Skip to content

refactor: unify interwave and intrawave pipeline implementation #3552

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

Open
AviralGoelAMD wants to merge 7 commits into
base: develop
Choose a base branch
from
+40 −339
Open

refactor: unify interwave and intrawave pipeline implementation #3552

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
b06f6f5
refactor: remove Default scheduler implementation as it not used anymore
AviralGoelAMD Jan 10, 2026
09aa536
refactor: remove dead code from gemm universal kernel
AviralGoelAMD Jan 10, 2026
958c99e
chore: add descriptive comments about amd intrinsic hardware sync ins…
AviralGoelAMD Jan 10, 2026
2240aa5
fix: label existing memory pipeline for aquant as intrawave
AviralGoelAMD Jan 9, 2026
0f2c913
refactor: unify interwave and intrawave pipeline implementation at wo…
AviralGoelAMD Jan 12, 2026
7c78c99
Merge branch 'develop' into aviralgoel/memory_pipeline_refactor_2
AviralGoelAMD Jan 13, 2026
d45b830
Merge branch 'develop' into aviralgoel/memory_pipeline_refactor_2
AviralGoelAMD Jan 13, 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
36 changes: 27 additions & 9 deletions include/ck_tile/ops/gemm/block/block_universal_gemm_as_bs_cr.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -224,21 +224,28 @@ struct BlockUniversalGemmAsBsCr
}

// C += A * B
template <typename CBlockTensor,
template <bool DoLocalPrefetch = true,
typename CBlockTensor,
typename ASmemBlockWindow,
typename BSmemBlockWindow,
bool ALoadTranspose = false,
bool BLoadTranspose = false>
CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
const ASmemBlockWindow&,
const BSmemBlockWindow&,
bool_constant<ALoadTranspose> = {},
bool_constant<BLoadTranspose> = {})
const ASmemBlockWindow& a_block_window,
const BSmemBlockWindow& b_block_window,
bool_constant<ALoadTranspose> a_load_tr = {},
bool_constant<BLoadTranspose> b_load_tr = {})
{
static_assert(std::is_same_v<CDataType, typename CBlockTensor::DataType>,
"The CDataType as defined in traits should be the same as correspoinding "
"C block tensor data type!");

// LocalPrefetch can be controlled by pipeline via DoLocalPrefetch parameter
if constexpr(DoLocalPrefetch)
{
LocalPrefetch(a_block_window, b_block_window, a_load_tr, b_load_tr);
}

// hot loop:
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
Expand Down Expand Up @@ -275,6 +282,7 @@ struct BlockUniversalGemmAsBsCr
});
});
});
block_sync_lds();
}
};

Expand Down Expand Up @@ -355,7 +363,8 @@ struct BlockUniversalGemmAsBsCr
}

// C += A * B
template <typename CBlockTensor,
template <bool DoLocalPrefetch = true,
typename CBlockTensor,
typename ASmemBlockWindow,
typename BSmemBlockWindow,
bool ALoadTranspose = false,
Expand All @@ -372,7 +381,14 @@ struct BlockUniversalGemmAsBsCr

// hot loop:
static_for<0, KRepeat, 1>{}([&](auto kIter) {
LocalPrefetch<kIter.value>(a_block_window, b_block_window, a_load_tr, b_load_tr);
// Note: Interwave scheduler requires LocalPrefetch as part of its design
// The DoLocalPrefetch flag is provided for API consistency but should typically be
// true
if constexpr(DoLocalPrefetch)
{
LocalPrefetch<kIter.value>(
a_block_window, b_block_window, a_load_tr, b_load_tr);
}
Comment on lines +384 to +391
Copy link

Copilot AI Jan 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The comment states that LocalPrefetch is required for Interwave scheduler design, but the code still allows it to be disabled via DoLocalPrefetch=false. Consider either enforcing this requirement with a static_assert or updating the comment to clarify what happens if DoLocalPrefetch is false for Interwave.

Copilot uses AI. Check for mistakes.
__builtin_amdgcn_sched_barrier(
0); // Complete scheduling all pending instruction groups before this point

Expand Down Expand Up @@ -494,7 +510,8 @@ struct BlockUniversalGemmAsBsCr
}

// C += A * B
template <typename CBlockTensor,
template <bool DoLocalPrefetch = true,
typename CBlockTensor,
typename ASmemBlockWindow,
typename BSmemBlockWindow,
bool ALoadTranspose = false,
Expand All @@ -505,7 +522,8 @@ struct BlockUniversalGemmAsBsCr
bool_constant<ALoadTranspose> a_load_tr = {},
bool_constant<BLoadTranspose> b_load_tr = {})
{
block_gemm_impl_(c_block_tensor, a_block_window, b_block_window, a_load_tr, b_load_tr);
block_gemm_impl_.template operator()<DoLocalPrefetch>(
c_block_tensor, a_block_window, b_block_window, a_load_tr, b_load_tr);
}

// C = A * B
Expand Down
19 changes: 10 additions & 9 deletions include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -569,9 +569,9 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
load_tile_with_elementwise(b_copy_dram_window, b_element_func);
move_tile_window(b_copy_dram_window, b_dram_tile_window_step);

block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);

block_sync_lds();
block_gemm.template operator()<false>(
c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
// block_sync_lds();
Copy link

Copilot AI Jan 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Consider removing the commented-out block_sync_lds() calls rather than leaving them in the code. If these are no longer needed due to the refactoring (where synchronization is now handled within the BlockGemm operator), they should be removed to avoid confusion. If they might be needed for debugging or future reference, add a comment explaining why they are commented out.

Suggested change
// block_sync_lds();

Copilot uses AI. Check for mistakes.

block_gemm.LocalPrefetch(
a_lds_gemm_window, b_lds_gemm_window, is_a_load_tr_v, is_b_load_tr_v);
Expand All @@ -586,12 +586,14 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
{
// Leak last MFMA block to epilogue region, cover the potential lds-shuffle
// latency
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
block_gemm.template operator()<false>(
c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
}
else
{
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
block_sync_lds();
block_gemm.template operator()<false>(
c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
// block_sync_lds();
Copy link

Copilot AI Jan 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Consider removing the commented-out block_sync_lds() calls rather than leaving them in the code. If these are no longer needed due to the refactoring (where synchronization is now handled within the BlockGemm operator), they should be removed to avoid confusion. If they might be needed for debugging or future reference, add a comment explaining why they are commented out.

Suggested change
// block_sync_lds();

Copilot uses AI. Check for mistakes.

if constexpr(is_a_col_major && !is_a_load_tr_v())
{
Expand All @@ -616,9 +618,8 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
Base::LocalPrefill(b_copy_lds_window, elementwise_Bs_res);
}
block_sync_lds();
block_gemm.LocalPrefetch(
a_lds_gemm_window, b_lds_gemm_window, is_a_load_tr_v, is_b_load_tr_v);
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
block_gemm.template operator()<true>(
c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
}
// __builtin_amdgcn_sched_barrier(0);
return c_block_tile;
Expand Down
Loading