util: move u_cpu_detect to util
[mesa.git] / src / util / u_cpu_detect.c
diff --git a/src/util/u_cpu_detect.c b/src/util/u_cpu_detect.c
new file mode 100644 (file)
index 0000000..4dbb4d8
--- /dev/null
@@ -0,0 +1,593 @@
+/**************************************************************************
+ * 
+ * Copyright 2008 Dennis Smit
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.  IN NO EVENT SHALL
+ * AUTHORS, COPYRIGHT HOLDERS, AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ * 
+ **************************************************************************/
+
+/**
+ * @file
+ * CPU feature detection.
+ *
+ * @author Dennis Smit
+ * @author Based on the work of Eric Anholt <anholt@FreeBSD.org>
+ */
+
+#include "pipe/p_config.h"
+
+#include "util/u_debug.h"
+#include "u_cpu_detect.h"
+#include "c11/threads.h"
+
+#if defined(PIPE_ARCH_PPC)
+#if defined(PIPE_OS_APPLE)
+#include <sys/sysctl.h>
+#else
+#include <signal.h>
+#include <setjmp.h>
+#endif
+#endif
+
+#if defined(PIPE_OS_NETBSD) || defined(PIPE_OS_OPENBSD)
+#include <sys/param.h>
+#include <sys/sysctl.h>
+#include <machine/cpu.h>
+#endif
+
+#if defined(PIPE_OS_FREEBSD) || defined(PIPE_OS_DRAGONFLY)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
+
+#if defined(PIPE_OS_LINUX)
+#include <signal.h>
+#include <fcntl.h>
+#include <elf.h>
+#endif
+
+#ifdef PIPE_OS_UNIX
+#include <unistd.h>
+#endif
+
+#if defined(HAS_ANDROID_CPUFEATURES)
+#include <cpu-features.h>
+#endif
+
+#if defined(PIPE_OS_WINDOWS)
+#include <windows.h>
+#if defined(PIPE_CC_MSVC)
+#include <intrin.h>
+#endif
+#endif
+
+
+#ifdef DEBUG
+DEBUG_GET_ONCE_BOOL_OPTION(dump_cpu, "GALLIUM_DUMP_CPU", FALSE)
+#endif
+
+
+struct util_cpu_caps util_cpu_caps;
+
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+static int has_cpuid(void);
+#endif
+
+
+#if defined(PIPE_ARCH_PPC) && !defined(PIPE_OS_APPLE)
+static jmp_buf  __lv_powerpc_jmpbuf;
+static volatile sig_atomic_t __lv_powerpc_canjump = 0;
+
+static void
+sigill_handler(int sig)
+{
+   if (!__lv_powerpc_canjump) {
+      signal (sig, SIG_DFL);
+      raise (sig);
+   }
+
+   __lv_powerpc_canjump = 0;
+   longjmp(__lv_powerpc_jmpbuf, 1);
+}
+#endif
+
+#if defined(PIPE_ARCH_PPC)
+static void
+check_os_altivec_support(void)
+{
+#if defined(PIPE_OS_APPLE)
+   int sels[2] = {CTL_HW, HW_VECTORUNIT};
+   int has_vu = 0;
+   int len = sizeof (has_vu);
+   int err;
+
+   err = sysctl(sels, 2, &has_vu, &len, NULL, 0);
+
+   if (err == 0) {
+      if (has_vu != 0) {
+         util_cpu_caps.has_altivec = 1;
+      }
+   }
+#else /* !PIPE_OS_APPLE */
+   /* not on Apple/Darwin, do it the brute-force way */
+   /* this is borrowed from the libmpeg2 library */
+   signal(SIGILL, sigill_handler);
+   if (setjmp(__lv_powerpc_jmpbuf)) {
+      signal(SIGILL, SIG_DFL);
+   } else {
+      boolean enable_altivec = TRUE;    /* Default: enable  if available, and if not overridden */
+      boolean enable_vsx = TRUE;
+#ifdef DEBUG
+      /* Disabling Altivec code generation is not the same as disabling VSX code generation,
+       * which can be done simply by passing -mattr=-vsx to the LLVM compiler; cf.
+       * lp_build_create_jit_compiler_for_module().
+       * If you want to disable Altivec code generation, the best place to do it is here.
+       */
+      char *env_control = getenv("GALLIVM_ALTIVEC");    /* 1=enable (default); 0=disable */
+      if (env_control && env_control[0] == '0') {
+         enable_altivec = FALSE;
+      }
+#endif
+      /* VSX instructions can be explicitly enabled/disabled via GALLIVM_VSX=1 or 0 */
+      char *env_vsx = getenv("GALLIVM_VSX");
+      if (env_vsx && env_vsx[0] == '0') {
+         enable_vsx = FALSE;
+      }
+      if (enable_altivec) {
+         __lv_powerpc_canjump = 1;
+
+         __asm __volatile
+            ("mtspr 256, %0\n\t"
+             "vand %%v0, %%v0, %%v0"
+             :
+             : "r" (-1));
+
+         util_cpu_caps.has_altivec = 1;
+
+         if (enable_vsx) {
+            __asm __volatile("xxland %vs0, %vs0, %vs0");
+            util_cpu_caps.has_vsx = 1;
+         }
+         signal(SIGILL, SIG_DFL);
+      } else {
+         util_cpu_caps.has_altivec = 0;
+      }
+   }
+#endif /* !PIPE_OS_APPLE */
+}
+#endif /* PIPE_ARCH_PPC */
+
+
+#if defined(PIPE_ARCH_X86) || defined (PIPE_ARCH_X86_64)
+static int has_cpuid(void)
+{
+#if defined(PIPE_ARCH_X86)
+#if defined(PIPE_OS_GCC)
+   int a, c;
+
+   __asm __volatile
+      ("pushf\n"
+       "popl %0\n"
+       "movl %0, %1\n"
+       "xorl $0x200000, %0\n"
+       "push %0\n"
+       "popf\n"
+       "pushf\n"
+       "popl %0\n"
+       : "=a" (a), "=c" (c)
+       :
+       : "cc");
+
+   return a != c;
+#else
+   /* FIXME */
+   return 1;
+#endif
+#elif defined(PIPE_ARCH_X86_64)
+   return 1;
+#else
+   return 0;
+#endif
+}
+
+
+/**
+ * @sa cpuid.h included in gcc-4.3 onwards.
+ * @sa http://msdn.microsoft.com/en-us/library/hskdteyh.aspx
+ */
+static inline void
+cpuid(uint32_t ax, uint32_t *p)
+{
+#if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
+   __asm __volatile (
+     "xchgl %%ebx, %1\n\t"
+     "cpuid\n\t"
+     "xchgl %%ebx, %1"
+     : "=a" (p[0]),
+       "=S" (p[1]),
+       "=c" (p[2]),
+       "=d" (p[3])
+     : "0" (ax)
+   );
+#elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
+   __asm __volatile (
+     "cpuid\n\t"
+     : "=a" (p[0]),
+       "=b" (p[1]),
+       "=c" (p[2]),
+       "=d" (p[3])
+     : "0" (ax)
+   );
+#elif defined(PIPE_CC_MSVC)
+   __cpuid(p, ax);
+#else
+   p[0] = 0;
+   p[1] = 0;
+   p[2] = 0;
+   p[3] = 0;
+#endif
+}
+
+/**
+ * @sa cpuid.h included in gcc-4.4 onwards.
+ * @sa http://msdn.microsoft.com/en-us/library/hskdteyh%28v=vs.90%29.aspx
+ */
+static inline void
+cpuid_count(uint32_t ax, uint32_t cx, uint32_t *p)
+{
+#if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
+   __asm __volatile (
+     "xchgl %%ebx, %1\n\t"
+     "cpuid\n\t"
+     "xchgl %%ebx, %1"
+     : "=a" (p[0]),
+       "=S" (p[1]),
+       "=c" (p[2]),
+       "=d" (p[3])
+     : "0" (ax), "2" (cx)
+   );
+#elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
+   __asm __volatile (
+     "cpuid\n\t"
+     : "=a" (p[0]),
+       "=b" (p[1]),
+       "=c" (p[2]),
+       "=d" (p[3])
+     : "0" (ax), "2" (cx)
+   );
+#elif defined(PIPE_CC_MSVC)
+   __cpuidex(p, ax, cx);
+#else
+   p[0] = 0;
+   p[1] = 0;
+   p[2] = 0;
+   p[3] = 0;
+#endif
+}
+
+
+static inline uint64_t xgetbv(void)
+{
+#if defined(PIPE_CC_GCC)
+   uint32_t eax, edx;
+
+   __asm __volatile (
+     ".byte 0x0f, 0x01, 0xd0" // xgetbv isn't supported on gcc < 4.4
+     : "=a"(eax),
+       "=d"(edx)
+     : "c"(0)
+   );
+
+   return ((uint64_t)edx << 32) | eax;
+#elif defined(PIPE_CC_MSVC) && defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
+   return _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
+#else
+   return 0;
+#endif
+}
+
+
+#if defined(PIPE_ARCH_X86)
+PIPE_ALIGN_STACK static inline boolean sse2_has_daz(void)
+{
+   struct {
+      uint32_t pad1[7];
+      uint32_t mxcsr_mask;
+      uint32_t pad2[128-8];
+   } PIPE_ALIGN_VAR(16) fxarea;
+
+   fxarea.mxcsr_mask = 0;
+#if defined(PIPE_CC_GCC)
+   __asm __volatile ("fxsave %0" : "+m" (fxarea));
+#elif defined(PIPE_CC_MSVC) || defined(PIPE_CC_ICL)
+   _fxsave(&fxarea);
+#else
+   fxarea.mxcsr_mask = 0;
+#endif
+   return !!(fxarea.mxcsr_mask & (1 << 6));
+}
+#endif
+
+#endif /* X86 or X86_64 */
+
+#if defined(PIPE_ARCH_ARM)
+static void
+check_os_arm_support(void)
+{
+   /*
+    * On Android, the cpufeatures library is preferred way of checking
+    * CPU capabilities. However, it is not available for standalone Mesa
+    * builds, i.e. when Android build system (Android.mk-based) is not
+    * used. Because of this we cannot use PIPE_OS_ANDROID here, but rather
+    * have a separate macro that only gets enabled from respective Android.mk.
+    */
+#if defined(HAS_ANDROID_CPUFEATURES)
+   AndroidCpuFamily cpu_family = android_getCpuFamily();
+   uint64_t cpu_features = android_getCpuFeatures();
+
+   if (cpu_family == ANDROID_CPU_FAMILY_ARM) {
+      if (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON)
+         util_cpu_caps.has_neon = 1;
+   }
+#elif defined(PIPE_OS_LINUX)
+    Elf32_auxv_t aux;
+    int fd;
+
+    fd = open("/proc/self/auxv", O_RDONLY | O_CLOEXEC);
+    if (fd >= 0) {
+       while (read(fd, &aux, sizeof(Elf32_auxv_t)) == sizeof(Elf32_auxv_t)) {
+          if (aux.a_type == AT_HWCAP) {
+             uint32_t hwcap = aux.a_un.a_val;
+
+             util_cpu_caps.has_neon = (hwcap >> 12) & 1;
+             break;
+          }
+       }
+       close (fd);
+    }
+#endif /* PIPE_OS_LINUX */
+}
+#endif /* PIPE_ARCH_ARM */
+
+static void
+get_cpu_topology(void)
+{
+   uint32_t regs[4];
+
+   /* Default. This is correct if L3 is not present or there is only one. */
+   util_cpu_caps.cores_per_L3 = util_cpu_caps.nr_cpus;
+
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   /* AMD Zen */
+   if (util_cpu_caps.x86_cpu_type == 0x17) {
+      /* Query the L3 cache topology information. */
+      cpuid_count(0x8000001D, 3, regs);
+      unsigned cache_level = (regs[0] >> 5) & 0x7;
+      unsigned cores_per_cache = ((regs[0] >> 14) & 0xfff) + 1;
+
+      if (cache_level == 3)
+         util_cpu_caps.cores_per_L3 = cores_per_cache;
+   }
+#endif
+}
+
+static void
+util_cpu_detect_once(void)
+{
+   memset(&util_cpu_caps, 0, sizeof util_cpu_caps);
+
+   /* Count the number of CPUs in system */
+#if defined(PIPE_OS_WINDOWS)
+   {
+      SYSTEM_INFO system_info;
+      GetSystemInfo(&system_info);
+      util_cpu_caps.nr_cpus = system_info.dwNumberOfProcessors;
+   }
+#elif defined(PIPE_OS_UNIX) && defined(_SC_NPROCESSORS_ONLN)
+   util_cpu_caps.nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+   if (util_cpu_caps.nr_cpus == ~0)
+      util_cpu_caps.nr_cpus = 1;
+#elif defined(PIPE_OS_BSD)
+   {
+      int mib[2], ncpu;
+      int len;
+
+      mib[0] = CTL_HW;
+      mib[1] = HW_NCPU;
+
+      len = sizeof (ncpu);
+      sysctl(mib, 2, &ncpu, &len, NULL, 0);
+      util_cpu_caps.nr_cpus = ncpu;
+   }
+#else
+   util_cpu_caps.nr_cpus = 1;
+#endif
+
+   /* Make the fallback cacheline size nonzero so that it can be
+    * safely passed to align().
+    */
+   util_cpu_caps.cacheline = sizeof(void *);
+
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   if (has_cpuid()) {
+      uint32_t regs[4];
+      uint32_t regs2[4];
+
+      util_cpu_caps.cacheline = 32;
+
+      /* Get max cpuid level */
+      cpuid(0x00000000, regs);
+
+      if (regs[0] >= 0x00000001) {
+         unsigned int cacheline;
+
+         cpuid (0x00000001, regs2);
+
+         util_cpu_caps.x86_cpu_type = (regs2[0] >> 8) & 0xf;
+         /* Add "extended family". */
+         if (util_cpu_caps.x86_cpu_type == 0xf)
+             util_cpu_caps.x86_cpu_type += ((regs2[0] >> 20) & 0xff);
+
+         /* general feature flags */
+         util_cpu_caps.has_tsc    = (regs2[3] >>  4) & 1; /* 0x0000010 */
+         util_cpu_caps.has_mmx    = (regs2[3] >> 23) & 1; /* 0x0800000 */
+         util_cpu_caps.has_sse    = (regs2[3] >> 25) & 1; /* 0x2000000 */
+         util_cpu_caps.has_sse2   = (regs2[3] >> 26) & 1; /* 0x4000000 */
+         util_cpu_caps.has_sse3   = (regs2[2] >>  0) & 1; /* 0x0000001 */
+         util_cpu_caps.has_ssse3  = (regs2[2] >>  9) & 1; /* 0x0000020 */
+         util_cpu_caps.has_sse4_1 = (regs2[2] >> 19) & 1;
+         util_cpu_caps.has_sse4_2 = (regs2[2] >> 20) & 1;
+         util_cpu_caps.has_popcnt = (regs2[2] >> 23) & 1;
+         util_cpu_caps.has_avx    = ((regs2[2] >> 28) & 1) && // AVX
+                                    ((regs2[2] >> 27) & 1) && // OSXSAVE
+                                    ((xgetbv() & 6) == 6);    // XMM & YMM
+         util_cpu_caps.has_f16c   = ((regs2[2] >> 29) & 1) && util_cpu_caps.has_avx;
+         util_cpu_caps.has_fma    = ((regs2[2] >> 12) & 1) && util_cpu_caps.has_avx;
+         util_cpu_caps.has_mmx2   = util_cpu_caps.has_sse; /* SSE cpus supports mmxext too */
+#if defined(PIPE_ARCH_X86_64)
+         util_cpu_caps.has_daz = 1;
+#else
+         util_cpu_caps.has_daz = util_cpu_caps.has_sse3 ||
+            (util_cpu_caps.has_sse2 && sse2_has_daz());
+#endif
+
+         cacheline = ((regs2[1] >> 8) & 0xFF) * 8;
+         if (cacheline > 0)
+            util_cpu_caps.cacheline = cacheline;
+      }
+      if (util_cpu_caps.has_avx && regs[0] >= 0x00000007) {
+         uint32_t regs7[4];
+         cpuid_count(0x00000007, 0x00000000, regs7);
+         util_cpu_caps.has_avx2 = (regs7[1] >> 5) & 1;
+      }
+
+      // check for avx512
+      if (((regs2[2] >> 27) & 1) && // OSXSAVE
+          (xgetbv() & (0x7 << 5)) && // OPMASK: upper-256 enabled by OS
+          ((xgetbv() & 6) == 6)) { // XMM/YMM enabled by OS
+         uint32_t regs3[4];
+         cpuid_count(0x00000007, 0x00000000, regs3);
+         util_cpu_caps.has_avx512f    = (regs3[1] >> 16) & 1;
+         util_cpu_caps.has_avx512dq   = (regs3[1] >> 17) & 1;
+         util_cpu_caps.has_avx512ifma = (regs3[1] >> 21) & 1;
+         util_cpu_caps.has_avx512pf   = (regs3[1] >> 26) & 1;
+         util_cpu_caps.has_avx512er   = (regs3[1] >> 27) & 1;
+         util_cpu_caps.has_avx512cd   = (regs3[1] >> 28) & 1;
+         util_cpu_caps.has_avx512bw   = (regs3[1] >> 30) & 1;
+         util_cpu_caps.has_avx512vl   = (regs3[1] >> 31) & 1;
+         util_cpu_caps.has_avx512vbmi = (regs3[2] >>  1) & 1;
+      }
+
+      if (regs[1] == 0x756e6547 && regs[2] == 0x6c65746e && regs[3] == 0x49656e69) {
+         /* GenuineIntel */
+         util_cpu_caps.has_intel = 1;
+      }
+
+      cpuid(0x80000000, regs);
+
+      if (regs[0] >= 0x80000001) {
+
+         cpuid(0x80000001, regs2);
+
+         util_cpu_caps.has_mmx  |= (regs2[3] >> 23) & 1;
+         util_cpu_caps.has_mmx2 |= (regs2[3] >> 22) & 1;
+         util_cpu_caps.has_3dnow = (regs2[3] >> 31) & 1;
+         util_cpu_caps.has_3dnow_ext = (regs2[3] >> 30) & 1;
+
+         util_cpu_caps.has_xop = util_cpu_caps.has_avx &&
+                                 ((regs2[2] >> 11) & 1);
+      }
+
+      if (regs[0] >= 0x80000006) {
+         /* should we really do this if the clflush size above worked? */
+         unsigned int cacheline;
+         cpuid(0x80000006, regs2);
+         cacheline = regs2[2] & 0xFF;
+         if (cacheline > 0)
+            util_cpu_caps.cacheline = cacheline;
+      }
+
+      if (!util_cpu_caps.has_sse) {
+         util_cpu_caps.has_sse2 = 0;
+         util_cpu_caps.has_sse3 = 0;
+         util_cpu_caps.has_ssse3 = 0;
+         util_cpu_caps.has_sse4_1 = 0;
+      }
+   }
+#endif /* PIPE_ARCH_X86 || PIPE_ARCH_X86_64 */
+
+#if defined(PIPE_ARCH_ARM)
+   check_os_arm_support();
+#endif
+
+#if defined(PIPE_ARCH_PPC)
+   check_os_altivec_support();
+#endif /* PIPE_ARCH_PPC */
+
+   get_cpu_topology();
+
+#ifdef DEBUG
+   if (debug_get_option_dump_cpu()) {
+      debug_printf("util_cpu_caps.nr_cpus = %u\n", util_cpu_caps.nr_cpus);
+
+      debug_printf("util_cpu_caps.x86_cpu_type = %u\n", util_cpu_caps.x86_cpu_type);
+      debug_printf("util_cpu_caps.cacheline = %u\n", util_cpu_caps.cacheline);
+
+      debug_printf("util_cpu_caps.has_tsc = %u\n", util_cpu_caps.has_tsc);
+      debug_printf("util_cpu_caps.has_mmx = %u\n", util_cpu_caps.has_mmx);
+      debug_printf("util_cpu_caps.has_mmx2 = %u\n", util_cpu_caps.has_mmx2);
+      debug_printf("util_cpu_caps.has_sse = %u\n", util_cpu_caps.has_sse);
+      debug_printf("util_cpu_caps.has_sse2 = %u\n", util_cpu_caps.has_sse2);
+      debug_printf("util_cpu_caps.has_sse3 = %u\n", util_cpu_caps.has_sse3);
+      debug_printf("util_cpu_caps.has_ssse3 = %u\n", util_cpu_caps.has_ssse3);
+      debug_printf("util_cpu_caps.has_sse4_1 = %u\n", util_cpu_caps.has_sse4_1);
+      debug_printf("util_cpu_caps.has_sse4_2 = %u\n", util_cpu_caps.has_sse4_2);
+      debug_printf("util_cpu_caps.has_avx = %u\n", util_cpu_caps.has_avx);
+      debug_printf("util_cpu_caps.has_avx2 = %u\n", util_cpu_caps.has_avx2);
+      debug_printf("util_cpu_caps.has_f16c = %u\n", util_cpu_caps.has_f16c);
+      debug_printf("util_cpu_caps.has_popcnt = %u\n", util_cpu_caps.has_popcnt);
+      debug_printf("util_cpu_caps.has_3dnow = %u\n", util_cpu_caps.has_3dnow);
+      debug_printf("util_cpu_caps.has_3dnow_ext = %u\n", util_cpu_caps.has_3dnow_ext);
+      debug_printf("util_cpu_caps.has_xop = %u\n", util_cpu_caps.has_xop);
+      debug_printf("util_cpu_caps.has_altivec = %u\n", util_cpu_caps.has_altivec);
+      debug_printf("util_cpu_caps.has_vsx = %u\n", util_cpu_caps.has_vsx);
+      debug_printf("util_cpu_caps.has_neon = %u\n", util_cpu_caps.has_neon);
+      debug_printf("util_cpu_caps.has_daz = %u\n", util_cpu_caps.has_daz);
+      debug_printf("util_cpu_caps.has_avx512f = %u\n", util_cpu_caps.has_avx512f);
+      debug_printf("util_cpu_caps.has_avx512dq = %u\n", util_cpu_caps.has_avx512dq);
+      debug_printf("util_cpu_caps.has_avx512ifma = %u\n", util_cpu_caps.has_avx512ifma);
+      debug_printf("util_cpu_caps.has_avx512pf = %u\n", util_cpu_caps.has_avx512pf);
+      debug_printf("util_cpu_caps.has_avx512er = %u\n", util_cpu_caps.has_avx512er);
+      debug_printf("util_cpu_caps.has_avx512cd = %u\n", util_cpu_caps.has_avx512cd);
+      debug_printf("util_cpu_caps.has_avx512bw = %u\n", util_cpu_caps.has_avx512bw);
+      debug_printf("util_cpu_caps.has_avx512vl = %u\n", util_cpu_caps.has_avx512vl);
+      debug_printf("util_cpu_caps.has_avx512vbmi = %u\n", util_cpu_caps.has_avx512vbmi);
+   }
+#endif
+}
+
+static once_flag cpu_once_flag = ONCE_FLAG_INIT;
+
+void
+util_cpu_detect(void)
+{
+   call_once(&cpu_once_flag, util_cpu_detect_once);
+}