| Index: src/opts/SkUtils_opts_SSE2.cpp
|
| diff --git a/src/opts/SkUtils_opts_SSE2.cpp b/src/opts/SkUtils_opts_SSE2.cpp
|
| index a3c5aa5dfd7f1cf417abda23f10d662d80bf8045..bd2f9b29a44ad612a93b14ee57c16b16d54b1f8f 100644
|
| --- a/src/opts/SkUtils_opts_SSE2.cpp
|
| +++ b/src/opts/SkUtils_opts_SSE2.cpp
|
| @@ -67,3 +67,33 @@ void sk_memset32_SSE2(uint32_t *dst, uint32_t value, int count)
|
| --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 *dst128 = reinterpret_cast<__m128i*>(dst);
|
| + const __m128i *src128 = reinterpret_cast<const __m128i*>(src);
|
| + while (count >= 16) {
|
| + __m128i a = _mm_loadu_si128(src128++);
|
| + __m128i b = _mm_loadu_si128(src128++);
|
| + __m128i c = _mm_loadu_si128(src128++);
|
| + __m128i d = _mm_loadu_si128(src128++);
|
| +
|
| + _mm_store_si128(dst128++, a);
|
| + _mm_store_si128(dst128++, b);
|
| + _mm_store_si128(dst128++, c);
|
| + _mm_store_si128(dst128++, d);
|
| + count -= 16;
|
| + }
|
| + dst = reinterpret_cast<uint32_t*>(dst128);
|
| + src = reinterpret_cast<const uint32_t*>(src128);
|
| + }
|
| + while (count > 0) {
|
| + *dst++ = *src++;
|
| + --count;
|
| + }
|
| +}
|
|
|