From c5cdd51ce4cf5749ca158d466cf3647283d7c4be Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ville=20Pietil=C3=A4?= <> Date: Wed, 7 Jan 2026 05:41:22 -0500 Subject: [PATCH] Fix factory for regular WMMA conv bwd weight. --- .../builder/factory/conv_bwd_weight_wmma_factory.hpp | 8 ++++---- .../conv/ck/test_ckb_conv_bwd_weight_wmma_cshuffle.cpp | 2 +- experimental/builder/test/impl/conv_algorithm_types.hpp | 2 +- .../builder/test/utils/conv_algorithm_type_utils.hpp | 2 +- 4 files changed, 7 insertions(+), 7 deletions(-) diff --git a/experimental/builder/include/ck_tile/builder/factory/conv_bwd_weight_wmma_factory.hpp b/experimental/builder/include/ck_tile/builder/factory/conv_bwd_weight_wmma_factory.hpp index 817432081b..32161a234a 100644 --- a/experimental/builder/include/ck_tile/builder/factory/conv_bwd_weight_wmma_factory.hpp +++ b/experimental/builder/include/ck_tile/builder/factory/conv_bwd_weight_wmma_factory.hpp @@ -51,13 +51,13 @@ struct ConvBwdWeightWmmaFactory static_assert(InputVectorTransferLimits, "Invalid A block transfer config"); static_assert(InputVectorTransferLimits, "Invalid B block transfer config"); static_assert(OutputVectorTransferLimits, "Invalid C block transfer config"); - static_assert(AccessOrderLimits4D, + static_assert(AccessOrderLimits3D, "Invalid A thread cluster access order"); - static_assert(AccessOrderLimits4D, + static_assert(AccessOrderLimits3D, "Invalid B thread cluster access order"); - static_assert(AccessOrderLimits4D, + static_assert(AccessOrderLimits3D, "Invalid A source access order"); - static_assert(AccessOrderLimits4D, + static_assert(AccessOrderLimits3D, "Invalid B source access order"); // The forward convolution kernel class instance. diff --git a/experimental/builder/test/conv/ck/test_ckb_conv_bwd_weight_wmma_cshuffle.cpp b/experimental/builder/test/conv/ck/test_ckb_conv_bwd_weight_wmma_cshuffle.cpp index 0981ea6c11..ff350ac804 100644 --- a/experimental/builder/test/conv/ck/test_ckb_conv_bwd_weight_wmma_cshuffle.cpp +++ b/experimental/builder/test/conv/ck/test_ckb_conv_bwd_weight_wmma_cshuffle.cpp @@ -22,7 +22,7 @@ constexpr auto SIGNATURE = ckt::ConvSignature{.spatial_dim = 3, constexpr auto ALGORITHM = cku::ConvAlgorithm_DeviceGroupedConvBwdWeight_Wmma_CShuffle{} .with_thread_block(cku::ThreadBlock_64_32x32x32) .with_gemm_config(cku::GemmParams_Wmma_16x16_2x1_per_wave) - .with_transfer(cku::BwdTransfer_4x64x1) + .with_transfer(cku::BwdTransfer_4x8x1_4x16x1_v3) .with_bwd_specialization(ckb::ConvSpecialization::DEFAULT) .with_prefetch_config(1, ckb::PipelineScheduler::DEFAULT) .with_gridwise_gemm_pipeline(ckb::PipelineVersion::V1); diff --git a/experimental/builder/test/impl/conv_algorithm_types.hpp b/experimental/builder/test/impl/conv_algorithm_types.hpp index 797f440bdc..27ba1ec3b6 100644 --- a/experimental/builder/test/impl/conv_algorithm_types.hpp +++ b/experimental/builder/test/impl/conv_algorithm_types.hpp @@ -669,7 +669,7 @@ using ConvAlgorithm_DeviceGroupedConvBwdWeight_TwoStage_Wmma_CShuffle_V3 = using ConvAlgorithm_DeviceGroupedConvBwdWeight_Wmma_CShuffle = ConvAlgorithmTemplate, + Transfer_<>, ConvSpecializationBwdWeight_, GridGemm_, Prefetch_>; diff --git a/experimental/builder/test/utils/conv_algorithm_type_utils.hpp b/experimental/builder/test/utils/conv_algorithm_type_utils.hpp index b705ab1e47..d80d6a1b8c 100644 --- a/experimental/builder/test/utils/conv_algorithm_type_utils.hpp +++ b/experimental/builder/test/utils/conv_algorithm_type_utils.hpp @@ -428,7 +428,7 @@ inline std::string to_string(t)) << "," << to_string(static_cast(t)) - << "," << to_string(static_cast>(t)); + << "," << to_string(static_cast>(t)); return oss.str(); }