make all phi kernels to 2(host/device) static libraries directly (#43247)

* make all phi kernels to 2(host/device) static libraries directly

* fix calling kernel_declare

* fix compile

* fix cpu compile

* fix rocm compile

* fix xpu compile

* fix xpu kp compile

* fix inference compile

cmake/phi.cmake

@@ -61,7 +61,9 @@ endfunction()
 
 # call kernel_declare need to make sure whether the target of input exists
 function(kernel_declare TARGET_LIST)
+  # message("TARGET LIST ${TARGET_LIST}")
   foreach(kernel_path ${TARGET_LIST})
+    # message("kernel path ${kernel_path}" )
     file(READ ${kernel_path} kernel_impl)
     string(
       REGEX
@@ -111,6 +113,7 @@ function(kernel_declare TARGET_LIST)
 endfunction()
 
 function(kernel_library TARGET)
+  return()
   set(common_srcs)
   set(cpu_srcs)
   set(gpu_srcs)

paddle/fluid/inference/CMakeLists.txt

@@ -37,6 +37,7 @@ endif()
 # fluid_modules exclude API-interface of inference/api and inference/capi_exp
 get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
 get_property(phi_modules GLOBAL PROPERTY PHI_MODULES)
+get_property(phi_kernels GLOBAL PROPERTY PHI_KERNELS)
 set(utils_modules stringpiece pretty_log string_helper benchmark)
 
 add_subdirectory(api)
@@ -50,7 +51,7 @@ set(STATIC_INFERENCE_API
     reset_tensor_array
     analysis_config
     paddle_pass_builder
-    activation_functions
+    phi
     ${mkldnn_quantizer_cfg})
 
 #windows GPU static library over the limit, so not create_static_lib, and cc_library is dummy
@@ -59,7 +60,7 @@ if(WIN32 AND WITH_GPU)
                                    ${utils_modules})
 else()
   create_static_lib(paddle_inference ${fluid_modules} ${phi_modules}
-                    ${STATIC_INFERENCE_API} ${utils_modules})
+                    ${phi_kernels} ${STATIC_INFERENCE_API} ${utils_modules})
 endif()
 
 if(NOT APPLE)

paddle/fluid/operators/CMakeLists.txt

@@ -108,7 +108,7 @@ register_operators(EXCLUDES py_layer_op py_func_op warpctc_op dgc_op load_combin
 
 op_library(run_program_op SRCS run_program_op.cc run_program_op.cu.cc DEPS executor_cache ${OP_HEADER_DEPS})
 target_link_libraries(run_program_op cuda_graph_with_memory_pool)
-op_library(quantize_linear_op DEPS cast_kernel)
+op_library(quantize_linear_op DEPS phi)
 op_library(save_combine_op DEPS string_array)
 op_library(load_combine_op DEPS string_array)
 

paddle/phi/CMakeLists.txt

@@ -38,6 +38,7 @@ set(PHI_DEPS
     string_tensor
     api_scalar
     api_int_array)
+
 get_property(phi_kernels GLOBAL PROPERTY PHI_KERNELS)
 set(PHI_DEPS ${PHI_DEPS} ${phi_kernels})
 

paddle/phi/api/lib/CMakeLists.txt

@@ -335,11 +335,12 @@ cc_library(
 cc_library(
   api_gen_utils
   SRCS api_gen_utils.cc
-  DEPS phi_tensor_raw selected_rows sparse_csr_tensor sparse_coo_tensor)
+  DEPS phi_tensor_raw selected_rows sparse_csr_tensor sparse_coo_tensor
+       infermeta_utils)
 cc_library(
   phi_data_transform
   SRCS data_transform.cc
-  DEPS phi_tensor_raw transfer_layout_kernel cast_kernel copy_kernel tensor)
+  DEPS phi_tensor_raw phi tensor)
 cc_library(
   api_custom_impl
   SRCS api_custom_impl.cc
@@ -404,7 +405,7 @@ cc_library(
 cc_library(
   tensor_copy
   SRCS tensor_copy.cc
-  DEPS phi_tensor_raw copy_kernel kernel_dispatch api_gen_utils)
+  DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils)
 cc_library(
   api_scalar
   SRCS scalar.cc

paddle/phi/kernels/CMakeLists.txt

@@ -41,124 +41,104 @@ set(COMMON_KERNEL_DEPS
     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)
+set(COMMON_KERNEL_DEPS ${COMMON_KERNEL_DEPS} infermeta infermeta_utils)
+set(COMMON_KERNEL_DEPS ${COMMON_KERNEL_DEPS} switch_autotune)
 
-# [ 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})
+set(COMMON_KERNEL_DEPS
+    ${COMMON_KERNEL_DEPS}
+    threadpool
+    jit_kernel_helper
+    softmax
+    cross_entropy
+    matrix_bit_code
+    lapack_function
+    lstm_compute
+    gru_compute
+    deformable_conv_functor
+    matrix_reduce
+    segment_pooling
+    gather_scatter_kernel
+    pooling
+    maxouting
+    matrix_inverse
+    phi_dynload_warpctc
+    sequence_padding
+    sequence_scale)
 
-# phi sparse kernels
-add_subdirectory(sparse)
-# phi selected_rows kernels
-add_subdirectory(selected_rows)
+set(COMMON_KERNEL_DEPS
+    ${COMMON_KERNEL_DEPS}
+    dense_tensor
+    string_tensor
+    sparse_coo_tensor
+    sparse_csr_tensor
+    kernel_context
+    kernel_factory
+    arg_map_context
+    convert_utils
+    lod_utils
+    custom_kernel
+    string_infermeta
+    utf8proc)
 
 copy_if_different(${kernel_declare_file} ${kernel_declare_file_final})
 # For strings kernels
 add_subdirectory(strings)
+
+file(GLOB kernel_h "*.h" "selected_rows/*.h" "sparse/*.h" "strings/*.h")
+file(GLOB kernel_impl_h "impl/*.h" "selected_rows/impl/*.h")
+file(GLOB kernel_primitive_h "primitive/*.h")
+file(
+  GLOB
+  kernel_cc
+  "*.cc"
+  "cpu/*.cc"
+  "selected_rows/*.cc"
+  "selected_rows/cpu/*.cc"
+  "sparse/*.cc"
+  "sparse/cpu/*.cc"
+  "strings/*.cc"
+  "strings/cpu/*.cc")
+
+file(
+  GLOB
+  kernel_cu
+  "gpu/*.cu"
+  "gpu/*.cu.cc"
+  "gpudnn/*.cu"
+  "kps/*.cu"
+  "selected_rows/gpu/*.cu"
+  "sparse/gpu/*.cu"
+  "strings/*.cu"
+  "strings/gpu/*.cu")
+
+# file(GLOB kernel_cudnn "gpudnn/*.cu")
+# file(GLOB kernel_kps "kps/*.cu")
+file(GLOB kernel_xpu "xpu/*.cc")
+
+add_library(phi_cpu ${kernel_cc})
+kernel_declare("${kernel_cc}")
+target_link_libraries(phi_cpu ${COMMON_KERNEL_DEPS})
+set_property(GLOBAL PROPERTY PHI_KERNELS phi_cpu)
+
+if(WITH_GPU OR WITH_ROCM)
+  if(WITH_GPU)
+    add_library(phi_gpu ${kernel_cu})
+  elseif(WITH_ROCM)
+    hip_add_library(phi_gpu STATIC ${kernel_cu})
+  endif()
+  kernel_declare("${kernel_cu}")
+  target_link_libraries(phi_gpu ${COMMON_KERNEL_DEPS})
+  set_property(GLOBAL PROPERTY PHI_KERNELS phi_cpu phi_gpu)
+endif()
+
+if(WITH_XPU)
+  if(WITH_XPU_KP)
+    file(GLOB kernel_xpu_kps "kps/*.cu")
+    xpu_add_library(phi_xpu STATIC ${kernel_xpu} ${kernel_xpu_kps})
+  else()
+    add_library(phi_xpu ${kernel_xpu})
+  endif()
+  kernel_declare(${kernel_xpu})
+  target_link_libraries(phi_xpu ${COMMON_KERNEL_DEPS})
+  set_property(GLOBAL PROPERTY PHI_KERNELS phi_cpu phi_xpu)
+endif()

paddle/phi/tests/core/CMakeLists.txt

@@ -11,7 +11,7 @@ cc_test(test_type_info SRCS test_type_info.cc)
 cc_test(
   test_kernel_factory
   SRCS test_kernel_factory.cc
-  DEPS kernel_factory scale_kernel)
+  DEPS kernel_factory phi)
 cc_test(
   test_sparse_coo_tensor
   SRCS test_sparse_coo_tensor.cc