1 /**************************************************************************
3 * Copyright 2008 Dennis Smit
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * on the rights to use, copy, modify, merge, publish, distribute, sub
10 * license, and/or sell copies of the Software, and to permit persons to whom
11 * the Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
20 * AUTHORS, COPYRIGHT HOLDERS, AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
21 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
22 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
23 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 **************************************************************************/
29 * CPU feature detection.
32 * @author Based on the work of Eric Anholt <anholt@FreeBSD.org>
35 #include "pipe/p_config.h"
37 #include "util/u_debug.h"
38 #include "u_cpu_detect.h"
39 #include "c11/threads.h"
41 #if defined(PIPE_ARCH_PPC)
42 #if defined(PIPE_OS_APPLE)
43 #include <sys/sysctl.h>
50 #if defined(PIPE_OS_NETBSD) || defined(PIPE_OS_OPENBSD)
51 #include <sys/param.h>
52 #include <sys/sysctl.h>
53 #include <machine/cpu.h>
56 #if defined(PIPE_OS_FREEBSD) || defined(PIPE_OS_DRAGONFLY)
57 #include <sys/types.h>
58 #include <sys/sysctl.h>
59 #if __has_include(<sys/auxv.h>)
61 #define HAVE_ELF_AUX_INFO
65 #if defined(PIPE_OS_LINUX)
75 #if defined(HAS_ANDROID_CPUFEATURES)
76 #include <cpu-features.h>
79 #if defined(PIPE_OS_WINDOWS)
81 #if defined(PIPE_CC_MSVC)
88 DEBUG_GET_ONCE_BOOL_OPTION(dump_cpu
, "GALLIUM_DUMP_CPU", FALSE
)
92 struct util_cpu_caps util_cpu_caps
;
94 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
95 static int has_cpuid(void);
99 #if defined(PIPE_ARCH_PPC) && !defined(PIPE_OS_APPLE) && !defined(PIPE_OS_LINUX)
100 static jmp_buf __lv_powerpc_jmpbuf
;
101 static volatile sig_atomic_t __lv_powerpc_canjump
= 0;
104 sigill_handler(int sig
)
106 if (!__lv_powerpc_canjump
) {
107 signal (sig
, SIG_DFL
);
111 __lv_powerpc_canjump
= 0;
112 longjmp(__lv_powerpc_jmpbuf
, 1);
116 #if defined(PIPE_ARCH_PPC)
118 check_os_altivec_support(void)
120 #if defined(PIPE_OS_APPLE)
121 int sels
[2] = {CTL_HW
, HW_VECTORUNIT
};
123 int len
= sizeof (has_vu
);
126 err
= sysctl(sels
, 2, &has_vu
, &len
, NULL
, 0);
130 util_cpu_caps
.has_altivec
= 1;
133 #elif defined(PIPE_OS_LINUX) /* !PIPE_OS_APPLE */
134 #if defined(PIPE_ARCH_PPC_64)
139 int fd
= open("/proc/self/auxv", O_RDONLY
| O_CLOEXEC
);
141 while (read(fd
, &aux
, sizeof(aux
)) == sizeof(aux
)) {
142 if (aux
.a_type
== AT_HWCAP
) {
143 char *env_vsx
= getenv("GALLIVM_VSX");
144 uint64_t hwcap
= aux
.a_un
.a_val
;
145 util_cpu_caps
.has_altivec
= (hwcap
>> 28) & 1;
146 if (!env_vsx
|| env_vsx
[0] != '0') {
147 util_cpu_caps
.has_vsx
= (hwcap
>> 7) & 1;
154 #else /* !PIPE_OS_APPLE && !PIPE_OS_LINUX */
155 /* not on Apple/Darwin or Linux, do it the brute-force way */
156 /* this is borrowed from the libmpeg2 library */
157 signal(SIGILL
, sigill_handler
);
158 if (setjmp(__lv_powerpc_jmpbuf
)) {
159 signal(SIGILL
, SIG_DFL
);
161 boolean enable_altivec
= TRUE
; /* Default: enable if available, and if not overridden */
162 boolean enable_vsx
= TRUE
;
164 /* Disabling Altivec code generation is not the same as disabling VSX code generation,
165 * which can be done simply by passing -mattr=-vsx to the LLVM compiler; cf.
166 * lp_build_create_jit_compiler_for_module().
167 * If you want to disable Altivec code generation, the best place to do it is here.
169 char *env_control
= getenv("GALLIVM_ALTIVEC"); /* 1=enable (default); 0=disable */
170 if (env_control
&& env_control
[0] == '0') {
171 enable_altivec
= FALSE
;
174 /* VSX instructions can be explicitly enabled/disabled via GALLIVM_VSX=1 or 0 */
175 char *env_vsx
= getenv("GALLIVM_VSX");
176 if (env_vsx
&& env_vsx
[0] == '0') {
179 if (enable_altivec
) {
180 __lv_powerpc_canjump
= 1;
184 "vand %%v0, %%v0, %%v0"
188 util_cpu_caps
.has_altivec
= 1;
191 __asm
__volatile("xxland %vs0, %vs0, %vs0");
192 util_cpu_caps
.has_vsx
= 1;
194 signal(SIGILL
, SIG_DFL
);
196 util_cpu_caps
.has_altivec
= 0;
199 #endif /* !PIPE_OS_APPLE && !PIPE_OS_LINUX */
201 #endif /* PIPE_ARCH_PPC */
204 #if defined(PIPE_ARCH_X86) || defined (PIPE_ARCH_X86_64)
205 static int has_cpuid(void)
207 #if defined(PIPE_ARCH_X86)
208 #if defined(PIPE_OS_GCC)
215 "xorl $0x200000, %0\n"
229 #elif defined(PIPE_ARCH_X86_64)
238 * @sa cpuid.h included in gcc-4.3 onwards.
239 * @sa http://msdn.microsoft.com/en-us/library/hskdteyh.aspx
242 cpuid(uint32_t ax
, uint32_t *p
)
244 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
246 "xchgl %%ebx, %1\n\t"
255 #elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
264 #elif defined(PIPE_CC_MSVC)
275 * @sa cpuid.h included in gcc-4.4 onwards.
276 * @sa http://msdn.microsoft.com/en-us/library/hskdteyh%28v=vs.90%29.aspx
279 cpuid_count(uint32_t ax
, uint32_t cx
, uint32_t *p
)
281 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
283 "xchgl %%ebx, %1\n\t"
292 #elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
301 #elif defined(PIPE_CC_MSVC)
302 __cpuidex(p
, ax
, cx
);
312 static inline uint64_t xgetbv(void)
314 #if defined(PIPE_CC_GCC)
318 ".byte 0x0f, 0x01, 0xd0" // xgetbv isn't supported on gcc < 4.4
324 return ((uint64_t)edx
<< 32) | eax
;
325 #elif defined(PIPE_CC_MSVC) && defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
326 return _xgetbv(_XCR_XFEATURE_ENABLED_MASK
);
333 #if defined(PIPE_ARCH_X86)
334 PIPE_ALIGN_STACK
static inline boolean
sse2_has_daz(void)
339 uint32_t pad2
[128-8];
340 } PIPE_ALIGN_VAR(16) fxarea
;
342 fxarea
.mxcsr_mask
= 0;
343 #if defined(PIPE_CC_GCC)
344 __asm
__volatile ("fxsave %0" : "+m" (fxarea
));
345 #elif defined(PIPE_CC_MSVC) || defined(PIPE_CC_ICL)
348 fxarea
.mxcsr_mask
= 0;
350 return !!(fxarea
.mxcsr_mask
& (1 << 6));
354 #endif /* X86 or X86_64 */
356 #if defined(PIPE_ARCH_ARM)
358 check_os_arm_support(void)
361 * On Android, the cpufeatures library is preferred way of checking
362 * CPU capabilities. However, it is not available for standalone Mesa
363 * builds, i.e. when Android build system (Android.mk-based) is not
364 * used. Because of this we cannot use PIPE_OS_ANDROID here, but rather
365 * have a separate macro that only gets enabled from respective Android.mk.
367 #if defined(__ARM_NEON) || defined(__ARM_NEON__)
368 util_cpu_caps
.has_neon
= 1;
369 #elif defined(PIPE_OS_FREEBSD) && defined(HAVE_ELF_AUX_INFO)
370 unsigned long hwcap
= 0;
371 elf_aux_info(AT_HWCAP
, &hwcap
, sizeof(hwcap
));
372 if (hwcap
& HWCAP_NEON
)
373 util_cpu_caps
.has_neon
= 1;
374 #elif defined(HAS_ANDROID_CPUFEATURES)
375 AndroidCpuFamily cpu_family
= android_getCpuFamily();
376 uint64_t cpu_features
= android_getCpuFeatures();
378 if (cpu_family
== ANDROID_CPU_FAMILY_ARM
) {
379 if (cpu_features
& ANDROID_CPU_ARM_FEATURE_NEON
)
380 util_cpu_caps
.has_neon
= 1;
382 #elif defined(PIPE_OS_LINUX)
386 fd
= open("/proc/self/auxv", O_RDONLY
| O_CLOEXEC
);
388 while (read(fd
, &aux
, sizeof(Elf32_auxv_t
)) == sizeof(Elf32_auxv_t
)) {
389 if (aux
.a_type
== AT_HWCAP
) {
390 uint32_t hwcap
= aux
.a_un
.a_val
;
392 util_cpu_caps
.has_neon
= (hwcap
>> 12) & 1;
398 #endif /* PIPE_OS_LINUX */
401 #elif defined(PIPE_ARCH_AARCH64)
403 check_os_arm_support(void)
405 util_cpu_caps
.has_neon
= true;
407 #endif /* PIPE_ARCH_ARM || PIPE_ARCH_AARCH64 */
410 get_cpu_topology(void)
412 /* Default. This is correct if L3 is not present or there is only one. */
413 util_cpu_caps
.cores_per_L3
= util_cpu_caps
.nr_cpus
;
415 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
417 if (util_cpu_caps
.x86_cpu_type
== 0x17) {
420 /* Query the L3 cache topology information. */
421 cpuid_count(0x8000001D, 3, regs
);
422 unsigned cache_level
= (regs
[0] >> 5) & 0x7;
423 unsigned cores_per_cache
= ((regs
[0] >> 14) & 0xfff) + 1;
425 if (cache_level
== 3)
426 util_cpu_caps
.cores_per_L3
= cores_per_cache
;
432 util_cpu_detect_once(void)
434 memset(&util_cpu_caps
, 0, sizeof util_cpu_caps
);
436 /* Count the number of CPUs in system */
437 #if defined(PIPE_OS_WINDOWS)
439 SYSTEM_INFO system_info
;
440 GetSystemInfo(&system_info
);
441 util_cpu_caps
.nr_cpus
= system_info
.dwNumberOfProcessors
;
443 #elif defined(PIPE_OS_UNIX) && defined(_SC_NPROCESSORS_ONLN)
444 util_cpu_caps
.nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
445 if (util_cpu_caps
.nr_cpus
== ~0)
446 util_cpu_caps
.nr_cpus
= 1;
447 #elif defined(PIPE_OS_BSD)
456 sysctl(mib
, 2, &ncpu
, &len
, NULL
, 0);
457 util_cpu_caps
.nr_cpus
= ncpu
;
460 util_cpu_caps
.nr_cpus
= 1;
463 /* Make the fallback cacheline size nonzero so that it can be
464 * safely passed to align().
466 util_cpu_caps
.cacheline
= sizeof(void *);
468 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
473 util_cpu_caps
.cacheline
= 32;
475 /* Get max cpuid level */
476 cpuid(0x00000000, regs
);
478 if (regs
[0] >= 0x00000001) {
479 unsigned int cacheline
;
481 cpuid (0x00000001, regs2
);
483 util_cpu_caps
.x86_cpu_type
= (regs2
[0] >> 8) & 0xf;
484 /* Add "extended family". */
485 if (util_cpu_caps
.x86_cpu_type
== 0xf)
486 util_cpu_caps
.x86_cpu_type
+= ((regs2
[0] >> 20) & 0xff);
488 /* general feature flags */
489 util_cpu_caps
.has_tsc
= (regs2
[3] >> 4) & 1; /* 0x0000010 */
490 util_cpu_caps
.has_mmx
= (regs2
[3] >> 23) & 1; /* 0x0800000 */
491 util_cpu_caps
.has_sse
= (regs2
[3] >> 25) & 1; /* 0x2000000 */
492 util_cpu_caps
.has_sse2
= (regs2
[3] >> 26) & 1; /* 0x4000000 */
493 util_cpu_caps
.has_sse3
= (regs2
[2] >> 0) & 1; /* 0x0000001 */
494 util_cpu_caps
.has_ssse3
= (regs2
[2] >> 9) & 1; /* 0x0000020 */
495 util_cpu_caps
.has_sse4_1
= (regs2
[2] >> 19) & 1;
496 util_cpu_caps
.has_sse4_2
= (regs2
[2] >> 20) & 1;
497 util_cpu_caps
.has_popcnt
= (regs2
[2] >> 23) & 1;
498 util_cpu_caps
.has_avx
= ((regs2
[2] >> 28) & 1) && // AVX
499 ((regs2
[2] >> 27) & 1) && // OSXSAVE
500 ((xgetbv() & 6) == 6); // XMM & YMM
501 util_cpu_caps
.has_f16c
= ((regs2
[2] >> 29) & 1) && util_cpu_caps
.has_avx
;
502 util_cpu_caps
.has_fma
= ((regs2
[2] >> 12) & 1) && util_cpu_caps
.has_avx
;
503 util_cpu_caps
.has_mmx2
= util_cpu_caps
.has_sse
; /* SSE cpus supports mmxext too */
504 #if defined(PIPE_ARCH_X86_64)
505 util_cpu_caps
.has_daz
= 1;
507 util_cpu_caps
.has_daz
= util_cpu_caps
.has_sse3
||
508 (util_cpu_caps
.has_sse2
&& sse2_has_daz());
511 cacheline
= ((regs2
[1] >> 8) & 0xFF) * 8;
513 util_cpu_caps
.cacheline
= cacheline
;
515 if (util_cpu_caps
.has_avx
&& regs
[0] >= 0x00000007) {
517 cpuid_count(0x00000007, 0x00000000, regs7
);
518 util_cpu_caps
.has_avx2
= (regs7
[1] >> 5) & 1;
522 if (((regs2
[2] >> 27) & 1) && // OSXSAVE
523 (xgetbv() & (0x7 << 5)) && // OPMASK: upper-256 enabled by OS
524 ((xgetbv() & 6) == 6)) { // XMM/YMM enabled by OS
526 cpuid_count(0x00000007, 0x00000000, regs3
);
527 util_cpu_caps
.has_avx512f
= (regs3
[1] >> 16) & 1;
528 util_cpu_caps
.has_avx512dq
= (regs3
[1] >> 17) & 1;
529 util_cpu_caps
.has_avx512ifma
= (regs3
[1] >> 21) & 1;
530 util_cpu_caps
.has_avx512pf
= (regs3
[1] >> 26) & 1;
531 util_cpu_caps
.has_avx512er
= (regs3
[1] >> 27) & 1;
532 util_cpu_caps
.has_avx512cd
= (regs3
[1] >> 28) & 1;
533 util_cpu_caps
.has_avx512bw
= (regs3
[1] >> 30) & 1;
534 util_cpu_caps
.has_avx512vl
= (regs3
[1] >> 31) & 1;
535 util_cpu_caps
.has_avx512vbmi
= (regs3
[2] >> 1) & 1;
538 if (regs
[1] == 0x756e6547 && regs
[2] == 0x6c65746e && regs
[3] == 0x49656e69) {
540 util_cpu_caps
.has_intel
= 1;
543 cpuid(0x80000000, regs
);
545 if (regs
[0] >= 0x80000001) {
547 cpuid(0x80000001, regs2
);
549 util_cpu_caps
.has_mmx
|= (regs2
[3] >> 23) & 1;
550 util_cpu_caps
.has_mmx2
|= (regs2
[3] >> 22) & 1;
551 util_cpu_caps
.has_3dnow
= (regs2
[3] >> 31) & 1;
552 util_cpu_caps
.has_3dnow_ext
= (regs2
[3] >> 30) & 1;
554 util_cpu_caps
.has_xop
= util_cpu_caps
.has_avx
&&
555 ((regs2
[2] >> 11) & 1);
558 if (regs
[0] >= 0x80000006) {
559 /* should we really do this if the clflush size above worked? */
560 unsigned int cacheline
;
561 cpuid(0x80000006, regs2
);
562 cacheline
= regs2
[2] & 0xFF;
564 util_cpu_caps
.cacheline
= cacheline
;
567 if (!util_cpu_caps
.has_sse
) {
568 util_cpu_caps
.has_sse2
= 0;
569 util_cpu_caps
.has_sse3
= 0;
570 util_cpu_caps
.has_ssse3
= 0;
571 util_cpu_caps
.has_sse4_1
= 0;
574 #endif /* PIPE_ARCH_X86 || PIPE_ARCH_X86_64 */
576 #if defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
577 check_os_arm_support();
580 #if defined(PIPE_ARCH_PPC)
581 check_os_altivec_support();
582 #endif /* PIPE_ARCH_PPC */
587 if (debug_get_option_dump_cpu()) {
588 debug_printf("util_cpu_caps.nr_cpus = %u\n", util_cpu_caps
.nr_cpus
);
590 debug_printf("util_cpu_caps.x86_cpu_type = %u\n", util_cpu_caps
.x86_cpu_type
);
591 debug_printf("util_cpu_caps.cacheline = %u\n", util_cpu_caps
.cacheline
);
593 debug_printf("util_cpu_caps.has_tsc = %u\n", util_cpu_caps
.has_tsc
);
594 debug_printf("util_cpu_caps.has_mmx = %u\n", util_cpu_caps
.has_mmx
);
595 debug_printf("util_cpu_caps.has_mmx2 = %u\n", util_cpu_caps
.has_mmx2
);
596 debug_printf("util_cpu_caps.has_sse = %u\n", util_cpu_caps
.has_sse
);
597 debug_printf("util_cpu_caps.has_sse2 = %u\n", util_cpu_caps
.has_sse2
);
598 debug_printf("util_cpu_caps.has_sse3 = %u\n", util_cpu_caps
.has_sse3
);
599 debug_printf("util_cpu_caps.has_ssse3 = %u\n", util_cpu_caps
.has_ssse3
);
600 debug_printf("util_cpu_caps.has_sse4_1 = %u\n", util_cpu_caps
.has_sse4_1
);
601 debug_printf("util_cpu_caps.has_sse4_2 = %u\n", util_cpu_caps
.has_sse4_2
);
602 debug_printf("util_cpu_caps.has_avx = %u\n", util_cpu_caps
.has_avx
);
603 debug_printf("util_cpu_caps.has_avx2 = %u\n", util_cpu_caps
.has_avx2
);
604 debug_printf("util_cpu_caps.has_f16c = %u\n", util_cpu_caps
.has_f16c
);
605 debug_printf("util_cpu_caps.has_popcnt = %u\n", util_cpu_caps
.has_popcnt
);
606 debug_printf("util_cpu_caps.has_3dnow = %u\n", util_cpu_caps
.has_3dnow
);
607 debug_printf("util_cpu_caps.has_3dnow_ext = %u\n", util_cpu_caps
.has_3dnow_ext
);
608 debug_printf("util_cpu_caps.has_xop = %u\n", util_cpu_caps
.has_xop
);
609 debug_printf("util_cpu_caps.has_altivec = %u\n", util_cpu_caps
.has_altivec
);
610 debug_printf("util_cpu_caps.has_vsx = %u\n", util_cpu_caps
.has_vsx
);
611 debug_printf("util_cpu_caps.has_neon = %u\n", util_cpu_caps
.has_neon
);
612 debug_printf("util_cpu_caps.has_daz = %u\n", util_cpu_caps
.has_daz
);
613 debug_printf("util_cpu_caps.has_avx512f = %u\n", util_cpu_caps
.has_avx512f
);
614 debug_printf("util_cpu_caps.has_avx512dq = %u\n", util_cpu_caps
.has_avx512dq
);
615 debug_printf("util_cpu_caps.has_avx512ifma = %u\n", util_cpu_caps
.has_avx512ifma
);
616 debug_printf("util_cpu_caps.has_avx512pf = %u\n", util_cpu_caps
.has_avx512pf
);
617 debug_printf("util_cpu_caps.has_avx512er = %u\n", util_cpu_caps
.has_avx512er
);
618 debug_printf("util_cpu_caps.has_avx512cd = %u\n", util_cpu_caps
.has_avx512cd
);
619 debug_printf("util_cpu_caps.has_avx512bw = %u\n", util_cpu_caps
.has_avx512bw
);
620 debug_printf("util_cpu_caps.has_avx512vl = %u\n", util_cpu_caps
.has_avx512vl
);
621 debug_printf("util_cpu_caps.has_avx512vbmi = %u\n", util_cpu_caps
.has_avx512vbmi
);
626 static once_flag cpu_once_flag
= ONCE_FLAG_INIT
;
629 util_cpu_detect(void)
631 call_once(&cpu_once_flag
, util_cpu_detect_once
);