1 /*****************************************************************************
3 *****************************************************************************
4 * Copyright (C) 2003-2014 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Laurent Aimar <fenrir@via.ecp.fr>
8 * Fiona Glaser <fiona@x264.com>
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
24 * This program is also available under a commercial proprietary license.
25 * For more information, contact us at licensing@x264.com.
26 *****************************************************************************/
28 #define _GNU_SOURCE // for sched_getaffinity
32 #if HAVE_POSIXTHREAD && SYS_LINUX
36 #include <kernel/OS.h>
38 #if SYS_MACOSX || SYS_FREEBSD
39 #include <sys/types.h>
40 #include <sys/sysctl.h>
43 #include <sys/param.h>
44 #include <sys/sysctl.h>
45 #include <machine/cpu.h>
48 const x264_cpu_name_t x264_cpu_names[] =
51 // {"MMX", X264_CPU_MMX}, // we don't support asm on mmx1 cpus anymore
52 // {"CMOV", X264_CPU_CMOV}, // we require this unconditionally, so don't print it
53 #define MMX2 X264_CPU_MMX|X264_CPU_MMX2|X264_CPU_CMOV
56 {"SSE", MMX2|X264_CPU_SSE},
57 #define SSE2 MMX2|X264_CPU_SSE|X264_CPU_SSE2
58 {"SSE2Slow", SSE2|X264_CPU_SSE2_IS_SLOW},
60 {"SSE2Fast", SSE2|X264_CPU_SSE2_IS_FAST},
61 {"SSE3", SSE2|X264_CPU_SSE3},
62 {"SSSE3", SSE2|X264_CPU_SSE3|X264_CPU_SSSE3},
63 {"SSE4.1", SSE2|X264_CPU_SSE3|X264_CPU_SSSE3|X264_CPU_SSE4},
64 {"SSE4", SSE2|X264_CPU_SSE3|X264_CPU_SSSE3|X264_CPU_SSE4},
65 {"SSE4.2", SSE2|X264_CPU_SSE3|X264_CPU_SSSE3|X264_CPU_SSE4|X264_CPU_SSE42},
66 #define AVX SSE2|X264_CPU_SSE3|X264_CPU_SSSE3|X264_CPU_SSE4|X264_CPU_SSE42|X264_CPU_AVX
68 {"XOP", AVX|X264_CPU_XOP},
69 {"FMA4", AVX|X264_CPU_FMA4},
70 {"AVX2", AVX|X264_CPU_AVX2},
71 {"FMA3", AVX|X264_CPU_FMA3},
75 {"Cache32", X264_CPU_CACHELINE_32},
76 {"Cache64", X264_CPU_CACHELINE_64},
77 {"LZCNT", X264_CPU_LZCNT},
78 {"BMI1", X264_CPU_BMI1},
79 {"BMI2", X264_CPU_BMI1|X264_CPU_BMI2},
80 {"SlowCTZ", X264_CPU_SLOW_CTZ},
81 {"SlowAtom", X264_CPU_SLOW_ATOM},
82 {"SlowPshufb", X264_CPU_SLOW_PSHUFB},
83 {"SlowPalignr", X264_CPU_SLOW_PALIGNR},
84 {"SlowShuffle", X264_CPU_SLOW_SHUFFLE},
85 {"UnalignedStack", X264_CPU_STACK_MOD4},
87 {"Altivec", X264_CPU_ALTIVEC},
89 {"ARMv6", X264_CPU_ARMV6},
90 {"NEON", X264_CPU_NEON},
91 {"FastNeonMRC", X264_CPU_FAST_NEON_MRC},
96 #if (ARCH_PPC && SYS_LINUX) || (ARCH_ARM && !HAVE_NEON)
99 static sigjmp_buf jmpbuf;
100 static volatile sig_atomic_t canjump = 0;
102 static void sigill_handler( int sig )
106 signal( sig, SIG_DFL );
111 siglongjmp( jmpbuf, 1 );
116 int x264_cpu_cpuid_test( void );
117 void x264_cpu_cpuid( uint32_t op, uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx );
118 void x264_cpu_xgetbv( uint32_t op, uint32_t *eax, uint32_t *edx );
120 uint32_t x264_cpu_detect( void )
123 uint32_t eax, ebx, ecx, edx;
124 uint32_t vendor[4] = {0};
125 uint32_t max_extended_cap, max_basic_cap;
129 if( !x264_cpu_cpuid_test() )
133 x264_cpu_cpuid( 0, &eax, vendor+0, vendor+2, vendor+1 );
135 if( max_basic_cap == 0 )
138 x264_cpu_cpuid( 1, &eax, &ebx, &ecx, &edx );
144 cpu |= X264_CPU_MMX2|X264_CPU_SSE;
146 cpu |= X264_CPU_CMOV;
150 cpu |= X264_CPU_SSE2;
152 cpu |= X264_CPU_SSE3;
154 cpu |= X264_CPU_SSSE3;
156 cpu |= X264_CPU_SSE4;
158 cpu |= X264_CPU_SSE42;
159 /* Check OXSAVE and AVX bits */
160 if( (ecx&0x18000000) == 0x18000000 )
162 /* Check for OS support */
163 x264_cpu_xgetbv( 0, &eax, &edx );
164 if( (eax&0x6) == 0x6 )
168 cpu |= X264_CPU_FMA3;
172 if( max_basic_cap >= 7 )
174 x264_cpu_cpuid( 7, &eax, &ebx, &ecx, &edx );
175 /* AVX2 requires OS support, but BMI1/2 don't. */
176 if( (cpu&X264_CPU_AVX) && (ebx&0x00000020) )
177 cpu |= X264_CPU_AVX2;
180 cpu |= X264_CPU_BMI1;
182 cpu |= X264_CPU_BMI2;
186 if( cpu & X264_CPU_SSSE3 )
187 cpu |= X264_CPU_SSE2_IS_FAST;
189 x264_cpu_cpuid( 0x80000000, &eax, &ebx, &ecx, &edx );
190 max_extended_cap = eax;
192 if( max_extended_cap >= 0x80000001 )
194 x264_cpu_cpuid( 0x80000001, &eax, &ebx, &ecx, &edx );
197 cpu |= X264_CPU_LZCNT; /* Supported by Intel chips starting with Haswell */
198 if( ecx&0x00000040 ) /* SSE4a, AMD only */
200 int family = ((eax>>8)&0xf) + ((eax>>20)&0xff);
201 cpu |= X264_CPU_SSE2_IS_FAST; /* Phenom and later CPUs have fast SSE units */
204 cpu &= ~X264_CPU_SSE2_IS_FAST; /* SSSE3 doesn't imply fast SSE anymore... */
205 cpu |= X264_CPU_SSE2_IS_SLOW; /* Bobcat has 64-bit SIMD units */
206 cpu |= X264_CPU_SLOW_PALIGNR; /* palignr is insanely slow on Bobcat */
210 cpu |= X264_CPU_SLOW_PSHUFB; /* Jaguar's pshufb isn't that slow, but it's slow enough
211 * compared to alternate instruction sequences that this
212 * is equal or faster on almost all such functions. */
216 if( cpu & X264_CPU_AVX )
218 if( ecx&0x00000800 ) /* XOP */
220 if( ecx&0x00010000 ) /* FMA4 */
221 cpu |= X264_CPU_FMA4;
224 if( !strcmp((char*)vendor, "AuthenticAMD") )
227 cpu |= X264_CPU_MMX2;
228 if( !(cpu&X264_CPU_LZCNT) )
229 cpu |= X264_CPU_SLOW_CTZ;
230 if( (cpu&X264_CPU_SSE2) && !(cpu&X264_CPU_SSE2_IS_FAST) )
231 cpu |= X264_CPU_SSE2_IS_SLOW; /* AMD CPUs come in two types: terrible at SSE and great at it */
235 if( !strcmp((char*)vendor, "GenuineIntel") )
237 x264_cpu_cpuid( 1, &eax, &ebx, &ecx, &edx );
238 int family = ((eax>>8)&0xf) + ((eax>>20)&0xff);
239 int model = ((eax>>4)&0xf) + ((eax>>12)&0xf0);
242 /* 6/9 (pentium-m "banias"), 6/13 (pentium-m "dothan"), and 6/14 (core1 "yonah")
243 * theoretically support sse2, but it's significantly slower than mmx for
244 * almost all of x264's functions, so let's just pretend they don't. */
245 if( model == 9 || model == 13 || model == 14 )
247 cpu &= ~(X264_CPU_SSE2|X264_CPU_SSE3);
248 assert(!(cpu&(X264_CPU_SSSE3|X264_CPU_SSE4)));
250 /* Detect Atom CPU */
251 else if( model == 28 )
253 cpu |= X264_CPU_SLOW_ATOM;
254 cpu |= X264_CPU_SLOW_CTZ;
255 cpu |= X264_CPU_SLOW_PSHUFB;
257 /* Conroe has a slow shuffle unit. Check the model number to make sure not
258 * to include crippled low-end Penryns and Nehalems that don't have SSE4. */
259 else if( (cpu&X264_CPU_SSSE3) && !(cpu&X264_CPU_SSE4) && model < 23 )
260 cpu |= X264_CPU_SLOW_SHUFFLE;
264 if( (!strcmp((char*)vendor, "GenuineIntel") || !strcmp((char*)vendor, "CyrixInstead")) && !(cpu&X264_CPU_SSE42))
266 /* cacheline size is specified in 3 places, any of which may be missing */
267 x264_cpu_cpuid( 1, &eax, &ebx, &ecx, &edx );
268 cache = (ebx&0xff00)>>5; // cflush size
269 if( !cache && max_extended_cap >= 0x80000006 )
271 x264_cpu_cpuid( 0x80000006, &eax, &ebx, &ecx, &edx );
272 cache = ecx&0xff; // cacheline size
274 if( !cache && max_basic_cap >= 2 )
276 // Cache and TLB Information
277 static const char cache32_ids[] = { 0x0a, 0x0c, 0x41, 0x42, 0x43, 0x44, 0x45, 0x82, 0x83, 0x84, 0x85, 0 };
278 static const char cache64_ids[] = { 0x22, 0x23, 0x25, 0x29, 0x2c, 0x46, 0x47, 0x49, 0x60, 0x66, 0x67,
279 0x68, 0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7c, 0x7f, 0x86, 0x87, 0 };
283 x264_cpu_cpuid( 2, buf+0, buf+1, buf+2, buf+3 );
286 for( int j = 0; j < 4; j++ )
290 if( strchr( cache32_ids, buf[j]&0xff ) )
292 if( strchr( cache64_ids, buf[j]&0xff ) )
296 } while( ++i < max );
300 cpu |= X264_CPU_CACHELINE_32;
301 else if( cache == 64 )
302 cpu |= X264_CPU_CACHELINE_64;
304 x264_log( NULL, X264_LOG_WARNING, "unable to determine cacheline size\n" );
307 #if BROKEN_STACK_ALIGNMENT
308 cpu |= X264_CPU_STACK_MOD4;
316 #if SYS_MACOSX || SYS_OPENBSD
317 #include <sys/sysctl.h>
318 uint32_t x264_cpu_detect( void )
323 int selectors[2] = { CTL_MACHDEP, CPU_ALTIVEC };
325 int selectors[2] = { CTL_HW, HW_VECTORUNIT };
328 size_t length = sizeof( has_altivec );
329 int error = sysctl( selectors, 2, &has_altivec, &length, NULL, 0 );
331 if( error == 0 && has_altivec != 0 )
332 cpu |= X264_CPU_ALTIVEC;
339 uint32_t x264_cpu_detect( void )
341 static void (*oldsig)( int );
343 oldsig = signal( SIGILL, sigill_handler );
344 if( sigsetjmp( jmpbuf, 1 ) )
346 signal( SIGILL, oldsig );
351 asm volatile( "mtspr 256, %0\n\t"
357 signal( SIGILL, oldsig );
359 return X264_CPU_ALTIVEC;
365 void x264_cpu_neon_test( void );
366 int x264_cpu_fast_neon_mrc_test( void );
368 uint32_t x264_cpu_detect( void )
372 flags |= X264_CPU_ARMV6;
374 // don't do this hack if compiled with -mfpu=neon
376 static void (* oldsig)( int );
377 oldsig = signal( SIGILL, sigill_handler );
378 if( sigsetjmp( jmpbuf, 1 ) )
380 signal( SIGILL, oldsig );
385 x264_cpu_neon_test();
387 signal( SIGILL, oldsig );
390 flags |= X264_CPU_NEON;
392 // fast neon -> arm (Cortex-A9) detection relies on user access to the
393 // cycle counter; this assumes ARMv7 performance counters.
394 // NEON requires at least ARMv7, ARMv8 may require changes here, but
395 // hopefully this hacky detection method will have been replaced by then.
396 // Note that there is potential for a race condition if another program or
397 // x264 instance disables or reinits the counters while x264 is using them,
398 // which may result in incorrect detection and the counters stuck enabled.
399 // right now Apple does not seem to support performance counters for this test
401 flags |= x264_cpu_fast_neon_mrc_test() ? X264_CPU_FAST_NEON_MRC : 0;
403 // TODO: write dual issue test? currently it's A8 (dual issue) vs. A9 (fast mrc)
410 uint32_t x264_cpu_detect( void )
417 int x264_cpu_num_processors( void )
423 return x264_pthread_num_processors_np();
425 #elif SYS_CYGWIN || SYS_SunOS
426 return sysconf( _SC_NPROCESSORS_ONLN );
430 memset( &p_aff, 0, sizeof(p_aff) );
431 if( sched_getaffinity( 0, sizeof(p_aff), &p_aff ) )
434 return CPU_COUNT(&p_aff);
437 for( unsigned int bit = 0; bit < 8 * sizeof(p_aff); bit++ )
438 np += (((uint8_t *)&p_aff)[bit / 8] >> (bit % 8)) & 1;
444 get_system_info( &info );
445 return info.cpu_count;
447 #elif SYS_MACOSX || SYS_FREEBSD || SYS_OPENBSD
449 size_t length = sizeof( ncpu );
451 int mib[2] = { CTL_HW, HW_NCPU };
452 if( sysctl(mib, 2, &ncpu, &length, NULL, 0) )
454 if( sysctlbyname("hw.ncpu", &ncpu, &length, NULL, 0) )