/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/block/blockwise_gemm_smfmac_xdlops.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/block/blockwise_gemm_smfmac_xdlops.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/block/blockwise_gemm_smfmac_xdlops.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/loop_scheduler.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
 #include "ck/tensor_description/tensor_adaptor.hpp"
  
 namespace ck {
  
 template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
 __host__ __device__ static constexpr auto
 MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(const TileDesc_K0_MN_K1&)
 {
     constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
     constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
  
     return transform_tensor_descriptor(
         TileDesc_K0_MN_K1{},
         make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
                    make_unmerge_transform(
                        make_tuple(Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
         make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
         make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
 }
  
 template <index_t BlockSize,
           typename FloatA,
           typename FloatB,
           typename FloatAcc,
           typename AK0MK1BlockDesc,
           typename BK0NK1BlockDesc,
           index_t MPerXDL,
           index_t NPerXDL,
           index_t MRepeat,
           index_t NRepeat,
           index_t KPack,
           typename ComputeTypeA = FloatA,
           typename ComputeTypeB = FloatB>
 struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
 {
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
     static constexpr auto I3 = Number<3>{};
  
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
  
     static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
     static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
     static_assert(MWaves > 0);
     static_assert(NWaves > 0);
     static constexpr index_t WaveSize = BlockSize / MWaves / NWaves;
  
     static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
     static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
     static constexpr index_t KPerBlock =
         BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
  
     static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
     static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
     static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
     static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
  
     static constexpr auto xdlops_gemm =
         SparseXdlopsGemm<ComputeTypeA, MPerXDL, NPerXDL, KPack, ComputeTypeB>{};
  
     static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
  
     StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
                               FloatAcc,
                               MRepeat * NRepeat,
                               xdlops_gemm.GetRegSizePerXdlops(),
                               true>
         c_thread_buf_;
  
     __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
  
     __device__ static auto GetWaveIdx()
     {
         const index_t thread_id = ThisThreadBlock::GetThreadId();
  
         constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
             make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
             make_tuple(Sequence<0, 1, 2>{}),
             make_tuple(Sequence<0>{}));
  
         return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
     }
  
     __device__ static auto CalculateAThreadOriginDataIndex()
     {
         const auto wave_idx     = GetWaveIdx();
         const auto waveId_m     = wave_idx[I0];
         const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
  
         return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPerThread * xdlops_a_idx[I0]);
     }
  
     __device__ static auto CalculateBThreadOriginDataIndex()
     {
         const auto wave_idx     = GetWaveIdx();
         const auto waveId_n     = wave_idx[I1];
         const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
  
         return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPerThread * xdlops_b_idx[I0]);
     }
  
     template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
     __device__ static auto
     CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
     {
         const auto wave_idx = GetWaveIdx();
         const auto waveId_m = wave_idx[I0];
         const auto waveId_n = wave_idx[I1];
  
         const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
  
         constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
             make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
             make_tuple(Sequence<0>{}),
             make_tuple(Sequence<0, 1, 2>{}));
  
         constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
             make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
             make_tuple(Sequence<0>{}),
             make_tuple(Sequence<0, 1, 2>{}));
  
         const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
             make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
         const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
             make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
  
         return make_tuple(c_thread_m, c_thread_n);
     }
  
     template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
     __device__ static auto
     CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
     {
         const auto wave_idx = GetWaveIdx();
         const auto waveId_m = wave_idx[I0];
         const auto waveId_n = wave_idx[I1];
  
         const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
  
         return make_tuple(Number<m0>{},
                           Number<n0>{},
                           waveId_m,
                           waveId_n,
                           blk_idx[I0],
                           blk_idx[I1],
                           blk_idx[I2],
                           blk_idx[I3]);
     }
  
     __host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1()
     {
 #if defined(__HIP_DEVICE_COMPILE__)
         static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
                           BK0NK1BlockDesc::IsKnownAtCompileTime(),
                       "wrong! Desc should be known at compile-time");
  
         static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
                       "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
  
         static_assert(MPerBlock % (MPerXDL * MRepeat) == 0,
                       "MPerBlock must be divisible by MPerXDL * MRepeat");
         static_assert(NPerBlock % (NPerXDL * NRepeat) == 0,
                       "NPerBlock must be divisible by NPerXDL * NRepeat");
  
         static_assert(
             KPack % (16 * sizeof(ComputeTypeA)) == 0,
             "KPack must be divisbile by number of elements processed in single smfmac instruction");
 #endif
     }
  
     __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
     {
         constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
  
         constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
         constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
         constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
         constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
  
         return make_naive_tensor_descriptor_packed(
             make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
     }
  
     __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
     {
         constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
  
         constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
         constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
         constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
         constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
  
         return make_naive_tensor_descriptor_packed(
             make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
     }
  
     __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
     {
         constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
             make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
                                                            Number<NRepeat>{},
                                                            Number<MWaves>{},
                                                            Number<NWaves>{},
                                                            Number<MPerXDL>{},
                                                            Number<NPerXDL>{}));
  
         return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
     }
  
     __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
     {
         constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
             make_naive_tensor_descriptor_packed(make_tuple(I1,
                                                            Number<MRepeat>{},
                                                            Number<NRepeat>{},
                                                            Number<MWaves>{},
                                                            Number<NWaves>{},
                                                            Number<MPerXDL>{},
                                                            Number<NPerXDL>{}));
  
         return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
             c_block_desc_g_m0_n0_m1_n1_m2_n2);
     }
  
     template <typename CGridDesc_M_N>
     __host__ __device__ static constexpr auto
     MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
     {
         const auto M = c_grid_desc_m_n.GetLength(I0);
         const auto N = c_grid_desc_m_n.GetLength(I1);
  
         const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
             c_grid_desc_m_n,
             make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
                        make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
             make_tuple(Sequence<0>{}, Sequence<1>{}),
             make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
  
         return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
     }
  
     template <typename CGridDesc_G_M_N>
     __host__ __device__ static constexpr auto
     MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
     {
         const auto G = c_grid_desc_g_m_n.GetLength(I0);
         const auto M = c_grid_desc_g_m_n.GetLength(I1);
         const auto N = c_grid_desc_g_m_n.GetLength(I2);
  
         const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
             c_grid_desc_g_m_n,
             make_tuple(make_pass_through_transform(G),
                        make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
                        make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
             make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
             make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
  
         return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
             c_grid_desc_g_m0_n0_m1_n1_m2_n2);
     }
  
     __host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
     {
         return transform_tensor_descriptor(
             AK0MK1BlockDesc{},
             make_tuple(
                 make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
                 make_unmerge_transform(
                     make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}))),
             make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
             make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
     }
  
     __host__ __device__ static constexpr auto MakeBBlockDescriptor_N0_N1_N2_K()
     {
         return transform_tensor_descriptor(
             BK0NK1BlockDesc{},
             make_tuple(
                 make_merge_transform_v3_division_mod(make_tuple(Number<B_K0>{}, Number<B_K1>{})),
                 make_unmerge_transform(
                     make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{}))),
             make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
             make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
     }
  
     static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
     static constexpr auto b_block_desc_n0_n1_n2_k = MakeBBlockDescriptor_N0_N1_N2_K();
  
     // Prepares data in a_thread_buf by squeezing values by ommiting zeros to adjust it to 2:4
     // structural sparsity. The indexes of non-zero elements are stored in idx_buf and used later in
     // smfmac instruction
     template <typename AThreadBuf, typename IdxBuf, int32_t num_elems>
     __device__ void SetIdxSqueezeA(AThreadBuf& a_thread_buf, IdxBuf& idx_buf)
     {
         static constexpr int32_t bit_clear_masks[4] = {0b11, 0b1100, 0b110000, 0b11000000};
         static constexpr int32_t processed_elems    = 16 / sizeof(ComputeTypeA);
  
         static_for<0, num_elems, processed_elems>{}([&](auto i) {
             constexpr int idx_reg_num  = i / (16 * sizeof(ComputeTypeA));
             constexpr int idx_reg_part = (i % 32) / processed_elems;
  
             vector_type<ComputeTypeA, processed_elems> a_thread_vec;
             static_for<0, processed_elems, 1>{}([&](auto j) {
                 a_thread_vec.template AsType<ComputeTypeA>()(j) = a_thread_buf
                     [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, i + j))>{}];
             });
  
             uint8_t idx = 0b11101110; // set to last 2 elems for both 4-elems subgroups by default
             for(int j = 0; j < processed_elems; j += 4)
             {
                 int32_t a_pos                 = idx_reg_part * processed_elems + j;
                 int32_t nonzero_pos           = 0;
                 ComputeTypeA nonzero_elems[2] = {a_thread_vec[j + 2], a_thread_vec[j + 3]};
                 for(int k = 0; k < 3; k += 1)
                 {
                     if(a_thread_vec[j + k] != 0.0f)
                     {
                         nonzero_elems[nonzero_pos] = a_thread_vec[j + k];
                         idx &= ~bit_clear_masks[j / 2 + nonzero_pos];
                         idx |= k << 2 * (j / 2 + nonzero_pos);
                         ++nonzero_pos;
                     }
                 }
                 a_thread_vec[j / 2]     = nonzero_elems[0];
                 a_thread_vec[j / 2 + 1] = nonzero_elems[1];
             }
             IdxBuf[idx_reg_num].AsType<int8x4_t>()[Number<idx_reg_part>{}] = idx;
  
             static_for<0, processed_elems / 2, 1>{}([&](auto j) {
                 a_thread_buf[Number<a_thread_desc_.CalculateOffset(
                     make_tuple(0, 0, 0, i / 2 + j))>{}] = a_thread_vec[j];
             });
         });
     }
  
     template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
     __device__ void Run(const ABlockBuffer& a_block_buf,
                         const BBlockBuffer& b_block_buf,
                         CThreadBuffer& c_thread_buf) const
     {
         auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
             a_thread_desc_.GetElementSpaceSize());
         auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
             b_thread_desc_.GetElementSpaceSize());
         static constexpr int32_t elems_per_idx = 16 * sizeof(ComputeTypeA);
         auto idx_buf = make_static_buffer<AddressSpaceEnum::Vgpr, int32_t>(
             (a_thread_desc_.GetElementSpaceSize() + elems_per_idx - 1) / elems_per_idx);
  
         static_for<0, MRepeat, 1>{}([&](auto m0) {
             // read A
             a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
                                make_tuple(m0, I0, I0, I0),
                                a_block_buf,
                                a_thread_desc_,
                                make_tuple(I0, I0, I0, I0),
                                a_thread_buf);
  
             SetIdxSqueezeA(a_thread_buf, idx_buf, a_thread_desc_.GetElementSpaceSize());
  
             static_for<0, NRepeat, 1>{}([&](auto n0) {
                 // read B
                 b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
                                    make_tuple(n0, I0, I0, I0),
                                    b_block_buf,
                                    b_thread_desc_,
                                    make_tuple(I0, I0, I0, I0),
                                    b_thread_buf);
  
                 static_for<0, KPerThread, KPack>{}([&](auto k) {
                     // a_thread_vec is smaller because it's structurally sparse 2:4
                     vector_type<ComputeTypeA, KPack / 2> a_thread_vec;
                     vector_type<ComputeTypeB, KPack> b_thread_vec;
                     vector_type<int32_t, KPack / elems_per_idx> idx_vec;
  
                     static_for<0, KPack / 2, 1>{}([&](auto i) {
                         a_thread_vec.template AsType<ComputeTypeA>()(i) =
                             a_thread_buf[Number<a_thread_desc_.CalculateOffset(
                                 make_tuple(0, 0, 0, k / 2 + i))>{}];
                     });
  
                     static_for<0, KPack, 1>{}([&](auto i) {
                         b_thread_vec.template AsType<ComputeTypeB>()(2 * i) = b_thread_buf
                             [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
                     });
  
                     static_for<0, KPack / elems_per_idx, 1>{}([&](auto i) {
                         idx_vec.template AsType<int32_t>()(i) = idx_buf[k / elems_per_idx + i];
                     });
  
                     // A is smaller because it's structurally sparse 2:4
                     using mfma_input_type_a =
                         typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops / 2>::type;
                     using mfma_input_type_b =
                         typename vector_type<ComputeTypeB, xdlops_gemm.K1PerXdlops>::type;
                     using mfma_input_type_idx = typename vector_type<int32_t, 1>::type;
  
                     constexpr index_t c_offset =
                         c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
  
                     xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type_a>(),
                                     b_thread_vec.template AsType<mfma_input_type_b>(),
                                     idx_vec.template AsType<mfma_input_type_idx>(),
                                     c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
                 });
             });
         });
     }
  
     protected:
     // A[M0, M1, M2, KPerThread]
     static constexpr auto a_thread_desc_ =
         make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
  
     // B[N0, N1, N2, KPerThread]
     static constexpr auto b_thread_desc_ =
         make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
  
     // C[M, N, NumRegXdlops]
     static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
         make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
  
     using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
                                                          ComputeTypeA,
                                                          decltype(a_block_desc_m0_m1_m2_k),
                                                          decltype(a_thread_desc_),
                                                          Sequence<1, 1, 1, KPerThread>,
                                                          Sequence<0, 1, 2, 3>,
                                                          3,
                                                          A_K1,
                                                          A_K1>;
  
     using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
                                                          ComputeTypeB,
                                                          decltype(b_block_desc_n0_n1_n2_k),
                                                          decltype(b_thread_desc_),
                                                          Sequence<1, 1, 1, KPerThread>,
                                                          Sequence<0, 1, 2, 3>,
                                                          3,
                                                          B_K1,
                                                          B_K1>;
  
     AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
     BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
 };
  
 } // namespace ck