michael@0: /*
michael@0:  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS. All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: #include "libyuv/cpu_id.h"
michael@0: 
michael@0: #ifdef _MSC_VER
michael@0: #include <intrin.h>  // For __cpuidex()
michael@0: #endif
michael@0: #if !defined(__pnacl__) && !defined(__CLR_VER) && \
michael@0:     !defined(__native_client__) && defined(_M_X64) && \
michael@0:     defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
michael@0: #include <immintrin.h>  // For _xgetbv()
michael@0: #endif
michael@0: 
michael@0: #if !defined(__native_client__)
michael@0: #include <stdlib.h>  // For getenv()
michael@0: #endif
michael@0: 
michael@0: // For ArmCpuCaps() but unittested on all platforms
michael@0: #include <stdio.h>
michael@0: #include <string.h>
michael@0: 
michael@0: #include "libyuv/basic_types.h"  // For CPU_X86
michael@0: 
michael@0: #ifdef __cplusplus
michael@0: namespace libyuv {
michael@0: extern "C" {
michael@0: #endif
michael@0: 
michael@0: // For functions that use the stack and have runtime checks for overflow,
michael@0: // use SAFEBUFFERS to avoid additional check.
michael@0: #if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
michael@0: #define SAFEBUFFERS __declspec(safebuffers)
michael@0: #else
michael@0: #define SAFEBUFFERS
michael@0: #endif
michael@0: 
michael@0: // Low level cpuid for X86. Returns zeros on other CPUs.
michael@0: #if !defined(__pnacl__) && !defined(__CLR_VER) && \
michael@0:     (defined(_M_IX86) || defined(_M_X64) || \
michael@0:     defined(__i386__) || defined(__x86_64__))
michael@0: LIBYUV_API
michael@0: void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
michael@0: #if defined(_MSC_VER)
michael@0: #if (_MSC_FULL_VER >= 160040219)
michael@0:   __cpuidex((int*)(cpu_info), info_eax, info_ecx);
michael@0: #elif defined(_M_IX86)
michael@0:   __asm {
michael@0:     mov        eax, info_eax
michael@0:     mov        ecx, info_ecx
michael@0:     mov        edi, cpu_info
michael@0:     cpuid
michael@0:     mov        [edi], eax
michael@0:     mov        [edi + 4], ebx
michael@0:     mov        [edi + 8], ecx
michael@0:     mov        [edi + 12], edx
michael@0:   }
michael@0: #else
michael@0:   if (info_ecx == 0) {
michael@0:     __cpuid((int*)(cpu_info), info_eax);
michael@0:   } else {
michael@0:     cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0;
michael@0:   }
michael@0: #endif
michael@0: #else  // defined(_MSC_VER)
michael@0:   uint32 info_ebx, info_edx;
michael@0:   asm volatile (  // NOLINT
michael@0: #if defined( __i386__) && defined(__PIC__)
michael@0:     // Preserve ebx for fpic 32 bit.
michael@0:     "mov %%ebx, %%edi                          \n"
michael@0:     "cpuid                                     \n"
michael@0:     "xchg %%edi, %%ebx                         \n"
michael@0:     : "=D" (info_ebx),
michael@0: #else
michael@0:     "cpuid                                     \n"
michael@0:     : "=b" (info_ebx),
michael@0: #endif  //  defined( __i386__) && defined(__PIC__)
michael@0:       "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx));
michael@0:   cpu_info[0] = info_eax;
michael@0:   cpu_info[1] = info_ebx;
michael@0:   cpu_info[2] = info_ecx;
michael@0:   cpu_info[3] = info_edx;
michael@0: #endif  // defined(_MSC_VER)
michael@0: }
michael@0: 
michael@0: #if !defined(__native_client__)
michael@0: #define HAS_XGETBV
michael@0: // X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
michael@0: int TestOsSaveYmm() {
michael@0:   uint32 xcr0 = 0u;
michael@0: #if defined(_MSC_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
michael@0:   xcr0 = (uint32)(_xgetbv(_XCR_XFEATURE_ENABLED_MASK));
michael@0: #elif defined(_MSC_VER) && defined(_M_IX86)
michael@0:   __asm {
michael@0:     xor        ecx, ecx    // xcr 0
michael@0:     _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
michael@0:     mov        xcr0, eax
michael@0:   }
michael@0: #elif defined(__i386__) || defined(__x86_64__)
michael@0:   asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
michael@0: #endif  // defined(_MSC_VER)
michael@0:   return((xcr0 & 6) == 6);  // Is ymm saved?
michael@0: }
michael@0: #endif  // !defined(__native_client__)
michael@0: #else
michael@0: LIBYUV_API
michael@0: void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
michael@0:   cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
michael@0: }
michael@0: #endif
michael@0: 
michael@0: // based on libvpx arm_cpudetect.c
michael@0: // For Arm, but public to allow testing on any CPU
michael@0: LIBYUV_API SAFEBUFFERS
michael@0: int ArmCpuCaps(const char* cpuinfo_name) {
michael@0:   FILE* f = fopen(cpuinfo_name, "r");
michael@0:   if (f) {
michael@0:     char cpuinfo_line[512];
michael@0:     while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
michael@0:       if (memcmp(cpuinfo_line, "Features", 8) == 0) {
michael@0:         char* p = strstr(cpuinfo_line, " neon");
michael@0:         if (p && (p[5] == ' ' || p[5] == '\n')) {
michael@0:           fclose(f);
michael@0:           return kCpuHasNEON;
michael@0:         }
michael@0:       }
michael@0:     }
michael@0:     fclose(f);
michael@0:   }
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: #if defined(__mips__) && defined(__linux__)
michael@0: static int MipsCpuCaps(const char* search_string) {
michael@0:   const char* file_name = "/proc/cpuinfo";
michael@0:   char cpuinfo_line[256];
michael@0:   FILE* f = NULL;
michael@0:   if ((f = fopen(file_name, "r")) != NULL) {
michael@0:     while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) {
michael@0:       if (strstr(cpuinfo_line, search_string) != NULL) {
michael@0:         fclose(f);
michael@0:         return kCpuHasMIPS_DSP;
michael@0:       }
michael@0:     }
michael@0:     fclose(f);
michael@0:   }
michael@0:   /* Did not find string in the proc file, or not Linux ELF. */
michael@0:   return 0;
michael@0: }
michael@0: #endif
michael@0: 
michael@0: // CPU detect function for SIMD instruction sets.
michael@0: LIBYUV_API
michael@0: int cpu_info_ = kCpuInit;  // cpu_info is not initialized yet.
michael@0: 
michael@0: // Test environment variable for disabling CPU features. Any non-zero value
michael@0: // to disable. Zero ignored to make it easy to set the variable on/off.
michael@0: #if !defined(__native_client__) && !defined(_M_ARM)
michael@0: 
michael@0: static LIBYUV_BOOL TestEnv(const char* name) {
michael@0:   const char* var = getenv(name);
michael@0:   if (var) {
michael@0:     if (var[0] != '0') {
michael@0:       return LIBYUV_TRUE;
michael@0:     }
michael@0:   }
michael@0:   return LIBYUV_FALSE;
michael@0: }
michael@0: #else  // nacl does not support getenv().
michael@0: static LIBYUV_BOOL TestEnv(const char*) {
michael@0:   return LIBYUV_FALSE;
michael@0: }
michael@0: #endif
michael@0: 
michael@0: LIBYUV_API SAFEBUFFERS
michael@0: int InitCpuFlags(void) {
michael@0: #if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
michael@0: 
michael@0:   uint32 cpu_info1[4] = { 0, 0, 0, 0 };
michael@0:   uint32 cpu_info7[4] = { 0, 0, 0, 0 };
michael@0:   CpuId(1, 0, cpu_info1);
michael@0:   CpuId(7, 0, cpu_info7);
michael@0:   cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
michael@0:               ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
michael@0:               ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
michael@0:               ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
michael@0:               ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
michael@0:               ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
michael@0:               kCpuHasX86;
michael@0: #ifdef HAS_XGETBV
michael@0:   if ((cpu_info1[2] & 0x18000000) == 0x18000000 &&  // AVX and OSSave
michael@0:       TestOsSaveYmm()) {  // Saves YMM.
michael@0:     cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
michael@0:                  kCpuHasAVX;
michael@0:   }
michael@0: #endif
michael@0:   // Environment variable overrides for testing.
michael@0:   if (TestEnv("LIBYUV_DISABLE_X86")) {
michael@0:     cpu_info_ &= ~kCpuHasX86;
michael@0:   }
michael@0:   if (TestEnv("LIBYUV_DISABLE_SSE2")) {
michael@0:     cpu_info_ &= ~kCpuHasSSE2;
michael@0:   }
michael@0:   if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
michael@0:     cpu_info_ &= ~kCpuHasSSSE3;
michael@0:   }
michael@0:   if (TestEnv("LIBYUV_DISABLE_SSE41")) {
michael@0:     cpu_info_ &= ~kCpuHasSSE41;
michael@0:   }
michael@0:   if (TestEnv("LIBYUV_DISABLE_SSE42")) {
michael@0:     cpu_info_ &= ~kCpuHasSSE42;
michael@0:   }
michael@0:   if (TestEnv("LIBYUV_DISABLE_AVX")) {
michael@0:     cpu_info_ &= ~kCpuHasAVX;
michael@0:   }
michael@0:   if (TestEnv("LIBYUV_DISABLE_AVX2")) {
michael@0:     cpu_info_ &= ~kCpuHasAVX2;
michael@0:   }
michael@0:   if (TestEnv("LIBYUV_DISABLE_ERMS")) {
michael@0:     cpu_info_ &= ~kCpuHasERMS;
michael@0:   }
michael@0:   if (TestEnv("LIBYUV_DISABLE_FMA3")) {
michael@0:     cpu_info_ &= ~kCpuHasFMA3;
michael@0:   }
michael@0: #elif defined(__mips__) && defined(__linux__)
michael@0:   // Linux mips parse text file for dsp detect.
michael@0:   cpu_info_ = MipsCpuCaps("dsp");  // set kCpuHasMIPS_DSP.
michael@0: #if defined(__mips_dspr2)
michael@0:   cpu_info_ |= kCpuHasMIPS_DSPR2;
michael@0: #endif
michael@0:   cpu_info_ |= kCpuHasMIPS;
michael@0: 
michael@0:   if (getenv("LIBYUV_DISABLE_MIPS")) {
michael@0:     cpu_info_ &= ~kCpuHasMIPS;
michael@0:   }
michael@0:   if (getenv("LIBYUV_DISABLE_MIPS_DSP")) {
michael@0:     cpu_info_ &= ~kCpuHasMIPS_DSP;
michael@0:   }
michael@0:   if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
michael@0:     cpu_info_ &= ~kCpuHasMIPS_DSPR2;
michael@0:   }
michael@0: #elif defined(__arm__)
michael@0: #if defined(__linux__) && (defined(__ARM_NEON__) || defined(LIBYUV_NEON)) && \
michael@0:     !defined(__native_client__)
michael@0:   // Linux arm parse text file for neon detect.
michael@0:   cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
michael@0: #elif defined(__ARM_NEON__) || defined(__native_client__)
michael@0:   // gcc -mfpu=neon defines __ARM_NEON__
michael@0:   // Enable Neon if you want support for Neon and Arm, and use MaskCpuFlags
michael@0:   // to disable Neon on devices that do not have it.
michael@0:   cpu_info_ = kCpuHasNEON;
michael@0: #endif
michael@0:   cpu_info_ |= kCpuHasARM;
michael@0:   if (TestEnv("LIBYUV_DISABLE_NEON")) {
michael@0:     cpu_info_ &= ~kCpuHasNEON;
michael@0:   }
michael@0: #endif  // __arm__
michael@0:   if (TestEnv("LIBYUV_DISABLE_ASM")) {
michael@0:     cpu_info_ = 0;
michael@0:   }
michael@0:   return cpu_info_;
michael@0: }
michael@0: 
michael@0: LIBYUV_API
michael@0: void MaskCpuFlags(int enable_flags) {
michael@0:   cpu_info_ = InitCpuFlags() & enable_flags;
michael@0: }
michael@0: 
michael@0: #ifdef __cplusplus
michael@0: }  // extern "C"
michael@0: }  // namespace libyuv
michael@0: #endif