AVX2 hash using vex128 as first step.

BUG=none
TEST=BenchmarkDjb2_Opt
R=ryanpetrie@google.com

Review URL: https://webrtc-codereview.appspot.com/2219004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@792 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 791
+Version: 792
 License: BSD
 License File: LICENSE
 

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 791
+#define LIBYUV_VERSION 792
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/compare.cc

@@ -10,6 +10,8 @@
 
 #include "libyuv/compare.h"
 
+#include <stdio.h>  // printf
+
 #include <float.h>
 #include <math.h>
 #ifdef _OPENMP
@@ -34,9 +36,13 @@ uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
     (defined(_M_IX86) || \
     (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))))
 #define HAS_HASHDJB2_SSE41
-
 uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);
 
+#if _MSC_VER >= 1700
+#define HAS_HASHDJB2_AVX2
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed);
+#endif
+
 #endif  // HAS_HASHDJB2_SSE41
 
 // hash seed of 5381 recommended.
@@ -48,6 +54,11 @@ uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
     HashDjb2_SSE = HashDjb2_SSE41;
   }
 #endif
+#if defined(HAS_HASHDJB2_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HashDjb2_SSE = HashDjb2_AVX2;
+  }
+#endif
 
   const int kBlockSize = 1 << 15;  // 32768;
   while (count >= static_cast<uint64>(kBlockSize)) {

source/compare_win.cc

@@ -184,6 +184,46 @@ uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
     ret
   }
 }
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+__declspec(naked) __declspec(align(16))
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+    movdqa     xmm6, kHash16x33
+
+    align      16
+  wloop:
+    vpmovzxbd  xmm3, dword ptr [eax]  // src[0-3]
+    pmulld     xmm0, xmm6  // hash *= 33 ^ 16
+    vpmovzxbd  xmm4, dword ptr [eax + 4]  // src[4-7]
+    pmulld     xmm3, kHashMul0
+    vpmovzxbd  xmm2, dword ptr [eax + 8]  // src[8-11]
+    pmulld     xmm4, kHashMul1
+    vpmovzxbd  xmm1, dword ptr [eax + 12]  // src[12-15]
+    pmulld     xmm2, kHashMul2
+    lea        eax, [eax + 16]
+    pmulld     xmm1, kHashMul3
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    sub        ecx, 16
+    paddd      xmm1, xmm3
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
 #endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
 
 #ifdef __cplusplus