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#include "SkCpu.h"
9#include "SkOnce.h"
10
11#if !defined(__has_include)
12    #define __has_include(x) 0
13#endif
14
15#if defined(SK_CPU_X86)
16    #if defined(SK_BUILD_FOR_WIN32)
17        #include <intrin.h>
18        static void cpuid (uint32_t abcd[4]) { __cpuid  ((int*)abcd, 1);    }
19        static void cpuid7(uint32_t abcd[4]) { __cpuidex((int*)abcd, 7, 0); }
20        static uint64_t xgetbv(uint32_t xcr) { return _xgetbv(xcr); }
21    #else
22        #include <cpuid.h>
23        #if !defined(__cpuid_count)  // Old Mac Clang doesn't have this defined.
24            #define  __cpuid_count(eax, ecx, a, b, c, d) \
25                __asm__("cpuid" : "=a"(a), "=b"(b), "=c"(c), "=d"(d) : "0"(eax), "2"(ecx))
26        #endif
27        static void cpuid (uint32_t abcd[4]) { __get_cpuid(1, abcd+0, abcd+1, abcd+2, abcd+3); }
28        static void cpuid7(uint32_t abcd[4]) {
29            __cpuid_count(7, 0, abcd[0], abcd[1], abcd[2], abcd[3]);
30        }
31        static uint64_t xgetbv(uint32_t xcr) {
32            uint32_t eax, edx;
33            __asm__ __volatile__ ( "xgetbv" : "=a"(eax), "=d"(edx) : "c"(xcr));
34            return (uint64_t)(edx) << 32 | eax;
35        }
36    #endif
37
38    static uint32_t read_cpu_features() {
39        uint32_t features = 0;
40        uint32_t abcd[4] = {0,0,0,0};
41
42        // You might want to refer to http://www.sandpile.org/x86/cpuid.htm
43
44        cpuid(abcd);
45        if (abcd[3] & (1<<25)) { features |= SkCpu:: SSE1; }
46        if (abcd[3] & (1<<26)) { features |= SkCpu:: SSE2; }
47        if (abcd[2] & (1<< 0)) { features |= SkCpu:: SSE3; }
48        if (abcd[2] & (1<< 9)) { features |= SkCpu::SSSE3; }
49        if (abcd[2] & (1<<19)) { features |= SkCpu::SSE41; }
50        if (abcd[2] & (1<<20)) { features |= SkCpu::SSE42; }
51
52        if ((abcd[2] & (3<<26)) == (3<<26)         // XSAVE + OSXSAVE
53             && (xgetbv(0) & (3<<1)) == (3<<1)) {  // XMM and YMM state enabled.
54            if (abcd[2] & (1<<28)) { features |= SkCpu:: AVX; }
55            if (abcd[2] & (1<<29)) { features |= SkCpu::F16C; }
56            if (abcd[2] & (1<<12)) { features |= SkCpu:: FMA; }
57
58            cpuid7(abcd);
59            if (abcd[1] & (1<<5)) { features |= SkCpu::AVX2; }
60            if (abcd[1] & (1<<3)) { features |= SkCpu::BMI1; }
61            if (abcd[1] & (1<<8)) { features |= SkCpu::BMI2; }
62
63            if ((xgetbv(0) & (7<<5)) == (7<<5)) {  // All ZMM state bits enabled too.
64                if (abcd[1] & (1<<16)) { features |= SkCpu::AVX512F; }
65                if (abcd[1] & (1<<17)) { features |= SkCpu::AVX512DQ; }
66                if (abcd[1] & (1<<21)) { features |= SkCpu::AVX512IFMA; }
67                if (abcd[1] & (1<<26)) { features |= SkCpu::AVX512PF; }
68                if (abcd[1] & (1<<27)) { features |= SkCpu::AVX512ER; }
69                if (abcd[1] & (1<<28)) { features |= SkCpu::AVX512CD; }
70                if (abcd[1] & (1<<30)) { features |= SkCpu::AVX512BW; }
71                if (abcd[1] & (1<<31)) { features |= SkCpu::AVX512VL; }
72            }
73        }
74        return features;
75    }
76
77#elif defined(SK_CPU_ARM64) && __has_include(<asm/hwcap.h>) && __has_include(<sys/auxv.h>)
78    #include <asm/hwcap.h>
79    #include <sys/auxv.h>
80
81    static uint32_t read_cpu_features() {
82        uint32_t features = 0;
83        uint32_t hwcaps = getauxval(AT_HWCAP);
84        if (hwcaps & HWCAP_CRC32) { features |= SkCpu::CRC32; }
85        return features;
86    }
87
88#elif defined(SK_CPU_ARM32) && __has_include(<asm/hwcap.h>) && __has_include(<sys/auxv.h>)
89    // asm/hwcap.h and sys/auxv.h won't be present on NDK builds before API v21.
90    #include <asm/hwcap.h>
91    #include <sys/auxv.h>
92
93    static uint32_t read_cpu_features() {
94        uint32_t features = 0;
95        uint32_t hwcaps = getauxval(AT_HWCAP);
96        if (hwcaps & HWCAP_VFPv4) { features |= SkCpu::NEON|SkCpu::NEON_FMA|SkCpu::VFP_FP16; }
97        return features;
98    }
99
100#elif defined(SK_CPU_ARM32) && __has_include(<cpu-features.h>)
101    #include <cpu-features.h>
102
103    static uint32_t read_cpu_features() {
104        uint32_t features = 0;
105        uint64_t cpu_features = android_getCpuFeatures();
106        if (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON)     { features |= SkCpu::NEON; }
107        if (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON_FMA) { features |= SkCpu::NEON_FMA; }
108        if (cpu_features & ANDROID_CPU_ARM_FEATURE_VFP_FP16) { features |= SkCpu::VFP_FP16; }
109        return features;
110    }
111
112#else
113    static uint32_t read_cpu_features() {
114        return 0;
115    }
116
117#endif
118
119uint32_t SkCpu::gCachedFeatures = 0;
120
121void SkCpu::CacheRuntimeFeatures() {
122    static SkOnce once;
123    once([] { gCachedFeatures = read_cpu_features(); });
124}
125