2 # This file is part of the GROMACS molecular simulation package.
4 # Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
5 # Copyright (c) 2017,2018,2019,2020, by the GROMACS development team, led by
6 # Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 # and including many others, as listed in the AUTHORS file in the
8 # top-level source directory and at http://www.gromacs.org.
10 # GROMACS is free software; you can redistribute it and/or
11 # modify it under the terms of the GNU Lesser General Public License
12 # as published by the Free Software Foundation; either version 2.1
13 # of the License, or (at your option) any later version.
15 # GROMACS is distributed in the hope that it will be useful,
16 # but WITHOUT ANY WARRANTY; without even the implied warranty of
17 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 # Lesser General Public License for more details.
20 # You should have received a copy of the GNU Lesser General Public
21 # License along with GROMACS; if not, see
22 # http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 # Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 # If you want to redistribute modifications to GROMACS, please
26 # consider that scientific software is very special. Version
27 # control is crucial - bugs must be traceable. We will be happy to
28 # consider code for inclusion in the official distribution, but
29 # derived work must not be called official GROMACS. Details are found
30 # in the README & COPYING files - if they are missing, get the
31 # official version at http://www.gromacs.org.
33 # To help us fund GROMACS development, we humbly ask that you cite
34 # the research papers on the package. Check out http://www.gromacs.org.
39 # Issue a fatal error with an appropriate message, when the toolchain
40 # was not able to compile code for SIMD support.
43 # SIMD_STRING A string describing the kind of SIMD support that didn't work.
44 # ALTERNATIVE_SUGGESTION A string describing anything the user could try other than getting a new compiler.
45 # SUGGEST_BINUTILS_UPDATE True when there's information that the compiler was OK, but something else was not.
46 function(gmx_give_fatal_error_when_simd_support_not_found SIMD_STRING ALTERNATIVE_SUGGESTION SUGGEST_BINUTILS_UPDATE)
47 if(SUGGEST_BINUTILS_UPDATE)
48 set(_msg "Found a compiler flag for ${SIMD_STRING} support, but some other problem exists. Update your assembler and/or linker, e.g. in the binutils package of your distribution.")
50 set(_msg "Cannot find ${SIMD_STRING} compiler flag. Use a newer compiler, or ${ALTERNATIVE_SUGGESTION}.")
52 message(FATAL_ERROR ${_msg})
55 macro(gmx_manage_simd)
57 set(GMX_SIMD_ACCURACY_BITS_SINGLE 22 CACHE STRING "Target mantissa bits for SIMD single math")
59 # Note that we typically restrict double precision target accuracy to be twice that
60 # of single. This means we only need one more N-R iteration for 1/sqrt(x) and 1(x),
61 # and the first iteration can sometimes be done as a pair in single precision. This should
62 # be plenty enough for Molecular Dynamics applications. Many of our double precision math
63 # functions still achieve very close to full double precision, but we do not guarantee that
64 # they will be able to achieve higher accuracy if you set this beyond 44 bits. GROMACS will
65 # work - but some unit tests might fail.
67 set(GMX_SIMD_ACCURACY_BITS_DOUBLE 44 CACHE STRING "Target mantissa bits for SIMD double math")
68 mark_as_advanced(GMX_SIMD_ACCURACY_BITS_SINGLE)
69 mark_as_advanced(GMX_SIMD_ACCURACY_BITS_DOUBLE)
71 if(${GMX_SIMD_ACCURACY_BITS_SINGLE} GREATER 22)
72 message(STATUS "Note: Full mantissa accuracy (including least significant bit) requested for SIMD single math. Presently we cannot get the least significant bit correct since that would require different algorithms - reducing to 22 bits.")
73 set(GMX_SIMD_ACCURACY_BITS_SINGLE 22 CACHE STRING "Target mantissa bits for SIMD single math" FORCE)
76 if(${GMX_SIMD_ACCURACY_BITS_DOUBLE} GREATER 51)
77 message(STATUS "Note: Full mantissa accuracy (including least significant bit) requested for SIMD double math. Presently we cannot get the least significant bit correct since that would require different algorithms - reducing to 51 bits.")
78 set(GMX_SIMD_ACCURACY_BITS_DOUBLE 51 CACHE STRING "Target mantissa bits for SIMD double math" FORCE)
82 # Section to set (and test) compiler flags for SIMD.
84 # If the user chose the (default) automatic behaviour, then detection
85 # is run to suggest a SIMD choice suitable for the build
86 # host. Otherwise, the users's choice is always honoured. The compiler
87 # flags will be set based on that choice.
90 set(GMX_SIMD_ACTIVE ${GMX_SIMD})
91 if(GMX_SIMD STREQUAL "AUTO")
92 include(gmxDetectSimd)
93 gmx_detect_simd(GMX_SUGGESTED_SIMD)
94 set(GMX_SIMD_ACTIVE ${GMX_SUGGESTED_SIMD})
97 if(GMX_SIMD_ACTIVE STREQUAL "NONE")
98 # nothing to do configuration-wise
99 set(SIMD_STATUS_MESSAGE "SIMD instructions disabled")
100 elseif(GMX_SIMD_ACTIVE STREQUAL "SSE2")
102 gmx_find_simd_sse2_flags(SIMD_SSE2_C_SUPPORTED SIMD_SSE2_CXX_SUPPORTED
103 SIMD_SSE2_C_FLAGS SIMD_SSE2_CXX_FLAGS)
105 if(NOT SIMD_SSE2_C_SUPPORTED OR NOT SIMD_SSE2_CXX_SUPPORTED)
106 gmx_give_fatal_error_when_simd_support_not_found("SSE2" "disable SIMD support (slow)" "${SUGGEST_BINUTILS_UPDATE}")
109 # If multiple flags are neeed, make them into a list
110 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_SSE2_C_FLAGS})
111 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_SSE2_CXX_FLAGS})
112 set(GMX_SIMD_X86_${GMX_SIMD_ACTIVE} 1)
113 set(SIMD_STATUS_MESSAGE "Enabling SSE2 SIMD instructions using CXX flags: ${SIMD_SSE2_CXX_FLAGS}")
115 elseif(GMX_SIMD_ACTIVE STREQUAL "SSE4.1")
117 gmx_find_simd_sse4_1_flags(SIMD_SSE4_1_C_SUPPORTED SIMD_SSE4_1_CXX_SUPPORTED
118 SIMD_SSE4_1_C_FLAGS SIMD_SSE4_1_CXX_FLAGS)
120 if(NOT SIMD_SSE4_1_C_SUPPORTED OR NOT SIMD_SSE4_1_CXX_SUPPORTED)
121 gmx_give_fatal_error_when_simd_support_not_found("SSE4.1" "choose SSE2 SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
124 # If multiple flags are neeed, make them into a list
125 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_SSE4_1_C_FLAGS})
126 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_SSE4_1_CXX_FLAGS})
127 set(GMX_SIMD_X86_SSE4_1 1)
128 set(SIMD_STATUS_MESSAGE "Enabling SSE4.1 SIMD instructions using CXX flags: ${SIMD_SSE4_1_CXX_FLAGS}")
130 elseif(GMX_SIMD_ACTIVE STREQUAL "AVX_128_FMA")
132 gmx_find_simd_avx_128_fma_flags(SIMD_AVX_128_FMA_C_SUPPORTED SIMD_AVX_128_FMA_CXX_SUPPORTED
133 SIMD_AVX_128_FMA_C_FLAGS SIMD_AVX_128_FMA_CXX_FLAGS)
135 if(NOT SIMD_AVX_128_FMA_C_SUPPORTED OR NOT SIMD_AVX_128_FMA_CXX_SUPPORTED)
136 gmx_give_fatal_error_when_simd_support_not_found("128-bit AVX with FMA support" "choose SSE4.1 SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
139 # If multiple flags are neeed, make them into a list
140 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_AVX_128_FMA_C_FLAGS})
141 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_AVX_128_FMA_CXX_FLAGS})
142 set(GMX_SIMD_X86_${GMX_SIMD_ACTIVE} 1)
143 set(SIMD_STATUS_MESSAGE "Enabling 128-bit AMD FMA SIMD instructions using CXX flags: ${SIMD_AVX_128_FMA_CXX_FLAGS}")
145 elseif(GMX_SIMD_ACTIVE STREQUAL "AVX_256")
147 gmx_find_simd_avx_flags(SIMD_AVX_C_SUPPORTED SIMD_AVX_CXX_SUPPORTED
148 SIMD_AVX_C_FLAGS SIMD_AVX_CXX_FLAGS)
150 if(NOT SIMD_AVX_C_SUPPORTED OR NOT SIMD_AVX_CXX_SUPPORTED)
151 gmx_give_fatal_error_when_simd_support_not_found("AVX" "choose SSE4.1 SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
154 # If multiple flags are neeed, make them into a list
155 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_AVX_C_FLAGS})
156 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_AVX_CXX_FLAGS})
157 set(GMX_SIMD_X86_${GMX_SIMD_ACTIVE} 1)
158 set(SIMD_STATUS_MESSAGE "Enabling 256-bit AVX SIMD instructions using CXX flags: ${SIMD_AVX_CXX_FLAGS}")
160 elseif(GMX_SIMD_ACTIVE MATCHES "AVX2_")
162 gmx_find_simd_avx2_flags(SIMD_AVX2_C_SUPPORTED SIMD_AVX2_CXX_SUPPORTED
163 SIMD_AVX2_C_FLAGS SIMD_AVX2_CXX_FLAGS)
165 if(NOT SIMD_AVX2_C_SUPPORTED OR NOT SIMD_AVX2_CXX_SUPPORTED)
166 gmx_give_fatal_error_when_simd_support_not_found("AVX2" "choose AVX SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
169 # If multiple flags are neeed, make them into a list
170 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_AVX2_C_FLAGS})
171 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_AVX2_CXX_FLAGS})
172 set(GMX_SIMD_X86_${GMX_SIMD_ACTIVE} 1)
174 if(GMX_SIMD_ACTIVE STREQUAL "AVX2_128")
175 set(SIMD_STATUS_MESSAGE "Enabling 128-bit AVX2 SIMD instructions using CXX flags: ${SIMD_AVX2_CXX_FLAGS}")
177 set(SIMD_STATUS_MESSAGE "Enabling 256-bit AVX2 SIMD instructions using CXX flags: ${SIMD_AVX2_CXX_FLAGS}")
180 elseif(GMX_SIMD_ACTIVE STREQUAL "MIC")
182 # No flags needed. Not testing.
183 set(GMX_SIMD_X86_MIC 1)
184 set(SIMD_STATUS_MESSAGE "Enabling MIC (Xeon Phi) SIMD instructions without special flags.")
186 elseif(GMX_SIMD_ACTIVE STREQUAL "AVX_512")
188 gmx_find_simd_avx_512_flags(SIMD_AVX_512_C_SUPPORTED SIMD_AVX_512_CXX_SUPPORTED
189 SIMD_AVX_512_C_FLAGS SIMD_AVX_512_CXX_FLAGS)
191 if(NOT SIMD_AVX_512_C_SUPPORTED OR NOT SIMD_AVX_512_CXX_SUPPORTED)
192 gmx_give_fatal_error_when_simd_support_not_found("AVX 512F" "choose a lower level of SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
195 # If multiple flags are neeed, make them into a list
196 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_AVX_512_C_FLAGS})
197 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_AVX_512_CXX_FLAGS})
198 set(GMX_SIMD_X86_${GMX_SIMD_ACTIVE} 1)
199 set(SIMD_STATUS_MESSAGE "Enabling 512-bit AVX-512 SIMD instructions using CXX flags: ${SIMD_AVX_512_CXX_FLAGS}")
201 elseif(GMX_SIMD_ACTIVE STREQUAL "AVX_512_KNL")
203 gmx_find_simd_avx_512_knl_flags(SIMD_AVX_512_KNL_C_SUPPORTED SIMD_AVX_512_KNL_CXX_SUPPORTED
204 SIMD_AVX_512_KNL_C_FLAGS SIMD_AVX_512_KNL_CXX_FLAGS)
206 if(NOT SIMD_AVX_512_KNL_C_SUPPORTED OR NOT SIMD_AVX_512_KNL_CXX_SUPPORTED)
207 gmx_give_fatal_error_when_simd_support_not_found("AVX 512ER" "choose a lower level of SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
210 # If multiple flags are neeed, make them into a list
211 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_AVX_512_KNL_C_FLAGS})
212 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_AVX_512_KNL_CXX_FLAGS})
213 set(GMX_SIMD_X86_${GMX_SIMD_ACTIVE} 1)
214 set(SIMD_STATUS_MESSAGE "Enabling 512-bit AVX-512-KNL SIMD instructions using CXX flags: ${SIMD_AVX_512_KNL_CXX_FLAGS}")
216 elseif(GMX_SIMD_ACTIVE STREQUAL "ARM_NEON")
219 message(FATAL_ERROR "ARM_NEON SIMD support is not available for a double precision build because the architecture lacks double-precision support")
222 gmx_find_simd_arm_neon_flags(SIMD_ARM_NEON_C_SUPPORTED SIMD_ARM_NEON_CXX_SUPPORTED
223 SIMD_ARM_NEON_C_FLAGS SIMD_ARM_NEON_CXX_FLAGS)
225 if(NOT SIMD_ARM_NEON_C_SUPPORTED OR NOT SIMD_ARM_NEON_CXX_SUPPORTED)
226 gmx_give_fatal_error_when_simd_support_not_found("ARM NEON" "disable SIMD support (slower)" "${SUGGEST_BINUTILS_UPDATE}")
229 # If multiple flags are neeed, make them into a list
230 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_ARM_NEON_C_FLAGS})
231 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_ARM_NEON_CXX_FLAGS})
232 set(GMX_SIMD_${GMX_SIMD_ACTIVE} 1)
233 set(SIMD_STATUS_MESSAGE "Enabling 32-bit ARM NEON SIMD instructions using CXX flags: ${SIMD_ARM_NEON_CXX_FLAGS}")
235 elseif(GMX_SIMD_ACTIVE STREQUAL "ARM_NEON_ASIMD")
237 gmx_find_simd_arm_neon_asimd_flags(SIMD_ARM_NEON_ASIMD_C_SUPPORTED SIMD_ARM_NEON_ASIMD_CXX_SUPPORTED
238 SIMD_ARM_NEON_ASIMD_C_FLAGS SIMD_ARM_NEON_ASIMD_CXX_FLAGS)
240 if(NOT SIMD_ARM_NEON_ASIMD_C_SUPPORTED OR NOT SIMD_ARM_NEON_ASIMD_CXX_SUPPORTED)
241 gmx_give_fatal_error_when_simd_support_not_found("ARM (AArch64) NEON Advanced SIMD" "particularly gcc version 4.9 or later, or disable SIMD support (slower)" "${SUGGEST_BINUTILS_UPDATE}")
244 # If multiple flags are neeed, make them into a list
245 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_ARM_NEON_ASIMD_C_FLAGS})
246 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_ARM_NEON_ASIMD_CXX_FLAGS})
247 set(GMX_SIMD_${GMX_SIMD_ACTIVE} 1)
248 set(SIMD_STATUS_MESSAGE "Enabling ARM (AArch64) NEON Advanced SIMD instructions using CXX flags: ${SIMD_ARM_NEON_ASIMD_CXX_FLAGS}")
250 elseif(GMX_SIMD_ACTIVE STREQUAL "ARM_SVE")
252 # Note that GMX_RELAXED_DOUBLE_PRECISION is enabled by default in the top-level CMakeLists.txt
253 gmx_option_multichoice(
254 GMX_SIMD_ARM_SVE_LENGTH
255 "SVE vector length in bits"
257 auto 128 256 512 1024 2048)
259 if (GMX_SIMD_ARM_SVE_LENGTH STREQUAL "AUTO")
260 if (NOT GMX_SIMD_ARM_SVE_DETECTED_LENGTH)
261 # Read the vector length and cache it
262 if(NOT EXISTS "/proc/sys/abi/sve_default_vector_length")
263 message(FATAL_ERROR "cannot automatically determine the SVE vector length, please explicitly set it via -DGMX_SIMD_ARM_SVE_LENGTH=<bits>")
265 file(READ "/proc/sys/abi/sve_default_vector_length" GMX_SIMD_ARM_SVE_DETECTED_LENGTH_IN_BYTES)
266 message(STATUS "Detected SVE vector length in bytes : ${GMX_SIMD_ARM_SVE_DETECTED_LENGTH_IN_BYTES}")
267 math(EXPR GMX_SIMD_ARM_SVE_DETECTED_LENGTH "${GMX_SIMD_ARM_SVE_DETECTED_LENGTH_IN_BYTES} * 8")
268 set(GMX_SIMD_ARM_SVE_DETECTED_LENGTH ${GMX_SIMD_ARM_SVE_DETECTED_LENGTH} CACHE STRING "Detected length in bits for SVE vectors")
269 message(STATUS "Detected SVE vector length of ${GMX_SIMD_ARM_SVE_DETECTED_LENGTH} bits")
271 set(GMX_SIMD_ARM_SVE_LENGTH_VALUE ${GMX_SIMD_ARM_SVE_DETECTED_LENGTH})
273 set(GMX_SIMD_ARM_SVE_LENGTH_VALUE ${GMX_SIMD_ARM_SVE_LENGTH})
276 gmx_find_simd_arm_sve_flags(SIMD_ARM_SVE_C_SUPPORTED SIMD_ARM_SVE_CXX_SUPPORTED
277 SIMD_ARM_SVE_C_FLAGS SIMD_ARM_SVE_CXX_FLAGS)
279 if(NOT SIMD_ARM_SVE_C_SUPPORTED OR NOT SIMD_ARM_SVE_CXX_SUPPORTED)
280 gmx_give_fatal_error_when_simd_support_not_found("ARM (AArch64) SVE SIMD" "particularly gcc version 10.1 or later, or disable SIMD support (slower)" "${SUGGEST_BINUTILS_UPDATE}")
283 # If multiple flags are neeed, make them into a list
284 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_ARM_SVE_C_FLAGS})
285 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_ARM_SVE_CXX_FLAGS})
286 set(GMX_SIMD_${GMX_SIMD_ACTIVE} 1)
287 set(SIMD_STATUS_MESSAGE "Enabling ARM (AArch64) SVE Advanced SIMD instructions using CXX flags: ${SIMD_ARM_SVE_CXX_FLAGS}")
289 elseif(GMX_SIMD_ACTIVE STREQUAL "IBM_VMX")
291 gmx_find_simd_ibm_vmx_flags(SIMD_IBM_VMX_C_SUPPORTED SIMD_IBM_VMX_CXX_SUPPORTED
292 SIMD_IBM_VMX_C_FLAGS SIMD_IBM_VMX_CXX_FLAGS)
294 if(NOT SIMD_IBM_VMX_C_SUPPORTED OR NOT SIMD_IBM_VMX_CXX_SUPPORTED)
295 gmx_give_fatal_error_when_simd_support_not_found("IBM VMX" "disable SIMD support (slower)" "${SUGGEST_BINUTILS_UPDATE}")
298 # If multiple flags are neeed, make them into a list
299 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_IBM_VMX_C_FLAGS})
300 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_IBM_VMX_CXX_FLAGS})
301 set(GMX_SIMD_${GMX_SIMD_ACTIVE} 1)
302 set(SIMD_STATUS_MESSAGE "Enabling IBM VMX SIMD instructions using CXX flags: ${SIMD_IBM_VMX_CXX_FLAGS}")
304 elseif(GMX_SIMD_ACTIVE STREQUAL "IBM_VSX")
306 # IBM_VSX and gcc > 9 do not work together, so we need to prevent people from
307 # choosing a combination that might fail. Issue #3380.
308 if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU" AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 10)
309 message(FATAL_ERROR "IBM_VSX does not work together with gcc > 9. Disable SIMD support (slower), or use an older version of the GNU compiler")
313 gmx_find_simd_ibm_vsx_flags(SIMD_IBM_VSX_C_SUPPORTED SIMD_IBM_VSX_CXX_SUPPORTED
314 SIMD_IBM_VSX_C_FLAGS SIMD_IBM_VSX_CXX_FLAGS)
316 # Usually we check also for the C compiler here, but a C compiler
317 # is not required for SIMD support on this platform. cmake through
318 # at least version 3.7 cannot pass this check with the C compiler
319 # in the latest xlc 13.1.5, but the C++ compiler has different
320 # behaviour and is OK. See Issue #2102.
321 if(NOT SIMD_IBM_VSX_CXX_SUPPORTED)
322 gmx_give_fatal_error_when_simd_support_not_found("IBM VSX" "disable SIMD support (slower)" "${SUGGEST_BINUTILS_UPDATE}")
325 # If multiple flags are neeed, make them into a list
326 string(REPLACE " " ";" SIMD_C_FLAGS ${SIMD_IBM_VSX_C_FLAGS})
327 string(REPLACE " " ";" SIMD_CXX_FLAGS ${SIMD_IBM_VSX_CXX_FLAGS})
328 set(GMX_SIMD_${GMX_SIMD_ACTIVE} 1)
329 set(SIMD_STATUS_MESSAGE "Enabling IBM VSX SIMD instructions using CXX flags: ${SIMD_IBM_VSX_CXX_FLAGS}")
331 elseif(GMX_SIMD_ACTIVE STREQUAL "SPARC64_HPC_ACE")
333 # Note that GMX_RELAXED_DOUBLE_PRECISION is enabled by default in the top-level CMakeLists.txt
335 set(GMX_SIMD_${GMX_SIMD_ACTIVE} 1)
336 set(SIMD_STATUS_MESSAGE "Enabling Sparc64 HPC-ACE SIMD instructions without special flags.")
338 elseif(GMX_SIMD_ACTIVE STREQUAL "REFERENCE")
340 # NB: This file handles settings for the SIMD module, so in the interest
341 # of proper modularization, please do NOT put any verlet kernel settings in this file.
343 if(GMX_SIMD_REF_FLOAT_WIDTH)
344 add_definitions(-DGMX_SIMD_REF_FLOAT_WIDTH=${GMX_SIMD_REF_FLOAT_WIDTH})
346 if(GMX_SIMD_REF_DOUBLE_WIDTH)
347 add_definitions(-DGMX_SIMD_REF_DOUBLE_WIDTH=${GMX_SIMD_REF_DOUBLE_WIDTH})
350 set(GMX_SIMD_${GMX_SIMD_ACTIVE} 1)
351 set(SIMD_STATUS_MESSAGE "Enabling reference (emulated) SIMD instructions without special flags.")
354 gmx_invalid_option_value(GMX_SIMD_ACTIVE)
358 gmx_check_if_changed(SIMD_CHANGED GMX_SIMD_ACTIVE)
359 if (SIMD_CHANGED AND DEFINED SIMD_STATUS_MESSAGE)
360 message(STATUS "${SIMD_STATUS_MESSAGE}")
363 # While AVX-512 is a more recent SIMD ISA than AVX2, some Intel CPUs only have
364 # a single AVX-512 FMA unit, but two AVX2 FMA units, and then it is better to
365 # use AVX2. The only way to test this is to execute a small timing loop.
366 # To be able to recommend the user whether s/he should try AVX-512 instead of
367 # AVX2, we need to compile a single file with AVX512 flags. We do this
368 # automatically, but this option provides a way to turn it off in case it
369 # breaks something. The actual test source file is built if
370 # SIMD_AVX_512_CXX_SUPPORTED is set, so it will always be included if we have
372 set(GMX_ENABLE_AVX512_TESTS ON CACHE BOOL "Compile AVX512 code to test FMA units, even when not using AVX512 SIMD")
373 mark_as_advanced(GMX_ENABLE_AVX512_TESTS)
375 if(GMX_ENABLE_AVX512_TESTS AND
376 (GMX_SIMD_ACTIVE STREQUAL "AVX_256" OR GMX_SIMD_ACTIVE STREQUAL "AVX2_256" OR GMX_SIMD_ACTIVE STREQUAL "AVX2_128"))
377 if(NOT DEFINED SIMD_AVX_512_CXX_SUPPORTED)
378 message(STATUS "Detecting flags to enable runtime detection of AVX-512 units on newer CPUs")
379 set(SIMD_AVX_512_REPORT_STATUS 1)
381 gmx_find_simd_avx_512_flags(SIMD_AVX_512_C_SUPPORTED SIMD_AVX_512_CXX_SUPPORTED
382 SIMD_AVX_512_C_FLAGS SIMD_AVX_512_CXX_FLAGS)
383 if(SIMD_AVX_512_REPORT_STATUS)
384 if(SIMD_AVX_512_CXX_SUPPORTED)
385 message(STATUS "Detecting flags to enable runtime detection of AVX-512 units on newer CPUs - ${SIMD_AVX_512_CXX_FLAGS}")
387 message(STATUS "Detecting flags to enable runtime detection of AVX-512 units on newer CPUs - not supported")
390 # Since we might be overriding AVX2 architecture flags with the AVX512 flags for the
391 # files where it is used, we also check for a flag not to warn about the first (unused) arch.
392 # To avoid spamming the user with lots of gromacs tests we just call the CMake flag test directly.
394 foreach(_testflag "-Wno-unused-command-line-argument" "-wd10121")
395 string(REGEX REPLACE "[^a-zA-Z0-9]+" "_" FLAG_ACCEPTED_VARIABLE "${_testflag}_FLAG_ACCEPTED")
396 check_cxx_compiler_flag("${_testflag}" ${FLAG_ACCEPTED_VARIABLE})
397 if(${FLAG_ACCEPTED_VARIABLE})
398 set(CXX_NO_UNUSED_OPTION_WARNING_FLAGS "${_testflag}")
401 endforeach(_testflag)
405 # By default, 32-bit windows cannot pass SIMD (SSE/AVX) arguments in registers,
406 # and even on 64-bit (all platforms) it is only used for a handful of arguments.
407 # The __vectorcall (MSVC, from MSVC2013) or __regcall (ICC) calling conventions
408 # enable this, which is critical to enable 32-bit SIMD and improves performance
410 # Check if the compiler supports one of these, and in that case set gmx_simdcall
411 # to that string. If we do not have any such calling convention modifier, set it
412 # to an empty string.
414 # Update 2015-11-04: As of version 3.6, clang has added support for __vectorcall
415 # (also on Linux). This appears to be buggy for the reference SIMD
416 # implementation when using the Debug build (when functions are not inlined)
417 # while it seems works fine for the actual SIMD implementations. This is likely
418 # because the reference build ends up passing lots of structures with arrays
419 # rather than actual vector data. For now we disable __vectorcall with clang
420 # when using the reference build.
422 # xlc 13.1.5 does not seem recognize any attribute, and warns about invalid ones
423 # so we avoid searching for any.
425 if(NOT DEFINED GMX_SIMD_CALLING_CONVENTION)
426 if(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND GMX_SIMD_ACTIVE STREQUAL "REFERENCE")
427 set(CALLCONV_LIST __regcall " ")
428 elseif(CMAKE_CXX_COMPILER_ID MATCHES "XL")
429 set(CALLCONV_LIST " ")
431 set(CALLCONV_LIST __vectorcall __regcall " ")
433 foreach(callconv ${CALLCONV_LIST})
434 set(callconv_compile_var "_callconv_${callconv}")
435 # Some compilers warn about targets for which attributes are
436 # ignored (e.g. clang on ARM), and in such cases we want this
437 # check to lead to using no attribute in subsequent GROMACS
438 # compilation, to avoid issuing the warning for lots of files.
439 check_cxx_source_compiles("
440 #pragma GCC diagnostic error \"-Wignored-attributes\"
441 int ${callconv} f(int i) {return i;} int main(void) {return f(0);}
442 " ${callconv_compile_var})
443 if(${callconv_compile_var})
444 set(GMX_SIMD_CALLING_CONVENTION_VALUE "${callconv}" CACHE INTERNAL "Calling convention for SIMD routines" FORCE)
448 # If the build is not using SIMD, then we should not manage the
449 # calling convention. Doing so seems to confuse
450 # clang-static-analyzer in at least version 6.0.
451 if(GMX_SIMD_ACTIVE STREQUAL "NONE")
452 set(GMX_SIMD_CALLING_CONVENTION " ")
454 set(GMX_SIMD_CALLING_CONVENTION ${GMX_SIMD_CALLING_CONVENTION_VALUE})
458 if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
459 # GCC bug 49001, 54412 on Windows (just warn, since it might be fixed in later versions)
460 if((CMAKE_CXX_COMPILER_VERSION VERSION_LESS "4.9.0" OR CMAKE_SIZEOF_VOID_P EQUAL 8)
461 AND (WIN32 OR CYGWIN)
462 AND (GMX_SIMD_ACTIVE MATCHES "AVX") AND NOT (GMX_SIMD_ACTIVE STREQUAL "AVX_128_FMA"))
463 message(WARNING "GCC on Windows (GCC older than 4.9 in 32-bit mode, or any version in 64-bit mode) with 256-bit AVX will probably crash. You might want to choose a different GMX_SIMD or a different compiler.")
467 string(TOUPPER "${CMAKE_BUILD_TYPE}" _cmake_build_type)
468 if (_cmake_build_type STREQUAL "TSAN" AND NOT (GMX_SIMD_ACTIVE STREQUAL "NONE" OR GMX_SIMD_ACTIVE STREQUAL "REFERENCE" OR GMX_SIMD_ACTIVE MATCHES "AVX_512" OR GMX_SIMD_ACTIVE STREQUAL AVX2_256))
469 message(WARNING "TSAN is only tested with SIMD None, Reference, AVX2_256, and AVX_512. It is known to detect (harmless) memory races with SSE and AVX.")