CMakeLists.txt

project( spamm-chunk C ASM Fortran )
cmake_minimum_required( VERSION 2.8 )

include( FindOpenMP )
include( CheckCSourceCompiles )
include( CheckCCompilerFlag )
include( CheckFunctionExists )

if( NOT CMAKE_BUILD_TYPE )
  set( CMAKE_BUILD_TYPE Debug )
  message( STATUS "Setting default build type to ${CMAKE_BUILD_TYPE}" )
endif()

if( CMAKE_Fortran_COMPILER_ID STREQUAL "GNU" )
  set( TEMP_FLAGS
    "-O0 -g -gdwarf-3"
    "-std=f2008"
    "-fvar-tracking-assignments -fcheck=all"
    "-ffree-line-length-none"
    "-Wall -Wimplicit-interface -Wuninitialized -Wimplicit-procedure"
    )
  string( TOUPPER ${CMAKE_BUILD_TYPE} UPPER_BUILD_TYPE )
  if( UPPER_BUILD_TYPE STREQUAL "DEBUG" )
    try_compile( HAVE_ADDRESS_SANITIZER
      ${CMAKE_BINARY_DIR}/CMakeTmp
      ${CMAKE_SOURCE_DIR}/cmake-tests/address-sanitizer
      address-sanitizer
      OUTPUT_VARIABLE ADDRESS_SANITIZER_OUTPUT
      )
    if( HAVE_ADDRESS_SANITIZER )
      message( STATUS "Adding address sanitizer compiler flag" )
      list( APPEND TEMP_FLAGS "-fsanitize=address" )
    else()
      message( STATUS "Address sanitizer not supported" )
    endif()
  endif()
  string( REGEX REPLACE ";" " " CMAKE_Fortran_FLAGS_DEBUG "${TEMP_FLAGS}" )
  set( CMAKE_Fortran_FLAGS_RELEASE "-Ofast -g -ffree-line-length-none" )
elseif( CMAKE_Fortran_COMPILER_ID STREQUAL "Intel" )
  set( CMAKE_Fortran_FLAGS_DEBUG
    "-stand f08 -O0 -g"
    "-extend-source"
    "-debug all -check all"
    "-warn unused -traceback"
    )
  string( REGEX REPLACE ";" " " CMAKE_Fortran_FLAGS_DEBUG "${CMAKE_Fortran_FLAGS_DEBUG}" )
  set( CMAKE_Fortran_FLAGS_RELEASE "-O3 -no-prec-div -fp-model fast=2 -vec-report -g" )
else()
  message( STATUS "Unknown Fortran compiler ${CMAKE_Fortran_COMPILER_ID}" )
endif()

check_c_compiler_flag( "-std=gnu99" C99_CONFORM )
if( NOT C99_CONFORM )
  message( WARNING "Can not set C compiler to C99 standard" )
else()
  set( STDC "-std=gnu99" )
endif()

include( FindDoxygen )
if( DOXYGEN_DOT_FOUND )
  set( HAVE_DOT "YES" )
else()
  set( HAVE_DOT "NO" )
endif()

configure_file( Doxyfile.in Doxyfile )

if( DOXYGEN_FOUND )
  add_custom_target( doc
    COMMAND ${DOXYGEN_EXECUTABLE}
    )
endif()

check_function_exists( clock_gettime FOUND_GETTIME )
if( NOT FOUND_GETTIME )
  message( WARNING "Could not find gettime" )
endif()

check_function_exists( ceil FOUND_CEIL )
if( NOT FOUND_CEIL )
  # Try linking with m.
  set( CMAKE_REQUIRED_LIBRARIES "-lm" )
  check_function_exists( ceil FOUND_CEIL_WITH_M )
  if( NOT FOUND_CEIL_WITH_M )
    message( FATAL_ERROR "Could not find ceil() function" )
  endif()
  set( MATHLIB "-lm" )
endif()

# Check for supported clocks. Thanks to Jonathan Lifflander
# <jliffl2@illinois.edu> for suggesting this test.
foreach( CLOCKTYPE_
    CLOCK_MONOTONIC_RAW
    CLOCK_MONOTONIC
    CLOCK_REALTIME )
  check_c_source_compiles(
    "#include <time.h>
    int main () {
    struct timespec p;
    clock_gettime(${CLOCKTYPE_}, &p);
    return 0;
    }"
    TEST_COMPILED )
  if( TEST_COMPILED )
    message( STATUS "Setting clocktype to ${CLOCKTYPE_}" )
    set( CLOCKTYPE ${CLOCKTYPE_}
      CACHE STRING "The default clock type for clock_gettime()" )
    break()
  endif()
endforeach()

if( NOT TEST_COMPILED )
  message( FATAL_ERROR "could not find a suitable clock type" )
endif()

set( ENABLE_MIC FALSE
  CACHE BOOL "Enable Xeon Phi target" )
if( ENABLE_MIC )
  set( MIC_ALLOC TRUE )
  set( MIC_CFLAGS "-mmic" )
  set( MIC_LDFLAGS "-static" )
endif()

set( FORCE_OMP_NUM_THREADS FALSE
  CACHE STRING
  "Set the number of threads in OpenMP parallel sections to N" )

set( CHUNK_IMPLEMENTATION tree
  CACHE STRING "The chunk implementation (tiled, tree), default = tree" )

set( BLOCK_MULTIPLY block
  CACHE STRING
  "The dense multiply implementation (block, blas), default = block" )

if( BLOCK_MULTIPLY STREQUAL "block" )
  set( BLOCK_MULTIPLY TRUE )
elseif( BLOCK_MULTIPLY STREQUAL "blas" )
  set( BLOCK_BLAS TRUE )
else()
  message( FATAL_ERROR "Unknown multiply implementation" )
endif()

set( CHUNK_TREE_MAX_TIER 6
  CACHE STRING
  "The maximum tier at which new tasks are created when recursing the chunk
  tree" )

set( CHUNK_BLOCK_TRANSPOSE 32
  CACHE STRING
  "The minimum size N of the block matrix that will transpose a matrix to
  allow for loop vectorization" )

set( BLOCK_SLEEP 0
  CACHE STRING "Number of nanoseconds to sleep after block multiply" )

set( MEASURE_COMPLEXITY FALSE
  CACHE BOOL "Measure the product complexity" )

set( DEBUG_OUTPUT FALSE
  CACHE BOOL "Enable a lot of extra output" )

set( DEBUG_OUTPUT_MATRIX FALSE
  CACHE BOOL "Print the matrices for debugging" )

set( CONFIG_BLOCK_SIZE 4
  CACHE STRING "The size of the basic dense sub-matrix blocks" )

set( CONFIG_CHUNK_SIZE 32
  CACHE STRING "The size of the chunk sub-matrix" )

set( CONFIG_REAL_KIND "double"
  CACHE STRING "The real kind (single, double)" )
if( CONFIG_REAL_KIND STREQUAL "single" )
  set( REAL_KIND_EXPRESSION "kind(1e0)" )
elseif( CONFIG_REAL_KIND STREQUAL "double" )
  set( REAL_KIND_EXPRESSION "kind(1d0)" )
else()
  message( FATAL_ERROR "CONFIG_REAL_KIND = [ single, double ]" )
endif()

configure_file( config.h.in config.h )

enable_testing()

add_subdirectory( src )
add_subdirectory( tests )