ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 21:21:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add column to image kernel (#930)

Browse Source 
* Add column to image kernel

* Minor fixes for dtypes and client examples

* Disable tests for disabled dtypes

* Disable add instances functions for disabled data types

* Minor stylistic fixes

* Revert "Disable add instances functions for disabled data types"

This reverts commit 728b869563.

* Instances reduction

* Add comments in device_column_to_image_impl

* Update changelog and Copyrights

* Improve changelog
This commit is contained in:
Bartłomiej Kocot
2023-09-27 17:19:06 +02:00
committed by GitHub
parent 11676c7e49
commit e2243a4d1e
41 changed files with 3040 additions and 1201 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
include/ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp → include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp
View File

@@ -16,6 +16,36 @@
namespace ck {
template <typename InputGridDesc,
typename InputDataType,
typename OutputGridDesc,
typename OutputDataType,
typename Block2ETileMap,
typename GridwiseTensorRearrangeKernel>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_tensor_rearrange(const InputGridDesc in_grid_desc,
const InputDataType* __restrict__ p_in_global,
const OutputGridDesc out_grid_desc,
OutputDataType* __restrict__ p_out_global,
const Block2ETileMap block_2_tile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
GridwiseTensorRearrangeKernel::Run(
in_grid_desc, p_in_global, out_grid_desc, p_out_global, block_2_tile_map);
#else
ignore = in_grid_desc;
ignore = p_in_global;
ignore = out_grid_desc;
ignore = p_out_global;
ignore = block_2_tile_map;
#endif
}
template <typename InputGridDesc,
typename InputDataType,
typename OutputGridDesc,
@@ -25,8 +55,9 @@ template <typename InputGridDesc,
index_t KPerBlock,
typename ThreadClusterLengths,
index_t ScalarPerVector,
InMemoryDataOperationEnum DstInMemOp,
typename Block2ETileMap>
struct GridwiseImageToColumn
struct GridwiseTensorRearrange
{
static constexpr auto I0 = Number<0>{};
@@ -55,27 +86,27 @@ struct GridwiseImageToColumn
auto out_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_out_global, out_grid_desc.GetElementSpaceSize());
auto copy_global_to_global = ThreadGroupTensorSliceTransfer_v7<
ThisThreadBlock,
Tuple<InputDataType>,
Tuple<OutputDataType>,
decltype(tie(in_grid_desc)),
decltype(tie(out_grid_desc)),
tensor_operation::element_wise::PassThrough,
Sequence<static_cast<index_t>(InMemoryDataOperationEnum::Set)>,
Sequence<MPerBlock, KPerBlock>,
ThreadClusterLengths,
Sequence<0, 1>,
Sequence<0, 1>,
I1,
ScalarPerVector,
Sequence<true>,
Sequence<true>>{
in_grid_desc,
make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
out_grid_desc,
make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
tensor_operation::element_wise::PassThrough{}};
auto copy_global_to_global =
ThreadGroupTensorSliceTransfer_v7<ThisThreadBlock,
Tuple<InputDataType>,
Tuple<OutputDataType>,
decltype(tie(in_grid_desc)),
decltype(tie(out_grid_desc)),
tensor_operation::element_wise::PassThrough,
Sequence<static_cast<index_t>(DstInMemOp)>,
Sequence<MPerBlock, KPerBlock>,
ThreadClusterLengths,
Sequence<0, 1>,
Sequence<0, 1>,
I1,
ScalarPerVector,
Sequence<true>,
Sequence<true>>{
in_grid_desc,
make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
out_grid_desc,
make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
tensor_operation::element_wise::PassThrough{}};
copy_global_to_global.Run(
tie(in_grid_desc), tie(in_global_buf), tie(out_grid_desc), tie(out_global_buf));