1 /* 2 * Copyright 2011 The LibYuv Project Authors. All rights reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 #include "libyuv/cpu_id.h" 12 13 #if defined(_MSC_VER) 14 #include <intrin.h> // For __cpuidex() 15 #endif 16 #if !defined(__pnacl__) && !defined(__CLR_VER) && \ 17 !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \ 18 defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) 19 #include <immintrin.h> // For _xgetbv() 20 #endif 21 22 #if !defined(__native_client__) 23 #include <stdlib.h> // For getenv() 24 #endif 25 26 // For ArmCpuCaps() but unittested on all platforms 27 #include <stdio.h> 28 #include <string.h> 29 30 #include "libyuv/basic_types.h" // For CPU_X86 31 32 #ifdef __cplusplus 33 namespace libyuv { 34 extern "C" { 35 #endif 36 37 // For functions that use the stack and have runtime checks for overflow, 38 // use SAFEBUFFERS to avoid additional check. 39 #if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) && \ 40 !defined(__clang__) 41 #define SAFEBUFFERS __declspec(safebuffers) 42 #else 43 #define SAFEBUFFERS 44 #endif 45 46 // Low level cpuid for X86. 47 #if (defined(_M_IX86) || defined(_M_X64) || \ 48 defined(__i386__) || defined(__x86_64__)) && \ 49 !defined(__pnacl__) && !defined(__CLR_VER) 50 LIBYUV_API 51 void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) { 52 #if defined(_MSC_VER) 53 // Visual C version uses intrinsic or inline x86 assembly. 54 #if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) 55 __cpuidex((int*)(cpu_info), info_eax, info_ecx); 56 #elif defined(_M_IX86) 57 __asm { 58 mov eax, info_eax 59 mov ecx, info_ecx 60 mov edi, cpu_info 61 cpuid 62 mov [edi], eax 63 mov [edi + 4], ebx 64 mov [edi + 8], ecx 65 mov [edi + 12], edx 66 } 67 #else // Visual C but not x86 68 if (info_ecx == 0) { 69 __cpuid((int*)(cpu_info), info_eax); 70 } else { 71 cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0; 72 } 73 #endif 74 // GCC version uses inline x86 assembly. 75 #else // defined(_MSC_VER) 76 uint32 info_ebx, info_edx; 77 asm volatile ( 78 #if defined( __i386__) && defined(__PIC__) 79 // Preserve ebx for fpic 32 bit. 80 "mov %%ebx, %%edi \n" 81 "cpuid \n" 82 "xchg %%edi, %%ebx \n" 83 : "=D" (info_ebx), 84 #else 85 "cpuid \n" 86 : "=b" (info_ebx), 87 #endif // defined( __i386__) && defined(__PIC__) 88 "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx)); 89 cpu_info[0] = info_eax; 90 cpu_info[1] = info_ebx; 91 cpu_info[2] = info_ecx; 92 cpu_info[3] = info_edx; 93 #endif // defined(_MSC_VER) 94 } 95 #else // (defined(_M_IX86) || defined(_M_X64) ... 96 LIBYUV_API 97 void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) { 98 cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0; 99 } 100 #endif 101 102 // For VS2010 and earlier emit can be used: 103 // _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0 // For VS2010 and earlier. 104 // __asm { 105 // xor ecx, ecx // xcr 0 106 // xgetbv 107 // mov xcr0, eax 108 // } 109 // For VS2013 and earlier 32 bit, the _xgetbv(0) optimizer produces bad code. 110 // https://code.google.com/p/libyuv/issues/detail?id=529 111 #if defined(_M_IX86) && (_MSC_VER < 1900) 112 #pragma optimize("g", off) 113 #endif 114 #if (defined(_M_IX86) || defined(_M_X64) || \ 115 defined(__i386__) || defined(__x86_64__)) && \ 116 !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__) 117 #define HAS_XGETBV 118 // X86 CPUs have xgetbv to detect OS saves high parts of ymm registers. 119 int GetXCR0() { 120 uint32 xcr0 = 0u; 121 #if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) 122 xcr0 = (uint32)(_xgetbv(0)); // VS2010 SP1 required. 123 #elif defined(__i386__) || defined(__x86_64__) 124 asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx"); 125 #endif // defined(__i386__) || defined(__x86_64__) 126 return xcr0; 127 } 128 #endif // defined(_M_IX86) || defined(_M_X64) .. 129 // Return optimization to previous setting. 130 #if defined(_M_IX86) && (_MSC_VER < 1900) 131 #pragma optimize("g", on) 132 #endif 133 134 // based on libvpx arm_cpudetect.c 135 // For Arm, but public to allow testing on any CPU 136 LIBYUV_API SAFEBUFFERS 137 int ArmCpuCaps(const char* cpuinfo_name) { 138 char cpuinfo_line[512]; 139 FILE* f = fopen(cpuinfo_name, "r"); 140 if (!f) { 141 // Assume Neon if /proc/cpuinfo is unavailable. 142 // This will occur for Chrome sandbox for Pepper or Render process. 143 return kCpuHasNEON; 144 } 145 while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) { 146 if (memcmp(cpuinfo_line, "Features", 8) == 0) { 147 char* p = strstr(cpuinfo_line, " neon"); 148 if (p && (p[5] == ' ' || p[5] == '\n')) { 149 fclose(f); 150 return kCpuHasNEON; 151 } 152 // aarch64 uses asimd for Neon. 153 p = strstr(cpuinfo_line, " asimd"); 154 if (p && (p[6] == ' ' || p[6] == '\n')) { 155 fclose(f); 156 return kCpuHasNEON; 157 } 158 } 159 } 160 fclose(f); 161 return 0; 162 } 163 164 // CPU detect function for SIMD instruction sets. 165 LIBYUV_API 166 int cpu_info_ = 0; // cpu_info is not initialized yet. 167 168 // Test environment variable for disabling CPU features. Any non-zero value 169 // to disable. Zero ignored to make it easy to set the variable on/off. 170 #if !defined(__native_client__) && !defined(_M_ARM) 171 172 static LIBYUV_BOOL TestEnv(const char* name) { 173 const char* var = getenv(name); 174 if (var) { 175 if (var[0] != '0') { 176 return LIBYUV_TRUE; 177 } 178 } 179 return LIBYUV_FALSE; 180 } 181 #else // nacl does not support getenv(). 182 static LIBYUV_BOOL TestEnv(const char*) { 183 return LIBYUV_FALSE; 184 } 185 #endif 186 187 LIBYUV_API SAFEBUFFERS 188 int InitCpuFlags(void) { 189 // TODO(fbarchard): swap kCpuInit logic so 0 means uninitialized. 190 int cpu_info = 0; 191 #if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86) 192 uint32 cpu_info0[4] = { 0, 0, 0, 0 }; 193 uint32 cpu_info1[4] = { 0, 0, 0, 0 }; 194 uint32 cpu_info7[4] = { 0, 0, 0, 0 }; 195 CpuId(0, 0, cpu_info0); 196 CpuId(1, 0, cpu_info1); 197 if (cpu_info0[0] >= 7) { 198 CpuId(7, 0, cpu_info7); 199 } 200 cpu_info = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) | 201 ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) | 202 ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) | 203 ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) | 204 ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) | 205 ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) | 206 kCpuHasX86; 207 208 #ifdef HAS_XGETBV 209 // AVX requires CPU has AVX, XSAVE and OSXSave for xgetbv 210 if (((cpu_info1[2] & 0x1c000000) == 0x1c000000) && // AVX and OSXSave 211 ((GetXCR0() & 6) == 6)) { // Test OS saves YMM registers 212 cpu_info |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) | kCpuHasAVX; 213 214 // Detect AVX512bw 215 if ((GetXCR0() & 0xe0) == 0xe0) { 216 cpu_info |= (cpu_info7[1] & 0x40000000) ? kCpuHasAVX3 : 0; 217 } 218 } 219 #endif 220 221 // Environment variable overrides for testing. 222 if (TestEnv("LIBYUV_DISABLE_X86")) { 223 cpu_info &= ~kCpuHasX86; 224 } 225 if (TestEnv("LIBYUV_DISABLE_SSE2")) { 226 cpu_info &= ~kCpuHasSSE2; 227 } 228 if (TestEnv("LIBYUV_DISABLE_SSSE3")) { 229 cpu_info &= ~kCpuHasSSSE3; 230 } 231 if (TestEnv("LIBYUV_DISABLE_SSE41")) { 232 cpu_info &= ~kCpuHasSSE41; 233 } 234 if (TestEnv("LIBYUV_DISABLE_SSE42")) { 235 cpu_info &= ~kCpuHasSSE42; 236 } 237 if (TestEnv("LIBYUV_DISABLE_AVX")) { 238 cpu_info &= ~kCpuHasAVX; 239 } 240 if (TestEnv("LIBYUV_DISABLE_AVX2")) { 241 cpu_info &= ~kCpuHasAVX2; 242 } 243 if (TestEnv("LIBYUV_DISABLE_ERMS")) { 244 cpu_info &= ~kCpuHasERMS; 245 } 246 if (TestEnv("LIBYUV_DISABLE_FMA3")) { 247 cpu_info &= ~kCpuHasFMA3; 248 } 249 if (TestEnv("LIBYUV_DISABLE_AVX3")) { 250 cpu_info &= ~kCpuHasAVX3; 251 } 252 #endif 253 #if defined(__mips__) && defined(__linux__) 254 #if defined(__mips_dspr2) 255 cpu_info |= kCpuHasDSPR2; 256 #endif 257 cpu_info |= kCpuHasMIPS; 258 if (getenv("LIBYUV_DISABLE_DSPR2")) { 259 cpu_info &= ~kCpuHasDSPR2; 260 } 261 #endif 262 #if defined(__arm__) || defined(__aarch64__) 263 // gcc -mfpu=neon defines __ARM_NEON__ 264 // __ARM_NEON__ generates code that requires Neon. NaCL also requires Neon. 265 // For Linux, /proc/cpuinfo can be tested but without that assume Neon. 266 #if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__) 267 cpu_info = kCpuHasNEON; 268 // For aarch64(arm64), /proc/cpuinfo's feature is not complete, e.g. no neon 269 // flag in it. 270 // So for aarch64, neon enabling is hard coded here. 271 #endif 272 #if defined(__aarch64__) 273 cpu_info = kCpuHasNEON; 274 #else 275 // Linux arm parse text file for neon detect. 276 cpu_info = ArmCpuCaps("/proc/cpuinfo"); 277 #endif 278 cpu_info |= kCpuHasARM; 279 if (TestEnv("LIBYUV_DISABLE_NEON")) { 280 cpu_info &= ~kCpuHasNEON; 281 } 282 #endif // __arm__ 283 if (TestEnv("LIBYUV_DISABLE_ASM")) { 284 cpu_info = 0; 285 } 286 cpu_info |= kCpuInitialized; 287 cpu_info_ = cpu_info; 288 return cpu_info; 289 } 290 291 // Note that use of this function is not thread safe. 292 LIBYUV_API 293 void MaskCpuFlags(int enable_flags) { 294 cpu_info_ = InitCpuFlags() & enable_flags; 295 } 296 297 #ifdef __cplusplus 298 } // extern "C" 299 } // namespace libyuv 300 #endif 301