Skip to content

[XPU] support fp16 for group_norm_grad #71569

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 2 commits into from
Mar 13, 2025
Merged

[XPU] support fp16 for group_norm_grad #71569

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a5ed9f6
[XPU] support fp16 for group_norm_grad
lj970926 Mar 11, 2025
1803fdc
refine code
lj970926 Mar 11, 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
3 changes: 2 additions & 1 deletion paddle/phi/backends/xpu/xpu3_op_list.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1668,7 +1668,8 @@ XPUOpMap& get_kl3_ops() {
XPUKernelSet({phi::DataType::FLOAT32,
phi::DataType::FLOAT16,
phi::DataType::BFLOAT16})},
{"group_norm_grad", XPUKernelSet({phi::DataType::FLOAT32})},
{"group_norm_grad",
XPUKernelSet({phi::DataType::FLOAT32, phi::DataType::FLOAT16})},
{"meshgrid",
XPUKernelSet({phi::DataType::FLOAT32,
phi::DataType::INT32,
Expand Down
117 changes: 89 additions & 28 deletions paddle/phi/kernels/xpu/group_norm_grad_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,8 @@ void GroupNormGradKernel(const Context& dev_ctx,
DenseTensor* d_scale,
DenseTensor* d_bias) {
using XPUType = typename XPUTypeTrait<T>::Type;
xpu::ctx_guard RAII_GUARD(dev_ctx.x_context());
int ret = xpu::SUCCESS;
const DataLayout data_layout = common::StringToDataLayout(data_layout_str);
const auto scale_ptr = scale.get_ptr();
const auto bias_ptr = bias.get_ptr();
Expand All @@ -66,49 +68,108 @@ void GroupNormGradKernel(const Context& dev_ctx,
auto* y_data = y.data<T>();
auto* d_x_data = d_x->data<T>();
auto* d_y_data = d_y.data<T>();
auto* mean_data = mean.data<T>();
auto* var_data = var.data<T>();
T* d_scale_data = nullptr;
float* d_scale_data_fp32 = 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>();
if (!std::is_same_v<XPUType, float>) {
d_scale_data_fp32 = RAII_GUARD.alloc_l3_or_gm<float>(d_scale->numel());
} else {
d_scale_data_fp32 = reinterpret_cast<float*>(d_scale_data);
}
}
T* d_bias_data = nullptr;
float* d_bias_data_fp32 = nullptr;
if (d_bias) {
dev_ctx.template Alloc<T>(d_bias);
set_zero(dev_ctx, d_bias, static_cast<T>(0));
d_bias_data = d_bias->data<T>();
if (!std::is_same_v<XPUType, float>) {
d_bias_data_fp32 = RAII_GUARD.alloc_l3_or_gm<float>(d_bias->numel());
} else {
d_bias_data_fp32 = reinterpret_cast<float*>(d_bias_data);
}
}

const T* scale_data = nullptr;
if (scale_ptr) scale_data = scale_ptr->data<T>();
const T* bias_data = nullptr;
if (bias_ptr) bias_data = bias_ptr->data<T>();
const float* scale_data = nullptr;
if (scale_ptr) {
if (!std::is_same_v<XPUType, float>) {
float* scale_data_tmp =
RAII_GUARD.alloc_l3_or_gm<float>(scale_ptr->numel());
ret = xpu::cast<XPUType, float>(
dev_ctx.x_context(),
reinterpret_cast<const XPUType*>(scale_ptr->data<T>()),
scale_data_tmp,
scale_ptr->numel());
PADDLE_ENFORCE_XDNN_SUCCESS(ret, "cast");
scale_data = scale_data_tmp;
} else {
scale_data = scale_ptr->data<float>();
}
}
const float* bias_data = nullptr;
if (bias_ptr) {
if (!std::is_same_v<XPUType, float>) {
float* bias_data_tmp =
RAII_GUARD.alloc_l3_or_gm<float>(bias_ptr->numel());
ret = xpu::cast<XPUType, float>(
dev_ctx.x_context(),
reinterpret_cast<const XPUType*>(bias_ptr->data<T>()),
bias_data_tmp,
bias_ptr->numel());
PADDLE_ENFORCE_XDNN_SUCCESS(ret, "cast");
bias_data = bias_data_tmp;
} else {
bias_data = bias_ptr->data<float>();
}
}

int r = xpu::group_norm_grad<XPUType>(
dev_ctx.x_context(),
reinterpret_cast<const XPUType*>(x_data),
reinterpret_cast<const XPUType*>(y_data),
reinterpret_cast<const XPUType*>(d_y_data),
reinterpret_cast<XPUType*>(d_x_data),
N,
C,
L,
1,
groups,
static_cast<XPUType>(epsilon),
reinterpret_cast<const XPUType*>(scale_data),
reinterpret_cast<const XPUType*>(bias_data),
reinterpret_cast<const XPUType*>(mean_data),
reinterpret_cast<const XPUType*>(var_data),
reinterpret_cast<XPUType*>(d_scale_data),
reinterpret_cast<XPUType*>(d_bias_data),
channel_first);
PADDLE_ENFORCE_XDNN_SUCCESS(r, "group_norm_grad");
ret =
xpu::group_norm_grad<XPUType>(dev_ctx.x_context(),
reinterpret_cast<const XPUType*>(x_data),
reinterpret_cast<const XPUType*>(y_data),
reinterpret_cast<const XPUType*>(d_y_data),
reinterpret_cast<XPUType*>(d_x_data),
N,
C,
L,
1,
groups,
epsilon,
scale_data,
bias_data,
mean.data<float>(),
var.data<float>(),
d_scale_data_fp32,
d_bias_data_fp32,
channel_first);
PADDLE_ENFORCE_XDNN_SUCCESS(ret, "group_norm_grad");
if (!std::is_same_v<XPUType, float>) {
if (d_scale) {
ret = xpu::cast<float, XPUType>(dev_ctx.x_context(),
d_scale_data_fp32,
reinterpret_cast<XPUType*>(d_scale_data),
d_scale->numel());
PADDLE_ENFORCE_XDNN_SUCCESS(ret, "cast");
}

if (d_bias) {
ret = xpu::cast<float, XPUType>(dev_ctx.x_context(),
d_bias_data_fp32,
reinterpret_cast<XPUType*>(d_bias_data),
d_bias->numel());
PADDLE_ENFORCE_XDNN_SUCCESS(ret, "cast");
}
}
}

} // namespace phi

PD_REGISTER_KERNEL(
group_norm_grad, XPU, ALL_LAYOUT, phi::GroupNormGradKernel, float) {}
PD_REGISTER_KERNEL(group_norm_grad,
XPU,
ALL_LAYOUT,
phi::GroupNormGradKernel,
float,
phi::dtype::float16) {}
84 changes: 81 additions & 3 deletions test/xpu/test_group_norm_op_xpu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,6 @@
from get_test_cover_info import (
XPUOpTestWrapper,
create_test_class,
get_xpu_op_support_types,
)
from op_test import OpTest
from op_test_xpu import XPUOpTest
Expand Down Expand Up @@ -99,10 +98,89 @@ def init_test_case(self):
self.attrs = {'epsilon': 1e-5, 'groups': 2, 'data_layout': "NHWC"}


support_types = get_xpu_op_support_types('group_norm')
for stype in support_types:
for stype in ["float32"]:
create_test_class(globals(), XPUTestGroupNormOp, stype)


class TestGroupNormFP16(unittest.TestCase):
def setUp(self):
self.shape = [2, 100, 3, 5]
self.data_format = "NCHW"
self.epsilon = 1e-5
self.groups = 2

def test_dygraph(self):
paddle.disable_static()
inp = np.random.random(self.shape).astype("float16")
if self.data_format == "NHWC":
inp = np.transpose(inp, (0, 2, 3, 1))
scale = np.random.random([self.shape[1]]).astype("float16")
bias = np.random.random([self.shape[1]]).astype("float16")
inp_fp16 = paddle.to_tensor(inp, stop_gradient=False)
scale_fp16 = paddle.to_tensor(scale, stop_gradient=False)
bias_fp16 = paddle.to_tensor(bias, stop_gradient=False)

inp_fp32 = paddle.to_tensor(inp.astype("float32"), stop_gradient=False)
scale_fp32 = paddle.to_tensor(
scale.astype("float32"), stop_gradient=False
)
bias_fp32 = paddle.to_tensor(
bias.astype("float32"), stop_gradient=False
)

out_fp32 = paddle.nn.functional.group_norm(
inp_fp32,
self.groups,
self.epsilon,
scale_fp32,
bias_fp32,
self.data_format,
)
out_fp32.mean().backward()
inp_grad_fp32 = inp_fp32.grad.numpy()
scale_grad_fp32 = scale_fp32.grad.numpy()
bias_grad_fp32 = bias_fp32.grad.numpy()

out_fp16 = paddle.nn.functional.group_norm(
inp_fp16,
self.groups,
self.epsilon,
scale_fp16,
bias_fp16,
self.data_format,
)
out_fp16.mean().backward()
inp_grad_fp16 = inp_fp16.grad.numpy()
scale_grad_fp16 = scale_fp16.grad.numpy()
bias_grad_fp16 = bias_fp16.grad.numpy()

np.testing.assert_allclose(
out_fp32.numpy(),
out_fp16.numpy().astype("float32"),
atol=0.001,
rtol=0.001,
)
np.testing.assert_allclose(
inp_grad_fp32,
inp_grad_fp16.astype("float32"),
atol=0.001,
rtol=0.001,
)
np.testing.assert_allclose(
scale_grad_fp32,
scale_grad_fp16.astype("float32"),
atol=1e-4,
rtol=1e-4,
)
np.testing.assert_allclose(
bias_grad_fp32,
bias_grad_fp16.astype("float32"),
atol=1e-4,
rtol=1e-4,
)
paddle.enable_static()


if __name__ == "__main__":
paddle.enable_static()
unittest.main()