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

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

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/gemm/pipeline/tile_gemm_shape.hpp Source File

Go to the documentation of this file.

 // Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 // SPDX-License-Identifier: MIT
  
 #pragma once
  
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/concat.hpp"
  
 namespace ck_tile {
  
 template <typename BlockTile_,
           typename BlockWarps_,
           typename WarpTile_,
           bool PermuteA_ = false,
           bool PermuteB_ = false>
 struct TileGemmShape
 {
     using BlockTile  = remove_cvref_t<BlockTile_>;
     using BlockWarps = remove_cvref_t<BlockWarps_>;
     using WarpTile   = remove_cvref_t<WarpTile_>;
  
     static constexpr index_t NumWarps = reduce_on_sequence(BlockWarps{}, multiplies{}, number<1>{});
  
     static constexpr index_t kM = BlockTile::at(number<0>{});
     static constexpr index_t kN = BlockTile::at(number<1>{});
     static constexpr index_t kK = BlockTile::at(number<2>{});
  
     static constexpr bool PermuteA = PermuteA_;
     static constexpr bool PermuteB = PermuteB_;
  
     static constexpr index_t flatNPerWarp  = BlockWarps::at(number<1>{});
     static constexpr index_t flatKPerWarp  = WarpTile::at(number<2>{}) * WarpTile::at(number<1>{});
     static constexpr index_t flatKPerBlock = flatKPerWarp * kK / WarpTile::at(number<2>{});
  
     CK_TILE_HOST static std::string GetName()
     {
         // clang-format off
         return concat('_', "tile_gemm_shape",
                       concat('x', kM, kN, kK, NumWarps),
                       concat('x', BlockWarps::at(number<0>{}), BlockWarps::at(number<1>{}), BlockWarps::at(number<2>{})),
                       concat('x', (WarpTile::at(number<0>{})), WarpTile::at(number<1>{}), WarpTile::at(number<2>{})));
         // clang-format on
     }
 };
  
 template <typename PrecType, index_t M_Warp_Tile, bool IsFlatMM = false>
 constexpr index_t get_k_warp_tile()
 {
 #if CK_TILE_USE_WMMA
     return 16;
 #else
 #if defined(CK_GFX950_SUPPORT)
     constexpr bool is_8bit_float =
         std::is_same_v<PrecType, fp8_t> || std::is_same_v<PrecType, bf8_t>;
     if constexpr(M_Warp_Tile == 32)
         return is_8bit_float ? 64 : 16;
     else
         return is_8bit_float ? 128 : 32;
 #else
     if constexpr(M_Warp_Tile == 32)
         return (sizeof(PrecType) == 2 || IsFlatMM == false) ? 16 : 32;
     else
         return (sizeof(PrecType) == 2 || IsFlatMM == false) ? 32 : 64;
 #endif
 #endif
 }
  
 } // namespace ck_tile