port I420ToBGRA to AVX2.

BUG=269
TESTED=c:\intelsde\sde -ast -hsw -- out\release\libyuv_unittest.exe --gtest_filter=*I420ToBGRA*
R=brucedawson@google.com, harryjin@google.com, magjed@chromium.org

Review URL: https://webrtc-codereview.appspot.com/26869004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1127 16f28f9a-4ce2-e073-06de-1de4eb20be90

7 files changed, 114 insertions, 10 deletions

diff --git a/README.chromium b/README.chromium
index 600d33f..d86bde5 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1125
+Version: 1127
 License: BSD
 License File: LICENSE
 
diff --git a/include/libyuv/row.h b/include/libyuv/row.h
index 07d3875..5a4c497 100644
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@@ -212,6 +212,7 @@ extern "C" {
 #define HAS_ARGBTOYJROW_AVX2
 #define HAS_ARGBTOYROW_AVX2
 #define HAS_I422TOARGBROW_AVX2
+#define HAS_I422TOBGRAROW_AVX2
 #define HAS_INTERPOLATEROW_AVX2
 #define HAS_MERGEUVROW_AVX2
 #define HAS_MIRRORROW_AVX2
@@ -1055,6 +1056,11 @@ void I422ToARGBRow_AVX2(const uint8* src_y,
                         const uint8* src_v,
                         uint8* dst_argb,
                         int width);
+void I422ToBGRARow_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,
@@ -1137,6 +1143,11 @@ void I422ToARGBRow_Any_AVX2(const uint8* src_y,
                             const uint8* src_v,
                             uint8* dst_argb,
                             int width);
+void I422ToBGRARow_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,
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index 349b523..88f948f 100644
--- a/include/libyuv/version.h
+++ b/include/libyuv/version.h
@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1125
+#define LIBYUV_VERSION 1127
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/source/convert_from.cc b/source/convert_from.cc
index 2220005..e044e03 100644
--- a/source/convert_from.cc
+++ b/source/convert_from.cc
@@ -551,6 +551,14 @@ int I420ToBGRA(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOBGRAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToBGRARow = I422ToBGRARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_I422TOBGRAROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     I422ToBGRARow = I422ToBGRARow_Any_NEON;
diff --git a/source/planar_functions.cc b/source/planar_functions.cc
index ebed839..b21192b 100644
--- a/source/planar_functions.cc
+++ b/source/planar_functions.cc
@@ -793,14 +793,6 @@ int I422ToBGRA(const uint8* src_y, int src_stride_y,
     height = 1;
     src_stride_y = src_stride_u = src_stride_v = dst_stride_bgra = 0;
   }
-#if defined(HAS_I422TOBGRAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
-    I422ToBGRARow = I422ToBGRARow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToBGRARow = I422ToBGRARow_NEON;
-    }
-  }
-#endif
 #if defined(HAS_I422TOBGRAROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
     I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
@@ -809,6 +801,22 @@ int I422ToBGRA(const uint8* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_I422TOBGRAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToBGRARow = I422ToBGRARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#endif
 #if defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
       IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
diff --git a/source/row_any.cc b/source/row_any.cc
index 9d8a5e5..b1ede4e 100644
--- a/source/row_any.cc
+++ b/source/row_any.cc
@@ -65,6 +65,9 @@ YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
 #ifdef HAS_I422TOARGBROW_AVX2
 YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
 #endif  // HAS_I422TOARGBROW_AVX2
+#ifdef HAS_I422TOBGRAROW_AVX2
+YANY(I422ToBGRARow_Any_AVX2, I422ToBGRARow_AVX2, I422ToBGRARow_C, 1, 4, 15)
+#endif  // HAS_I422TOBGRAROW_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)
diff --git a/source/row_win.cc b/source/row_win.cc
index 969d78c..d0a1059 100644
--- a/source/row_win.cc
+++ b/source/row_win.cc
@@ -1600,6 +1600,80 @@ void I422ToARGBRow_AVX2(const uint8* y_buf,
     ret
   }
 }
+
+// 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,
+                        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:
+    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
+// TODO(fbarchard): Switch register order to match SSSE3.
+    vpackuswb  ymm2, ymm2, ymm2           // B
+    vpackuswb  ymm1, ymm1, ymm1           // G
+    vpackuswb  ymm0, ymm0, ymm0           // R
+
+    // Step 3: Weave into BGRA
+    vpunpcklbw ymm1, ymm1, ymm2           // GB
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm0, ymm5, ymm0           // AR
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm2, ymm0, ymm1           // ARGB first 8 pixels
+    vpunpckhwd ymm0, ymm0, ymm1           // ARGB next 8 pixels
+    vmovdqu    [edx], ymm2
+    vmovdqu    [edx + 32], ymm0
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
 #endif  // HAS_I422TOARGBROW_AVX2
 
 #ifdef HAS_I422TOARGBROW_SSSE3