0, xmm1, 0x88
shufps xmm2, xmm1, 0xdd
pavgb xmm0, xmm2
// step 2 - convert to U and V
// from here down is very similar to Y code except
// instead of 8 different pixels, its 4 pixels of U and 4 of V
movdqa xmm1, xmm0
pmaddubsw xmm0, xmm7 // U
pmaddubsw xmm1, xmm6 // V
palignr xmm2, xmm0, 2 // AR xx AR xx AR xx AR xx
paddw xmm2, xmm0 // BGRA xx BGRA xx BGRA xx BGRA xx
pand xmm2, xmm4 // BGRA 00 BGRA 00 BGRA 00 BGRA 00
palignr xmm3, xmm1, 2
paddw xmm3, xmm1
pand xmm3, xmm4 // BGRA 00 BGRA 00 BGRA 00 BGRA 00
psraw xmm2, 8
psraw xmm3, 8
packsswb xmm2, xmm3 // BGRA BGRA BGRA BGRA BGRA BGRA BGRA BGRA
paddb xmm2, xmm5 // -> unsigned
packuswb xmm2, xmm2 // 8 bytes. 4 U, 4 V
// step 3 - store 4 U and 4 V values
movd dword ptr [edx], xmm2 // U
lea edx, [edx + 4]
pshufd xmm0, xmm2, 0x55 // V
movd dword ptr [edi], xmm0
lea edi, [edi + 4]
sub ecx, 8
ja convertloop
pop edi
pop esi
ret
}
}
static inline int RGBToY(uint8 r, uint8 g, uint8 b) {
return (( 66 * r + 129 * g + 25 * b + 128) >> 8) + 16;
}
static inline int RGBToU(uint8 r, uint8 g, uint8 b) {
return ((-38 * r - 74 * g + 112 * b + 128) >> 8) + 128;
}
static inline int RGBToV(uint8 r, uint8 g, uint8 b) {
return ((112 * r - 94 * g - 18 * b + 128) >> 8) + 128;
}
void ARGBToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) {
for (int x = 0; x < width; ++x) {
dst_y[0] = RGBToY(src_argb0[2], src_argb0[1], src_argb0[0]);
src_argb0 += 4;
dst_y += 1;
}
}
void ARGBToUVRow_C(const uint8* src_argb0, int src_stride_argb,
uint8* dst_u, uint8* dst_v, int width) {
const uint8* src_argb1 = src_argb0 + src_stride_argb;
for (int x = 0; x < width - 1; x += 2) {
uint8 ab = (src_argb0[0] + src_argb0[4] + src_argb1[0] + src_argb1[4]) >> 2;
uint8 ag = (src_argb0[1] + src_argb0[5] + src_argb1[1] + src_argb1[5]) >> 2;
uint8 ar = (src_argb0[2] + src_argb0[6] + src_argb1[2] + src_argb1[6]) >> 2;
dst_u[0] = RGBToU(ar, ag, ab);
dst_v[0] = RGBToV(ar, ag, ab);
src_argb0 += 8;
src_argb1 += 8;
dst_u += 1;
dst_v += 1;
}
if (width & 1) {
uint8 ab = (src_argb0[0] + src_argb1[0]) >> 1;
uint8 ag = (src_argb0[1] + src_argb1[1]) >> 1;
uint8 ar = (src_argb0[2] + src_argb1[2]) >> 1;
dst_u[0] = RGBToU(ar, ag, ab);
dst_v[0] = RGBToV(ar, ag, ab);
}
}
__declspec(naked) __declspec(naked)
void FastConvertYUVToRGB32Row(const uint8* y_buf, void FastConvertYUVToRGB32Row(const uint8* y_buf,
const uint8* u_buf, const uint8* u_buf,
...@@ -200,4 +370,6 @@ void FastConvertYToRGB32Row(const uint8* y_buf, ...@@ -200,4 +370,6 @@ void FastConvertYToRGB32Row(const uint8* y_buf,
} }
} }
#endif
} // extern "C" } // extern "C"
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment