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_BSD)
51 #include <sys/param.h>
52 #include <sys/sysctl.h>
53 #include <machine/cpu.h>
56 #if defined(PIPE_OS_FREEBSD)
57 #if __has_include(<sys/auxv.h>)
59 #define HAVE_ELF_AUX_INFO
63 #if defined(PIPE_OS_LINUX)
73 #if defined(HAS_ANDROID_CPUFEATURES)
74 #include <cpu-features.h>
77 #if defined(PIPE_OS_WINDOWS)
79 #if defined(PIPE_CC_MSVC)
86 DEBUG_GET_ONCE_BOOL_OPTION(dump_cpu
, "GALLIUM_DUMP_CPU", FALSE
)
90 struct util_cpu_caps util_cpu_caps
;
92 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
93 static int has_cpuid(void);
97 #if defined(PIPE_ARCH_PPC) && !defined(PIPE_OS_APPLE) && !defined(PIPE_OS_BSD) && !defined(PIPE_OS_LINUX)
98 static jmp_buf __lv_powerpc_jmpbuf
;
99 static volatile sig_atomic_t __lv_powerpc_canjump
= 0;
102 sigill_handler(int sig
)
104 if (!__lv_powerpc_canjump
) {
105 signal (sig
, SIG_DFL
);
109 __lv_powerpc_canjump
= 0;
110 longjmp(__lv_powerpc_jmpbuf
, 1);
114 #if defined(PIPE_ARCH_PPC)
116 check_os_altivec_support(void)
118 #if defined(__ALTIVEC__)
119 util_cpu_caps
.has_altivec
= 1;
122 util_cpu_caps
.has_vsx
= 1;
124 #if defined(__ALTIVEC__) && defined(__VSX__)
126 #elif defined(PIPE_OS_APPLE) || defined(PIPE_OS_NETBSD) || defined(PIPE_OS_OPENBSD)
128 int sels
[2] = {CTL_HW
, HW_VECTORUNIT
};
130 int sels
[2] = {CTL_MACHDEP
, CPU_ALTIVEC
};
133 int len
= sizeof (has_vu
);
136 err
= sysctl(sels
, 2, &has_vu
, &len
, NULL
, 0);
140 util_cpu_caps
.has_altivec
= 1;
143 #elif defined(PIPE_OS_FREEBSD) /* !PIPE_OS_APPLE && !PIPE_OS_NETBSD && !PIPE_OS_OPENBSD */
144 unsigned long hwcap
= 0;
145 #ifdef HAVE_ELF_AUX_INFO
146 elf_aux_info(AT_HWCAP
, &hwcap
, sizeof(hwcap
));
148 size_t len
= sizeof(hwcap
);
149 sysctlbyname("hw.cpu_features", &hwcap
, &len
, NULL
, 0);
151 if (hwcap
& PPC_FEATURE_HAS_ALTIVEC
)
152 util_cpu_caps
.has_altivec
= 1;
153 if (hwcap
& PPC_FEATURE_HAS_VSX
)
154 util_cpu_caps
.has_vsx
= 1;
155 #elif defined(PIPE_OS_LINUX) /* !PIPE_OS_FREEBSD */
156 #if defined(PIPE_ARCH_PPC_64)
161 int fd
= open("/proc/self/auxv", O_RDONLY
| O_CLOEXEC
);
163 while (read(fd
, &aux
, sizeof(aux
)) == sizeof(aux
)) {
164 if (aux
.a_type
== AT_HWCAP
) {
165 char *env_vsx
= getenv("GALLIVM_VSX");
166 uint64_t hwcap
= aux
.a_un
.a_val
;
167 util_cpu_caps
.has_altivec
= (hwcap
>> 28) & 1;
168 if (!env_vsx
|| env_vsx
[0] != '0') {
169 util_cpu_caps
.has_vsx
= (hwcap
>> 7) & 1;
176 #else /* !PIPE_OS_APPLE && !PIPE_OS_BSD && !PIPE_OS_LINUX */
177 /* not on Apple/Darwin or Linux, do it the brute-force way */
178 /* this is borrowed from the libmpeg2 library */
179 signal(SIGILL
, sigill_handler
);
180 if (setjmp(__lv_powerpc_jmpbuf
)) {
181 signal(SIGILL
, SIG_DFL
);
183 boolean enable_altivec
= TRUE
; /* Default: enable if available, and if not overridden */
184 boolean enable_vsx
= TRUE
;
186 /* Disabling Altivec code generation is not the same as disabling VSX code generation,
187 * which can be done simply by passing -mattr=-vsx to the LLVM compiler; cf.
188 * lp_build_create_jit_compiler_for_module().
189 * If you want to disable Altivec code generation, the best place to do it is here.
191 char *env_control
= getenv("GALLIVM_ALTIVEC"); /* 1=enable (default); 0=disable */
192 if (env_control
&& env_control
[0] == '0') {
193 enable_altivec
= FALSE
;
196 /* VSX instructions can be explicitly enabled/disabled via GALLIVM_VSX=1 or 0 */
197 char *env_vsx
= getenv("GALLIVM_VSX");
198 if (env_vsx
&& env_vsx
[0] == '0') {
201 if (enable_altivec
) {
202 __lv_powerpc_canjump
= 1;
206 "vand %%v0, %%v0, %%v0"
210 util_cpu_caps
.has_altivec
= 1;
213 __asm
__volatile("xxland %vs0, %vs0, %vs0");
214 util_cpu_caps
.has_vsx
= 1;
216 signal(SIGILL
, SIG_DFL
);
218 util_cpu_caps
.has_altivec
= 0;
221 #endif /* !PIPE_OS_APPLE && !PIPE_OS_LINUX */
223 #endif /* PIPE_ARCH_PPC */
226 #if defined(PIPE_ARCH_X86) || defined (PIPE_ARCH_X86_64)
227 static int has_cpuid(void)
229 #if defined(PIPE_ARCH_X86)
230 #if defined(PIPE_OS_GCC)
237 "xorl $0x200000, %0\n"
251 #elif defined(PIPE_ARCH_X86_64)
260 * @sa cpuid.h included in gcc-4.3 onwards.
261 * @sa http://msdn.microsoft.com/en-us/library/hskdteyh.aspx
264 cpuid(uint32_t ax
, uint32_t *p
)
266 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
268 "xchgl %%ebx, %1\n\t"
277 #elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
286 #elif defined(PIPE_CC_MSVC)
297 * @sa cpuid.h included in gcc-4.4 onwards.
298 * @sa http://msdn.microsoft.com/en-us/library/hskdteyh%28v=vs.90%29.aspx
301 cpuid_count(uint32_t ax
, uint32_t cx
, uint32_t *p
)
303 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
305 "xchgl %%ebx, %1\n\t"
314 #elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
323 #elif defined(PIPE_CC_MSVC)
324 __cpuidex(p
, ax
, cx
);
334 static inline uint64_t xgetbv(void)
336 #if defined(PIPE_CC_GCC)
340 ".byte 0x0f, 0x01, 0xd0" // xgetbv isn't supported on gcc < 4.4
346 return ((uint64_t)edx
<< 32) | eax
;
347 #elif defined(PIPE_CC_MSVC) && defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
348 return _xgetbv(_XCR_XFEATURE_ENABLED_MASK
);
355 #if defined(PIPE_ARCH_X86)
356 PIPE_ALIGN_STACK
static inline boolean
sse2_has_daz(void)
361 uint32_t pad2
[128-8];
362 } PIPE_ALIGN_VAR(16) fxarea
;
364 fxarea
.mxcsr_mask
= 0;
365 #if defined(PIPE_CC_GCC)
366 __asm
__volatile ("fxsave %0" : "+m" (fxarea
));
367 #elif defined(PIPE_CC_MSVC) || defined(PIPE_CC_ICL)
370 fxarea
.mxcsr_mask
= 0;
372 return !!(fxarea
.mxcsr_mask
& (1 << 6));
376 #endif /* X86 or X86_64 */
378 #if defined(PIPE_ARCH_ARM)
380 check_os_arm_support(void)
383 * On Android, the cpufeatures library is preferred way of checking
384 * CPU capabilities. However, it is not available for standalone Mesa
385 * builds, i.e. when Android build system (Android.mk-based) is not
386 * used. Because of this we cannot use PIPE_OS_ANDROID here, but rather
387 * have a separate macro that only gets enabled from respective Android.mk.
389 #if defined(__ARM_NEON) || defined(__ARM_NEON__)
390 util_cpu_caps
.has_neon
= 1;
391 #elif defined(PIPE_OS_FREEBSD) && defined(HAVE_ELF_AUX_INFO)
392 unsigned long hwcap
= 0;
393 elf_aux_info(AT_HWCAP
, &hwcap
, sizeof(hwcap
));
394 if (hwcap
& HWCAP_NEON
)
395 util_cpu_caps
.has_neon
= 1;
396 #elif defined(HAS_ANDROID_CPUFEATURES)
397 AndroidCpuFamily cpu_family
= android_getCpuFamily();
398 uint64_t cpu_features
= android_getCpuFeatures();
400 if (cpu_family
== ANDROID_CPU_FAMILY_ARM
) {
401 if (cpu_features
& ANDROID_CPU_ARM_FEATURE_NEON
)
402 util_cpu_caps
.has_neon
= 1;
404 #elif defined(PIPE_OS_LINUX)
408 fd
= open("/proc/self/auxv", O_RDONLY
| O_CLOEXEC
);
410 while (read(fd
, &aux
, sizeof(Elf32_auxv_t
)) == sizeof(Elf32_auxv_t
)) {
411 if (aux
.a_type
== AT_HWCAP
) {
412 uint32_t hwcap
= aux
.a_un
.a_val
;
414 util_cpu_caps
.has_neon
= (hwcap
>> 12) & 1;
420 #endif /* PIPE_OS_LINUX */
423 #elif defined(PIPE_ARCH_AARCH64)
425 check_os_arm_support(void)
427 util_cpu_caps
.has_neon
= true;
429 #endif /* PIPE_ARCH_ARM || PIPE_ARCH_AARCH64 */
432 get_cpu_topology(void)
434 /* Default. This is correct if L3 is not present or there is only one. */
435 util_cpu_caps
.cores_per_L3
= util_cpu_caps
.nr_cpus
;
437 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
439 if (util_cpu_caps
.x86_cpu_type
== 0x17) {
442 /* Query the L3 cache topology information. */
443 cpuid_count(0x8000001D, 3, regs
);
444 unsigned cache_level
= (regs
[0] >> 5) & 0x7;
445 unsigned cores_per_cache
= ((regs
[0] >> 14) & 0xfff) + 1;
447 if (cache_level
== 3)
448 util_cpu_caps
.cores_per_L3
= cores_per_cache
;
454 util_cpu_detect_once(void)
456 memset(&util_cpu_caps
, 0, sizeof util_cpu_caps
);
458 /* Count the number of CPUs in system */
459 #if defined(PIPE_OS_WINDOWS)
461 SYSTEM_INFO system_info
;
462 GetSystemInfo(&system_info
);
463 util_cpu_caps
.nr_cpus
= system_info
.dwNumberOfProcessors
;
465 #elif defined(PIPE_OS_UNIX) && defined(_SC_NPROCESSORS_ONLN)
466 util_cpu_caps
.nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
467 if (util_cpu_caps
.nr_cpus
== ~0)
468 util_cpu_caps
.nr_cpus
= 1;
469 #elif defined(PIPE_OS_BSD)
478 sysctl(mib
, 2, &ncpu
, &len
, NULL
, 0);
479 util_cpu_caps
.nr_cpus
= ncpu
;
482 util_cpu_caps
.nr_cpus
= 1;
485 /* Make the fallback cacheline size nonzero so that it can be
486 * safely passed to align().
488 util_cpu_caps
.cacheline
= sizeof(void *);
490 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
495 util_cpu_caps
.cacheline
= 32;
497 /* Get max cpuid level */
498 cpuid(0x00000000, regs
);
500 if (regs
[0] >= 0x00000001) {
501 unsigned int cacheline
;
503 cpuid (0x00000001, regs2
);
505 util_cpu_caps
.x86_cpu_type
= (regs2
[0] >> 8) & 0xf;
506 /* Add "extended family". */
507 if (util_cpu_caps
.x86_cpu_type
== 0xf)
508 util_cpu_caps
.x86_cpu_type
+= ((regs2
[0] >> 20) & 0xff);
510 /* general feature flags */
511 util_cpu_caps
.has_tsc
= (regs2
[3] >> 4) & 1; /* 0x0000010 */
512 util_cpu_caps
.has_mmx
= (regs2
[3] >> 23) & 1; /* 0x0800000 */
513 util_cpu_caps
.has_sse
= (regs2
[3] >> 25) & 1; /* 0x2000000 */
514 util_cpu_caps
.has_sse2
= (regs2
[3] >> 26) & 1; /* 0x4000000 */
515 util_cpu_caps
.has_sse3
= (regs2
[2] >> 0) & 1; /* 0x0000001 */
516 util_cpu_caps
.has_ssse3
= (regs2
[2] >> 9) & 1; /* 0x0000020 */
517 util_cpu_caps
.has_sse4_1
= (regs2
[2] >> 19) & 1;
518 util_cpu_caps
.has_sse4_2
= (regs2
[2] >> 20) & 1;
519 util_cpu_caps
.has_popcnt
= (regs2
[2] >> 23) & 1;
520 util_cpu_caps
.has_avx
= ((regs2
[2] >> 28) & 1) && // AVX
521 ((regs2
[2] >> 27) & 1) && // OSXSAVE
522 ((xgetbv() & 6) == 6); // XMM & YMM
523 util_cpu_caps
.has_f16c
= ((regs2
[2] >> 29) & 1) && util_cpu_caps
.has_avx
;
524 util_cpu_caps
.has_fma
= ((regs2
[2] >> 12) & 1) && util_cpu_caps
.has_avx
;
525 util_cpu_caps
.has_mmx2
= util_cpu_caps
.has_sse
; /* SSE cpus supports mmxext too */
526 #if defined(PIPE_ARCH_X86_64)
527 util_cpu_caps
.has_daz
= 1;
529 util_cpu_caps
.has_daz
= util_cpu_caps
.has_sse3
||
530 (util_cpu_caps
.has_sse2
&& sse2_has_daz());
533 cacheline
= ((regs2
[1] >> 8) & 0xFF) * 8;
535 util_cpu_caps
.cacheline
= cacheline
;
537 if (util_cpu_caps
.has_avx
&& regs
[0] >= 0x00000007) {
539 cpuid_count(0x00000007, 0x00000000, regs7
);
540 util_cpu_caps
.has_avx2
= (regs7
[1] >> 5) & 1;
544 if (((regs2
[2] >> 27) & 1) && // OSXSAVE
545 (xgetbv() & (0x7 << 5)) && // OPMASK: upper-256 enabled by OS
546 ((xgetbv() & 6) == 6)) { // XMM/YMM enabled by OS
548 cpuid_count(0x00000007, 0x00000000, regs3
);
549 util_cpu_caps
.has_avx512f
= (regs3
[1] >> 16) & 1;
550 util_cpu_caps
.has_avx512dq
= (regs3
[1] >> 17) & 1;
551 util_cpu_caps
.has_avx512ifma
= (regs3
[1] >> 21) & 1;
552 util_cpu_caps
.has_avx512pf
= (regs3
[1] >> 26) & 1;
553 util_cpu_caps
.has_avx512er
= (regs3
[1] >> 27) & 1;
554 util_cpu_caps
.has_avx512cd
= (regs3
[1] >> 28) & 1;
555 util_cpu_caps
.has_avx512bw
= (regs3
[1] >> 30) & 1;
556 util_cpu_caps
.has_avx512vl
= (regs3
[1] >> 31) & 1;
557 util_cpu_caps
.has_avx512vbmi
= (regs3
[2] >> 1) & 1;
560 if (regs
[1] == 0x756e6547 && regs
[2] == 0x6c65746e && regs
[3] == 0x49656e69) {
562 util_cpu_caps
.has_intel
= 1;
565 cpuid(0x80000000, regs
);
567 if (regs
[0] >= 0x80000001) {
569 cpuid(0x80000001, regs2
);
571 util_cpu_caps
.has_mmx
|= (regs2
[3] >> 23) & 1;
572 util_cpu_caps
.has_mmx2
|= (regs2
[3] >> 22) & 1;
573 util_cpu_caps
.has_3dnow
= (regs2
[3] >> 31) & 1;
574 util_cpu_caps
.has_3dnow_ext
= (regs2
[3] >> 30) & 1;
576 util_cpu_caps
.has_xop
= util_cpu_caps
.has_avx
&&
577 ((regs2
[2] >> 11) & 1);
580 if (regs
[0] >= 0x80000006) {
581 /* should we really do this if the clflush size above worked? */
582 unsigned int cacheline
;
583 cpuid(0x80000006, regs2
);
584 cacheline
= regs2
[2] & 0xFF;
586 util_cpu_caps
.cacheline
= cacheline
;
589 if (!util_cpu_caps
.has_sse
) {
590 util_cpu_caps
.has_sse2
= 0;
591 util_cpu_caps
.has_sse3
= 0;
592 util_cpu_caps
.has_ssse3
= 0;
593 util_cpu_caps
.has_sse4_1
= 0;
596 #endif /* PIPE_ARCH_X86 || PIPE_ARCH_X86_64 */
598 #if defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
599 check_os_arm_support();
602 #if defined(PIPE_ARCH_PPC)
603 check_os_altivec_support();
604 #endif /* PIPE_ARCH_PPC */
609 if (debug_get_option_dump_cpu()) {
610 debug_printf("util_cpu_caps.nr_cpus = %u\n", util_cpu_caps
.nr_cpus
);
612 debug_printf("util_cpu_caps.x86_cpu_type = %u\n", util_cpu_caps
.x86_cpu_type
);
613 debug_printf("util_cpu_caps.cacheline = %u\n", util_cpu_caps
.cacheline
);
615 debug_printf("util_cpu_caps.has_tsc = %u\n", util_cpu_caps
.has_tsc
);
616 debug_printf("util_cpu_caps.has_mmx = %u\n", util_cpu_caps
.has_mmx
);
617 debug_printf("util_cpu_caps.has_mmx2 = %u\n", util_cpu_caps
.has_mmx2
);
618 debug_printf("util_cpu_caps.has_sse = %u\n", util_cpu_caps
.has_sse
);
619 debug_printf("util_cpu_caps.has_sse2 = %u\n", util_cpu_caps
.has_sse2
);
620 debug_printf("util_cpu_caps.has_sse3 = %u\n", util_cpu_caps
.has_sse3
);
621 debug_printf("util_cpu_caps.has_ssse3 = %u\n", util_cpu_caps
.has_ssse3
);
622 debug_printf("util_cpu_caps.has_sse4_1 = %u\n", util_cpu_caps
.has_sse4_1
);
623 debug_printf("util_cpu_caps.has_sse4_2 = %u\n", util_cpu_caps
.has_sse4_2
);
624 debug_printf("util_cpu_caps.has_avx = %u\n", util_cpu_caps
.has_avx
);
625 debug_printf("util_cpu_caps.has_avx2 = %u\n", util_cpu_caps
.has_avx2
);
626 debug_printf("util_cpu_caps.has_f16c = %u\n", util_cpu_caps
.has_f16c
);
627 debug_printf("util_cpu_caps.has_popcnt = %u\n", util_cpu_caps
.has_popcnt
);
628 debug_printf("util_cpu_caps.has_3dnow = %u\n", util_cpu_caps
.has_3dnow
);
629 debug_printf("util_cpu_caps.has_3dnow_ext = %u\n", util_cpu_caps
.has_3dnow_ext
);
630 debug_printf("util_cpu_caps.has_xop = %u\n", util_cpu_caps
.has_xop
);
631 debug_printf("util_cpu_caps.has_altivec = %u\n", util_cpu_caps
.has_altivec
);
632 debug_printf("util_cpu_caps.has_vsx = %u\n", util_cpu_caps
.has_vsx
);
633 debug_printf("util_cpu_caps.has_neon = %u\n", util_cpu_caps
.has_neon
);
634 debug_printf("util_cpu_caps.has_daz = %u\n", util_cpu_caps
.has_daz
);
635 debug_printf("util_cpu_caps.has_avx512f = %u\n", util_cpu_caps
.has_avx512f
);
636 debug_printf("util_cpu_caps.has_avx512dq = %u\n", util_cpu_caps
.has_avx512dq
);
637 debug_printf("util_cpu_caps.has_avx512ifma = %u\n", util_cpu_caps
.has_avx512ifma
);
638 debug_printf("util_cpu_caps.has_avx512pf = %u\n", util_cpu_caps
.has_avx512pf
);
639 debug_printf("util_cpu_caps.has_avx512er = %u\n", util_cpu_caps
.has_avx512er
);
640 debug_printf("util_cpu_caps.has_avx512cd = %u\n", util_cpu_caps
.has_avx512cd
);
641 debug_printf("util_cpu_caps.has_avx512bw = %u\n", util_cpu_caps
.has_avx512bw
);
642 debug_printf("util_cpu_caps.has_avx512vl = %u\n", util_cpu_caps
.has_avx512vl
);
643 debug_printf("util_cpu_caps.has_avx512vbmi = %u\n", util_cpu_caps
.has_avx512vbmi
);
648 static once_flag cpu_once_flag
= ONCE_FLAG_INIT
;
651 util_cpu_detect(void)
653 call_once(&cpu_once_flag
, util_cpu_detect_once
);