senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 1 | /* |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 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. |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 6 | */ |
| 7 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 8 | #include <emmintrin.h> |
| 9 | #include "SkUtils_opts_SSE2.h" |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 10 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 11 | void 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) { |
commit-bot@chromium.org | 9772a52 | 2013-08-05 20:25:57 +0000 | [diff] [blame] | 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; |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 31 | count -= 32; |
| 32 | } |
| 33 | dst = reinterpret_cast<uint16_t*>(d); |
| 34 | } |
| 35 | while (count > 0) { |
| 36 | *dst++ = value; |
| 37 | --count; |
| 38 | } |
| 39 | } |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 40 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 41 | void 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) { |
commit-bot@chromium.org | 9772a52 | 2013-08-05 20:25:57 +0000 | [diff] [blame] | 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; |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 61 | count -= 16; |
| 62 | } |
| 63 | dst = reinterpret_cast<uint32_t*>(d); |
| 64 | } |
| 65 | while (count > 0) { |
| 66 | *dst++ = value; |
| 67 | --count; |
| 68 | } |
| 69 | } |