Skip to content

improve group norm cpu precision and performance #33176

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
May 31, 2021
Merged

improve group norm cpu precision and performance #33176

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
262596a
improve group norm cpu precision and performance
jeff41404 May 27, 2021
987dd89
add unit test to group norm
jeff41404 May 28, 2021
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
75 changes: 71 additions & 4 deletions paddle/fluid/operators/group_norm_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,8 @@ limitations under the License. */

#pragma once
#include <algorithm>
#include <array>
#include <numeric>
#include <string>
#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/framework/eigen.h"
Expand Down Expand Up @@ -73,6 +75,11 @@ class GroupNormKernel : public framework::OpKernel<T> {
auto* iter_y_data = y_data;
for (int bid = 0; bid < x_dims[0]; bid++) {
for (int gid = 0; gid < groups; gid++) {
const int64_t M = 8;
std::array<T, M> x_mean_arr;
std::array<T, M> x_var_arr;
std::fill(x_mean_arr.begin(), x_mean_arr.end(), T(0));
std::fill(x_var_arr.begin(), x_var_arr.end(), T(0));
T x_mean = 0, x_var = 0;
int number =
std::min(group_size, static_cast<int>(C - gid * group_size));
Expand All @@ -83,15 +90,75 @@ class GroupNormKernel : public framework::OpKernel<T> {

if (data_layout == DataLayout::kNCHW) {
for (int cid = 0; cid < number; cid++) {
for (int imid = 0; imid < imsize; imid++, iter_x_data++) {
int imid;
for (imid = 0; imid < imsize - (imsize % M);
imid += M, iter_x_data += M) {
// TODO(gaoxiang) ：Because AVX/AVX2/AVX512 can not directly used
// in template class/function, before we complete high
// performance cpu vector extension, temporarily unrolling
// loop to get high precision and performance
x_mean_arr[0] += iter_x_data[0];
x_var_arr[0] += iter_x_data[0] * iter_x_data[0];
x_mean_arr[1] += iter_x_data[1];
x_var_arr[1] += iter_x_data[1] * iter_x_data[1];
x_mean_arr[2] += iter_x_data[2];
x_var_arr[2] += iter_x_data[2] * iter_x_data[2];
x_mean_arr[3] += iter_x_data[3];
x_var_arr[3] += iter_x_data[3] * iter_x_data[3];
x_mean_arr[4] += iter_x_data[4];
x_var_arr[4] += iter_x_data[4] * iter_x_data[4];
x_mean_arr[5] += iter_x_data[5];
x_var_arr[5] += iter_x_data[5] * iter_x_data[5];
x_mean_arr[6] += iter_x_data[6];
x_var_arr[6] += iter_x_data[6] * iter_x_data[6];
x_mean_arr[7] += iter_x_data[7];
x_var_arr[7] += iter_x_data[7] * iter_x_data[7];
}
x_mean =
std::accumulate(x_mean_arr.cbegin(), x_mean_arr.cend(), x_mean);
x_var =
std::accumulate(x_var_arr.cbegin(), x_var_arr.cend(), x_var);
std::fill(x_mean_arr.begin(), x_mean_arr.end(), T(0));
std::fill(x_var_arr.begin(), x_var_arr.end(), T(0));
for (; imid < imsize; imid++, iter_x_data++) {
x_mean += iter_x_data[0];
x_var += iter_x_data[0] * iter_x_data[0];
}
}
} else {
for (int cid = 0; cid < number; cid++) {
iter_x_data = tmp_x + cid;
for (int imid = 0; imid < imsize; imid++, iter_x_data += C) {
int imid;
for (imid = 0; imid < imsize - (imsize % M);
imid += M, iter_x_data += M * C) {
// TODO(gaoxiang) ：Because AVX/AVX2/AVX512 can not directly used
// in template class/function, before we complete high
// performance cpu vector extension, temporarily unrolling
// loop to get high precision and performance
x_mean_arr[0] += iter_x_data[0 * C];
x_var_arr[0] += iter_x_data[0 * C] * iter_x_data[0 * C];
x_mean_arr[1] += iter_x_data[1 * C];
x_var_arr[1] += iter_x_data[1 * C] * iter_x_data[1 * C];
x_mean_arr[2] += iter_x_data[2 * C];
x_var_arr[2] += iter_x_data[2 * C] * iter_x_data[2 * C];
x_mean_arr[3] += iter_x_data[3 * C];
x_var_arr[3] += iter_x_data[3 * C] * iter_x_data[3 * C];
x_mean_arr[4] += iter_x_data[4 * C];
x_var_arr[4] += iter_x_data[4 * C] * iter_x_data[4 * C];
x_mean_arr[5] += iter_x_data[5 * C];
x_var_arr[5] += iter_x_data[5 * C] * iter_x_data[5 * C];
x_mean_arr[6] += iter_x_data[6 * C];
x_var_arr[6] += iter_x_data[6 * C] * iter_x_data[6 * C];
x_mean_arr[7] += iter_x_data[7 * C];
x_var_arr[7] += iter_x_data[7 * C] * iter_x_data[7 * C];
}
x_mean =
std::accumulate(x_mean_arr.cbegin(), x_mean_arr.cend(), x_mean);
x_var =
std::accumulate(x_var_arr.cbegin(), x_var_arr.cend(), x_var);
std::fill(x_mean_arr.begin(), x_mean_arr.end(), T(0));
std::fill(x_var_arr.begin(), x_var_arr.end(), T(0));
for (; imid < imsize; imid++, iter_x_data += C) {
x_mean += iter_x_data[0];
x_var += iter_x_data[0] * iter_x_data[0];
}
Expand All @@ -101,8 +168,8 @@ class GroupNormKernel : public framework::OpKernel<T> {

x_mean /= number * imsize;
x_var /= number * imsize;
x_var = x_var - x_mean * x_mean;
T var_inv = 1.0 / sqrt(x_var + epsilon);
x_var = std::max(x_var - x_mean * x_mean, T(0));
T var_inv = T(1) / std::sqrt(x_var + epsilon);
mean_data[bid * groups + gid] = x_mean;
var_data[bid * groups + gid] = x_var;

Expand Down
10 changes: 10 additions & 0 deletions python/paddle/fluid/tests/unittests/test_group_norm_op_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,15 @@ def test_weight_bias_false():
weight_attr=False,
bias_attr=False)

def test_nn_exception():
with fluid.dygraph.guard(p):

def attr_data_format():
out = paddle.nn.GroupNorm(
num_groups=2, num_channels=2, data_format="NHWC")

self.assertRaises(ValueError, attr_data_format)

x = np.random.randn(*shape).astype("float32")
y1 = compute_v1(x)
y2 = compute_v2(x)
Expand All @@ -61,6 +70,7 @@ def test_weight_bias_false():
print("y1:", y1, "\ty2:", y2)
self.assertTrue(result)
test_weight_bias_false()
test_nn_exception()

def test_static(self):
places = [fluid.CPUPlace()]
Expand Down
2 changes: 1 addition & 1 deletion python/paddle/nn/layer/norm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -375,7 +375,7 @@ def __init__(self,
self._num_channels = num_channels
self._num_groups = num_groups
if data_format != 'NCHW':
raise ValueError("unsupported data layout:" + data_layout)
raise ValueError("unsupported data layout:" + data_format)

param_shape = [self._num_channels]

Expand Down