ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK-Tile] Universal gemm memory bound pipeline (#1558)

Browse Source 
* CK-Tile GEMM with memory bound pipeline.

* Memory bound gemm pipeline.

* Fix not closed namespace.

* Block gemm mem pipeline draft.

* Do not use ck_tile:: within ck_tile namespace.

* Refactoring & Move Layout info to pipeline problem.

* Get hot loop and TailNum information before lunching kernel.

* Fixes in pipeline.

* Add comment to load_tile_raw and change variable naming style.

* Few small changes & formatting.

* Do not use macro.

* Add gtests.

* Use AccDataType for Output of MFMA instruction.

* Formatting.

* Refactor gemm examples.

* Switch over to current block gemm.

* Use currently available pipeline policy.

* Refactoring and review comment.s

* Fixes after merge.

* Add missing include.

* Add load tile overload which accepts output tensor as parameter.

* This give 8% perf boost at the cost of using more registers.

* Rename example.

* Small changes.

* Fix compilation err and lower K.

* Support different layouts for A/B

* Fix vector size for different layouts.

* Rename Alignment into VectorSize

* Unblock tests.
This commit is contained in:
Adam Osewski
2024-10-30 10:05:15 +01:00
committed by GitHub
parent 3d60953477
commit 24d996aae1
29 changed files with 1660 additions and 561 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
include/ck_tile/ops/gemm/block/block_gemm_areg_bgmem_creg_v1.hpp
View File

@@ -32,7 +32,7 @@ struct BlockGemmARegBGmemCRegV1
BlockGemmProblem<ADataType, BDataType, CDataType, kBlockSize, BlockGemmShape>,
BlockGemmARegBGmemCRegV1DefaultPolicy>;
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetStaticLdsSize()
CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
{
return sizeof(BDataType) *
Policy::template MakeBSmemBlockDescriptor<Problem>().get_element_space_size();

30
include/ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1.hpp
View File

@@ -24,19 +24,19 @@ struct BlockGemmASmemBSmemCRegV1
static constexpr index_t kBlockSize = Problem::kBlockSize;
// C += A * B
template <typename CBlockTensor, typename ABlockWindowTmp, typename BBlockWindowTmp>
template <typename CBlockTensor, typename ABlockWindow, typename BBlockWindow>
CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
const ABlockWindowTmp& a_block_window_tmp,
const BBlockWindowTmp& b_block_window_tmp) const
const ABlockWindow& a_block_window,
const BBlockWindow& b_block_window) const
{
static_assert(std::is_same_v<ADataType, typename ABlockWindowTmp::DataType> &&
std::is_same_v<BDataType, typename BBlockWindowTmp::DataType> &&
static_assert(std::is_same_v<ADataType, typename ABlockWindow::DataType> &&
std::is_same_v<BDataType, typename BBlockWindow::DataType> &&
std::is_same_v<CDataType, typename CBlockTensor::DataType>,
"wrong!");
constexpr index_t MPerBlock = ABlockWindowTmp{}.get_window_lengths()[number<0>{}];
constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
constexpr index_t KPerBlock = ABlockWindowTmp{}.get_window_lengths()[number<1>{}];
constexpr index_t MPerBlock = ABlockWindow{}.get_window_lengths()[number<0>{}];
constexpr index_t NPerBlock = BBlockWindow{}.get_window_lengths()[number<0>{}];
constexpr index_t KPerBlock = ABlockWindow{}.get_window_lengths()[number<1>{}];
static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
KPerBlock == BlockGemmShape::kK,
@@ -62,9 +62,9 @@ struct BlockGemmASmemBSmemCRegV1
// construct A-warp-window
auto a_warp_window_tmp = make_tile_window(
a_block_window_tmp.get_bottom_tensor_view(),
a_block_window.get_bottom_tensor_view(),
make_tuple(number<WG::kM>{}, number<WG::kK>{}),
a_block_window_tmp.get_window_origin() + multi_index<2>{iMWarp * WG::kM, 0},
a_block_window.get_window_origin() + multi_index<2>{iMWarp * WG::kM, 0},
make_static_tile_distribution(typename WG::AWarpDstrEncoding{}));
#if 0 // FIXME: using array will cause register spill
@@ -97,9 +97,9 @@ struct BlockGemmASmemBSmemCRegV1
// construct B-warp-window
auto b_warp_window_tmp = make_tile_window(
b_block_window_tmp.get_bottom_tensor_view(),
b_block_window.get_bottom_tensor_view(),
make_tuple(number<WG::kN>{}, number<WG::kK>{}),
b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
b_block_window.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
#if 0 // FIXME: using array will cause register spill
@@ -200,12 +200,12 @@ struct BlockGemmASmemBSmemCRegV1
}
// C = A * B
template <typename ABlockTensorTmp, typename BBlockWindowTmp>
template <typename ABlockTensorTmp, typename BBlockWindow>
CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
const BBlockWindowTmp& b_block_window_tmp) const
const BBlockWindow& b_block_window) const
{
auto c_block_tensor = MakeCBlockTile();
operator()(c_block_tensor, a_block_tensor_tmp, b_block_window_tmp);
operator()(c_block_tensor, a_block_tensor_tmp, b_block_window);
return c_block_tensor;
}
};