diff --git a/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp b/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp index cfb42c6e17..a90a6340a4 100644 --- a/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp +++ b/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp @@ -121,7 +121,8 @@ using DeviceOpInstance = 2, // CShuffleNXdlPerWavePerShuffle S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock - MaskingSpec>; // MaskingSpecialization + MaskingSpec, // MaskingSpecialization + 1>; // Ref Gemm0: fp16 in, fp32 out using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm + int D0sTransferSrcScalarPerVector = 4, + LoopScheduler LoopSched = LoopScheduler::Default> struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle : public DeviceBatchedGemmSoftmaxGemmPermute; + MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle, + D0sTransferSrcScalarPerVector>; // Argument // FIXME: constness @@ -530,6 +532,11 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle using D0DataType = remove_cvref_t>; // D0 pointer p_d0s_grid_(i) = static_cast(p_acc0_biases[i]); + // for check + d0s_nl_ns_lengths_strides_[i].push_back( + acc0_biases_gs_ms_ns_lengths[i][NumDimG + NumDimM]); + d0s_nl_ns_lengths_strides_[i].push_back( + acc0_biases_gs_ms_ns_strides[i][NumDimG + NumDimM]); }); if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_, @@ -608,6 +615,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle std::vector b_nz_kz_strides_; std::vector b1_nz_kz_strides_; std::vector c_mz_gemm1nz_strides_; + std::array, NumD0Tensor> d0s_nl_ns_lengths_strides_; index_t batch_count_; ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_; @@ -772,6 +780,20 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle { return false; } + for(int i = 0; i < NumD0Tensor; i++) + { + if(arg.d0s_nl_ns_lengths_strides_[i][1] == 1 && + arg.d0s_nl_ns_lengths_strides_[i][0] % D0sTransferSrcScalarPerVector != 0) + { + std::cout << "first" << std::endl; + return false; + } + if(arg.d0s_nl_ns_lengths_strides_[i][1] != 1 && D0sTransferSrcScalarPerVector != 1) + { + std::cout << "second" << std::endl; + return false; + } + } return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp index 74171ea9d9..135b9da6a0 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp @@ -80,7 +80,8 @@ template + int D0sTransferSrcScalarPerVector = 4, + PipelineVersion PipelineVer = PipelineVersion::v1> struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle { static_assert(LoopSched == LoopScheduler::Default, @@ -621,13 +622,13 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 = make_naive_tensor_descriptor_packed(make_tuple(I1, // MBlockId I1, // NBlockID - I1, // MRepeat - I1, // NRepeat - I1, // MWaveId - I1, // NWaveId - I1, // MPerXdl - I1, // NGroupNum - I1, // NInputNum + m0, // MRepeat + n0, // NRepeat + m1, // MWaveId + n1, // NWaveId + m2, // MPerXdl + n2, // NGroupNum + n3, // NInputNum n4)); // registerNum auto d0s_thread_buf = generate_tuple( @@ -644,9 +645,6 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle const auto wave_id = GetGemm0WaveIdx(); const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63 - constexpr auto acc0_thread_desc = make_naive_tensor_descriptor_packed( - make_tuple(Number{}, Number{}, n2, n4)); - auto d0s_threadwise_copy = generate_tuple( [&](auto i) { using D0DataType = remove_cvref_t>; @@ -655,10 +653,19 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle D0DataType, decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]), decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5), - Sequence, + Sequence, Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>, 9, - n4, + D0sTransferSrcScalarPerVector, 1, false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], make_multi_index(block_work_idx[I0], // MBlockId @@ -884,62 +891,35 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle // multiple d if constexpr(NumD0Tensor) { - static_for<0, MXdlPerWave, 1>{}([&](auto mr) { - static_for<0, NXdlPerWave, 1>{}([&](auto nr) { - static_for<0, n2, 1>{}([&](auto groupid) { - static_for<0, NumD0Tensor, 1>{}([&](auto i) { - d0s_threadwise_copy(i).Run( - d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - d0s_grid_buf[i], - d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, - make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0), - d0s_thread_buf(i)); - }); + static_assert(NXdlPerWave == n0); + static_assert(MXdlPerWave == m0); - static_for<0, n4, 1>{}([&](auto i) { - constexpr index_t c_offset = acc0_thread_desc.CalculateOffset( - make_tuple(mr, nr, groupid, i)); + static_for<0, NumD0Tensor, 1>{}([&](auto i) { + d0s_threadwise_copy(i).Run(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], + d0s_grid_buf[i], + d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, + make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0), + d0s_thread_buf(i)); + }); + static_for<0, m0 * n0 * n2 * n4, 1>{}([&](auto i) { + // get reference to src data + const auto src_data_refs = generate_tie( + // return type should be lvalue + [&](auto iSrc) -> const auto& { return d0s_thread_buf[iSrc][i]; }, + Number{}); - // get reference to src data - const auto src_data_refs = generate_tie( - // return type should be lvalue - [&](auto iSrc) -> const auto& { - return d0s_thread_buf[iSrc][i]; - }, - Number{}); + // get reference to dst data + auto dst_data_refs = generate_tie( + // return type should be lvalue + [&](auto) -> auto& { return acc_thread_buf(i); }, + Number<2>{}); - // get reference to dst data - auto dst_data_refs = generate_tie( - // return type should be lvalue - [&](auto) -> auto& { - return acc_thread_buf(Number{}); - }, - Number<2>{}); - - unpack2(c0de_element_op, dst_data_refs, src_data_refs); - }); - static_for<0, NumD0Tensor, 1>{}([&](auto i) { - d0s_threadwise_copy(i).MoveSrcSliceWindow( - d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - make_multi_index(0, 0, 0, 0, 0, 0, 0, 1, 0, 0)); - }); - }); - static_for<0, NumD0Tensor, 1>{}([&](auto i) { - d0s_threadwise_copy(i).MoveSrcSliceWindow( - d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - make_multi_index(0, 0, 0, 1, 0, 0, 0, -n2.value, 0, 0)); - }); - }); - static_for<0, NumD0Tensor, 1>{}([&](auto i) { - d0s_threadwise_copy(i).MoveSrcSliceWindow( - d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - make_multi_index(0, 0, 1, -NXdlPerWave, 0, 0, 0, 0, 0, 0)); - }); + unpack2(c0de_element_op, dst_data_refs, src_data_refs); }); static_for<0, NumD0Tensor, 1>{}([&](auto i) { d0s_threadwise_copy(i).MoveSrcSliceWindow( d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i], - make_multi_index(0, 1, -MXdlPerWave, 0, 0, 0, 0, 0, 0, 0)); + make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0)); }); } else diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp index a541a2d224..498bf58fb3 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp @@ -41,10 +41,11 @@ template , ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, 1>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, @@ -58,8 +59,9 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_ DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>, // Padded fallback kernel - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, 1>, + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp index 5e481603ca..744bd6456d 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp @@ -41,10 +41,11 @@ template , ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, 1>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, #if CK_WORKAROUND_SWDEV_388832 @@ -60,6 +61,7 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16 DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>, // Padded fallback kernel + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, 1>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple, ck::Tuple<>, F32, F16, PassThrough, PassThrough, ScaleAdd, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> // clang-format on diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp index 501ea85f60..b342612d1c 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp @@ -45,6 +45,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_ // #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | // #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | // #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, @@ -58,8 +59,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_ DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>, // Padded fallback kernel - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, F32, BF16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp index 91ab541bf9..3fd0c07370 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp @@ -45,6 +45,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_ // #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | // #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | // #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, #if CK_WORKAROUND_SWDEV_388832 @@ -60,8 +61,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_ DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>, DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>, // Padded fallback kernel - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>, - DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> + DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec> // clang-format on >;