Optimize yuv alpha blend AVX2 code to do 32 pixels at time.

source/cpu_id.cc

Index: source/cpu_id.cc
diff --git a/source/cpu_id.cc b/source/cpu_id.cc
index 2fdb82ab0e3bd6ccaeb3e7461e7a86cd0c3f4329..56b3c403ddfc051204ed79fd52b3ca917b2eee21 100644
--- a/source/cpu_id.cc
+++ b/source/cpu_id.cc
@@ -106,14 +106,17 @@ void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
 // X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
 int GetXCR0() {
   uint32 xcr0 = 0u;
-#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
+#if defined(_MSC_VER)
+//0 && (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)

[harryjin, 2015/12/08 07:48:58]

? Remove?

[fbarchard, 2015/12/08 19:24:04]

Done.

   xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
+  printf("xgetbv xcr0 %d\n", xcr0);

[harryjin, 2015/12/08 07:48:58]

Remove the printf?

[fbarchard, 2015/12/08 19:24:04]

Done.

 #elif defined(_M_IX86) && defined(_MSC_VER) && !defined(__clang__)
   __asm {
     xor        ecx, ecx    // xcr 0
     _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
     mov        xcr0, eax
   }
+  printf("asm xcr0 %d\n", xcr0);

[harryjin, 2015/12/08 07:48:58]

same here?

[fbarchard, 2015/12/08 19:24:04]

Done.

 #elif defined(__i386__) || defined(__x86_64__)
   asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
 #endif  // defined(__i386__) || defined(__x86_64__)
@@ -197,8 +200,8 @@ int InitCpuFlags(void) {
 
 #ifdef HAS_XGETBV
   // AVX requires CPU has AVX, XSAVE and OSXSave for xgetbv
-  if ((cpu_info1[2] & 0x1c000000) == 0x1c000000 &&  // AVX and OSXSave
-      (GetXCR0() & 6) == 6) {  // Test OD saves YMM registers
+  if (((cpu_info1[2] & 0x1c000000) == 0x1c000000) &&  // AVX and OSXSave
+      ((GetXCR0() & 6) == 6)) {  // Test OS saves YMM registers
     cpu_info |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) | kCpuHasAVX;
 
     // Detect AVX512bw