include(GNUInstallDirs)

get_target_property(POPC_EXECUTABLE popc_bin LOCATION)

add_library(popsparse SHARED
  codelets.cpp
  BlockSparseMatMul.cpp
  BSMatrix.cpp
  HyperGraph.cpp
  HyperGraphBlock.cpp
  HyperGraphStripV0.cpp
  HyperGraphStrip.cpp
  HyperGraphBlockZoltan.cpp
  HyperGraphBlockNaive.cpp
  HyperGraphBlockGroup.cpp
  HyperGraphBlockGroup2.cpp
  HyperGraphPartitioner.cpp
  ZoltanPartitioner.cpp
  BalancedPartitioner.cpp
  BSOps.cpp
  BSUtils.cpp
  BSNonLinearity.cpp
  Embedding.cpp
  FullyConnected.cpp
  FullyConnectedOnTile.hpp
  FullyConnectedOnTile.cpp
  FullyConnectedOptions.hpp
  FullyConnectedOptions.cpp
  FullyConnectedParams.cpp
  FullyConnectedPlan.cpp
  FullyConnectedPlan.hpp
  FullyConnectedPNMapping.cpp
  FullyConnectedPNMapping.hpp
  FullyConnectedTensorMetaData.hpp
  FullyConnectedUtils.hpp
  FullyConnectedUtils.cpp
  FullyConnectedVector.hpp
  MatMul.cpp
  MatMulOptions.cpp
  MatMulOptions.hpp
  MatMulParams.cpp
  MatMulTensorMetaData.hpp
  MatMulUtils.cpp
  MatMulUtils.hpp
  PerformanceEstimation.hpp
  PlanningCacheImpl.hpp
  PlanningCache.cpp
  popsparseCycleEstimators.hpp
  popsparseCycleEstimators.cpp
  SparseCodeletMetaInfoScale.hpp
  SparseMetaInfo.hpp
  SparseStorageInternal.hpp
  SparseFormatsValidate.cpp
  SparseFormatsValidate.hpp
  SparsityParams.cpp
  StaticMatMul.cpp
  StaticMatMulPartitioner.cpp
  StaticMatMulPartitioner.hpp
  SparseTensor.cpp
  SparsePartitioner.cpp
  SparsePartitionerImpl.cpp
  SparsePartitionerImpl.hpp
  SparsePartitionerOptions.cpp
  SparsePartitionerOptions.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/codelets.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/experimental/BlockSparseMatMul.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/Embedding.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/FullyConnectedParams.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/FullyConnected.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/MatMul.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/MatMulParams.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/PlanningCache.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/SparseTensor.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/SparsityParams.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/SparseStorageFormats.hpp
  ${CMAKE_SOURCE_DIR}/include/popsparse/SparsePartitioner.hpp
  )

# Enable auto vectorization and common microarchitecture extensions available
# in modern x86-64 processors
TargetAddMarchExtensions(popsparse)

target_link_libraries(popsparse
  PUBLIC
    poplar poputil popops poplin ${CMAKE_DL_LIBS} ZOLTAN::zoltan ZOLTAN::simpi
  PRIVATE
    poplibs_support
    popsolver
    Boost::boost
)

target_include_directories(popsparse
  PUBLIC
    $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
    $<INSTALL_INTERFACE:include>
  PRIVATE
    # Needed for popsparse to access ConvReduce functionality for the timebeing
    $<TARGET_PROPERTY:poplin,INCLUDE_DIRECTORIES>
    $<TARGET_PROPERTY:poputil,INCLUDE_DIRECTORIES>
    .
)

add_gp_library(
  NAME popsparse
  CPP_SOURCES
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/StaticSparseDenseElementWise.cpp  
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/StaticSparseDenseMatMulBlock.cpp  
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMatMulElementWise.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMatMulElementWiseTranspose.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMatMulGradWElementWise.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMatMulGradAElementWise.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMatMulBlock.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMatMulBlockGradA.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMatMulBlockGradW.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMatMulBlockAmpGradW.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/BlockTransposeGradW.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMultiSliceBlock.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseDenseMultiSliceElementWise.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SparseGather.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/Utils.cpp

  ASM_SOURCES
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticSparseDenseElementWise_half_half.S  
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticSparseDenseElementWise_float_float.S  
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticBlock4x4SparseDenseMatMul_half_half.S  
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticBlock8x8SparseDenseMatMul_half_half.S  
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticBlock16x16SparseDenseMatMul_half_half.S  
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticBlock4x4SparseDenseMatMul_float_float.S     
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticBlock8x8SparseDenseMatMul_float_float.S     
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticBlock16x8SparseDenseMatMul_float_float.S     
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulElementWise_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulElementWise_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseGatherElementWise.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMultiSliceBlock.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMultiSliceElementWise.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMultiUpdateAddElementWise.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMultiUpdateAddBlock.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulGradWElementWise_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulGradWElementWise_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulGradAElementWise_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulGradAElementWise_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulTranspElementWise_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulTranspElementWise_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMul_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMul_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMul_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMul_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMul_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMul_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMul_half_half_2ampsets.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMul_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMul_half_half.S

    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulGradA_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulGradA_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulGradA_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulGradA_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulGradA_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMulGradA_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMulGradA_half_half_2ampsets.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMulGradA_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulGradA_half_half.S

    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockMatMulGradWCommon.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulGradW_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulGradW_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulGradW_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulGradW_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulGradW_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulGradW_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMulGradW_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMulGradW_half_float.S

    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMulAmpGradW_half_half_2ampsets.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMulAmpGradW_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block16x16SparseDenseMatMulAmpGradW_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulAmpGradW_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulAmpGradW_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block8x8SparseDenseMatMulAmpGradW_float_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulAmpGradW_half_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulAmpGradW_half_float.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Block4x4SparseDenseMatMulAmpGradW_float_float.S

    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockTransposeGradW_half.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockTransposeGradW_float.S

  HEADERS
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/StaticBlockSparseDenseMatMul.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulGradAElementWise.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulElementWise.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulTranspElementWise.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulGradWElementWise.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/SparseDenseMatMulStructs.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockSparseDenseMatMul.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockSparseMatMulStructs.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockSparseMatMulStructs.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockSparseMatMulGradW.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockSparseMatMulAmpGradW.h.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/BlockTransposeGradW.h.S

    ${CMAKE_CURRENT_SOURCE_DIR}/SparseMetaInfo.hpp
    ${CMAKE_CURRENT_SOURCE_DIR}/SparseCodeletMetaInfoScale.hpp
)

install(TARGETS popsparse
        EXPORT popsparse
        DESTINATION ${CMAKE_INSTALL_LIBDIR}
        COMPONENT popsparse
        INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})

install(EXPORT popsparse
        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/popsparse
        FILE popsparse-targets.cmake
        COMPONENT popsparse)

install(DIRECTORY ${PROJECT_SOURCE_DIR}/include/popsparse
        DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
        COMPONENT popsparse)

install(DIRECTORY ${PROJECT_SOURCE_DIR}/include/popsparse/experimental
        DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/popsparse
        COMPONENT popsparse)