diff options
Diffstat (limited to 'media/libyuv/source/cpu_id.cc')
| -rw-r--r-- | media/libyuv/source/cpu_id.cc | 240 |
1 files changed, 120 insertions, 120 deletions
diff --git a/media/libyuv/source/cpu_id.cc b/media/libyuv/source/cpu_id.cc index 84927ebc3e..13e3da7bb1 100644 --- a/media/libyuv/source/cpu_id.cc +++ b/media/libyuv/source/cpu_id.cc @@ -13,22 +13,16 @@ #if defined(_MSC_VER) #include <intrin.h> // For __cpuidex() #endif -#if !defined(__pnacl__) && !defined(__CLR_VER) && \ +#if !defined(__pnacl__) && !defined(__CLR_VER) && \ !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \ defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) #include <immintrin.h> // For _xgetbv() #endif -#if !defined(__native_client__) -#include <stdlib.h> // For getenv() -#endif - // For ArmCpuCaps() but unittested on all platforms -#include <stdio.h> +#include <stdio.h> // For fopen() #include <string.h> -#include "libyuv/basic_types.h" // For CPU_X86 - #ifdef __cplusplus namespace libyuv { extern "C" { @@ -43,16 +37,20 @@ extern "C" { #define SAFEBUFFERS #endif +// cpu_info_ variable for SIMD instruction sets detected. +LIBYUV_API int cpu_info_ = 0; + +// TODO(fbarchard): Consider using int for cpuid so casting is not needed. // Low level cpuid for X86. -#if (defined(_M_IX86) || defined(_M_X64) || \ - defined(__i386__) || defined(__x86_64__)) && \ +#if (defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || \ + defined(__x86_64__)) && \ !defined(__pnacl__) && !defined(__CLR_VER) LIBYUV_API -void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) { +void CpuId(int info_eax, int info_ecx, int* cpu_info) { #if defined(_MSC_VER) // Visual C version uses intrinsic or inline x86 assembly. #if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) - __cpuidex((int*)(cpu_info), info_eax, info_ecx); + __cpuidex(cpu_info, info_eax, info_ecx); #elif defined(_M_IX86) __asm { mov eax, info_eax @@ -66,26 +64,26 @@ void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) { } #else // Visual C but not x86 if (info_ecx == 0) { - __cpuid((int*)(cpu_info), info_eax); + __cpuid(cpu_info, info_eax); } else { - cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0; + cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0u; } #endif // GCC version uses inline x86 assembly. #else // defined(_MSC_VER) - uint32 info_ebx, info_edx; - asm volatile ( -#if defined( __i386__) && defined(__PIC__) - // Preserve ebx for fpic 32 bit. - "mov %%ebx, %%edi \n" - "cpuid \n" - "xchg %%edi, %%ebx \n" - : "=D" (info_ebx), + int info_ebx, info_edx; + asm volatile( +#if defined(__i386__) && defined(__PIC__) + // Preserve ebx for fpic 32 bit. + "mov %%ebx, %%edi \n" + "cpuid \n" + "xchg %%edi, %%ebx \n" + : "=D"(info_ebx), #else - "cpuid \n" - : "=b" (info_ebx), + "cpuid \n" + : "=b"(info_ebx), #endif // defined( __i386__) && defined(__PIC__) - "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx)); + "+a"(info_eax), "+c"(info_ecx), "=d"(info_edx)); cpu_info[0] = info_eax; cpu_info[1] = info_ebx; cpu_info[2] = info_ecx; @@ -94,7 +92,9 @@ void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) { } #else // (defined(_M_IX86) || defined(_M_X64) ... LIBYUV_API -void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) { +void CpuId(int eax, int ecx, int* cpu_info) { + (void)eax; + (void)ecx; cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0; } #endif @@ -108,33 +108,34 @@ void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) { // } // For VS2013 and earlier 32 bit, the _xgetbv(0) optimizer produces bad code. // https://code.google.com/p/libyuv/issues/detail?id=529 -#if defined(_M_IX86) && (_MSC_VER < 1900) +#if defined(_M_IX86) && defined(_MSC_VER) && (_MSC_VER < 1900) #pragma optimize("g", off) #endif -#if (defined(_M_IX86) || defined(_M_X64) || \ - defined(__i386__) || defined(__x86_64__)) && \ +#if (defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || \ + defined(__x86_64__)) && \ !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__) -#define HAS_XGETBV // X86 CPUs have xgetbv to detect OS saves high parts of ymm registers. -int GetXCR0() { - uint32 xcr0 = 0u; +static int GetXCR0() { + int xcr0 = 0; #if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 160040219) - xcr0 = (uint32)(_xgetbv(0)); // VS2010 SP1 required. + xcr0 = (int)_xgetbv(0); // VS2010 SP1 required. NOLINT #elif defined(__i386__) || defined(__x86_64__) - asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx"); + asm(".byte 0x0f, 0x01, 0xd0" : "=a"(xcr0) : "c"(0) : "%edx"); #endif // defined(__i386__) || defined(__x86_64__) return xcr0; } +#else +// xgetbv unavailable to query for OSSave support. Return 0. +#define GetXCR0() 0 #endif // defined(_M_IX86) || defined(_M_X64) .. // Return optimization to previous setting. -#if defined(_M_IX86) && (_MSC_VER < 1900) +#if defined(_M_IX86) && defined(_MSC_VER) && (_MSC_VER < 1900) #pragma optimize("g", on) #endif -// based on libvpx arm_cpudetect.c +// Based on libvpx arm_cpudetect.c // For Arm, but public to allow testing on any CPU -LIBYUV_API SAFEBUFFERS -int ArmCpuCaps(const char* cpuinfo_name) { +LIBYUV_API SAFEBUFFERS int ArmCpuCaps(const char* cpuinfo_name) { char cpuinfo_line[512]; FILE* f = fopen(cpuinfo_name, "r"); if (!f) { @@ -151,7 +152,7 @@ int ArmCpuCaps(const char* cpuinfo_name) { } // aarch64 uses asimd for Neon. p = strstr(cpuinfo_line, " asimd"); - if (p && (p[6] == ' ' || p[6] == '\n')) { + if (p) { fclose(f); return kCpuHasNEON; } @@ -161,103 +162,102 @@ int ArmCpuCaps(const char* cpuinfo_name) { return 0; } -// CPU detect function for SIMD instruction sets. -LIBYUV_API -int cpu_info_ = 0; // cpu_info is not initialized yet. - -// Test environment variable for disabling CPU features. Any non-zero value -// to disable. Zero ignored to make it easy to set the variable on/off. -#if !defined(__native_client__) && !defined(_M_ARM) - -static LIBYUV_BOOL TestEnv(const char* name) { - const char* var = getenv(name); - if (var) { - if (var[0] != '0') { - return LIBYUV_TRUE; +// TODO(fbarchard): Consider read_msa_ir(). +LIBYUV_API SAFEBUFFERS int MipsCpuCaps(const char* cpuinfo_name) { + char cpuinfo_line[512]; + int flag = 0x0; + FILE* f = fopen(cpuinfo_name, "r"); + if (!f) { + // Assume nothing if /proc/cpuinfo is unavailable. + // This will occur for Chrome sandbox for Pepper or Render process. + return 0; + } + while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) { + if (memcmp(cpuinfo_line, "cpu model", 9) == 0) { + // Workaround early kernel without MSA in ASEs line. + if (strstr(cpuinfo_line, "Loongson-2K")) { + flag |= kCpuHasMSA; + } + } + if (memcmp(cpuinfo_line, "ASEs implemented", 16) == 0) { + if (strstr(cpuinfo_line, "msa")) { + flag |= kCpuHasMSA; + } + // ASEs is the last line, so we can break here. + break; } } - return LIBYUV_FALSE; + fclose(f); + return flag; } -#else // nacl does not support getenv(). -static LIBYUV_BOOL TestEnv(const char*) { - return LIBYUV_FALSE; + +// TODO(fbarchard): Consider read_loongarch_ir(). +#define LOONGARCH_CFG2 0x2 +#define LOONGARCH_CFG2_LSX (1 << 6) +#define LOONGARCH_CFG2_LASX (1 << 7) + +#if defined(__loongarch__) +LIBYUV_API SAFEBUFFERS int LoongarchCpuCaps(void) { + int flag = 0x0; + uint32_t cfg2 = 0; + + __asm__ volatile("cpucfg %0, %1 \n\t" : "+&r"(cfg2) : "r"(LOONGARCH_CFG2)); + + if (cfg2 & LOONGARCH_CFG2_LSX) + flag |= kCpuHasLSX; + + if (cfg2 & LOONGARCH_CFG2_LASX) + flag |= kCpuHasLASX; + return flag; } #endif -LIBYUV_API SAFEBUFFERS -int InitCpuFlags(void) { - // TODO(fbarchard): swap kCpuInit logic so 0 means uninitialized. +static SAFEBUFFERS int GetCpuFlags(void) { int cpu_info = 0; -#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86) - uint32 cpu_info0[4] = { 0, 0, 0, 0 }; - uint32 cpu_info1[4] = { 0, 0, 0, 0 }; - uint32 cpu_info7[4] = { 0, 0, 0, 0 }; +#if !defined(__pnacl__) && !defined(__CLR_VER) && \ + (defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || \ + defined(_M_IX86)) + int cpu_info0[4] = {0, 0, 0, 0}; + int cpu_info1[4] = {0, 0, 0, 0}; + int cpu_info7[4] = {0, 0, 0, 0}; CpuId(0, 0, cpu_info0); CpuId(1, 0, cpu_info1); if (cpu_info0[0] >= 7) { CpuId(7, 0, cpu_info7); } - cpu_info = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) | + cpu_info = kCpuHasX86 | ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) | ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) | ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) | ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) | - ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) | - ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) | - kCpuHasX86; + ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0); -#ifdef HAS_XGETBV - // AVX requires CPU has AVX, XSAVE and OSXSave for xgetbv + // AVX requires OS saves YMM registers. if (((cpu_info1[2] & 0x1c000000) == 0x1c000000) && // AVX and OSXSave ((GetXCR0() & 6) == 6)) { // Test OS saves YMM registers - cpu_info |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) | kCpuHasAVX; + cpu_info |= kCpuHasAVX | ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) | + ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) | + ((cpu_info1[2] & 0x20000000) ? kCpuHasF16C : 0); // Detect AVX512bw if ((GetXCR0() & 0xe0) == 0xe0) { - cpu_info |= (cpu_info7[1] & 0x40000000) ? kCpuHasAVX3 : 0; + cpu_info |= (cpu_info7[1] & 0x40000000) ? kCpuHasAVX512BW : 0; + cpu_info |= (cpu_info7[1] & 0x80000000) ? kCpuHasAVX512VL : 0; + cpu_info |= (cpu_info7[2] & 0x00000002) ? kCpuHasAVX512VBMI : 0; + cpu_info |= (cpu_info7[2] & 0x00000040) ? kCpuHasAVX512VBMI2 : 0; + cpu_info |= (cpu_info7[2] & 0x00000800) ? kCpuHasAVX512VNNI : 0; + cpu_info |= (cpu_info7[2] & 0x00001000) ? kCpuHasAVX512VBITALG : 0; + cpu_info |= (cpu_info7[2] & 0x00004000) ? kCpuHasAVX512VPOPCNTDQ : 0; + cpu_info |= (cpu_info7[2] & 0x00000100) ? kCpuHasGFNI : 0; } } #endif - - // Environment variable overrides for testing. - if (TestEnv("LIBYUV_DISABLE_X86")) { - cpu_info &= ~kCpuHasX86; - } - if (TestEnv("LIBYUV_DISABLE_SSE2")) { - cpu_info &= ~kCpuHasSSE2; - } - if (TestEnv("LIBYUV_DISABLE_SSSE3")) { - cpu_info &= ~kCpuHasSSSE3; - } - if (TestEnv("LIBYUV_DISABLE_SSE41")) { - cpu_info &= ~kCpuHasSSE41; - } - if (TestEnv("LIBYUV_DISABLE_SSE42")) { - cpu_info &= ~kCpuHasSSE42; - } - if (TestEnv("LIBYUV_DISABLE_AVX")) { - cpu_info &= ~kCpuHasAVX; - } - if (TestEnv("LIBYUV_DISABLE_AVX2")) { - cpu_info &= ~kCpuHasAVX2; - } - if (TestEnv("LIBYUV_DISABLE_ERMS")) { - cpu_info &= ~kCpuHasERMS; - } - if (TestEnv("LIBYUV_DISABLE_FMA3")) { - cpu_info &= ~kCpuHasFMA3; - } - if (TestEnv("LIBYUV_DISABLE_AVX3")) { - cpu_info &= ~kCpuHasAVX3; - } -#endif #if defined(__mips__) && defined(__linux__) -#if defined(__mips_dspr2) - cpu_info |= kCpuHasDSPR2; -#endif + cpu_info = MipsCpuCaps("/proc/cpuinfo"); cpu_info |= kCpuHasMIPS; - if (getenv("LIBYUV_DISABLE_DSPR2")) { - cpu_info &= ~kCpuHasDSPR2; - } +#endif +#if defined(__loongarch__) && defined(__linux__) + cpu_info = LoongarchCpuCaps(); + cpu_info |= kCpuHasLOONGARCH; #endif #if defined(__arm__) || defined(__aarch64__) // gcc -mfpu=neon defines __ARM_NEON__ @@ -276,22 +276,22 @@ int InitCpuFlags(void) { cpu_info = ArmCpuCaps("/proc/cpuinfo"); #endif cpu_info |= kCpuHasARM; - if (TestEnv("LIBYUV_DISABLE_NEON")) { - cpu_info &= ~kCpuHasNEON; - } #endif // __arm__ - if (TestEnv("LIBYUV_DISABLE_ASM")) { - cpu_info = 0; - } - cpu_info |= kCpuInitialized; - cpu_info_ = cpu_info; + cpu_info |= kCpuInitialized; return cpu_info; } // Note that use of this function is not thread safe. LIBYUV_API -void MaskCpuFlags(int enable_flags) { - cpu_info_ = InitCpuFlags() & enable_flags; +int MaskCpuFlags(int enable_flags) { + int cpu_info = GetCpuFlags() & enable_flags; + SetCpuFlags(cpu_info); + return cpu_info; +} + +LIBYUV_API +int InitCpuFlags(void) { + return MaskCpuFlags(-1); } #ifdef __cplusplus |