1 /* 2 * Copyright (c) 2010 The WebM 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 #ifndef VPX_PORTS_X86_H_ 12 #define VPX_PORTS_X86_H_ 13 #include <stdlib.h> 14 15 #if defined(_MSC_VER) 16 #include <intrin.h> /* For __cpuidex, __rdtsc */ 17 #endif 18 19 #include "vpx_config.h" 20 #include "vpx/vpx_integer.h" 21 22 #ifdef __cplusplus 23 extern "C" { 24 #endif 25 26 typedef enum { 27 VPX_CPU_UNKNOWN = -1, 28 VPX_CPU_AMD, 29 VPX_CPU_AMD_OLD, 30 VPX_CPU_CENTAUR, 31 VPX_CPU_CYRIX, 32 VPX_CPU_INTEL, 33 VPX_CPU_NEXGEN, 34 VPX_CPU_NSC, 35 VPX_CPU_RISE, 36 VPX_CPU_SIS, 37 VPX_CPU_TRANSMETA, 38 VPX_CPU_TRANSMETA_OLD, 39 VPX_CPU_UMC, 40 VPX_CPU_VIA, 41 42 VPX_CPU_LAST 43 } vpx_cpu_t; 44 45 #if defined(__GNUC__) && __GNUC__ || defined(__ANDROID__) 46 #if ARCH_X86_64 47 #define cpuid(func, func2, ax, bx, cx, dx) \ 48 __asm__ __volatile__("cpuid \n\t" \ 49 : "=a"(ax), "=b"(bx), "=c"(cx), "=d"(dx) \ 50 : "a"(func), "c"(func2)); 51 #else 52 #define cpuid(func, func2, ax, bx, cx, dx) \ 53 __asm__ __volatile__( \ 54 "mov %%ebx, %%edi \n\t" \ 55 "cpuid \n\t" \ 56 "xchg %%edi, %%ebx \n\t" \ 57 : "=a"(ax), "=D"(bx), "=c"(cx), "=d"(dx) \ 58 : "a"(func), "c"(func2)); 59 #endif 60 #elif defined(__SUNPRO_C) || \ 61 defined(__SUNPRO_CC) /* end __GNUC__ or __ANDROID__*/ 62 #if ARCH_X86_64 63 #define cpuid(func, func2, ax, bx, cx, dx) \ 64 asm volatile( \ 65 "xchg %rsi, %rbx \n\t" \ 66 "cpuid \n\t" \ 67 "movl %ebx, %edi \n\t" \ 68 "xchg %rsi, %rbx \n\t" \ 69 : "=a"(ax), "=D"(bx), "=c"(cx), "=d"(dx) \ 70 : "a"(func), "c"(func2)); 71 #else 72 #define cpuid(func, func2, ax, bx, cx, dx) \ 73 asm volatile( \ 74 "pushl %ebx \n\t" \ 75 "cpuid \n\t" \ 76 "movl %ebx, %edi \n\t" \ 77 "popl %ebx \n\t" \ 78 : "=a"(ax), "=D"(bx), "=c"(cx), "=d"(dx) \ 79 : "a"(func), "c"(func2)); 80 #endif 81 #else /* end __SUNPRO__ */ 82 #if ARCH_X86_64 83 #if defined(_MSC_VER) && _MSC_VER > 1500 84 #define cpuid(func, func2, a, b, c, d) \ 85 do { \ 86 int regs[4]; \ 87 __cpuidex(regs, func, func2); \ 88 a = regs[0]; \ 89 b = regs[1]; \ 90 c = regs[2]; \ 91 d = regs[3]; \ 92 } while (0) 93 #else 94 #define cpuid(func, func2, a, b, c, d) \ 95 do { \ 96 int regs[4]; \ 97 __cpuid(regs, func); \ 98 a = regs[0]; \ 99 b = regs[1]; \ 100 c = regs[2]; \ 101 d = regs[3]; \ 102 } while (0) 103 #endif 104 #else 105 #define cpuid(func, func2, a, b, c, d) \ 106 __asm mov eax, func __asm mov ecx, func2 __asm cpuid __asm mov a, \ 107 eax __asm mov b, ebx __asm mov c, ecx __asm mov d, edx 108 #endif 109 #endif /* end others */ 110 111 // NaCl has no support for xgetbv or the raw opcode. 112 #if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__)) 113 static INLINE uint64_t xgetbv(void) { 114 const uint32_t ecx = 0; 115 uint32_t eax, edx; 116 // Use the raw opcode for xgetbv for compatibility with older toolchains. 117 __asm__ volatile(".byte 0x0f, 0x01, 0xd0\n" 118 : "=a"(eax), "=d"(edx) 119 : "c"(ecx)); 120 return ((uint64_t)edx << 32) | eax; 121 } 122 #elif (defined(_M_X64) || defined(_M_IX86)) && defined(_MSC_FULL_VER) && \ 123 _MSC_FULL_VER >= 160040219 // >= VS2010 SP1 124 #include <immintrin.h> 125 #define xgetbv() _xgetbv(0) 126 #elif defined(_MSC_VER) && defined(_M_IX86) 127 static INLINE uint64_t xgetbv(void) { 128 uint32_t eax_, edx_; 129 __asm { 130 xor ecx, ecx // ecx = 0 131 // Use the raw opcode for xgetbv for compatibility with older toolchains. 132 __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0 133 mov eax_, eax 134 mov edx_, edx 135 } 136 return ((uint64_t)edx_ << 32) | eax_; 137 } 138 #else 139 #define xgetbv() 0U // no AVX for older x64 or unrecognized toolchains. 140 #endif 141 142 #if defined(_MSC_VER) && _MSC_VER >= 1700 143 #undef NOMINMAX 144 #define NOMINMAX 145 #ifndef WIN32_LEAN_AND_MEAN 146 #define WIN32_LEAN_AND_MEAN 147 #endif 148 #include <windows.h> 149 #if WINAPI_FAMILY_PARTITION(WINAPI_FAMILY_APP) 150 #define getenv(x) NULL 151 #endif 152 #endif 153 154 #define HAS_MMX 0x01 155 #define HAS_SSE 0x02 156 #define HAS_SSE2 0x04 157 #define HAS_SSE3 0x08 158 #define HAS_SSSE3 0x10 159 #define HAS_SSE4_1 0x20 160 #define HAS_AVX 0x40 161 #define HAS_AVX2 0x80 162 #ifndef BIT 163 #define BIT(n) (1 << n) 164 #endif 165 166 static INLINE int x86_simd_caps(void) { 167 unsigned int flags = 0; 168 unsigned int mask = ~0; 169 unsigned int max_cpuid_val, reg_eax, reg_ebx, reg_ecx, reg_edx; 170 char *env; 171 (void)reg_ebx; 172 173 /* See if the CPU capabilities are being overridden by the environment */ 174 env = getenv("VPX_SIMD_CAPS"); 175 176 if (env && *env) return (int)strtol(env, NULL, 0); 177 178 env = getenv("VPX_SIMD_CAPS_MASK"); 179 180 if (env && *env) mask = (unsigned int)strtoul(env, NULL, 0); 181 182 /* Ensure that the CPUID instruction supports extended features */ 183 cpuid(0, 0, max_cpuid_val, reg_ebx, reg_ecx, reg_edx); 184 185 if (max_cpuid_val < 1) return 0; 186 187 /* Get the standard feature flags */ 188 cpuid(1, 0, reg_eax, reg_ebx, reg_ecx, reg_edx); 189 190 if (reg_edx & BIT(23)) flags |= HAS_MMX; 191 192 if (reg_edx & BIT(25)) flags |= HAS_SSE; /* aka xmm */ 193 194 if (reg_edx & BIT(26)) flags |= HAS_SSE2; /* aka wmt */ 195 196 if (reg_ecx & BIT(0)) flags |= HAS_SSE3; 197 198 if (reg_ecx & BIT(9)) flags |= HAS_SSSE3; 199 200 if (reg_ecx & BIT(19)) flags |= HAS_SSE4_1; 201 202 // bits 27 (OSXSAVE) & 28 (256-bit AVX) 203 if ((reg_ecx & (BIT(27) | BIT(28))) == (BIT(27) | BIT(28))) { 204 if ((xgetbv() & 0x6) == 0x6) { 205 flags |= HAS_AVX; 206 207 if (max_cpuid_val >= 7) { 208 /* Get the leaf 7 feature flags. Needed to check for AVX2 support */ 209 cpuid(7, 0, reg_eax, reg_ebx, reg_ecx, reg_edx); 210 211 if (reg_ebx & BIT(5)) flags |= HAS_AVX2; 212 } 213 } 214 } 215 216 return flags & mask; 217 } 218 219 // Note: 220 // 32-bit CPU cycle counter is light-weighted for most function performance 221 // measurement. For large function (CPU time > a couple of seconds), 64-bit 222 // counter should be used. 223 // 32-bit CPU cycle counter 224 static INLINE unsigned int x86_readtsc(void) { 225 #if defined(__GNUC__) && __GNUC__ 226 unsigned int tsc; 227 __asm__ __volatile__("rdtsc\n\t" : "=a"(tsc) :); 228 return tsc; 229 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC) 230 unsigned int tsc; 231 asm volatile("rdtsc\n\t" : "=a"(tsc) :); 232 return tsc; 233 #else 234 #if ARCH_X86_64 235 return (unsigned int)__rdtsc(); 236 #else 237 __asm rdtsc; 238 #endif 239 #endif 240 } 241 // 64-bit CPU cycle counter 242 static INLINE uint64_t x86_readtsc64(void) { 243 #if defined(__GNUC__) && __GNUC__ 244 uint32_t hi, lo; 245 __asm__ __volatile__("rdtsc" : "=a"(lo), "=d"(hi)); 246 return ((uint64_t)hi << 32) | lo; 247 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC) 248 uint_t hi, lo; 249 asm volatile("rdtsc\n\t" : "=a"(lo), "=d"(hi)); 250 return ((uint64_t)hi << 32) | lo; 251 #else 252 #if ARCH_X86_64 253 return (uint64_t)__rdtsc(); 254 #else 255 __asm rdtsc; 256 #endif 257 #endif 258 } 259 260 #if defined(__GNUC__) && __GNUC__ 261 #define x86_pause_hint() __asm__ __volatile__("pause \n\t") 262 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC) 263 #define x86_pause_hint() asm volatile("pause \n\t") 264 #else 265 #if ARCH_X86_64 266 #define x86_pause_hint() _mm_pause(); 267 #else 268 #define x86_pause_hint() __asm pause 269 #endif 270 #endif 271 272 #if defined(__GNUC__) && __GNUC__ 273 static void x87_set_control_word(unsigned short mode) { 274 __asm__ __volatile__("fldcw %0" : : "m"(*&mode)); 275 } 276 static unsigned short x87_get_control_word(void) { 277 unsigned short mode; 278 __asm__ __volatile__("fstcw %0\n\t" : "=m"(*&mode) :); 279 return mode; 280 } 281 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC) 282 static void x87_set_control_word(unsigned short mode) { 283 asm volatile("fldcw %0" : : "m"(*&mode)); 284 } 285 static unsigned short x87_get_control_word(void) { 286 unsigned short mode; 287 asm volatile("fstcw %0\n\t" : "=m"(*&mode) :); 288 return mode; 289 } 290 #elif ARCH_X86_64 291 /* No fldcw intrinsics on Windows x64, punt to external asm */ 292 extern void vpx_winx64_fldcw(unsigned short mode); 293 extern unsigned short vpx_winx64_fstcw(void); 294 #define x87_set_control_word vpx_winx64_fldcw 295 #define x87_get_control_word vpx_winx64_fstcw 296 #else 297 static void x87_set_control_word(unsigned short mode) { 298 __asm { fldcw mode } 299 } 300 static unsigned short x87_get_control_word(void) { 301 unsigned short mode; 302 __asm { fstcw mode } 303 return mode; 304 } 305 #endif 306 307 static INLINE unsigned int x87_set_double_precision(void) { 308 unsigned int mode = x87_get_control_word(); 309 x87_set_control_word((mode & ~0x300) | 0x200); 310 return mode; 311 } 312 313 extern void vpx_reset_mmx_state(void); 314 315 #ifdef __cplusplus 316 } // extern "C" 317 #endif 318 319 #endif // VPX_PORTS_X86_H_ 320
