m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #9167 from kexinzhao/pool2d_fp16

Browse Source 
Add float16 support for pool 2d operator
shanyi15-patch-2
Kexin Zhao 7 years ago committed by GitHub
parent 3f5705c346 dfec1df14f
commit df99b16a16
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 152 additions and 102 deletions
Show Stats Download Patch File Download Diff File

8
paddle/fluid/operators/conv_cudnn_op.cu.cc
Unescape View File

@ -28,6 +28,8 @@ using ScopedTensorDescriptor = platform::ScopedTensorDescriptor;
using ScopedFilterDescriptor = platform::ScopedFilterDescriptor;
using ScopedConvolutionDescriptor = platform::ScopedConvolutionDescriptor;
using DataLayout = platform::DataLayout;
template <typename T>
using ScalingParamType = typename platform::CudnnDataType<T>::ScalingParamType;
static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES =
static_cast<size_t>(1024) * 1024 * 1024;
@ -134,8 +136,7 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
// ------------------- cudnn conv forward ---------------------
typename platform::CudnnDataType<T>::ScalingParamType alpha = 1.0f,
beta = 0.0f;
ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
for (int i = 0; i < groups; i++) {
PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward(
handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,
@ -282,8 +283,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
// ------------------- cudnn conv backward data ---------------------
typename platform::CudnnDataType<T>::ScalingParamType alpha = 1.0f,
beta = 0.0f;
ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
if (input_grad) {
T* input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
// Because beta is zero, it is unnecessary to reset input_grad.

20
paddle/fluid/operators/pool_cudnn_op.cu.cc
Unescape View File

@ -24,6 +24,8 @@ using ScopedTensorDescriptor = platform::ScopedTensorDescriptor;
using ScopedPoolingDescriptor = platform::ScopedPoolingDescriptor;
using DataLayout = platform::DataLayout;
using PoolingMode = platform::PoolingMode;
template <typename T>
using ScalingParamType = typename platform::CudnnDataType<T>::ScalingParamType;
template <typename T>
class PoolCUDNNOpKernel : public framework::OpKernel<T> {
@ -78,8 +80,7 @@ class PoolCUDNNOpKernel : public framework::OpKernel<T> {
// ------------------- cudnn pool algorithm ---------------------
auto handle = ctx.cuda_device_context().cudnn_handle();
T alpha = 1.0f, beta = 0.0f;
ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
PADDLE_ENFORCE(platform::dynload::cudnnPoolingForward(
handle, cudnn_pool_desc, &alpha, cudnn_input_desc, input_data, &beta,
cudnn_output_desc, output_data));
@ -144,8 +145,7 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel<T> {
// ------------------- cudnn pool algorithm ---------------------
auto handle = ctx.cuda_device_context().cudnn_handle();
T alpha = 1.0f, beta = 0.0f;
ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
if (input_grad) {
T *input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
// Because beta is zero, it is unnecessary to reset input_grad.
@ -162,17 +162,19 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel<T> {
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_KERNEL(pool2d, CUDNN, ::paddle::platform::CUDAPlace,
REGISTER_OP_KERNEL(pool2d, CUDNN, plat::CUDAPlace,
ops::PoolCUDNNOpKernel<float>,
ops::PoolCUDNNOpKernel<double>);
REGISTER_OP_KERNEL(pool2d_grad, CUDNN, ::paddle::platform::CUDAPlace,
ops::PoolCUDNNOpKernel<double>,
ops::PoolCUDNNOpKernel<plat::float16>);
REGISTER_OP_KERNEL(pool2d_grad, CUDNN, plat::CUDAPlace,
ops::PoolCUDNNGradOpKernel<float>,
ops::PoolCUDNNGradOpKernel<double>);
REGISTER_OP_KERNEL(pool3d, CUDNN, ::paddle::platform::CUDAPlace,
REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace,
ops::PoolCUDNNOpKernel<float>,
ops::PoolCUDNNOpKernel<double>);
REGISTER_OP_KERNEL(pool3d_grad, CUDNN, ::paddle::platform::CUDAPlace,
REGISTER_OP_KERNEL(pool3d_grad, CUDNN, plat::CUDAPlace,
ops::PoolCUDNNGradOpKernel<float>,
ops::PoolCUDNNGradOpKernel<double>);

10
paddle/fluid/operators/pool_op.cc
Unescape View File

@ -124,11 +124,15 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
}
#endif
auto input_data_type = framework::ToDataType(ctx.Input<Tensor>("X")->type());
if (input_data_type == framework::proto::VarType::FP16) {
PADDLE_ENFORCE_EQ(library_, framework::LibraryType::kCUDNN,
"float16 can only be used when CUDNN is used");
}
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout_, library_);
return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_,
library_);
}
Pool2dOpMaker::Pool2dOpMaker(OpProto *proto, OpAttrChecker *op_checker)

6
python/paddle/fluid/tests/unittests/op_test.py
Unescape View File

@ -483,9 +483,9 @@ class OpTest(unittest.TestCase):
input: input numpy array
Returns:
input: if the dtype of input is np.float16, its dtype will be
changed to np.uint16 so that the internal memory will be
reinterpreted input as of dtype np.uint16.
input: The dtype of input will be changed to np.uint16 if
it is originally np.float16, such that the internal memory
of input will be reinterpreted as of dtype np.uint16.
"""
if input.dtype == np.float16:
input.dtype = np.uint16

69
python/paddle/fluid/tests/unittests/test_conv2d_op.py
Unescape View File

@ -63,12 +63,13 @@ def conv2d_forward_naive(input, filter, group, conv_param):
class TestConv2dOp(OpTest):
def setUp(self):
self.op_type = "conv2d"
self.use_cudnn = False
self.use_mkldnn = False
self.init_op_type()
self.dtype = np.float32
self.init_kernel_type()
self.init_group()
self.init_dilation()
self.init_data_type()
self.init_test_case()
conv2d_param = {
@ -159,17 +160,14 @@ class TestConv2dOp(OpTest):
f_c = self.input_size[1] / self.groups
self.filter_size = [6, f_c, 3, 3]
def init_data_type(self):
self.dtype = np.float32
def init_dilation(self):
self.dilations = [1, 1]
def init_group(self):
self.groups = 1
def init_op_type(self):
self.op_type = "conv2d"
def init_kernel_type(self):
pass
class TestWithPad(TestConv2dOp):
@ -241,13 +239,13 @@ class TestWithInput1x1Filter1x1(TestConv2dOp):
#----------------Conv2dCUDNN----------------
class TestCUDNN(TestConv2dOp):
def init_op_type(self):
def init_kernel_type(self):
self.use_cudnn = True
self.op_type = "conv2d"
class TestFP16CUDNN(TestCUDNN):
def init_data_type(self):
class TestFP16CUDNN(TestConv2dOp):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
@ -258,13 +256,13 @@ class TestFP16CUDNN(TestCUDNN):
class TestCUDNNWithPad(TestWithPad):
def init_op_type(self):
def init_kernel_type(self):
self.use_cudnn = True
self.op_type = "conv2d"
class TestFP16CUDNNWithPad(TestCUDNNWithPad):
def init_data_type(self):
class TestFP16CUDNNWithPad(TestWithPad):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
@ -275,13 +273,13 @@ class TestFP16CUDNNWithPad(TestCUDNNWithPad):
class TestCUDNNWithStride(TestWithStride):
def init_op_type(self):
def init_kernel_type(self):
self.use_cudnn = True
self.op_type = "conv2d"
class TestFP16CUDNNWithStride(TestCUDNNWithStride):
def init_data_type(self):
class TestFP16CUDNNWithStride(TestWithStride):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
@ -292,13 +290,13 @@ class TestFP16CUDNNWithStride(TestCUDNNWithStride):
class TestCUDNNWithGroup(TestWithGroup):
def init_op_type(self):
def init_kernel_type(self):
self.use_cudnn = True
self.op_type = "conv2d"
class TestFP16CUDNNWithGroup(TestCUDNNWithGroup):
def init_data_type(self):
class TestFP16CUDNNWithGroup(TestWithGroup):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
@ -309,13 +307,13 @@ class TestFP16CUDNNWithGroup(TestCUDNNWithGroup):
class TestCUDNNWith1x1(TestWith1x1):
def init_op_type(self):
def init_kernel_type(self):
self.use_cudnn = True
self.op_type = "conv2d"
class TestFP16CUDNNWith1x1(TestCUDNNWith1x1):
def init_data_type(self):
class TestFP16CUDNNWith1x1(TestWith1x1):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
@ -326,13 +324,13 @@ class TestFP16CUDNNWith1x1(TestCUDNNWith1x1):
class TestCUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
def init_op_type(self):
def init_kernel_type(self):
self.use_cudnn = True
self.op_type = "conv2d"
class TestFP16CUDNNWithInput1x1Filter1x1(TestCUDNNWithInput1x1Filter1x1):
def init_data_type(self):
class TestFP16CUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
@ -375,21 +373,18 @@ class TestDepthwiseConv2(TestConv2dOp):
#----------------Conv2dMKLDNN----------------
class TestMKLDNN(TestConv2dOp):
def init_op_type(self):
def init_kernel_type(self):
self.use_mkldnn = True
self.op_type = "conv2d"
class TestMKLDNNWithPad(TestWithPad):
def init_op_type(self):
def init_kernel_type(self):
self.use_mkldnn = True
self.op_type = "conv2d"
class TestMKLDNNWithStride(TestWithStride):
def init_op_type(self):
def init_kernel_type(self):
self.use_mkldnn = True
self.op_type = "conv2d"
if __name__ == '__main__':

141
python/paddle/fluid/tests/unittests/test_pool2d_op.py
View File

File diff suppressed because it is too large Load Diff
Write Preview
Loading…
Cancel
Save