|
1 |
|
2 /* |
|
3 * Copyright 2009 The Android Open Source Project |
|
4 * |
|
5 * Use of this source code is governed by a BSD-style license that can be |
|
6 * found in the LICENSE file. |
|
7 */ |
|
8 |
|
9 |
|
10 #include <emmintrin.h> |
|
11 #include "SkUtils_opts_SSE2.h" |
|
12 |
|
13 void sk_memset16_SSE2(uint16_t *dst, uint16_t value, int count) |
|
14 { |
|
15 SkASSERT(dst != NULL && count >= 0); |
|
16 |
|
17 // dst must be 2-byte aligned. |
|
18 SkASSERT((((size_t) dst) & 0x01) == 0); |
|
19 |
|
20 if (count >= 32) { |
|
21 while (((size_t)dst) & 0x0F) { |
|
22 *dst++ = value; |
|
23 --count; |
|
24 } |
|
25 __m128i *d = reinterpret_cast<__m128i*>(dst); |
|
26 __m128i value_wide = _mm_set1_epi16(value); |
|
27 while (count >= 32) { |
|
28 _mm_store_si128(d , value_wide); |
|
29 _mm_store_si128(d + 1, value_wide); |
|
30 _mm_store_si128(d + 2, value_wide); |
|
31 _mm_store_si128(d + 3, value_wide); |
|
32 d += 4; |
|
33 count -= 32; |
|
34 } |
|
35 dst = reinterpret_cast<uint16_t*>(d); |
|
36 } |
|
37 while (count > 0) { |
|
38 *dst++ = value; |
|
39 --count; |
|
40 } |
|
41 } |
|
42 |
|
43 void sk_memset32_SSE2(uint32_t *dst, uint32_t value, int count) |
|
44 { |
|
45 SkASSERT(dst != NULL && count >= 0); |
|
46 |
|
47 // dst must be 4-byte aligned. |
|
48 SkASSERT((((size_t) dst) & 0x03) == 0); |
|
49 |
|
50 if (count >= 16) { |
|
51 while (((size_t)dst) & 0x0F) { |
|
52 *dst++ = value; |
|
53 --count; |
|
54 } |
|
55 __m128i *d = reinterpret_cast<__m128i*>(dst); |
|
56 __m128i value_wide = _mm_set1_epi32(value); |
|
57 while (count >= 16) { |
|
58 _mm_store_si128(d , value_wide); |
|
59 _mm_store_si128(d + 1, value_wide); |
|
60 _mm_store_si128(d + 2, value_wide); |
|
61 _mm_store_si128(d + 3, value_wide); |
|
62 d += 4; |
|
63 count -= 16; |
|
64 } |
|
65 dst = reinterpret_cast<uint32_t*>(d); |
|
66 } |
|
67 while (count > 0) { |
|
68 *dst++ = value; |
|
69 --count; |
|
70 } |
|
71 } |