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

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

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/warp/smfmac_xdlops_gemm.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/math.hpp"
 #include "ck/utility/amd_smfmac.hpp"
  
 namespace ck {
  
 enum struct SmfmacInstr
 {
     smfmac_f32_16x16x32f16 = 0,
     smfmac_f32_32x32x16f16,
     smfmac_f32_16x16x32bf16,
     smfmac_f32_32x32x16bf16,
 };
  
 template <SmfmacInstr instr>
 struct smfmac_type;
  
 template <>
 struct smfmac<SmfmacInstr::smfmac_f32_16x16x32f16>
 {
     static constexpr index_t group_size          = 4;
     static constexpr index_t num_groups_per_blk  = 1;
     static constexpr index_t num_regs_per_blk    = 4;
     static constexpr index_t num_threads_per_blk = 16;
     static constexpr index_t wave_size           = 64;
     static constexpr index_t num_input_blks      = 4;
     static constexpr index_t num_output_blks     = 1;
     static constexpr index_t m_per_blk           = 16;
     static constexpr index_t n_per_blk           = 16;
     static constexpr index_t k_per_blk           = 8;
     static constexpr bool is_k_reduction         = true;
  
     template <index_t MPerXdlops,
               index_t NPerXdlops,
               index_t idx_part,
               class FloatA,
               class FloatB,
               class FloatC>
     __device__ void run(const FloatA& a, const FloatB& b, const index_t& idx, FloatC& reg_c) const
     {
         intrin_smfmac_f32_16x16x32f16<MPerXdlops, NPerXdlops>::Run<FloatC, idx_part>(
             a, b, idx, reg_c);
     }
 };
  
 template <>
 struct smfmac<SmfmacInstr::smfmac_f32_32x32x16f16>
 {
     static constexpr index_t group_size          = 4;
     static constexpr index_t num_groups_per_blk  = 4;
     static constexpr index_t num_regs_per_blk    = 16;
     static constexpr index_t num_threads_per_blk = 32;
     static constexpr index_t wave_size           = 64;
     static constexpr index_t num_input_blks      = 2;
     static constexpr index_t num_output_blks     = 1;
     static constexpr index_t m_per_blk           = 32;
     static constexpr index_t n_per_blk           = 32;
     static constexpr index_t k_per_blk           = 16;
     static constexpr bool is_k_reduction         = true;
  
     template <index_t MPerXdlops,
               index_t NPerXdlops,
               index_t idx_part,
               class FloatA,
               class FloatB,
               class FloatC>
     __device__ void run(const FloatA& a, const FloatB& b, const index_t& idx, FloatC& reg_c) const
     {
         intrin_smfmac_f32_32x32x16f16<MPerXdlops, NPerXdlops>::Run<FloatC, idx_part>(
             a, b, idx, reg_c);
     }
 };
  
 template <>
 struct smfmac<SmfmacInstr::smfmac_f32_16x16x32bf16>
 {
     static constexpr index_t group_size          = 4;
     static constexpr index_t num_groups_per_blk  = 1;
     static constexpr index_t num_regs_per_blk    = 4;
     static constexpr index_t num_threads_per_blk = 16;
     static constexpr index_t wave_size           = 64;
     static constexpr index_t num_input_blks      = 4;
     static constexpr index_t num_output_blks     = 1;
     static constexpr index_t m_per_blk           = 16;
     static constexpr index_t n_per_blk           = 16;
     static constexpr index_t k_per_blk           = 8;
     static constexpr bool is_k_reduction         = true;
  
     template <index_t MPerXdlops,
               index_t NPerXdlops,
               index_t idx_part,
               class FloatA,
               class FloatB,
               class FloatC>
     __device__ void run(const FloatA& a, const FloatB& b, const index_t& idx, FloatC& reg_c) const
     {
         intrin_smfmac_f32_16x16x32bf16<MPerXdlops, NPerXdlops>::Run<FloatC, idx_part>(
             a, b, idx, reg_c);
     }
 };
  
 template <>
 struct smfmac<SmfmacInstr::smfmac_f32_32x32x16bf16>
 {
     static constexpr index_t group_size          = 4;
     static constexpr index_t num_groups_per_blk  = 4;
     static constexpr index_t num_regs_per_blk    = 16;
     static constexpr index_t num_threads_per_blk = 32;
     static constexpr index_t wave_size           = 64;
     static constexpr index_t num_input_blks      = 2;
     static constexpr index_t num_output_blks     = 1;
     static constexpr index_t m_per_blk           = 32;
     static constexpr index_t n_per_blk           = 32;
     static constexpr index_t k_per_blk           = 16;
     static constexpr bool is_k_reduction         = true;
  
     template <index_t MPerXdlops,
               index_t NPerXdlops,
               index_t idx_part,
               class FloatA,
               class FloatB,
               class FloatC>
     __device__ void run(const FloatA& a, const FloatB& b, const index_t& idx, FloatC& reg_c) const
     {
         intrin_smfmac_f32_32x32x16bf16<MPerXdlops, NPerXdlops>::Run<FloatC, idx_part>(
             a, b, idx, reg_c);
     }
 };
  
 template <typename base_type,
           index_t MPerXdlops,
           index_t NPerXdlops,
           typename additional_type = base_type>
 struct SmfmacSelector
 {
     template <typename base_type_,
               index_t MPerXdlops_,
               index_t NPerXdlops_,
               typename additional_type_ = base_type_>
     static constexpr auto GetSmfmac();
  
     template <>
     static constexpr auto GetSmfmac<half_t, 16, 16>()
     {
         return SmfmacInstr::smfmac_f32_16x16x32f16;
     }
  
     template <>
     static constexpr auto GetSmfmac<half_t, 32, 32>()
     {
         return SmfmacInstr::smfmac_f32_32x32x16f16;
     }
  
     template <>
     static constexpr auto GetSmfmac<bhalf_t, 16, 16>()
     {
         return SmfmacInstr::smfmac_f32_16x16x32bf16;
     }
  
     template <>
     static constexpr auto GetSmfmac<bhalf_t, 32, 32>()
     {
         return SmfmacInstr::smfmac_f32_32x32x16bf16;
     }
  
     static constexpr auto selected_smfmac =
         smfmac_type<GetSmfmac<base_type, MPerXdlops, NPerXdlops, additional_type>()>{};
  
     __host__ __device__ constexpr SmfmacSelector()
     {
         static_assert(selected_smfmac.group_size * selected_smfmac.num_groups_per_blk ==
                           selected_smfmac.num_regs_per_blk,
                       "wrong! num_regs_per_blk");
  
         static_assert(selected_smfmac.num_threads_per_blk == selected_smfmac.n_per_blk,
                       "n_per_blk != num_threads_per_blk");
  
         static_assert(selected_smfmac.num_regs_per_blk * selected_smfmac.num_input_blks ==
                           selected_smfmac.m_per_blk,
                       "m_per_blk != num_input_blks * num_regs_per_blk");
  
         static_assert(selected_smfmac.num_output_blks == selected_smfmac.num_input_blks ||
                           selected_smfmac.num_output_blks == 1,
                       "incorrect num_output_blks");
  
         static_assert(selected_smfmac.num_regs_per_blk * selected_smfmac.wave_size ==
                           selected_smfmac.m_per_blk * selected_smfmac.n_per_blk,
                       "num_regs_per_blk incorrect");
  
         static_assert(selected_smfmac.is_k_reduction ||
                           (selected_smfmac.num_input_blks == selected_smfmac.num_output_blks),
                       "is_k_reduction wrong!");
     }
  
     static constexpr index_t GetKPerXdlops()
     {
         return (selected_smfmac.is_k_reduction ? selected_smfmac.num_input_blks : 1) *
                selected_smfmac.k_per_blk;
     }
  
     static constexpr index_t GetK1PerXdlops() { return selected_smfmac.k_per_blk; }
 };
  
 template <typename base_type,
           index_t MPerXdlops,
           index_t NPerXdlops,
           index_t KPack,
           typename additional_type = base_type>
 struct SparseXdlopsGemm
 {
     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>{};
  
     using CIndex   = MultiIndex<2>;
     using CIndex4D = MultiIndex<4>;
  
     __device__ static constexpr index_t GetNumBlks() { return smfmac_instr.num_output_blks; }
  
     __device__ static constexpr index_t GetNumXdlops()
     {
         return MPerXdlops * NPerXdlops /
                (smfmac_instr.m_per_blk * smfmac_instr.n_per_blk * smfmac_instr.num_output_blks);
     }
  
     __host__ __device__ constexpr SparseXdlopsGemm()
     {
         static_assert(NPerXdlops == 16 || NPerXdlops == 32,
                       "Only support GemmNPerXdlops == 16 or 32 for smfmac xdlops");
  
         static_assert(MPerXdlops == 16 || MPerXdlops == 32,
                       "Only support GemmMPerXdlops == 16 or 32 for smfmac xdlops");
  
         static_assert(KPack % smfmac_instr.k_per_blk == 0, "KPack cannot be divided by k_per_blk");
     }
  
     // XDL output supporting C = A * B
     // M2_N2 -> M2_M3_M4_N2
     template <typename CDesc_M0_N0_M1_N1_M2_N2>
     __host__ __device__ static constexpr auto
     MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CDesc_M0_N0_M1_N1_M2_N2& c_desc_m0_n0_m1_n1_m2_n2)
     {
         const auto M0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I0);
         const auto N0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I1);
         const auto M1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I2);
         const auto N1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I3);
  
         return transform_tensor_descriptor(
             c_desc_m0_n0_m1_n1_m2_n2,
             make_tuple(make_pass_through_transform(M0),
                        make_pass_through_transform(N0),
                        make_pass_through_transform(M1),
                        make_pass_through_transform(N1),
                        make_unmerge_transform(make_tuple(Number<smfmac_instr.num_groups_per_blk>{},
                                                          Number<smfmac_instr.num_input_blks>{},
                                                          Number<smfmac_instr.group_size>{})),
                        make_pass_through_transform(Number<smfmac_instr.num_threads_per_blk>{})),
             make_tuple(Sequence<0>{},
                        Sequence<1>{},
                        Sequence<2>{},
                        Sequence<3>{},
                        Sequence<4>{},
                        Sequence<5>{}),
             make_tuple(Sequence<0>{},
                        Sequence<1>{},
                        Sequence<2>{},
                        Sequence<3>{},
                        Sequence<4, 5, 6>{},
                        Sequence<7>{}));
     }
  
     template <typename CDesc_G_M0_N0_M1_N1_M2_N2>
     __host__ __device__ static constexpr auto MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
         const CDesc_G_M0_N0_M1_N1_M2_N2& c_desc_g_m0_n0_m1_n1_m2_n2)
     {
         const auto G  = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I0);
         const auto M0 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I1);
         const auto N0 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I2);
         const auto M1 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I3);
         const auto N1 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I4);
  
         return transform_tensor_descriptor(
             c_desc_g_m0_n0_m1_n1_m2_n2,
             make_tuple(make_pass_through_transform(G),
                        make_pass_through_transform(M0),
                        make_pass_through_transform(N0),
                        make_pass_through_transform(M1),
                        make_pass_through_transform(N1),
                        make_unmerge_transform(make_tuple(smfmac_instr.num_groups_per_blk,
                                                          smfmac_instr.num_input_blks,
                                                          smfmac_instr.group_size)),
                        make_pass_through_transform(smfmac_instr.num_threads_per_blk)),
             make_tuple(Sequence<0>{},
                        Sequence<1>{},
                        Sequence<2>{},
                        Sequence<3>{},
                        Sequence<4>{},
                        Sequence<5>{},
                        Sequence<6>{}),
             make_tuple(Sequence<0>{},
                        Sequence<1>{},
                        Sequence<2>{},
                        Sequence<3>{},
                        Sequence<4>{},
                        Sequence<5, 6, 7>{},
                        Sequence<8>{}));
     }
  
     __device__ static constexpr index_t GetRegSizePerXdlops()
     {
         return MPerXdlops * NPerXdlops / smfmac_instr.wave_size;
     }
  
     __device__ static constexpr index_t GetWaveSize() { return smfmac_instr.wave_size; }
  
     template <class FloatA, class FloatB, class Idx, class FloatC>
     __device__ void
     Run(const FloatA& p_a_wave, const FloatB& p_b_wave, const Idx& idx, FloatC& p_c_thread) const
     {
         static_assert(is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value,
                       "base base_type must be half or bfloat16!");
  
         static_for<0, KPack / smfmac_instr.k_per_blk, 1>{}([&](auto k) {
             smfmac_instr.template run<MPerXdlops, NPerXdlops, k % 4>(
                 p_a_wave[k], p_b_wave[k], idx[k / 4], p_c_thread);
         });
     }
  
     __device__ static auto GetLaneId() { return get_thread_local_1d_id() % smfmac_instr.wave_size; }
  
     __device__ static auto GetBlkIdx()
     {
         const auto laneId = GetLaneId();
  
         constexpr auto threadidx_to_blk_idx_adaptor = make_single_stage_tensor_adaptor(
             make_tuple(make_merge_transform(
                 make_tuple(1, smfmac_instr.num_input_blks, smfmac_instr.num_threads_per_blk))),
             make_tuple(Sequence<0, 1, 2>{}),
             make_tuple(Sequence<0>{}));
  
         const auto blk_idx =
             threadidx_to_blk_idx_adaptor.CalculateBottomIndex(make_multi_index(laneId));
  
         const auto blk_id = blk_idx[I1];
         const auto blk_td = blk_idx[I2];
  
         return make_tuple(blk_id, blk_td);
     }
  
     __host__ __device__ static auto CalculateAThreadOriginDataIndex()
     {
         const auto laneId  = GetLaneId();
         const auto blk_idx = GetBlkIdx();
  
         const auto blk_id = blk_idx[I0];
         const auto blk_td = blk_idx[I1];
  
         if constexpr(smfmac_instr.is_k_reduction)
         {
             return make_tuple(blk_id, blk_td);
         }
         else
         {
             return make_tuple(0, laneId);
         }
     }
  
     __host__ __device__ static auto CalculateBThreadOriginDataIndex()
     {
         const auto laneId  = GetLaneId();
         const auto blk_idx = GetBlkIdx();
  
         const auto blk_id = blk_idx[I0];
         const auto blk_td = blk_idx[I1];
  
         if constexpr(smfmac_instr.is_k_reduction)
         {
             return make_tuple(blk_id, blk_td);
         }
         else
         {
             return make_tuple(0, laneId);
         }
     }
  
     __device__ static CIndex GetBeginOfThreadBlk(index_t xdlops_i, index_t blk_i)
     {
         const auto blk_idx = GetBlkIdx();
  
         const auto blk_id = blk_idx[I0];
         const auto blk_td = blk_idx[I1];
  
         index_t n_offset = blk_i * smfmac_instr.n_per_blk + blk_td;
         index_t m_offset = xdlops_i * smfmac_instr.m_per_blk + blk_id * smfmac_instr.group_size;
  
         return CIndex{m_offset, n_offset};
     }
  
     __device__ static CIndex4D GetBeginOfThreadBlk4D(index_t /* xdlops_i */, index_t /* blk_i */)
     {
         const auto blk_idx = GetBlkIdx();
  
         const auto blk_id = blk_idx[I0];
         const auto blk_td = blk_idx[I1];
  
         return CIndex4D{I0, blk_id, I0, blk_td};
     }
  
     static constexpr auto smfmac =
         SmfmacSelector<base_type, MPerXdlops, NPerXdlops, additional_type>{};
  
     static constexpr auto smfmac_instr = smfmac.selected_smfmac;
  
     static constexpr auto KPerXdlops  = smfmac.GetKPerXdlops();
     static constexpr auto K1PerXdlops = smfmac.GetK1PerXdlops();
     static constexpr auto K0PerXdlops = KPerXdlops / K1PerXdlops;
  
     __host__ __device__ static constexpr auto GetCM0M1M2NThreadBlkLengths()
     {
         return make_tuple(
             Number<smfmac_instr.num_groups_per_blk>{}, I1, Number<smfmac_instr.group_size>{}, I1);
     }
 };
  
 } // namespace ck