/*
* This file is part of the GROMACS molecular simulation package.
*
- * Copyright (c) 2012,2013, by the GROMACS development team, led by
+ * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
* Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
* and including many others, as listed in the AUTHORS file in the
* top-level source directory and at http://www.gromacs.org.
* To help us fund GROMACS development, we humbly ask that you cite
* the research papers on the package. Check out http://www.gromacs.org.
*/
+#include "gmxpre.h"
+
#ifdef HAVE_CONFIG_H
-#include <config.h>
+#include "config.h"
#endif
#ifdef HAVE_SCHED_H
-#define _GNU_SOURCE
-#include <sched.h>
+# ifndef _GNU_SOURCE
+# define _GNU_SOURCE 1
+# endif
+# include <sched.h>
#endif
+#include <ctype.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
-#include <ctype.h>
-#ifdef _MSC_VER
+#ifdef GMX_NATIVE_WINDOWS
/* MSVC definition for __cpuid() */
-#include <intrin.h>
+ #ifdef _MSC_VER
+ #include <intrin.h>
+ #endif
/* sysinfo functions */
-#include <windows.h>
+ #include <windows.h>
#endif
#ifdef HAVE_UNISTD_H
/* sysconf() definition */
-#include <unistd.h>
+ #include <unistd.h>
#endif
-#include "gmx_cpuid.h"
+#include "gromacs/legacyheaders/gmx_cpuid.h"
"GenuineIntel",
"AuthenticAMD",
"Fujitsu",
- "IBM"
+ "IBM",
+ "ARM"
};
const char *
"GenuineIntel",
"AuthenticAMD",
"Fujitsu",
- "ibm" /* Used on BlueGene/Q */
+ "ibm", /* Used on BlueGene/Q */
+ "AArch64"
};
const char *
"ssse3",
"tdt",
"x2apic",
- "xop"
+ "xop",
+ "arm_neon",
+ "arm_neon_asimd"
};
const char *
-gmx_cpuid_acceleration_string[GMX_CPUID_NACCELERATIONS] =
+gmx_cpuid_simd_string[GMX_CPUID_NSIMD] =
{
"CannotDetect",
"None",
+ "Reference",
"SSE2",
"SSE4.1",
"AVX_128_FMA",
"AVX_256",
+ "AVX2_256",
"Sparc64 HPC-ACE",
- "IBM_QPX"
+ "IBM_QPX",
+ "ARM_NEON",
+ "ARM_NEON_ASIMD"
};
/* Max length of brand string */
-#define GMX_CPUID_BRAND_MAXLEN 256
+#define GMX_CPUID_STRLEN 256
/* Contents of the abstract datatype */
struct gmx_cpuid
{
enum gmx_cpuid_vendor vendor;
- char brand[GMX_CPUID_BRAND_MAXLEN];
+ char brand[GMX_CPUID_STRLEN];
int family;
int model;
int stepping;
-/* What type of acceleration was compiled in, if any?
- * This is set from Cmake. Note that the SSE2 and SSE4_1 macros are set for
- * AVX too, so it is important that they appear last in the list.
- */
-#ifdef GMX_X86_AVX_256
-static const
-enum gmx_cpuid_acceleration
- compiled_acc = GMX_CPUID_ACCELERATION_X86_AVX_256;
-#elif defined GMX_X86_AVX_128_FMA
-static const
-enum gmx_cpuid_acceleration
- compiled_acc = GMX_CPUID_ACCELERATION_X86_AVX_128_FMA;
-#elif defined GMX_X86_SSE4_1
-static const
-enum gmx_cpuid_acceleration
- compiled_acc = GMX_CPUID_ACCELERATION_X86_SSE4_1;
-#elif defined GMX_X86_SSE2
-static const
-enum gmx_cpuid_acceleration
- compiled_acc = GMX_CPUID_ACCELERATION_X86_SSE2;
-#elif defined GMX_CPU_ACCELERATION_SPARC64_HPC_ACE
-static const
-enum gmx_cpuid_acceleration
- compiled_acc = GMX_CPUID_ACCELERATION_SPARC64_HPC_ACE;
-#elif defined GMX_CPU_ACCELERATION_IBM_QPX
-static const
-enum gmx_cpuid_acceleration
- compiled_acc = GMX_CPUID_ACCELERATION_IBM_QPX;
+/* What type of SIMD was compiled in, if any? */
+#ifdef GMX_SIMD_X86_AVX2_256
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_X86_AVX2_256;
+#elif defined GMX_SIMD_X86_AVX_256
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_X86_AVX_256;
+#elif defined GMX_SIMD_X86_AVX_128_FMA
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_X86_AVX_128_FMA;
+#elif defined GMX_SIMD_X86_SSE4_1
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_X86_SSE4_1;
+#elif defined GMX_SIMD_X86_SSE2
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_X86_SSE2;
+#elif defined GMX_SIMD_ARM_NEON
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_ARM_NEON;
+#elif defined GMX_SIMD_ARM_NEON_ASIMD
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_ARM_NEON_ASIMD;
+#elif defined GMX_SIMD_SPARC64_HPC_ACE
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_SPARC64_HPC_ACE;
+#elif defined GMX_SIMD_IBM_QPX
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_IBM_QPX;
+#elif defined GMX_SIMD_REFERENCE
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_REFERENCE;
#else
-static const
-enum gmx_cpuid_acceleration
- compiled_acc = GMX_CPUID_ACCELERATION_NONE;
+static const enum gmx_cpuid_simd compiled_simd = GMX_CPUID_SIMD_NONE;
#endif
/* Death and horror!
* Apparently this is an x86 platform where we don't know how to call cpuid.
*
- * This is REALLY bad, since we will lose all Gromacs acceleration.
+ * This is REALLY bad, since we will lose all Gromacs SIMD support.
*/
*eax = 0;
*ebx = 0;
{
int fn, max_stdfn, max_extfn;
unsigned int eax, ebx, ecx, edx;
- char str[GMX_CPUID_BRAND_MAXLEN];
+ char str[GMX_CPUID_STRLEN];
char * p;
/* Find largest standard/extended function input value */
{
p++;
}
- strncpy(cpuid->brand, p, GMX_CPUID_BRAND_MAXLEN);
+ strncpy(cpuid->brand, p, GMX_CPUID_STRLEN);
}
else
{
- strncpy(cpuid->brand, "Unknown CPU brand", GMX_CPUID_BRAND_MAXLEN);
+ strncpy(cpuid->brand, "Unknown CPU brand", GMX_CPUID_STRLEN);
}
/* Find basic CPU properties */
}
}
}
+ free(unique);
return nunique;
}
/* Query APIC information on AMD */
if (max_extfn >= 0x80000008)
{
-#if (defined HAVE_SCHED_H && defined HAVE_SCHED_SETAFFINITY && defined HAVE_SYSCONF && defined __linux__)
+#if (defined HAVE_SCHED_AFFINITY && defined HAVE_SYSCONF && defined __linux__)
/* Linux */
unsigned int i;
cpu_set_t cpuset, save_cpuset;
if (max_stdfn >= 0xB)
{
/* Query x2 APIC information from cores */
-#if (defined HAVE_SCHED_H && defined HAVE_SCHED_SETAFFINITY && defined HAVE_SYSCONF && defined __linux__)
+#if (defined HAVE_SCHED_AFFINITY && defined HAVE_SYSCONF && defined __linux__)
/* Linux */
unsigned int i;
cpu_set_t cpuset, save_cpuset;
-
static void
chomp_substring_before_colon(const char *in, char *s, int maxlength)
{
}
}
+static int
+cpuid_check_arm(gmx_cpuid_t cpuid)
+{
+#if defined(__linux__) || defined(__linux)
+ FILE *fp;
+ char buffer[GMX_CPUID_STRLEN], buffer2[GMX_CPUID_STRLEN], buffer3[GMX_CPUID_STRLEN];
+
+ if ( (fp = fopen("/proc/cpuinfo", "r")) != NULL)
+ {
+ while ( (fgets(buffer, sizeof(buffer), fp) != NULL))
+ {
+ chomp_substring_before_colon(buffer, buffer2, GMX_CPUID_STRLEN);
+ chomp_substring_after_colon(buffer, buffer3, GMX_CPUID_STRLEN);
+
+ if (!strcmp(buffer2, "Processor"))
+ {
+ strncpy(cpuid->brand, buffer3, GMX_CPUID_STRLEN);
+ }
+ else if (!strcmp(buffer2, "CPU architecture"))
+ {
+ cpuid->family = strtol(buffer3, NULL, 10);
+ if (!strcmp(buffer3, "AArch64"))
+ {
+ cpuid->family = 8;
+ }
+ }
+ else if (!strcmp(buffer2, "CPU part"))
+ {
+ cpuid->model = strtol(buffer3, NULL, 16);
+ }
+ else if (!strcmp(buffer2, "CPU revision"))
+ {
+ cpuid->stepping = strtol(buffer3, NULL, 10);
+ }
+ else if (!strcmp(buffer2, "Features") && strstr(buffer3, "neon"))
+ {
+ cpuid->feature[GMX_CPUID_FEATURE_ARM_NEON] = 1;
+ }
+ else if (!strcmp(buffer2, "Features") && strstr(buffer3, "asimd"))
+ {
+ cpuid->feature[GMX_CPUID_FEATURE_ARM_NEON_ASIMD] = 1;
+ }
+ }
+ }
+ fclose(fp);
+#else
+# ifdef __aarch64__
+ /* Strange 64-bit non-linux platform. However, since NEON ASIMD is present on all
+ * implementations of AArch64 this far, we assume it is present for now.
+ */
+ cpuid->feature[GMX_CPUID_FEATURE_ARM_NEON_ASIMD] = 1;
+# else
+ /* Strange 32-bit non-linux platform. We cannot assume that neon is present. */
+ cpuid->feature[GMX_CPUID_FEATURE_ARM_NEON] = 0;
+# endif
+#endif
+ return 0;
+}
+
+
/* Try to find the vendor of the current CPU, so we know what specific
* detection routine to call.
*/
unsigned int eax, ebx, ecx, edx;
char vendorstring[13];
FILE * fp;
- char buffer[255], before_colon[255], after_colon[255];
+ char buffer[GMX_CPUID_STRLEN];
+ char before_colon[GMX_CPUID_STRLEN];
+ char after_colon[GMX_CPUID_STRLEN];
/* Set default first */
vendor = GMX_CPUID_VENDOR_UNKNOWN;
while ( (vendor == GMX_CPUID_VENDOR_UNKNOWN) && (fgets(buffer, sizeof(buffer), fp) != NULL))
{
chomp_substring_before_colon(buffer, before_colon, sizeof(before_colon));
- /* Intel/AMD use "vendor_id", IBM "vendor"(?) or "model". Fujitsu "manufacture". Add others if you have them! */
+ /* Intel/AMD use "vendor_id", IBM "vendor"(?) or "model". Fujitsu "manufacture".
+ * On ARM there does not seem to be a vendor, but ARM or AArch64 is listed in the Processor string.
+ * Add others if you have them!
+ */
if (!strcmp(before_colon, "vendor_id")
|| !strcmp(before_colon, "vendor")
|| !strcmp(before_colon, "manufacture")
- || !strcmp(before_colon, "model"))
+ || !strcmp(before_colon, "model")
+ || !strcmp(before_colon, "Processor"))
{
chomp_substring_after_colon(buffer, after_colon, sizeof(after_colon));
for (i = GMX_CPUID_VENDOR_UNKNOWN; i < GMX_CPUID_NVENDORS; i++)
}
}
fclose(fp);
+#elif defined(__arm__) || defined (__arm) || defined(__aarch64__)
+ /* If we are using ARM on something that is not linux we have to trust the compiler,
+ * and we cannot get the extra info that might be present in /proc/cpuinfo.
+ */
+ vendor = GMX_CPUID_VENDOR_ARM;
#endif
-
return vendor;
}
gmx_cpuid_t cpuid;
int i;
FILE * fp;
- char buffer[255], buffer2[255];
+ char buffer[GMX_CPUID_STRLEN], buffer2[GMX_CPUID_STRLEN];
int found_brand;
cpuid = malloc(sizeof(*cpuid));
cpuid_check_amd_x86(cpuid);
break;
#endif
+ case GMX_CPUID_VENDOR_ARM:
+ cpuid_check_arm(cpuid);
+ break;
default:
/* Default value */
- strncpy(cpuid->brand, "Unknown CPU brand", GMX_CPUID_BRAND_MAXLEN);
+ strncpy(cpuid->brand, "Unknown CPU brand", GMX_CPUID_STRLEN);
#if defined(__linux__) || defined(__linux)
/* General Linux. Try to get CPU type from /proc/cpuinfo */
if ( (fp = fopen("/proc/cpuinfo", "r")) != NULL)
/* Intel uses "model name", Fujitsu and IBM "cpu". */
if (!strcmp(buffer2, "model name") || !strcmp(buffer2, "cpu"))
{
- chomp_substring_after_colon(buffer, cpuid->brand, GMX_CPUID_BRAND_MAXLEN);
+ chomp_substring_after_colon(buffer, cpuid->brand, GMX_CPUID_STRLEN);
found_brand = 1;
}
}
-enum gmx_cpuid_acceleration
-gmx_cpuid_acceleration_suggest (gmx_cpuid_t cpuid)
+enum gmx_cpuid_simd
+gmx_cpuid_simd_suggest (gmx_cpuid_t cpuid)
{
- enum gmx_cpuid_acceleration tmpacc;
+ enum gmx_cpuid_simd tmpsimd;
- tmpacc = GMX_CPUID_ACCELERATION_NONE;
+ tmpsimd = GMX_CPUID_SIMD_NONE;
if (gmx_cpuid_vendor(cpuid) == GMX_CPUID_VENDOR_INTEL)
{
- if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_X86_AVX))
+ if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_X86_AVX2))
+ {
+ tmpsimd = GMX_CPUID_SIMD_X86_AVX2_256;
+ }
+ else if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_X86_AVX))
{
- tmpacc = GMX_CPUID_ACCELERATION_X86_AVX_256;
+ tmpsimd = GMX_CPUID_SIMD_X86_AVX_256;
}
else if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_X86_SSE4_1))
{
- tmpacc = GMX_CPUID_ACCELERATION_X86_SSE4_1;
+ tmpsimd = GMX_CPUID_SIMD_X86_SSE4_1;
}
else if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_X86_SSE2))
{
- tmpacc = GMX_CPUID_ACCELERATION_X86_SSE2;
+ tmpsimd = GMX_CPUID_SIMD_X86_SSE2;
}
}
else if (gmx_cpuid_vendor(cpuid) == GMX_CPUID_VENDOR_AMD)
{
if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_X86_AVX))
{
- tmpacc = GMX_CPUID_ACCELERATION_X86_AVX_128_FMA;
+ tmpsimd = GMX_CPUID_SIMD_X86_AVX_128_FMA;
}
else if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_X86_SSE4_1))
{
- tmpacc = GMX_CPUID_ACCELERATION_X86_SSE4_1;
+ tmpsimd = GMX_CPUID_SIMD_X86_SSE4_1;
}
else if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_X86_SSE2))
{
- tmpacc = GMX_CPUID_ACCELERATION_X86_SSE2;
+ tmpsimd = GMX_CPUID_SIMD_X86_SSE2;
}
}
else if (gmx_cpuid_vendor(cpuid) == GMX_CPUID_VENDOR_FUJITSU)
{
if (strstr(gmx_cpuid_brand(cpuid), "SPARC64"))
{
- tmpacc = GMX_CPUID_ACCELERATION_SPARC64_HPC_ACE;
+ tmpsimd = GMX_CPUID_SIMD_SPARC64_HPC_ACE;
}
}
else if (gmx_cpuid_vendor(cpuid) == GMX_CPUID_VENDOR_IBM)
{
if (strstr(gmx_cpuid_brand(cpuid), "A2"))
{
- tmpacc = GMX_CPUID_ACCELERATION_IBM_QPX;
+ tmpsimd = GMX_CPUID_SIMD_IBM_QPX;
+ }
+ }
+ else if (gmx_cpuid_vendor(cpuid) == GMX_CPUID_VENDOR_ARM)
+ {
+ if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_ARM_NEON_ASIMD))
+ {
+ tmpsimd = GMX_CPUID_SIMD_ARM_NEON_ASIMD;
+ }
+ else if (gmx_cpuid_feature(cpuid, GMX_CPUID_FEATURE_ARM_NEON))
+ {
+ tmpsimd = GMX_CPUID_SIMD_ARM_NEON;
}
}
- return tmpacc;
+ return tmpsimd;
}
int
-gmx_cpuid_acceleration_check(gmx_cpuid_t cpuid,
- FILE * log,
- int print_to_stderr)
+gmx_cpuid_simd_check(gmx_cpuid_t cpuid,
+ FILE * log,
+ int print_to_stderr)
{
int rc;
char str[1024];
- enum gmx_cpuid_acceleration acc;
+ enum gmx_cpuid_simd simd;
- acc = gmx_cpuid_acceleration_suggest(cpuid);
+ simd = gmx_cpuid_simd_suggest(cpuid);
- rc = (acc != compiled_acc);
+ rc = (simd != compiled_simd);
gmx_cpuid_formatstring(cpuid, str, 1023);
str[1023] = '\0';
if (log != NULL)
{
fprintf(log,
- "\nDetecting CPU-specific acceleration.\nPresent hardware specification:\n"
+ "\nDetecting CPU SIMD instructions.\nPresent hardware specification:\n"
"%s"
- "Acceleration most likely to fit this hardware: %s\n"
- "Acceleration selected at GROMACS compile time: %s\n\n",
+ "SIMD instructions most likely to fit this hardware: %s\n"
+ "SIMD instructions selected at GROMACS compile time: %s\n\n",
str,
- gmx_cpuid_acceleration_string[acc],
- gmx_cpuid_acceleration_string[compiled_acc]);
+ gmx_cpuid_simd_string[simd],
+ gmx_cpuid_simd_string[compiled_simd]);
}
if (rc != 0)
if (log != NULL)
{
fprintf(log, "\nBinary not matching hardware - you might be losing performance.\n"
- "Acceleration most likely to fit this hardware: %s\n"
- "Acceleration selected at GROMACS compile time: %s\n\n",
- gmx_cpuid_acceleration_string[acc],
- gmx_cpuid_acceleration_string[compiled_acc]);
+ "SIMD instructions most likely to fit this hardware: %s\n"
+ "SIMD instructions selected at GROMACS compile time: %s\n\n",
+ gmx_cpuid_simd_string[simd],
+ gmx_cpuid_simd_string[compiled_simd]);
}
if (print_to_stderr)
{
- fprintf(stderr, "Compiled acceleration: %s (Gromacs could use %s on this machine, which is better)\n",
- gmx_cpuid_acceleration_string[compiled_acc],
- gmx_cpuid_acceleration_string[acc]);
+ fprintf(stderr, "Compiled SIMD instructions: %s (Gromacs could use %s on this machine, which is better)\n",
+ gmx_cpuid_simd_string[compiled_simd],
+ gmx_cpuid_simd_string[simd]);
}
}
return rc;
main(int argc, char **argv)
{
gmx_cpuid_t cpuid;
- enum gmx_cpuid_acceleration acc;
+ enum gmx_cpuid_simd simd;
int i, cnt;
if (argc < 2)
"-model Print CPU model version.\n"
"-stepping Print CPU stepping version.\n"
"-features Print CPU feature flags.\n"
- "-acceleration Print suggested GROMACS acceleration.\n",
+ "-simd Print suggested GROMACS SIMD instructions.\n",
argv[0]);
exit(0);
}
}
printf("\n");
}
- else if (!strncmp(argv[1], "-acceleration", 3))
+ else if (!strncmp(argv[1], "-simd", 3))
{
- acc = gmx_cpuid_acceleration_suggest(cpuid);
- fprintf(stdout, "%s\n", gmx_cpuid_acceleration_string[acc]);
+ simd = gmx_cpuid_simd_suggest(cpuid);
+ fprintf(stdout, "%s\n", gmx_cpuid_simd_string[simd]);
}
gmx_cpuid_done(cpuid);