Skip to content

[Kernel] Fix GroupNormGradKernel when d_x is nullptr #72358

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

Merged
merged 8 commits into from
Apr 25, 2025
Merged

[Kernel] Fix GroupNormGradKernel when d_x is nullptr #72358

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
62e774d
fix GroupNormGradKernel when d_x is nullptr
HydrogenSulfate Apr 18, 2025
71ef877
update UT
HydrogenSulfate Apr 21, 2025
7a4e38f
fix relative import
HydrogenSulfate Apr 22, 2025
7daa7c7
fix coverage
HydrogenSulfate Apr 23, 2025
32252c0
fix rtol and atol
HydrogenSulfate Apr 23, 2025
b10975b
fix for coverage
HydrogenSulfate Apr 24, 2025
670b3ac
update UT
HydrogenSulfate Apr 24, 2025
25d39a7
add nhwc for coverage
HydrogenSulfate Apr 25, 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
90 changes: 51 additions & 39 deletions paddle/phi/kernels/cpu/group_norm_grad_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,19 +52,25 @@ void GroupNormGradKernel(const Context& dev_ctx,
data_layout == DataLayout::kNCHW ? x_dims[1] : x_dims[x_dims.size() - 1]);
const int group_size = C / groups;

dev_ctx.template Alloc<T>(d_x);
phi::funcs::SetConstant<CPUContext, T> set_zero;

auto* x_data = y.data<T>();
auto* d_x_data = d_x->data<T>();
auto* y_data = d_y.data<T>();
auto* var_data = var.data<T>();

T* d_x_data = nullptr;
if (d_x) {
dev_ctx.template Alloc<T>(d_x);
d_x_data = d_x->data<T>();
}

T* d_scale_data = nullptr;
if (d_scale) {
dev_ctx.template Alloc<T>(d_scale);
set_zero(dev_ctx, d_scale, static_cast<T>(0));
d_scale_data = d_scale->data<T>();
}

T* d_bias_data = nullptr;
if (d_bias) {
dev_ctx.template Alloc<T>(d_bias);
Expand Down Expand Up @@ -124,22 +130,23 @@ void GroupNormGradKernel(const Context& dev_ctx,
d_scale_data[gid * group_size + cid] += val * dval;
}
}

for (int cid = 0; cid < number; cid++) {
for (int imid = 0; imid < imsize;
imid++, iter_d_x_data++, tmp_x++, tmp_y++) {
T v_y = tmp_x[0];
T dly = tmp_y[0];
T dss = dp_scale;
T dbs = dp_bias;
T v_scale = 1., v_bias = 0.;
if (scale_data) v_scale = scale_data[gid * group_size + cid];
if (bias_data) v_bias = bias_data[gid * group_size + cid];
v_y -= v_bias;
if (v_scale != 0) v_y /= v_scale;
iter_d_x_data[0] =
(dly * v_scale - number_inv * dss * v_y - number_inv * dbs) *
var_inv;
if (d_x_data) {
for (int cid = 0; cid < number; cid++) {
for (int imid = 0; imid < imsize;
imid++, iter_d_x_data++, tmp_x++, tmp_y++) {
T v_y = tmp_x[0];
T dly = tmp_y[0];
T dss = dp_scale;
T dbs = dp_bias;
T v_scale = 1., v_bias = 0.;
if (scale_data) v_scale = scale_data[gid * group_size + cid];
if (bias_data) v_bias = bias_data[gid * group_size + cid];
v_y -= v_bias;
if (v_scale != 0) v_y /= v_scale;
iter_d_x_data[0] =
(dly * v_scale - number_inv * dss * v_y - number_inv * dbs) *
var_inv;
}
}
}
} else {
Expand All @@ -162,35 +169,40 @@ void GroupNormGradKernel(const Context& dev_ctx,
d_scale_data[gid * group_size + cid] += val * dval;
}
}

for (int cid = 0; cid < number; cid++) {
tmp_x = x_src_data + cid;
tmp_y = y_src_data + cid;
iter_d_x_data = tmp_d_x + cid;
for (int imid = 0; imid < imsize;
imid++, iter_d_x_data += C, tmp_x += C, tmp_y += C) {
T v_y = tmp_x[0];
T dly = tmp_y[0];
T dss = dp_scale;
T dbs = dp_bias;
T v_scale = 1.0, v_bias = 0.;
if (scale_data) v_scale = scale_data[gid * group_size + cid];
if (bias_data) v_bias = bias_data[gid * group_size + cid];
v_y -= v_bias;
if (v_scale != 0) v_y /= v_scale;
iter_d_x_data[0] =
(dly * v_scale - number_inv * dss * v_y - number_inv * dbs) *
var_inv;
if (d_x_data) {
for (int cid = 0; cid < number; cid++) {
tmp_x = x_src_data + cid;
tmp_y = y_src_data + cid;
iter_d_x_data = tmp_d_x + cid;
for (int imid = 0; imid < imsize;
imid++, iter_d_x_data += C, tmp_x += C, tmp_y += C) {
T v_y = tmp_x[0];
T dly = tmp_y[0];
T dss = dp_scale;
T dbs = dp_bias;
T v_scale = 1.0, v_bias = 0.;
if (scale_data) v_scale = scale_data[gid * group_size + cid];
if (bias_data) v_bias = bias_data[gid * group_size + cid];
v_y -= v_bias;
if (v_scale != 0) v_y /= v_scale;
iter_d_x_data[0] =
(dly * v_scale - number_inv * dss * v_y - number_inv * dbs) *
var_inv;
}
}
}
iter_x_data = iter_x_data_backup + group_size;
iter_y_data = iter_y_data_backup + group_size;
iter_d_x_data = iter_d_x_data_backup + group_size;
if (d_x_data) {
iter_d_x_data = iter_d_x_data_backup + group_size;
}
}
}
if (data_layout == DataLayout::kNHWC) {
iter_x_data = x_data + (bid + 1) * C * imsize;
iter_d_x_data = d_x_data + (bid + 1) * C * imsize;
if (d_x_data) {
iter_d_x_data = d_x_data + (bid + 1) * C * imsize;
}
iter_y_data = y_data + (bid + 1) * C * imsize;
}
}
Expand Down
2 changes: 1 addition & 1 deletion test/dygraph_to_static/test_deal_inplace.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,7 +90,7 @@ def run_test(self, dygraph_fn, *inputs, static_n_times=1):
dygraph_out.numpy(),
static_out.numpy(),
rtol=1e-5,
atol=1e-8,
atol=1e-6,
err_msg=f"Run {i}-th check failed.",
)

Expand Down
51 changes: 51 additions & 0 deletions test/legacy_test/test_group_norm_op_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@
import unittest

import numpy as np
from utils import dygraph_guard

import paddle
from paddle import base
Expand Down Expand Up @@ -610,5 +611,55 @@ def test_one_dim_input_static_API():
self.assertRaises(ValueError, test_one_dim_input_static_API)


class TestGroupNormWithOptionalgradX(unittest.TestCase):
def test_group_norm_cpu_with_optional_grad(self):
with dygraph_guard():
origin_device = paddle.device.get_device()
paddle.device.set_device("cpu")
x = paddle.randn([16, 32])
x.stop_gradient = False
gpn = paddle.nn.GroupNorm(num_groups=8, num_channels=32)
y = gpn(x)
dw_ref, db_ref, dx_ref = paddle.grad(y, [gpn.weight, gpn.bias, x])
try:
dw, db, dx = (
paddle.grad(y, gpn.weight)[0],
paddle.grad(y, gpn.bias)[0],
paddle.grad(y, x)[0],
)
except Exception as e:
raise e
finally:
paddle.device.set_device(origin_device)
np.testing.assert_equal(dw.numpy(), dw_ref.numpy())
np.testing.assert_equal(db.numpy(), db_ref.numpy())
np.testing.assert_equal(dx.numpy(), dx_ref.numpy())

def test_group_norm_cpu_with_optional_grad_nhwc(self):
with dygraph_guard():
origin_device = paddle.device.get_device()
paddle.device.set_device("cpu")
x = paddle.randn([4, 32, 32, 32])
x.stop_gradient = False
gpn = paddle.nn.GroupNorm(
num_groups=8, num_channels=32, data_format="NHWC"
)
y = gpn(x)
dw_ref, db_ref, dx_ref = paddle.grad(y, [gpn.weight, gpn.bias, x])
try:
dw, db, dx = (
paddle.grad(y, gpn.weight)[0],
paddle.grad(y, gpn.bias)[0],
paddle.grad(y, x)[0],
)
except Exception as e:
raise e
finally:
paddle.device.set_device(origin_device)
np.testing.assert_equal(dw.numpy(), dw_ref.numpy())
np.testing.assert_equal(db.numpy(), db_ref.numpy())
np.testing.assert_equal(dx.numpy(), dx_ref.numpy())


if __name__ == '__main__':
unittest.main()
Loading