| 1 | /* |
| 2 | * Copyright 2016 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 | #ifndef SkCpu_DEFINED |
| 9 | #define SkCpu_DEFINED |
| 10 | |
| 11 | #include "include/core/SkTypes.h" |
| 12 | |
| 13 | struct SkCpu { |
| 14 | enum { |
| 15 | SSE1 = 1 << 0, |
| 16 | SSE2 = 1 << 1, |
| 17 | SSE3 = 1 << 2, |
| 18 | SSSE3 = 1 << 3, |
| 19 | SSE41 = 1 << 4, |
| 20 | SSE42 = 1 << 5, |
| 21 | AVX = 1 << 6, |
| 22 | F16C = 1 << 7, |
| 23 | FMA = 1 << 8, |
| 24 | AVX2 = 1 << 9, |
| 25 | BMI1 = 1 << 10, |
| 26 | BMI2 = 1 << 11, |
| 27 | // Handy alias for all the cool Haswell+ instructions. |
| 28 | HSW = AVX2 | BMI1 | BMI2 | F16C | FMA, |
| 29 | |
| 30 | AVX512F = 1 << 12, |
| 31 | AVX512DQ = 1 << 13, |
| 32 | AVX512IFMA = 1 << 14, |
| 33 | AVX512PF = 1 << 15, |
| 34 | AVX512ER = 1 << 16, |
| 35 | AVX512CD = 1 << 17, |
| 36 | AVX512BW = 1 << 18, |
| 37 | AVX512VL = 1 << 19, |
| 38 | |
| 39 | // Handy alias for all the cool Skylake Xeon+ instructions. |
| 40 | SKX = AVX512F | AVX512DQ | AVX512CD | AVX512BW | AVX512VL, |
| 41 | }; |
| 42 | enum { |
| 43 | NEON = 1 << 0, |
| 44 | NEON_FMA = 1 << 1, |
| 45 | VFP_FP16 = 1 << 2, |
| 46 | CRC32 = 1 << 3, |
| 47 | ASIMDHP = 1 << 4, |
| 48 | }; |
| 49 | |
| 50 | static void CacheRuntimeFeatures(); |
| 51 | static bool Supports(uint32_t); |
| 52 | private: |
| 53 | static uint32_t gCachedFeatures; |
| 54 | }; |
| 55 | |
| 56 | inline bool SkCpu::Supports(uint32_t mask) { |
| 57 | uint32_t features = gCachedFeatures; |
| 58 | |
| 59 | // If we mask in compile-time known lower limits, the compiler can |
| 60 | // often compile away this entire function. |
| 61 | #if SK_CPU_X86 |
| 62 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1 |
| 63 | features |= SSE1; |
| 64 | #endif |
| 65 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
| 66 | features |= SSE2; |
| 67 | #endif |
| 68 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE3 |
| 69 | features |= SSE3; |
| 70 | #endif |
| 71 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 |
| 72 | features |= SSSE3; |
| 73 | #endif |
| 74 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41 |
| 75 | features |= SSE41; |
| 76 | #endif |
| 77 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE42 |
| 78 | features |= SSE42; |
| 79 | #endif |
| 80 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX |
| 81 | features |= AVX; |
| 82 | #endif |
| 83 | // F16C goes here if we add SK_CPU_SSE_LEVEL_F16C |
| 84 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2 |
| 85 | features |= AVX2; |
| 86 | #endif |
| 87 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SKX |
| 88 | features |= (AVX512F | AVX512DQ | AVX512CD | AVX512BW | AVX512VL); |
| 89 | #endif |
| 90 | // FMA doesn't fit neatly into this total ordering. |
| 91 | // It's available on Haswell+ just like AVX2, but it's technically a different bit. |
| 92 | // TODO: circle back on this if we find ourselves limited by lack of compile-time FMA |
| 93 | |
| 94 | #if defined(SK_CPU_LIMIT_AVX) |
| 95 | features &= (SSE1 | SSE2 | SSE3 | SSSE3 | SSE41 | SSE42 | AVX); |
| 96 | #elif defined(SK_CPU_LIMIT_SSE41) |
| 97 | features &= (SSE1 | SSE2 | SSE3 | SSSE3 | SSE41); |
| 98 | #elif defined(SK_CPU_LIMIT_SSE2) |
| 99 | features &= (SSE1 | SSE2); |
| 100 | #endif |
| 101 | |
| 102 | #else |
| 103 | #if defined(SK_ARM_HAS_NEON) |
| 104 | features |= NEON; |
| 105 | #endif |
| 106 | |
| 107 | #if defined(SK_CPU_ARM64) |
| 108 | features |= NEON|NEON_FMA|VFP_FP16; |
| 109 | #endif |
| 110 | |
| 111 | #if defined(SK_ARM_HAS_CRC32) |
| 112 | features |= CRC32; |
| 113 | #endif |
| 114 | |
| 115 | #endif |
| 116 | return (features & mask) == mask; |
| 117 | } |
| 118 | |
| 119 | #endif//SkCpu_DEFINED |
| 120 | |