FFmpeg
cpu.c
Go to the documentation of this file.
1 /*
2  * CPU detection code, extracted from mmx.h
3  * (c)1997-99 by H. Dietz and R. Fisher
4  * Converted to C and improved by Fabrice Bellard.
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
23 #include <stdlib.h>
24 #include <string.h>
25 
26 #include "libavutil/x86/asm.h"
27 #include "libavutil/x86/cpu.h"
28 #include "libavutil/cpu.h"
29 #include "libavutil/cpu_internal.h"
30 
31 #if HAVE_X86ASM
32 
33 #define cpuid(index, eax, ebx, ecx, edx) \
34  ff_cpu_cpuid(index, &eax, &ebx, &ecx, &edx)
35 
36 #define xgetbv(index, eax, edx) \
37  ff_cpu_xgetbv(index, &eax, &edx)
38 
39 #elif HAVE_INLINE_ASM
40 
41 /* ebx saving is necessary for PIC. gcc seems unable to see it alone */
42 #define cpuid(index, eax, ebx, ecx, edx) \
43  __asm__ volatile ( \
44  "mov %%"FF_REG_b", %%"FF_REG_S" \n\t" \
45  "cpuid \n\t" \
46  "xchg %%"FF_REG_b", %%"FF_REG_S \
47  : "=a" (eax), "=S" (ebx), "=c" (ecx), "=d" (edx) \
48  : "0" (index), "2"(0))
49 
50 #define xgetbv(index, eax, edx) \
51  __asm__ (".byte 0x0f, 0x01, 0xd0" : "=a"(eax), "=d"(edx) : "c" (index))
52 
53 #define get_eflags(x) \
54  __asm__ volatile ("pushfl \n" \
55  "pop %0 \n" \
56  : "=r"(x))
57 
58 #define set_eflags(x) \
59  __asm__ volatile ("push %0 \n" \
60  "popfl \n" \
61  :: "r"(x))
62 
63 #endif /* HAVE_INLINE_ASM */
64 
65 #if ARCH_X86_64
66 
67 #define cpuid_test() 1
68 
69 #elif HAVE_X86ASM
70 
71 #define cpuid_test ff_cpu_cpuid_test
72 
73 #elif HAVE_INLINE_ASM
74 
75 static int cpuid_test(void)
76 {
77  x86_reg a, c;
78 
79  /* Check if CPUID is supported by attempting to toggle the ID bit in
80  * the EFLAGS register. */
81  get_eflags(a);
82  set_eflags(a ^ 0x200000);
83  get_eflags(c);
84 
85  return a != c;
86 }
87 #endif
88 
89 /* Function to test if multimedia instructions are supported... */
90 int ff_get_cpu_flags_x86(void)
91 {
92  int rval = 0;
93 
94 #ifdef cpuid
95 
96  int eax, ebx, ecx, edx;
97  int max_std_level, max_ext_level, std_caps = 0, ext_caps = 0;
98  int family = 0, model = 0;
99  union { int i[3]; char c[12]; } vendor;
100  int xcr0_lo = 0, xcr0_hi = 0;
101 
102  if (!cpuid_test())
103  return 0; /* CPUID not supported */
104 
105  cpuid(0, max_std_level, vendor.i[0], vendor.i[2], vendor.i[1]);
106 
107  if (max_std_level >= 1) {
108  cpuid(1, eax, ebx, ecx, std_caps);
109  family = ((eax >> 8) & 0xf) + ((eax >> 20) & 0xff);
110  model = ((eax >> 4) & 0xf) + ((eax >> 12) & 0xf0);
111  if (std_caps & (1 << 15))
112  rval |= AV_CPU_FLAG_CMOV;
113  if (std_caps & (1 << 23))
114  rval |= AV_CPU_FLAG_MMX;
115  if (std_caps & (1 << 25))
116  rval |= AV_CPU_FLAG_MMXEXT;
117 #if HAVE_SSE
118  if (std_caps & (1 << 25))
119  rval |= AV_CPU_FLAG_SSE;
120  if (std_caps & (1 << 26))
121  rval |= AV_CPU_FLAG_SSE2;
122  if (ecx & 1)
123  rval |= AV_CPU_FLAG_SSE3;
124  if (ecx & 0x00000200 )
125  rval |= AV_CPU_FLAG_SSSE3;
126  if (ecx & 0x00080000 )
127  rval |= AV_CPU_FLAG_SSE4;
128  if (ecx & 0x00100000 )
129  rval |= AV_CPU_FLAG_SSE42;
130  if (ecx & 0x02000000 )
131  rval |= AV_CPU_FLAG_AESNI;
132 #if HAVE_AVX
133  /* Check OXSAVE and AVX bits */
134  if ((ecx & 0x18000000) == 0x18000000) {
135  /* Check for OS support */
136  xgetbv(0, xcr0_lo, xcr0_hi);
137  if ((xcr0_lo & 0x6) == 0x6) {
138  rval |= AV_CPU_FLAG_AVX;
139  if (ecx & 0x00001000)
140  rval |= AV_CPU_FLAG_FMA3;
141  }
142  }
143 #endif /* HAVE_AVX */
144 #endif /* HAVE_SSE */
145  }
146  if (max_std_level >= 7) {
147  cpuid(7, eax, ebx, ecx, edx);
148 #if HAVE_AVX2
149  if ((rval & AV_CPU_FLAG_AVX) && (ebx & 0x00000020))
150  rval |= AV_CPU_FLAG_AVX2;
151 #if HAVE_AVX512 /* F, CD, BW, DQ, VL */
152  if ((xcr0_lo & 0xe0) == 0xe0) { /* OPMASK/ZMM state */
153  if ((rval & AV_CPU_FLAG_AVX2) && (ebx & 0xd0030000) == 0xd0030000) {
154  rval |= AV_CPU_FLAG_AVX512;
155 #if HAVE_AVX512ICL
156  if ((ebx & 0xd0200000) == 0xd0200000 && (ecx & 0x5f42) == 0x5f42)
157  rval |= AV_CPU_FLAG_AVX512ICL;
158 #endif /* HAVE_AVX512ICL */
159  }
160  }
161 #endif /* HAVE_AVX512 */
162 #endif /* HAVE_AVX2 */
163  /* BMI1/2 don't need OS support */
164  if (ebx & 0x00000008) {
165  rval |= AV_CPU_FLAG_BMI1;
166  if (ebx & 0x00000100)
167  rval |= AV_CPU_FLAG_BMI2;
168  }
169  }
170 
171  cpuid(0x80000000, max_ext_level, ebx, ecx, edx);
172 
173  if (max_ext_level >= 0x80000001) {
174  cpuid(0x80000001, eax, ebx, ecx, ext_caps);
175  if (ext_caps & (1U << 31))
176  rval |= AV_CPU_FLAG_3DNOW;
177  if (ext_caps & (1 << 30))
178  rval |= AV_CPU_FLAG_3DNOWEXT;
179  if (ext_caps & (1 << 23))
180  rval |= AV_CPU_FLAG_MMX;
181  if (ext_caps & (1 << 22))
182  rval |= AV_CPU_FLAG_MMXEXT;
183 
184  if (!strncmp(vendor.c, "AuthenticAMD", 12)) {
185  /* Allow for selectively disabling SSE2 functions on AMD processors
186  with SSE2 support but not SSE4a. This includes Athlon64, some
187  Opteron, and some Sempron processors. MMX, SSE, or 3DNow! are faster
188  than SSE2 often enough to utilize this special-case flag.
189  AV_CPU_FLAG_SSE2 and AV_CPU_FLAG_SSE2SLOW are both set in this case
190  so that SSE2 is used unless explicitly disabled by checking
191  AV_CPU_FLAG_SSE2SLOW. */
192  if (rval & AV_CPU_FLAG_SSE2 && !(ecx & 0x00000040))
193  rval |= AV_CPU_FLAG_SSE2SLOW;
194 
195  /* Similar to the above but for AVX functions on AMD processors.
196  This is necessary only for functions using YMM registers on Bulldozer
197  and Jaguar based CPUs as they lack 256-bit execution units. SSE/AVX
198  functions using XMM registers are always faster on them.
199  AV_CPU_FLAG_AVX and AV_CPU_FLAG_AVXSLOW are both set so that AVX is
200  used unless explicitly disabled by checking AV_CPU_FLAG_AVXSLOW. */
201  if ((family == 0x15 || family == 0x16) && (rval & AV_CPU_FLAG_AVX))
202  rval |= AV_CPU_FLAG_AVXSLOW;
203 
204  /* Zen 3 and earlier have slow gather */
205  if ((family <= 0x19) && (rval & AV_CPU_FLAG_AVX2))
206  rval |= AV_CPU_FLAG_SLOW_GATHER;
207  }
208 
209  /* XOP and FMA4 use the AVX instruction coding scheme, so they can't be
210  * used unless the OS has AVX support. */
211  if (rval & AV_CPU_FLAG_AVX) {
212  if (ecx & 0x00000800)
213  rval |= AV_CPU_FLAG_XOP;
214  if (ecx & 0x00010000)
215  rval |= AV_CPU_FLAG_FMA4;
216  }
217  }
218 
219  if (!strncmp(vendor.c, "GenuineIntel", 12)) {
220  if (family == 6 && (model == 9 || model == 13 || model == 14)) {
221  /* 6/9 (pentium-m "banias"), 6/13 (pentium-m "dothan"), and
222  * 6/14 (core1 "yonah") theoretically support sse2, but it's
223  * usually slower than mmx, so let's just pretend they don't.
224  * AV_CPU_FLAG_SSE2 is disabled and AV_CPU_FLAG_SSE2SLOW is
225  * enabled so that SSE2 is not used unless explicitly enabled
226  * by checking AV_CPU_FLAG_SSE2SLOW. The same situation
227  * applies for AV_CPU_FLAG_SSE3 and AV_CPU_FLAG_SSE3SLOW. */
228  if (rval & AV_CPU_FLAG_SSE2)
229  rval ^= AV_CPU_FLAG_SSE2SLOW | AV_CPU_FLAG_SSE2;
230  if (rval & AV_CPU_FLAG_SSE3)
231  rval ^= AV_CPU_FLAG_SSE3SLOW | AV_CPU_FLAG_SSE3;
232  }
233  /* The Atom processor has SSSE3 support, which is useful in many cases,
234  * but sometimes the SSSE3 version is slower than the SSE2 equivalent
235  * on the Atom, but is generally faster on other processors supporting
236  * SSSE3. This flag allows for selectively disabling certain SSSE3
237  * functions on the Atom. */
238  if (family == 6 && model == 28)
239  rval |= AV_CPU_FLAG_ATOM;
240 
241  /* Conroe has a slow shuffle unit. Check the model number to ensure not
242  * to include crippled low-end Penryns and Nehalems that lack SSE4. */
243  if ((rval & AV_CPU_FLAG_SSSE3) && !(rval & AV_CPU_FLAG_SSE4) &&
244  family == 6 && model < 23)
245  rval |= AV_CPU_FLAG_SSSE3SLOW;
246 
247  /* Haswell has slow gather */
248  if ((rval & AV_CPU_FLAG_AVX2) && family == 6 && model < 70)
249  rval |= AV_CPU_FLAG_SLOW_GATHER;
250  }
251 
252 #endif /* cpuid */
253 
254  return rval;
255 }
256 
257 size_t ff_get_cpu_max_align_x86(void)
258 {
259  int flags = av_get_cpu_flags();
260 
261  if (flags & AV_CPU_FLAG_AVX512)
262  return 64;
263  if (flags & (AV_CPU_FLAG_AVX2 |
264  AV_CPU_FLAG_AVX |
265  AV_CPU_FLAG_XOP |
266  AV_CPU_FLAG_FMA4 |
267  AV_CPU_FLAG_FMA3 |
268  AV_CPU_FLAG_AVXSLOW))
269  return 32;
270  if (flags & (AV_CPU_FLAG_AESNI |
271  AV_CPU_FLAG_SSE42 |
272  AV_CPU_FLAG_SSE4 |
273  AV_CPU_FLAG_SSSE3 |
274  AV_CPU_FLAG_SSE3 |
275  AV_CPU_FLAG_SSE2 |
276  AV_CPU_FLAG_SSE |
277  AV_CPU_FLAG_ATOM |
278  AV_CPU_FLAG_SSSE3SLOW |
279  AV_CPU_FLAG_SSE3SLOW |
280  AV_CPU_FLAG_SSE2SLOW))
281  return 16;
282 
283  return 8;
284 }
AV_CPU_FLAG_SSSE3SLOW
#define AV_CPU_FLAG_SSSE3SLOW
SSSE3 supported, but usually not faster.
Definition: cpu.h:42
AV_CPU_FLAG_SSE3
#define AV_CPU_FLAG_SSE3
Prescott SSE3 functions.
Definition: cpu.h:38
cpu.h
AV_CPU_FLAG_SSE3SLOW
#define AV_CPU_FLAG_SSE3SLOW
SSE3 supported, but usually not faster.
Definition: cpu.h:39
AV_CPU_FLAG_3DNOW
#define AV_CPU_FLAG_3DNOW
AMD 3DNOW.
Definition: cpu.h:32
av_get_cpu_flags
int av_get_cpu_flags(void)
Return the flags which specify extensions supported by the CPU.
Definition: cpu.c:103
AV_CPU_FLAG_BMI1
#define AV_CPU_FLAG_BMI1
Bit Manipulation Instruction Set 1.
Definition: cpu.h:54
AV_CPU_FLAG_SSSE3
#define AV_CPU_FLAG_SSSE3
Conroe SSSE3 functions.
Definition: cpu.h:41
AV_CPU_FLAG_XOP
#define AV_CPU_FLAG_XOP
Bulldozer XOP functions.
Definition: cpu.h:49
AV_CPU_FLAG_3DNOWEXT
#define AV_CPU_FLAG_3DNOWEXT
AMD 3DNowExt.
Definition: cpu.h:37
AV_CPU_FLAG_SLOW_GATHER
#define AV_CPU_FLAG_SLOW_GATHER
CPU has slow gathers.
Definition: cpu.h:58
AV_CPU_FLAG_AVX512
#define AV_CPU_FLAG_AVX512
AVX-512 functions: requires OS support even if YMM/ZMM registers aren't used.
Definition: cpu.h:56
ff_get_cpu_max_align_x86
size_t ff_get_cpu_max_align_x86(void)
Definition: cpu.c:257
cpu_internal.h
AV_CPU_FLAG_SSE4
#define AV_CPU_FLAG_SSE4
Penryn SSE4.1 functions.
Definition: cpu.h:44
AV_CPU_FLAG_AVX512ICL
#define AV_CPU_FLAG_AVX512ICL
F/CD/BW/DQ/VL/VNNI/IFMA/VBMI/VBMI2/VPOPCNTDQ/BITALG/GFNI/VAES/VPCLMULQDQ.
Definition: cpu.h:57
AV_CPU_FLAG_CMOV
#define AV_CPU_FLAG_CMOV
supports cmov instruction
Definition: cpu.h:51
c
Undefined Behavior In the C some operations are like signed integer dereferencing freed accessing outside allocated Undefined Behavior must not occur in a C it is not safe even if the output of undefined operations is unused The unsafety may seem nit picking but Optimizing compilers have in fact optimized code on the assumption that no undefined Behavior occurs Optimizing code based on wrong assumptions can and has in some cases lead to effects beyond the output of computations The signed integer overflow problem in speed critical code Code which is highly optimized and works with signed integers sometimes has the problem that often the output of the computation does not c
Definition: undefined.txt:32
AV_CPU_FLAG_SSE2
#define AV_CPU_FLAG_SSE2
PIV SSE2 functions.
Definition: cpu.h:34
AV_CPU_FLAG_AVXSLOW
#define AV_CPU_FLAG_AVXSLOW
AVX supported, but slow when using YMM registers (e.g. Bulldozer)
Definition: cpu.h:48
AV_CPU_FLAG_AVX
#define AV_CPU_FLAG_AVX
AVX functions: requires OS support even if YMM registers aren't used.
Definition: cpu.h:47
AV_CPU_FLAG_FMA4
#define AV_CPU_FLAG_FMA4
Bulldozer FMA4 functions.
Definition: cpu.h:50
cpu.h
AV_CPU_FLAG_AVX2
#define AV_CPU_FLAG_AVX2
AVX2 functions: requires OS support even if YMM registers aren't used.
Definition: cpu.h:52
AV_CPU_FLAG_SSE2SLOW
#define AV_CPU_FLAG_SSE2SLOW
SSE2 supported, but usually not faster.
Definition: cpu.h:35
asm.h
AV_CPU_FLAG_FMA3
#define AV_CPU_FLAG_FMA3
Haswell FMA3 functions.
Definition: cpu.h:53
a
The reader does not expect b to be semantically here and if the code is changed by maybe adding a a division or other the signedness will almost certainly be mistaken To avoid this confusion a new type was SUINT is the C unsigned type but it holds a signed int to use the same example SUINT a
Definition: undefined.txt:41
i
#define i(width, name, range_min, range_max)
Definition: cbs_h2645.c:255
AV_CPU_FLAG_SSE42
#define AV_CPU_FLAG_SSE42
Nehalem SSE4.2 functions.
Definition: cpu.h:45
xf
#define xf(width, name, var, range_min, range_max, subs,...)
Definition: cbs_av1.c:598
ff_get_cpu_flags_x86
int ff_get_cpu_flags_x86(void)
Definition: cpu.c:90
AV_CPU_FLAG_ATOM
#define AV_CPU_FLAG_ATOM
Atom processor, some SSSE3 instructions are slower.
Definition: cpu.h:43
U
#define U(x)
Definition: vpx_arith.h:37
AV_CPU_FLAG_MMX
#define AV_CPU_FLAG_MMX
standard MMX
Definition: cpu.h:29
AV_CPU_FLAG_AESNI
#define AV_CPU_FLAG_AESNI
Advanced Encryption Standard functions.
Definition: cpu.h:46
AV_CPU_FLAG_SSE
#define AV_CPU_FLAG_SSE
SSE functions.
Definition: cpu.h:33
AV_CPU_FLAG_MMXEXT
#define AV_CPU_FLAG_MMXEXT
SSE integer functions or AMD MMX ext.
Definition: cpu.h:30
AV_CPU_FLAG_BMI2
#define AV_CPU_FLAG_BMI2
Bit Manipulation Instruction Set 2.
Definition: cpu.h:55
x86_reg
int x86_reg
Definition: asm.h:72
flags
#define flags(name, subs,...)
Definition: cbs_av1.c:482
Generated on Thu Sep 26 2024 23:15:53 for FFmpeg by   doxygen 1.8.17