2 # This file is part of the GROMACS molecular simulation package.
4 # Copyright (c) 2012,2013,2014,2015,2016,2017, by the GROMACS development team, led by
5 # Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 # and including many others, as listed in the AUTHORS file in the
7 # top-level source 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.
35 # include avx test source, used if the AVX flags are set below
36 include(gmxTestAVXMaskload)
37 include(gmxFindFlagsForSource)
39 # Macro that manages setting the respective C and C++ toolchain
40 # variables so that subsequent tests for SIMD support can work.
41 macro(prepare_x86_toolchain TOOLCHAIN_C_FLAGS_VARIABLE TOOLCHAIN_CXX_FLAGS_VARIABLE)
42 # On OS X, we often want to use gcc instead of clang, since gcc
43 # supports OpenMP (until clang 3.8, or so, plus whenever Apple
44 # support it in their version). However, by default gcc uses the
45 # external system assembler, which does not support AVX, so we
46 # need to tell the linker to use the clang compilers assembler
47 # instead - and this has to happen before we detect AVX flags.
48 if(APPLE AND CMAKE_C_COMPILER_ID STREQUAL "GNU")
49 gmx_test_cflag(GNU_C_USE_CLANG_AS "-Wa,-q" ${TOOLCHAIN_C_FLAGS_VARIABLE})
51 if(APPLE AND CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
52 gmx_test_cxxflag(GNU_CXX_USE_CLANG_AS "-Wa,-q" ${TOOLCHAIN_CXX_FLAGS_VARIABLE})
56 # Macro that manages setting the respective C and C++ toolchain
57 # variables so that subsequent tests for SIMD support can work.
58 macro(prepare_power_vsx_toolchain TOOLCHAIN_C_FLAGS_VARIABLE TOOLCHAIN_CXX_FLAGS_VARIABLE)
59 if(${CMAKE_CXX_COMPILER_ID} MATCHES "GNU" OR ${CMAKE_C_COMPILER_ID} MATCHES "GNU")
60 # VSX uses the same function API as Altivec/VMX, so make sure we tune for the current CPU and not VMX.
61 # By putting these flags here rather than in the general compiler flags file we can safely assume
62 # that we are at least on Power7 since that is when VSX appeared.
63 if(BUILD_CPU_BRAND MATCHES "POWER7")
64 gmx_test_cflag(GNU_C_VSX_POWER7 "-mcpu=power7 -mtune=power7" ${TOOLCHAIN_C_FLAGS_VARIABLE})
65 gmx_test_cflag(GNU_CXX_VSX_POWER7 "-mcpu=power7 -mtune=power7" ${TOOLCHAIN_CXX_FLAGS_VARIABLE})
67 # Enable power8 vector extensions on all platforms except old Power7.
68 gmx_test_cflag(GNU_C_VSX_POWER8 "-mcpu=power8 -mpower8-vector -mpower8-fusion -mdirect-move" ${TOOLCHAIN_C_FLAGS_VARIABLE})
69 gmx_test_cflag(GNU_CXX_VSX_POWER8 "-mcpu=power8 -mpower8-vector -mpower8-fusion -mdirect-move" ${TOOLCHAIN_CXX_FLAGS_VARIABLE})
71 # Altivec was originally single-only, and it took a while for compilers
72 # to support the double-precision features in VSX.
73 if(GMX_DOUBLE AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS "4.9")
74 message(FATAL_ERROR "Using VSX SIMD in double precision with GCC requires GCC-4.9 or later.")
77 if(${CMAKE_CXX_COMPILER_ID} MATCHES "XL" OR ${CMAKE_C_COMPILER_ID} MATCHES "XL")
78 if(CMAKE_CXX_COMPILER_VERSION VERSION_LESS "13.1.5" OR CMAKE_C_COMPILER_VERSION VERSION_LESS "13.1.5")
79 message(FATAL_ERROR "Using VSX SIMD requires XL compiler version 13.1.5 or later.")
84 # Issue a fatal error with an appropriate message, when the toolchain
85 # was not able to compile code for SIMD support.
88 # SIMD_STRING A string describing the kind of SIMD support that didn't work.
89 # ALTERNATIVE_SUGGESTION A string describing anything the user could try other than getting a new compiler.
90 # SUGGEST_BINUTILS_UPDATE True when there's information that the compiler was OK, but something else was not.
91 function(gmx_give_fatal_error_when_simd_support_not_found SIMD_STRING ALTERNATIVE_SUGGESTION SUGGEST_BINUTILS_UPDATE)
92 if(SUGGEST_BINUTILS_UPDATE)
93 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.")
95 set(_msg "Cannot find ${SIMD_STRING} compiler flag. Use a newer compiler, or ${ALTERNATIVE_SUGGESTION}.")
97 message(FATAL_ERROR ${_msg})
100 macro(gmx_manage_simd)
102 set(GMX_SIMD_ACCURACY_BITS_SINGLE 22 CACHE STRING "Target mantissa bits for SIMD single math")
104 # Note that we typically restrict double precision target accuracy to be twice that
105 # of single. This means we only need one more N-R iteration for 1/sqrt(x) and 1(x),
106 # and the first iteration can sometimes be done as a pair in single precision. This should
107 # be plenty enough for Molecular Dynamics applications. Many of our double precision math
108 # functions still achieve very close to full double precision, but we do not guarantee that
109 # they will be able to achieve higher accuracy if you set this beyond 44 bits. GROMACS will
110 # work - but some unit tests might fail.
112 set(GMX_SIMD_ACCURACY_BITS_DOUBLE 44 CACHE STRING "Target mantissa bits for SIMD double math")
113 mark_as_advanced(GMX_SIMD_ACCURACY_BITS_SINGLE)
114 mark_as_advanced(GMX_SIMD_ACCURACY_BITS_DOUBLE)
116 if(${GMX_SIMD_ACCURACY_BITS_SINGLE} GREATER 22)
117 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.")
118 set(GMX_SIMD_ACCURACY_BITS_SINGLE 22 CACHE STRING "Target mantissa bits for SIMD single math" FORCE)
121 if(${GMX_SIMD_ACCURACY_BITS_DOUBLE} GREATER 51)
122 message(STATUS "Note: Full mantissa accuracy (including least significant bit) requested for SIMD double math. Presently we cannot get the least significant bit correct since that would require different algorithms - reducing to 51 bits.")
123 set(GMX_SIMD_ACCURACY_BITS_DOUBLE 51 CACHE STRING "Target mantissa bits for SIMD double math" FORCE)
127 # Section to set (and test) compiler flags for SIMD.
129 # The flags will be set based on the GMX_SIMD choice provided by the user.
130 # Automatic detection of the architecture on the build host is done prior to
131 # calling this macro.
134 if(GMX_SIMD STREQUAL "NONE")
135 # nothing to do configuration-wise
136 set(SIMD_STATUS_MESSAGE "SIMD instructions disabled")
137 elseif(GMX_SIMD STREQUAL "SSE2")
140 "#include<xmmintrin.h>
141 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_rsqrt_ps(x);return _mm_movemask_ps(x);}"
142 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
143 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
144 "-msse2" "/arch:SSE2" "-hgnu")
146 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
147 gmx_give_fatal_error_when_simd_support_not_found("SSE2" "disable SIMD support (slow)" "${SUGGEST_BINUTILS_UPDATE}")
150 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
151 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
152 set(GMX_SIMD_X86_${GMX_SIMD} 1)
153 set(SIMD_STATUS_MESSAGE "Enabling SSE2 SIMD instructions")
155 elseif(GMX_SIMD STREQUAL "SSE4.1")
157 # Note: MSVC enables SSE4.1 with the SSE2 flag, so we include that in testing.
159 "#include<smmintrin.h>
160 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_dp_ps(x,x,0x77);return _mm_movemask_ps(x);}"
161 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
162 SIMD_SSE_4_1_C_FLAGS SIMD_SSE_4_1_CXX_FLAGS
163 "-msse4.1" "/arch:SSE4.1" "/arch:SSE2" "-hgnu")
165 if(NOT SIMD_SSE_4_1_C_FLAGS OR NOT SIMD_SSE_4_1_CXX_FLAGS)
166 gmx_give_fatal_error_when_simd_support_not_found("SSE4.1" "choose SSE2 SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
169 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
170 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
171 set(GMX_SIMD_X86_SSE4_1 1)
172 set(SIMD_STATUS_MESSAGE "Enabling SSE4.1 SIMD instructions")
174 elseif(GMX_SIMD STREQUAL "AVX_128_FMA")
176 prepare_x86_toolchain(TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS)
178 # We don't have the full compiler version string yet (BUILD_C_COMPILER),
179 # so we can't distinguish vanilla from Apple clang versions, but catering for a few rare AMD
180 # hackintoshes is not worth the effort.
181 if (APPLE AND (CMAKE_C_COMPILER_ID STREQUAL "Clang" OR
182 CMAKE_CXX_COMPILER_ID STREQUAL "Clang"))
183 message(WARNING "Due to a known compiler bug, Clang up to version 3.2 (and Apple Clang up to version 4.1) produces incorrect code with AVX_128_FMA SIMD. As we cannot work around this bug on OS X, you will have to select a different compiler or SIMD instruction set.")
186 # clang <=3.2 contains a bug that causes incorrect code to be generated for the
187 # vfmaddps instruction and therefore the bug is triggered with AVX_128_FMA.
188 # (see: http://llvm.org/bugs/show_bug.cgi?id=15040).
189 # We can work around this by not using the integrated assembler (except on OS X
190 # which has an outdated assembler that does not support AVX instructions).
191 if (CMAKE_C_COMPILER_ID MATCHES "Clang" AND CMAKE_C_COMPILER_VERSION VERSION_LESS "3.3")
192 # we assume that we have an external assembler that supports AVX
193 message(STATUS "Clang ${CMAKE_C_COMPILER_VERSION} detected, enabling FMA bug workaround")
194 set(TOOLCHAIN_C_FLAGS "${TOOLCHAIN_C_FLAGS} -no-integrated-as")
196 if (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS "3.3")
197 # we assume that we have an external assembler that supports AVX
198 message(STATUS "Clang ${CMAKE_CXX_COMPILER_VERSION} detected, enabling FMA bug workaround")
199 set(TOOLCHAIN_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS} -no-integrated-as")
202 # AVX128/FMA on AMD is a bit complicated. We need to do detection in three stages:
203 # 1) Find the flags required for generic AVX support
204 # 2) Find the flags necessary to enable fused-multiply add support
205 # 3) Optional: Find a flag to enable the AMD XOP instructions
207 ### STAGE 1: Find the generic AVX flag
209 "#include<immintrin.h>
210 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_permute_ps(x,1);return 0;}"
211 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
212 SIMD_GENERIC_AVX_C_FLAGS SIMD_GENERIC_AVX_CXX_FLAGS
213 "-mavx" "/arch:AVX" "-hgnu")
215 ### STAGE 2: Find the fused-multiply add flag.
216 # GCC requires x86intrin.h for FMA support. MSVC 2010 requires intrin.h for FMA support.
217 check_include_file(x86intrin.h HAVE_X86INTRIN_H ${SIMD_C_FLAGS})
218 check_include_file(intrin.h HAVE_INTRIN_H ${SIMD_C_FLAGS})
220 set(INCLUDE_X86INTRIN_H "#include <x86intrin.h>")
223 set(INCLUDE_INTRIN_H "#include <xintrin.h>")
227 "#include<immintrin.h>
228 ${INCLUDE_X86INTRIN_H}
230 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_macc_ps(x,x,x);return _mm_movemask_ps(x);}"
231 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
232 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
235 # We only need to check the last (FMA) test; that will always fail if the generic AVX test failed
236 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
237 gmx_give_fatal_error_when_simd_support_not_found("128-bit AVX with FMA support" "choose SSE4.1 SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
240 ### STAGE 3: Optional: Find the XOP instruction flag (No point in yelling if this does not work)
242 "#include<immintrin.h>
243 ${INCLUDE_X86INTRIN_H}
245 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_frcz_ps(x);return _mm_movemask_ps(x);}"
246 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
247 SIMD_AVX_128_XOP_C_FLAGS SIMD_AVX_128_XOP_CXX_FLAGS
250 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
251 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
252 set(GMX_SIMD_X86_${GMX_SIMD} 1)
253 set(SIMD_STATUS_MESSAGE "Enabling 128-bit AVX SIMD GROMACS SIMD (with fused-multiply add)")
255 gmx_test_avx_gcc_maskload_bug(GMX_SIMD_X86_AVX_GCC_MASKLOAD_BUG "${SIMD_C_FLAGS}")
257 elseif(GMX_SIMD STREQUAL "AVX_256")
259 prepare_x86_toolchain(TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS)
262 "#include<immintrin.h>
263 int main(){__m256 x=_mm256_set1_ps(0.5);x=_mm256_add_ps(x,x);return _mm256_movemask_ps(x);}"
264 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
265 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
266 "-mavx" "/arch:AVX" "-hgnu")
268 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
269 gmx_give_fatal_error_when_simd_support_not_found("AVX" "choose SSE4.1 SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
272 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
273 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
274 set(GMX_SIMD_X86_${GMX_SIMD} 1)
275 set(SIMD_STATUS_MESSAGE "Enabling 256-bit AVX SIMD instructions")
277 gmx_test_avx_gcc_maskload_bug(GMX_SIMD_X86_AVX_GCC_MASKLOAD_BUG "${SIMD_C_FLAGS}")
279 elseif(GMX_SIMD STREQUAL "AVX2_256")
281 prepare_x86_toolchain(TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS)
284 "#include<immintrin.h>
285 int main(){__m256i x=_mm256_set1_epi32(5);x=_mm256_add_epi32(x,x);return _mm256_movemask_epi8(x);}"
286 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
287 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
288 "-march=core-avx2" "-mavx2" "/arch:AVX" "-hgnu") # no AVX2-specific flag for MSVC yet
290 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
291 gmx_give_fatal_error_when_simd_support_not_found("AVX2" "choose AVX SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
294 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
295 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
296 set(GMX_SIMD_X86_${GMX_SIMD} 1)
297 set(SIMD_STATUS_MESSAGE "Enabling 256-bit AVX2 SIMD instructions")
299 # No need to test for Maskload bug - it was fixed before gcc added AVX2 support
301 elseif(GMX_SIMD STREQUAL "MIC")
303 # No flags needed. Not testing.
304 set(GMX_SIMD_X86_MIC 1)
305 set(SIMD_STATUS_MESSAGE "Enabling MIC (Xeon Phi) SIMD instructions")
307 elseif(GMX_SIMD STREQUAL "AVX_512")
309 prepare_x86_toolchain(TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS)
312 "#include<immintrin.h>
313 int main(){__m512 y,x=_mm512_set1_ps(0.5);y=_mm512_fmadd_ps(x,x,x);return (int)_mm512_cmp_ps_mask(x,y,_CMP_LT_OS);}"
314 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
315 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
316 "-xCORE-AVX512" "-mavx512f -mfma" "-mavx512f" "/arch:AVX" "-hgnu") # no AVX_512F flags known for MSVC yet
318 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
319 gmx_give_fatal_error_when_simd_support_not_found("AVX 512F" "choose a lower level of SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
322 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
323 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
324 set(GMX_SIMD_X86_${GMX_SIMD} 1)
325 set(SIMD_STATUS_MESSAGE "Enabling 512-bit AVX-512 SIMD instructions")
327 elseif(GMX_SIMD STREQUAL "AVX_512_KNL")
329 prepare_x86_toolchain(TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS)
332 "#include<immintrin.h>
333 int main(){__m512 y,x=_mm512_set1_ps(0.5);y=_mm512_rsqrt28_ps(x);return (int)_mm512_cmp_ps_mask(x,y,_CMP_LT_OS);}"
334 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
335 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
336 "-xMIC-AVX512" "-mavx512er -mfma" "-mavx512er" "/arch:AVX" "-hgnu") # no AVX_512ER flags known for MSVC yet
338 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
339 gmx_give_fatal_error_when_simd_support_not_found("AVX 512ER" "choose a lower level of SIMD (slower)" "${SUGGEST_BINUTILS_UPDATE}")
342 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
343 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
344 set(GMX_SIMD_X86_${GMX_SIMD} 1)
345 set(SIMD_STATUS_MESSAGE "Enabling 512-bit AVX-512-KNL SIMD instructions")
347 elseif(GMX_SIMD STREQUAL "ARM_NEON")
350 message(FATAL_ERROR "ARM_NEON SIMD support is not available for a double precision build because the architecture lacks double-precision support")
354 "#include<arm_neon.h>
355 int main(){float32x4_t x=vdupq_n_f32(0.5);x=vmlaq_f32(x,x,x);return vgetq_lane_f32(x,0)>0;}"
356 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
357 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
358 "-mfpu=neon-vfpv4" "-mfpu=neon" "")
360 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
361 gmx_give_fatal_error_when_simd_support_not_found("ARM NEON" "disable SIMD support (slower)" "${SUGGEST_BINUTILS_UPDATE}")
364 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
365 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
366 set(GMX_SIMD_${GMX_SIMD} 1)
367 set(SIMD_STATUS_MESSAGE "Enabling 32-bit ARM NEON SIMD instructions")
369 elseif(GMX_SIMD STREQUAL "ARM_NEON_ASIMD")
372 "#include<arm_neon.h>
373 int main(){float64x2_t x=vdupq_n_f64(0.5);x=vfmaq_f64(x,x,x);x=vrndnq_f64(x);return vgetq_lane_f64(x,0)>0;}"
374 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
375 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
378 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
379 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)" "${SUGGEST_BINUTILS_UPDATE}")
382 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
383 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
384 set(GMX_SIMD_${GMX_SIMD} 1)
385 set(SIMD_STATUS_MESSAGE "Enabling ARM (AArch64) NEON Advanced SIMD instructions")
387 elseif(GMX_SIMD STREQUAL "IBM_QPX")
389 try_compile(TEST_QPX ${CMAKE_BINARY_DIR}
390 "${CMAKE_SOURCE_DIR}/cmake/TestQPX.c")
393 message(WARNING "IBM QPX SIMD instructions selected. This will work, but SIMD kernels are only available for the Verlet cut-off scheme. The plain C kernels that are used for the group cut-off scheme kernels will be slow, so please consider using the Verlet cut-off scheme.")
394 set(GMX_SIMD_${GMX_SIMD} 1)
395 set(SIMD_STATUS_MESSAGE "Enabling IBM QPX SIMD instructions")
398 gmx_give_fatal_error_when_simd_support_not_found("IBM QPX" "or 'cmake .. -DCMAKE_TOOLCHAIN_FILE=Platform/BlueGeneQ-static-XL-CXX' to set up the tool chain" "${SUGGEST_BINUTILS_UPDATE}")
401 elseif(GMX_SIMD STREQUAL "IBM_VMX")
405 int main(){vector float x,y=vec_ctf(vec_splat_s32(1),0);x=vec_madd(y,y,y);return vec_all_ge(y,x);}"
406 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
407 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
408 "-maltivec -mabi=altivec" "-qarch=auto -qaltivec")
410 if(NOT SIMD_${GMX_SIMD}_C_FLAGS OR NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
411 gmx_give_fatal_error_when_simd_support_not_found("IBM VMX" "disable SIMD support (slower)" "${SUGGEST_BINUTILS_UPDATE}")
414 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
415 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
416 set(GMX_SIMD_${GMX_SIMD} 1)
417 set(SIMD_STATUS_MESSAGE "Enabling IBM VMX SIMD instructions")
419 elseif(GMX_SIMD STREQUAL "IBM_VSX")
421 prepare_power_vsx_toolchain(TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS)
425 int main(){vector double x,y=vec_splats(1.0);x=vec_madd(y,y,y);return vec_all_ge(y,x);}"
426 TOOLCHAIN_C_FLAGS TOOLCHAIN_CXX_FLAGS
427 SIMD_${GMX_SIMD}_C_FLAGS SIMD_${GMX_SIMD}_CXX_FLAGS
428 "-mvsx" "-maltivec -mabi=altivec" "-qarch=auto -qaltivec")
430 # Usually we check also for the C compiler here, but a C compiler
431 # is not required for SIMD support on this platform. cmake through
432 # at least version 3.7 cannot pass this check with the C compiler
433 # in the latest xlc 13.1.5, but the C++ compiler has different
434 # behaviour and is OK. See Redmine #2102.
435 if(NOT SIMD_${GMX_SIMD}_CXX_FLAGS)
436 gmx_give_fatal_error_when_simd_support_not_found("IBM VSX" "disable SIMD support (slower)" "${SUGGEST_BINUTILS_UPDATE}")
439 set(SIMD_C_FLAGS "${TOOLCHAIN_C_FLAGS}")
440 set(SIMD_CXX_FLAGS "${TOOLCHAIN_CXX_FLAGS}")
441 set(GMX_SIMD_${GMX_SIMD} 1)
442 set(SIMD_STATUS_MESSAGE "Enabling IBM VSX SIMD instructions")
444 elseif(GMX_SIMD STREQUAL "SPARC64_HPC_ACE")
446 # Note that GMX_RELAXED_DOUBLE_PRECISION is enabled by default in the top-level CMakeLists.txt
448 set(GMX_SIMD_${GMX_SIMD} 1)
449 set(SIMD_STATUS_MESSAGE "Enabling Sparc64 HPC-ACE SIMD instructions")
451 elseif(GMX_SIMD STREQUAL "REFERENCE")
453 # NB: This file handles settings for the SIMD module, so in the interest
454 # of proper modularization, please do NOT put any verlet kernel settings in this file.
456 if(GMX_SIMD_REF_FLOAT_WIDTH)
457 add_definitions(-DGMX_SIMD_REF_FLOAT_WIDTH=${GMX_SIMD_REF_FLOAT_WIDTH})
459 if(GMX_SIMD_REF_DOUBLE_WIDTH)
460 add_definitions(-DGMX_SIMD_REF_DOUBLE_WIDTH=${GMX_SIMD_REF_DOUBLE_WIDTH})
463 set(GMX_SIMD_${GMX_SIMD} 1)
464 set(SIMD_STATUS_MESSAGE "Enabling reference (emulated) SIMD instructions.")
467 gmx_invalid_option_value(GMX_SIMD)
471 gmx_check_if_changed(SIMD_CHANGED GMX_SIMD)
472 if (SIMD_CHANGED AND DEFINED SIMD_STATUS_MESSAGE)
473 message(STATUS "${SIMD_STATUS_MESSAGE}")
476 # By default, 32-bit windows cannot pass SIMD (SSE/AVX) arguments in registers,
477 # and even on 64-bit (all platforms) it is only used for a handful of arguments.
478 # The __vectorcall (MSVC, from MSVC2013) or __regcall (ICC) calling conventions
479 # enable this, which is critical to enable 32-bit SIMD and improves performance
481 # Check if the compiler supports one of these, and in that case set gmx_simdcall
482 # to that string. If we do not have any such calling convention modifier, set it
483 # to an empty string.
485 # Update 2015-11-04: As of version 3.6, clang has added support for __vectorcall
486 # (also on Linux). This appears to be buggy for the reference SIMD
487 # implementation when using the Debug build (when functions are not inlined)
488 # while it seems works fine for the actual SIMD implementations. This is likely
489 # because the reference build ends up passing lots of structures with arrays
490 # rather than actual vector data. For now we disable __vectorcall with clang
491 # when using the reference build.
493 # xlc 13.1.5 does not seem recognize any attribute, and warns about invalid ones
494 # so we avoid searching for any.
496 if(NOT DEFINED GMX_SIMD_CALLING_CONVENTION)
498 set(CALLCONV_LIST " ")
499 elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND GMX_SIMD STREQUAL "REFERENCE")
500 set(CALLCONV_LIST __regcall " ")
501 elseif(CMAKE_CXX_COMPILER_ID MATCHES "XL")
502 set(CALLCONV_LIST " ")
504 set(CALLCONV_LIST __vectorcall __regcall " ")
506 foreach(callconv ${CALLCONV_LIST})
507 set(callconv_compile_var "_callconv_${callconv}")
508 check_c_source_compiles("int ${callconv} f(int i) {return i;} int main(void) {return f(0);}" ${callconv_compile_var})
509 if(${callconv_compile_var})
510 set(GMX_SIMD_CALLING_CONVENTION "${callconv}" CACHE INTERNAL "Calling convention for SIMD routines" FORCE)