michael@0: 
michael@0: /*
michael@0:  * Copyright 2009 The Android Open Source Project
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: 
michael@0: 
michael@0: #include <emmintrin.h>
michael@0: #include "SkUtils_opts_SSE2.h"
michael@0: 
michael@0: void sk_memset16_SSE2(uint16_t *dst, uint16_t value, int count)
michael@0: {
michael@0:     SkASSERT(dst != NULL && count >= 0);
michael@0: 
michael@0:     // dst must be 2-byte aligned.
michael@0:     SkASSERT((((size_t) dst) & 0x01) == 0);
michael@0: 
michael@0:     if (count >= 32) {
michael@0:         while (((size_t)dst) & 0x0F) {
michael@0:             *dst++ = value;
michael@0:             --count;
michael@0:         }
michael@0:         __m128i *d = reinterpret_cast<__m128i*>(dst);
michael@0:         __m128i value_wide = _mm_set1_epi16(value);
michael@0:         while (count >= 32) {
michael@0:             _mm_store_si128(d    , value_wide);
michael@0:             _mm_store_si128(d + 1, value_wide);
michael@0:             _mm_store_si128(d + 2, value_wide);
michael@0:             _mm_store_si128(d + 3, value_wide);
michael@0:             d += 4;
michael@0:             count -= 32;
michael@0:         }
michael@0:         dst = reinterpret_cast<uint16_t*>(d);
michael@0:     }
michael@0:     while (count > 0) {
michael@0:         *dst++ = value;
michael@0:         --count;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void sk_memset32_SSE2(uint32_t *dst, uint32_t value, int count)
michael@0: {
michael@0:     SkASSERT(dst != NULL && count >= 0);
michael@0: 
michael@0:     // dst must be 4-byte aligned.
michael@0:     SkASSERT((((size_t) dst) & 0x03) == 0);
michael@0: 
michael@0:     if (count >= 16) {
michael@0:         while (((size_t)dst) & 0x0F) {
michael@0:             *dst++ = value;
michael@0:             --count;
michael@0:         }
michael@0:         __m128i *d = reinterpret_cast<__m128i*>(dst);
michael@0:         __m128i value_wide = _mm_set1_epi32(value);
michael@0:         while (count >= 16) {
michael@0:             _mm_store_si128(d    , value_wide);
michael@0:             _mm_store_si128(d + 1, value_wide);
michael@0:             _mm_store_si128(d + 2, value_wide);
michael@0:             _mm_store_si128(d + 3, value_wide);
michael@0:             d += 4;
michael@0:             count -= 16;
michael@0:         }
michael@0:         dst = reinterpret_cast<uint32_t*>(d);
michael@0:     }
michael@0:     while (count > 0) {
michael@0:         *dst++ = value;
michael@0:         --count;
michael@0:     }
michael@0: }