Skip to content

[bf16] add bf16 kernel: transpose & unbind #39457

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 3 commits into from
Feb 11, 2022
Merged

[bf16] add bf16 kernel: transpose & unbind #39457

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1464c82
add transpose unbind
zhangbo9674 Feb 9, 2022
dd03a55
add unittest
zhangbo9674 Feb 10, 2022
c91035a
refine transpose unittest
zhangbo9674 Feb 11, 2022
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
16 changes: 12 additions & 4 deletions paddle/fluid/operators/transpose_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -356,7 +356,9 @@ REGISTER_OP_CPU_KERNEL(
ops::TransposeKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<float>>,
ops::TransposeKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<double>>);
paddle::platform::complex<double>>,
ops::TransposeKernel<paddle::platform::CPUDeviceContext,
paddle::platform::bfloat16>);
REGISTER_OP_CPU_KERNEL(
transpose_grad,
ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, bool>,
Expand All @@ -365,7 +367,9 @@ REGISTER_OP_CPU_KERNEL(
ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<float>>,
ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<double>>);
paddle::platform::complex<double>>,
ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::bfloat16>);

REGISTER_OPERATOR(transpose2, ops::Transpose2Op, ops::Transpose2OpMaker,
ops::Transpose2GradMaker<paddle::framework::OpDesc>,
Expand All @@ -383,7 +387,9 @@ REGISTER_OP_CPU_KERNEL(
ops::TransposeKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<float>>,
ops::TransposeKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<double>>);
paddle::platform::complex<double>>,
ops::TransposeKernel<paddle::platform::CPUDeviceContext,
paddle::platform::bfloat16>);
REGISTER_OP_CPU_KERNEL(
transpose2_grad,
ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, bool>,
Expand All @@ -394,4 +400,6 @@ REGISTER_OP_CPU_KERNEL(
ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<float>>,
ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<double>>);
paddle::platform::complex<double>>,
ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::bfloat16>);
9 changes: 9 additions & 0 deletions paddle/fluid/operators/transpose_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@ limitations under the License. */

#include "paddle/fluid/operators/transpose_op.cu.h"
#include "paddle/fluid/operators/transpose_op.h"
#include "paddle/fluid/platform/bfloat16.h"
#include "paddle/fluid/platform/float16.h"

namespace paddle {
Expand Down Expand Up @@ -87,6 +88,8 @@ REGISTER_OP_CUDA_KERNEL(
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext, float>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext, double>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext, plat::float16>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext,
plat::bfloat16>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext,
paddle::platform::complex<float>>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext,
Expand All @@ -98,6 +101,8 @@ REGISTER_OP_CUDA_KERNEL(
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext, double>,
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext,
plat::float16>,
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext,
plat::bfloat16>,
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext,
paddle::platform::complex<float>>,
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext,
Expand All @@ -111,6 +116,8 @@ REGISTER_OP_CUDA_KERNEL(
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext, float>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext, double>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext, plat::float16>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext,
plat::bfloat16>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext,
paddle::platform::complex<float>>,
ops::TransposeGPUKernel<paddle::platform::CUDADeviceContext,
Expand All @@ -124,6 +131,8 @@ REGISTER_OP_CUDA_KERNEL(
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext, double>,
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext,
plat::float16>,
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext,
plat::bfloat16>,
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext,
paddle::platform::complex<float>>,
ops::TransposeGradGPUKernel<paddle::platform::CUDADeviceContext,
Expand Down
3 changes: 2 additions & 1 deletion paddle/fluid/operators/unbind_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,4 +85,5 @@ REGISTER_OP_CPU_KERNEL(
ops::UnbindOpKernel<plat::CPUDeviceContext, float>,
ops::UnbindOpKernel<plat::CPUDeviceContext, int64_t>,
ops::UnbindOpKernel<plat::CPUDeviceContext, int>,
ops::UnbindOpKernel<plat::CPUDeviceContext, plat::float16>);
ops::UnbindOpKernel<plat::CPUDeviceContext, plat::float16>,
ops::UnbindOpKernel<plat::CPUDeviceContext, plat::bfloat16>);
3 changes: 2 additions & 1 deletion paddle/fluid/operators/unbind_op.cu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,4 +20,5 @@ REGISTER_OP_CUDA_KERNEL(
ops::UnbindOpKernel<plat::CUDADeviceContext, float>,
ops::UnbindOpKernel<plat::CUDADeviceContext, int64_t>,
ops::UnbindOpKernel<plat::CUDADeviceContext, int>,
ops::UnbindOpKernel<plat::CUDADeviceContext, plat::float16>);
ops::UnbindOpKernel<plat::CUDADeviceContext, plat::float16>,
ops::UnbindOpKernel<plat::CUDADeviceContext, plat::bfloat16>);
36 changes: 35 additions & 1 deletion python/paddle/fluid/tests/unittests/test_transpose_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,10 +16,11 @@

import unittest
import numpy as np
from op_test import OpTest
from op_test import OpTest, convert_float_to_uint16
import paddle
import paddle.fluid as fluid
from paddle.fluid import Program, program_guard
import paddle.fluid.core as core

paddle.enable_static()

Expand Down Expand Up @@ -113,6 +114,39 @@ def initTestCase(self):
self.axis = (6, 1, 3, 5, 0, 2, 4, 7)


class TestTransposeBF16Op(OpTest):
def setUp(self):
self.init_op_type()
self.initTestCase()
self.dtype = np.uint16
x = np.random.random(self.shape).astype("float32")

self.inputs = {'X': convert_float_to_uint16(x)}
self.attrs = {
'axis': list(self.axis),
'use_mkldnn': self.use_mkldnn,
}
self.outputs = {
'XShape': convert_float_to_uint16(
np.random.random(self.shape).astype("float32")),
'Out': self.inputs['X'].transpose(self.axis)
}

def init_op_type(self):
self.op_type = "transpose2"
self.use_mkldnn = False

def test_check_output(self):
self.check_output(no_check_set=['XShape'])

def test_check_grad(self):
pass

def initTestCase(self):
self.shape = (3, 2)
self.axis = (1, 0)


class TestTransposeOpBool(TestTransposeOp):
def test_check_grad(self):
pass
Expand Down
28 changes: 27 additions & 1 deletion python/paddle/fluid/tests/unittests/test_unbind_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@

import unittest
import numpy as np
from op_test import OpTest
from op_test import OpTest, convert_float_to_uint16
import paddle.fluid as fluid
import paddle.tensor as tensor
from paddle.fluid import compiler, Program, program_guard, core
Expand Down Expand Up @@ -154,6 +154,32 @@ def outReshape(self):
self.out[1] = self.out[1].reshape((3, 2))


class TestUnbindBF16Op(OpTest):
def setUp(self):
self._set_op_type()
self.dtype = self.get_dtype()
self.axis = 0
self.num = 3
x = np.arange(12).reshape(3, 2, 2).astype(self.dtype)
self.out = np.split(x, self.num, self.axis)
self.inputs = {'X': convert_float_to_uint16(x)}
self.attrs = {'axis': self.axis}
self.outputs = {'Out': [('out%d' % i, convert_float_to_uint16(self.out[i])) \
for i in range(len(self.out))]}

def get_dtype(self):
return np.uint16

def _set_op_type(self):
self.op_type = "unbind"

def test_check_output(self):
self.check_output()

def test_check_grad(self):
pass


class TestUnbindAxisError(unittest.TestCase):
def test_errors(self):
with program_guard(Program(), Program()):
Expand Down