ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-05 22:22:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Fix flatmm on gfx11 and gfx12 (#2790)

Browse Source 
1. Correct shuffle_b and MakeBFlatDramTileDistribution according to WMMA warp layout
2. Add FlatmmConfig16_Wmma for gfx11 and gfx12
This commit is contained in:
linqunAMD
2025-09-10 08:28:00 +08:00
committed by GitHub
parent 82890192dd
commit df4ee556d6
14 changed files with 224 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
example/ck_tile/17_grouped_gemm/grouped_gemm.hpp
View File

@@ -190,6 +190,30 @@ struct GemmConfigPreshufflePrefill : public GemmConfigBase
static constexpr bool kPadK = true;
};
template <typename PrecType>
struct GemmConfigPreshuffleDecode_Wmma : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 32 / sizeof(PrecType);
static constexpr ck_tile::index_t N_Tile = 64;
static constexpr ck_tile::index_t K_Tile = 256 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 4;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = 16;
static constexpr bool kPadK = true;
static constexpr int kBlockPerCu = 1;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_PRESHUFFLE_V2;
static constexpr bool Preshuffle = true;
static constexpr bool DoubleSmemBuffer = true;
};
template <ck_tile::index_t PipelineId>
struct PipelineTypeTraits;
@@ -266,16 +290,43 @@ template <typename GemmConfig, typename T>
auto shuffle_b(const ck_tile::HostTensor<T>& t)
{
assert(t.get_lengths().size() == 2);
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[0];
constexpr int divisor = GemmConfig::N_Warp_Tile == 32 ? 2 : 4;
ck_tile::HostTensor<T> t_view({n_ / GemmConfig::N_Warp_Tile,
GemmConfig::N_Warp_Tile,
k_ / GemmConfig::K_Warp_Tile,
divisor,
GemmConfig::K_Warp_Tile / divisor});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 3, 1, 4});
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[0];
if(ck_tile::is_gfx12_supported())
{
// TODO: Please modify it once kABK0PerLane is changed in WmmaTraitsBase<gfx12>
constexpr int divisor = 2;
constexpr int kABK0PerLane = 2;
ck_tile::HostTensor<T> t_view({n_ / GemmConfig::N_Warp_Tile,
GemmConfig::N_Warp_Tile,
k_ / GemmConfig::K_Warp_Tile,
divisor,
kABK0PerLane,
GemmConfig::K_Warp_Tile / divisor / kABK0PerLane});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 4, 1, 3, 5});
}
else
{
int divisor = 1;
if(ck_tile::is_gfx11_supported())
{
divisor = 1;
}
else
{
assert(is_wave32() == false);
divisor = GemmConfig::N_Warp_Tile == 32 ? 2 : 4;
}
ck_tile::HostTensor<T> t_view({n_ / GemmConfig::N_Warp_Tile,
GemmConfig::N_Warp_Tile,
k_ / GemmConfig::K_Warp_Tile,
divisor,
GemmConfig::K_Warp_Tile / divisor});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 3, 1, 4});
}
}
template <typename GemmConfig,