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" SIMD_C_FLAGS)
49 if(APPLE AND ${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
50 gmx_test_cxxflag(GNU_CXX_USE_CLANG_AS "-Wa,-q" SIMD_CXX_FLAGS)
57 # To improve backward compatibility on x86 SIMD architectures,
58 # we set the flags for all SIMD instructions that are supported, not only
59 # the most recent instruction set. I.e., if your machine supports AVX2_256,
60 # we will set flags both for AVX2_256, AVX_256, SSE4.1, and SSE2 support.
62 if(${GMX_SIMD} STREQUAL "NONE")
63 # nothing to do configuration-wise
64 set(SIMD_STATUS_MESSAGE "SIMD instructions disabled")
65 elseif(${GMX_SIMD} STREQUAL "SSE2")
67 gmx_find_cflag_for_source(CFLAGS_SSE2 "C compiler SSE2 flag"
68 "#include<xmmintrin.h>
69 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_rsqrt_ps(x);return _mm_movemask_ps(x);}"
71 "-msse2" "/arch:SSE2" "-hgnu")
72 gmx_find_cxxflag_for_source(CXXFLAGS_SSE2 "C++ compiler SSE2 flag"
73 "#include<xmmintrin.h>
74 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_rsqrt_ps(x);return _mm_movemask_ps(x);}"
76 "-msse2" "/arch:SSE2" "-hgnu")
78 if(NOT CFLAGS_SSE2 OR NOT CXXFLAGS_SSE2)
79 message(FATAL_ERROR "Cannot find SSE2 compiler flag. Use a newer compiler, or disable SIMD (slower).")
82 set(GMX_SIMD_X86_SSE2 1)
83 set(SIMD_STATUS_MESSAGE "Enabling SSE2 SIMD instructions")
85 elseif(${GMX_SIMD} STREQUAL "SSE4.1")
87 # Note: MSVC enables SSE4.1 with the SSE2 flag, so we include that in testing.
88 gmx_find_cflag_for_source(CFLAGS_SSE4_1 "C compiler SSE4.1 flag"
89 "#include<smmintrin.h>
90 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_dp_ps(x,x,0x77);return _mm_movemask_ps(x);}"
92 "-msse4.1" "/arch:SSE4.1" "/arch:SSE2" "-hgnu")
93 gmx_find_cxxflag_for_source(CXXFLAGS_SSE4_1 "C++ compiler SSE4.1 flag"
94 "#include<smmintrin.h>
95 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_dp_ps(x,x,0x77);return _mm_movemask_ps(x);}"
97 "-msse4.1" "/arch:SSE4.1" "/arch:SSE2" "-hgnu")
99 if(NOT CFLAGS_SSE4_1 OR NOT CXXFLAGS_SSE4_1)
100 message(FATAL_ERROR "Cannot find SSE4.1 compiler flag. "
101 "Use a newer compiler, or choose SSE2 SIMD (slower).")
104 if(CMAKE_C_COMPILER_ID MATCHES "Intel" AND CMAKE_C_COMPILER_VERSION VERSION_EQUAL "11.1")
105 message(FATAL_ERROR "You are using Intel compiler version 11.1, which produces incorrect results with SSE4.1 SIMD. You need to use a newer compiler (e.g. icc >= 12.0) or in worst case try a lower level of SIMD if performance is not critical.")
108 set(GMX_SIMD_X86_SSE4_1 1)
109 set(SIMD_STATUS_MESSAGE "Enabling SSE4.1 SIMD instructions")
111 elseif(${GMX_SIMD} STREQUAL "AVX_128_FMA")
113 gmx_use_clang_as_with_gnu_compilers_on_osx()
115 # AVX128/FMA on AMD is a bit complicated. We need to do detection in three stages:
116 # 1) Find the flags required for generic AVX support
117 # 2) Find the flags necessary to enable fused-multiply add support
118 # 3) Optional: Find a flag to enable the AMD XOP instructions
120 ### STAGE 1: Find the generic AVX flag
121 gmx_find_cflag_for_source(CFLAGS_AVX_128 "C compiler AVX (128 bit) flag"
122 "#include<immintrin.h>
123 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_permute_ps(x,1);return 0;}"
125 "-mavx" "/arch:AVX" "-hgnu")
126 gmx_find_cxxflag_for_source(CXXFLAGS_AVX_128 "C++ compiler AVX (128 bit) flag"
127 "#include<immintrin.h>
128 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_permute_ps(x,1);return 0;}"
130 "-mavx" "/arch:AVX" "-hgnu")
132 ### STAGE 2: Find the fused-multiply add flag.
133 # GCC requires x86intrin.h for FMA support. MSVC 2010 requires intrin.h for FMA support.
134 check_include_file(x86intrin.h HAVE_X86INTRIN_H ${SIMD_C_FLAGS})
135 check_include_file(intrin.h HAVE_INTRIN_H ${SIMD_C_FLAGS})
137 set(INCLUDE_X86INTRIN_H "#include <x86intrin.h>")
140 set(INCLUDE_INTRIN_H "#include <xintrin.h>")
143 gmx_find_cflag_for_source(CFLAGS_AVX_128_FMA "C compiler AVX (128 bit) FMA4 flag"
144 "#include<immintrin.h>
145 ${INCLUDE_X86INTRIN_H}
147 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_macc_ps(x,x,x);return _mm_movemask_ps(x);}"
150 gmx_find_cxxflag_for_source(CXXFLAGS_AVX_128_FMA "C++ compiler AVX (128 bit) FMA4 flag"
151 "#include<immintrin.h>
152 ${INCLUDE_X86INTRIN_H}
154 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_macc_ps(x,x,x);return _mm_movemask_ps(x);}"
158 # We only need to check the last (FMA) test; that will always fail if the basic AVX128 test failed
159 if(NOT CFLAGS_AVX_128_FMA OR NOT CXXFLAGS_AVX_128_FMA)
160 message(FATAL_ERROR "Cannot find compiler flags for 128 bit AVX with FMA support. Use a newer compiler, or choose SSE4.1 SIMD (slower).")
163 ### STAGE 3: Optional: Find the XOP instruction flag (No point in yelling if this does not work)
164 gmx_find_cflag_for_source(CFLAGS_AVX_128_XOP "C compiler AVX (128 bit) XOP flag"
165 "#include<immintrin.h>
166 ${INCLUDE_X86INTRIN_H}
168 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_frcz_ps(x);return _mm_movemask_ps(x);}"
171 gmx_find_cxxflag_for_source(CXXFLAGS_AVX_128_XOP "C++ compiler AVX (128 bit) XOP flag"
172 "#include<immintrin.h>
173 ${INCLUDE_X86INTRIN_H}
175 int main(){__m128 x=_mm_set1_ps(0.5);x=_mm_frcz_ps(x);return _mm_movemask_ps(x);}"
179 # We don't have the full compiler version string yet (BUILD_C_COMPILER),
180 # so we can't distinguish vanilla from Apple clang versions, but catering for a few rare AMD
181 # hackintoshes is not worth the effort.
182 if (APPLE AND (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR
183 ${CMAKE_CXX_COMPILER_ID} STREQUAL "Clang"))
184 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.")
188 if (GMX_USE_CLANG_C_FMA_BUG_WORKAROUND)
189 # we assume that we have an external assembler that supports AVX
190 message(STATUS "Clang ${CMAKE_C_COMPILER_VERSION} detected, enabling FMA bug workaround")
191 set(EXTRA_C_FLAGS "${EXTRA_C_FLAGS} -no-integrated-as")
193 if (GMX_USE_CLANG_CXX_FMA_BUG_WORKAROUND)
194 # we assume that we have an external assembler that supports AVX
195 message(STATUS "Clang ${CMAKE_CXX_COMPILER_VERSION} detected, enabling FMA bug workaround")
196 set(EXTRA_CXX_FLAGS "${EXTRA_CXX_FLAGS} -no-integrated-as")
199 gmx_test_avx_gcc_maskload_bug(GMX_SIMD_X86_AVX_GCC_MASKLOAD_BUG "${SIMD_C_FLAGS}")
201 set(GMX_SIMD_X86_AVX_128_FMA 1)
202 set(SIMD_STATUS_MESSAGE "Enabling 128-bit AVX SIMD Gromacs SIMD (with fused-multiply add)")
204 elseif(${GMX_SIMD} STREQUAL "AVX_256")
206 gmx_use_clang_as_with_gnu_compilers_on_osx()
208 gmx_find_cflag_for_source(CFLAGS_AVX "C compiler AVX (256 bit) flag"
209 "#include<immintrin.h>
210 int main(){__m256 x=_mm256_set1_ps(0.5);x=_mm256_add_ps(x,x);return _mm256_movemask_ps(x);}"
212 "-mavx" "/arch:AVX" "-hgnu")
213 gmx_find_cxxflag_for_source(CXXFLAGS_AVX "C++ compiler AVX (256 bit) flag"
214 "#include<immintrin.h>
215 int main(){__m256 x=_mm256_set1_ps(0.5);x=_mm256_add_ps(x,x);return _mm256_movemask_ps(x);}"
217 "-mavx" "/arch:AVX" "-hgnu")
219 if(NOT CFLAGS_AVX OR NOT CXXFLAGS_AVX)
220 message(FATAL_ERROR "Cannot find AVX compiler flag. Use a newer compiler, or choose SSE4.1 SIMD (slower).")
223 gmx_test_avx_gcc_maskload_bug(GMX_SIMD_X86_AVX_GCC_MASKLOAD_BUG "${SIMD_C_FLAGS}")
225 set(GMX_SIMD_X86_AVX_256 1)
226 set(SIMD_STATUS_MESSAGE "Enabling 256-bit AVX SIMD instructions")
228 elseif(${GMX_SIMD} STREQUAL "AVX2_256")
230 gmx_use_clang_as_with_gnu_compilers_on_osx()
232 gmx_find_cflag_for_source(CFLAGS_AVX2 "C compiler AVX2 flag"
233 "#include<immintrin.h>
234 int main(){__m256i x=_mm256_set1_epi32(5);x=_mm256_add_epi32(x,x);return _mm256_movemask_epi8(x);}"
236 "-march=core-avx2" "-mavx2" "/arch:AVX" "-hgnu") # no AVX2-specific flag for MSVC yet
237 gmx_find_cxxflag_for_source(CXXFLAGS_AVX2 "C++ compiler AVX2 flag"
238 "#include<immintrin.h>
239 int main(){__m256i x=_mm256_set1_epi32(5);x=_mm256_add_epi32(x,x);return _mm256_movemask_epi8(x);}"
241 "-march=core-avx2" "-mavx2" "/arch:AVX" "-hgnu") # no AVX2-specific flag for MSVC yet
243 if(NOT CFLAGS_AVX2 OR NOT CXXFLAGS_AVX2)
244 message(FATAL_ERROR "Cannot find AVX2 compiler flag. Use a newer compiler, or choose AVX SIMD (slower).")
247 # No need to test for Maskload bug - it was fixed before gcc added AVX2 support
249 set(GMX_SIMD_X86_AVX2_256 1)
250 set(SIMD_STATUS_MESSAGE "Enabling 256-bit AVX2 SIMD instructions")
252 elseif(${GMX_SIMD} STREQUAL "ARM_NEON")
254 gmx_find_cflag_for_source(CFLAGS_ARM_NEON "C compiler 32-bit ARM NEON flag"
255 "#include<arm_neon.h>
256 int main(){float32x4_t x=vdupq_n_f32(0.5);x=vmlaq_f32(x,x,x);return vgetq_lane_f32(x,0)>0;}"
259 gmx_find_cxxflag_for_source(CXXFLAGS_ARM_NEON "C++ compiler 32-bit ARM NEON flag"
260 "#include<arm_neon.h>
261 int main(){float32x4_t x=vdupq_n_f32(0.5);x=vmlaq_f32(x,x,x);return vgetq_lane_f32(x,0)>0;}"
263 "-mfpu=neon" "-D__STDC_CONSTANT_MACROS" "")
265 if(NOT CFLAGS_ARM_NEON OR NOT CXXFLAGS_ARM_NEON)
266 message(FATAL_ERROR "Cannot find ARM 32-bit NEON compiler flag. Use a newer compiler, or disable NEON SIMD.")
269 set(GMX_SIMD_ARM_NEON 1)
270 set(SIMD_STATUS_MESSAGE "Enabling 32-bit ARM NEON SIMD instructions")
272 elseif(${GMX_SIMD} STREQUAL "ARM_NEON_ASIMD")
273 # Gcc-4.8.1 appears to have a bug where the c++ compiler requires
274 # -D__STDC_CONSTANT_MACROS if we include arm_neon.h
276 gmx_find_cflag_for_source(CFLAGS_ARM_NEON_ASIMD "C compiler ARM NEON Advanced SIMD flag"
277 "#include<arm_neon.h>
278 int main(){float64x2_t x=vdupq_n_f64(0.5);x=vfmaq_f64(x,x,x);return vgetq_lane_f64(x,0)>0;}"
281 gmx_find_cxxflag_for_source(CXXFLAGS_ARM_NEON_ASIMD "C++ compiler ARM NEON Advanced SIMD flag"
282 "#include<arm_neon.h>
283 int main(){float64x2_t x=vdupq_n_f64(0.5);x=vfmaq_f64(x,x,x);return vgetq_lane_f64(x,0)>0;}"
285 "-D__STDC_CONSTANT_MACROS" "")
287 if(NOT CFLAGS_ARM_NEON_ASIMD OR NOT CXXFLAGS_ARM_NEON_ASIMD)
288 message(FATAL_ERROR "Cannot find ARM (AArch64) NEON Advanced SIMD compiler flag. Use a newer compiler, or disable SIMD.")
291 if(CMAKE_C_COMPILER_ID MATCHES "GNU" AND CMAKE_C_COMPILER_VERSION VERSION_LESS "4.9")
292 message(WARNING "At least gcc-4.8.1 has many bugs for ARM (AArch64) NEON Advanced SIMD compilation. You might need gcc version 4.9 or later.")
295 if(CMAKE_C_COMPILER_ID MATCHES "Clang" AND CMAKE_C_COMPILER_VERSION VERSION_LESS "3.4")
296 message(FATAL_ERROR "Clang version 3.4 or later is required for ARM (AArch64) NEON Advanced SIMD.")
299 set(GMX_SIMD_ARM_NEON_ASIMD 1)
300 set(SIMD_STATUS_MESSAGE "Enabling ARM (AArch64) NEON Advanced SIMD instructions")
302 elseif(${GMX_SIMD} STREQUAL "IBM_QPX")
304 try_compile(TEST_QPX ${CMAKE_BINARY_DIR}
305 "${CMAKE_SOURCE_DIR}/cmake/TestQPX.c")
308 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.")
309 set(GMX_SIMD_IBM_QPX 1)
310 set(SIMD_STATUS_MESSAGE "Enabling IBM QPX SIMD instructions")
313 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.")
316 elseif(${GMX_SIMD} STREQUAL "IBM_VMX")
318 gmx_find_cflag_for_source(CFLAGS_IBM_VMX "C compiler IBM VMX SIMD flag"
320 int main(){vector float x,y=vec_ctf(vec_splat_s32(1),0);x=vec_madd(y,y,y);return vec_all_ge(y,x);}"
322 "-maltivec -mabi=altivec" "-qarch=auto -qaltivec")
323 gmx_find_cxxflag_for_source(CXXFLAGS_IBM_VMX "C++ compiler IBM VMX SIMD flag"
325 int main(){vector float x,y=vec_ctf(vec_splat_s32(1),0);x=vec_madd(y,y,y);return vec_all_ge(y,x);}"
327 "-maltivec -mabi=altivec" "-qarch=auto -qaltivec")
329 if(NOT CFLAGS_IBM_VMX OR NOT CXXFLAGS_IBM_VMX)
330 message(FATAL_ERROR "Cannot find IBM VMX SIMD compiler flag. Use a newer compiler, or disable VMX SIMD.")
333 set(GMX_SIMD_IBM_VMX 1)
334 set(SIMD_STATUS_MESSAGE "Enabling IBM VMX SIMD instructions")
336 elseif(${GMX_SIMD} STREQUAL "SPARC64_HPC_ACE")
338 # Note that GMX_RELAXED_DOUBLE_PRECISION is enabled by default in the top-level CMakeLists.txt
340 set(GMX_SIMD_SPARC64_HPC_ACE 1)
341 set(SIMD_STATUS_MESSAGE "Enabling Sparc64 HPC-ACE SIMD instructions")
343 elseif(${GMX_SIMD} STREQUAL "REFERENCE")
345 # NB: This file handles settings for the SIMD module, so in the interest
346 # of proper modularization, please do NOT put any verlet kernel settings in this file.
348 if(GMX_SIMD_REF_FLOAT_WIDTH)
349 add_definitions(-DGMX_SIMD_REF_FLOAT_WIDTH=${GMX_SIMD_REF_FLOAT_WIDTH})
351 if(GMX_SIMD_REF_DOUBLE_WIDTH)
352 add_definitions(-DGMX_SIMD_REF_DOUBLE_WIDTH=${GMX_SIMD_REF_DOUBLE_WIDTH})
355 set(GMX_SIMD_REFERENCE 1)
356 set(SIMD_STATUS_MESSAGE "Enabling reference (emulated) SIMD instructions.")
359 gmx_invalid_option_value(GMX_SIMD)
363 gmx_check_if_changed(SIMD_CHANGED GMX_SIMD)
364 if (SIMD_CHANGED AND DEFINED SIMD_STATUS_MESSAGE)
365 message(STATUS "${SIMD_STATUS_MESSAGE}")
368 # By default, 32-bit windows cannot pass SIMD (SSE/AVX) arguments in registers,
369 # and even on 64-bit (all platforms) it is only used for a handful of arguments.
370 # The __vectorcall (MSVC, from MSVC2013) or __regcall (ICC) calling conventions
371 # enable this, which is critical to enable 32-bit SIMD and improves performance
373 # Check if the compiler supports one of these, and in that case set gmx_simdcall
374 # to that string. If we do not have any such calling convention modifier, set it
375 # to an empty string.
376 if(NOT DEFINED GMX_SIMD_CALLING_CONVENTION)
377 foreach(callconv __vectorcall __regcall "")
378 set(callconv_compile_var "_callconv_${callconv}")
379 check_c_source_compiles("int ${callconv} f(int i) {return i;} int main(void) {return f(0);}" ${callconv_compile_var})
380 if(${callconv_compile_var})
381 set(GMX_SIMD_CALLING_CONVENTION "${callconv}" CACHE INTERNAL "Calling convention for SIMD routines" FORCE)