vtk-m/vtkm/internal/Configure.h.in
Robert Maynard 3eec5e86df ICC: disable vectorization as both ivdep and simd generate bad code.
We are disabling the entire vectorization hints for ICC as it generates
both bad code, and dramatically decreases compile time.

The compiler does not check for aliasing or dependencies that might cause
incorrect results after vectorization, and it does not protect against illegal
memory references. #pragma ivdep overrides potential dependencies, but the
compiler still performs a dependency analysis, and will not vectorize if it
finds a proven dependency that would affect results. With #pragma simd, the
compiler does no such analysis, and tries to vectorize regardless.
2016-03-23 14:10:18 -04:00

264 lines
9.2 KiB
C

//============================================================================
// Copyright (c) Kitware, Inc.
// All rights reserved.
// See LICENSE.txt for details.
// This software is distributed WITHOUT ANY WARRANTY; without even
// the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR
// PURPOSE. See the above copyright notice for more information.
//
// Copyright 2014 Sandia Corporation.
// Copyright 2014 UT-Battelle, LLC.
// Copyright 2014 Los Alamos National Security.
//
// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
// the U.S. Government retains certain rights in this software.
//
// Under the terms of Contract DE-AC52-06NA25396 with Los Alamos National
// Laboratory (LANL), the U.S. Government retains certain rights in
// this software.
//============================================================================
#ifndef vtk_m_internal_Configure_h
#define vtk_m_internal_Configure_h
#ifdef __CUDACC__
#define VTKM_CUDA
#endif
#ifdef _OPENMP
#define VTKM_OPENMP
#endif
#if defined(_MSC_VER)
//MSVC 2015+ can use a clang frontend, so we want to label it only as MSVC
//and not MSVC and clang
#define VTKM_MSVC
#elif defined(__INTEL_COMPILER)
//Intel 14+ on OSX uses a clang frontend, so again we want to label them as
//intel only, and not intel and clang
#define VTKM_ICC
#elif defined(__PGI)
// PGI reports as GNUC as it generates the same ABI, so we need to check for
// it before gcc.
#define VTKM_PGI
#elif defined(__clang__)
//Check for clang before GCC, as clang says it is GNUC since it has ABI
//compliance
#define VTKM_CLANG
#elif defined(__GNUC__)
// Several compilers pretend to be GCC but have minor differences. Try to
// compensate for that, by checking for those compilers first
#define VTKM_GCC
#endif
#if defined(unix) || defined(__unix) || defined(__unix__) || (defined(__APPLE__) && defined(__MACH__))
#include <unistd.h>
# ifdef _POSIX_VERSION
# define VTKM_POSIX _POSIX_VERSION
# endif
#endif
#if !defined(VTKM_USE_DOUBLE_PRECISION) && !defined(VTKM_NO_DOUBLE_PRECISION)
#cmakedefine VTKM_USE_DOUBLE_PRECISION
#endif
#if defined(VTKM_USE_DOUBLE_PRECISION) && defined(VTKM_NO_DOUBLE_PRECISION)
# error Both VTKM_USE_DOUBLE_PRECISION and VTKM_NO_DOUBLE_PRECISION defined. Do not know what to do.
#endif
#if !defined(VTKM_USE_64BIT_IDS) && !defined(VTKM_NO_64BIT_IDS)
#cmakedefine VTKM_USE_64BIT_IDS
#endif
#if defined(VTKM_USE_64BIT_IDS) && defined(VTKM_NO_64BIT_IDS)
# error Both VTKM_USE_64BIT_IDS and VTKM_NO_64BIT_IDS defined. Do not know what to do.
#endif
#define VTKM_SIZE_FLOAT @VTKm_SIZE_FLOAT@
#define VTKM_SIZE_DOUBLE @VTKm_SIZE_DOUBLE@
#define VTKM_SIZE_CHAR @VTKm_SIZE_CHAR@
#define VTKM_SIZE_SHORT @VTKm_SIZE_SHORT@
#define VTKM_SIZE_INT @VTKm_SIZE_INT@
#define VTKM_SIZE_LONG @VTKm_SIZE_LONG@
#define VTKM_SIZE_LONG_LONG @VTKm_SIZE_LONG_LONG@
#ifdef VTKM_USE_DOUBLE_PRECISION
# ifndef VTKM_SIZE_SCALAR
# define VTKM_SIZE_SCALAR VTKM_SIZE_DOUBLE
# endif
# ifndef VTKM_ALIGNMENT_TWO_SCALAR
# define VTKM_ALIGNMENT_TWO_SCALAR 16
# endif
# ifndef VTKM_ALIGNMENT_FOUR_SCALAR
# define VTKM_ALIGNMENT_FOUR_SCALAR 8
# endif
#else
# ifndef VTKM_SIZE_SCALAR
# define VTKM_SIZE_SCALAR VTKM_SIZE_FLOAT
# define VTKM_ALIGNMENT_SCALAR VTKM_SIZE_SCALAR
# endif
# ifndef VTKM_ALIGNMENT_TWO_SCALAR
# define VTKM_ALIGNMENT_TWO_SCALAR 8
# endif
# ifndef VTKM_ALIGNMENT_FOUR_SCALAR
# define VTKM_ALIGNMENT_FOUR_SCALAR 16
# endif
#endif
#ifdef VTKM_USE_64BIT_IDS
# ifndef VTKM_SIZE_ID
# define VTKM_SIZE_ID 8
# endif
#else
# ifndef VTKM_SIZE_ID
# define VTKM_SIZE_ID 4
# endif
#endif
// Define a pair of macros, VTKM_THIRDPARTY_PRE_INCLUDE and VTKM_THIRDPARTY_POST_INCLUDE,
// that should be wrapped around any #include for a boost or thrust header file. Mostly
// this is used to set pragmas that dissable warnings that VTK-m checks for
// but boost and thrust does not.
#if (defined(VTKM_GCC) || defined(VTKM_CLANG))
#define VTK_M_THIRDPARTY_GCC_WARNING_PRAGMAS \
_Pragma("GCC diagnostic ignored \"-Wconversion\"") \
_Pragma("GCC diagnostic ignored \"-Wshadow\"") \
_Pragma("GCC diagnostic ignored \"-Wunused-parameter\"")
// Newer versions of clang have an unused-local-typedef warning, but not older
// versions. This checks for the apple version of clang, which is different
// than other clang compiled versions. If using a non-apple version of clang,
// you might need to extend this condition.
#if defined(VTKM_CLANG) && (__apple_build_version__ >= 7000072)
#define VTK_M_THIRDPARTY_CLANG_WARNING_PRAGMAS \
_Pragma("GCC diagnostic ignored \"-Wunused-local-typedef\"")
#else
#define VTK_M_THIRDPARTY_CLANG_WARNING_PRAGMAS
#endif
// Older versions of GCC don't support the push/pop pragmas. Right now we are
// not checking for GCC 3 or earlier. I'm not sure we have a use case for that.
#if defined(VTKM_GCC) && (__GNUC__ == 4 && __GNUC_MINOR__ < 6)
#define VTK_M_THIRDPARTY_WARNINGS_PUSH
#define VTK_M_THRIDPARTY_WARNINGS_POP
#else
#define VTK_M_THIRDPARTY_WARNINGS_PUSH _Pragma("GCC diagnostic push")
#define VTK_M_THRIDPARTY_WARNINGS_POP _Pragma("GCC diagnostic pop")
#endif
#define VTKM_THIRDPARTY_PRE_INCLUDE \
VTK_M_THIRDPARTY_WARNINGS_PUSH \
VTK_M_THIRDPARTY_GCC_WARNING_PRAGMAS \
VTK_M_THIRDPARTY_CLANG_WARNING_PRAGMAS
#define VTKM_THIRDPARTY_POST_INCLUDE \
VTK_M_THRIDPARTY_WARNINGS_POP
#elif (defined(VTKM_MSVC))
#define VTKM_THIRDPARTY_PRE_INCLUDE \
__pragma(warning(push)) \
__pragma(warning(disable:4267))
#define VTKM_THIRDPARTY_POST_INCLUDE \
__pragma(warning(pop))
#else
#define VTKM_THIRDPARTY_PRE_INCLUDE
#define VTKM_THIRDPARTY_POST_INCLUDE
#endif
//Determine if current compiler supports vectorization pragma's
//if so set the define VTKM_COMPILER_SUPPORTS_VECTOR_PRAGMAS
//
//While ICC 14 does support simd pragma, in testing we find that the
//simd code produces SIGBUS and isn't safe to use.
//
#if ( defined(VTKM_GCC) && ( __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9) ) ) || \
( defined(VTKM_ICC) && (__INTEL_COMPILER >= 1500) ) || \
( defined(VTKM_CLANG) && defined(__apple_build_version__) && (__apple_build_version__ >= 7000000) ) || \
( defined(VTKM_CLANG) && !defined(__apple_build_version__) && (__clang_major__ > 3) ) || \
( defined(VTKM_CLANG) && !defined(__apple_build_version__) && (__clang_major__ == 3 && __clang_minor__ >= 5) )
#define VTKM_COMPILER_SUPPORTS_VECTOR_PRAGMAS 1
#endif
// Define a pair of macros, VTKM_VECTORIZATION_PRE_LOOP and VTKM_VECTORIZATION_IN_LOOP,
// that should be wrapped around any "for"/"while" that you want vectorized.
// This is used to set per compiler pragmas for vectorization, and to disable
// any warnings that about vectorization failures.
#if defined(VTKM_COMPILER_SUPPORTS_VECTOR_PRAGMAS)
#if defined(VTKM_CLANG)
//clang only needs pre loop
#define VTKM_VECTORIZATION_PRE_LOOP \
_Pragma("clang loop vectorize(enable) interleave(enable)")
#define VTKM_VECTORIZATION_IN_LOOP
#elif defined(VTKM_ICC) && defined(NDEBUG)
//Note: icc can't do vectorization in debug builds
//icc needs pre and in loop
//For ICC we want ivdep over simd for the following reason:
//#pragma simd is a more powerful combo of '#pragma vector always' and '#pragma ivdep'
//The compiler does not check for aliasing or dependencies that might cause
//incorrect results after vectorization, and it does not protect against illegal
//memory references. #pragma ivdep overrides potential dependencies, but the
//compiler still performs a dependency analysis, and will not vectorize if it
//finds a proven dependency that would affect results. With #pragma simd, the
//compiler does no such analysis, and tries to vectorize regardless.
//
//Final: We are currently disabling all vectorization with ICC
// in the short-term. Both ivdep and simd cause a horrible decrease in compile
// time, and generates bad vectorization code.
#define VTKM_VECTORIZATION_PRE_LOOP
#define VTKM_VECTORIZATION_IN_LOOP
#elif defined(VTKM_GCC)
//gcc only needs in loop
#define VTKM_VECTORIZATION_PRE_LOOP \
_Pragma("GCC ivdep")
#define VTKM_VECTORIZATION_IN_LOOP
#else
// Compiler is unknown so we don't define any vectortization pragmas
#define VTKM_VECTORIZATION_PRE_LOOP
#define VTKM_VECTORIZATION_IN_LOOP
#endif
#else
// Compiler doesn't support any vectortization pragmas
#define VTKM_VECTORIZATION_PRE_LOOP
#define VTKM_VECTORIZATION_IN_LOOP
#endif
//Mark if we are building with CUDA enabled
#ifndef VTKM_ENABLE_CUDA
#cmakedefine VTKM_ENABLE_CUDA
#endif
//Mark if we are building with TBB enabled
#ifndef VTKM_ENABLE_TBB
#cmakedefine VTKM_ENABLE_TBB
#endif
//Mark if we are building with interop enabled
#ifndef VTKM_ENABLE_OPENGL_INTEROP
#cmakedefine VTKM_ENABLE_OPENGL_INTEROP
#endif
// Determine whether we will use variadic templates (a new feature in C++11).
// Currently have VARIADIC_TEMPLATE support off.
#cmakedefine VTKM_NO_VARIADIC_TEMPLATE
#if !defined(VTKM_USE_VARIADIC_TEMPLATE) && !defined(VTKM_NO_VARIADIC_TEMPLATE)
// Currently using Boost to determine support.
VTKM_THIRDPARTY_PRE_INCLUDE
# include <boost/config.hpp>
VTKM_THIRDPARTY_POST_INCLUDE
# if defined(BOOST_HAS_VARIADIC_TMPL)
# define VTKM_USE_VARIADIC_TEMPLATE 1
# endif
#endif
#if defined(VTKM_USE_VARIADIC_TEMPLATE) && defined(VTKM_NO_VARIADIC_TEMPLATE)
# error Both VTKM_USE_VARIADIC_TEMPLATE and VTKM_NO_VARIADIC_TEMPLATE defined. Do not know what to do.
#endif
#endif //vtkm_internal_Configure_h