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

add batched_transpose implement (#1660)

Browse Source 
* add batched_transpose implement

---------

Co-authored-by: root <root@ctr-ubbsmc16.amd.com>
Co-authored-by: ThruptiRajLakshmanaGowda <tlakshma@amd.com>
Co-authored-by: ThomasNing <thomas.ning@amd.com>
This commit is contained in:
fangche123
2025-01-29 08:22:02 +08:00
committed by GitHub
parent d6a4605e1c
commit c5fff071e5
14 changed files with 760 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

129
include/ck_tile/ops/batched_transpose/kernel/batched_transpose_kernel.hpp Normal file
View File

@@ -0,0 +1,129 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common.hpp"
#include "ck_tile/ops/elementwise.hpp"
#include "ck_tile/host/hip_check_error.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
struct BatchedTransposeHostArgs
{
const void* p_input;
void* p_output;
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;
};
template <typename Pipeline_>
struct BatchedTransposeKernel
{
using Pipeline = remove_cvref_t<Pipeline_>;
using Problem = remove_cvref_t<typename Pipeline::Problem>;
using Type = typename Problem::InputType;
struct BatchedTransposeKargs
{
const void* p_input;
void* p_output;
index_t batch;
index_t height;
index_t width;
index_t dim_stride;
};
using Kargs = BatchedTransposeKargs;
using Hargs = BatchedTransposeHostArgs;
CK_TILE_HOST static constexpr auto GridSize(const Hargs& h)
{
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;
return dim3(grid_size_x, grid_size_y, grid_size_z);
}
CK_TILE_HOST static constexpr auto MakeKargs(const Hargs& h)
{
Kargs k;
k.p_input = h.p_input;
k.p_output = h.p_output;
k.batch = h.batch;
k.height = h.height;
k.width = h.width;
k.dim_stride = h.dim_stride;
return k;
}
CK_TILE_HOST_DEVICE static constexpr auto BlockSize() { return Problem::kBlockSize; }
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 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<1>{});
return pad_tensor_view(x_dram_naive,
make_tuple(number<kMPerBlock>{}, number<kNPerBlock>{}),
sequence<kPadM, kPadN>{});
}();
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<1>{});
return pad_tensor_view(y_dram_naive,
make_tuple(number<kNPerBlock>{}, number<kMPerBlock>{}),
sequence<kPadN, kPadM>{});
}();
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 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)});
Pipeline{}(x_block_window, y_block_window);
}
};
} // namespace ck_tile

52
include/ck_tile/ops/batched_transpose/pipeline/batched_transpose_pipeline.hpp Normal file
View File

@@ -0,0 +1,52 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/batched_transpose/pipeline/batched_transpose_policy.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
template <typename Problem_, typename Policy_ = BatchedTransposePolicy>
struct BatchedTransposePipeline
{
// TODO: this kernel only support warp per row
using Problem = remove_cvref_t<Problem_>;
using Policy = remove_cvref_t<Policy_>;
using InputType = ck_tile::remove_cvref_t<typename Problem::InputType>;
static constexpr ck_tile::index_t kMPerBlock = Problem::kMPerBlock;
static constexpr ck_tile::index_t kNPerBlock = Problem::kNPerBlock;
static constexpr index_t AlignmentM = Problem::AlignmentM;
static constexpr index_t AlignmentN = Problem::AlignmentN;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
template <typename InputWindow, typename OutputWindow>
CK_TILE_DEVICE auto operator()(const InputWindow& input_window, OutputWindow& out_window)
{
auto inp_win =
make_tile_window(input_window, Policy::template MakeInputDistribution<Problem>());
auto out_win =
make_tile_window(out_window, Policy::template MakeOutputDistribution<Problem>());
auto x = load_tile(inp_win); // x->thread input_win->block
auto y = make_static_distributed_tensor<InputType>(
Policy::template MakeOutputDistribution<Problem>());
constexpr auto span_2d_x = decltype(x)::get_distributed_spans();
sweep_tile_span(span_2d_x[number<0>{}], [&](auto idx0) {
sweep_tile_span(span_2d_x[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx1, idx0);
y(i_j_idx) = x(i_j_idx);
});
});
store_tile(out_win, y);
}
};
} // namespace ck_tile

44
include/ck_tile/ops/batched_transpose/pipeline/batched_transpose_policy.hpp Normal file
View File

@@ -0,0 +1,44 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/softmax.hpp"
#include "ck_tile/ops/topk.hpp"
namespace ck_tile {
struct BatchedTransposePolicy
{
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeInputDistribution()
{
using S = Problem;
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<>,
tuple<sequence<S::kMWarpPerBlock, S::kMThreadPerWarp, S::kMPerThread>,
sequence<S::kNWarpPerBlock, S::kNThreadPerWarp, S::kNPerThread>>,
tuple<sequence<1, 2>, sequence<1, 2>>,
tuple<sequence<0, 0>, sequence<1, 1>>,
sequence<1, 2>,
sequence<2, 2>>{});
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeOutputDistribution()
{
using S = Problem;
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<>,
tuple<sequence<S::kNWarpPerBlock, S::kNThreadPerWarp, S::kNPerThread>,
sequence<S::kMWarpPerBlock, S::kMThreadPerWarp, S::kMPerThread>>,
tuple<sequence<2, 1>, sequence<2, 1>>,
tuple<sequence<0, 0>, sequence<1, 1>>,
sequence<2, 1>,
sequence<2, 2>>{});
}
};
} // namespace ck_tile

48
include/ck_tile/ops/batched_transpose/pipeline/batched_transpose_problem.hpp Normal file
View File

@@ -0,0 +1,48 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include <string>
#include <type_traits>
#define VectorLoadSize 16
namespace ck_tile {
template <typename InputType_,
typename BlockTile, // Sequence<...
typename WarpTile, // Sequence<...
typename ThreadTile, // Sequence<...
bool kPadM_ = true,
bool kPadN_ = true>
struct BatchedTransposeProblem
{
using InputType = remove_cvref_t<InputType_>;
static constexpr index_t kMPerThread = ThreadTile::at(number<0>{});
static constexpr index_t kNPerThread = ThreadTile::at(number<1>{});
static constexpr index_t kMPerWarp = WarpTile::at(number<0>{});
static constexpr index_t kNPerWarp = WarpTile::at(number<1>{});
static constexpr index_t kMThreadPerWarp = kMPerWarp / kMPerThread;
static constexpr index_t kNThreadPerWarp = kNPerWarp / kNPerThread;
static constexpr index_t kMPerBlock = BlockTile::at(number<0>{});
static constexpr index_t kNPerBlock = BlockTile::at(number<1>{});
static constexpr index_t kMWarpPerBlock = kMPerBlock / kMPerWarp;
static constexpr index_t kNWarpPerBlock = kNPerBlock / kNPerWarp;
static constexpr index_t kBlockSize =
kMThreadPerWarp * kNThreadPerWarp * kMWarpPerBlock * kNWarpPerBlock;
static constexpr bool kPadM = kPadM_;
static constexpr bool kPadN = kPadN_;
static constexpr index_t AlignmentM = kPadM ? VectorLoadSize / sizeof(InputType) : 1; // TODO
static constexpr index_t AlignmentN = kPadN ? VectorLoadSize / sizeof(InputType) : 1;
};
} // namespace ck_tile