# ckProfiler set(PROFILER_SOURCES profiler.cpp profile_gemm.cpp # profile_gemm_splitk.cpp # profile_gemm_bilinear.cpp # profile_gemm_bias_add_reduce.cpp # profile_gemm_add_add_fastgelu.cpp # profile_gemm_add_multiply.cpp # profile_gemm_add_fastgelu.cpp # profile_gemm_add_relu_add_layernorm.cpp # profile_gemm_fastgelu.cpp # profile_gemm_reduce.cpp # profile_batched_gemm.cpp # profile_batched_gemm_gemm.cpp # profile_batched_gemm_add_relu_gemm_add.cpp # profile_batched_gemm_reduce.cpp # profile_grouped_gemm.cpp # profile_conv_fwd.cpp # profile_conv_fwd_bias_relu.cpp # profile_conv_fwd_bias_relu_add.cpp # profile_conv_bwd_data.cpp # profile_grouped_conv_fwd.cpp # profile_grouped_conv_bwd_weight.cpp # profile_reduce.cpp # profile_groupnorm.cpp # profile_layernorm.cpp # profile_avg_pool2d_fwd.cpp # profile_max_pool3d_fwd.cpp # profile_softmax.cpp # profile_batchnorm_fwd.cpp # profile_batchnorm_bwd.cpp # profile_batchnorm_infer.cpp # profile_grouped_gemm_fastgelu.cpp # profile_contraction_bilinear.cpp # profile_contraction_scale.cpp ) set(PROFILER_EXECUTABLE ckProfiler) add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES}) target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)