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

[rocm-libraries] ROCm/rocm-libraries#5516 (commit ff3afda)

Browse Source 
[CK_TILE, CK_BUILDER] Add bwd data to CK Tile profiler
 (#5516)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

## Motivation

We want close the performance gap between old CK and CK Tile for bwd
data convolutions. To achieve this, we need tow things

- Configurations for the old CK kernel instances such that we can map
them into CK Tile instances.
- Support in CK profiler to run the CK Tile instance with the same API
as for old CK instances.

## Technical Details

Extracted kernel configurations from old CK. The codegen python script
for CK Tile convs is extended to support also bwd data. The generated
instances are added to the CMake build (target
`device_grouped_conv_bwd_data_tile_instances`).
A new profiler op (`grouped_conv_bwd_data_tile`) has been added to the
CK Profiler. The API is same as for old CK's profiler op
`grouped_conv_bwd_data`.
This commit is contained in:
Ville Pietilä
2026-03-25 14:36:11 +00:00
committed by assistant-librarian[bot]
parent 1834e318da
commit ec2dbfbfde
29 changed files with 1588 additions and 956 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
include/ck_tile/ops/grouped_convolution/utils/transform_conv_bwd_data_to_gemm.hpp
View File

@@ -634,22 +634,40 @@ struct TransformConvBwdDataToGemm
constexpr auto CStride = I1;
// TODO Add support for NumGroupsToMerge > 1
return make_naive_tensor_descriptor(
make_tuple(N_, Di_, Hi_, Wi_, C_),
make_tuple(NStride, DiStride, HiStride, WiStride, CStride),
number<VectorSizeC>{},
I1);
if constexpr(ConvSpec == ConvolutionSpecialization::Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor(make_tuple(N_ * Di_ * Hi_ * Wi_, C_),
make_tuple(WiStride, CStride),
number<VectorSizeC>{},
I1);
}
else
{
return make_naive_tensor_descriptor(
make_tuple(N_, Di_, Hi_, Wi_, C_),
make_tuple(NStride, DiStride, HiStride, WiStride, CStride),
number<VectorSizeC>{},
I1);
}
}
template <index_t NDim = NDimSpatial, typename std::enable_if<NDim == 3, bool>::type = false>
CK_TILE_HOST auto make_wei_grid_desc() const
{
// GKZYXC
return make_naive_tensor_descriptor(
make_tuple(K_, Z_, Y_, X_, C_),
make_tuple(C_ * X_ * Y_ * Z_, C_ * X_ * Y_, C_ * X_, C_, I1),
number<VectorSizeB>{},
I1);
if constexpr(ConvSpec == ConvolutionSpecialization::Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor(
make_tuple(K_, C_), make_tuple(C_, I1), number<VectorSizeB>{}, I1);
}
else
{
return make_naive_tensor_descriptor(
make_tuple(K_, Z_, Y_, X_, C_),
make_tuple(C_ * X_ * Y_ * Z_, C_ * X_ * Y_, C_ * X_, C_, I1),
number<VectorSizeB>{},
I1);
}
}
// TODO: implement ck_tile::tensor_layout::convolution that describe packed/strided dimemsion as
// properties