diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_mx.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_mx.hpp index 2f5bcf2fd2..ffb97c2297 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_mx.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_mx.hpp @@ -170,8 +170,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_mx 1) { @@ -338,242 +337,14 @@ struct DeviceGemmMX_Xdl_CShuffleV3 : public DeviceGemmMX 1) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Full) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two) - { - const auto kernel = kernel_gemm_xdl_cshuffle_v3< - GridwiseGemm, - true, - InMemoryDataOperationEnum::AtomicAdd, - minimum_occupancy, - TailNumber::Two>; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Three) - { - const auto kernel = kernel_gemm_xdl_cshuffle_v3< - GridwiseGemm, - true, - InMemoryDataOperationEnum::AtomicAdd, - minimum_occupancy, - TailNumber::Three>; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Four) - { - const auto kernel = kernel_gemm_xdl_cshuffle_v3< - GridwiseGemm, - true, - InMemoryDataOperationEnum::AtomicAdd, - minimum_occupancy, - TailNumber::Four>; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Five) - { - const auto kernel = kernel_gemm_xdl_cshuffle_v3< - GridwiseGemm, - true, - InMemoryDataOperationEnum::AtomicAdd, - minimum_occupancy, - TailNumber::Five>; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six) - { - const auto kernel = kernel_gemm_xdl_cshuffle_v3< - GridwiseGemm, - true, - InMemoryDataOperationEnum::AtomicAdd, - minimum_occupancy, - TailNumber::Six>; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Seven) - { - const auto kernel = kernel_gemm_xdl_cshuffle_v3< - GridwiseGemm, - true, - InMemoryDataOperationEnum::AtomicAdd, - minimum_occupancy, - TailNumber::Seven>; - Run(kernel); - } - } - } - else - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Full) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Three) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Four) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Five) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - - if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == - TailNumber::Seven) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - } - } // Tail number could be Odd or Even - else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4) + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) { if(arg.KBatch > 1) { if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) { - const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds< + const auto kernel = kernel_gemm_xdl_cshuffle_v3< GridwiseGemm, true, InMemoryDataOperationEnum::AtomicAdd, @@ -583,7 +354,7 @@ struct DeviceGemmMX_Xdl_CShuffleV3 : public DeviceGemmMX 1) - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - else - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } - else - { - if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - else - { - const auto kernel = - kernel_gemm_xdl_cshuffle_v3; - Run(kernel); - } - } + throw std::runtime_error("wrong! BlkGemmPipelineVer"); } } else