1/*
2 * Copyright 2009 The Android Open Source Project
3 *
4 * Use of this source code is governed by a BSD-style license that can be
5 * found in the LICENSE file.
6 */
7
8#include <emmintrin.h>
9#include "SkUtils_opts_SSE2.h"
10
11void sk_memset16_SSE2(uint16_t *dst, uint16_t value, int count)
12{
13    SkASSERT(dst != NULL && count >= 0);
14
15    // dst must be 2-byte aligned.
16    SkASSERT((((size_t) dst) & 0x01) == 0);
17
18    if (count >= 32) {
19        while (((size_t)dst) & 0x0F) {
20            *dst++ = value;
21            --count;
22        }
23        __m128i *d = reinterpret_cast<__m128i*>(dst);
24        __m128i value_wide = _mm_set1_epi16(value);
25        while (count >= 32) {
26            _mm_store_si128(d    , value_wide);
27            _mm_store_si128(d + 1, value_wide);
28            _mm_store_si128(d + 2, value_wide);
29            _mm_store_si128(d + 3, value_wide);
30            d += 4;
31            count -= 32;
32        }
33        dst = reinterpret_cast<uint16_t*>(d);
34    }
35    while (count > 0) {
36        *dst++ = value;
37        --count;
38    }
39}
40
41void sk_memset32_SSE2(uint32_t *dst, uint32_t value, int count)
42{
43    SkASSERT(dst != NULL && count >= 0);
44
45    // dst must be 4-byte aligned.
46    SkASSERT((((size_t) dst) & 0x03) == 0);
47
48    if (count >= 16) {
49        while (((size_t)dst) & 0x0F) {
50            *dst++ = value;
51            --count;
52        }
53        __m128i *d = reinterpret_cast<__m128i*>(dst);
54        __m128i value_wide = _mm_set1_epi32(value);
55        while (count >= 16) {
56            _mm_store_si128(d    , value_wide);
57            _mm_store_si128(d + 1, value_wide);
58            _mm_store_si128(d + 2, value_wide);
59            _mm_store_si128(d + 3, value_wide);
60            d += 4;
61            count -= 16;
62        }
63        dst = reinterpret_cast<uint32_t*>(d);
64    }
65    while (count > 0) {
66        *dst++ = value;
67        --count;
68    }
69}
70
71void sk_memcpy32_SSE2(uint32_t *dst, const uint32_t *src, int count)
72{
73    if (count >= 16) {
74        while (((size_t)dst) & 0x0F) {
75            *dst++ = *src++;
76            --count;
77        }
78        __m128i *dst128 = reinterpret_cast<__m128i*>(dst);
79        const __m128i *src128 = reinterpret_cast<const __m128i*>(src);
80        while (count >= 16) {
81            __m128i a =  _mm_loadu_si128(src128++);
82            __m128i b =  _mm_loadu_si128(src128++);
83            __m128i c =  _mm_loadu_si128(src128++);
84            __m128i d =  _mm_loadu_si128(src128++);
85
86            _mm_store_si128(dst128++, a);
87            _mm_store_si128(dst128++, b);
88            _mm_store_si128(dst128++, c);
89            _mm_store_si128(dst128++, d);
90            count -= 16;
91        }
92        dst = reinterpret_cast<uint32_t*>(dst128);
93        src = reinterpret_cast<const uint32_t*>(src128);
94    }
95    while (count > 0) {
96        *dst++ = *src++;
97        --count;
98    }
99}
100