OLD | NEW |
(Empty) | |
| 1 /* |
| 2 * Copyright 2014 Google Inc. |
| 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 "SkBenchmark.h" |
| 9 #include "SkRandom.h" |
| 10 #include "SkTemplates.h" |
| 11 |
| 12 template <typename Memcpy32> |
| 13 class Memcpy32Bench : public SkBenchmark { |
| 14 public: |
| 15 explicit Memcpy32Bench(int count, Memcpy32 memcpy32, const char* name) |
| 16 : fCount(count) |
| 17 , fMemcpy32(memcpy32) |
| 18 , fName(SkStringPrintf("%s_%d", name, count)) {} |
| 19 |
| 20 virtual const char* onGetName() SK_OVERRIDE { |
| 21 return fName.c_str(); |
| 22 } |
| 23 |
| 24 virtual bool isSuitableFor(Backend backend) SK_OVERRIDE { |
| 25 return backend == kNonRendering_Backend; |
| 26 } |
| 27 |
| 28 virtual void onPreDraw() SK_OVERRIDE { |
| 29 fDst.reset(fCount); |
| 30 fSrc.reset(fCount); |
| 31 |
| 32 SkRandom rand; |
| 33 for (int i = 0; i < fCount; i++) { |
| 34 fSrc[i] = rand.nextU(); |
| 35 } |
| 36 } |
| 37 |
| 38 virtual void onDraw(const int loops, SkCanvas*) SK_OVERRIDE { |
| 39 for (int i = 0; i < loops; i++) { |
| 40 fMemcpy32(fDst, fSrc, fCount); |
| 41 } |
| 42 } |
| 43 |
| 44 private: |
| 45 SkAutoTMalloc<uint32_t> fDst, fSrc; |
| 46 |
| 47 int fCount; |
| 48 Memcpy32 fMemcpy32; |
| 49 const SkString fName; |
| 50 }; |
| 51 |
| 52 template <typename Memcpy32> |
| 53 static Memcpy32Bench<Memcpy32>* Bench(int count, Memcpy32 memcpy32, const char*
name) { |
| 54 return new Memcpy32Bench<Memcpy32>(count, memcpy32, name); |
| 55 } |
| 56 #define BENCH(memcpy32, count) DEF_BENCH(return Bench(count, memcpy32, #memcpy32
); ) |
| 57 |
| 58 |
| 59 // Let the libc developers do what they think is best. |
| 60 static void memcpy32_memcpy(uint32_t* dst, const uint32_t* src, int count) { |
| 61 memcpy(dst, src, sizeof(uint32_t) * count); |
| 62 } |
| 63 BENCH(memcpy32_memcpy, 10) |
| 64 BENCH(memcpy32_memcpy, 100) |
| 65 BENCH(memcpy32_memcpy, 1000) |
| 66 BENCH(memcpy32_memcpy, 10000) |
| 67 BENCH(memcpy32_memcpy, 100000) |
| 68 |
| 69 // Let the compiler's autovectorizer do what it thinks is best. |
| 70 static void memcpy32_autovectorize(uint32_t* dst, const uint32_t* src, int count
) { |
| 71 while (count --> 0) { |
| 72 *dst++ = *src++; |
| 73 } |
| 74 } |
| 75 BENCH(memcpy32_autovectorize, 10) |
| 76 BENCH(memcpy32_autovectorize, 100) |
| 77 BENCH(memcpy32_autovectorize, 1000) |
| 78 BENCH(memcpy32_autovectorize, 10000) |
| 79 BENCH(memcpy32_autovectorize, 100000) |
| 80 |
| 81 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
| 82 |
| 83 // Align dst to 16 bytes, then use aligned stores. src isn't algined, so use un
aligned loads. |
| 84 static void memcpy32_sse2_align(uint32_t* dst, const uint32_t* src, int count) { |
| 85 if (count >= 16) { |
| 86 while (uintptr_t(dst) & 0xF) { |
| 87 *dst++ = *src++; |
| 88 count--; |
| 89 } |
| 90 |
| 91 __m128i* dst128 = reinterpret_cast<__m128i*>(dst); |
| 92 const __m128i* src128 = reinterpret_cast<const __m128i*>(src); |
| 93 while (count >= 16) { |
| 94 __m128i a = _mm_loadu_si128(src128++); |
| 95 __m128i b = _mm_loadu_si128(src128++); |
| 96 __m128i c = _mm_loadu_si128(src128++); |
| 97 __m128i d = _mm_loadu_si128(src128++); |
| 98 |
| 99 _mm_store_si128(dst128++, a); |
| 100 _mm_store_si128(dst128++, b); |
| 101 _mm_store_si128(dst128++, c); |
| 102 _mm_store_si128(dst128++, d); |
| 103 |
| 104 count -= 16; |
| 105 } |
| 106 |
| 107 dst = reinterpret_cast<uint32_t*>(dst128); |
| 108 src = reinterpret_cast<const uint32_t*>(src128); |
| 109 } |
| 110 |
| 111 while (count --> 0) { |
| 112 *dst++ = *src++; |
| 113 } |
| 114 } |
| 115 BENCH(memcpy32_sse2_align, 10) |
| 116 BENCH(memcpy32_sse2_align, 100) |
| 117 BENCH(memcpy32_sse2_align, 1000) |
| 118 BENCH(memcpy32_sse2_align, 10000) |
| 119 BENCH(memcpy32_sse2_align, 100000) |
| 120 |
| 121 // Leave both dst and src unaliged, and so use unaligned stores for dst and unal
igned loads for src. |
| 122 static void memcpy32_sse2_unalign(uint32_t* dst, const uint32_t* src, int count)
{ |
| 123 __m128i* dst128 = reinterpret_cast<__m128i*>(dst); |
| 124 const __m128i* src128 = reinterpret_cast<const __m128i*>(src); |
| 125 while (count >= 16) { |
| 126 __m128i a = _mm_loadu_si128(src128++); |
| 127 __m128i b = _mm_loadu_si128(src128++); |
| 128 __m128i c = _mm_loadu_si128(src128++); |
| 129 __m128i d = _mm_loadu_si128(src128++); |
| 130 |
| 131 _mm_storeu_si128(dst128++, a); |
| 132 _mm_storeu_si128(dst128++, b); |
| 133 _mm_storeu_si128(dst128++, c); |
| 134 _mm_storeu_si128(dst128++, d); |
| 135 |
| 136 count -= 16; |
| 137 } |
| 138 |
| 139 dst = reinterpret_cast<uint32_t*>(dst128); |
| 140 src = reinterpret_cast<const uint32_t*>(src128); |
| 141 while (count --> 0) { |
| 142 *dst++ = *src++; |
| 143 } |
| 144 } |
| 145 BENCH(memcpy32_sse2_unalign, 10) |
| 146 BENCH(memcpy32_sse2_unalign, 100) |
| 147 BENCH(memcpy32_sse2_unalign, 1000) |
| 148 BENCH(memcpy32_sse2_unalign, 10000) |
| 149 BENCH(memcpy32_sse2_unalign, 100000) |
| 150 |
| 151 #endif // SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
| 152 |
| 153 #undef BENCH |
| 154 |
OLD | NEW |