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.
6 * This file is part of FFmpeg.
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.
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.
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
26 #include "libavutil/x86/asm.h"
27 #include "libavutil/x86/cpu.h"
28 #include "libavutil/cpu.h"
29 #include "libavutil/cpu_internal.h"
33 #define cpuid(index, eax, ebx, ecx, edx) \
34 ff_cpu_cpuid(index, &eax, &ebx, &ecx, &edx)
36 #define xgetbv(index, eax, edx) \
37 ff_cpu_xgetbv(index, &eax, &edx)
41 /* ebx saving is necessary for PIC. gcc seems unable to see it alone */
42 #define cpuid(index, eax, ebx, ecx, edx) \
44 "mov %%"FF_REG_b", %%"FF_REG_S" \n\t" \
46 "xchg %%"FF_REG_b", %%"FF_REG_S \
47 : "=a" (eax), "=S" (ebx), "=c" (ecx), "=d" (edx) \
48 : "0" (index), "2"(0))
50 #define xgetbv(index, eax, edx) \
51 __asm__ (".byte 0x0f, 0x01, 0xd0" : "=a"(eax), "=d"(edx) : "c" (index))
53 #define get_eflags(x) \
54 __asm__ volatile ("pushfl \n" \
58 #define set_eflags(x) \
59 __asm__ volatile ("push %0 \n" \
63 #endif /* HAVE_INLINE_ASM */
67 #define cpuid_test() 1
71 #define cpuid_test ff_cpu_cpuid_test
75 static int cpuid_test(void)
79 /* Check if CPUID is supported by attempting to toggle the ID bit in
80 * the EFLAGS register. */
82 set_eflags(a ^ 0x200000);
89 /* Function to test if multimedia instructions are supported... */
90 int ff_get_cpu_flags_x86(void)
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;
103 return 0; /* CPUID not supported */
105 cpuid(0, max_std_level, vendor.i[0], vendor.i[2], vendor.i[1]);
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;
118 if (std_caps & (1 << 25))
119 rval |= AV_CPU_FLAG_SSE;
120 if (std_caps & (1 << 26))
121 rval |= AV_CPU_FLAG_SSE2;
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 & 0x01000000 )
131 rval |= AV_CPU_FLAG_AESNI;
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;
143 #endif /* HAVE_AVX */
144 #endif /* HAVE_SSE */
146 if (max_std_level >= 7) {
147 cpuid(7, eax, ebx, ecx, edx);
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;
157 #endif /* HAVE_AVX512 */
158 #endif /* HAVE_AVX2 */
159 /* BMI1/2 don't need OS support */
160 if (ebx & 0x00000008) {
161 rval |= AV_CPU_FLAG_BMI1;
162 if (ebx & 0x00000100)
163 rval |= AV_CPU_FLAG_BMI2;
167 cpuid(0x80000000, max_ext_level, ebx, ecx, edx);
169 if (max_ext_level >= 0x80000001) {
170 cpuid(0x80000001, eax, ebx, ecx, ext_caps);
171 if (ext_caps & (1U << 31))
172 rval |= AV_CPU_FLAG_3DNOW;
173 if (ext_caps & (1 << 30))
174 rval |= AV_CPU_FLAG_3DNOWEXT;
175 if (ext_caps & (1 << 23))
176 rval |= AV_CPU_FLAG_MMX;
177 if (ext_caps & (1 << 22))
178 rval |= AV_CPU_FLAG_MMXEXT;
180 if (!strncmp(vendor.c, "AuthenticAMD", 12)) {
181 /* Allow for selectively disabling SSE2 functions on AMD processors
182 with SSE2 support but not SSE4a. This includes Athlon64, some
183 Opteron, and some Sempron processors. MMX, SSE, or 3DNow! are faster
184 than SSE2 often enough to utilize this special-case flag.
185 AV_CPU_FLAG_SSE2 and AV_CPU_FLAG_SSE2SLOW are both set in this case
186 so that SSE2 is used unless explicitly disabled by checking
187 AV_CPU_FLAG_SSE2SLOW. */
188 if (rval & AV_CPU_FLAG_SSE2 && !(ecx & 0x00000040))
189 rval |= AV_CPU_FLAG_SSE2SLOW;
191 /* Similar to the above but for AVX functions on AMD processors.
192 This is necessary only for functions using YMM registers on Bulldozer
193 and Jaguar based CPUs as they lack 256-bit execution units. SSE/AVX
194 functions using XMM registers are always faster on them.
195 AV_CPU_FLAG_AVX and AV_CPU_FLAG_AVXSLOW are both set so that AVX is
196 used unless explicitly disabled by checking AV_CPU_FLAG_AVXSLOW. */
197 if ((family == 0x15 || family == 0x16) && (rval & AV_CPU_FLAG_AVX))
198 rval |= AV_CPU_FLAG_AVXSLOW;
201 /* XOP and FMA4 use the AVX instruction coding scheme, so they can't be
202 * used unless the OS has AVX support. */
203 if (rval & AV_CPU_FLAG_AVX) {
204 if (ecx & 0x00000800)
205 rval |= AV_CPU_FLAG_XOP;
206 if (ecx & 0x00010000)
207 rval |= AV_CPU_FLAG_FMA4;
211 if (!strncmp(vendor.c, "GenuineIntel", 12)) {
212 if (family == 6 && (model == 9 || model == 13 || model == 14)) {
213 /* 6/9 (pentium-m "banias"), 6/13 (pentium-m "dothan"), and
214 * 6/14 (core1 "yonah") theoretically support sse2, but it's
215 * usually slower than mmx, so let's just pretend they don't.
216 * AV_CPU_FLAG_SSE2 is disabled and AV_CPU_FLAG_SSE2SLOW is
217 * enabled so that SSE2 is not used unless explicitly enabled
218 * by checking AV_CPU_FLAG_SSE2SLOW. The same situation
219 * applies for AV_CPU_FLAG_SSE3 and AV_CPU_FLAG_SSE3SLOW. */
220 if (rval & AV_CPU_FLAG_SSE2)
221 rval ^= AV_CPU_FLAG_SSE2SLOW | AV_CPU_FLAG_SSE2;
222 if (rval & AV_CPU_FLAG_SSE3)
223 rval ^= AV_CPU_FLAG_SSE3SLOW | AV_CPU_FLAG_SSE3;
225 /* The Atom processor has SSSE3 support, which is useful in many cases,
226 * but sometimes the SSSE3 version is slower than the SSE2 equivalent
227 * on the Atom, but is generally faster on other processors supporting
228 * SSSE3. This flag allows for selectively disabling certain SSSE3
229 * functions on the Atom. */
230 if (family == 6 && model == 28)
231 rval |= AV_CPU_FLAG_ATOM;
233 /* Conroe has a slow shuffle unit. Check the model number to ensure not
234 * to include crippled low-end Penryns and Nehalems that lack SSE4. */
235 if ((rval & AV_CPU_FLAG_SSSE3) && !(rval & AV_CPU_FLAG_SSE4) &&
236 family == 6 && model < 23)
237 rval |= AV_CPU_FLAG_SSSE3SLOW;
245 size_t ff_get_cpu_max_align_x86(void)
247 int flags = av_get_cpu_flags();
249 if (flags & AV_CPU_FLAG_AVX512)
251 if (flags & (AV_CPU_FLAG_AVX2 |
256 AV_CPU_FLAG_AVXSLOW))
258 if (flags & (AV_CPU_FLAG_AESNI |
266 AV_CPU_FLAG_SSSE3SLOW |
267 AV_CPU_FLAG_SSE3SLOW |
268 AV_CPU_FLAG_SSE2SLOW))