gmx_detect_target_architecture()
include(gmxDetectAcceleration)
gmx_detect_acceleration(GMX_SUGGESTED_CPU_ACCELERATION)
+if("${GMX_SUGGESTED_CPU_ACCELERATION}" STREQUAL "AVX2_256")
+ message(STATUS "Changing acceleration from AVX2 to AVX (until AVX2 patches commited).")
+ set(GMX_SUGGESTED_CPU_ACCELERATION "AVX_256")
+endif()
gmx_option_multichoice(
GMX_CPU_ACCELERATION
"Acceleration for CPU kernels and compiler optimization"
"${GMX_SUGGESTED_CPU_ACCELERATION}"
- None SSE2 SSE4.1 AVX_128_FMA AVX_256 IBM_QPX Sparc64_HPC_ACE Reference)
+ None SSE2 SSE4.1 AVX_128_FMA AVX_256 AVX2_256 IBM_QPX Sparc64_HPC_ACE Reference)
gmx_option_multichoice(
GMX_FFT_LIBRARY
set(GMX_INTERNAL_XDR 1)
endif(NOT GMX_SYSTEM_XDR)
-# include avx test source, used if the AVX flags are set below
-include(gmxTestAVXMaskload)
-
-# Process nonbonded accelerated kernels settings
-#
-# Note that for the backward-compatible x86 SIMD architectures, the
-# GMX_CPU_ACCELERATION determines the maximum level of the instruction
-# set used (e.g. GMX_CPU_ACCLERATION=SSE4.1 implies
-# SSE2). Accordingly, there are a set of CMake variables
-# GMX_<arch>_<feature-set> that are exported to the C code to specify
-# CPU features that should be used. This means that the logic for
-# requiring such backward compatibility is all located here.
-if(${GMX_CPU_ACCELERATION} STREQUAL "NONE")
- # nothing to do
- set(ACCELERATION_STATUS_MESSAGE "CPU acceleration disabled")
-
-elseif(${GMX_CPU_ACCELERATION} STREQUAL "SSE2")
-
- GMX_TEST_CFLAG(GNU_SSE2_CFLAG "-msse2" ACCELERATION_C_FLAGS)
- if(NOT GNU_SSE2_CFLAG AND GMX_NATIVE_WINDOWS)
- GMX_TEST_CFLAG(MSVC_SSE2_CFLAG "/arch:SSE2" ACCELERATION_C_FLAGS)
- endif(NOT GNU_SSE2_CFLAG AND GMX_NATIVE_WINDOWS)
-
- GMX_TEST_CXXFLAG(GNU_SSE2_CXXFLAG "-msse2" ACCELERATION_CXX_FLAGS)
- if(NOT GNU_SSE2_CXXFLAG AND GMX_NATIVE_WINDOWS)
- GMX_TEST_CXXFLAG(MSVC_SSE2_CXXFLAG "/arch:SSE2" ACCELERATION_CXX_FLAGS)
- endif(NOT GNU_SSE2_CXXFLAG AND GMX_NATIVE_WINDOWS)
-
- # We dont warn for lacking SSE2 flag support, since that is probably standard today.
-
- # Only test the include after we have tried to add the correct flag for SSE2 support
- check_include_file(emmintrin.h HAVE_EMMINTRIN_H ${ACCELERATION_C_FLAGS})
-
- if(NOT HAVE_EMMINTRIN_H)
- message(FATAL_ERROR "Cannot find emmintrin.h, which is required for SSE2 intrinsics support.")
- endif(NOT HAVE_EMMINTRIN_H)
-
- set(GMX_CPU_ACCELERATION_X86_SSE2 1)
- # The user should not be able to set this orthogonally to the acceleration
- set(GMX_X86_SSE2 1)
- set(ACCELERATION_STATUS_MESSAGE
- "Enabling SSE2 Gromacs acceleration")
-
-elseif(${GMX_CPU_ACCELERATION} STREQUAL "SSE4.1")
-
- GMX_TEST_CFLAG(GNU_SSE4_CFLAG "-msse4.1" ACCELERATION_C_FLAGS)
- if (NOT GNU_SSE4_CFLAG AND GMX_NATIVE_WINDOWS)
- GMX_TEST_CFLAG(MSVC_SSE4_CFLAG "/arch:SSE4.1" ACCELERATION_C_FLAGS)
- endif(NOT GNU_SSE4_CFLAG AND GMX_NATIVE_WINDOWS)
- if (NOT GNU_SSE4_CFLAG AND NOT MSVC_SSE4_CFLAG)
- # Not surprising if we end up here! MSVC current does not support the SSE4.1 flag. However, it appears to accept SSE4.1
- # intrinsics when SSE2 support is enabled, so we try that instead first.
- if (GMX_NATIVE_WINDOWS)
- GMX_TEST_CFLAG(MSVC_SSE2_CFLAG "/arch:SSE2" ACCELERATION_C_FLAGS)
- message(WARNING "Neither SSE4.1 or SSE2 seems to be supported by your Windows compiler. Something is likely broken.")
- else()
- message(WARNING "No C SSE4.1 flag found. Consider a newer compiler, or use SSE2 for slightly lower performance")
- endif()
- endif(NOT GNU_SSE4_CFLAG AND NOT MSVC_SSE4_CFLAG)
-
- GMX_TEST_CXXFLAG(GNU_SSE4_CXXFLAG "-msse4.1" ACCELERATION_CXX_FLAGS)
- if (NOT GNU_SSE4_CXXFLAG AND GMX_NATIVE_WINDOWS)
- GMX_TEST_CXXFLAG(MSVC_SSE4_CXXFLAG "/arch:SSE4.1" ACCELERATION_CXX_FLAGS)
- endif(NOT GNU_SSE4_CXXFLAG AND GMX_NATIVE_WINDOWS)
- if (NOT GNU_SSE4_CXXFLAG AND NOT MSVC_SSE4_CXXFLAG)
- message(WARNING "No C++ SSE4.1 flag found. Consider a newer compiler, or use SSE2 for slightly lower performance.")
- # Not surprising if we end up here! MSVC current does not support the SSE4.1 flag. However, it appears to accept SSE4.1
- # intrinsics when SSE2 support is enabled, so we try that instead.
- if (GMX_NATIVE_WINDOWS)
- GMX_TEST_CXXFLAG(MSVC_SSE2_CXXFLAG "/arch:SSE2" ACCELERATION_CXX_FLAGS)
- endif()
- endif(NOT GNU_SSE4_CXXFLAG AND NOT MSVC_SSE4_CXXFLAG)
-
- # This must come after we have added the -msse4.1 flag on some platforms.
- check_include_file(smmintrin.h HAVE_SMMINTRIN_H ${ACCELERATION_C_FLAGS})
-
- if(NOT HAVE_SMMINTRIN_H)
- message(FATAL_ERROR "Cannot find smmintrin.h, which is required for SSE4.1 intrinsics support.")
- endif(NOT HAVE_SMMINTRIN_H)
-
- if(CMAKE_C_COMPILER_ID MATCHES "Intel" AND CMAKE_C_COMPILER_VERSION VERSION_EQUAL "11.1")
- message(FATAL_ERROR "You are using Intel compiler version 11.1, and that compiler is known to produce incorrect results with SSE4.1 acceleration. You need to use another compiler (e.g. icc 12 or newer) or different acceleration (probably slower simulations).")
- endif()
-
- set(GMX_CPU_ACCELERATION_X86_SSE4_1 1)
- # The user should not be able to set this orthogonally to the acceleration
- set(GMX_X86_SSE4_1 1)
- set(GMX_X86_SSE2 1)
- set(ACCELERATION_STATUS_MESSAGE
- "Enabling SSE4.1 Gromacs acceleration")
-
-elseif(${GMX_CPU_ACCELERATION} STREQUAL "AVX_128_FMA" OR ${GMX_CPU_ACCELERATION} STREQUAL "AVX_256")
-
- # Set the AVX compiler flag for both these choices!
-
- GMX_TEST_CFLAG(GNU_AVX_CFLAG "-mavx" ACCELERATION_C_FLAGS)
- if (NOT GNU_AVX_CFLAG AND GMX_NATIVE_WINDOWS)
- GMX_TEST_CFLAG(MSVC_AVX_CFLAG "/arch:AVX" ACCELERATION_C_FLAGS)
- endif (NOT GNU_AVX_CFLAG AND GMX_NATIVE_WINDOWS)
- if (NOT GNU_AVX_CFLAG AND NOT MSVC_AVX_CFLAG)
- message(WARNING "No C AVX flag found. Consider a newer compiler, or try SSE4.1 (lower performance) giving the -DGMX_CPU_ACCELERATION=SSE4.1 to cmake.")
- endif (NOT GNU_AVX_CFLAG AND NOT MSVC_AVX_CFLAG)
-
- GMX_TEST_CXXFLAG(GNU_AVX_CXXFLAG "-mavx" ACCELERATION_CXX_FLAGS)
- if (NOT GNU_AVX_CXXFLAG AND GMX_NATIVE_WINDOWS)
- GMX_TEST_CXXFLAG(MSVC_AVX_CXXFLAG "/arch:AVX" ACCELERATION_CXX_FLAGS)
- endif (NOT GNU_AVX_CXXFLAG AND GMX_NATIVE_WINDOWS)
- if (NOT GNU_AVX_CXXFLAG AND NOT MSVC_AVX_CXXFLAG)
- message(WARNING "No C++ AVX flag found. Consider a newer compiler, or try SSE4.1 (lower performance) giving the -DGMX_CPU_ACCELERATION=SSE4.1 to cmake.")
- endif (NOT GNU_AVX_CXXFLAG AND NOT MSVC_AVX_CXXFLAG)
-
- # Set the FMA4 flags (MSVC doesn't require any)
- if(${GMX_CPU_ACCELERATION} STREQUAL "AVX_128_FMA" AND NOT MSVC)
- GMX_TEST_CFLAG(GNU_FMA_CFLAG "-mfma4" ACCELERATION_C_FLAGS)
- if (NOT GNU_FMA_CFLAG)
- message(WARNING "No C FMA4 flag found. Consider a newer compiler, or try SSE4.1 (lower performance).")
- endif(NOT GNU_FMA_CFLAG)
- GMX_TEST_CFLAG(GNU_XOP_CFLAG "-mxop" ACCELERATION_C_FLAGS)
- # No big deal if we do not have xop, so no point yelling warnings about it.
- GMX_TEST_CXXFLAG(GNU_FMA_CXXFLAG "-mfma4" ACCELERATION_CXX_FLAGS)
- if (NOT GNU_FMA_CXXFLAG)
- message(WARNING "No C++ FMA flag found. Consider a newer compiler, or try SSE4.1 (lower performance).")
- endif (NOT GNU_FMA_CXXFLAG)
- GMX_TEST_CXXFLAG(GNU_XOP_CXXFLAG "-mxop" ACCELERATION_CXX_FLAGS)
- # No big deal if we do not have xop, so no point yelling warnings about it.
- endif()
- # Only test the header after we have tried to add the flag for AVX support
- check_include_file(immintrin.h HAVE_IMMINTRIN_H ${ACCELERATION_C_FLAGS})
-
- if(NOT HAVE_IMMINTRIN_H)
- message(FATAL_ERROR "Cannot find immintrin.h, which is required for AVX intrinsics support. Consider switching compiler.")
- endif(NOT HAVE_IMMINTRIN_H)
-
- if(${GMX_CPU_ACCELERATION} STREQUAL "AVX_256")
- try_compile(TEST_AVX ${CMAKE_BINARY_DIR}
- "${CMAKE_SOURCE_DIR}/cmake/TestAVX.c"
- COMPILE_DEFINITIONS "${ACCELERATION_C_FLAGS}")
- if(NOT TEST_AVX)
- message(FATAL_ERROR "Cannot compile AVX intrinsics. Consider switching compiler.")
- endif()
- endif()
-
- # GCC requires x86intrin.h for FMA support. MSVC 2010 requires intrin.h for FMA support.
- check_include_file(x86intrin.h HAVE_X86INTRIN_H ${ACCELERATION_C_FLAGS})
- check_include_file(intrin.h HAVE_INTRIN_H ${ACCELERATION_C_FLAGS})
-
- # The user should not be able to set this orthogonally to the acceleration
- set(GMX_X86_SSE4_1 1)
- set(GMX_X86_SSE2 1)
-
- # But just enable one of the choices internally...
- if(${GMX_CPU_ACCELERATION} STREQUAL "AVX_128_FMA")
- # We don't have the full compiler version string yet (BUILD_C_COMPILER),
- # so we can't distinguish vanilla and Apple clang, but catering for AMD
- # hackintoshes is not worth the effort.
- if (APPLE AND (${CMAKE_C_COMPILER_ID} STREQUAL "Clang" OR
- ${CMAKE_CXX_COMPILER_ID} STREQUAL "Clang"))
- 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 can not work around this bug on OS X, you will have to select a different compiler or CPU acceleration.")
- endif()
-
- if (GMX_USE_CLANG_C_FMA_BUG_WORKAROUND)
- # we assume that we have an external assembler that supports AVX
- message(STATUS "Clang ${CMAKE_C_COMPILER_VERSION} detected, enabling FMA bug workaround")
- set(EXTRA_C_FLAGS "${EXTRA_C_FLAGS} -no-integrated-as")
- endif()
- if (GMX_USE_CLANG_CXX_FMA_BUG_WORKAROUND)
- # we assume that we have an external assembler that supports AVX
- message(STATUS "Clang ${CMAKE_CXX_COMPILER_VERSION} detected, enabling FMA bug workaround")
- set(EXTRA_CXX_FLAGS "${EXTRA_CXX_FLAGS} -no-integrated-as")
- endif()
-
- set(GMX_CPU_ACCELERATION_X86_AVX_128_FMA 1)
- set(GMX_X86_AVX_128_FMA 1)
- set(ACCELERATION_STATUS_MESSAGE
- "Enabling 128-bit AVX Gromacs acceleration (with fused-multiply add)")
-
- else()
- # If we are not doing AVX_128, it must be AVX_256...
- set(GMX_CPU_ACCELERATION_X86_AVX_256 1)
- set(GMX_X86_AVX_256 1)
- set(ACCELERATION_STATUS_MESSAGE
- "Enabling 256-bit AVX Gromacs acceleration")
- endif()
-
- # Unfortunately gcc-4.5.2 and gcc-4.6.0 has a bug where they use the wrong datatype for the formal
- # parameter of the mask for maskload/maskstore arguments. Check if this is present, since we can work around it.
- gmx_test_avx_gcc_maskload_bug(${ACCELERATION_C_FLAGS} GMX_X86_AVX_GCC_MASKLOAD_BUG)
-
-elseif(${GMX_CPU_ACCELERATION} STREQUAL "IBM_QPX")
- try_compile(TEST_QPX ${CMAKE_BINARY_DIR}
- "${CMAKE_SOURCE_DIR}/cmake/TestQPX.c")
+##################################################
+# Process CPU acceleration settings
+##################################################
+# This checks what flags to add in order to
+# support the SIMD instructions we need, and sets
+# correct defines for the acceleration supported.
+include(gmxTestCPUAcceleration)
+gmx_test_cpu_acceleration()
- if (TEST_QPX)
- 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.")
- set(GMX_CPU_ACCELERATION_IBM_QPX 1)
- else()
- 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.")
- endif()
-elseif(${GMX_CPU_ACCELERATION} STREQUAL "SPARC64_HPC_ACE")
- set(GMX_CPU_ACCELERATION_SPARC64_HPC_ACE 1)
-elseif(${GMX_CPU_ACCELERATION} STREQUAL "REFERENCE")
- add_definitions(-DGMX_SIMD_REFERENCE_PLAIN_C)
- if(${GMX_NBNXN_REF_KERNEL_TYPE} STREQUAL "4xn")
- if(${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "2" OR ${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "4" OR ${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "8")
- add_definitions(-DGMX_NBNXN_SIMD_4XN -DGMX_SIMD_REF_WIDTH=${GMX_NBNXN_REF_KERNEL_WIDTH})
- else()
- message(FATAL_ERROR "Unsupported width for 4xn reference kernels")
- endif()
- elseif(${GMX_NBNXN_REF_KERNEL_TYPE} STREQUAL "2xnn")
- if(${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "8" OR ${GMX_NBNXN_REF_KERNEL_WIDTH} STREQUAL "16")
- add_definitions(-DGMX_NBNXN_SIMD_2XNN -DGMX_SIMD_REF_WIDTH=${GMX_NBNXN_REF_KERNEL_WIDTH})
- else()
- message(FATAL_ERROR "Unsupported width for 2xn reference kernels")
- endif()
- else()
- message(FATAL_ERROR "Unsupported kernel type")
- endif()
-else()
- gmx_invalid_option_value(GMX_CPU_ACCELERATION)
-endif()
-gmx_check_if_changed(ACCELERATION_CHANGED GMX_CPU_ACCELERATION)
-if (ACCELERATION_CHANGED AND DEFINED ACCELERATION_STATUS_MESSAGE)
- message(STATUS "${ACCELERATION_STATUS_MESSAGE}")
-endif()
# Process QM/MM Settings
if(${GMX_QMMM_PROGRAM} STREQUAL "GAUSSIAN")
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff