Home | History | Annotate | Download | only in vpx_ports
      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 0x001
    155 #define HAS_SSE 0x002
    156 #define HAS_SSE2 0x004
    157 #define HAS_SSE3 0x008
    158 #define HAS_SSSE3 0x010
    159 #define HAS_SSE4_1 0x020
    160 #define HAS_AVX 0x040
    161 #define HAS_AVX2 0x080
    162 #define HAS_AVX512 0x100
    163 #ifndef BIT
    164 #define BIT(n) (1u << n)
    165 #endif
    166 
    167 static INLINE int x86_simd_caps(void) {
    168   unsigned int flags = 0;
    169   unsigned int mask = ~0;
    170   unsigned int max_cpuid_val, reg_eax, reg_ebx, reg_ecx, reg_edx;
    171   char *env;
    172   (void)reg_ebx;
    173 
    174   /* See if the CPU capabilities are being overridden by the environment */
    175   env = getenv("VPX_SIMD_CAPS");
    176 
    177   if (env && *env) return (int)strtol(env, NULL, 0);
    178 
    179   env = getenv("VPX_SIMD_CAPS_MASK");
    180 
    181   if (env && *env) mask = (unsigned int)strtoul(env, NULL, 0);
    182 
    183   /* Ensure that the CPUID instruction supports extended features */
    184   cpuid(0, 0, max_cpuid_val, reg_ebx, reg_ecx, reg_edx);
    185 
    186   if (max_cpuid_val < 1) return 0;
    187 
    188   /* Get the standard feature flags */
    189   cpuid(1, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
    190 
    191   if (reg_edx & BIT(23)) flags |= HAS_MMX;
    192 
    193   if (reg_edx & BIT(25)) flags |= HAS_SSE; /* aka xmm */
    194 
    195   if (reg_edx & BIT(26)) flags |= HAS_SSE2; /* aka wmt */
    196 
    197   if (reg_ecx & BIT(0)) flags |= HAS_SSE3;
    198 
    199   if (reg_ecx & BIT(9)) flags |= HAS_SSSE3;
    200 
    201   if (reg_ecx & BIT(19)) flags |= HAS_SSE4_1;
    202 
    203   // bits 27 (OSXSAVE) & 28 (256-bit AVX)
    204   if ((reg_ecx & (BIT(27) | BIT(28))) == (BIT(27) | BIT(28))) {
    205     if ((xgetbv() & 0x6) == 0x6) {
    206       flags |= HAS_AVX;
    207 
    208       if (max_cpuid_val >= 7) {
    209         /* Get the leaf 7 feature flags. Needed to check for AVX2 support */
    210         cpuid(7, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
    211 
    212         if (reg_ebx & BIT(5)) flags |= HAS_AVX2;
    213 
    214         // bits 16 (AVX-512F) & 17 (AVX-512DQ) & 28 (AVX-512CD) &
    215         // 30 (AVX-512BW) & 32 (AVX-512VL)
    216         if ((reg_ebx & (BIT(16) | BIT(17) | BIT(28) | BIT(30) | BIT(31))) ==
    217             (BIT(16) | BIT(17) | BIT(28) | BIT(30) | BIT(31)))
    218           flags |= HAS_AVX512;
    219       }
    220     }
    221   }
    222 
    223   return flags & mask;
    224 }
    225 
    226 // Note:
    227 //  32-bit CPU cycle counter is light-weighted for most function performance
    228 //  measurement. For large function (CPU time > a couple of seconds), 64-bit
    229 //  counter should be used.
    230 // 32-bit CPU cycle counter
    231 static INLINE unsigned int x86_readtsc(void) {
    232 #if defined(__GNUC__) && __GNUC__
    233   unsigned int tsc;
    234   __asm__ __volatile__("rdtsc\n\t" : "=a"(tsc) :);
    235   return tsc;
    236 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
    237   unsigned int tsc;
    238   asm volatile("rdtsc\n\t" : "=a"(tsc) :);
    239   return tsc;
    240 #else
    241 #if ARCH_X86_64
    242   return (unsigned int)__rdtsc();
    243 #else
    244   __asm rdtsc;
    245 #endif
    246 #endif
    247 }
    248 // 64-bit CPU cycle counter
    249 static INLINE uint64_t x86_readtsc64(void) {
    250 #if defined(__GNUC__) && __GNUC__
    251   uint32_t hi, lo;
    252   __asm__ __volatile__("rdtsc" : "=a"(lo), "=d"(hi));
    253   return ((uint64_t)hi << 32) | lo;
    254 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
    255   uint_t hi, lo;
    256   asm volatile("rdtsc\n\t" : "=a"(lo), "=d"(hi));
    257   return ((uint64_t)hi << 32) | lo;
    258 #else
    259 #if ARCH_X86_64
    260   return (uint64_t)__rdtsc();
    261 #else
    262   __asm rdtsc;
    263 #endif
    264 #endif
    265 }
    266 
    267 #if defined(__GNUC__) && __GNUC__
    268 #define x86_pause_hint() __asm__ __volatile__("pause \n\t")
    269 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
    270 #define x86_pause_hint() asm volatile("pause \n\t")
    271 #else
    272 #if ARCH_X86_64
    273 #define x86_pause_hint() _mm_pause();
    274 #else
    275 #define x86_pause_hint() __asm pause
    276 #endif
    277 #endif
    278 
    279 #if defined(__GNUC__) && __GNUC__
    280 static void x87_set_control_word(unsigned short mode) {
    281   __asm__ __volatile__("fldcw %0" : : "m"(*&mode));
    282 }
    283 static unsigned short x87_get_control_word(void) {
    284   unsigned short mode;
    285   __asm__ __volatile__("fstcw %0\n\t" : "=m"(*&mode) :);
    286   return mode;
    287 }
    288 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
    289 static void x87_set_control_word(unsigned short mode) {
    290   asm volatile("fldcw %0" : : "m"(*&mode));
    291 }
    292 static unsigned short x87_get_control_word(void) {
    293   unsigned short mode;
    294   asm volatile("fstcw %0\n\t" : "=m"(*&mode) :);
    295   return mode;
    296 }
    297 #elif ARCH_X86_64
    298 /* No fldcw intrinsics on Windows x64, punt to external asm */
    299 extern void vpx_winx64_fldcw(unsigned short mode);
    300 extern unsigned short vpx_winx64_fstcw(void);
    301 #define x87_set_control_word vpx_winx64_fldcw
    302 #define x87_get_control_word vpx_winx64_fstcw
    303 #else
    304 static void x87_set_control_word(unsigned short mode) {
    305   __asm { fldcw mode }
    306 }
    307 static unsigned short x87_get_control_word(void) {
    308   unsigned short mode;
    309   __asm { fstcw mode }
    310   return mode;
    311 }
    312 #endif
    313 
    314 static INLINE unsigned int x87_set_double_precision(void) {
    315   unsigned int mode = x87_get_control_word();
    316   x87_set_control_word((mode & ~0x300) | 0x200);
    317   return mode;
    318 }
    319 
    320 extern void vpx_reset_mmx_state(void);
    321 
    322 #ifdef __cplusplus
    323 }  // extern "C"
    324 #endif
    325 
    326 #endif  // VPX_PORTS_X86_H_
    327