2 * Copyright 2011 The LibYuv Project Authors. All rights reserved.
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.
11 #include "libyuv/cpu_id.h"
13 #if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
14 #include <intrin.h> // For __cpuidex()
16 #if !defined(__pnacl__) && !defined(__CLR_VER) && \
17 !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \
18 defined(_MSC_VER) && !defined(__clang__) && (_MSC_FULL_VER >= 160040219)
19 #include <immintrin.h> // For _xgetbv()
22 #if !defined(__native_client__)
23 #include <stdlib.h> // For getenv()
26 // For ArmCpuCaps() but unittested on all platforms
30 #include "libyuv/basic_types.h" // For CPU_X86
37 // For functions that use the stack and have runtime checks for overflow,
38 // use SAFEBUFFERS to avoid additional check.
39 #if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
40 #define SAFEBUFFERS __declspec(safebuffers)
45 // Low level cpuid for X86.
46 #if (defined(_M_IX86) || defined(_M_X64) || \
47 defined(__i386__) || defined(__x86_64__)) && \
48 !defined(__pnacl__) && !defined(__CLR_VER)
50 void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
51 #if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
52 // Visual C version uses intrinsic or inline x86 assembly.
53 #if (_MSC_FULL_VER >= 160040219)
54 __cpuidex((int*)(cpu_info), info_eax, info_ecx);
55 #elif defined(_M_IX86)
68 __cpuid((int*)(cpu_info), info_eax);
70 cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0;
73 // GCC version uses inline x86 assembly.
74 #else // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
75 uint32 info_ebx, info_edx;
76 asm volatile ( // NOLINT
77 #if defined( __i386__) && defined(__PIC__)
78 // Preserve ebx for fpic 32 bit.
81 "xchg %%edi, %%ebx \n"
86 #endif // defined( __i386__) && defined(__PIC__)
87 "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx));
88 cpu_info[0] = info_eax;
89 cpu_info[1] = info_ebx;
90 cpu_info[2] = info_ecx;
91 cpu_info[3] = info_edx;
92 #endif // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
94 #else // (defined(_M_IX86) || defined(_M_X64) ...
96 void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
97 cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
101 // TODO(fbarchard): Enable xgetbv when validator supports it.
102 #if (defined(_M_IX86) || defined(_M_X64) || \
103 defined(__i386__) || defined(__x86_64__)) && \
104 !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__)
106 // X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
107 int TestOsSaveYmm() {
109 #if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
110 xcr0 = (uint32)(_xgetbv(0)); // VS2010 SP1 required.
111 #elif defined(_M_IX86) && defined(_MSC_VER) && !defined(__clang__)
113 xor ecx, ecx // xcr 0
114 _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0 // For VS2010 and earlier.
117 #elif defined(__i386__) || defined(__x86_64__)
118 asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
119 #endif // defined(__i386__) || defined(__x86_64__)
120 return((xcr0 & 6) == 6); // Is ymm saved?
122 #endif // defined(_M_IX86) || defined(_M_X64) ..
124 // based on libvpx arm_cpudetect.c
125 // For Arm, but public to allow testing on any CPU
126 LIBYUV_API SAFEBUFFERS
127 int ArmCpuCaps(const char* cpuinfo_name) {
128 char cpuinfo_line[512];
129 FILE* f = fopen(cpuinfo_name, "r");
131 // Assume Neon if /proc/cpuinfo is unavailable.
132 // This will occur for Chrome sandbox for Pepper or Render process.
135 while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
136 if (memcmp(cpuinfo_line, "Features", 8) == 0) {
137 char* p = strstr(cpuinfo_line, " neon");
138 if (p && (p[5] == ' ' || p[5] == '\n')) {
142 // aarch64 uses asimd for Neon.
143 p = strstr(cpuinfo_line, " asimd");
144 if (p && (p[6] == ' ' || p[6] == '\n')) {
154 #if defined(__mips__) && defined(__linux__)
155 static int MipsCpuCaps(const char* search_string) {
156 char cpuinfo_line[512];
157 const char* file_name = "/proc/cpuinfo";
158 FILE* f = fopen(file_name, "r");
160 // Assume DSP if /proc/cpuinfo is unavailable.
161 // This will occur for Chrome sandbox for Pepper or Render process.
162 return kCpuHasMIPS_DSP;
164 while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) {
165 if (strstr(cpuinfo_line, search_string) != NULL) {
167 return kCpuHasMIPS_DSP;
175 // CPU detect function for SIMD instruction sets.
177 int cpu_info_ = kCpuInit; // cpu_info is not initialized yet.
179 // Test environment variable for disabling CPU features. Any non-zero value
180 // to disable. Zero ignored to make it easy to set the variable on/off.
181 #if !defined(__native_client__) && !defined(_M_ARM)
183 static LIBYUV_BOOL TestEnv(const char* name) {
184 const char* var = getenv(name);
192 #else // nacl does not support getenv().
193 static LIBYUV_BOOL TestEnv(const char*) {
198 LIBYUV_API SAFEBUFFERS
199 int InitCpuFlags(void) {
200 #if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
202 uint32 cpu_info0[4] = { 0, 0, 0, 0 };
203 uint32 cpu_info1[4] = { 0, 0, 0, 0 };
204 uint32 cpu_info7[4] = { 0, 0, 0, 0 };
205 CpuId(0, 0, cpu_info0);
206 CpuId(1, 0, cpu_info1);
207 if (cpu_info0[0] >= 7) {
208 CpuId(7, 0, cpu_info7);
210 cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
211 ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
212 ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
213 ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
214 ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
215 ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
219 if ((cpu_info1[2] & 0x18000000) == 0x18000000 && // AVX and OSSave
220 TestOsSaveYmm()) { // Saves YMM.
221 cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
225 // Environment variable overrides for testing.
226 if (TestEnv("LIBYUV_DISABLE_X86")) {
227 cpu_info_ &= ~kCpuHasX86;
229 if (TestEnv("LIBYUV_DISABLE_SSE2")) {
230 cpu_info_ &= ~kCpuHasSSE2;
232 if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
233 cpu_info_ &= ~kCpuHasSSSE3;
235 if (TestEnv("LIBYUV_DISABLE_SSE41")) {
236 cpu_info_ &= ~kCpuHasSSE41;
238 if (TestEnv("LIBYUV_DISABLE_SSE42")) {
239 cpu_info_ &= ~kCpuHasSSE42;
241 if (TestEnv("LIBYUV_DISABLE_AVX")) {
242 cpu_info_ &= ~kCpuHasAVX;
244 if (TestEnv("LIBYUV_DISABLE_AVX2")) {
245 cpu_info_ &= ~kCpuHasAVX2;
247 if (TestEnv("LIBYUV_DISABLE_ERMS")) {
248 cpu_info_ &= ~kCpuHasERMS;
250 if (TestEnv("LIBYUV_DISABLE_FMA3")) {
251 cpu_info_ &= ~kCpuHasFMA3;
254 #if defined(__mips__) && defined(__linux__)
255 // Linux mips parse text file for dsp detect.
256 cpu_info_ = MipsCpuCaps("dsp"); // set kCpuHasMIPS_DSP.
257 #if defined(__mips_dspr2)
258 cpu_info_ |= kCpuHasMIPS_DSPR2;
260 cpu_info_ |= kCpuHasMIPS;
262 if (getenv("LIBYUV_DISABLE_MIPS")) {
263 cpu_info_ &= ~kCpuHasMIPS;
265 if (getenv("LIBYUV_DISABLE_MIPS_DSP")) {
266 cpu_info_ &= ~kCpuHasMIPS_DSP;
268 if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
269 cpu_info_ &= ~kCpuHasMIPS_DSPR2;
272 #if defined(__arm__) || defined(__aarch64__)
273 // gcc -mfpu=neon defines __ARM_NEON__
274 // __ARM_NEON__ generates code that requires Neon. NaCL also requires Neon.
275 // For Linux, /proc/cpuinfo can be tested but without that assume Neon.
276 #if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__)
277 cpu_info_ = kCpuHasNEON;
278 // For aarch64(arm64), /proc/cpuinfo's feature is not complete, e.g. no neon
280 // So for aarch64, neon enabling is hard coded here.
282 #if defined(__aarch64__)
283 cpu_info_ = kCpuHasNEON;
285 // Linux arm parse text file for neon detect.
286 cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
288 cpu_info_ |= kCpuHasARM;
289 if (TestEnv("LIBYUV_DISABLE_NEON")) {
290 cpu_info_ &= ~kCpuHasNEON;
293 if (TestEnv("LIBYUV_DISABLE_ASM")) {
300 void MaskCpuFlags(int enable_flags) {
301 cpu_info_ = InitCpuFlags() & enable_flags;
306 } // namespace libyuv