ARGBUnattenuate_AVX2 ported to GCC. Minor cleanup of constants to use broadcast…

ARGBUnattenuate_AVX2 ported to GCC. Minor cleanup of constants to use broadcast to make 16 byte constant instead of 32 byte.
BUG=269
TESTED=try bots
R=harryjin@google.com

Review URL: https://webrtc-codereview.appspot.com/30999004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1163 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1162
+Version: 1163
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -205,6 +205,7 @@ extern "C" {
 #define HAS_ARGBSUBTRACTROW_AVX2
 #define HAS_ARGBMULTIPLYROW_AVX2
 #define HAS_ARGBATTENUATEROW_AVX2
+#define HAS_ARGBUNATTENUATEROW_AVX2
 #endif
 
 // The following are require VS2012.
@@ -218,10 +219,7 @@ extern "C" {
 #define HAS_I422TOABGRROW_AVX2
 #define HAS_INTERPOLATEROW_AVX2
 #define HAS_MIRRORROW_AVX2
-
-// Effects:
 #define HAS_ARGBMIRRORROW_AVX2
-#define HAS_ARGBUNATTENUATEROW_AVX2
 #endif  // defined(VISUALC_HAS_AVX2)
 
 // The following are Yasm x86 only:

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1162
+#define LIBYUV_VERSION 1163
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/row_posix.cc

@@ -3379,7 +3379,7 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
     "jge       40b                             \n"
     "jmp       49f                             \n"
 
-    // 4 pixel unaligned loop.
+    // 4 pixel loop.
     LABELALIGN
   "41:                                         \n"
     "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
@@ -3449,7 +3449,6 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
 
 #ifdef HAS_ARGBATTENUATEROW_SSE2
 // Attenuate 4 pixels at a time.
-// aligned to 16 bytes
 void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
   asm volatile (
     "pcmpeqb   %%xmm4,%%xmm4                   \n"
@@ -3497,14 +3496,13 @@ void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
 #ifdef HAS_ARGBATTENUATEROW_SSSE3
 // Shuffle table duplicating alpha
 static uvec8 kShuffleAlpha0 = {
-  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u
 };
 static uvec8 kShuffleAlpha1 = {
   11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
-  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u
 };
 // Attenuate 4 pixels at a time.
-// aligned to 16 bytes
 void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
   asm volatile (
     "pcmpeqb   %%xmm3,%%xmm3                   \n"
@@ -3551,9 +3549,8 @@ void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
 
 #ifdef HAS_ARGBATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
-static const ulvec8 kShuffleAlpha_AVX2 = {
-  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
-  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
+static const uvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
 };
 // Attenuate 8 pixels at a time.
 void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
@@ -3597,7 +3594,6 @@ void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
 
 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
 // Unattenuate 4 pixels at a time.
-// aligned to 16 bytes
 void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
                              int width) {
   uintptr_t alpha = 0;
@@ -3647,6 +3643,80 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
 }
 #endif  // HAS_ARGBUNATTENUATEROW_SSE2
 
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
+};
+// Unattenuate 8 pixels at a time.
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    "sub        %0,%1                          \n"
+    "vbroadcastf128 %5,%%ymm5                  \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    // replace VPGATHER
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "vpunpckldq %%xmm1,%%xmm0,%%xmm6           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
+    "movzb     " MEMACCESS2(0x13,0) ",%3       \n"
+    "vpunpckldq %%xmm3,%%xmm2,%%xmm7           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
+    "movzb     " MEMACCESS2(0x17,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
+    "movzb     " MEMACCESS2(0x1b,0) ",%3       \n"
+    "vpunpckldq %%xmm1,%%xmm0,%%xmm0           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x1f,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
+    "vpunpckldq %%xmm3,%%xmm2,%%xmm2           \n"
+    "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3          \n"
+    "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0          \n"
+    "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3     \n"
+    // end of VPGATHER
+
+    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
+    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
+    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
+    "vpunpcklwd %%ymm3,%%ymm3,%%ymm2           \n"
+    "vpunpckhwd %%ymm3,%%ymm3,%%ymm3           \n"
+    "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
+    "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    "sub        $0x8,%2                        \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8),  // %4
+    "m"(kUnattenShuffleAlpha_AVX2)  // %5
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_AVX2
+
 #ifdef HAS_ARGBGRAYROW_SSSE3
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
 void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
@@ -3841,7 +3911,6 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
 
 #ifdef HAS_ARGBQUANTIZEROW_SSE2
 // Quantize 4 ARGB pixels (16 bytes).
-// aligned to 16 bytes
 void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
                           int interval_offset, int width) {
   asm volatile (
@@ -3894,7 +3963,6 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
 
 #ifdef HAS_ARGBSHADEROW_SSE2
 // Shade 4 pixels at a time by specified value.
-// Aligned to 16 bytes.
 void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
                        uint32 value) {
   asm volatile (

source/row_win.cc

@@ -72,7 +72,6 @@ static const vec16 kUVBiasR = { BR, BR, BR, BR, BR, BR, BR, BR };
 // 64 bit
 #if defined(_M_X64)
 
-// Aligned destination version.
 __declspec(align(16))
 void I422ToARGBRow_SSSE3(const uint8* y_buf,
                          const uint8* u_buf,
@@ -165,7 +164,7 @@ static const lvec32 kPermdARGBToY_AVX = {
 // 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,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
 };
 
 // Constants for BGRA.
@@ -1845,7 +1844,7 @@ void I422ToABGRRow_AVX2(const uint8* y_buf,
     __asm packuswb   xmm2, xmm2           /* R */                              \
   }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void I444ToARGBRow_SSSE3(const uint8* y_buf,
@@ -1888,7 +1887,7 @@ void I444ToARGBRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void I422ToRGB24Row_SSSE3(const uint8* y_buf,
@@ -1935,7 +1934,7 @@ void I422ToRGB24Row_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void I422ToRAWRow_SSSE3(const uint8* y_buf,
@@ -2055,7 +2054,7 @@ void I422ToRGB565Row_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void I422ToARGBRow_SSSE3(const uint8* y_buf,
@@ -2098,7 +2097,7 @@ void I422ToARGBRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 // Similar to I420 but duplicate UV once more.
 __declspec(naked) __declspec(align(16))
@@ -2144,7 +2143,7 @@ void I411ToARGBRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void NV12ToARGBRow_SSSE3(const uint8* y_buf,
@@ -2182,7 +2181,7 @@ void NV12ToARGBRow_SSSE3(const uint8* y_buf,
   }
 }
 
-// 8 pixels, dest aligned 16.
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
 __declspec(naked) __declspec(align(16))
 void NV21ToARGBRow_SSSE3(const uint8* y_buf,
@@ -2423,8 +2422,7 @@ void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
 
 #ifdef HAS_MIRRORROW_AVX2
 // Shuffle table for reversing the bytes.
-static const ulvec8 kShuffleMirror_AVX2 = {
-  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u,
+static const uvec8 kShuffleMirror_AVX2 = {
   15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
 };
 
@@ -2434,7 +2432,7 @@ void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
     mov       eax, [esp + 4]   // src
     mov       edx, [esp + 8]   // dst
     mov       ecx, [esp + 12]  // width
-    vmovdqa   ymm5, kShuffleMirror_AVX2
+    vbroadcastf128 ymm5, kShuffleMirror_AVX2
     lea       eax, [eax - 32]
 
     align      4
@@ -3711,7 +3709,6 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
 
 #ifdef HAS_ARGBATTENUATEROW_SSE2
 // Attenuate 4 pixels at a time.
-// Aligned to 16 bytes.
 __declspec(naked) __declspec(align(16))
 void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
   __asm {
@@ -3805,8 +3802,7 @@ void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
 #ifdef HAS_ARGBATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
 static const uvec8 kShuffleAlpha_AVX2 = {
-  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
-  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
 };
 __declspec(naked) __declspec(align(16))
 void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
@@ -3846,7 +3842,6 @@ void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
 
 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
 // Unattenuate 4 pixels at a time.
-// Aligned to 16 bytes.
 __declspec(naked) __declspec(align(16))
 void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
                              int width) {
@@ -3896,7 +3891,7 @@ void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
 
 #ifdef HAS_ARGBUNATTENUATEROW_AVX2
 // Shuffle table duplicating alpha.
-static const ulvec8 kUnattenShuffleAlpha_AVX2 = {
+static const uvec8 kUnattenShuffleAlpha_AVX2 = {
   0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
 };
 // TODO(fbarchard): Enable USE_GATHER for future hardware if faster.
@@ -4185,7 +4180,6 @@ void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
 
 #ifdef HAS_ARGBQUANTIZEROW_SSE2
 // Quantize 4 ARGB pixels (16 bytes).
-// Aligned to 16 bytes.
 __declspec(naked) __declspec(align(16))
 void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
                           int interval_offset, int width) {
@@ -4232,7 +4226,6 @@ void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
 
 #ifdef HAS_ARGBSHADEROW_SSE2
 // Shade 4 pixels at a time by specified value.
-// Aligned to 16 bytes.
 __declspec(naked) __declspec(align(16))
 void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
                        uint32 value) {
@@ -4738,8 +4731,7 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
 // area is the number of pixels in the area being averaged.
 // dst points to pixel to store result to.
 // count is number of averaged pixels to produce.
-// Does 4 pixels at a time, requires CumulativeSum pointers to be 16 byte
-// aligned.
+// Does 4 pixels at a time.
 void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
                                     int width, int area, uint8* dst,
                                     int count) {