Port I422ToABGR to AVX2.

BUG=269
TESTED=intelsde on I422ToABGR
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/23149004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1138 16f28f9a-4ce2-e073-06de-1de4eb20be90

include/libyuv/row.h

@@ -202,6 +202,7 @@ extern "C" {
 #define HAS_ARGBTOYROW_AVX2
 #define HAS_I422TOARGBROW_AVX2
 #define HAS_I422TORGBAROW_AVX2
+#define HAS_I422TOABGRROW_AVX2
 #define HAS_INTERPOLATEROW_AVX2
 #define HAS_MERGEUVROW_AVX2
 #define HAS_MIRRORROW_AVX2
@@ -1058,6 +1059,11 @@ void I422ToRGBARow_AVX2(const uint8* src_y,
                         const uint8* src_v,
                         uint8* dst_argb,
                         int width);
+void I422ToABGRRow_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,
@@ -1150,6 +1156,11 @@ void I422ToRGBARow_Any_AVX2(const uint8* src_y,
                             const uint8* src_v,
                             uint8* dst_argb,
                             int width);
+void I422ToABGRRow_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,

source/convert_from.cc

@@ -620,6 +620,14 @@ int I420ToABGR(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOABGRROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToABGRRow = I422ToABGRRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOABGRROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     I422ToABGRRow = I422ToABGRRow_Any_NEON;

source/planar_functions.cc

@@ -886,6 +886,14 @@ int I422ToABGR(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOABGRROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToABGRRow = I422ToABGRRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_AVX2;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);

source/row_any.cc

@@ -71,6 +71,9 @@ YANY(I422ToBGRARow_Any_AVX2, I422ToBGRARow_AVX2, I422ToBGRARow_C, 1, 4, 15)
 #ifdef HAS_I422TORGBAROW_AVX2
 YANY(I422ToRGBARow_Any_AVX2, I422ToRGBARow_AVX2, I422ToRGBARow_C, 1, 4, 15)
 #endif  // HAS_I422TORGBAROW_AVX2
+#ifdef HAS_I422TOABGRROW_AVX2
+YANY(I422ToABGRRow_Any_AVX2, I422ToABGRRow_AVX2, I422ToABGRRow_C, 1, 4, 15)
+#endif  // HAS_I422TOABGRROW_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

@@ -1613,6 +1613,7 @@ void I422ToARGBRow_AVX2(const uint8* y_buf,
 
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
 __declspec(naked) __declspec(align(16))
 void I422ToBGRARow_AVX2(const uint8* y_buf,
                         const uint8* u_buf,
@@ -1658,6 +1659,7 @@ void I422ToBGRARow_AVX2(const uint8* y_buf,
 
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
 __declspec(naked) __declspec(align(16))
 void I422ToRGBARow_AVX2(const uint8* y_buf,
                         const uint8* u_buf,
@@ -1701,6 +1703,51 @@ void I422ToRGBARow_AVX2(const uint8* y_buf,
   }
 }
 
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
+__declspec(naked) __declspec(align(16))
+void I422ToABGRRow_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      4
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2
+
+    // Step 3: Weave into ABGR
+    vpunpcklbw ymm1, ymm2, ymm1           // RG
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm0, ymm5           // BA
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm1, ymm2           // RGBA first 8 pixels
+    vpunpckhwd ymm1, ymm1, ymm2           // RGBA next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
 #endif  // HAS_I422TOARGBROW_AVX2
 
 #ifdef HAS_I422TOARGBROW_SSSE3
@@ -1827,8 +1874,8 @@ void I444ToARGBRow_SSSE3(const uint8* y_buf,
     punpcklbw  xmm0, xmm1           // BG
     punpcklbw  xmm2, xmm5           // RA
     movdqa     xmm1, xmm0
-    punpcklwd  xmm0, xmm2           // RGBA first 4 pixels
-    punpckhwd  xmm1, xmm2           // RGBA next 4 pixels
+    punpcklwd  xmm0, xmm2           // ABGR first 4 pixels
+    punpckhwd  xmm1, xmm2           // ABGR next 4 pixels
     movdqu     [edx], xmm0
     movdqu     [edx + 16], xmm1
     lea        edx,  [edx + 32]