From 8596ce24cab80229080aa9e4de55226e94e30ebc Mon Sep 17 00:00:00 2001 From: Bartlomiej Kocot Date: Wed, 24 Sep 2025 09:07:42 +0000 Subject: [PATCH] fixes --- include/ck/tensor_description/multi_index_transform.hpp | 8 ++++++++ .../transform_conv_bwd_data_to_gemm_v1.hpp | 6 ++++++ .../profiler/profile_grouped_conv_bwd_data_impl.hpp | 4 ++-- 3 files changed, 16 insertions(+), 2 deletions(-) diff --git a/include/ck/tensor_description/multi_index_transform.hpp b/include/ck/tensor_description/multi_index_transform.hpp index ce7e58bd31..e24227ecc3 100644 --- a/include/ck/tensor_description/multi_index_transform.hpp +++ b/include/ck/tensor_description/multi_index_transform.hpp @@ -1553,6 +1553,14 @@ struct UnMerge } }; +/** + * @brief Transformation struct for convolution backward data output indices to GEMM indices. + * + * This struct is responsible for mapping the output tensor indices (N, Ho, Wo, K) from the + * convolution backward data operation to the corresponding indices (K0, M, K1) used in the + * implicit GEMM computation. It encapsulates the necessary parameters and transformation logic + * required to efficiently perform the index conversion. + */ struct ConvBwdDataImplicitGemmOutTransform { static constexpr auto I0 = Number<0>{}; diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp index 9d88cf9370..03c1945d95 100644 --- a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp +++ b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp @@ -13,6 +13,12 @@ namespace ck { namespace tensor_operation { +/** + * @brief Enable custom tensor transform for convolution backward data output. + * + * When set to 1, this macro enables a custom transformation of the output tensor + * in convolution backward data operations. + */ #define CK_USE_CUSTOM_TENSOR_TRANSFORM_FOR_BWD_DATA_OUT 1 template < diff --git a/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp b/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp index 70bdcb7fab..29b2fece6b 100644 --- a/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp +++ b/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp @@ -137,8 +137,8 @@ bool profile_grouped_conv_bwd_data_impl(int do_verification, auto invoker_ptr = op_ptr->MakeInvokerPointer(); - float avg_time = invoker_ptr->Run(argument_ptr.get(), - StreamConfig{nullptr, time_kernel, 0, 500, 100}); + float avg_time = + invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel}); std::size_t flop = conv_param.GetFlops(); std::size_t num_btype = conv_param.GetByte();