Resubmit - Use SSE2 block differ for chromoting

remoting/host/differ_block.cc

Index: remoting/host/differ_block.cc
diff --git a/remoting/host/differ_block.cc b/remoting/host/differ_block.cc
index f7b785d7c09831a0f16ede3b590f8569e5572b99..88a4f8b615dd73340f6297847658fe7a5ff75781 100644
--- a/remoting/host/differ_block.cc
+++ b/remoting/host/differ_block.cc
@@ -4,76 +4,13 @@
 
 #include "remoting/host/differ_block.h"
 
-#include <stdlib.h>
-
-#if !defined(USE_SSE2)
-#if defined(__SSE2__) || defined(ARCH_CPU_X86_64) || defined(_MSC_VER)
-#define USE_SSE2 1
-#else
-#define USE_SSE2 0
-#endif
-#endif
-
-#if USE_SSE2
-#include <emmintrin.h>
-#endif
+#include "build/build_config.h"
+#include "media/base/cpu_features.h"
+#include "remoting/host/differ_block_internal.h"
 
 namespace remoting {
 
-#if USE_SSE2
-int BlockDifference(const uint8* image1, const uint8* image2, int stride) {
-  __m128i acc = _mm_setzero_si128();
-  __m128i v0;
-  __m128i v1;
-  __m128i sad;
-  for (int y = 0; y < kBlockHeight; ++y) {
-    const __m128i* i1 = reinterpret_cast<const __m128i*>(image1);
-    const __m128i* i2 = reinterpret_cast<const __m128i*>(image2);
-    v0 = _mm_loadu_si128(i1);
-    v1 = _mm_loadu_si128(i2);
-    sad = _mm_sad_epu8(v0, v1);
-    acc = _mm_adds_epu16(acc, sad);
-    v0 = _mm_loadu_si128(i1 + 1);
-    v1 = _mm_loadu_si128(i2 + 1);
-    sad = _mm_sad_epu8(v0, v1);
-    acc = _mm_adds_epu16(acc, sad);
-    v0 = _mm_loadu_si128(i1 + 2);
-    v1 = _mm_loadu_si128(i2 + 2);
-    sad = _mm_sad_epu8(v0, v1);
-    acc = _mm_adds_epu16(acc, sad);
-    v0 = _mm_loadu_si128(i1 + 3);
-    v1 = _mm_loadu_si128(i2 + 3);
-    sad = _mm_sad_epu8(v0, v1);
-    acc = _mm_adds_epu16(acc, sad);
-    v0 = _mm_loadu_si128(i1 + 4);
-    v1 = _mm_loadu_si128(i2 + 4);
-    sad = _mm_sad_epu8(v0, v1);
-    acc = _mm_adds_epu16(acc, sad);
-    v0 = _mm_loadu_si128(i1 + 5);
-    v1 = _mm_loadu_si128(i2 + 5);
-    sad = _mm_sad_epu8(v0, v1);
-    acc = _mm_adds_epu16(acc, sad);
-    v0 = _mm_loadu_si128(i1 + 6);
-    v1 = _mm_loadu_si128(i2 + 6);
-    sad = _mm_sad_epu8(v0, v1);
-    acc = _mm_adds_epu16(acc, sad);
-    v0 = _mm_loadu_si128(i1 + 7);
-    v1 = _mm_loadu_si128(i2 + 7);
-    sad = _mm_sad_epu8(v0, v1);
-    acc = _mm_adds_epu16(acc, sad);
-    sad = _mm_shuffle_epi32(acc, 0xEE);  // [acc3, acc2, acc3, acc2]
-    sad = _mm_adds_epu16(sad, acc);
-    int diff = _mm_cvtsi128_si32(sad);
-    if (diff) {
-      return 1;
-    }
-    image1 += stride;
-    image2 += stride;
-  }
-  return 0;
-}
-#else
-int BlockDifference(const uint8* image1, const uint8* image2, int stride) {
+int BlockDifference_C(const uint8* image1, const uint8* image2, int stride) {
   // Number of uint64s in each row of the block.
   // This must be an integral number.
   int int64s_per_row = (kBlockWidth * kBytesPerPixel) / sizeof(uint64);
@@ -96,6 +33,27 @@ int BlockDifference(const uint8* image1, const uint8* image2, int stride) {
   }
   return 0;
 }
+
+int BlockDifference(const uint8* image1, const uint8* image2, int stride) {
+  static int (*diff_proc)(const uint8*, const uint8*, int) = NULL;
+
+  if (!diff_proc) {
+#if defined(ARCH_CPU_ARM_FAMILY)
+    // For ARM processors, always use C version.
+    // TODO(hclam): Implement a NEON version.
+    diff_proc = &BlockDifference_C;
+#else
+    // For x86 processors, check if SSE2 is supported.
+    if (media::hasSSE2() && kBlockWidth == 32)
+      diff_proc = &BlockDifference_SSE2_W32;
+    else if (media::hasSSE2() && kBlockWidth == 16)
+      diff_proc = &BlockDifference_SSE2_W16;
+    else
+      diff_proc = &BlockDifference_C;
 #endif
+  }
+
+  return diff_proc(image1, image2, stride);
+}
 
 }  // namespace remoting