2 # This file is part of the GROMACS molecular simulation package.
4 # Copyright (c) 2012,2013,2014, 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)
40 macro(gmx_use_clang_as_with_gnu_compilers_on_osx)
41 # On OS X, we often want to use gcc instead of clang, since gcc supports
42 # OpenMP. However, by default gcc uses the external system assembler, which
43 # does not support AVX, so we need to tell the linker to use the clang
44 # compilers assembler instead - and this has to happen before we detect AVX
46 if(APPLE AND ${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
47 gmx_test_cflag(GNU_C_USE_CLANG_AS "-Wa,-q" ACCELERATION_C_FLAGS)
49 if(APPLE AND ${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
50 gmx_test_cxxflag(GNU_CXX_USE_CLANG_AS "-Wa,-q" ACCELERATION_CXX_FLAGS)
55 macro(gmx_test_cpu_acceleration)
57 # To improve backward compatibility on x86 SIMD architectures,
58 # we set the flags for all accelerations that are supported, not only
59 # the most recent instruction set. I.e., if your machine supports AVX2_256,
60 # we will set flags both for AVX2_256, AVX_256, SSE4.1, and SSE2 support.
62 if(${GMX_CPU_ACCELERATION} STREQUAL "NONE")
63 # nothing to do configuration-wise
64 set(ACCELERATION_STATUS_MESSAGE "CPU SIMD acceleration disabled")
65 elseif(${GMX_CPU_ACCELERATION} STREQUAL "SSE2")
67 gmx_find_cflag_for_source(CFLAGS_SSE2 "C compiler SSE2 flag"
68 "#include<xmmintrin.h>
69 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_rsqrt_ps(x);return 0;}"
71 "-msse2" "/arch:SSE2")
72 gmx_find_cxxflag_for_source(CXXFLAGS_SSE2 "C++ compiler SSE2 flag"
73 "#include<xmmintrin.h>
74 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_rsqrt_ps(x);return 0;}"
75 ACCELERATION_CXX_FLAGS
76 "-msse2" "/arch:SSE2")
78 if(NOT CFLAGS_SSE2 OR NOT CXXFLAGS_SSE2)
79 message(FATAL_ERROR "Cannot find SSE2 compiler flag. Use a newer compiler, or disable acceleration (slower).")
82 set(GMX_CPU_ACCELERATION_X86_SSE2 1)
85 set(ACCELERATION_STATUS_MESSAGE "Enabling SSE2 SIMD Gromacs acceleration")
87 elseif(${GMX_CPU_ACCELERATION} STREQUAL "SSE4.1")
89 # Note: MSVC enables SSE4.1 with the SSE2 flag, so we include that in testing.
90 gmx_find_cflag_for_source(CFLAGS_SSE4_1 "C compiler SSE4.1 flag"
91 "#include<smmintrin.h>
92 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_dp_ps(x,x,0x77);return 0;}"
94 "-msse4.1" "/arch:SSE4.1" "/arch:SSE2")
95 gmx_find_cxxflag_for_source(CXXFLAGS_SSE4_1 "C++ compiler SSE4.1 flag"
96 "#include<smmintrin.h>
97 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_dp_ps(x,x,0x77);return 0;}"
98 ACCELERATION_CXX_FLAGS
99 "-msse4.1" "/arch:SSE4.1" "/arch:SSE2")
101 if(NOT CFLAGS_SSE4_1 OR NOT CXXFLAGS_SSE4_1)
102 message(FATAL_ERROR "Cannot find SSE4.1 compiler flag. "
103 "Use a newer compiler, or choose SSE2 acceleration (slower).")
106 if(CMAKE_C_COMPILER_ID MATCHES "Intel" AND CMAKE_C_COMPILER_VERSION VERSION_EQUAL "11.1")
107 message(FATAL_ERROR "You are using Intel compiler version 11.1, which produces incorrect results with SSE4.1 acceleration. You need to use a newer compiler (e.g. icc >= 12.0) or in worst case try a lower level of acceleration if performance is not critical.")
110 set(GMX_CPU_ACCELERATION_X86_SSE4_1 1)
111 set(GMX_X86_SSE4_1 1)
113 set(ACCELERATION_STATUS_MESSAGE "Enabling SSE4.1 SIMD Gromacs acceleration")
115 elseif(${GMX_CPU_ACCELERATION} STREQUAL "AVX_128_FMA")
117 gmx_use_clang_as_with_gnu_compilers_on_osx()
119 # AVX128/FMA on AMD is a bit complicated. We need to do detection in three stages:
120 # 1) Find the flags required for generic AVX support
121 # 2) Find the flags necessary to enable fused-multiply add support
122 # 3) Optional: Find a flag to enable the AMD XOP instructions
124 ### STAGE 1: Find the generic AVX flag
125 gmx_find_cflag_for_source(CFLAGS_AVX_128 "C compiler AVX (128 bit) flag"
126 "#include<immintrin.h>
127 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_permute_ps(x,1);return 0;}"
130 gmx_find_cxxflag_for_source(CXXFLAGS_AVX_128 "C++ compiler AVX (128 bit) flag"
131 "#include<immintrin.h>
132 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_permute_ps(x,1);return 0;}"
133 ACCELERATION_CXX_FLAGS
136 ### STAGE 2: Find the fused-multiply add flag.
137 # GCC requires x86intrin.h for FMA support. MSVC 2010 requires intrin.h for FMA support.
138 check_include_file(x86intrin.h HAVE_X86INTRIN_H ${ACCELERATION_C_FLAGS})
139 check_include_file(intrin.h HAVE_INTRIN_H ${ACCELERATION_C_FLAGS})
141 set(INCLUDE_X86INTRIN_H "#include <x86intrin.h>")
144 set(INCLUDE_INTRIN_H "#include <xintrin.h>")
147 gmx_find_cflag_for_source(CFLAGS_AVX_128_FMA "C compiler AVX (128 bit) FMA4 flag"
148 "#include<immintrin.h>
149 ${INCLUDE_X86INTRIN_H}
151 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_macc_ps(x,x,x);return 0;}"
154 gmx_find_cxxflag_for_source(CXXFLAGS_AVX_128_FMA "C++ compiler AVX (128 bit) FMA4 flag"
155 "#include<immintrin.h>
156 ${INCLUDE_X86INTRIN_H}
158 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_macc_ps(x,x,x);return 0;}"
159 ACCELERATION_CXX_FLAGS
162 # We only need to check the last (FMA) test; that will always fail if the basic AVX128 test failed
163 if(NOT CFLAGS_AVX_128_FMA OR NOT CXXFLAGS_AVX_128_FMA)
164 message(FATAL_ERROR "Cannot find compiler flags for 128 bit AVX with FMA support. Use a newer compiler, or choose SSE4.1 acceleration (slower).")
167 ### STAGE 3: Optional: Find the XOP instruction flag (No point in yelling if this does not work)
168 gmx_find_cflag_for_source(CFLAGS_AVX_128_XOP "C compiler AVX (128 bit) XOP flag"
169 "#include<immintrin.h>
170 ${INCLUDE_X86INTRIN_H}
172 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_frcz_ps(x);return 0;}"
175 gmx_find_cxxflag_for_source(CXXFLAGS_AVX_128_XOP "C++ compiler AVX (128 bit) XOP flag"
176 "#include<immintrin.h>
177 ${INCLUDE_X86INTRIN_H}
179 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_frcz_ps(x);return 0;}"
180 ACCELERATION_CXX_FLAGS
183 # We don't have the full compiler version string yet (BUILD_C_COMPILER),
184 # so we can't distinguish vanilla from Apple clang versions, but catering for a few rare AMD
185 # hackintoshes is not worth the effort.
186 if (APPLE AND (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR
187 ${CMAKE_CXX_COMPILER_ID} STREQUAL "Clang"))
188 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 acceleration. As we cannot work around this bug on OS X, you will have to select a different compiler or CPU acceleration.")
192 if (GMX_USE_CLANG_C_FMA_BUG_WORKAROUND)
193 # we assume that we have an external assembler that supports AVX
194 message(STATUS "Clang ${CMAKE_C_COMPILER_VERSION} detected, enabling FMA bug workaround")
195 set(EXTRA_C_FLAGS "${EXTRA_C_FLAGS} -no-integrated-as")
197 if (GMX_USE_CLANG_CXX_FMA_BUG_WORKAROUND)
198 # we assume that we have an external assembler that supports AVX
199 message(STATUS "Clang ${CMAKE_CXX_COMPILER_VERSION} detected, enabling FMA bug workaround")
200 set(EXTRA_CXX_FLAGS "${EXTRA_CXX_FLAGS} -no-integrated-as")
203 gmx_test_avx_gcc_maskload_bug(GMX_X86_AVX_GCC_MASKLOAD_BUG "${ACCELERATION_C_FLAGS}")
205 set(GMX_CPU_ACCELERATION_X86_AVX_128_FMA 1)
206 set(GMX_X86_AVX_128_FMA 1)
207 set(GMX_X86_SSE4_1 1)
210 set(ACCELERATION_STATUS_MESSAGE "Enabling 128-bit AVX SIMD Gromacs acceleration (with fused-multiply add)")
212 elseif(${GMX_CPU_ACCELERATION} STREQUAL "AVX_256")
214 gmx_use_clang_as_with_gnu_compilers_on_osx()
216 gmx_find_cflag_for_source(CFLAGS_AVX "C compiler AVX (256 bit) flag"
217 "#include<immintrin.h>
218 int main(){__m256 x=_mm256_set1_ps(0.5);x=_mm256_add_ps(x,x);return 0;}"
221 gmx_find_cxxflag_for_source(CXXFLAGS_AVX "C++ compiler AVX (256 bit) flag"
222 "#include<immintrin.h>
223 int main(){__m256 x=_mm256_set1_ps(0.5);x=_mm256_add_ps(x,x);return 0;}"
224 ACCELERATION_CXX_FLAGS
227 if(NOT CFLAGS_AVX OR NOT CXXFLAGS_AVX)
228 message(FATAL_ERROR "Cannot find AVX compiler flag. Use a newer compiler, or choose SSE4.1 acceleration (slower).")
231 gmx_test_avx_gcc_maskload_bug(GMX_X86_AVX_GCC_MASKLOAD_BUG "${ACCELERATION_C_FLAGS}")
233 set(GMX_CPU_ACCELERATION_X86_AVX_256 1)
234 set(GMX_X86_AVX_256 1)
235 set(GMX_X86_SSE4_1 1)
238 set(ACCELERATION_STATUS_MESSAGE "Enabling 256-bit AVX SIMD Gromacs acceleration")
240 elseif(${GMX_CPU_ACCELERATION} STREQUAL "AVX2_256")
242 # Comment out this line for AVX2 development
243 message(FATAL_ERROR "AVX2_256 is disabled until the implementation has been commited.")
245 gmx_use_clang_as_with_gnu_compilers_on_osx()
247 gmx_find_cflag_for_source(CFLAGS_AVX2 "C compiler AVX2 flag"
248 "#include<immintrin.h>
249 int main(){__m256 x=_mm256_set1_ps(0.5);x=_mm256_fmadd_ps(x,x,x);return 0;}"
251 "-march=core-avx2" "-mavx2" "/arch:AVX") # no AVX2-specific flag for MSVC yet
252 gmx_find_cxxflag_for_source(CXXFLAGS_AVX2 "C++ compiler AVX2 flag"
253 "#include<immintrin.h>
254 int main(){__m256 x=_mm256_set1_ps(0.5);x=_mm256_fmadd_ps(x,x,x);return 0;}"
255 ACCELERATION_CXX_FLAGS
256 "-march=core-avx2" "-mavx2" "/arch:AVX") # no AVX2-specific flag for MSVC yet
258 if(NOT CFLAGS_AVX2 OR NOT CXXFLAGS_AVX2)
259 message(FATAL_ERROR "Cannot find AVX2 compiler flag. Use a newer compiler, or choose AVX acceleration (slower).")
262 # No need to test for Maskload bug - it was fixed before gcc added AVX2 support
264 set(GMX_CPU_ACCELERATION_X86_AVX2_256 1)
265 set(GMX_X86_AVX2_256 1)
266 set(GMX_X86_AVX_256 1)
267 set(GMX_X86_SSE4_1 1)
270 set(ACCELERATION_STATUS_MESSAGE "Enabling 256-bit AVX2 Gromacs acceleration")
272 elseif(${GMX_CPU_ACCELERATION} STREQUAL "IBM_QPX")
274 try_compile(TEST_QPX ${CMAKE_BINARY_DIR}
275 "${CMAKE_SOURCE_DIR}/cmake/TestQPX.c")
278 message(WARNING "IBM QPX acceleration was selected. This will work, but SIMD-accelerated 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.")
279 set(GMX_CPU_ACCELERATION_IBM_QPX 1)
280 set(ACCELERATION_STATUS_MESSAGE "Enabling IBM QPX SIMD acceleration")
283 message(FATAL_ERROR "Cannot compile the requested IBM QPX intrinsics. If you are compiling for BlueGene/Q with the XL compilers, use 'cmake .. -DCMAKE_TOOLCHAIN_FILE=Platform/BlueGeneQ-static-XL-C' to set up the tool chain.")
286 elseif(${GMX_CPU_ACCELERATION} STREQUAL "SPARC64_HPC_ACE")
288 set(GMX_CPU_ACCELERATION_SPARC64_HPC_ACE 1)
289 set(ACCELERATION_STATUS_MESSAGE "Enabling Sparc64 HPC-ACE SIMD acceleration")
291 elseif(${GMX_CPU_ACCELERATION} STREQUAL "REFERENCE")
293 add_definitions(-DGMX_SIMD_REFERENCE_PLAIN_C)
294 if(${GMX_NBNXN_REF_KERNEL_TYPE} STREQUAL "4xn")
295 if(${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "2" OR ${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "4" OR ${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "8")
296 add_definitions(-DGMX_NBNXN_SIMD_4XN -DGMX_SIMD_REF_WIDTH=${GMX_NBNXN_REF_KERNEL_WIDTH})
298 message(FATAL_ERROR "Unsupported width for 4xn reference kernels")
300 elseif(${GMX_NBNXN_REF_KERNEL_TYPE} STREQUAL "2xnn")
301 if(${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "8" OR ${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "16")
302 add_definitions(-DGMX_NBNXN_SIMD_2XNN -DGMX_SIMD_REF_WIDTH=${GMX_NBNXN_REF_KERNEL_WIDTH})
304 message(FATAL_ERROR "Unsupported width for 2xn reference kernels")
307 message(FATAL_ERROR "Unsupported kernel type")
311 gmx_invalid_option_value(GMX_CPU_ACCELERATION)
315 gmx_check_if_changed(ACCELERATION_CHANGED GMX_CPU_ACCELERATION)
316 if (ACCELERATION_CHANGED AND DEFINED ACCELERATION_STATUS_MESSAGE)
317 message(STATUS "${ACCELERATION_STATUS_MESSAGE}")