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
63 #if defined(PIPE_ARCH_PPC)
64 #include <machine/cpu.h>
68 #if defined(PIPE_OS_LINUX)
78 #if defined(HAS_ANDROID_CPUFEATURES)
79 #include <cpu-features.h>
82 #if defined(PIPE_OS_WINDOWS)
84 #if defined(PIPE_CC_MSVC)
91 DEBUG_GET_ONCE_BOOL_OPTION(dump_cpu
, "GALLIUM_DUMP_CPU", FALSE
)
95 struct util_cpu_caps util_cpu_caps
;
97 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
98 static int has_cpuid(void);
102 #if defined(PIPE_ARCH_PPC) && !defined(PIPE_OS_APPLE) && !defined(PIPE_OS_BSD) && !defined(PIPE_OS_LINUX)
103 static jmp_buf __lv_powerpc_jmpbuf
;
104 static volatile sig_atomic_t __lv_powerpc_canjump
= 0;
107 sigill_handler(int sig
)
109 if (!__lv_powerpc_canjump
) {
110 signal (sig
, SIG_DFL
);
114 __lv_powerpc_canjump
= 0;
115 longjmp(__lv_powerpc_jmpbuf
, 1);
119 #if defined(PIPE_ARCH_PPC)
121 check_os_altivec_support(void)
123 #if defined(__ALTIVEC__)
124 util_cpu_caps
.has_altivec
= 1;
127 util_cpu_caps
.has_vsx
= 1;
129 #if defined(__ALTIVEC__) && defined(__VSX__)
131 #elif defined(PIPE_OS_APPLE) || defined(PIPE_OS_NETBSD) || defined(PIPE_OS_OPENBSD)
133 int sels
[2] = {CTL_HW
, HW_VECTORUNIT
};
135 int sels
[2] = {CTL_MACHDEP
, CPU_ALTIVEC
};
138 int len
= sizeof (has_vu
);
141 err
= sysctl(sels
, 2, &has_vu
, &len
, NULL
, 0);
145 util_cpu_caps
.has_altivec
= 1;
148 #elif defined(PIPE_OS_FREEBSD) /* !PIPE_OS_APPLE && !PIPE_OS_NETBSD && !PIPE_OS_OPENBSD */
149 unsigned long hwcap
= 0;
150 #ifdef HAVE_ELF_AUX_INFO
151 elf_aux_info(AT_HWCAP
, &hwcap
, sizeof(hwcap
));
153 size_t len
= sizeof(hwcap
);
154 sysctlbyname("hw.cpu_features", &hwcap
, &len
, NULL
, 0);
156 if (hwcap
& PPC_FEATURE_HAS_ALTIVEC
)
157 util_cpu_caps
.has_altivec
= 1;
158 if (hwcap
& PPC_FEATURE_HAS_VSX
)
159 util_cpu_caps
.has_vsx
= 1;
160 #elif defined(PIPE_OS_LINUX) /* !PIPE_OS_FREEBSD */
161 #if defined(PIPE_ARCH_PPC_64)
166 int fd
= open("/proc/self/auxv", O_RDONLY
| O_CLOEXEC
);
168 while (read(fd
, &aux
, sizeof(aux
)) == sizeof(aux
)) {
169 if (aux
.a_type
== AT_HWCAP
) {
170 char *env_vsx
= getenv("GALLIVM_VSX");
171 uint64_t hwcap
= aux
.a_un
.a_val
;
172 util_cpu_caps
.has_altivec
= (hwcap
>> 28) & 1;
173 if (!env_vsx
|| env_vsx
[0] != '0') {
174 util_cpu_caps
.has_vsx
= (hwcap
>> 7) & 1;
181 #else /* !PIPE_OS_APPLE && !PIPE_OS_BSD && !PIPE_OS_LINUX */
182 /* not on Apple/Darwin or Linux, do it the brute-force way */
183 /* this is borrowed from the libmpeg2 library */
184 signal(SIGILL
, sigill_handler
);
185 if (setjmp(__lv_powerpc_jmpbuf
)) {
186 signal(SIGILL
, SIG_DFL
);
188 boolean enable_altivec
= TRUE
; /* Default: enable if available, and if not overridden */
189 boolean enable_vsx
= TRUE
;
191 /* Disabling Altivec code generation is not the same as disabling VSX code generation,
192 * which can be done simply by passing -mattr=-vsx to the LLVM compiler; cf.
193 * lp_build_create_jit_compiler_for_module().
194 * If you want to disable Altivec code generation, the best place to do it is here.
196 char *env_control
= getenv("GALLIVM_ALTIVEC"); /* 1=enable (default); 0=disable */
197 if (env_control
&& env_control
[0] == '0') {
198 enable_altivec
= FALSE
;
201 /* VSX instructions can be explicitly enabled/disabled via GALLIVM_VSX=1 or 0 */
202 char *env_vsx
= getenv("GALLIVM_VSX");
203 if (env_vsx
&& env_vsx
[0] == '0') {
206 if (enable_altivec
) {
207 __lv_powerpc_canjump
= 1;
211 "vand %%v0, %%v0, %%v0"
215 util_cpu_caps
.has_altivec
= 1;
218 __asm
__volatile("xxland %vs0, %vs0, %vs0");
219 util_cpu_caps
.has_vsx
= 1;
221 signal(SIGILL
, SIG_DFL
);
223 util_cpu_caps
.has_altivec
= 0;
226 #endif /* !PIPE_OS_APPLE && !PIPE_OS_LINUX */
228 #endif /* PIPE_ARCH_PPC */
231 #if defined(PIPE_ARCH_X86) || defined (PIPE_ARCH_X86_64)
232 static int has_cpuid(void)
234 #if defined(PIPE_ARCH_X86)
235 #if defined(PIPE_OS_GCC)
242 "xorl $0x200000, %0\n"
256 #elif defined(PIPE_ARCH_X86_64)
265 * @sa cpuid.h included in gcc-4.3 onwards.
266 * @sa http://msdn.microsoft.com/en-us/library/hskdteyh.aspx
269 cpuid(uint32_t ax
, uint32_t *p
)
271 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
273 "xchgl %%ebx, %1\n\t"
282 #elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
291 #elif defined(PIPE_CC_MSVC)
302 * @sa cpuid.h included in gcc-4.4 onwards.
303 * @sa http://msdn.microsoft.com/en-us/library/hskdteyh%28v=vs.90%29.aspx
306 cpuid_count(uint32_t ax
, uint32_t cx
, uint32_t *p
)
308 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
310 "xchgl %%ebx, %1\n\t"
319 #elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
328 #elif defined(PIPE_CC_MSVC)
329 __cpuidex(p
, ax
, cx
);
339 static inline uint64_t xgetbv(void)
341 #if defined(PIPE_CC_GCC)
345 ".byte 0x0f, 0x01, 0xd0" // xgetbv isn't supported on gcc < 4.4
351 return ((uint64_t)edx
<< 32) | eax
;
352 #elif defined(PIPE_CC_MSVC) && defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
353 return _xgetbv(_XCR_XFEATURE_ENABLED_MASK
);
360 #if defined(PIPE_ARCH_X86)
361 PIPE_ALIGN_STACK
static inline boolean
sse2_has_daz(void)
366 uint32_t pad2
[128-8];
367 } PIPE_ALIGN_VAR(16) fxarea
;
369 fxarea
.mxcsr_mask
= 0;
370 #if defined(PIPE_CC_GCC)
371 __asm
__volatile ("fxsave %0" : "+m" (fxarea
));
372 #elif defined(PIPE_CC_MSVC) || defined(PIPE_CC_ICL)
375 fxarea
.mxcsr_mask
= 0;
377 return !!(fxarea
.mxcsr_mask
& (1 << 6));
381 #endif /* X86 or X86_64 */
383 #if defined(PIPE_ARCH_ARM)
385 check_os_arm_support(void)
388 * On Android, the cpufeatures library is preferred way of checking
389 * CPU capabilities. However, it is not available for standalone Mesa
390 * builds, i.e. when Android build system (Android.mk-based) is not
391 * used. Because of this we cannot use PIPE_OS_ANDROID here, but rather
392 * have a separate macro that only gets enabled from respective Android.mk.
394 #if defined(__ARM_NEON) || defined(__ARM_NEON__)
395 util_cpu_caps
.has_neon
= 1;
396 #elif defined(PIPE_OS_FREEBSD) && defined(HAVE_ELF_AUX_INFO)
397 unsigned long hwcap
= 0;
398 elf_aux_info(AT_HWCAP
, &hwcap
, sizeof(hwcap
));
399 if (hwcap
& HWCAP_NEON
)
400 util_cpu_caps
.has_neon
= 1;
401 #elif defined(HAS_ANDROID_CPUFEATURES)
402 AndroidCpuFamily cpu_family
= android_getCpuFamily();
403 uint64_t cpu_features
= android_getCpuFeatures();
405 if (cpu_family
== ANDROID_CPU_FAMILY_ARM
) {
406 if (cpu_features
& ANDROID_CPU_ARM_FEATURE_NEON
)
407 util_cpu_caps
.has_neon
= 1;
409 #elif defined(PIPE_OS_LINUX)
413 fd
= open("/proc/self/auxv", O_RDONLY
| O_CLOEXEC
);
415 while (read(fd
, &aux
, sizeof(Elf32_auxv_t
)) == sizeof(Elf32_auxv_t
)) {
416 if (aux
.a_type
== AT_HWCAP
) {
417 uint32_t hwcap
= aux
.a_un
.a_val
;
419 util_cpu_caps
.has_neon
= (hwcap
>> 12) & 1;
425 #endif /* PIPE_OS_LINUX */
428 #elif defined(PIPE_ARCH_AARCH64)
430 check_os_arm_support(void)
432 util_cpu_caps
.has_neon
= true;
434 #endif /* PIPE_ARCH_ARM || PIPE_ARCH_AARCH64 */
437 get_cpu_topology(void)
439 /* Default. This is correct if L3 is not present or there is only one. */
440 util_cpu_caps
.cores_per_L3
= util_cpu_caps
.nr_cpus
;
442 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
444 if (util_cpu_caps
.x86_cpu_type
== 0x17) {
447 /* Query the L3 cache topology information. */
448 cpuid_count(0x8000001D, 3, regs
);
449 unsigned cache_level
= (regs
[0] >> 5) & 0x7;
450 unsigned cores_per_cache
= ((regs
[0] >> 14) & 0xfff) + 1;
452 if (cache_level
== 3)
453 util_cpu_caps
.cores_per_L3
= cores_per_cache
;
459 util_cpu_detect_once(void)
461 memset(&util_cpu_caps
, 0, sizeof util_cpu_caps
);
463 /* Count the number of CPUs in system */
464 #if defined(PIPE_OS_WINDOWS)
466 SYSTEM_INFO system_info
;
467 GetSystemInfo(&system_info
);
468 util_cpu_caps
.nr_cpus
= system_info
.dwNumberOfProcessors
;
470 #elif defined(PIPE_OS_UNIX) && defined(_SC_NPROCESSORS_ONLN)
471 util_cpu_caps
.nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
472 if (util_cpu_caps
.nr_cpus
== ~0)
473 util_cpu_caps
.nr_cpus
= 1;
474 #elif defined(PIPE_OS_BSD)
483 sysctl(mib
, 2, &ncpu
, &len
, NULL
, 0);
484 util_cpu_caps
.nr_cpus
= ncpu
;
487 util_cpu_caps
.nr_cpus
= 1;
490 /* Make the fallback cacheline size nonzero so that it can be
491 * safely passed to align().
493 util_cpu_caps
.cacheline
= sizeof(void *);
495 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
500 util_cpu_caps
.cacheline
= 32;
502 /* Get max cpuid level */
503 cpuid(0x00000000, regs
);
505 if (regs
[0] >= 0x00000001) {
506 unsigned int cacheline
;
508 cpuid (0x00000001, regs2
);
510 util_cpu_caps
.x86_cpu_type
= (regs2
[0] >> 8) & 0xf;
511 /* Add "extended family". */
512 if (util_cpu_caps
.x86_cpu_type
== 0xf)
513 util_cpu_caps
.x86_cpu_type
+= ((regs2
[0] >> 20) & 0xff);
515 /* general feature flags */
516 util_cpu_caps
.has_tsc
= (regs2
[3] >> 4) & 1; /* 0x0000010 */
517 util_cpu_caps
.has_mmx
= (regs2
[3] >> 23) & 1; /* 0x0800000 */
518 util_cpu_caps
.has_sse
= (regs2
[3] >> 25) & 1; /* 0x2000000 */
519 util_cpu_caps
.has_sse2
= (regs2
[3] >> 26) & 1; /* 0x4000000 */
520 util_cpu_caps
.has_sse3
= (regs2
[2] >> 0) & 1; /* 0x0000001 */
521 util_cpu_caps
.has_ssse3
= (regs2
[2] >> 9) & 1; /* 0x0000020 */
522 util_cpu_caps
.has_sse4_1
= (regs2
[2] >> 19) & 1;
523 util_cpu_caps
.has_sse4_2
= (regs2
[2] >> 20) & 1;
524 util_cpu_caps
.has_popcnt
= (regs2
[2] >> 23) & 1;
525 util_cpu_caps
.has_avx
= ((regs2
[2] >> 28) & 1) && // AVX
526 ((regs2
[2] >> 27) & 1) && // OSXSAVE
527 ((xgetbv() & 6) == 6); // XMM & YMM
528 util_cpu_caps
.has_f16c
= ((regs2
[2] >> 29) & 1) && util_cpu_caps
.has_avx
;
529 util_cpu_caps
.has_fma
= ((regs2
[2] >> 12) & 1) && util_cpu_caps
.has_avx
;
530 util_cpu_caps
.has_mmx2
= util_cpu_caps
.has_sse
; /* SSE cpus supports mmxext too */
531 #if defined(PIPE_ARCH_X86_64)
532 util_cpu_caps
.has_daz
= 1;
534 util_cpu_caps
.has_daz
= util_cpu_caps
.has_sse3
||
535 (util_cpu_caps
.has_sse2
&& sse2_has_daz());
538 cacheline
= ((regs2
[1] >> 8) & 0xFF) * 8;
540 util_cpu_caps
.cacheline
= cacheline
;
542 if (util_cpu_caps
.has_avx
&& regs
[0] >= 0x00000007) {
544 cpuid_count(0x00000007, 0x00000000, regs7
);
545 util_cpu_caps
.has_avx2
= (regs7
[1] >> 5) & 1;
549 if (((regs2
[2] >> 27) & 1) && // OSXSAVE
550 (xgetbv() & (0x7 << 5)) && // OPMASK: upper-256 enabled by OS
551 ((xgetbv() & 6) == 6)) { // XMM/YMM enabled by OS
553 cpuid_count(0x00000007, 0x00000000, regs3
);
554 util_cpu_caps
.has_avx512f
= (regs3
[1] >> 16) & 1;
555 util_cpu_caps
.has_avx512dq
= (regs3
[1] >> 17) & 1;
556 util_cpu_caps
.has_avx512ifma
= (regs3
[1] >> 21) & 1;
557 util_cpu_caps
.has_avx512pf
= (regs3
[1] >> 26) & 1;
558 util_cpu_caps
.has_avx512er
= (regs3
[1] >> 27) & 1;
559 util_cpu_caps
.has_avx512cd
= (regs3
[1] >> 28) & 1;
560 util_cpu_caps
.has_avx512bw
= (regs3
[1] >> 30) & 1;
561 util_cpu_caps
.has_avx512vl
= (regs3
[1] >> 31) & 1;
562 util_cpu_caps
.has_avx512vbmi
= (regs3
[2] >> 1) & 1;
565 if (regs
[1] == 0x756e6547 && regs
[2] == 0x6c65746e && regs
[3] == 0x49656e69) {
567 util_cpu_caps
.has_intel
= 1;
570 cpuid(0x80000000, regs
);
572 if (regs
[0] >= 0x80000001) {
574 cpuid(0x80000001, regs2
);
576 util_cpu_caps
.has_mmx
|= (regs2
[3] >> 23) & 1;
577 util_cpu_caps
.has_mmx2
|= (regs2
[3] >> 22) & 1;
578 util_cpu_caps
.has_3dnow
= (regs2
[3] >> 31) & 1;
579 util_cpu_caps
.has_3dnow_ext
= (regs2
[3] >> 30) & 1;
581 util_cpu_caps
.has_xop
= util_cpu_caps
.has_avx
&&
582 ((regs2
[2] >> 11) & 1);
585 if (regs
[0] >= 0x80000006) {
586 /* should we really do this if the clflush size above worked? */
587 unsigned int cacheline
;
588 cpuid(0x80000006, regs2
);
589 cacheline
= regs2
[2] & 0xFF;
591 util_cpu_caps
.cacheline
= cacheline
;
594 if (!util_cpu_caps
.has_sse
) {
595 util_cpu_caps
.has_sse2
= 0;
596 util_cpu_caps
.has_sse3
= 0;
597 util_cpu_caps
.has_ssse3
= 0;
598 util_cpu_caps
.has_sse4_1
= 0;
601 #endif /* PIPE_ARCH_X86 || PIPE_ARCH_X86_64 */
603 #if defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
604 check_os_arm_support();
607 #if defined(PIPE_ARCH_PPC)
608 check_os_altivec_support();
609 #endif /* PIPE_ARCH_PPC */
614 if (debug_get_option_dump_cpu()) {
615 debug_printf("util_cpu_caps.nr_cpus = %u\n", util_cpu_caps
.nr_cpus
);
617 debug_printf("util_cpu_caps.x86_cpu_type = %u\n", util_cpu_caps
.x86_cpu_type
);
618 debug_printf("util_cpu_caps.cacheline = %u\n", util_cpu_caps
.cacheline
);
620 debug_printf("util_cpu_caps.has_tsc = %u\n", util_cpu_caps
.has_tsc
);
621 debug_printf("util_cpu_caps.has_mmx = %u\n", util_cpu_caps
.has_mmx
);
622 debug_printf("util_cpu_caps.has_mmx2 = %u\n", util_cpu_caps
.has_mmx2
);
623 debug_printf("util_cpu_caps.has_sse = %u\n", util_cpu_caps
.has_sse
);
624 debug_printf("util_cpu_caps.has_sse2 = %u\n", util_cpu_caps
.has_sse2
);
625 debug_printf("util_cpu_caps.has_sse3 = %u\n", util_cpu_caps
.has_sse3
);
626 debug_printf("util_cpu_caps.has_ssse3 = %u\n", util_cpu_caps
.has_ssse3
);
627 debug_printf("util_cpu_caps.has_sse4_1 = %u\n", util_cpu_caps
.has_sse4_1
);
628 debug_printf("util_cpu_caps.has_sse4_2 = %u\n", util_cpu_caps
.has_sse4_2
);
629 debug_printf("util_cpu_caps.has_avx = %u\n", util_cpu_caps
.has_avx
);
630 debug_printf("util_cpu_caps.has_avx2 = %u\n", util_cpu_caps
.has_avx2
);
631 debug_printf("util_cpu_caps.has_f16c = %u\n", util_cpu_caps
.has_f16c
);
632 debug_printf("util_cpu_caps.has_popcnt = %u\n", util_cpu_caps
.has_popcnt
);
633 debug_printf("util_cpu_caps.has_3dnow = %u\n", util_cpu_caps
.has_3dnow
);
634 debug_printf("util_cpu_caps.has_3dnow_ext = %u\n", util_cpu_caps
.has_3dnow_ext
);
635 debug_printf("util_cpu_caps.has_xop = %u\n", util_cpu_caps
.has_xop
);
636 debug_printf("util_cpu_caps.has_altivec = %u\n", util_cpu_caps
.has_altivec
);
637 debug_printf("util_cpu_caps.has_vsx = %u\n", util_cpu_caps
.has_vsx
);
638 debug_printf("util_cpu_caps.has_neon = %u\n", util_cpu_caps
.has_neon
);
639 debug_printf("util_cpu_caps.has_daz = %u\n", util_cpu_caps
.has_daz
);
640 debug_printf("util_cpu_caps.has_avx512f = %u\n", util_cpu_caps
.has_avx512f
);
641 debug_printf("util_cpu_caps.has_avx512dq = %u\n", util_cpu_caps
.has_avx512dq
);
642 debug_printf("util_cpu_caps.has_avx512ifma = %u\n", util_cpu_caps
.has_avx512ifma
);
643 debug_printf("util_cpu_caps.has_avx512pf = %u\n", util_cpu_caps
.has_avx512pf
);
644 debug_printf("util_cpu_caps.has_avx512er = %u\n", util_cpu_caps
.has_avx512er
);
645 debug_printf("util_cpu_caps.has_avx512cd = %u\n", util_cpu_caps
.has_avx512cd
);
646 debug_printf("util_cpu_caps.has_avx512bw = %u\n", util_cpu_caps
.has_avx512bw
);
647 debug_printf("util_cpu_caps.has_avx512vl = %u\n", util_cpu_caps
.has_avx512vl
);
648 debug_printf("util_cpu_caps.has_avx512vbmi = %u\n", util_cpu_caps
.has_avx512vbmi
);
653 static once_flag cpu_once_flag
= ONCE_FLAG_INIT
;
656 util_cpu_detect(void)
658 call_once(&cpu_once_flag
, util_cpu_detect_once
);