14311f016612a814282029daa4bd102053a853d82mtklein/*
24311f016612a814282029daa4bd102053a853d82mtklein * Copyright 2016 Google Inc.
34311f016612a814282029daa4bd102053a853d82mtklein *
44311f016612a814282029daa4bd102053a853d82mtklein * Use of this source code is governed by a BSD-style license that can be
54311f016612a814282029daa4bd102053a853d82mtklein * found in the LICENSE file.
64311f016612a814282029daa4bd102053a853d82mtklein */
74311f016612a814282029daa4bd102053a853d82mtklein
84311f016612a814282029daa4bd102053a853d82mtklein#include "SkCpu.h"
9eb85fd746d6390f53e250583a0544bf59ed34b35mtklein#include "SkOnce.h"
104311f016612a814282029daa4bd102053a853d82mtklein
114ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein#if !defined(__has_include)
124ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein    #define __has_include(x) 0
134ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein#endif
144ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein
154311f016612a814282029daa4bd102053a853d82mtklein#if defined(SK_CPU_X86)
164311f016612a814282029daa4bd102053a853d82mtklein    #if defined(SK_BUILD_FOR_WIN32)
174311f016612a814282029daa4bd102053a853d82mtklein        #include <intrin.h>
184311f016612a814282029daa4bd102053a853d82mtklein        static void cpuid (uint32_t abcd[4]) { __cpuid  ((int*)abcd, 1);    }
194311f016612a814282029daa4bd102053a853d82mtklein        static void cpuid7(uint32_t abcd[4]) { __cpuidex((int*)abcd, 7, 0); }
204311f016612a814282029daa4bd102053a853d82mtklein        static uint64_t xgetbv(uint32_t xcr) { return _xgetbv(xcr); }
214311f016612a814282029daa4bd102053a853d82mtklein    #else
224311f016612a814282029daa4bd102053a853d82mtklein        #include <cpuid.h>
234311f016612a814282029daa4bd102053a853d82mtklein        #if !defined(__cpuid_count)  // Old Mac Clang doesn't have this defined.
244311f016612a814282029daa4bd102053a853d82mtklein            #define  __cpuid_count(eax, ecx, a, b, c, d) \
254311f016612a814282029daa4bd102053a853d82mtklein                __asm__("cpuid" : "=a"(a), "=b"(b), "=c"(c), "=d"(d) : "0"(eax), "2"(ecx))
264311f016612a814282029daa4bd102053a853d82mtklein        #endif
274311f016612a814282029daa4bd102053a853d82mtklein        static void cpuid (uint32_t abcd[4]) { __get_cpuid(1, abcd+0, abcd+1, abcd+2, abcd+3); }
284311f016612a814282029daa4bd102053a853d82mtklein        static void cpuid7(uint32_t abcd[4]) {
294311f016612a814282029daa4bd102053a853d82mtklein            __cpuid_count(7, 0, abcd[0], abcd[1], abcd[2], abcd[3]);
304311f016612a814282029daa4bd102053a853d82mtklein        }
314311f016612a814282029daa4bd102053a853d82mtklein        static uint64_t xgetbv(uint32_t xcr) {
324311f016612a814282029daa4bd102053a853d82mtklein            uint32_t eax, edx;
334311f016612a814282029daa4bd102053a853d82mtklein            __asm__ __volatile__ ( "xgetbv" : "=a"(eax), "=d"(edx) : "c"(xcr));
344311f016612a814282029daa4bd102053a853d82mtklein            return (uint64_t)(edx) << 32 | eax;
354311f016612a814282029daa4bd102053a853d82mtklein        }
364311f016612a814282029daa4bd102053a853d82mtklein    #endif
374311f016612a814282029daa4bd102053a853d82mtklein
384311f016612a814282029daa4bd102053a853d82mtklein    static uint32_t read_cpu_features() {
394311f016612a814282029daa4bd102053a853d82mtklein        uint32_t features = 0;
404311f016612a814282029daa4bd102053a853d82mtklein        uint32_t abcd[4] = {0,0,0,0};
414311f016612a814282029daa4bd102053a853d82mtklein
424311f016612a814282029daa4bd102053a853d82mtklein        // You might want to refer to http://www.sandpile.org/x86/cpuid.htm
434311f016612a814282029daa4bd102053a853d82mtklein
444311f016612a814282029daa4bd102053a853d82mtklein        cpuid(abcd);
454311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[3] & (1<<25)) { features |= SkCpu:: SSE1; }
464311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[3] & (1<<26)) { features |= SkCpu:: SSE2; }
474311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[2] & (1<< 0)) { features |= SkCpu:: SSE3; }
484311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[2] & (1<< 9)) { features |= SkCpu::SSSE3; }
494311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[2] & (1<<19)) { features |= SkCpu::SSE41; }
504311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[2] & (1<<20)) { features |= SkCpu::SSE42; }
514311f016612a814282029daa4bd102053a853d82mtklein
52c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein        if ((abcd[2] & (3<<26)) == (3<<26)         // XSAVE + OSXSAVE
53c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein             && (xgetbv(0) & (3<<1)) == (3<<1)) {  // XMM and YMM state enabled.
544311f016612a814282029daa4bd102053a853d82mtklein            if (abcd[2] & (1<<28)) { features |= SkCpu:: AVX; }
554311f016612a814282029daa4bd102053a853d82mtklein            if (abcd[2] & (1<<29)) { features |= SkCpu::F16C; }
564311f016612a814282029daa4bd102053a853d82mtklein            if (abcd[2] & (1<<12)) { features |= SkCpu:: FMA; }
574311f016612a814282029daa4bd102053a853d82mtklein
584311f016612a814282029daa4bd102053a853d82mtklein            cpuid7(abcd);
594311f016612a814282029daa4bd102053a853d82mtklein            if (abcd[1] & (1<<5)) { features |= SkCpu::AVX2; }
6078d5a3bac5cbde50cd12d8b9ab6dd269324b5272Mike Klein            if (abcd[1] & (1<<3)) { features |= SkCpu::BMI1; }
6178d5a3bac5cbde50cd12d8b9ab6dd269324b5272Mike Klein            if (abcd[1] & (1<<8)) { features |= SkCpu::BMI2; }
62c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein
63c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein            if ((xgetbv(0) & (7<<5)) == (7<<5)) {  // All ZMM state bits enabled too.
64c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein                if (abcd[1] & (1<<16)) { features |= SkCpu::AVX512F; }
65c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein                if (abcd[1] & (1<<17)) { features |= SkCpu::AVX512DQ; }
66c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein                if (abcd[1] & (1<<21)) { features |= SkCpu::AVX512IFMA; }
67c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein                if (abcd[1] & (1<<26)) { features |= SkCpu::AVX512PF; }
68c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein                if (abcd[1] & (1<<27)) { features |= SkCpu::AVX512ER; }
69c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein                if (abcd[1] & (1<<28)) { features |= SkCpu::AVX512CD; }
70c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein                if (abcd[1] & (1<<30)) { features |= SkCpu::AVX512BW; }
71c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein                if (abcd[1] & (1<<31)) { features |= SkCpu::AVX512VL; }
72c6a449d6bf80c5bb9e02aeaed049a99870e8b1e8Mike Klein            }
734311f016612a814282029daa4bd102053a853d82mtklein        }
744311f016612a814282029daa4bd102053a853d82mtklein        return features;
754311f016612a814282029daa4bd102053a853d82mtklein    }
764311f016612a814282029daa4bd102053a853d82mtklein
7730ec0b3735d5f728c2aea4184736a3e286a5ccdaMike Klein#elif defined(SK_CPU_ARM64) && __has_include(<asm/hwcap.h>) && __has_include(<sys/auxv.h>)
78f44703a87f532b3f593d91605d66d52c6bbc45c9Mike Klein    #include <asm/hwcap.h>
79f44703a87f532b3f593d91605d66d52c6bbc45c9Mike Klein    #include <sys/auxv.h>
804311f016612a814282029daa4bd102053a853d82mtklein
814311f016612a814282029daa4bd102053a853d82mtklein    static uint32_t read_cpu_features() {
824311f016612a814282029daa4bd102053a853d82mtklein        uint32_t features = 0;
83f44703a87f532b3f593d91605d66d52c6bbc45c9Mike Klein        uint32_t hwcaps = getauxval(AT_HWCAP);
84f44703a87f532b3f593d91605d66d52c6bbc45c9Mike Klein        if (hwcaps & HWCAP_CRC32) { features |= SkCpu::CRC32; }
85f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein        return features;
86f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein    }
87f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein
8830ec0b3735d5f728c2aea4184736a3e286a5ccdaMike Klein#elif defined(SK_CPU_ARM32) && __has_include(<asm/hwcap.h>) && __has_include(<sys/auxv.h>)
8930ec0b3735d5f728c2aea4184736a3e286a5ccdaMike Klein    // asm/hwcap.h and sys/auxv.h won't be present on NDK builds before API v21.
904ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein    #include <asm/hwcap.h>
914ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein    #include <sys/auxv.h>
924ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein
934ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein    static uint32_t read_cpu_features() {
944ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        uint32_t features = 0;
954ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        uint32_t hwcaps = getauxval(AT_HWCAP);
964ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        if (hwcaps & HWCAP_VFPv4) { features |= SkCpu::NEON|SkCpu::NEON_FMA|SkCpu::VFP_FP16; }
974ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        return features;
984ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein    }
994ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein
10030ec0b3735d5f728c2aea4184736a3e286a5ccdaMike Klein#elif defined(SK_CPU_ARM32) && __has_include(<cpu-features.h>)
1014ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein    #include <cpu-features.h>
1024ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein
1034ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein    static uint32_t read_cpu_features() {
1044ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        uint32_t features = 0;
1054ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        uint64_t cpu_features = android_getCpuFeatures();
1064ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        if (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON)     { features |= SkCpu::NEON; }
1074ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        if (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON_FMA) { features |= SkCpu::NEON_FMA; }
1084ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        if (cpu_features & ANDROID_CPU_ARM_FEATURE_VFP_FP16) { features |= SkCpu::VFP_FP16; }
1094ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein        return features;
1104ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein    }
1114ef8cb3527b7e3f453dccd39eea76e31eb2c33c7Mike Klein
1124311f016612a814282029daa4bd102053a853d82mtklein#else
1134311f016612a814282029daa4bd102053a853d82mtklein    static uint32_t read_cpu_features() {
1144311f016612a814282029daa4bd102053a853d82mtklein        return 0;
1154311f016612a814282029daa4bd102053a853d82mtklein    }
1164311f016612a814282029daa4bd102053a853d82mtklein
1174311f016612a814282029daa4bd102053a853d82mtklein#endif
1184311f016612a814282029daa4bd102053a853d82mtklein
1195608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtkleinuint32_t SkCpu::gCachedFeatures = 0;
120eb85fd746d6390f53e250583a0544bf59ed34b35mtklein
1215608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtkleinvoid SkCpu::CacheRuntimeFeatures() {
1225608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtklein    static SkOnce once;
1235608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtklein    once([] { gCachedFeatures = read_cpu_features(); });
1245608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtklein}
125