Skip to content

[CK_TILE] Stream-K XCD remapping #3652

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

Draft
arai713 wants to merge 1 commit into
base: develop
Choose a base branch
from
+137 −7
Draft

[CK_TILE] Stream-K XCD remapping #3652

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 8 additions & 3 deletions include/ck_tile/ops/gemm/kernel/streamk_gemm/streamk_gemm_kernel.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -673,10 +673,12 @@ struct StreamKKernel
__shared__ char smem_ptr_0[UniversalGemmKernel::GetSmemSize()];

index_t block_idx = ck_tile::get_block_1d_id();
index_t grid_size = kargs.tile_partitioner.grid_size().x;
index_t dp_num_loop = kargs.tile_partitioner.get_iters_per_tile();
index_t dp_ctas = kargs.tile_partitioner.get_dp_ctas();
bool is_dp_ctas = block_idx < kargs.tile_partitioner.get_dp_ctas();

block_idx = kargs.tile_partitioner.RemapXCD(block_idx, grid_size);
Copy link
Collaborator

@ecamartins ecamartins Jan 27, 2026

Choose a reason for hiding this comment

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

Since this change should only apply to gfx942 and gfx950, we likely need some kind of logic to determine whether we want to apply the remap or not. Also, depending on the variant of gfx942 there may be fewer than 8 XCDs, so I think some additional logic will be required to determine what value for num_xcds we want to use (rather than solely relying on the default of 8).

We want to avoid using preprocessor macros, so depending on what there is available in the HIP api, maybe we can query the number of XCDs from the device?

If not, we may need to consider other options. Perhaps we could consider some logic that follows a similar pattern to: include/ck_tile/ops/gemm/kernel/streamk_gemm/streamk_gemm_coherency.hpp.

// Check if at the data parallel section
if(is_dp_ctas)
{
Expand Down Expand Up @@ -706,11 +708,14 @@ struct StreamKKernel
__shared__ char smem_ptr_0[UniversalGemmKernel::GetSmemSize()];

index_t block_idx = ck_tile::get_block_1d_id();
index_t grid_size = kargs.tile_partitioner.grid_size().x;
index_t dp_num_loop = kargs.tile_partitioner.get_iters_per_tile();

// Data-parallel section
for(index_t tile_idx = block_idx; tile_idx < kargs.tile_partitioner.get_dp_tiles();
tile_idx += kargs.tile_partitioner.get_grid())
block_idx =
kargs.tile_partitioner.RemapXCD(block_idx, grid_size)
Copy link
Collaborator

@ecamartins ecamartins Jan 27, 2026

Choose a reason for hiding this comment

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

Same comment about how we shouldn't do this on all architectures. (See above)

// Data-parallel section
for(index_t tile_idx = block_idx; tile_idx < kargs.tile_partitioner.get_dp_tiles();
tile_idx += kargs.tile_partitioner.get_grid())
{
BaseGemm(kargs, tile_idx, dp_num_loop, 0, 0, kargs.K, smem_ptr_0);
block_sync_lds();
Expand Down
17 changes: 15 additions & 2 deletions include/ck_tile/ops/gemm/kernel/streamk_gemm/streamk_gemm_tile_partitioner.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -213,6 +213,19 @@ struct StreamKTilePartitionerBase
*/
CK_TILE_HOST index_t estimate_num_wgs_per_tile() const noexcept;

/**
* @brief XCDs access ids in round robin format, this function remaps the 1D ids to continguous
* XCD segments
*
* @param block_1d_id grid 1D id
* @param total_num_tiles size of the 1D grid
* @param NUM_XCDS number of XCDs
* @return index_t The id after XCD remap
*/
CK_TILE_HOST_DEVICE index_t RemapXCD(index_t block_1d_id,
index_t total_num_tiles,
index_t NUM_XCDS = 8) const noexcept;

protected:
index_t num_tiles_;
index_t grid_;
Expand Down Expand Up @@ -281,7 +294,7 @@ struct StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, true>
*
* @return dim_3 The launching grid size for the kernel.
*/
CK_TILE_HOST auto grid_size() const noexcept -> dim3;
CK_TILE_HOST_DEVICE auto grid_size() const noexcept -> dim3;

/**
* @brief Returns the total number of DP tiles per workgroup.
Expand Down Expand Up @@ -328,7 +341,7 @@ struct StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, false>
*
* @return dim_3 The launching grid size for the kernel.
*/
CK_TILE_HOST auto grid_size() const noexcept -> dim3;
CK_TILE_HOST_DEVICE auto grid_size() const noexcept -> dim3;

/**
* @brief Returns the total number of DP workgroups.
Expand Down
53 changes: 51 additions & 2 deletions include/ck_tile/ops/gemm/kernel/streamk_gemm/streamk_gemm_tile_partitioner_impl.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -279,6 +279,55 @@ StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::estimate_
return std::max(num_wgs_per_tile, 1);
}

/**
* @brief XCDs access ids in round robin format, this function remaps the 1D ids to continguous
* XCD segments
*
* @param block_1d_id grid 1D id
* @param total_num_tiles size of the 1D grid
* @param NUM_XCDS number of XCDs
* @return index_t The id after XCD remap
*/
template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
CK_TILE_HOST_DEVICE index_t
StreamKTilePartitionerBase<BlockGemmShapeType, ReductionStrategyType>::RemapXCD(
index_t block_1d_id, index_t total_num_tiles, index_t NUM_XCDS) const noexcept
{
// Number of ids per XCD in the new arrangement
index_t ids_per_xcd = (total_num_tiles + NUM_XCDS - 1) / NUM_XCDS;

// When total_num_tiles cannot divide NUM_XCDS, some xcds will have
// ids_per_xcd ids, the other will have ids_per_xcd - 1 ids.
// We calculate the number of xcds that have ids_per_xcd ids as tall_xcds
index_t tall_xcds = total_num_tiles % NUM_XCDS;
tall_xcds = (tall_xcds == 0) ? NUM_XCDS : tall_xcds;

// Compute current XCD and local id within the XCD
index_t xcd = block_1d_id % NUM_XCDS;
index_t local_id = block_1d_id / NUM_XCDS;

// Calculate new id based on the new grouping
if(xcd < tall_xcds)
{
block_1d_id = xcd * ids_per_xcd + local_id;
}
else
{
block_1d_id = tall_xcds * ids_per_xcd + (xcd - tall_xcds) * (ids_per_xcd - 1) + local_id;
}

/**
* original ids: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
* XCD 0 gets: [0, 8], XCD 1 gets: [1, 9], ...
*
* post-remap ids: [0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15]
* XCD 0 gets: [0, 1], XCD 1 gets: [2, 3], ...
*
* after remap the ids are continguous on each XCD
*/
return block_1d_id;
}

template <typename BlockGemmShapeType,
StreamKReductionStrategy ReductionStrategyType,
bool Persistent>
Expand All @@ -295,7 +344,7 @@ StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, true>::StreamK
}

template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
CK_TILE_HOST auto
CK_TILE_HOST_DEVICE auto
StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, true>::grid_size() const noexcept
-> dim3
{
Expand Down Expand Up @@ -337,7 +386,7 @@ StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, false>::Stream
}

template <typename BlockGemmShapeType, StreamKReductionStrategy ReductionStrategyType>
CK_TILE_HOST auto
CK_TILE_HOST_DEVICE auto
StreamKTilePartitioner<BlockGemmShapeType, ReductionStrategyType, false>::grid_size() const noexcept
-> dim3
{
Expand Down
46 changes: 46 additions & 0 deletions test/ck_tile/gemm_streamk/test_streamk_tile_partitioner.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -407,6 +407,52 @@ TEST(StreamKTilePartitionerBaseGetOutputTileIndex, TestAllMappings)
}
}

TEST(StreamKTilePartitionerBaseRemapXCD, SmallArray)
{
using Config = StreamKTilePartitionerBaseConfigSKOnly;

ck_tile::StreamKTilePartitioner_v2<Config::GemmShape,
ck_tile::StreamKReductionStrategy::Atomic,
true>
tile_partitioner{Config::M, Config::N, Config::K, Config::GRID};

const std::vector<ck_tile::index_t> initial_values = {
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
const std::vector<ck_tile::index_t> expected_values = {
0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15};

test_remap_xcd<Config::GemmShape>(initial_values, expected_values, tile_partitioner);
}

TEST(StreamKTilePartitionerBaseRemapXCD, MidArray)
{
using Config = StreamKTilePartitionerBaseConfigSKOnly;

ck_tile::StreamKTilePartitioner_v2<Config::GemmShape,
ck_tile::StreamKReductionStrategy::Atomic,
true>
tile_partitioner{Config::M, Config::N, Config::K, Config::GRID};

const std::vector<ck_tile::index_t> initial_values = {
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79,
80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95,
96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111,
112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126};
const std::vector<ck_tile::index_t> expected_values = {
0, 16, 32, 48, 64, 80, 96, 112, 1, 17, 33, 49, 65, 81, 97, 113, 2, 18, 34, 50,
66, 82, 98, 114, 3, 19, 35, 51, 67, 83, 99, 115, 4, 20, 36, 52, 68, 84, 100, 116,
5, 21, 37, 53, 69, 85, 101, 117, 6, 22, 38, 54, 70, 86, 102, 118, 7, 23, 39, 55,
71, 87, 103, 119, 8, 24, 40, 56, 72, 88, 104, 120, 9, 25, 41, 57, 73, 89, 105, 121,
10, 26, 42, 58, 74, 90, 106, 122, 11, 27, 43, 59, 75, 91, 107, 123, 12, 28, 44, 60,
76, 92, 108, 124, 13, 29, 45, 61, 77, 93, 109, 125, 14, 30, 46, 62, 78, 94, 110, 126,
15, 31, 47, 63, 79, 95, 111};
test_remap_xcd<Config::GemmShape>(initial_values, expected_values, tile_partitioner);
}

TEST(StreamKTilePartitionerBaseGetTileLocalCtaIndex, SKOnlyLargeK)
{
/*
Expand Down
17 changes: 17 additions & 0 deletions test/ck_tile/gemm_streamk/test_streamk_tile_partitioner_common.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -421,6 +421,23 @@ void test_get_tile_local_cta_idx(ck_tile::index_t tile_iter_start,
EXPECT_EQ(tile_local_cta_idx, expected_tile_local_cta_idx);
}

template <typename GemmShape>
Copy link
Collaborator

@ecamartins ecamartins Jan 27, 2026

Choose a reason for hiding this comment

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

Can we also add a test for when we don't use the default number of XCDs?

void test_remap_xcd(
const std::vector<ck_tile::index_t>& initial_values,
const std::vector<ck_tile::index_t>& expected_values,
ck_tile::StreamKTilePartitioner_v2<GemmShape, ck_tile::StreamKReductionStrategy::Atomic, true>&
tile_partitioner)
{
// ck_tile::index_t grid_size = tile_partitioner.grid_size().x;
std::vector<ck_tile::index_t> remapped_values(initial_values.size());
for(std::size_t i = 0; i < initial_values.size(); ++i)
{
remapped_values[i] = tile_partitioner.RemapXCD(initial_values[i], initial_values.size());
}

EXPECT_EQ(remapped_values, expected_values);
}

// Configs for TilePartitioner Child structs
struct StreamKTilePartitionerV2PersistentExpected
{
Expand Down