ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-24 23:05:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Grouped Convolution Backward Data Direct Load (#6624)

Browse Source 
## Proposed changes

Add Grouped Convolution Backward Data with Direct Load into
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffleV3 device implementation.
This enables direct global memory loading (bypassing LDS) for the
backward data convolution path on gfx950, following the same pattern
used in both backward weight and forward convolution.

Direct load convolution backward data improves performance by avoiding
LDS round-trips for certain configurations on gfx950, which supports a
wider range of instructions. Currently correctness is checked only at
usage point, but should be extended to a standalone UT in the future.
This commit is contained in:
jakpiase
2026-04-23 11:16:55 +02:00
committed by GitHub
parent eaeba5266b
commit 7f14d346f1
12 changed files with 1676 additions and 136 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
View File

@@ -32,6 +32,8 @@ add_instance_library(
xdl/device_grouped_conv2d_bwd_data_xdl_ngchw_gkcyx_ngkhw_f16_vec_transpose_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_ngchw_gkcyx_ngkhw_bf16_vec_transpose_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_ngchw_gkcyx_ngkhw_f32_vec_transpose_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_v3_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/device_grouped_conv2d_bwd_data_xdl_v3_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp

49
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_v3_nhwgc_gkyxc_nhwgk_bf16_instance.cpp Normal file
View File

@@ -0,0 +1,49 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_v3_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv2d_bwd_data_xdl_v3_nhwgk_gkyxc_nhwgc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_v3_bf16_instances<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_v3_bf16_instances<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

49
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_v3_nhwgc_gkyxc_nhwgk_f16_instance.cpp Normal file
View File

@@ -0,0 +1,49 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_v3_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv2d_bwd_data_xdl_v3_nhwgk_gkyxc_nhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_v3_f16_instances<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_v3_f16_instances<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck