mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-04 05:31:24 +00:00
Add permuteN optimzization when NRepeat % 2 == 0 on flatmm
This commit is contained in:
@@ -41,6 +41,7 @@ struct FlatmmConfig32
|
||||
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
|
||||
static constexpr ck_tile::index_t NumWaveGroups = 1;
|
||||
static constexpr bool DoubleSmemBuffer = false;
|
||||
static constexpr bool TiledMMAPermuteN = false; // disable PermuteN when NWarpTile != 16
|
||||
};
|
||||
|
||||
template <typename DataType>
|
||||
@@ -78,6 +79,9 @@ struct FlatmmConfig16
|
||||
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
|
||||
static constexpr ck_tile::index_t NumWaveGroups = 1;
|
||||
static constexpr bool DoubleSmemBuffer = false;
|
||||
|
||||
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
|
||||
static constexpr bool TiledMMAPermuteN = N_Repeat % 2 == 0;
|
||||
};
|
||||
|
||||
template <typename DataType>
|
||||
@@ -87,6 +91,10 @@ struct FlatmmConfig16_950 : public FlatmmConfig16<DataType>
|
||||
static constexpr ck_tile::index_t K_Tile = 256 / sizeof(DataType);
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(DataType) == 2 ? 32 : 128;
|
||||
static constexpr int kBlockPerCu = 1;
|
||||
|
||||
static constexpr int N_Repeat =
|
||||
N_Tile / FlatmmConfig16<DataType>::N_Warp_Tile / FlatmmConfig16<DataType>::N_Warp;
|
||||
static constexpr bool TiledMMAPermuteN = N_Repeat % 2 == 0;
|
||||
};
|
||||
|
||||
template <typename ADataType>
|
||||
|
||||
Reference in New Issue
Block a user