From a404cc8fafcd9afd2e1ab996b904fdcbc25c80a8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= Date: Mon, 12 Jun 2023 15:37:15 +0200 Subject: [PATCH] Add DeviceBatchedGemmMultipleD_Dl (#732) * Add DeviceBatchedGemmMultipleD_Dl * Fix batched_gemm tests * Fix comments * test_batched_gemm_multi_d fixes * Fix args for isSupported batchedGemmMultipleDDl * Disable tests for gfx90a [ROCm/composable_kernel commit: fc9f97568ffcecf36c3f59f4b0d6680720ddd099] --- .../device_batched_gemm_multiple_d_dl.hpp | 796 ++++++++++++++++++ .../gpu/batched_gemm_multi_d.hpp | 337 ++++++++ .../gpu/batched_gemm_multi_d/CMakeLists.txt | 18 + ..._d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp | 95 +++ ...f16_f16_gkm_gkn_gmn_irregular_instance.cpp | 84 ++ ..._d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp | 95 +++ ...f16_f16_gkm_gnk_gmn_irregular_instance.cpp | 83 ++ ..._d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp | 95 +++ ...f16_f16_gmk_gkn_gmn_irregular_instance.cpp | 83 ++ ..._d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp | 95 +++ ...f16_f16_gmk_gnk_gmn_irregular_instance.cpp | 83 ++ ...lti_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp | 93 ++ ...8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp | 90 ++ ...lti_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp | 93 ++ ...8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp | 90 ++ ...lti_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp | 93 ++ ...8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp | 90 ++ ...lti_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp | 93 ++ ...8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp | 90 ++ profiler/README.md | 27 + .../profiler/profile_batched_gemm_impl.hpp | 95 ++- profiler/src/CMakeLists.txt | 3 +- profiler/src/profile_batched_gemm.cpp | 103 ++- profiler/src/profile_batched_gemm_multi_d.cpp | 190 +++++ test/CMakeLists.txt | 1 + test/batched_gemm/batched_gemm_bf16.cpp | 94 ++- test/batched_gemm/batched_gemm_fp16.cpp | 94 ++- test/batched_gemm/batched_gemm_fp32.cpp | 94 ++- test/batched_gemm/batched_gemm_int8.cpp | 94 ++- test/batched_gemm_multi_d/CMakeLists.txt | 5 + .../test_batched_gemm_multi_d.cpp | 74 ++ 31 files changed, 3347 insertions(+), 123 deletions(-) create mode 100644 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp create mode 100644 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp create mode 100644 profiler/src/profile_batched_gemm_multi_d.cpp create mode 100644 test/batched_gemm_multi_d/CMakeLists.txt create mode 100644 test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp new file mode 100644 index 0000000000..0df365ceb2 --- /dev/null +++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp @@ -0,0 +1,796 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include + +#include "ck/utility/common_header.hpp" +#include "ck/tensor_description/tensor_descriptor.hpp" +#include "ck/tensor_description/tensor_descriptor_helper.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp" +#include "ck/host_utility/device_prop.hpp" +#include "ck/host_utility/kernel_launch.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { + +/* + * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix + * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly + * strided batched, but we can easily extend to other layouts. The returned offset can be either \p + * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB + * limitations. + * + * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2 + * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid + * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link + * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for + * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the + * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch. + */ + +template +__global__ void +#if CK_USE_LAUNCH_BOUNDS + __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU) +#endif + kernel_gemm_dl_multiple_d( + const ABDataType* __restrict__ p_a_grid, + const ABDataType* __restrict__ p_b_grid, + DsPointer p_ds_grid, + EDataType* __restrict__ p_e_grid, + const index_t batch_count, + const AElementwiseOperation a_element_op, + const BElementwiseOperation b_element_op, + const CDEElementwiseOperation cde_element_op, + const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1, + const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1, + const DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11, + const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11, + const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch, + const Block2CTileMap block_2_ctile_map) +{ +// TODO: Enable for gfx90a after complier fix +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \ + defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx1101__) || \ + defined(__gfx1102__)) + + const index_t num_blocks_per_batch = + __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); + const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch); + + const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx))); + const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx))); + const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx))); + + const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx); + + __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + + DsPointer p_ds_grid_grp; + + static constexpr index_t NumDTensor = DsGridDesc_M0_M10_M11_N0_N10_N11::Size(); + + static_for<0, NumDTensor, 1>{}( + [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; }); + + GridwiseGemm::Run(p_a_grid + a_batch_offset, + p_b_grid + b_batch_offset, + p_ds_grid_grp, + p_e_grid + e_batch_offset, + p_shared, + a_element_op, + b_element_op, + cde_element_op, + a_grid_desc_k0_m0_m1_k1, + b_grid_desc_k0_n0_n1_k1, + ds_grid_desc_m0_m10_m11_n0_n10_n11, + e_grid_desc_m0_m10_m11_n0_n10_n11, + block_2_ctile_map, + integral_constant{}, + integral_constant{}); +#else + ignore = p_a_grid; + ignore = p_b_grid; + ignore = p_ds_grid; + ignore = p_e_grid; + ignore = batch_count; + ignore = a_element_op; + ignore = b_element_op; + ignore = cde_element_op; + ignore = a_grid_desc_k0_m0_m1_k1; + ignore = b_grid_desc_k0_n0_n1_k1; + ignore = ds_grid_desc_m0_m10_m11_n0_n10_n11; + ignore = e_grid_desc_m0_m10_m11_n0_n10_n11; + ignore = compute_ptr_offset_of_batch; + ignore = block_2_ctile_map; + +#endif +} + +template && + is_same_v, + bool> = false> +struct DeviceBatchedGemmMultipleD_Dl : public DeviceBatchedGemmMultiD + +{ + using DeviceOp = DeviceBatchedGemmMultipleD_Dl; + static constexpr index_t NumDTensor = DsDataType::Size(); + + static constexpr auto I0 = Number<0>{}; + static constexpr auto I1 = Number<1>{}; + static constexpr auto I2 = Number<2>{}; + static constexpr auto I3 = Number<3>{}; + static constexpr auto I4 = Number<4>{}; + static constexpr auto I5 = Number<5>{}; + + static constexpr auto K1Number = Number{}; + + static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA) + { + const index_t K0 = K / K1; + + const auto a_grid_desc_m_k = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA)); + } + }(); + + if constexpr(GemmSpec == GemmSpecialization::MNPadding) + { + const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock; + + return transform_tensor_descriptor( + a_grid_desc_m_k, + make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), + make_right_pad_transform(M, PadM)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + else + { + return transform_tensor_descriptor( + a_grid_desc_m_k, + make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), + make_pass_through_transform(M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + } + + static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB) + { + const index_t K0 = K / K1; + + const auto b_grid_desc_k_n = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB)); + } + }(); + + if constexpr(GemmSpec == GemmSpecialization::MNPadding) + { + const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock; + + return transform_tensor_descriptor( + b_grid_desc_k_n, + make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), + make_right_pad_transform(N, PadN)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + else + { + return transform_tensor_descriptor( + b_grid_desc_k_n, + make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), + make_pass_through_transform(N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + } + + template + static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE) + { + const auto c_grid_desc_m_n = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE)); + } + }(); + + if constexpr(GemmSpec == GemmSpecialization::MNPadding) + { + const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock; + const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock; + + return transform_tensor_descriptor( + c_grid_desc_m_n, + make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + else + { + + return transform_tensor_descriptor( + c_grid_desc_m_n, + make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + } + + static auto MakeDsGridDescriptor_M_N(const std::array& MRaws, + const std::array& NRaws, + const std::array& DsStride) + { + return generate_tuple( + [&](auto i) { + using DLayout = remove_cvref_t>; + + return DeviceOp::MakeEGridDescriptor_M_N(MRaws[i], NRaws[i], DsStride[i]); + }, + Number{}); + } + + using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1)); + using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1)); + using DsGridDesc_M_N = decltype(MakeDsGridDescriptor_M_N({}, {}, {})); + using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N(1, 1, 1)); + + struct ComputePtrOffsetOfStridedBatch + { + ComputePtrOffsetOfStridedBatch(index_t BatchStrideA, + index_t BatchStrideB, + std::array BatchStrideDs, + index_t BatchStrideE) + : BatchStrideA_(BatchStrideA), + BatchStrideB_(BatchStrideB), + BatchStrideDs_(BatchStrideDs), + BatchStrideE_(BatchStrideE) + { + } + + __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideA_); + } + + __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideB_); + } + + __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const + { + std::array ds_offset; + static_for<0, NumDTensor, 1>{}([&](auto i) { + ds_offset[i] = g_idx * static_cast(BatchStrideDs_[i]); + }); + return ds_offset; + } + + __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideE_); + } + + private: + index_t BatchStrideA_; + index_t BatchStrideB_; + std::array BatchStrideDs_; + index_t BatchStrideE_; + }; + + // GridwiseGemm + using GridwiseGemm = + GridwiseGemmDlMultipleD_km_kn_mn; + + using AGridDesc_K0_M0_M1_K1 = + decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{})); + using BGridDesc_K0_N0_N1_K1 = + decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{})); + using DsGridDesc_M0_M10_M11_N0_N10_N11 = + decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{})); + using EGridDesc_M0_M10_M11_N0_N10_N11 = + decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(EGridDesc_M_N{})); + using DefaultBlock2CTileMap = + decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{})); + + // Argument + struct Argument : public BaseArgument + { + Argument(const void* p_a_grid, + const void* p_b_grid, + std::array p_ds_grid, + void* p_e_grid, + index_t M, + index_t N, + index_t K, + index_t Batch, + index_t StrideA, + index_t StrideB, + std::array StrideDs, + index_t StrideE, + index_t BatchStrideA, + index_t BatchStrideB, + const std::array& BatchStrideDs, + index_t BatchStrideE, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CDEElementwiseOperation cde_element_op) + : p_a_grid_{static_cast(p_a_grid)}, + p_b_grid_{static_cast(p_b_grid)}, + p_ds_grid_{}, + p_e_grid_{static_cast(p_e_grid)}, + K_(K), + Batch_(Batch), + a_grid_desc_k0_m0_m1_k1_{}, + b_grid_desc_k0_n0_n1_k1_{}, + e_grid_desc_m0_m10_m11_n0_n10_n11_{}, + compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideDs, BatchStrideE}, + block_2_ctile_map_{}, + a_element_op_{a_element_op}, + b_element_op_{b_element_op}, + cde_element_op_{cde_element_op} + { + a_grid_desc_k0_m_k1_ = + DeviceBatchedGemmMultipleD_Dl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA); + b_grid_desc_k0_n_k1_ = + DeviceBatchedGemmMultipleD_Dl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB); + static_for<0, NumDTensor, 1>{}([&](auto i) { + using DLayout = remove_cvref_t>; + using DDataType = remove_cvref_t>; + + // D pointer + p_ds_grid_(i) = static_cast(p_ds_grid[i]); + + // D desc + ds_grid_desc_m_n_(i) = + DeviceOp::MakeEGridDescriptor_M_N(M, N, StrideDs[i]); + }); + e_grid_desc_m_n_ = + DeviceBatchedGemmMultipleD_Dl::MakeEGridDescriptor_M_N(M, N, StrideE); + + if(GridwiseGemm::CheckValidity( + a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, e_grid_desc_m_n_)) + { + a_grid_desc_k0_m0_m1_k1_ = + GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_k0_m_k1_); + b_grid_desc_k0_n0_n1_k1_ = + GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_k0_n_k1_); + + ds_grid_desc_m0_m10_m11_n0_n10_n11_ = + GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n_); + + e_grid_desc_m0_m10_m11_n0_n10_n11_ = + GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n_); + + block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_); + } + } + + // private: + const ADataType* p_a_grid_; + const BDataType* p_b_grid_; + typename GridwiseGemm::DsGridPointer p_ds_grid_; + EDataType* p_e_grid_; + + index_t K_; + + // Batch + index_t Batch_; + + AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_; + BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_; + DsGridDesc_M_N ds_grid_desc_m_n_; + EGridDesc_M_N e_grid_desc_m_n_; + + AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_; + BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_; + DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11_; + EGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_; + + // for calculating batch offset + ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_; + + DefaultBlock2CTileMap block_2_ctile_map_; + + // TODO: unused since gridwise_gemm_dl_v1r3 does NOT support prologue for the time being. + AElementwiseOperation a_element_op_; + BElementwiseOperation b_element_op_; + CDEElementwiseOperation cde_element_op_; + }; + + // Invoker + struct Invoker : public BaseInvoker + { + using Argument = DeviceBatchedGemmMultipleD_Dl::Argument; + + float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) + { + { + std::cout << "arg.a_grid_desc_k0_m0_m1_k1_{" + << arg.a_grid_desc_k0_m_k1_.GetLength(I0) << ", " + << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", " + << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl; + + std::cout << "arg.b_grid_desc_k0_n0_n1_k1_{" + << arg.b_grid_desc_k0_n_k1_.GetLength(I0) << ", " + << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", " + << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl; + + std::cout << "arg.e_grid_desc_m_n_{ " << arg.e_grid_desc_m_n_.GetLength(I0) << ", " + << arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl; + } + + if(!GridwiseGemm::CheckValidity( + arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_)) + { + throw std::runtime_error( + "wrong! GridwiseGemmDlMultipleD_km_kn_mn has invalid setting"); + } + + const index_t grid_size = + GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_.GetLength(I0), + arg.e_grid_desc_m_n_.GetLength(I1)) * + arg.Batch_; + + auto launch_kernel = [&](auto has_main_k_block_loop, + auto has_double_tail_k_block_loop) { + constexpr bool has_main_loop = has_main_k_block_loop.value; + constexpr bool has_double_loop = has_double_tail_k_block_loop.value; + + const auto kernel = + kernel_gemm_dl_multiple_d; + + return launch_and_time_kernel(stream_config, + kernel, + dim3(grid_size), + dim3(BlockSize), + 0, + arg.p_a_grid_, + arg.p_b_grid_, + arg.p_ds_grid_, + arg.p_e_grid_, + arg.Batch_, + arg.a_element_op_, + arg.b_element_op_, + arg.cde_element_op_, + arg.a_grid_desc_k0_m0_m1_k1_, + arg.b_grid_desc_k0_n0_n1_k1_, + arg.ds_grid_desc_m0_m10_m11_n0_n10_n11_, + arg.e_grid_desc_m0_m10_m11_n0_n10_n11_, + arg.compute_ptr_offset_of_batch_, + arg.block_2_ctile_map_); + }; + + const auto K0 = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0); + const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0); + const bool has_double_tail_k_block_loop = + GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0); + + if(has_main_k_block_loop && has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else if(has_main_k_block_loop && !has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else if(!has_main_k_block_loop && has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + } + + // polymorphic + float Run(const BaseArgument* p_arg, + const StreamConfig& stream_config = StreamConfig{}) override + { + return Run(*dynamic_cast(p_arg), stream_config); + } + }; + + static constexpr bool IsValidCompilationParameter() + { + // TODO: properly implement this check + return true; + } + + static bool IsSupportedArgument(const Argument& arg) + { + // TODO: Enable for gfx90a after complier fix + if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" || + ck::get_device_name() == "gfx1030" || ck::get_device_name() == "gfx940" || + ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" || + ck::get_device_name() == "gfx1102") + { + bool pass = true; + pass = pass && arg.K_ % K1 == 0; + + pass = pass && GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_, + arg.b_grid_desc_k0_n_k1_, + arg.e_grid_desc_m_n_); + + return pass; + } + else + { + return false; + } + } + + // polymorphic + bool IsSupportedArgument(const BaseArgument* p_arg) override + { + return IsSupportedArgument(*dynamic_cast(p_arg)); + } + + static auto MakeArgument(const void* p_a, + const void* p_b, + std::array p_ds, + void* p_e, + index_t M, + index_t N, + index_t K, + index_t Batch, + index_t StrideA, + index_t StrideB, + std::array StrideDs, + index_t StrideE, + index_t BatchStrideA, + index_t BatchStrideB, + const std::array& BatchStrideDs, + index_t BatchStrideE, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CDEElementwiseOperation cde_element_op) + { + return Argument{p_a, + p_b, + p_ds, + p_e, + M, + N, + K, + Batch, + StrideA, + StrideB, + StrideDs, + StrideE, + BatchStrideA, + BatchStrideB, + BatchStrideDs, + BatchStrideE, + a_element_op, + b_element_op, + cde_element_op}; + } + + static auto MakeInvoker() { return Invoker{}; } + + // polymorphic + std::unique_ptr + MakeArgumentPointer(const void* p_a, + const void* p_b, + const std::array& p_ds, + void* p_e, + index_t M, + index_t N, + index_t K, + index_t Batch, + index_t StrideA, + index_t StrideB, + const std::array& StrideDs, + index_t StrideE, + index_t BatchStrideA, + index_t BatchStrideB, + const std::array& BatchStrideDs, + index_t BatchStrideE, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CDEElementwiseOperation cde_element_op) override + { + return std::make_unique(p_a, + p_b, + p_ds, + p_e, + M, + N, + K, + Batch, + StrideA, + StrideB, + StrideDs, + StrideE, + BatchStrideA, + BatchStrideB, + BatchStrideDs, + BatchStrideE, + a_element_op, + b_element_op, + cde_element_op); + } + + // polymorphic + std::unique_ptr MakeInvokerPointer() override + { + return std::make_unique(Invoker{}); + } + + // polymorphic + std::string GetTypeString() const override + { + auto str = std::stringstream(); + + // clang-format off + str << "DeviceBatchedGemmMultipleD_Dl" + << "<" + << BlockSize << ", " + << MPerBlock << ", " + << NPerBlock << ", " + << K0PerBlock << ", " + << K1 << ", " + << M1PerThread << ", " + << N1PerThread << ", " + << KPerThread + << ">"; + // clang-format on + + return str.str(); + } +}; + +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp new file mode 100644 index 0000000000..ae12f4c7ac --- /dev/null +++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp @@ -0,0 +1,337 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances( + std::vector>>& instances); + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances( + std::vector>>& instances); + +template +struct DeviceOperationInstanceFactory> +{ + using DeviceOp = DeviceBatchedGemmMultiD; + + static auto GetInstances() + { + std::vector> op_ptrs; + + if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances( + op_ptrs); + } + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances(op_ptrs); + add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances( + op_ptrs); + } + } + + return op_ptrs; + } +}; + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt new file mode 100644 index 0000000000..fda55a9303 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt @@ -0,0 +1,18 @@ +add_instance_library(device_batched_gemm_multi_d_instance + device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp +) diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp new file mode 100644 index 0000000000..3fe9f78b2a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp @@ -0,0 +1,95 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[k, m] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..4ab22bb03a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp @@ -0,0 +1,84 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[k, m] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp new file mode 100644 index 0000000000..80c890cdb3 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp @@ -0,0 +1,95 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[k, m] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..647c583036 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[k, m] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp new file mode 100644 index 0000000000..3ce582f1f9 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp @@ -0,0 +1,95 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..34c29d0380 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp new file mode 100644 index 0000000000..e0da8e3d5a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp @@ -0,0 +1,95 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..3edd3f78c6 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[m, k] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, F16, F16, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp new file mode 100644 index 0000000000..33234edc17 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp @@ -0,0 +1,93 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[k, m] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 4, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 4>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 4, 1, 4, 1, S<2, 4>, S<2, 4>, S<1, 1, 4, 4>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 4, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 4, 4, 1, 1, S<2, 4>, S<2, 4>, S<4, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 4, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 4, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 4, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 4, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 4, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 4, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 4, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 4, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 4, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..16107e1bde --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp @@ -0,0 +1,90 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[k, m] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<4, 4>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 2, 4, 1, S<4, 2>, S<2, 2>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 4, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 4, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 4, 4>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 4, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 4, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 4, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 4, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 4, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 4, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<1, 1, 4, 4>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp new file mode 100644 index 0000000000..3e4bdb0172 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp @@ -0,0 +1,93 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[k, m] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<2, 4>, S<2, 4>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..8107717162 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp @@ -0,0 +1,90 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[k, m] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<4, 4>, S<4, 2>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 2, 4, 1, S<4, 2>, S<2, 2>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<4, 2>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp new file mode 100644 index 0000000000..0e943c88c3 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp @@ -0,0 +1,93 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<2, 4>, S<2, 4>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..ea5e7c562d --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp @@ -0,0 +1,90 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<4, 4>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 8, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 2, 4, 1, S<4, 2>, S<2, 2>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<4, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp new file mode 100644 index 0000000000..000a4b0130 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp @@ -0,0 +1,93 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 4, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=16, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<2, 4>, S<2, 4>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp new file mode 100644 index 0000000000..24fb67619a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp @@ -0,0 +1,90 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; +using Empty_Tuple = ck::Tuple<>; + +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +// Compilation parameters for a[m, k] * b[n, k] = c[m, n] +using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances = std::tuple< + // clang-format off + // ##########################| ALayout| BLayout| DsLayout| CLayout| AData| BData| AccData| DsData| CData| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ##########################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ##########################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | + // ##########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // MPerBlock=128, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<4, 4>, S<4, 2>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 16, 4, 4, 8, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 2, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // // MPerBlock=128, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 16, 4, 4, 2, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // // MPerBlock=64, NPerBlock=128 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 16, 4, 2, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 4>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 64, 8, 4, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=32, NPerBlock=32 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 2, 4, 1, S<4, 2>, S<2, 2>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 32, 32, 32, 8, 4, 4, 2, 1, S<2, 2>, S<2, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<4, 1, 2, 4>, S<2, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 16, 16, 16, 16, 2, 2, 2, 1, S<4, 1>, S<4, 1>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=64 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=64, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceBatchedGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // clang-format on + >; + +void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/profiler/README.md b/profiler/README.md index 400a64a396..7a1fb29113 100644 --- a/profiler/README.md +++ b/profiler/README.md @@ -76,3 +76,30 @@ e_m_n: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1} .... Best Perf: 211.405 ms, 41.6077 TFlops, 15.2372 GB/s ``` + +## Profile batched gemm multiple D kernels +```bash +#arg1: tensor operation (batched_gemm_multi_d=Batched GEMM multi D); +#arg2: data type (0: fp16; 1: int8) +#arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n]; +# 1: A[g, m, k] * B[g, n, k] = C[g, m, n]; +# 2: A[g, k, m] * B[g, k, n] = C[g, m, n]; +# 3: A[g, k, m] * B[g, n, k] = C[g, m, n]) +#arg4: verification (0: no; 1: yes) +#arg5: initialization (0: no init; 1: integer value; 2: decimal value) +#arg6: print tensor value (0: no; 1: yes) +#arg7: time kernel (0=n0, 1=yes) +#arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount + +################ op datatype layout verify init log time M N K StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount +./bin/ckProfiler batched_gemm_multi_d 0 1 0 0 0 1 4096 4096 4096 4096 4096 4096 16777216 16777216 16777216 16 +``` + +Result (Radeon RX 6800 XT) +```bash +arg.a_grid_desc_k0_m0_m1_k1_{2048, 4096, 2} +arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2} +arg.e_grid_desc_m_n_{ 4096, 4096} +.... +Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s +``` diff --git a/profiler/include/profiler/profile_batched_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_impl.hpp index cdc94aa9a6..936c22f5d8 100644 --- a/profiler/include/profiler/profile_batched_gemm_impl.hpp +++ b/profiler/include/profiler/profile_batched_gemm_impl.hpp @@ -8,9 +8,11 @@ #include "ck/ck.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp" +#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp" #include "ck/library/utility/check_err.hpp" #include "ck/library/utility/device_memory.hpp" @@ -27,7 +29,11 @@ template + typename CLayout, + typename AElementOp, + typename BElementOp, + typename CElementOp, + typename DeviceOp> bool profile_batched_gemm_impl(int do_verification, int init_method, bool do_log, @@ -88,10 +94,6 @@ bool profile_batched_gemm_impl(int do_verification, b_g_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); } - using AElementOp = ck::tensor_operation::element_wise::PassThrough; - using BElementOp = ck::tensor_operation::element_wise::PassThrough; - using CElementOp = ck::tensor_operation::element_wise::PassThrough; - const auto a_element_op = AElementOp{}; const auto b_element_op = BElementOp{}; const auto c_element_op = CElementOp{}; @@ -124,16 +126,6 @@ bool profile_batched_gemm_impl(int do_verification, b_device_buf.ToDevice(b_g_k_n.mData.data()); c_device_buf.ToDevice(c_g_m_n_device_result.mData.data()); - using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemm; - // get device op instances const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< DeviceOp>::GetInstances(); @@ -148,23 +140,62 @@ bool profile_batched_gemm_impl(int do_verification, // profile device op instances for(auto& op_ptr : op_ptrs) { - auto argument_ptr = - op_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), - static_cast(b_device_buf.GetDeviceBuffer()), - static_cast(c_device_buf.GetDeviceBuffer()), - M, - N, - K, - StrideA, - StrideB, - StrideC, - BatchStrideA, - BatchStrideB, - BatchStrideC, - BatchCount, - ck::tensor_operation::element_wise::PassThrough{}, - ck::tensor_operation::element_wise::PassThrough{}, - ck::tensor_operation::element_wise::PassThrough{}); + std::unique_ptr argument_ptr; + // false branch for multi d dl kernel + if constexpr(std::is_same< + DeviceOp, + ck::tensor_operation::device::DeviceBatchedGemm>::value) + { + + argument_ptr = + op_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), + static_cast(b_device_buf.GetDeviceBuffer()), + static_cast(c_device_buf.GetDeviceBuffer()), + M, + N, + K, + StrideA, + StrideB, + StrideC, + BatchStrideA, + BatchStrideB, + BatchStrideC, + BatchCount, + ck::tensor_operation::element_wise::PassThrough{}, + ck::tensor_operation::element_wise::PassThrough{}, + ck::tensor_operation::element_wise::PassThrough{}); + } + else + { + argument_ptr = + op_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), + static_cast(b_device_buf.GetDeviceBuffer()), + {}, + static_cast(c_device_buf.GetDeviceBuffer()), + M, + N, + K, + BatchCount, + StrideA, + StrideB, + {}, + StrideC, + BatchStrideA, + BatchStrideB, + {}, + BatchStrideC, + ck::tensor_operation::element_wise::PassThrough{}, + ck::tensor_operation::element_wise::PassThrough{}, + ck::tensor_operation::element_wise::PassThrough{}); + } auto invoker_ptr = op_ptr->MakeInvokerPointer(); diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index c9fccc2588..6f768e0ae1 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -34,6 +34,7 @@ set(PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp profile_contraction_bilinear.cpp profile_contraction_scale.cpp + profile_batched_gemm_multi_d.cpp ) set(PROFILER_EXECUTABLE ckProfiler) @@ -77,5 +78,5 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgel target_link_libraries(${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) - +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) diff --git a/profiler/src/profile_batched_gemm.cpp b/profiler/src/profile_batched_gemm.cpp index dc83e25b48..222532b7bb 100644 --- a/profiler/src/profile_batched_gemm.cpp +++ b/profiler/src/profile_batched_gemm.cpp @@ -10,6 +10,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" #include "profiler_operation_registry.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + enum struct GemmMatrixLayout { MK_KN_MN, // 0 @@ -78,55 +80,72 @@ int profile_batched_gemm(int argc, char* argv[]) using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; - auto profile = [&](auto a_type, - auto b_type, - auto c_type, - auto a_layout, - auto b_layout, - auto c_layout) { - using ADataType = decltype(a_type); - using BDataType = decltype(b_type); - using CDataType = decltype(c_type); + auto profile = + [&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) { + using ADataType = decltype(a_type); + using BDataType = decltype(b_type); + using CDataType = decltype(c_type); - using ALayout = decltype(a_layout); - using BLayout = decltype(b_layout); - using CLayout = decltype(c_layout); + using ALayout = decltype(a_layout); + using BLayout = decltype(b_layout); + using CLayout = decltype(c_layout); - const int DefaultStrideA = ck::is_same_v ? K : M; - const int DefaultStrideB = ck::is_same_v ? N : K; - const int DefaultStrideC = ck::is_same_v ? N : M; + const int DefaultStrideA = ck::is_same_v ? K : M; + const int DefaultStrideB = ck::is_same_v ? N : K; + const int DefaultStrideC = ck::is_same_v ? N : M; - const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA; - const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB; - const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC; + const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA; + const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB; + const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC; - const int DefaultBatchStrideA = (ck::is_same_v ? M : K) * StrideA_; - const int DefaultBatchStrideB = (ck::is_same_v ? K : N) * StrideB_; - const int DefaultBatchStrideC = (ck::is_same_v ? M : N) * StrideC_; + const int DefaultBatchStrideA = (ck::is_same_v ? M : K) * StrideA_; + const int DefaultBatchStrideB = (ck::is_same_v ? K : N) * StrideB_; + const int DefaultBatchStrideC = (ck::is_same_v ? M : N) * StrideC_; - const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA; - const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB; - const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC; + const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA; + const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB; + const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC; - bool pass = ck::profiler:: - profile_batched_gemm_impl( - do_verification, - init_method, - do_log, - time_kernel, - M, - N, - K, - BatchStrideA_, - BatchStrideB_, - BatchStrideC_, - StrideA_, - StrideB_, - StrideC_, - BatchCount); + using AElementOp = ck::tensor_operation::element_wise::PassThrough; + using BElementOp = ck::tensor_operation::element_wise::PassThrough; + using CElementOp = ck::tensor_operation::element_wise::PassThrough; - return pass ? 0 : 1; - }; + using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemm; + + bool pass = ck::profiler::profile_batched_gemm_impl(do_verification, + init_method, + do_log, + time_kernel, + M, + N, + K, + BatchStrideA_, + BatchStrideB_, + BatchStrideC_, + StrideA_, + StrideB_, + StrideC_, + BatchCount); + + return pass ? 0 : 1; + }; if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN) { diff --git a/profiler/src/profile_batched_gemm_multi_d.cpp b/profiler/src/profile_batched_gemm_multi_d.cpp new file mode 100644 index 0000000000..98b462d950 --- /dev/null +++ b/profiler/src/profile_batched_gemm_multi_d.cpp @@ -0,0 +1,190 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include +#include + +#include "profiler/profile_batched_gemm_impl.hpp" +#include "profiler_operation_registry.hpp" + +#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp" + +enum struct GemmMatrixLayout +{ + MK_KN_MN, // 0 + MK_NK_MN, // 1 + KM_KN_MN, // 2 + KM_NK_MN, // 3 +}; + +enum struct GemmDataType +{ + F16_F16_F16, // 0 + INT8_INT8_INT8, // 1 +}; + +#define OP_NAME "batched_gemm_multi_d" +#define OP_DESC "Batched GEMM multi D" + +int profile_batched_gemm_multi_d(int argc, char* argv[]) +{ + if(argc != 18) + { + // clang-format off + printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"); + printf("arg2: data type (0: fp16; 1: int8)\n"); + printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n"); + printf(" 1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n"); + printf(" 2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n"); + printf(" 3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n"); + printf("arg4: verification (0: no; 1: yes)\n"); + printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"); + printf("arg6: print tensor value (0: no; 1: yes)\n"); + printf("arg7: time kernel (0=n0, 1=yes)\n"); + printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n"); + // clang-format on + exit(1); + } + + const auto data_type = static_cast(std::stoi(argv[2])); + const auto layout = static_cast(std::stoi(argv[3])); + const bool do_verification = std::stoi(argv[4]); + const int init_method = std::stoi(argv[5]); + const bool do_log = std::stoi(argv[6]); + const bool time_kernel = std::stoi(argv[7]); + + const int M = std::stoi(argv[8]); + const int N = std::stoi(argv[9]); + const int K = std::stoi(argv[10]); + + const int StrideA = std::stoi(argv[11]); + const int StrideB = std::stoi(argv[12]); + const int StrideC = std::stoi(argv[13]); + + const int BatchStrideA = std::stoi(argv[14]); + const int BatchStrideB = std::stoi(argv[15]); + const int BatchStrideC = std::stoi(argv[16]); + + const int BatchCount = std::stoi(argv[17]); + + using F16 = ck::half_t; + using INT8 = int8_t; + + using Row = ck::tensor_layout::gemm::RowMajor; + using Col = ck::tensor_layout::gemm::ColumnMajor; + + auto profile = + [&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) { + using ADataType = decltype(a_type); + using BDataType = decltype(b_type); + using CDataType = decltype(c_type); + using DsDataType = ck::Tuple<>; + + using ALayout = decltype(a_layout); + using BLayout = decltype(b_layout); + using CLayout = decltype(c_layout); + using DsLayout = ck::Tuple<>; + + const int DefaultStrideA = ck::is_same_v ? K : M; + const int DefaultStrideB = ck::is_same_v ? N : K; + const int DefaultStrideC = ck::is_same_v ? N : M; + + const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA; + const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB; + const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC; + + const int DefaultBatchStrideA = (ck::is_same_v ? M : K) * StrideA_; + const int DefaultBatchStrideB = (ck::is_same_v ? K : N) * StrideB_; + const int DefaultBatchStrideC = (ck::is_same_v ? M : N) * StrideC_; + + const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA; + const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB; + const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC; + + using AElementOp = ck::tensor_operation::element_wise::PassThrough; + using BElementOp = ck::tensor_operation::element_wise::PassThrough; + using CElementOp = ck::tensor_operation::element_wise::PassThrough; + + using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemmMultiD; + + bool pass = ck::profiler::profile_batched_gemm_impl(do_verification, + init_method, + do_log, + time_kernel, + M, + N, + K, + BatchStrideA_, + BatchStrideB_, + BatchStrideC_, + StrideA_, + StrideB_, + StrideC_, + BatchCount); + + return pass ? 0 : 1; + }; + + if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN) + { + return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{}); + } + else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN) + { + return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{}); + } + else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN) + { + return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{}); + } + else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN) + { + return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{}); + } + else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN) + { + return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{}); + } + else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN) + { + return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{}); + } + else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN) + { + return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{}); + } + else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN) + { + return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{}); + } + else + { + std::cout << "this data_type & layout is not implemented" << std::endl; + + return 1; + } +} + +REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_multi_d); diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index dad9b53cec..e3385b9dd4 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -58,6 +58,7 @@ add_subdirectory(elementwise_normalization) add_subdirectory(batchnorm) add_subdirectory(contraction) add_subdirectory(pool_fwd) +add_subdirectory(batched_gemm_multi_d) if(GPU_TARGETS MATCHES "gfx1100") add_subdirectory(wmma_op) endif() diff --git a/test/batched_gemm/batched_gemm_bf16.cpp b/test/batched_gemm/batched_gemm_bf16.cpp index fa1652e997..5d12a1e956 100644 --- a/test/batched_gemm/batched_gemm_bf16.cpp +++ b/test/batched_gemm/batched_gemm_bf16.cpp @@ -5,6 +5,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + namespace { using ADataType = ck::bhalf_t; using BDataType = ck::bhalf_t; @@ -12,6 +14,8 @@ using CDataType = ck::bhalf_t; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; } // namespace int main() @@ -23,21 +27,87 @@ int main() bool pass = true; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + using namespace ck::tensor_operation::device; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); std::cout << "test BatchedGEMM bf16: " << (pass ? "Pass" : "Fail") << std::endl; return pass ? 0 : 1; diff --git a/test/batched_gemm/batched_gemm_fp16.cpp b/test/batched_gemm/batched_gemm_fp16.cpp index 3df4912aa0..a2b61d951a 100644 --- a/test/batched_gemm/batched_gemm_fp16.cpp +++ b/test/batched_gemm/batched_gemm_fp16.cpp @@ -5,6 +5,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + namespace { using ADataType = ck::half_t; using BDataType = ck::half_t; @@ -12,6 +14,8 @@ using CDataType = ck::half_t; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; } // namespace int main() @@ -23,21 +27,87 @@ int main() bool pass = true; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + using namespace ck::tensor_operation::device; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); std::cout << "test BatchedGEMM fp16: " << (pass ? "Pass" : "Fail") << std::endl; return pass ? 0 : 1; diff --git a/test/batched_gemm/batched_gemm_fp32.cpp b/test/batched_gemm/batched_gemm_fp32.cpp index 2d808441b9..2b18d166e6 100644 --- a/test/batched_gemm/batched_gemm_fp32.cpp +++ b/test/batched_gemm/batched_gemm_fp32.cpp @@ -5,6 +5,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + namespace { using ADataType = float; using BDataType = float; @@ -12,6 +14,8 @@ using CDataType = float; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; } // namespace int main() @@ -23,21 +27,87 @@ int main() bool pass = true; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + using namespace ck::tensor_operation::device; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); std::cout << "test BatchedGEMM fp32: " << (pass ? "Pass" : "Fail") << std::endl; return pass ? 0 : 1; diff --git a/test/batched_gemm/batched_gemm_int8.cpp b/test/batched_gemm/batched_gemm_int8.cpp index ed233a5ada..f607eaa84b 100644 --- a/test/batched_gemm/batched_gemm_int8.cpp +++ b/test/batched_gemm/batched_gemm_int8.cpp @@ -5,6 +5,8 @@ #include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp" + namespace { using ADataType = int8_t; using BDataType = int8_t; @@ -12,6 +14,8 @@ using CDataType = int8_t; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; } // namespace int main() @@ -23,21 +27,87 @@ int main() bool pass = true; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + using namespace ck::tensor_operation::device; - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount); - pass = pass && - ck::profiler::profile_batched_gemm_impl( - true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount); + + pass = pass && ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount); std::cout << "test BatchedGEMM int8: " << (pass ? "Pass" : "Fail") << std::endl; return pass ? 0 : 1; diff --git a/test/batched_gemm_multi_d/CMakeLists.txt b/test/batched_gemm_multi_d/CMakeLists.txt new file mode 100644 index 0000000000..45a306551f --- /dev/null +++ b/test/batched_gemm_multi_d/CMakeLists.txt @@ -0,0 +1,5 @@ +# TODO: Enable for gfx90a after complier fix +if(NOT GPU_TARGETS MATCHES "gfx90a") + add_gtest_executable(test_batched_gemm_multi_d test_batched_gemm_multi_d.cpp) + target_link_libraries(test_batched_gemm_multi_d PRIVATE utility device_batched_gemm_multi_d_instance) +endif() diff --git a/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp b/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp new file mode 100644 index 0000000000..4a82654034 --- /dev/null +++ b/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp @@ -0,0 +1,74 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include + +#include "profiler/profile_batched_gemm_impl.hpp" +#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp" + +namespace { +using F16 = ck::half_t; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using Empty_Tuple = ck::Tuple<>; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +template +class TestBatchedGemmMultiD : public ::testing::Test +{ + protected: + using ALayout = std::tuple_element_t<0, Tuple>; + using BLayout = std::tuple_element_t<1, Tuple>; + using CLayout = std::tuple_element_t<2, Tuple>; + + static constexpr int M = 512; + static constexpr int N = 256; + static constexpr int K = 128; + static constexpr int BatchCount = 3; + + template + void Run() + { + using namespace ck::tensor_operation::device; + + const bool pass = + ck::profiler::profile_batched_gemm_impl>( + true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount); + EXPECT_TRUE(pass); + } +}; + +using KernelTypes = ::testing::Types, + std::tuple, + std::tuple, + std::tuple>; +} // namespace + +TYPED_TEST_SUITE(TestBatchedGemmMultiD, KernelTypes); + +TYPED_TEST(TestBatchedGemmMultiD, f16) { this->template Run(); } + +TYPED_TEST(TestBatchedGemmMultiD, int8) { this->template Run(); }