ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Vectorized Transpose for Batched Transpose CK Tile Operator (#2131)

Browse Source 
* Shared Memory for single data point

* CKTile Transpose vectorize CP1

* CKTile Transpose vectorize CP2

* CKTile Transpose vectorize CP2.1

* fixed the compile error of the transpose tile 2d

* Have the correct result for the current test sample

* Changes to printing tensor

* fp8 support added

* Debugging for transpose

* solving the corner issue

* Changed padding flag

* Intermideate Debugging

* Intermidiate Debugging

* Intermediate Debugging

* Finished debugging of the transpose op

* Code Cleanup

* Adding edge case smoke tests

* Adding Transpose test to CI/CD

* Adding Transpose test to CI/CD

* Adding Transpose test to CI/CD

* Addressing Review Comment

* Addressing Comments

* Addressing Comments

* Measuring Perf Tests

* Code Cleanup

* Changlog

* Added the running iterations

* clang format

* Fix the changelog

* Fix the compilation error

* change the printing factor

---------

Co-authored-by: ThruptiRajLakshmanaGowda <tlakshma@amd.com>
This commit is contained in:
Thomas Ning
2025-05-12 00:41:45 -07:00
committed by GitHub
parent d8faf1c6a1
commit 9d1e44e56a
14 changed files with 311 additions and 152 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
include/ck_tile/ops/batched_transpose/kernel/batched_transpose_kernel.hpp
View File

@@ -19,7 +19,6 @@ struct BatchedTransposeHostArgs
index_t batch;
index_t height;
index_t width;
// index_t dim_blocks;
index_t dim_stride;
index_t dim_block_h;
index_t dim_block_w;
@@ -28,8 +27,10 @@ struct BatchedTransposeHostArgs
template <typename Pipeline_>
struct BatchedTransposeKernel
{
using Pipeline = remove_cvref_t<Pipeline_>;
using Problem = remove_cvref_t<typename Pipeline::Problem>;
CK_TILE_DEVICE static index_t counter = 0;
using Pipeline = remove_cvref_t<Pipeline_>;
using Problem = remove_cvref_t<typename Pipeline::Problem>;
using Type = typename Problem::InputType;
@@ -46,11 +47,11 @@ struct BatchedTransposeKernel
using Kargs = BatchedTransposeKargs;
using Hargs = BatchedTransposeHostArgs;
CK_TILE_HOST static constexpr auto GridSize(const Hargs& h)
CK_TILE_HOST static constexpr auto GridSize(const Hargs& host_args)
{
size_t grid_size_x = (h.width + h.dim_block_w - 1) / h.dim_block_w;
size_t grid_size_y = (h.height + h.dim_block_h - 1) / h.dim_block_h;
size_t grid_size_z = h.batch;
size_t grid_size_x = (host_args.height + host_args.dim_block_h - 1) / host_args.dim_block_h;
size_t grid_size_y = (host_args.width + host_args.dim_block_w - 1) / host_args.dim_block_w;
size_t grid_size_z = host_args.batch;
return dim3(grid_size_x, grid_size_y, grid_size_z);
}
@@ -70,58 +71,52 @@ struct BatchedTransposeKernel
CK_TILE_DEVICE void operator()(Kargs kargs) const
{
static constexpr ck_tile::index_t kMPerBlock = Problem::kMPerBlock;
static constexpr ck_tile::index_t kNPerBlock = Problem::kNPerBlock;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr ck_tile::index_t VectorSizeInput = Problem::VectorSizeInput;
static constexpr ck_tile::index_t VectorSizeOutput = Problem::VectorSizeOutput;
static constexpr ck_tile::index_t kMPerBlock = Problem::kMPerBlock;
static constexpr ck_tile::index_t kNPerBlock = Problem::kNPerBlock;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
const auto iM = __builtin_amdgcn_readfirstlane(blockIdx.x * kMPerBlock);
const auto iN = __builtin_amdgcn_readfirstlane(blockIdx.y * kNPerBlock);
const auto iDim = blockIdx.z;
static constexpr ck_tile::index_t kMPerThread = Problem::kMPerThread;
static constexpr ck_tile::index_t kNPerThread = Problem::kNPerThread;
static_assert(kMPerThread == 1 && kNPerThread == 1);
const auto iDim = blockIdx.z;
const auto x_m_n = [&]() {
const auto x_dram_naive = make_naive_tensor_view<address_space_enum::global>(
static_cast<const Type*>(kargs.p_input) + iDim * kargs.dim_stride,
make_tuple(kargs.height, kargs.width),
make_tuple(kargs.width, 1),
number<kNPerThread>{}, // TODO thread load value
number<VectorSizeInput>{},
number<1>{});
return pad_tensor_view(x_dram_naive,
make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}),
sequence<kPadM, kPadN>{});
sequence<kPadN, kPadM>{});
}();
const auto iM = __builtin_amdgcn_readfirstlane(blockIdx.x * kMPerBlock);
const auto iN = __builtin_amdgcn_readfirstlane(blockIdx.y * kNPerBlock);
const auto y_n_m = [&]() {
const auto y_dram_naive = make_naive_tensor_view<address_space_enum::global>(
static_cast<Type*>(kargs.p_output) + iDim * kargs.dim_stride,
make_tuple(kargs.width, kargs.height),
make_tuple(kargs.height, 1),
number<kMPerThread>{},
number<VectorSizeOutput>{},
number<1>{});
return pad_tensor_view(y_dram_naive,
make_tuple(number<kNPerBlock>{}, number<kMPerBlock>{}),
sequence<kPadN, kPadM>{});
sequence<kPadM, kPadN>{});
}();
auto x_block_window =
make_tile_window(x_m_n,
make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}),
{static_cast<ck_tile::index_t>(iM * kMPerBlock),
static_cast<ck_tile::index_t>(iN * kNPerBlock)});
auto x_block_window = make_tile_window(
x_m_n,
make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}),
{static_cast<ck_tile::index_t>(iM), static_cast<ck_tile::index_t>(iN)});
auto y_block_window =
make_tile_window(y_n_m,
make_tuple(number<kNPerBlock>{}, number<kMPerBlock>{}),
{static_cast<ck_tile::index_t>(iN * kNPerBlock),
static_cast<ck_tile::index_t>(iM * kMPerBlock)});
auto y_block_window = make_tile_window(
y_n_m,
make_tuple(number<kNPerBlock>{}, number<kMPerBlock>{}),
{static_cast<ck_tile::index_t>(iN), static_cast<ck_tile::index_t>(iM)});
Pipeline{}(x_block_window, y_block_window);
}