Add ARGBExtractAlpha function

source/row_win.cc

Index: source/row_win.cc
diff --git a/source/row_win.cc b/source/row_win.cc
index a8c16c3c1ef3a4c0ed99814f9c3898636213e504..4a0a14fe258868c16151fd73ce11f4a85de0cb67 100644
--- a/source/row_win.cc
+++ b/source/row_win.cc
@@ -3532,6 +3532,33 @@ void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
 }
 #endif  // HAS_ARGBCOPYALPHAROW_AVX2
 
+#ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
+// width in pixels
+__declspec(naked)
+void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_a
+    mov        ecx, [esp + 12]  // width
+
+  extractloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm0, 24
+    psrld      xmm1, 24
+    packssdw   xmm0, xmm1
+    packuswb   xmm0, xmm1

[fbarchard1, 2016/05/25 21:34:35]

nit
should this be packuswb   xmm0, xmm0?

packssdw   xmm0, xmm1
took 32 bytes (8 ints) in xmm0 and xmm1 and made 16 bytes (8 shorts) in xmm0.
I guess the end result isnt different but packuswb   xmm0, xmm0 would take 16
bytes (8 shorts) and produce 8 bytes, using just the 1 register.
ditto for gcc version.

[magjed_chromium, 2016/05/26 08:28:33]

Done.

+    movq       qword ptr [edx], xmm0

[fbarchard1, 2016/05/25 21:34:35]

fyi could be a little faster if it writes full 16 bytes, so making loop do 64
bytes of source.  Probably best to keep the loop smaller though

+    lea        edx, [edx + 8]
+    sub        ecx, 8
+    jg         extractloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBEXTRACTALPHAROW_SSE2
+
 #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
 // width in pixels
 __declspec(naked)