2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013, by the GROMACS development team, led by
5 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
6 * others, as listed in the AUTHORS file in the top-level source
7 * directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
40 #include "visibility.h"
46 } /* fixes auto-indentation problems */
50 /* Currently identifiable CPU Vendors */
53 GMX_CPUID_VENDOR_CANNOTDETECT, /* Should only be used if something fails */
54 GMX_CPUID_VENDOR_UNKNOWN,
55 GMX_CPUID_VENDOR_INTEL,
57 GMX_CPUID_VENDOR_FUJITSU,
63 /* CPU feature/property list, to be used as indices into the feature array of the
64 * gmxcpuid_t data structure.
66 * To facilitate looking things up, we keep this list alphabetical.
67 * The list is NOT exhaustive - we have basically added stuff that might be
68 * useful in an application like Gromacs.
70 * AMD and Intel tend to share most architectural elements, and even if the
71 * flags might have to be detected in different ways (different cpuid registers),
72 * once the flag is present the functions should be identical. Unfortunately the
73 * trend right now (2012) seems to be that they are diverging. This means that
74 * we need to use specific flags to the compiler to maximize performance, and
75 * then the binaries might not be portable between Intel and AMD as they were
76 * before when we only needed to check for SSE and/or SSE2 support in Gromacs.
78 enum gmx_cpuid_feature
80 GMX_CPUID_FEATURE_CANNOTDETECT, /* Flag set if we could not detect on this CPU */
81 GMX_CPUID_FEATURE_X86_AES, /* x86 advanced encryption standard accel. */
82 GMX_CPUID_FEATURE_X86_APIC, /* APIC support */
83 GMX_CPUID_FEATURE_X86_AVX, /* Advanced vector extensions */
84 GMX_CPUID_FEATURE_X86_AVX2, /* AVX2 including gather support (not used yet) */
85 GMX_CPUID_FEATURE_X86_CLFSH, /* Supports CLFLUSH instruction */
86 GMX_CPUID_FEATURE_X86_CMOV, /* Conditional move insn support */
87 GMX_CPUID_FEATURE_X86_CX8, /* Supports CMPXCHG8B (8-byte compare-exchange) */
88 GMX_CPUID_FEATURE_X86_CX16, /* Supports CMPXCHG16B (16-byte compare-exchg) */
89 GMX_CPUID_FEATURE_X86_F16C, /* Supports 16-bit FP conversion instructions */
90 GMX_CPUID_FEATURE_X86_FMA, /* Fused-multiply add support (mainly for AVX) */
91 GMX_CPUID_FEATURE_X86_FMA4, /* 4-operand FMA, only on AMD for now */
92 GMX_CPUID_FEATURE_X86_HTT, /* Hyper-Threading supported */
93 GMX_CPUID_FEATURE_X86_LAHF_LM, /* LAHF/SAHF support in 64 bits */
94 GMX_CPUID_FEATURE_X86_MISALIGNSSE, /* Support for misaligned SSE data instructions */
95 GMX_CPUID_FEATURE_X86_MMX, /* MMX registers and instructions */
96 GMX_CPUID_FEATURE_X86_MSR, /* Supports Intel model-specific-registers */
97 GMX_CPUID_FEATURE_X86_NONSTOP_TSC, /* Invariant TSC (constant rate in ACPI states) */
98 GMX_CPUID_FEATURE_X86_PCID, /* Process context identifier support */
99 GMX_CPUID_FEATURE_X86_PCLMULDQ, /* Carry-less 64-bit multiplication supported */
100 GMX_CPUID_FEATURE_X86_PDCM, /* Perfmon and Debug Capability */
101 GMX_CPUID_FEATURE_X86_PDPE1GB, /* Support for 1GB pages */
102 GMX_CPUID_FEATURE_X86_POPCNT, /* Supports the POPCNT (population count) insn */
103 GMX_CPUID_FEATURE_X86_PSE, /* Supports 4MB-pages (page size extension) */
104 GMX_CPUID_FEATURE_X86_RDRND, /* RDRAND high-quality hardware random numbers */
105 GMX_CPUID_FEATURE_X86_RDTSCP, /* Serializing rdtscp instruction available */
106 GMX_CPUID_FEATURE_X86_SSE2, /* SSE 2 */
107 GMX_CPUID_FEATURE_X86_SSE3, /* SSE 3 */
108 GMX_CPUID_FEATURE_X86_SSE4A, /* SSE 4A */
109 GMX_CPUID_FEATURE_X86_SSE4_1, /* SSE 4.1 */
110 GMX_CPUID_FEATURE_X86_SSE4_2, /* SSE 4.2 */
111 GMX_CPUID_FEATURE_X86_SSSE3, /* Supplemental SSE3 */
112 GMX_CPUID_FEATURE_X86_TDT, /* TSC deadline timer */
113 GMX_CPUID_FEATURE_X86_X2APIC, /* Extended xAPIC Support */
114 GMX_CPUID_FEATURE_X86_XOP, /* AMD extended instructions, only AMD for now */
119 /* Currently supported acceleration instruction sets, intrinsics or other similar combinations
120 * in Gromacs. There is not always a 1-to-1 correspondence with feature flags; on some AMD
121 * hardware we prefer to use 128bit AVX instructions (although 256-bit ones could be executed),
122 * and we still haven't written the AVX2 kernels.
124 enum gmx_cpuid_acceleration
126 GMX_CPUID_ACCELERATION_CANNOTDETECT, /* Should only be used if something fails */
127 GMX_CPUID_ACCELERATION_NONE,
128 GMX_CPUID_ACCELERATION_X86_SSE2,
129 GMX_CPUID_ACCELERATION_X86_SSE4_1,
130 GMX_CPUID_ACCELERATION_X86_AVX_128_FMA,
131 GMX_CPUID_ACCELERATION_X86_AVX_256,
132 GMX_CPUID_ACCELERATION_SPARC64_HPC_ACE,
133 GMX_CPUID_ACCELERATION_IBM_QPX,
134 GMX_CPUID_NACCELERATIONS
137 /* Text strings corresponding to CPU vendors */
140 gmx_cpuid_vendor_string[GMX_CPUID_NVENDORS];
142 /* Text strings for CPU feature indices */
144 gmx_cpuid_feature_string[GMX_CPUID_NFEATURES];
146 /* Text strings for Gromacs acceleration/instruction sets */
148 gmx_cpuid_acceleration_string[GMX_CPUID_NACCELERATIONS];
151 /* Abstract data type with CPU detection information. Set by gmx_cpuid_init(). */
152 typedef struct gmx_cpuid *
156 /* Fill the data structure by using CPU detection instructions.
157 * Return 0 on success, 1 if something bad happened.
160 gmx_cpuid_init (gmx_cpuid_t * cpuid);
163 /* Return the vendor id as enumerated type. Use gmx_cpuid_vendor_string[]
164 * to get the corresponding text string.
167 enum gmx_cpuid_vendor
168 gmx_cpuid_vendor (gmx_cpuid_t cpuid);
171 /* Return a constant pointer to the processor brand string. */
173 gmx_cpuid_brand (gmx_cpuid_t cpuid);
176 /* Return processor family version. For a chip of version 1.2.3, this is 1 */
179 gmx_cpuid_family (gmx_cpuid_t cpuid);
181 /* Return processor model version, For a chip of version 1.2.3, this is 2. */
184 gmx_cpuid_model (gmx_cpuid_t cpuid);
186 /* Return processor stepping version, For a chip of version 1.2.3, this is 3. */
188 gmx_cpuid_stepping (gmx_cpuid_t cpuid);
191 /* Check whether a particular CPUID feature is set.
192 * Returns 0 if flag "feature" is not set, 1 if the flag is set. We cannot use
193 * gmx_bool here since this file must be possible to compile without simple.h.
197 gmx_cpuid_feature (gmx_cpuid_t cpuid,
198 enum gmx_cpuid_feature feature);
201 /* Return pointers to cpu topology information.
203 * Important: CPU topology requires more OS support than most other
204 * functions in this file, including support for thread pinning to hardware.
205 * This means it will not work on some platforms, including e.g. Mac OS X.
206 * Thus, it is IMPERATIVE that you check the return value from this routine
207 * before doing anything with the information. It is only if the return
208 * value is zero that the data is valid.
210 * For the returned values we have:
211 * - nprocessors Total number of logical processors reported by OS
212 * - npackages Usually number of CPU sockets
213 * - ncores_per_package Number of cores in each package
214 * - nhwthreads_per_core Number of hardware threads per core; 2 for hyperthreading.
215 * - package_id Array with the package index for each logical cpu
216 * - core_id Array with local core index for each logical cpu
217 * - hwthread_id Array with local hwthread index for each logical cpu
218 * - locality_order Array with logical cpu numbers, sorted in order
219 * of physical and logical locality in the system.
221 * All arrays are of length nprocessors.
225 gmx_cpuid_topology(gmx_cpuid_t cpuid,
228 int * ncores_per_package,
229 int * nhwthreads_per_core,
230 const int ** package_id,
231 const int ** core_id,
232 const int ** hwthread_id,
233 const int ** locality_order);
235 /* Enumerated values for x86 SMT enabled-status. Note that this does not refer
236 * to Hyper-Threading support (that is the flag GMX_CPUID_FEATURE_X86_HTT), but
237 * whether Hyper-Threading is _enabled_ and _used_ in bios right now.
239 enum gmx_cpuid_x86_smt
241 GMX_CPUID_X86_SMT_CANNOTDETECT,
242 GMX_CPUID_X86_SMT_DISABLED,
243 GMX_CPUID_X86_SMT_ENABLED
246 /* Returns the status of x86 SMT support. IMPORTANT: There are non-zero
247 * return values for this routine that still do not indicate supported and
248 * enabled smt/Hyper-Threading. You need to carefully check the return value
249 * against the enumerated type values to see what you are getting.
251 * Long-term, this functionality will move to a new hardware topology detection
252 * layer, but that will require a lot of new code and a working interface to the
253 * hwloc library. Surprisingly, there is no simple way to find out that
254 * Hyper-Threading is actually turned on without fully enumerating and checking
255 * all the cores, which we presently can only do on Linux. This means a couple
258 * 1) If you want to know whether your CPU _supports_ Hyper-Threading in the
259 * first place, check the GMX_CPUID_FEATURE_X86_HTT flag instead!
260 * 2) There are several scenarios where this routine will say that it cannot
261 * detect whether SMT is enabled and used right now.
262 * 3) If you need support on non-Linux x86, you have to write it :-)
263 * 4) Don't invest too much efforts, since this will be replaced with
264 * full hardware topology detection in the future.
265 * 5) Don't worry if the detection does not work. It is not a catastrophe, but
266 * but we get slightly better performance on x86 if we use Hyper-Threading
267 * cores in direct space, but not reciprocal space.
269 * Since this routine presently only supports Hyper-Threading we say X86_SMT
270 * in order not to give the impression we can detect any SMT. We haven't
271 * even tested the performance on other SMT implementations, so it is not
272 * obvious we shouldn't use SMT there.
274 * Note that you can get more complete topology information from
275 * gmx_cpuid_topology(), although that requires slightly more OS support.
278 enum gmx_cpuid_x86_smt
279 gmx_cpuid_x86_smt(gmx_cpuid_t cpuid);
282 /* Formats a text string (up to n characters) from the data structure.
283 * The output will have max 80 chars between newline characters.
286 gmx_cpuid_formatstring (gmx_cpuid_t cpuid,
291 /* Suggests a suitable gromacs acceleration based on the support in the
294 enum gmx_cpuid_acceleration
295 gmx_cpuid_acceleration_suggest (gmx_cpuid_t cpuid);
298 /* Check if this binary was compiled with the same acceleration as we
299 * would suggest for the current hardware. Always print stats to the log file
300 * if it is non-NULL, and print a warning in stdout if we don't have a match.
303 gmx_cpuid_acceleration_check (gmx_cpuid_t cpuid,
307 /* Release resources used by data structure. Note that the pointer to the
308 * CPU brand string will no longer be valid once this routine has been called.
311 gmx_cpuid_done (gmx_cpuid_t cpuid);
321 #endif /* GMX_CPUID_H_ */