#include "pipe/p_config.h"
-#include "u_debug.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 <machine/cpu.h>
#endif
-#if defined(PIPE_OS_FREEBSD)
+#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)
if (setjmp(__lv_powerpc_jmpbuf)) {
signal(SIGILL, SIG_DFL);
} else {
- __lv_powerpc_canjump = 1;
+ 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));
+ __asm __volatile
+ ("mtspr 256, %0\n\t"
+ "vand %%v0, %%v0, %%v0"
+ :
+ : "r" (-1));
- signal(SIGILL, SIG_DFL);
- util_cpu_caps.has_altivec = 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 */
}
* @sa cpuid.h included in gcc-4.3 onwards.
* @sa http://msdn.microsoft.com/en-us/library/hskdteyh.aspx
*/
-static INLINE void
+static inline void
cpuid(uint32_t ax, uint32_t *p)
{
-#if (defined(PIPE_CC_GCC) || defined(PIPE_CC_SUNPRO)) && defined(PIPE_ARCH_X86)
+#if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
__asm __volatile (
"xchgl %%ebx, %1\n\t"
"cpuid\n\t"
"=d" (p[3])
: "0" (ax)
);
-#elif (defined(PIPE_CC_GCC) || defined(PIPE_CC_SUNPRO)) && defined(PIPE_ARCH_X86_64)
+#elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
__asm __volatile (
"cpuid\n\t"
: "=a" (p[0]),
#endif
}
-static INLINE uint64_t xgetbv(void)
+/**
+ * @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;
#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 */
-void
-util_cpu_detect(void)
+#if defined(PIPE_ARCH_ARM)
+static void
+check_os_arm_support(void)
{
- static boolean util_cpu_detect_initialized = FALSE;
+ /*
+ * 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(util_cpu_detect_initialized)
- return;
+ 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 */
}
#elif defined(PIPE_OS_UNIX) && defined(_SC_NPROCESSORS_ONLN)
util_cpu_caps.nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
- if (util_cpu_caps.nr_cpus == -1)
+ if (util_cpu_caps.nr_cpus == ~0)
util_cpu_caps.nr_cpus = 1;
#elif defined(PIPE_OS_BSD)
{
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 = 8 + ((regs2[0] >> 20) & 255); /* use extended family (P4, IA64) */
+ 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_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_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_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);
- util_cpu_caps.cacheline = regs2[2] & 0xFF;
+ cacheline = regs2[2] & 0xFF;
+ if (cacheline > 0)
+ util_cpu_caps.cacheline = cacheline;
}
if (!util_cpu_caps.has_sse) {
}
#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.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;
- util_cpu_detect_initialized = TRUE;
+void
+util_cpu_detect(void)
+{
+ call_once(&cpu_once_flag, util_cpu_detect_once);
}