Skip to content

feat(ck-tile): Enabled the unification framework #11646

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
illsilin merged 2 commits into from
Sep 14, 2026
Merged

feat(ck-tile): Enabled the unification framework #11646

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0f39175
Enabled the unification framework
chris-tsiaousis-hpc Sep 3, 2026
035087b
Fix compilation issue for FMHA with K=48 on gfx1201
chris-tsiaousis-hpc Sep 3, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion projects/composablekernel/include/ck_tile/core/arch/arch.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,7 +47,7 @@
// get rerouted to the UnificationDispatcher. The latter is necessary because some pipelines bypass
// the WarpGemmDispatcher in favor of directly using named WarpGemms.
#ifndef USE_NEW_UNIFIED_FRAMEWORK
#define USE_NEW_UNIFIED_FRAMEWORK 0
#define USE_NEW_UNIFIED_FRAMEWORK 1
#endif

namespace ck_tile {
Expand Down
29 changes: 20 additions & 9 deletions projects/composablekernel/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -579,17 +579,22 @@ struct BlockFmhaPipelineQRKSVS
{
static_for<1, kGemm0KItersPerBlock, 1>{}([&](auto i_tail_k_iter) {
constexpr index_t kTailKIters = i_tail_k_iter;
constexpr index_t kTailK0 = kTailKIters * kGemm0WarpK;

if(gemm0_tail_k_iters == kTailKIters)
{
// gfx12 WMMA accumulates (c = a*b + c), so a K=kTailK0 tail
// equals kTailKIters accumulating warp-K(=kGemm0WarpK) sub-GEMMs.
// Slicing the Q A-operand register tile at warp-K granularity
// divides the register layout cleanly for any tail multiple of
// kGemm0WarpK (16/32/48), avoiding the reverse_slice_sequence
// static_assert that a monolithic K=48 slice triggers on WMMA.
using Gemm0TailProblem = BlockGemmProblem<
QDataType,
KDataType,
SaccDataType,
Problem::kNumGemm0Warps * get_warp_size(),
TileGemmShape<
sequence<kM0, kN0, kTailK0>,
sequence<kM0, kN0, kGemm0WarpK>,
typename BlockFmhaShape::Gemm0BlockWarps,
sequence<BlockFmhaShape::Gemm0WarpTile::at(number<0>{}),
BlockFmhaShape::Gemm0WarpTile::at(number<1>{}),
Expand All @@ -598,14 +603,20 @@ struct BlockFmhaPipelineQRKSVS
BlockGemmARegBSmemCRegV2<Gemm0TailProblem,
typename BlockGemm0::Policy>{};

auto q_slice =
get_slice_tile(q_tile,
sequence<0, i_k0 * kK0>{},
sequence<kM0, i_k0 * kK0 + kTailK0>{});
auto k_tail_window = make_tile_window(
k_lds, make_tuple(number<kN0>{}, number<kTailK0>{}), {0, 0});
static_for<0, kTailKIters, 1>{}([&](auto j) {
constexpr index_t kSubKOffset = j.value * kGemm0WarpK;

gemm_0_tail(s_acc, q_slice, k_tail_window);
auto q_slice = get_slice_tile(
q_tile,
sequence<0, i_k0 * kK0 + kSubKOffset>{},
sequence<kM0, i_k0 * kK0 + kSubKOffset + kGemm0WarpK>{});
auto k_tail_window = make_tile_window(
k_lds,
make_tuple(number<kN0>{}, number<kGemm0WarpK>{}),
{0, kSubKOffset});

gemm_0_tail(s_acc, q_slice, k_tail_window);
});
}
});
return;
Expand Down
Loading