Port ARGBToJ420 to AVX2

source/row_win.cc

Index: source/row_win.cc
diff --git a/source/row_win.cc b/source/row_win.cc
index b843998d89c0da59b6e41e3113dd96519db85d90..a8c16c3c1ef3a4c0ed99814f9c3898636213e504 100644
--- a/source/row_win.cc
+++ b/source/row_win.cc
@@ -1505,7 +1505,7 @@ void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
     pmaddubsw  xmm3, xmm6
     phaddw     xmm0, xmm2
     phaddw     xmm1, xmm3
-    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm0, xmm5  // +.5 rounding -> unsigned
     paddw      xmm1, xmm5
     psraw      xmm0, 8
     psraw      xmm1, 8
@@ -1590,6 +1590,73 @@ void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
 }
 #endif  // HAS_ARGBTOUVROW_AVX2
 
+#ifdef HAS_ARGBTOUVJROW_AVX2
+__declspec(naked)
+void ARGBToUVJRow_AVX2(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]  // width
+    vbroadcastf128 ymm5, xmmword ptr kAddUV128
+    vbroadcastf128 ymm6, xmmword ptr kARGBToV
+    vbroadcastf128 ymm7, xmmword ptr kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 32x2 argb pixels to 16x1 */
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    vpavgb     ymm2, ymm2, [eax + esi + 64]
+    vpavgb     ymm3, ymm3, [eax + esi + 96]
+    lea        eax,  [eax + 128]
+    vshufps    ymm4, ymm0, ymm1, 0x88
+    vshufps    ymm0, ymm0, ymm1, 0xdd
+    vpavgb     ymm0, ymm0, ymm4  // mutated by vshufps
+    vshufps    ymm4, ymm2, ymm3, 0x88
+    vshufps    ymm2, ymm2, ymm3, 0xdd
+    vpavgb     ymm2, ymm2, ymm4  // mutated by vshufps
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 32 different pixels, its 16 pixels of U and 16 of V
+    vpmaddubsw ymm1, ymm0, ymm7  // U
+    vpmaddubsw ymm3, ymm2, ymm7
+    vpmaddubsw ymm0, ymm0, ymm6  // V
+    vpmaddubsw ymm2, ymm2, ymm6
+    vphaddw    ymm1, ymm1, ymm3  // mutates
+    vphaddw    ymm0, ymm0, ymm2
+    vpaddw     ymm1, ymm1, ymm5  // +.5 rounding -> unsigned
+    vpaddw     ymm0, ymm0, ymm5
+    vpsraw     ymm1, ymm1, 8
+    vpsraw     ymm0, ymm0, 8
+    vpacksswb  ymm0, ymm1, ymm0  // mutates
+    vpermq     ymm0, ymm0, 0xd8  // For vpacksswb
+    vpshufb    ymm0, ymm0, ymmword ptr kShufARGBToUV_AVX  // for vshufps/vphaddw
+
+    // step 3 - store 16 U and 16 V values
+    vextractf128 [edx], ymm0, 0 // U
+    vextractf128 [edx + edi], ymm0, 1 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOUVJROW_AVX2
+
 __declspec(naked)
 void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
                           uint8* dst_u, uint8* dst_v, int width) {