I420ToARGB for Haswell.

BUG=216
TEST=I420ToARGB
Review URL: https://webrtc-codereview.appspot.com/1314004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@660 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 659
+Version: 660
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -133,12 +133,12 @@ extern "C" {
 #define HAS_ARGBCOLORTABLEROW_X86
 // Visual C 2012 required for AVX2.
 #if _MSC_VER >= 1700
-// TODO(fbarchard): Hook these up to all functions. e.g. format conversion.
 #define HAS_ARGBSHUFFLEROW_AVX2
 #define HAS_ARGBTOUVROW_AVX2
-#define HAS_ARGBTOYROW_AVX2
 #define HAS_ARGBTOYJROW_AVX2
+#define HAS_ARGBTOYROW_AVX2
 #define HAS_HALFROW_AVX2
+#define HAS_I422TOARGBROW_AVX2
 #define HAS_MERGEUVROW_AVX2
 #define HAS_MIRRORROW_AVX2
 #define HAS_SPLITUVROW_AVX2
@@ -826,7 +826,11 @@ void I422ToRGB565Row_C(const uint8* src_y,
 void YToARGBRow_C(const uint8* src_y,
                   uint8* dst_argb,
                   int width);
-
+void I422ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
 void I444ToARGBRow_SSSE3(const uint8* src_y,
                          const uint8* src_u,
                          const uint8* src_v,
@@ -950,6 +954,11 @@ void I422ToRGBARow_Unaligned_SSSE3(const uint8* src_y,
                                    const uint8* src_v,
                                    uint8* dst_rgba,
                                    int width);
+void I422ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
 void I444ToARGBRow_Any_SSSE3(const uint8* src_y,
                              const uint8* src_u,
                              const uint8* src_v,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 659
+#define LIBYUV_VERSION 660
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/convert_argb.cc

@@ -152,14 +152,24 @@ int I422ToARGB(const uint8* src_y, int src_stride_y,
       }
     }
   }
-#elif defined(HAS_I422TOARGBROW_NEON)
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGBRow = I422ToARGBRow_NEON;
     }
   }
-#elif defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
       IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
       IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&

source/convert_from.cc

@@ -557,14 +557,24 @@ int I420ToARGB(const uint8* src_y, int src_stride_y,
       }
     }
   }
-#elif defined(HAS_I422TOARGBROW_NEON)
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGBRow = I422ToARGBRow_NEON;
     }
   }
-#elif defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
       IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
       IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&

source/format_conversion.cc

@@ -410,14 +410,24 @@ int I420ToBayer(const uint8* src_y, int src_stride_y,
       I422ToARGBRow = I422ToARGBRow_SSSE3;
     }
   }
-#elif defined(HAS_I422TOARGBROW_NEON)
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       I422ToARGBRow = I422ToARGBRow_NEON;
     }
   }
-#elif defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
       IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
       IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&

source/row_any.cc

@@ -60,6 +60,9 @@ YANY(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, I422ToRAWRow_C, 1, 3, 7)
 YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15)
 YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
 #endif  // HAS_I422TOARGBROW_SSSE3
+#ifdef HAS_I422TOARGBROW_AVX2
+YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
+#endif  // HAS_I422TOARGBROW_AVX2
 #ifdef HAS_I422TOARGBROW_NEON
 YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7)
 YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7)

source/row_win.cc

@@ -2068,8 +2068,6 @@ void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
 }
 #endif  // HAS_ARGBTOYROW_SSSE3
 
-#ifdef HAS_I422TOARGBROW_SSSE3
-
 #define YG 74 /* static_cast<int8>(1.164 * 64 + 0.5) */
 
 #define UB 127 /* min(63,static_cast<int8>(2.018 * 64)) */
@@ -2085,6 +2083,104 @@ void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
 #define BG UG * 128 + VG * 128
 #define BR UR * 128 + VR * 128
 
+#ifdef HAS_I422TOARGBROW_AVX2
+
+static const lvec8 kUVToB_AVX = {
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB,
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
+};
+
+static const lvec8 kUVToR_AVX = {
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR,
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
+};
+
+static const lvec8 kUVToG_AVX = {
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
+};
+
+static const lvec16 kYToRgb_AVX = { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG };
+static const lvec16 kYSub16_AVX = { 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16 };
+static const lvec16 kUVBiasB_AVX = { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB};
+static const lvec16 kUVBiasG_AVX = { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG};
+static const lvec16 kUVBiasR_AVX = { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR};
+
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_AVX2(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpxor      ymm4, ymm4, ymm4
+
+    align      16
+ convertloop:
+    vmovq      xmm0, qword ptr [esi]          //  U
+    vmovq      xmm1, qword ptr [esi + edi]    //  V
+    lea        esi,  [esi + 8]                                           
+    vpunpcklbw ymm0, ymm0, ymm1               // UV                              
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm0, ymm0, ymm0              // UVUV
+    vpmaddubsw ymm2, ymm0, kUVToB_AVX        // scale B UV                       
+    vpmaddubsw ymm1, ymm0, kUVToG_AVX        // scale G UV                       
+    vpmaddubsw ymm0, ymm0, kUVToR_AVX        // scale R UV                       
+    vpsubw     ymm2, ymm2, kUVBiasB_AVX      // unbias back to signed           
+    vpsubw     ymm1, ymm1, kUVBiasG_AVX                                            
+    vpsubw     ymm0, ymm0, kUVBiasR_AVX                                            
+
+    // Step 2: Find Y contribution to 16 R,G,B values                         
+    vmovdqu    xmm3, [eax]                  // NOLINT  
+    lea        eax, [eax + 16]                                            
+    vpermq     ymm3, ymm3, 0xd8
+    vpunpcklbw ymm3, ymm3, ymm4
+    vpsubsw    ymm3, ymm3, kYSub16_AVX                                             
+    vpmullw    ymm3, ymm3, kYToRgb_AVX                                             
+    vpaddsw    ymm2, ymm2, ymm3           // B += Y                         
+    vpaddsw    ymm1, ymm1, ymm3           // G += Y                          
+    vpaddsw    ymm0, ymm0, ymm3           // R += Y                         
+    vpsraw     ymm2, ymm2, 6                                                   
+    vpsraw     ymm1, ymm1, 6                                                   
+    vpsraw     ymm0, ymm0, 6                                                   
+    vpackuswb  ymm2, ymm2, ymm2           // B
+    vpackuswb  ymm1, ymm1, ymm1           // G
+    vpackuswb  ymm0, ymm0, ymm0           // R  
+
+    // Step 3: Weave into ARGB
+    vpunpcklbw ymm2, ymm2, ymm1           // BG
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklbw ymm0, ymm0, ymm5           // RA
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm1, ymm2, ymm0           // BGRA first 4 pixels
+    vpunpckhwd ymm2, ymm2, ymm0           // BGRA next 4 pixels
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_I422TOARGBROW_AVX2
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+
 static const vec8 kUVToB = {
   UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
 };

unit_test/convert_test.cc

@@ -497,11 +497,10 @@ TESTPLANARTOB(I420, 2, 2, UYVY, 2, 4, 1, ARGB, 4)
 TESTPLANARTOB(I422, 2, 1, YUY2, 2, 4, 0, ARGB, 4)
 TESTPLANARTOB(I422, 2, 1, UYVY, 2, 4, 0, ARGB, 4)
 TESTPLANARTOB(I420, 2, 2, I400, 1, 1, 0, ARGB, 4)
-// TODO(fbarchard): These tests fail is width is odd.
+TESTPLANARTOB(I420, 2, 2, BayerBGGR, 1, 1, 2, ARGB, 4)
-// TESTPLANARTOB(I420, 2, 2, BayerBGGR, 1, 1, 2, ARGB, 4)
+TESTPLANARTOB(I420, 2, 2, BayerRGGB, 1, 1, 2, ARGB, 4)
-// TESTPLANARTOB(I420, 2, 2, BayerRGGB, 1, 1, 2, ARGB, 4)
+TESTPLANARTOB(I420, 2, 2, BayerGBRG, 1, 1, 2, ARGB, 4)
-// TESTPLANARTOB(I420, 2, 2, BayerGBRG, 1, 1, 2, ARGB, 4)
+TESTPLANARTOB(I420, 2, 2, BayerGRBG, 1, 1, 2, ARGB, 4)
-// TESTPLANARTOB(I420, 2, 2, BayerGRBG, 1, 1, 2, ARGB, 4)
 
 #define TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,       \
                          W1280, DIFF, N, NEG, OFF)                             \