Skip to content

Group_norm backward kernel optimization part 2 #1719

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 8 commits into
base: main
Choose a base branch
from
Open

Group_norm backward kernel optimization part 2 #1719

Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
a47738f
add dummy gamma
yucai-intel Jun 4, 2025
ff893a0
revise
yucai-intel Jun 4, 2025
21e37fd
Update GroupNormKernels.cpp
yucai-intel Jun 9, 2025
2b5b486
Merge branch 'main' into yucai/gn_dummy
yucai-intel Jun 9, 2025
dfd8aa9
fix err
yucai-intel Jun 10, 2025
1ca5264
Update GroupNormKernels.cpp
yucai-intel Jun 10, 2025
5b5be4d
Update GroupNormKernels.cpp
yucai-intel Jun 11, 2025
7035673
Merge branch 'main' into yucai/gn_dummy
yucai-intel Jun 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 21 additions & 26 deletions src/ATen/native/xpu/sycl/GroupNormKernels.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@
#include <ATen/native/xpu/sycl/GroupReduceUtils.h>
#include <ATen/native/xpu/sycl/Loops.h>
#include <ATen/native/xpu/sycl/SharedReduceOps.h>
#include <ATen/ops/ones.h>
#include <comm/MemoryFormat.h>
#include <comm/XPUMathCompat.h>
#include <comm/xpu_aten.h>
Expand Down Expand Up @@ -1408,7 +1409,7 @@ void group_norm_backward_kernel_impl(
Tensor c3 = at::empty({N, G}, X.options().dtype(kAccType));
T_ACC* c2_data = c2.mutable_data_ptr<T_ACC>();
T_ACC* c3_data = c3.mutable_data_ptr<T_ACC>();

Tensor dummy_gamma = at::ones({1, G, D}, X.options().dtype(kAccType));
if (gamma.defined()) {
auto iter = TensorIteratorConfig()
.check_all_same_dtype(std::is_same<T, T_ACC>::value)
Expand All @@ -1417,6 +1418,14 @@ void group_norm_backward_kernel_impl(
.add_owned_const_input(gamma.view({1, G, D}))
.build();
gpu_kernel(iter, GroupNormBackwardC1Functor<T, T_ACC>());
} else {
auto iter = TensorIteratorConfig()
.check_all_same_dtype(std::is_same<T, T_ACC>::value)
.add_output(c1)
.add_owned_const_input(rstd.view({N, G, 1}))
.add_owned_const_input(dummy_gamma.view({1, G, D}))
.build();
gpu_kernel(iter, GroupNormBackwardC1Functor<T, T_ACC>());
}

wg_size = (C / G) < get_group_reduce_group_size(simd)
Expand All @@ -1440,31 +1449,17 @@ void group_norm_backward_kernel_impl(
c2_data,
c3_data);

if (gamma.defined()) {
auto iter = TensorIteratorConfig()
.check_all_same_dtype(std::is_same<T, T_ACC>::value)
.resize_outputs(false)
.add_owned_output(dX.view({N * G, D, HxW}))
.add_owned_const_input(dY.view({N * G, D, HxW}))
.add_owned_const_input(X.view({N * G, D, HxW}))
.add_owned_const_input(c1.view({N * G, D, 1}))
.add_owned_const_input(c2.view({N * G, 1, 1}))
.add_owned_const_input(c3.view({N * G, 1, 1}))
.build();
gpu_kernel(iter, GroupNormBackwardDXFunctor<T, T_ACC>());
} else {
auto iter = TensorIteratorConfig()
.check_all_same_dtype(std::is_same<T, T_ACC>::value)
.resize_outputs(false)
.add_owned_output(dX.view({N * G, D * HxW}))
.add_owned_const_input(dY.view({N * G, D * HxW}))
.add_owned_const_input(X.view({N * G, D * HxW}))
.add_owned_const_input(rstd.view({N * G, 1}))
.add_owned_const_input(c2.view({N * G, 1}))
.add_owned_const_input(c3.view({N * G, 1}))
.build();
gpu_kernel(iter, GroupNormBackwardDXFunctor<T, T_ACC>());
}
auto iter = TensorIteratorConfig()
.check_all_same_dtype(std::is_same<T, T_ACC>::value)
.resize_outputs(false)
.add_owned_output(dX.view({N * G, D, HxW}))
.add_owned_const_input(dY.view({N * G, D, HxW}))
.add_owned_const_input(X.view({N * G, D, HxW}))
.add_owned_const_input(c1.view({N * G, D, 1}))
.add_owned_const_input(c2.view({N * G, 1, 1}))
.add_owned_const_input(c3.view({N * G, 1, 1}))
.build();
gpu_kernel(iter, GroupNormBackwardDXFunctor<T, T_ACC>());
}

if (dgamma.defined() || dbeta.defined()) {
Expand Down