/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/fmha/pipeline/tile_fmha_shape.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/fmha/pipeline/tile_fmha_shape.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/fmha/pipeline/tile_fmha_shape.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_tile/core.hpp"
  
 namespace ck_tile {
  
 static CK_TILE_HOST_DEVICE constexpr index_t ceil_to_qualified_tile_length(index_t len)
 {
     if(len == 96)
         return 128;
     if(len == 160)
         return 256;
     if(len == 192)
         return 192;
  
     // only length of 96, 160 and power-of-two is supported
     if(!(len & (len - 1)))
         return len;
  
     return 0;
 };
  
 template <typename BlockTile_, // sequence<...
           typename Gemm0BlockWarps_,
           typename Gemm0WarpTile_,
           typename Gemm1BlockWarps_,
           typename Gemm1WarpTile_,
           bool IsVLayoutRowMajor_>
 struct TileFmhaShape
 {
     using BlockTile       = remove_cvref_t<BlockTile_>;
     using Gemm0BlockWarps = remove_cvref_t<Gemm0BlockWarps_>;
     using Gemm0WarpTile   = remove_cvref_t<Gemm0WarpTile_>;
     using Gemm1BlockWarps = remove_cvref_t<Gemm1BlockWarps_>;
     using Gemm1WarpTile   = remove_cvref_t<Gemm1WarpTile_>;
  
     static constexpr index_t NumGemm0Warps =
         reduce_on_sequence(Gemm0BlockWarps{}, multiplies{}, number<1>{});
     static constexpr index_t NumGemm1Warps =
         reduce_on_sequence(Gemm1BlockWarps{}, multiplies{}, number<1>{});
     static_assert(NumGemm1Warps % NumGemm0Warps == 0);
  
     static constexpr index_t NumWarps = max(NumGemm0Warps, NumGemm1Warps);
  
     static constexpr index_t kM0 = BlockTile::at(number<0>{}); // tile size along q seqlen
     static constexpr index_t kN0 = BlockTile::at(number<1>{}); // tile size along k seqlen
     static constexpr index_t kK0 = BlockTile::at(number<2>{}); // tile size along qk gemm unroll
     static constexpr index_t kN1 = BlockTile::at(number<3>{}); // tile size along v head_dim
     static constexpr index_t kK1 = BlockTile::at(number<4>{}); // tile size along kv gemm unroll
     static constexpr index_t kQKHeaddim =
         BlockTile::at(number<5>{}); // total length of K0, used for pipeline that need load Q at
                                     // once (or repeately load Q as a whole tile)
     static_assert(kQKHeaddim % kK0 == 0, "kQKHeaddim should be divisible by kK0");
  
     static constexpr index_t kSubQKHeaddim = ceil_to_qualified_tile_length(kQKHeaddim);
  
     // v, rowmajor : seqlen*hdim, colmajor : hdim*seqlen
     static constexpr bool IsVLayoutRowMajor = IsVLayoutRowMajor_;
     using VLayout                           = std::conditional_t<IsVLayoutRowMajor,
                                                                  ck_tile::tensor_layout::gemm::RowMajor,
                                                                  ck_tile::tensor_layout::gemm::ColumnMajor>;
 };
  
 template <typename BlockTile_, // sequence<...
           typename Gemm0BlockWarps_,
           typename Gemm0WarpTile_,
           typename Gemm1BlockWarps_,
           typename Gemm1WarpTile_,
           typename Gemm2BlockWarps_,
           typename Gemm2WarpTile_,
           typename Gemm3BlockWarps_,
           typename Gemm3WarpTile_,
           typename Gemm4BlockWarps_,
           typename Gemm4WarpTile_,
           index_t kMaxSeqLenQ_ = 0>
 struct TileFmhaBwdShape
 {
     using BlockTile       = remove_cvref_t<BlockTile_>;
     using Gemm0BlockWarps = remove_cvref_t<Gemm0BlockWarps_>;
     using Gemm0WarpTile   = remove_cvref_t<Gemm0WarpTile_>;
     using Gemm1BlockWarps = remove_cvref_t<Gemm1BlockWarps_>;
     using Gemm1WarpTile   = remove_cvref_t<Gemm1WarpTile_>;
     using Gemm2BlockWarps = remove_cvref_t<Gemm2BlockWarps_>;
     using Gemm2WarpTile   = remove_cvref_t<Gemm2WarpTile_>;
     using Gemm3BlockWarps = remove_cvref_t<Gemm3BlockWarps_>;
     using Gemm3WarpTile   = remove_cvref_t<Gemm3WarpTile_>;
     using Gemm4BlockWarps = remove_cvref_t<Gemm4BlockWarps_>;
     using Gemm4WarpTile   = remove_cvref_t<Gemm4WarpTile_>;
  
     static constexpr index_t NumWarps =
         reduce_on_sequence(Gemm0BlockWarps{}, multiplies{}, number<1>{});
  
     static_assert(NumWarps == reduce_on_sequence(Gemm1BlockWarps{}, multiplies{}, number<1>{}) &&
                   NumWarps == reduce_on_sequence(Gemm4BlockWarps{}, multiplies{}, number<1>{}));
  
     static constexpr index_t kM0 = BlockTile::at(number<0>{}); // tile size along q seqlen
     static constexpr index_t kN0 = BlockTile::at(number<1>{}); // tile size along k seqlen
     static constexpr index_t kK0 =
         BlockTile::at(number<2>{}); // tile size along gemm0(Q@K^T) unroll
     static constexpr index_t kK1 =
         BlockTile::at(number<3>{}); // tile size along gemm1(P^T@dO) unroll
     static constexpr index_t kK2 =
         BlockTile::at(number<4>{}); // tile size along gemm2(dO@V^T) unroll
     static constexpr index_t kK3 =
         BlockTile::at(number<5>{}); // tile size along gemm3(dS^T@Q) unroll
     static constexpr index_t kK4 = BlockTile::at(number<6>{}); // tile size along gemm4(dS@K) unroll
     static constexpr index_t kQKHeaddim =
         BlockTile::at(number<7>{}); // Q & K headdim, used for pipeline that need load Q/Q^T or
                                     // K/K^T at once
     static constexpr index_t kVHeaddim = BlockTile::at(number<8>{}); // V headdim, used for pipeline
                                                                      // that need load V at once
  
     static constexpr index_t kMaxSeqLenQ = kMaxSeqLenQ_;
     static_assert(kMaxSeqLenQ == kM0 || kMaxSeqLenQ == 0,
                   "kMaxSeqLenQ should be equal to kM0 or 0, if 0, it means seq len Q is unlimited");
 };
  
 } // namespace ck_tile