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 
12 #ifndef VPX_PORTS_X86_H_
13 #define VPX_PORTS_X86_H_
14 #include <stdlib.h>
15 
16 #if defined(_MSC_VER)
17 #include <intrin.h>  /* For __cpuidex, __rdtsc */
18 #endif
19 
20 #include "vpx_config.h"
21 #include "vpx/vpx_integer.h"
22 
23 #ifdef __cplusplus
24 extern "C" {
25 #endif
26 
27 typedef enum {
28   VPX_CPU_UNKNOWN = -1,
29   VPX_CPU_AMD,
30   VPX_CPU_AMD_OLD,
31   VPX_CPU_CENTAUR,
32   VPX_CPU_CYRIX,
33   VPX_CPU_INTEL,
34   VPX_CPU_NEXGEN,
35   VPX_CPU_NSC,
36   VPX_CPU_RISE,
37   VPX_CPU_SIS,
38   VPX_CPU_TRANSMETA,
39   VPX_CPU_TRANSMETA_OLD,
40   VPX_CPU_UMC,
41   VPX_CPU_VIA,
42 
43   VPX_CPU_LAST
44 }  vpx_cpu_t;
45 
46 #if defined(__GNUC__) && __GNUC__ || defined(__ANDROID__)
47 #if ARCH_X86_64
48 #define cpuid(func, func2, ax, bx, cx, dx)\
49   __asm__ __volatile__ (\
50                         "cpuid           \n\t" \
51                         : "=a" (ax), "=b" (bx), "=c" (cx), "=d" (dx) \
52                         : "a" (func), "c" (func2));
53 #else
54 #define cpuid(func, func2, ax, bx, cx, dx)\
55   __asm__ __volatile__ (\
56                         "mov %%ebx, %%edi   \n\t" \
57                         "cpuid              \n\t" \
58                         "xchg %%edi, %%ebx  \n\t" \
59                         : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
60                         : "a" (func), "c" (func2));
61 #endif
62 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC) /* end __GNUC__ or __ANDROID__*/
63 #if ARCH_X86_64
64 #define cpuid(func, func2, ax, bx, cx, dx)\
65   asm volatile (\
66                 "xchg %rsi, %rbx \n\t" \
67                 "cpuid           \n\t" \
68                 "movl %ebx, %edi \n\t" \
69                 "xchg %rsi, %rbx \n\t" \
70                 : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
71                 : "a" (func), "c" (func2));
72 #else
73 #define cpuid(func, func2, ax, bx, cx, dx)\
74   asm volatile (\
75                 "pushl %ebx       \n\t" \
76                 "cpuid            \n\t" \
77                 "movl %ebx, %edi  \n\t" \
78                 "popl %ebx        \n\t" \
79                 : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
80                 : "a" (func), "c" (func2));
81 #endif
82 #else /* end __SUNPRO__ */
83 #if ARCH_X86_64
84 #if defined(_MSC_VER) && _MSC_VER > 1500
85 #define cpuid(func, func2, a, b, c, d) do {\
86     int regs[4];\
87     __cpuidex(regs, func, func2); \
88     a = regs[0];  b = regs[1];  c = regs[2];  d = regs[3];\
89   } while(0)
90 #else
91 #define cpuid(func, func2, a, b, c, d) do {\
92     int regs[4];\
93     __cpuid(regs, func); \
94     a = regs[0];  b = regs[1];  c = regs[2];  d = regs[3];\
95   } while (0)
96 #endif
97 #else
98 #define cpuid(func, func2, a, b, c, d)\
99   __asm mov eax, func\
100   __asm mov ecx, func2\
101   __asm cpuid\
102   __asm mov a, eax\
103   __asm mov b, ebx\
104   __asm mov c, ecx\
105   __asm mov d, edx
106 #endif
107 #endif /* end others */
108 
109 // NaCl has no support for xgetbv or the raw opcode.
110 #if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__))
xgetbv(void)111 static INLINE uint64_t xgetbv(void) {
112   const uint32_t ecx = 0;
113   uint32_t eax, edx;
114   // Use the raw opcode for xgetbv for compatibility with older toolchains.
115   __asm__ volatile (
116     ".byte 0x0f, 0x01, 0xd0\n"
117     : "=a"(eax), "=d"(edx) : "c" (ecx));
118   return ((uint64_t)edx << 32) | eax;
119 }
120 #elif (defined(_M_X64) || defined(_M_IX86)) && \
121       defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219  // >= VS2010 SP1
122 #include <immintrin.h>
123 #define xgetbv() _xgetbv(0)
124 #elif defined(_MSC_VER) && defined(_M_IX86)
xgetbv(void)125 static INLINE uint64_t xgetbv(void) {
126   uint32_t eax_, edx_;
127   __asm {
128     xor ecx, ecx  // ecx = 0
129     // Use the raw opcode for xgetbv for compatibility with older toolchains.
130     __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0
131     mov eax_, eax
132     mov edx_, edx
133   }
134   return ((uint64_t)edx_ << 32) | eax_;
135 }
136 #else
137 #define xgetbv() 0U  // no AVX for older x64 or unrecognized toolchains.
138 #endif
139 
140 #if defined(_MSC_VER) && _MSC_VER >= 1700
141 #include <windows.h>
142 #if WINAPI_FAMILY_PARTITION(WINAPI_FAMILY_APP)
143 #define getenv(x) NULL
144 #endif
145 #endif
146 
147 #define HAS_MMX     0x01
148 #define HAS_SSE     0x02
149 #define HAS_SSE2    0x04
150 #define HAS_SSE3    0x08
151 #define HAS_SSSE3   0x10
152 #define HAS_SSE4_1  0x20
153 #define HAS_AVX     0x40
154 #define HAS_AVX2    0x80
155 #ifndef BIT
156 #define BIT(n) (1<<n)
157 #endif
158 
159 static INLINE int
x86_simd_caps(void)160 x86_simd_caps(void) {
161   unsigned int flags = 0;
162   unsigned int mask = ~0;
163   unsigned int max_cpuid_val, reg_eax, reg_ebx, reg_ecx, reg_edx;
164   char *env;
165   (void)reg_ebx;
166 
167   /* See if the CPU capabilities are being overridden by the environment */
168   env = getenv("VPX_SIMD_CAPS");
169 
170   if (env && *env)
171     return (int)strtol(env, NULL, 0);
172 
173   env = getenv("VPX_SIMD_CAPS_MASK");
174 
175   if (env && *env)
176     mask = (unsigned int)strtoul(env, NULL, 0);
177 
178   /* Ensure that the CPUID instruction supports extended features */
179   cpuid(0, 0, max_cpuid_val, reg_ebx, reg_ecx, reg_edx);
180 
181   if (max_cpuid_val < 1)
182     return 0;
183 
184   /* Get the standard feature flags */
185   cpuid(1, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
186 
187   if (reg_edx & BIT(23)) flags |= HAS_MMX;
188 
189   if (reg_edx & BIT(25)) flags |= HAS_SSE; /* aka xmm */
190 
191   if (reg_edx & BIT(26)) flags |= HAS_SSE2; /* aka wmt */
192 
193   if (reg_ecx & BIT(0)) flags |= HAS_SSE3;
194 
195   if (reg_ecx & BIT(9)) flags |= HAS_SSSE3;
196 
197   if (reg_ecx & BIT(19)) flags |= HAS_SSE4_1;
198 
199   // bits 27 (OSXSAVE) & 28 (256-bit AVX)
200   if ((reg_ecx & (BIT(27) | BIT(28))) == (BIT(27) | BIT(28))) {
201     if ((xgetbv() & 0x6) == 0x6) {
202       flags |= HAS_AVX;
203 
204       if (max_cpuid_val >= 7) {
205         /* Get the leaf 7 feature flags. Needed to check for AVX2 support */
206         cpuid(7, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
207 
208         if (reg_ebx & BIT(5)) flags |= HAS_AVX2;
209       }
210     }
211   }
212 
213   return flags & mask;
214 }
215 
216 // Note:
217 //  32-bit CPU cycle counter is light-weighted for most function performance
218 //  measurement. For large function (CPU time > a couple of seconds), 64-bit
219 //  counter should be used.
220 // 32-bit CPU cycle counter
221 static INLINE unsigned int
x86_readtsc(void)222 x86_readtsc(void) {
223 #if defined(__GNUC__) && __GNUC__
224   unsigned int tsc;
225   __asm__ __volatile__("rdtsc\n\t":"=a"(tsc):);
226   return tsc;
227 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
228   unsigned int tsc;
229   asm volatile("rdtsc\n\t":"=a"(tsc):);
230   return tsc;
231 #else
232 #if ARCH_X86_64
233   return (unsigned int)__rdtsc();
234 #else
235   __asm  rdtsc;
236 #endif
237 #endif
238 }
239 // 64-bit CPU cycle counter
240 static INLINE uint64_t
x86_readtsc64(void)241 x86_readtsc64(void) {
242 #if defined(__GNUC__) && __GNUC__
243   uint32_t hi, lo;
244   __asm__ __volatile__("rdtsc" : "=a"(lo), "=d"(hi));
245   return ((uint64_t)hi << 32) | lo;
246 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
247   uint_t hi, lo;
248   asm volatile("rdtsc\n\t" : "=a"(lo), "=d"(hi));
249   return ((uint64_t)hi << 32) | lo;
250 #else
251 #if ARCH_X86_64
252   return (uint64_t)__rdtsc();
253 #else
254   __asm  rdtsc;
255 #endif
256 #endif
257 }
258 
259 #if defined(__GNUC__) && __GNUC__
260 #define x86_pause_hint()\
261   __asm__ __volatile__ ("pause \n\t")
262 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
263 #define x86_pause_hint()\
264   asm volatile ("pause \n\t")
265 #else
266 #if ARCH_X86_64
267 #define x86_pause_hint()\
268   _mm_pause();
269 #else
270 #define x86_pause_hint()\
271   __asm pause
272 #endif
273 #endif
274 
275 #if defined(__GNUC__) && __GNUC__
276 static void
x87_set_control_word(unsigned short mode)277 x87_set_control_word(unsigned short mode) {
278   __asm__ __volatile__("fldcw %0" : : "m"(*&mode));
279 }
280 static unsigned short
x87_get_control_word(void)281 x87_get_control_word(void) {
282   unsigned short mode;
283   __asm__ __volatile__("fstcw %0\n\t":"=m"(*&mode):);
284     return mode;
285 }
286 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
287 static void
x87_set_control_word(unsigned short mode)288 x87_set_control_word(unsigned short mode) {
289   asm volatile("fldcw %0" : : "m"(*&mode));
290 }
291 static unsigned short
x87_get_control_word(void)292 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 x87_set_control_word(unsigned short mode) {
306   __asm { fldcw mode }
307 }
308 static unsigned short
x87_get_control_word(void)309 x87_get_control_word(void) {
310   unsigned short mode;
311   __asm { fstcw mode }
312   return mode;
313 }
314 #endif
315 
316 static INLINE unsigned int
x87_set_double_precision(void)317 x87_set_double_precision(void) {
318   unsigned int mode = x87_get_control_word();
319   x87_set_control_word((mode&~0x300) | 0x200);
320   return mode;
321 }
322 
323 
324 extern void vpx_reset_mmx_state(void);
325 
326 #ifdef __cplusplus
327 }  // extern "C"
328 #endif
329 
330 #endif  // VPX_PORTS_X86_H_
331