SkCpu.cpp revision f1b6030b44a4a9523183c3809a165b6b5353fff5
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
114311f016612a814282029daa4bd102053a853d82mtklein#if defined(SK_CPU_X86)
124311f016612a814282029daa4bd102053a853d82mtklein    #if defined(SK_BUILD_FOR_WIN32)
134311f016612a814282029daa4bd102053a853d82mtklein        #include <intrin.h>
144311f016612a814282029daa4bd102053a853d82mtklein        static void cpuid (uint32_t abcd[4]) { __cpuid  ((int*)abcd, 1);    }
154311f016612a814282029daa4bd102053a853d82mtklein        static void cpuid7(uint32_t abcd[4]) { __cpuidex((int*)abcd, 7, 0); }
164311f016612a814282029daa4bd102053a853d82mtklein        static uint64_t xgetbv(uint32_t xcr) { return _xgetbv(xcr); }
174311f016612a814282029daa4bd102053a853d82mtklein    #else
184311f016612a814282029daa4bd102053a853d82mtklein        #include <cpuid.h>
194311f016612a814282029daa4bd102053a853d82mtklein        #if !defined(__cpuid_count)  // Old Mac Clang doesn't have this defined.
204311f016612a814282029daa4bd102053a853d82mtklein            #define  __cpuid_count(eax, ecx, a, b, c, d) \
214311f016612a814282029daa4bd102053a853d82mtklein                __asm__("cpuid" : "=a"(a), "=b"(b), "=c"(c), "=d"(d) : "0"(eax), "2"(ecx))
224311f016612a814282029daa4bd102053a853d82mtklein        #endif
234311f016612a814282029daa4bd102053a853d82mtklein        static void cpuid (uint32_t abcd[4]) { __get_cpuid(1, abcd+0, abcd+1, abcd+2, abcd+3); }
244311f016612a814282029daa4bd102053a853d82mtklein        static void cpuid7(uint32_t abcd[4]) {
254311f016612a814282029daa4bd102053a853d82mtklein            __cpuid_count(7, 0, abcd[0], abcd[1], abcd[2], abcd[3]);
264311f016612a814282029daa4bd102053a853d82mtklein        }
274311f016612a814282029daa4bd102053a853d82mtklein        static uint64_t xgetbv(uint32_t xcr) {
284311f016612a814282029daa4bd102053a853d82mtklein            uint32_t eax, edx;
294311f016612a814282029daa4bd102053a853d82mtklein            __asm__ __volatile__ ( "xgetbv" : "=a"(eax), "=d"(edx) : "c"(xcr));
304311f016612a814282029daa4bd102053a853d82mtklein            return (uint64_t)(edx) << 32 | eax;
314311f016612a814282029daa4bd102053a853d82mtklein        }
324311f016612a814282029daa4bd102053a853d82mtklein    #endif
334311f016612a814282029daa4bd102053a853d82mtklein
344311f016612a814282029daa4bd102053a853d82mtklein    static uint32_t read_cpu_features() {
354311f016612a814282029daa4bd102053a853d82mtklein        uint32_t features = 0;
364311f016612a814282029daa4bd102053a853d82mtklein        uint32_t abcd[4] = {0,0,0,0};
374311f016612a814282029daa4bd102053a853d82mtklein
384311f016612a814282029daa4bd102053a853d82mtklein        // You might want to refer to http://www.sandpile.org/x86/cpuid.htm
394311f016612a814282029daa4bd102053a853d82mtklein
404311f016612a814282029daa4bd102053a853d82mtklein        cpuid(abcd);
414311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[3] & (1<<25)) { features |= SkCpu:: SSE1; }
424311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[3] & (1<<26)) { features |= SkCpu:: SSE2; }
434311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[2] & (1<< 0)) { features |= SkCpu:: SSE3; }
444311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[2] & (1<< 9)) { features |= SkCpu::SSSE3; }
454311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[2] & (1<<19)) { features |= SkCpu::SSE41; }
464311f016612a814282029daa4bd102053a853d82mtklein        if (abcd[2] & (1<<20)) { features |= SkCpu::SSE42; }
474311f016612a814282029daa4bd102053a853d82mtklein
484311f016612a814282029daa4bd102053a853d82mtklein        if ((abcd[2] & (3<<26)) == (3<<26) && (xgetbv(0) & 6) == 6) {  // XSAVE + OSXSAVE
494311f016612a814282029daa4bd102053a853d82mtklein            if (abcd[2] & (1<<28)) { features |= SkCpu:: AVX; }
504311f016612a814282029daa4bd102053a853d82mtklein            if (abcd[2] & (1<<29)) { features |= SkCpu::F16C; }
514311f016612a814282029daa4bd102053a853d82mtklein            if (abcd[2] & (1<<12)) { features |= SkCpu:: FMA; }
524311f016612a814282029daa4bd102053a853d82mtklein
534311f016612a814282029daa4bd102053a853d82mtklein            cpuid7(abcd);
544311f016612a814282029daa4bd102053a853d82mtklein            if (abcd[1] & (1<<5)) { features |= SkCpu::AVX2; }
554311f016612a814282029daa4bd102053a853d82mtklein        }
564311f016612a814282029daa4bd102053a853d82mtklein        return features;
574311f016612a814282029daa4bd102053a853d82mtklein    }
584311f016612a814282029daa4bd102053a853d82mtklein
594311f016612a814282029daa4bd102053a853d82mtklein#elif defined(SK_CPU_ARM32)         && \
604311f016612a814282029daa4bd102053a853d82mtklein      defined(SK_BUILD_FOR_ANDROID) && \
614311f016612a814282029daa4bd102053a853d82mtklein     !defined(SK_BUILD_FOR_ANDROID_FRAMEWORK)
624311f016612a814282029daa4bd102053a853d82mtklein    #include <cpu-features.h>
634311f016612a814282029daa4bd102053a853d82mtklein
644311f016612a814282029daa4bd102053a853d82mtklein    static uint32_t read_cpu_features() {
654311f016612a814282029daa4bd102053a853d82mtklein        uint32_t features = 0;
664311f016612a814282029daa4bd102053a853d82mtklein
674311f016612a814282029daa4bd102053a853d82mtklein        uint64_t android_features = android_getCpuFeatures();
684311f016612a814282029daa4bd102053a853d82mtklein        if (android_features & ANDROID_CPU_ARM_FEATURE_NEON    ) { features |= SkCpu::NEON    ; }
694311f016612a814282029daa4bd102053a853d82mtklein        if (android_features & ANDROID_CPU_ARM_FEATURE_NEON_FMA) { features |= SkCpu::NEON_FMA; }
704311f016612a814282029daa4bd102053a853d82mtklein        if (android_features & ANDROID_CPU_ARM_FEATURE_VFP_FP16) { features |= SkCpu::VFP_FP16; }
714311f016612a814282029daa4bd102053a853d82mtklein        return features;
724311f016612a814282029daa4bd102053a853d82mtklein    }
734311f016612a814282029daa4bd102053a853d82mtklein
74f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein#elif defined(SK_CPU_ARM64)         && \
75f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein      defined(SK_BUILD_FOR_ANDROID) && \
76f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein     !defined(SK_BUILD_FOR_ANDROID_FRAMEWORK)
77f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein    #include <cpu-features.h>
78f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein
79f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein    static uint32_t read_cpu_features() {
80f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein        uint32_t features = 0;
81f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein
82f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein        uint64_t android_features = android_getCpuFeatures();
83f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein        if (android_features & ANDROID_CPU_ARM64_FEATURE_CRC32) { features |= SkCpu::CRC32; }
84f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein        return features;
85f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein    }
86f1b6030b44a4a9523183c3809a165b6b5353fff5mtklein
874311f016612a814282029daa4bd102053a853d82mtklein#else
884311f016612a814282029daa4bd102053a853d82mtklein    static uint32_t read_cpu_features() {
894311f016612a814282029daa4bd102053a853d82mtklein        return 0;
904311f016612a814282029daa4bd102053a853d82mtklein    }
914311f016612a814282029daa4bd102053a853d82mtklein
924311f016612a814282029daa4bd102053a853d82mtklein#endif
934311f016612a814282029daa4bd102053a853d82mtklein
945608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtkleinuint32_t SkCpu::gCachedFeatures = 0;
95eb85fd746d6390f53e250583a0544bf59ed34b35mtklein
965608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtkleinvoid SkCpu::CacheRuntimeFeatures() {
975608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtklein    static SkOnce once;
985608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtklein    once([] { gCachedFeatures = read_cpu_features(); });
995608e2ed2299496eee3c57e0fe426ae9bd0d07a4mtklein}
100