CMakeLists.txt

set(kernel_declare_file
    ${PADDLE_BINARY_DIR}/paddle/phi/kernels/declarations.h.tmp
    CACHE INTERNAL "declarations.h file")
set(kernel_declare_file_final
    ${PADDLE_BINARY_DIR}/paddle/phi/kernels/declarations.h)
file(
  WRITE ${kernel_declare_file}
  "// Generated by the paddle/phi/kernels/CMakeLists.txt.  DO NOT EDIT!\n\n#pragma once\n\n"
)
file(APPEND ${kernel_declare_file}
     "#include \"paddle/phi/core/kernel_registry.h\"\n\n")

# phi functors and functions called by kernels
add_subdirectory(funcs)

# kernel autotune
add_subdirectory(autotune)

# phi depends all phi kernel targets
set_property(GLOBAL PROPERTY PHI_KERNELS "")

# [ 1. Common kernel compilation dependencies ]
set(COMMON_KERNEL_DEPS
    dense_tensor
    sparse_coo_tensor
    sparse_csr_tensor
    kernel_context
    kernel_factory
    arg_map_context
    convert_utils
    lod_utils
    custom_kernel)
set(COMMON_KERNEL_DEPS
    ${COMMON_KERNEL_DEPS}
    eigen_function
    blas
    math_function
    im2col
    vol2col
    concat_and_split_functor
    selected_rows_functor)
# remove this dep after removing fluid deps on tensor creation
set(COMMON_KERNEL_DEPS ${COMMON_KERNEL_DEPS} phi_api_utils)
set(COMMON_KERNEL_DEPS ${COMMON_KERNEL_DEPS} infermeta)

# [ 2. Kernels that most kernels depend on ]
# There are a few kernels that are very basic operations, and most of the
# kernels depend on these kernels.
set(COMMON_BAISC_KERNELS empty_kernel full_kernel)
kernel_library(empty_kernel DEPS ${COMMON_KERNEL_DEPS})
kernel_library(full_kernel DEPS ${COMMON_KERNEL_DEPS} empty_kernel)

# [ 3. Kernels with special dependencies ]
# Some kernels depend on some targets that are not commonly used.
# These targets are not suitable for common dependencies.
# In this case, you need to manually generate them here.
set(AUTOTUNE_KERNELS conv_kernel conv_grad_kernel conv_grad_grad_kernel
                     conv_transpose_kernel conv_transpose_grad_kernel)
set(MANUAL_BUILD_KERNELS
    ${AUTOTUNE_KERNELS}
    cross_entropy_kernel
    adam_kernel
    adamw_kernel
    deformable_conv_kernel
    deformable_conv_grad_kernel
    eigh_kernel
    gumbel_softmax_kernel
    gumbel_softmax_grad_kernel
    hierarchical_sigmoid_kernel
    hierarchical_sigmoid_grad_kernel
    matrix_power_kernel
    matrix_power_grad_kernel
    maxout_kernel
    maxout_grad_kernel
    pool_kernel
    put_along_axis_kernel
    put_along_axis_grad_kernel
    segment_pool_kernel
    segment_pool_grad_kernel
    softmax_kernel
    softmax_grad_kernel
    take_along_axis_kernel
    take_along_axis_grad_kernel
    triangular_solve_grad_kernel
    determinant_grad_kernel
    reduce_sum_kernel
    reduce_mean_kernel
    rnn_kernel
    rnn_grad_kernel
    warpctc_kernel
    warpctc_grad_kernel)
foreach(src ${AUTOTUNE_KERNELS})
  kernel_library(${src} DEPS ${COMMON_KERNEL_DEPS} switch_autotune)
endforeach()
kernel_library(
  adam_kernel
  DEPS
  gflags
  glog
  flags
  ${COMMON_KERNEL_DEPS}
  selected_rows_functor
  threadpool
  jit_kernel_helper)
kernel_library(adamw_kernel DEPS ${COMMON_KERNEL_DEPS} adam_kernel)
kernel_library(cross_entropy_kernel DEPS ${COMMON_KERNEL_DEPS} softmax
               cross_entropy)
kernel_library(deformable_conv_kernel DEPS ${COMMON_KERNEL_DEPS}
               deformable_conv_functor)
kernel_library(deformable_conv_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               deformable_conv_functor)
kernel_library(determinant_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               matrix_inverse)
kernel_library(eigh_kernel DEPS ${COMMON_KERNEL_DEPS} lapack_function)
kernel_library(hierarchical_sigmoid_kernel DEPS ${COMMON_KERNEL_DEPS}
               matrix_bit_code)
kernel_library(hierarchical_sigmoid_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               matrix_bit_code)
kernel_library(gumbel_softmax_kernel DEPS ${COMMON_KERNEL_DEPS} softmax)
kernel_library(gumbel_softmax_grad_kernel DEPS ${COMMON_KERNEL_DEPS} softmax)
kernel_library(reduce_sum_kernel DEPS ${COMMON_KERNEL_DEPS} cast_kernel)
kernel_library(reduce_mean_kernel DEPS ${COMMON_KERNEL_DEPS} cast_kernel)
kernel_library(matrix_power_kernel DEPS ${COMMON_KERNEL_DEPS} matrix_inverse)
kernel_library(matrix_power_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               matrix_inverse)
kernel_library(maxout_kernel DEPS ${COMMON_KERNEL_DEPS} maxouting)
kernel_library(maxout_grad_kernel DEPS ${COMMON_KERNEL_DEPS} maxouting)
kernel_library(pool_kernel DEPS ${COMMON_KERNEL_DEPS} pooling)
kernel_library(put_along_axis_kernel DEPS ${COMMON_KERNEL_DEPS}
               gather_scatter_kernel)
kernel_library(put_along_axis_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               gather_scatter_kernel)
kernel_library(segment_pool_kernel DEPS ${COMMON_KERNEL_DEPS} segment_pooling)
kernel_library(segment_pool_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               segment_pooling)
kernel_library(softmax_kernel DEPS ${COMMON_KERNEL_DEPS} softmax)
kernel_library(softmax_grad_kernel DEPS ${COMMON_KERNEL_DEPS} softmax)
kernel_library(take_along_axis_kernel DEPS ${COMMON_KERNEL_DEPS}
               gather_scatter_kernel)
kernel_library(take_along_axis_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               gather_scatter_kernel)
kernel_library(triangular_solve_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               matrix_reduce)
kernel_library(rnn_kernel DEPS ${COMMON_KERNEL_DEPS} concat_and_split_functor
               lstm_compute gru_compute)
kernel_library(rnn_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               concat_and_split_functor lstm_compute gru_compute)
kernel_library(warpctc_kernel DEPS ${COMMON_KERNEL_DEPS} phi_dynload_warpctc
               sequence_padding sequence_scale)
kernel_library(warpctc_grad_kernel DEPS ${COMMON_KERNEL_DEPS}
               phi_dynload_warpctc sequence_padding sequence_scale)

# 4. auto parse and build kernel targets by cmake
register_kernels(EXCLUDES ${COMMON_BAISC_KERNELS} ${MANUAL_BUILD_KERNELS} DEPS
                 ${COMMON_KERNEL_DEPS} ${COMMON_BAISC_KERNELS})

# phi sparse kernels
add_subdirectory(sparse)
# phi selected_rows kernels
add_subdirectory(selected_rows)

copy_if_different(${kernel_declare_file} ${kernel_declare_file_final})
# For strings kernels
add_subdirectory(strings)