epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 1 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 2 | /* |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 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. |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 7 | */ |
| 8 | |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 9 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 10 | #include <emmintrin.h> |
| 11 | #include "SkUtils_opts_SSE2.h" |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 12 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 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++, value_wide); |
| 30 | _mm_store_si128(d++, value_wide); |
| 31 | _mm_store_si128(d++, value_wide); |
| 32 | count -= 32; |
| 33 | } |
| 34 | dst = reinterpret_cast<uint16_t*>(d); |
| 35 | } |
| 36 | while (count > 0) { |
| 37 | *dst++ = value; |
| 38 | --count; |
| 39 | } |
| 40 | } |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 41 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 42 | void sk_memset32_SSE2(uint32_t *dst, uint32_t value, int count) |
| 43 | { |
| 44 | SkASSERT(dst != NULL && count >= 0); |
| 45 | |
| 46 | // dst must be 4-byte aligned. |
| 47 | SkASSERT((((size_t) dst) & 0x03) == 0); |
| 48 | |
| 49 | if (count >= 16) { |
| 50 | while (((size_t)dst) & 0x0F) { |
| 51 | *dst++ = value; |
| 52 | --count; |
| 53 | } |
| 54 | __m128i *d = reinterpret_cast<__m128i*>(dst); |
| 55 | __m128i value_wide = _mm_set1_epi32(value); |
| 56 | while (count >= 16) { |
| 57 | _mm_store_si128(d++, value_wide); |
| 58 | _mm_store_si128(d++, value_wide); |
| 59 | _mm_store_si128(d++, value_wide); |
| 60 | _mm_store_si128(d++, value_wide); |
| 61 | count -= 16; |
| 62 | } |
| 63 | dst = reinterpret_cast<uint32_t*>(d); |
| 64 | } |
| 65 | while (count > 0) { |
| 66 | *dst++ = value; |
| 67 | --count; |
| 68 | } |
| 69 | } |