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 "include/core/SkStream.h"
9 #include "include/core/SkString.h"
10 #include "include/private/SkOnce.h"
11 #include "src/core/SkCpu.h"
12
13 #if defined(SK_CPU_X86)
14 #if defined(SK_BUILD_FOR_WIN)
15 #include <intrin.h>
cpuid(uint32_t abcd[4])16 static void cpuid (uint32_t abcd[4]) { __cpuid ((int*)abcd, 1); }
cpuid7(uint32_t abcd[4])17 static void cpuid7(uint32_t abcd[4]) { __cpuidex((int*)abcd, 7, 0); }
xgetbv(uint32_t xcr)18 static uint64_t xgetbv(uint32_t xcr) {
19 #if defined(_MSC_VER)
20 return _xgetbv(xcr);
21 #else
22 uint32_t eax, edx;
23
24 __asm__ volatile (
25 "xgetbv" : "=a"(eax), "=d"(edx) : "c"(xcr));
26 return (static_cast<uint64_t>(edx) << 32) | eax;
27 #endif // defined(_MSC_VER)
28 }
29 #else
30 #include <cpuid.h>
31 #if !defined(__cpuid_count) // Old Mac Clang doesn't have this defined.
32 #define __cpuid_count(eax, ecx, a, b, c, d) \
33 __asm__("cpuid" : "=a"(a), "=b"(b), "=c"(c), "=d"(d) : "0"(eax), "2"(ecx))
34 #endif
cpuid(uint32_t abcd[4])35 static void cpuid (uint32_t abcd[4]) { __get_cpuid(1, abcd+0, abcd+1, abcd+2, abcd+3); }
cpuid7(uint32_t abcd[4])36 static void cpuid7(uint32_t abcd[4]) {
37 __cpuid_count(7, 0, abcd[0], abcd[1], abcd[2], abcd[3]);
38 }
xgetbv(uint32_t xcr)39 static uint64_t xgetbv(uint32_t xcr) {
40 uint32_t eax, edx;
41 __asm__ __volatile__ ( "xgetbv" : "=a"(eax), "=d"(edx) : "c"(xcr));
42 return (uint64_t)(edx) << 32 | eax;
43 }
44 #endif
45
read_cpu_features()46 static uint32_t read_cpu_features() {
47 uint32_t features = 0;
48 uint32_t abcd[4] = {0,0,0,0};
49
50 // You might want to refer to http://www.sandpile.org/x86/cpuid.htm
51
52 cpuid(abcd);
53 if (abcd[3] & (1<<25)) { features |= SkCpu:: SSE1; }
54 if (abcd[3] & (1<<26)) { features |= SkCpu:: SSE2; }
55 if (abcd[2] & (1<< 0)) { features |= SkCpu:: SSE3; }
56 if (abcd[2] & (1<< 9)) { features |= SkCpu::SSSE3; }
57 if (abcd[2] & (1<<19)) { features |= SkCpu::SSE41; }
58 if (abcd[2] & (1<<20)) { features |= SkCpu::SSE42; }
59
60 if ((abcd[2] & (3<<26)) == (3<<26) // XSAVE + OSXSAVE
61 && (xgetbv(0) & (3<<1)) == (3<<1)) { // XMM and YMM state enabled.
62 if (abcd[2] & (1<<28)) { features |= SkCpu:: AVX; }
63 if (abcd[2] & (1<<29)) { features |= SkCpu::F16C; }
64 if (abcd[2] & (1<<12)) { features |= SkCpu:: FMA; }
65
66 cpuid7(abcd);
67 if (abcd[1] & (1<<5)) { features |= SkCpu::AVX2; }
68 if (abcd[1] & (1<<3)) { features |= SkCpu::BMI1; }
69 if (abcd[1] & (1<<8)) { features |= SkCpu::BMI2; }
70
71 if ((xgetbv(0) & (7<<5)) == (7<<5)) { // All ZMM state bits enabled too.
72 if (abcd[1] & (1<<16)) { features |= SkCpu::AVX512F; }
73 if (abcd[1] & (1<<17)) { features |= SkCpu::AVX512DQ; }
74 if (abcd[1] & (1<<21)) { features |= SkCpu::AVX512IFMA; }
75 if (abcd[1] & (1<<26)) { features |= SkCpu::AVX512PF; }
76 if (abcd[1] & (1<<27)) { features |= SkCpu::AVX512ER; }
77 if (abcd[1] & (1<<28)) { features |= SkCpu::AVX512CD; }
78 if (abcd[1] & (1<<30)) { features |= SkCpu::AVX512BW; }
79 if (abcd[1] & (1<<31)) { features |= SkCpu::AVX512VL; }
80 }
81 }
82 return features;
83 }
84
85 #elif defined(SK_CPU_ARM64) && __has_include(<sys/auxv.h>)
86 #include <sys/auxv.h>
87
read_cpu_features()88 static uint32_t read_cpu_features() {
89 const uint32_t kHWCAP_CRC32 = (1<< 7),
90 kHWCAP_ASIMDHP = (1<<10);
91
92 uint32_t features = 0;
93 uint32_t hwcaps = getauxval(AT_HWCAP);
94 if (hwcaps & kHWCAP_CRC32 ) { features |= SkCpu::CRC32; }
95 if (hwcaps & kHWCAP_ASIMDHP) { features |= SkCpu::ASIMDHP; }
96
97 // The Samsung Mongoose 3 core sets the ASIMDHP bit but doesn't support it.
98 for (int core = 0; features & SkCpu::ASIMDHP; core++) {
99 // These /sys files contain the core's MIDR_EL1 register, the source of
100 // CPU {implementer, variant, part, revision} you'd see in /proc/cpuinfo.
101 SkString path =
102 SkStringPrintf("/sys/devices/system/cpu/cpu%d/regs/identification/midr_el1", core);
103
104 // Can't use SkData::MakeFromFileName() here, I think because /sys can't be mmap()'d.
105 SkFILEStream midr_el1(path.c_str());
106 if (!midr_el1.isValid()) {
107 // This is our ordinary exit path.
108 // If we ask for MIDR_EL1 from a core that doesn't exist, we've checked all cores.
109 if (core == 0) {
110 // On the other hand, if we can't read MIDR_EL1 from any core, assume the worst.
111 features &= ~(SkCpu::ASIMDHP);
112 }
113 break;
114 }
115
116 const char kMongoose3[] = "0x00000000531f0020"; // 53 == Samsung.
117 char buf[SK_ARRAY_COUNT(kMongoose3) - 1]; // No need for the terminating \0.
118
119 if (SK_ARRAY_COUNT(buf) != midr_el1.read(buf, SK_ARRAY_COUNT(buf))
120 || 0 == memcmp(kMongoose3, buf, SK_ARRAY_COUNT(buf))) {
121 features &= ~(SkCpu::ASIMDHP);
122 }
123 }
124 return features;
125 }
126
127 #elif defined(SK_CPU_ARM32) && __has_include(<sys/auxv.h>) && \
128 (!defined(__ANDROID_API__) || __ANDROID_API__ >= 18)
129 // sys/auxv.h will always be present in the Android NDK due to unified
130 //headers, but getauxval is only defined for API >= 18.
131 #include <sys/auxv.h>
132
read_cpu_features()133 static uint32_t read_cpu_features() {
134 const uint32_t kHWCAP_NEON = (1<<12);
135 const uint32_t kHWCAP_VFPv4 = (1<<16);
136
137 uint32_t features = 0;
138 uint32_t hwcaps = getauxval(AT_HWCAP);
139 if (hwcaps & kHWCAP_NEON ) {
140 features |= SkCpu::NEON;
141 if (hwcaps & kHWCAP_VFPv4) { features |= SkCpu::NEON_FMA|SkCpu::VFP_FP16; }
142 }
143 return features;
144 }
145
146 #elif defined(SK_CPU_ARM32) && __has_include(<cpu-features.h>)
147 #include <cpu-features.h>
148
read_cpu_features()149 static uint32_t read_cpu_features() {
150 uint32_t features = 0;
151 uint64_t cpu_features = android_getCpuFeatures();
152 if (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON) { features |= SkCpu::NEON; }
153 if (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON_FMA) { features |= SkCpu::NEON_FMA; }
154 if (cpu_features & ANDROID_CPU_ARM_FEATURE_VFP_FP16) { features |= SkCpu::VFP_FP16; }
155 return features;
156 }
157
158 #else
read_cpu_features()159 static uint32_t read_cpu_features() {
160 return 0;
161 }
162
163 #endif
164
165 uint32_t SkCpu::gCachedFeatures = 0;
166
CacheRuntimeFeatures()167 void SkCpu::CacheRuntimeFeatures() {
168 static SkOnce once;
169 once([] { gCachedFeatures = read_cpu_features(); });
170 }
171