include(GNUInstallDirs)
get_target_property(POPC_EXECUTABLE popc_bin LOCATION)

add_library(popnn SHARED
  codelets.cpp
  CTCInference.cpp
  CTCInferenceConnection.cpp
  CTCInferenceConnection.hpp
  CTCLoss.cpp
  CTCInferencePlan.cpp
  CTCInferencePlan.hpp
  CTCLossPlan.cpp
  CTCLossPlan.hpp
  CTCPlanInternal.cpp
  CTCPlanInternal.hpp
  BatchNorm.cpp
  GroupNorm.cpp
  LogSoftmax.cpp
  Loss.cpp
  Lstm.cpp
  Rnn.cpp
  Gru.cpp
  HardSigmoid.cpp
  Norms.cpp
  NormsInternal.cpp
  NormsInternal.hpp
  Pooling.cpp
  PoolOptions.hpp
  PoolPlan.cpp
  PoolPlan.hpp
  PoolVertices.cpp
  PoolVertices.hpp
  popnnCycleEstimators.cpp
  NonLinearity.cpp
  PerformanceEstimation.hpp
  Recurrent.cpp
  SpatialSoftMax.cpp
  ROIAlign.cpp
  ${CMAKE_SOURCE_DIR}/include/popnn/codelets.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/BatchNorm.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/CTCPlan.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/CTCInference.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/CTCLoss.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/GroupNorm.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/InstanceNorm.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/LayerNorm.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/LogSoftmax.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/Loss.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/LstmDef.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/Lstm.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/GruDef.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/Gru.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/NonLinearityDef.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/NonLinearityDefUtil.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/NonLinearity.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/Norms.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/PoolingDef.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/Pooling.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/Recurrent.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/SpatialSoftMax.hpp
  ${CMAKE_SOURCE_DIR}/include/popnn/experimental/ROIAlign.hpp
  ${CMAKE_SOURCE_DIR}/include/poplibs_support/Compiler.hpp
  ${CMAKE_SOURCE_DIR}/include/poplibs_support/VectorUtils.hpp
)

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

target_link_libraries(popnn
  PUBLIC
    poplar poplin poputil ${CMAKE_DL_LIBS}
  PRIVATE
    popops
    popsolver
    poplibs_support
    Boost::boost
    TBB::tbb
    spdlog::spdlog_header_only
)

target_include_directories(popnn
  PUBLIC
    $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
    $<INSTALL_INTERFACE:include>
  PRIVATE
    # Needed for popsparse to access ConvPlan Constraints.
    $<TARGET_PROPERTY:poplin,INCLUDE_DIRECTORIES>
    .
)

set(codelet_asm_sources
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/CTCInferenceSort.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/CTCInferenceUpdate.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/LossTransform.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/Pooling.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/ReduceClassGather.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/ReduceClassSparse.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearity2D_gelu.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearity2DSwish.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearityGrad2D.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearityGrad2D_gelu.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearityGrad2DSwish.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearity1D_gelu.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearity1DSwish.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearityGrad1D.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearityGrad1D_gelu.S
  ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearityGrad1DSwish.S
)

add_gp_library(
  NAME popnn
  ASM_SOURCES
    ${codelet_asm_sources}
  CPP_SOURCES
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/CalcAccuracy.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/CTCLoss.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/CTCInference.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/LossCrossEntropyTransform.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/LossSumSquaredTransform.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/MaxPooling.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/MaxPoolingGrad.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/MaxPoolingGradientScale.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/NonLinearity2D.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/NonLinearityGrad2D.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/NonLinearityGrad1D.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/NonLinearity1D.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/ReduceMaxClassGather.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/ReduceMaxClassSparse.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/ReduceMaxNClassGather.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/ReduceMaxNClassSparse.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/ReduceMinClassGather.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/ReduceMinClassSparse.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SelectiveScaling.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/SumPooling.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/ROIAlign.cpp
  HEADERS
    PerformanceEstimation.hpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/LogOps.hpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/MinHeapView.hpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/NonLinearity.hpp
    ${CMAKE_SOURCE_DIR}/include/popops/EncodingConstants.hpp
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearitySwishCommon.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearityLoop_gelu.S
    ${CMAKE_CURRENT_SOURCE_DIR}/codelets/asm/NonLinearityGradLoop_gelu.S
    ${CMAKE_SOURCE_DIR}/include/poplibs_support/CTCInferenceDefs.hpp
)

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

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

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