Skip to content
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

[AMP OP&Test] support bf16 for batch norm #52407

Merged
merged 12 commits into from
Apr 12, 2023
Merged

[AMP OP&Test] support bf16 for batch norm #52407

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
b0a2094
[AMP OP&Test] support bf16 for batchnorm
GuoxiaWang Mar 31, 2023
5869056
Merge branch 'develop' into bf16_batchnorm
GuoxiaWang Apr 4, 2023
b594db7
codestyle
GuoxiaWang Apr 4, 2023
ea8c601
Update batch_norm_grad_kernel.cu
GuoxiaWang Apr 7, 2023
ce26f2d
Update batch_norm_kernel.cu
GuoxiaWang Apr 7, 2023
8edd991
fix codestyle
GuoxiaWang Apr 7, 2023
d1bcbff
fix
GuoxiaWang Apr 7, 2023
b0d06ee
fix
GuoxiaWang Apr 10, 2023
7b2bbbc
fix
GuoxiaWang Apr 10, 2023
41f251b
fix
GuoxiaWang Apr 10, 2023
5f8e189
fix
GuoxiaWang Apr 10, 2023
2649344
Update batch_norm_kernel.cc
GuoxiaWang Apr 11, 2023
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
18 changes: 18 additions & 0 deletions paddle/phi/kernels/batch_norm_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@

#include "paddle/phi/kernels/batch_norm_kernel.h"

#include "paddle/phi/backends/gpu/gpu_dnn.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/empty_kernel.h"

Expand Down Expand Up @@ -66,6 +67,22 @@ PD_REGISTER_KERNEL(batch_norm_infer,
float,
double) {}
#ifdef PADDLE_WITH_CUDA
#if CUDNN_VERSION_MIN(8, 1, 0)
PD_REGISTER_KERNEL(batch_norm_infer,
GPU,
ALL_LAYOUT,
phi::BatchNormInferKernel,
float,
double,
phi::dtype::bfloat16,
phi::dtype::float16) {
if (kernel_key.dtype() == phi::DataType::FLOAT16 ||
kernel_key.dtype() == phi::DataType::BFLOAT16) {
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32);
}
}
#else
PD_REGISTER_KERNEL(batch_norm_infer,
GPU,
ALL_LAYOUT,
Expand All @@ -79,6 +96,7 @@ PD_REGISTER_KERNEL(batch_norm_infer,
}
}
#endif
#endif
#ifdef PADDLE_WITH_HIP
PD_REGISTER_KERNEL(batch_norm_infer,
GPU,
Expand Down
36 changes: 35 additions & 1 deletion paddle/phi/kernels/gpu/batch_norm_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1314,14 +1314,18 @@ PD_REGISTER_KERNEL(batch_norm_grad_raw,
float,
phi::dtype::float16) {}
#else
#if CUDNN_VERSION_MIN(8, 1, 0)

PD_REGISTER_KERNEL(batch_norm_grad,
GPU,
ALL_LAYOUT,
phi::BatchNormGradKernel,
float,
double,
phi::dtype::bfloat16,
phi::dtype::float16) {
if (kernel_key.dtype() == phi::DataType::FLOAT16) {
if (kernel_key.dtype() == phi::DataType::FLOAT16 ||
kernel_key.dtype() == phi::DataType::BFLOAT16) {
kernel->OutputAt(0).SetDataType(phi::DataType::FLOAT32); // x_grad
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32); // scale_grad
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32); // bias_grad
Expand All @@ -1334,6 +1338,22 @@ PD_REGISTER_KERNEL(batch_norm_grad_raw,
phi::BatchNormGradRawKernel,
float,
double,
phi::dtype::bfloat16,
phi::dtype::float16) {
if (kernel_key.dtype() == phi::DataType::FLOAT16 ||
kernel_key.dtype() == phi::DataType::BFLOAT16) {
kernel->OutputAt(0).SetDataType(phi::DataType::FLOAT32); // x_grad
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32); // scale_grad
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32); // bias_grad
}
}
#else
PD_REGISTER_KERNEL(batch_norm_grad,
GPU,
ALL_LAYOUT,
phi::BatchNormGradKernel,
float,
double,
phi::dtype::float16) {
if (kernel_key.dtype() == phi::DataType::FLOAT16) {
kernel->OutputAt(0).SetDataType(phi::DataType::FLOAT32); // x_grad
Expand All @@ -1342,6 +1362,20 @@ PD_REGISTER_KERNEL(batch_norm_grad_raw,
}
}

PD_REGISTER_KERNEL(batch_norm_grad_raw,
GPU,
ALL_LAYOUT,
phi::BatchNormGradRawKernel,
float,
double,
phi::dtype::float16) {
if (kernel_key.dtype() == phi::DataType::FLOAT16) {
kernel->OutputAt(0).SetDataType(phi::DataType::FLOAT32); // x_grad
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32); // scale_grad
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32); // bias_grad
}
}
#endif
#endif

#ifdef PADDLE_WITH_HIP
Expand Down
24 changes: 24 additions & 0 deletions paddle/phi/kernels/gpu/batch_norm_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1221,6 +1221,7 @@ PD_REGISTER_KERNEL(batch_norm,
ALL_LAYOUT,
phi::BatchNormKernel,
float,
phi::dtype::bfloat16,
phi::dtype::float16) {
kernel->InputAt(1).SetDataType(phi::DataType::FLOAT32);
kernel->InputAt(2).SetDataType(phi::DataType::FLOAT32);
Expand All @@ -1232,6 +1233,28 @@ PD_REGISTER_KERNEL(batch_norm,
kernel->OutputAt(4).SetDataType(phi::DataType::FLOAT32);
}
#else
#if CUDNN_VERSION_MIN(8, 1, 0)
PD_REGISTER_KERNEL(batch_norm,
GPU,
ALL_LAYOUT,
phi::BatchNormKernel,
float,
double,
phi::dtype::bfloat16,
phi::dtype::float16) {
if (kernel_key.dtype() == phi::DataType::FLOAT16 ||
kernel_key.dtype() == phi::DataType::BFLOAT16) {
kernel->InputAt(1).SetDataType(phi::DataType::FLOAT32);
kernel->InputAt(2).SetDataType(phi::DataType::FLOAT32);
kernel->InputAt(3).SetDataType(phi::DataType::FLOAT32);
kernel->InputAt(4).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(3).SetDataType(phi::DataType::FLOAT32);
kernel->OutputAt(4).SetDataType(phi::DataType::FLOAT32);
}
}
#else
PD_REGISTER_KERNEL(batch_norm,
GPU,
ALL_LAYOUT,
Expand All @@ -1250,5 +1273,6 @@ PD_REGISTER_KERNEL(batch_norm,
kernel->OutputAt(4).SetDataType(phi::DataType::FLOAT32);
}
}
#endif

#endif
56 changes: 50 additions & 6 deletions python/paddle/fluid/tests/unittests/test_batch_norm_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,12 @@
import unittest

import numpy as np
from eager_op_test import OpTest, _set_use_system_allocator
from eager_op_test import (
OpTest,
_set_use_system_allocator,
convert_float_to_uint16,
convert_uint16_to_float,
)
from op import Operator

import paddle
Expand Down Expand Up @@ -239,7 +244,10 @@ def check_with_place(self, place, data_layout, dtype, shape):
raise ValueError("Unknown data layout.")
scale_shape = [c]

x_val = np.random.random_sample(x_shape).astype(dtype)
if dtype == np.uint16:
x_val = np.random.random_sample(x_shape).astype(np.float32)
else:
x_val = np.random.random_sample(x_shape).astype(dtype)
# generate some negative values to test case with relu fused
x_val = x_val - 0.5
scale_val = np.random.random_sample(scale_shape).astype(np.float32)
Expand All @@ -248,12 +256,20 @@ def check_with_place(self, place, data_layout, dtype, shape):
mean = np.zeros(scale_shape).astype(np.float32)
variance = np.ones(scale_shape).astype(np.float32)

y_out = _reference_testing(
x_val, scale_val, bias_val, mean, variance, epsilon, data_layout
).astype(dtype)
if dtype == np.uint16:
y_out = _reference_testing(
x_val, scale_val, bias_val, mean, variance, epsilon, data_layout
).astype(np.float32)
y_out = convert_float_to_uint16(y_out)
else:
y_out = _reference_testing(
x_val, scale_val, bias_val, mean, variance, epsilon, data_layout
).astype(dtype)
if self.fuse_with_relu:
y_out = np.maximum(y_out, 0)

if dtype == np.uint16:
x_val = convert_float_to_uint16(x_val)
scope = core.Scope()

# create input
Expand Down Expand Up @@ -324,6 +340,11 @@ def check_with_place(self, place, data_layout, dtype, shape):
y_tensor._set_dims(dims)

# check inference result
atol = 1e-3
if dtype == np.uint16:
y_tensor = convert_uint16_to_float(y_tensor)
y_out = convert_uint16_to_float(y_out)
atol = 1e-2
self.__assert_close(
y_tensor,
y_out,
Expand All @@ -335,7 +356,7 @@ def check_with_place(self, place, data_layout, dtype, shape):
+ str(np.dtype(dtype))
+ str(np.array(y_tensor))
+ str(y_out),
atol=1e-3,
atol=atol,
)

def test_check_output(self):
Expand Down Expand Up @@ -376,6 +397,29 @@ def test_check_output(self):
self.check_with_place(place, data_format, self.dtype, [2, 3])


@unittest.skipIf(
not core.is_compiled_with_cuda()
or not core.is_bfloat16_supported(core.CUDAPlace(0)),
"core is not compiled with CUDA or not support the bfloat16",
)
class TestBF16BatchNormOpInference(TestBatchNormOpInference):
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

是否需要加个skipif跳过不支持bf16的平台

def setUp(self):
self.dtype = np.uint16
self.use_mkldnn = False
self.fuse_with_relu = False
self.init_kernel_type()

def test_check_output(self):
places = [core.CUDAPlace(0)]
for place in places:
# for data_format in ["NCHW", "NHWC"]:
for data_format in ["NCHW"]:
self.check_with_place(
place, data_format, self.dtype, [2, 3, 4, 5]
)
self.check_with_place(place, data_format, self.dtype, [2, 3])


class TestBatchNormOpTraining(unittest.TestCase):
def setUp(self):
self.use_mkldnn = False
Expand Down