SSE2 implementation of memcpy32

src/opts/SkUtils_opts_SSE2.cpp

 /*
  * Copyright 2009 The Android Open Source Project
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include <emmintrin.h>
 #include "SkUtils_opts_SSE2.h"
 
@@ skipping 49 matching lines @@
             d += 4;
             count -= 16;
         }
         dst = reinterpret_cast<uint32_t*>(d);
     }
     while (count > 0) {
         *dst++ = value;
         --count;
     }
 }
+
+void sk_memcpy32_SSE2(uint32_t *dst, const uint32_t *src, int count)
+{
+    if (count >= 16) {
+        while (((size_t)dst) & 0x0F) {
+            *dst++ = *src++;
+            --count;
+        }
+        __m128i *d = reinterpret_cast<__m128i*>(dst);
+        const __m128i *s = reinterpret_cast<const __m128i*>(src);
+        while (count >= 16) {
+            __m128i src_pixel1 =  _mm_loadu_si128(s++);

[mtklein, 2014/05/15 15:38:30]

src_pixelN seems like a misleading name, as it's really 4 pixels.  I wouldn't
mind something neutral like 
__m128i a = ... ;
__m128i b = ... ;

[qiankun, 2014/05/20 09:35:30]

Done.

+            __m128i src_pixel2 =  _mm_loadu_si128(s++);
+            __m128i src_pixel3 =  _mm_loadu_si128(s++);
+            __m128i src_pixel4 =  _mm_loadu_si128(s++);
+
+            _mm_store_si128(d    , src_pixel1);

[mtklein, 2014/05/15 15:38:30]

Any chance you compared _mm_store and _mm_stream here?  Last I checked,
_mm_store was much, much better, but that may be an artifact of
microbenchmarking being really cache friendly.

[mtklein, 2014/05/15 15:38:30]

For symmetry with s, maybe _mm_store_si128(d++, ...) here instead of += 4 below?

[qiankun, 2014/05/20 09:35:30]

Done.

[qiankun, 2014/05/20 09:35:30]

At my side, _mm_store is better than _mm_stream too.

+            _mm_store_si128(d + 1, src_pixel2);
+            _mm_store_si128(d + 2, src_pixel3);
+            _mm_store_si128(d + 3, src_pixel4);
+            d += 4;
+            count -= 16;
+        }
+        dst = reinterpret_cast<uint32_t*>(d);
+        src = reinterpret_cast<const uint32_t*>(s);
+    }
+    while (count > 0) {
+        *dst++ = *src++;
+        --count;
+    }
+}