2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012, 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.
37 #include "visibility.h"
42 } /* fixes auto-indentation problems */
46 /* Currently identifiable CPU Vendors */
49 GMX_CPUID_VENDOR_CANNOTDETECT, /* Should only be used if something fails */
50 GMX_CPUID_VENDOR_UNKNOWN,
51 GMX_CPUID_VENDOR_INTEL,
57 /* CPU feature/property list, to be used as indices into the feature array of the
58 * gmxcpuid_t data structure.
60 * To facilitate looking things up, we keep this list alphabetical.
61 * The list is NOT exhaustive - we have basically added stuff that might be
62 * useful in an application like Gromacs.
64 * AMD and Intel tend to share most architectural elements, and even if the
65 * flags might have to be detected in different ways (different cpuid registers),
66 * once the flag is present the functions should be identical. Unfortunately the
67 * trend right now (2012) seems to be that they are diverging. This means that
68 * we need to use specific flags to the compiler to maximize performance, and
69 * then the binaries might not be portable between Intel and AMD as they were
70 * before when we only needed to check for SSE and/or SSE2 support in Gromacs.
72 enum gmx_cpuid_feature
74 GMX_CPUID_FEATURE_CANNOTDETECT, /* Flag set if we could not detect on this CPU */
75 GMX_CPUID_FEATURE_X86_AES, /* x86 advanced encryption standard accel. */
76 GMX_CPUID_FEATURE_X86_APIC, /* APIC support */
77 GMX_CPUID_FEATURE_X86_AVX, /* Advanced vector extensions */
78 GMX_CPUID_FEATURE_X86_AVX2, /* AVX2 including gather support (not used yet) */
79 GMX_CPUID_FEATURE_X86_CLFSH, /* Supports CLFLUSH instruction */
80 GMX_CPUID_FEATURE_X86_CMOV, /* Conditional move insn support */
81 GMX_CPUID_FEATURE_X86_CX8, /* Supports CMPXCHG8B (8-byte compare-exchange) */
82 GMX_CPUID_FEATURE_X86_CX16, /* Supports CMPXCHG16B (16-byte compare-exchg) */
83 GMX_CPUID_FEATURE_X86_F16C, /* Supports 16-bit FP conversion instructions */
84 GMX_CPUID_FEATURE_X86_FMA, /* Fused-multiply add support (mainly for AVX) */
85 GMX_CPUID_FEATURE_X86_FMA4, /* 4-operand FMA, only on AMD for now */
86 GMX_CPUID_FEATURE_X86_HTT, /* Hyper-Threading supported */
87 GMX_CPUID_FEATURE_X86_LAHF_LM, /* LAHF/SAHF support in 64 bits */
88 GMX_CPUID_FEATURE_X86_MISALIGNSSE, /* Support for misaligned SSE data instructions */
89 GMX_CPUID_FEATURE_X86_MMX, /* MMX registers and instructions */
90 GMX_CPUID_FEATURE_X86_MSR, /* Supports Intel model-specific-registers */
91 GMX_CPUID_FEATURE_X86_NONSTOP_TSC, /* Invariant TSC (constant rate in ACPI states) */
92 GMX_CPUID_FEATURE_X86_PCID, /* Process context identifier support */
93 GMX_CPUID_FEATURE_X86_PCLMULDQ, /* Carry-less 64-bit multiplication supported */
94 GMX_CPUID_FEATURE_X86_PDCM, /* Perfmon and Debug Capability */
95 GMX_CPUID_FEATURE_X86_PDPE1GB, /* Support for 1GB pages */
96 GMX_CPUID_FEATURE_X86_POPCNT, /* Supports the POPCNT (population count) insn */
97 GMX_CPUID_FEATURE_X86_PSE, /* Supports 4MB-pages (page size extension) */
98 GMX_CPUID_FEATURE_X86_RDRND, /* RDRAND high-quality hardware random numbers */
99 GMX_CPUID_FEATURE_X86_RDTSCP, /* Serializing rdtscp instruction available */
100 GMX_CPUID_FEATURE_X86_SSE2, /* SSE 2 */
101 GMX_CPUID_FEATURE_X86_SSE3, /* SSE 3 */
102 GMX_CPUID_FEATURE_X86_SSE4A, /* SSE 4A */
103 GMX_CPUID_FEATURE_X86_SSE4_1, /* SSE 4.1 */
104 GMX_CPUID_FEATURE_X86_SSE4_2, /* SSE 4.2 */
105 GMX_CPUID_FEATURE_X86_SSSE3, /* Supplemental SSE3 */
106 GMX_CPUID_FEATURE_X86_TDT, /* TSC deadline timer */
107 GMX_CPUID_FEATURE_X86_X2APIC, /* Extended xAPIC Support */
108 GMX_CPUID_FEATURE_X86_XOP, /* AMD extended instructions, only AMD for now */
113 /* Currently supported acceleration instruction sets, intrinsics or other similar combinations
114 * in Gromacs. There is not always a 1-to-1 correspondence with feature flags; on some AMD
115 * hardware we prefer to use 128bit AVX instructions (although 256-bit ones could be executed),
116 * and we still haven't written the AVX2 kernels.
118 enum gmx_cpuid_acceleration
120 GMX_CPUID_ACCELERATION_CANNOTDETECT, /* Should only be used if something fails */
121 GMX_CPUID_ACCELERATION_NONE,
122 GMX_CPUID_ACCELERATION_X86_SSE2,
123 GMX_CPUID_ACCELERATION_X86_SSE4_1,
124 GMX_CPUID_ACCELERATION_X86_AVX_128_FMA,
125 GMX_CPUID_ACCELERATION_X86_AVX_256,
126 GMX_CPUID_NACCELERATIONS
129 /* Text strings corresponding to CPU vendors */
132 gmx_cpuid_vendor_string[GMX_CPUID_NVENDORS];
134 /* Text strings for CPU feature indices */
136 gmx_cpuid_feature_string[GMX_CPUID_NFEATURES];
138 /* Text strings for Gromacs acceleration/instruction sets */
140 gmx_cpuid_acceleration_string[GMX_CPUID_NACCELERATIONS];
143 /* Abstract data type with CPU detection information. Set by gmx_cpuid_init(). */
144 typedef struct gmx_cpuid *
148 /* Fill the data structure by using CPU detection instructions.
149 * Return 0 on success, 1 if something bad happened.
152 gmx_cpuid_init (gmx_cpuid_t * cpuid);
155 /* Return the vendor id as enumerated type. Use gmx_cpuid_vendor_string[]
156 * to get the corresponding text string.
159 enum gmx_cpuid_vendor
160 gmx_cpuid_vendor (gmx_cpuid_t cpuid);
163 /* Return a constant pointer to the processor brand string. */
165 gmx_cpuid_brand (gmx_cpuid_t cpuid);
168 /* Return processor family version. For a chip of version 1.2.3, this is 1 */
171 gmx_cpuid_family (gmx_cpuid_t cpuid);
173 /* Return processor model version, For a chip of version 1.2.3, this is 2. */
176 gmx_cpuid_model (gmx_cpuid_t cpuid);
178 /* Return processor stepping version, For a chip of version 1.2.3, this is 3. */
180 gmx_cpuid_stepping (gmx_cpuid_t cpuid);
183 /* Check whether a particular CPUID feature is set.
184 * Returns 0 if flag "feature" is not set, 1 if the flag is set. We cannot use
185 * gmx_bool here since this file must be possible to compile without simple.h.
189 gmx_cpuid_feature (gmx_cpuid_t cpuid,
190 enum gmx_cpuid_feature feature);
193 /* Enumerated values for x86 SMT enabled-status. Note that this does not refer
194 * to Hyper-Threading support (that is the flag GMX_CPUID_FEATURE_X86_HTT), but
195 * whether Hyper-Threading is _enabled_ and _used_ in bios right now.
197 enum gmx_cpuid_x86_smt
199 GMX_CPUID_X86_SMT_CANNOTDETECT,
200 GMX_CPUID_X86_SMT_DISABLED,
201 GMX_CPUID_X86_SMT_ENABLED
204 /* Returns the status of x86 SMT support. IMPORTANT: There are non-zero
205 * return values for this routine that still do not indicate supported and
206 * enabled smt/Hyper-Threading. You need to carefully check the return value
207 * against the enumerated type values to see what you are getting.
209 * Long-term, this functionality will move to a new hardware topology detection
210 * layer, but that will require a lot of new code and a working interface to the
211 * hwloc library. Surprisingly, there is no simple way to find out that
212 * Hyper-Threading is actually turned on without fully enumerating and checking
213 * all the cores, which we presently can only do on Linux. This means a couple
216 * 1) If you want to know whether your CPU _supports_ Hyper-Threading in the
217 * first place, check the GMX_CPUID_FEATURE_X86_HTT flag instead!
218 * 2) There are several scenarios where this routine will say that it cannot
219 * detect whether SMT is enabled and used right now.
220 * 3) If you need support on non-Linux x86, you have to write it :-)
221 * 4) Don't invest too much efforts, since this will be replaced with
222 * full hardware topology detection in the future.
223 * 5) Don't worry if the detection does not work. It is not a catastrophe, but
224 * but we get slightly better performance on x86 if we use Hyper-Threading
225 * cores in direct space, but not reciprocal space.
227 * Since this routine presently only supports Hyper-Threading we say X86_SMT
228 * in order not to give the impression we can detect any SMT. We haven't
229 * even tested the performance on other SMT implementations, so it is not
230 * obvious we shouldn't use SMT there.
233 enum gmx_cpuid_x86_smt
234 gmx_cpuid_x86_smt(gmx_cpuid_t cpuid);
238 /* Formats a text string (up to n characters) from the data structure.
239 * The output will have max 80 chars between newline characters.
242 gmx_cpuid_formatstring (gmx_cpuid_t cpuid,
247 /* Suggests a suitable gromacs acceleration based on the support in the
250 enum gmx_cpuid_acceleration
251 gmx_cpuid_acceleration_suggest (gmx_cpuid_t cpuid);
254 /* Check if this binary was compiled with the same acceleration as we
255 * would suggest for the current hardware. Always print stats to the log file
256 * if it is non-NULL, and print a warning in stdout if we don't have a match.
259 gmx_cpuid_acceleration_check (gmx_cpuid_t cpuid,
263 /* Release resources used by data structure. Note that the pointer to the
264 * CPU brand string will no longer be valid once this routine has been called.
267 gmx_cpuid_done (gmx_cpuid_t cpuid);
277 #endif /* GMX_CPUID_H_ */