michael@0: /*
michael@0:  *  Copyright (c) 2010 The WebM 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: 
michael@0: #ifndef VPX_PORTS_X86_H
michael@0: #define VPX_PORTS_X86_H
michael@0: #include <stdlib.h>
michael@0: #include "vpx_config.h"
michael@0: 
michael@0: typedef enum {
michael@0:   VPX_CPU_UNKNOWN = -1,
michael@0:   VPX_CPU_AMD,
michael@0:   VPX_CPU_AMD_OLD,
michael@0:   VPX_CPU_CENTAUR,
michael@0:   VPX_CPU_CYRIX,
michael@0:   VPX_CPU_INTEL,
michael@0:   VPX_CPU_NEXGEN,
michael@0:   VPX_CPU_NSC,
michael@0:   VPX_CPU_RISE,
michael@0:   VPX_CPU_SIS,
michael@0:   VPX_CPU_TRANSMETA,
michael@0:   VPX_CPU_TRANSMETA_OLD,
michael@0:   VPX_CPU_UMC,
michael@0:   VPX_CPU_VIA,
michael@0: 
michael@0:   VPX_CPU_LAST
michael@0: }  vpx_cpu_t;
michael@0: 
michael@0: #if defined(__GNUC__) && __GNUC__ || defined(__ANDROID__)
michael@0: #if ARCH_X86_64
michael@0: #define cpuid(func,ax,bx,cx,dx)\
michael@0:   __asm__ __volatile__ (\
michael@0:                         "cpuid           \n\t" \
michael@0:                         : "=a" (ax), "=b" (bx), "=c" (cx), "=d" (dx) \
michael@0:                         : "a"  (func));
michael@0: #else
michael@0: #define cpuid(func,ax,bx,cx,dx)\
michael@0:   __asm__ __volatile__ (\
michael@0:                         "mov %%ebx, %%edi   \n\t" \
michael@0:                         "cpuid              \n\t" \
michael@0:                         "xchg %%edi, %%ebx  \n\t" \
michael@0:                         : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
michael@0:                         : "a" (func));
michael@0: #endif
michael@0: #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC) /* end __GNUC__ or __ANDROID__*/
michael@0: #if ARCH_X86_64
michael@0: #define cpuid(func,ax,bx,cx,dx)\
michael@0:   asm volatile (\
michael@0:                 "xchg %rsi, %rbx \n\t" \
michael@0:                 "cpuid           \n\t" \
michael@0:                 "movl %ebx, %edi \n\t" \
michael@0:                 "xchg %rsi, %rbx \n\t" \
michael@0:                 : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
michael@0:                 : "a"  (func));
michael@0: #else
michael@0: #define cpuid(func,ax,bx,cx,dx)\
michael@0:   asm volatile (\
michael@0:                 "pushl %ebx       \n\t" \
michael@0:                 "cpuid            \n\t" \
michael@0:                 "movl %ebx, %edi  \n\t" \
michael@0:                 "popl %ebx        \n\t" \
michael@0:                 : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
michael@0:                 : "a" (func));
michael@0: #endif
michael@0: #else /* end __SUNPRO__ */
michael@0: #if ARCH_X86_64
michael@0: void __cpuid(int CPUInfo[4], int info_type);
michael@0: #pragma intrinsic(__cpuid)
michael@0: #define cpuid(func,a,b,c,d) do{\
michael@0:     int regs[4];\
michael@0:     __cpuid(regs,func); a=regs[0];  b=regs[1];  c=regs[2];  d=regs[3];\
michael@0:   } while(0)
michael@0: #else
michael@0: #define cpuid(func,a,b,c,d)\
michael@0:   __asm mov eax, func\
michael@0:   __asm cpuid\
michael@0:   __asm mov a, eax\
michael@0:   __asm mov b, ebx\
michael@0:   __asm mov c, ecx\
michael@0:   __asm mov d, edx
michael@0: #endif
michael@0: #endif /* end others */
michael@0: 
michael@0: #define HAS_MMX     0x01
michael@0: #define HAS_SSE     0x02
michael@0: #define HAS_SSE2    0x04
michael@0: #define HAS_SSE3    0x08
michael@0: #define HAS_SSSE3   0x10
michael@0: #define HAS_SSE4_1  0x20
michael@0: #define HAS_AVX     0x40
michael@0: #define HAS_AVX2    0x80
michael@0: #ifndef BIT
michael@0: #define BIT(n) (1<<n)
michael@0: #endif
michael@0: 
michael@0: static int
michael@0: x86_simd_caps(void) {
michael@0:   unsigned int flags = 0;
michael@0:   unsigned int mask = ~0;
michael@0:   unsigned int reg_eax, reg_ebx, reg_ecx, reg_edx;
michael@0:   char *env;
michael@0:   (void)reg_ebx;
michael@0: 
michael@0:   /* See if the CPU capabilities are being overridden by the environment */
michael@0:   env = getenv("VPX_SIMD_CAPS");
michael@0: 
michael@0:   if (env && *env)
michael@0:     return (int)strtol(env, NULL, 0);
michael@0: 
michael@0:   env = getenv("VPX_SIMD_CAPS_MASK");
michael@0: 
michael@0:   if (env && *env)
michael@0:     mask = strtol(env, NULL, 0);
michael@0: 
michael@0:   /* Ensure that the CPUID instruction supports extended features */
michael@0:   cpuid(0, reg_eax, reg_ebx, reg_ecx, reg_edx);
michael@0: 
michael@0:   if (reg_eax < 1)
michael@0:     return 0;
michael@0: 
michael@0:   /* Get the standard feature flags */
michael@0:   cpuid(1, reg_eax, reg_ebx, reg_ecx, reg_edx);
michael@0: 
michael@0:   if (reg_edx & BIT(23)) flags |= HAS_MMX;
michael@0: 
michael@0:   if (reg_edx & BIT(25)) flags |= HAS_SSE; /* aka xmm */
michael@0: 
michael@0:   if (reg_edx & BIT(26)) flags |= HAS_SSE2; /* aka wmt */
michael@0: 
michael@0:   if (reg_ecx & BIT(0)) flags |= HAS_SSE3;
michael@0: 
michael@0:   if (reg_ecx & BIT(9)) flags |= HAS_SSSE3;
michael@0: 
michael@0:   if (reg_ecx & BIT(19)) flags |= HAS_SSE4_1;
michael@0: 
michael@0:   if (reg_ecx & BIT(28)) flags |= HAS_AVX;
michael@0: 
michael@0:   if (reg_ebx & BIT(5)) flags |= HAS_AVX2;
michael@0: 
michael@0:   return flags & mask;
michael@0: }
michael@0: 
michael@0: vpx_cpu_t vpx_x86_vendor(void);
michael@0: 
michael@0: #if ARCH_X86_64 && defined(_MSC_VER)
michael@0: unsigned __int64 __rdtsc(void);
michael@0: #pragma intrinsic(__rdtsc)
michael@0: #endif
michael@0: static unsigned int
michael@0: x86_readtsc(void) {
michael@0: #if defined(__GNUC__) && __GNUC__
michael@0:   unsigned int tsc;
michael@0:   __asm__ __volatile__("rdtsc\n\t":"=a"(tsc):);
michael@0:   return tsc;
michael@0: #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
michael@0:   unsigned int tsc;
michael@0:   asm volatile("rdtsc\n\t":"=a"(tsc):);
michael@0:   return tsc;
michael@0: #else
michael@0: #if ARCH_X86_64
michael@0:   return (unsigned int)__rdtsc();
michael@0: #else
michael@0:   __asm  rdtsc;
michael@0: #endif
michael@0: #endif
michael@0: }
michael@0: 
michael@0: 
michael@0: #if defined(__GNUC__) && __GNUC__
michael@0: #define x86_pause_hint()\
michael@0:   __asm__ __volatile__ ("pause \n\t")
michael@0: #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
michael@0: #define x86_pause_hint()\
michael@0:   asm volatile ("pause \n\t")
michael@0: #else
michael@0: #if ARCH_X86_64
michael@0: #define x86_pause_hint()\
michael@0:   _mm_pause();
michael@0: #else
michael@0: #define x86_pause_hint()\
michael@0:   __asm pause
michael@0: #endif
michael@0: #endif
michael@0: 
michael@0: #if defined(__GNUC__) && __GNUC__
michael@0: static void
michael@0: x87_set_control_word(unsigned short mode) {
michael@0:   __asm__ __volatile__("fldcw %0" : : "m"(*&mode));
michael@0: }
michael@0: static unsigned short
michael@0: x87_get_control_word(void) {
michael@0:   unsigned short mode;
michael@0:   __asm__ __volatile__("fstcw %0\n\t":"=m"(*&mode):);
michael@0:     return mode;
michael@0: }
michael@0: #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
michael@0: static void
michael@0: x87_set_control_word(unsigned short mode) {
michael@0:   asm volatile("fldcw %0" : : "m"(*&mode));
michael@0: }
michael@0: static unsigned short
michael@0: x87_get_control_word(void) {
michael@0:   unsigned short mode;
michael@0:   asm volatile("fstcw %0\n\t":"=m"(*&mode):);
michael@0:   return mode;
michael@0: }
michael@0: #elif ARCH_X86_64
michael@0: /* No fldcw intrinsics on Windows x64, punt to external asm */
michael@0: extern void           vpx_winx64_fldcw(unsigned short mode);
michael@0: extern unsigned short vpx_winx64_fstcw(void);
michael@0: #define x87_set_control_word vpx_winx64_fldcw
michael@0: #define x87_get_control_word vpx_winx64_fstcw
michael@0: #else
michael@0: static void
michael@0: x87_set_control_word(unsigned short mode) {
michael@0:   __asm { fldcw mode }
michael@0: }
michael@0: static unsigned short
michael@0: x87_get_control_word(void) {
michael@0:   unsigned short mode;
michael@0:   __asm { fstcw mode }
michael@0:   return mode;
michael@0: }
michael@0: #endif
michael@0: 
michael@0: static unsigned short
michael@0: x87_set_double_precision(void) {
michael@0:   unsigned short mode = x87_get_control_word();
michael@0:   x87_set_control_word((mode&~0x300) | 0x200);
michael@0:   return mode;
michael@0: }
michael@0: 
michael@0: 
michael@0: extern void vpx_reset_mmx_state(void);
michael@0: #endif
michael@0: