Skip to content

Commit

Permalink
Add conv yaml (#41354)
Browse files Browse the repository at this point in the history 
* update

* add conv yaml

* add backward

* remove useless code

* fix bug

* fix bug

* revert fluid dygraph conv2d

* remove useless infermeta function

* fix meta fn deluplicat error

* conv using custom impl

* remove amp include

* fix bug

* use cudnn = true

* fix test mkldnn caching bug
  • Loading branch information
@phlrain
phlrain committed Apr 6, 2022
1 parent 0b96793 commit 7ed7c6c
Show file tree
Hide file tree
Showing 19 changed files with 312 additions and 29 deletions.
6 changes: 0 additions & 6 deletions paddle/fluid/operators/conv_op.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -844,8 +844,6 @@ framework::OpKernelType ConvOpDoubleGrad::GetExpectedKernelType(
} // namespace paddle

namespace ops = paddle::operators;
DECLARE_INFER_SHAPE_FUNCTOR(conv2d, Conv2dInferShapeFunctor,
PD_INFER_META(phi::ConvInferMeta));
REGISTER_OPERATOR(conv2d, ops::ConvOp, ops::Conv2DOpMaker,
ops::ConvOpInferVarType,
ops::Conv2DGradMaker<paddle::framework::OpDesc>,
Expand All @@ -856,8 +854,6 @@ REGISTER_OPERATOR(conv2d_grad, ops::ConvOpGrad,
REGISTER_OPERATOR(conv2d_grad_grad, ops::ConvOpDoubleGrad);

// depthwise convolution op
DECLARE_INFER_SHAPE_FUNCTOR(depthwise_conv2d, DepthwiseConv2dInferShapeFunctor,
PD_INFER_META(phi::ConvInferMeta));
REGISTER_OPERATOR(depthwise_conv2d, ops::ConvOp, ops::Conv2DOpMaker,
ops::ConvOpInferVarType,
ops::Conv2DGradMaker<paddle::framework::OpDesc>,
Expand All @@ -867,8 +863,6 @@ REGISTER_OPERATOR(depthwise_conv2d_grad, ops::ConvOpGrad,
ops::Conv2DDoubleGradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(depthwise_conv2d_grad_grad, ops::ConvOpDoubleGrad);

DECLARE_INFER_SHAPE_FUNCTOR(conv3d, Conv3dInferShapeFunctor,
PD_INFER_META(phi::ConvInferMeta));
REGISTER_OPERATOR(conv3d, ops::ConvOp, ops::Conv3DOpMaker,
ops::ConvOpInferVarType,
ops::Conv3DGradMaker<paddle::framework::OpDesc>,
Expand Down
208 changes: 208 additions & 0 deletions paddle/phi/api/lib/api_custom_impl.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,213 @@ namespace experimental {

////////////////// Forward api impls //////////////////////

Tensor conv2d_impl(const Tensor& input,
const Tensor& filter,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
int groups,
const std::vector<int>& dilations,
const std::string& data_format,
bool use_addto,
int workspace_size_MB,
bool exhaustive_search) {
Backend kernel_backend = Backend::UNDEFINED;
DataLayout kernel_layout = DataLayout::UNDEFINED;
DataType kernel_data_type = DataType::UNDEFINED;

kernel_data_type = ParseDataType(input);

if (kernel_backend == Backend::UNDEFINED ||
kernel_layout == DataLayout::UNDEFINED ||
kernel_data_type == DataType::UNDEFINED) {
auto kernel_key_set = ParseKernelKeyByInputArgs(input, filter);
auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
if (kernel_backend == Backend::UNDEFINED) {
kernel_backend = kernel_key.backend();
}
if (kernel_layout == DataLayout::UNDEFINED) {
kernel_layout = kernel_key.layout();
}
if (kernel_data_type == DataType::UNDEFINED) {
kernel_data_type = kernel_key.dtype();
}
}

VLOG(6) << "conv2d API kernel key: [" << kernel_backend << ", "
<< kernel_layout << ", " << kernel_data_type << "]";
const auto& kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
"conv2d", {kernel_backend, kernel_layout, kernel_data_type}, true);
VLOG(6) << "conv2d API kernel: " << kernel;

auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);

phi::TensorArgDef args0 = kernel.InputAt(0);
phi::TensorArgDef args1 = kernel.InputAt(1);
if (kernel_backend == Backend::GPU) {
args0.backend = Backend::GPU;
args1.backend = Backend::GPU;
}

auto input_input = PrepareData(input, args0, {});
auto input_filter = PrepareData(filter, args1, {});

Tensor api_output;
auto kernel_out = SetKernelOutput(kernel_backend, &api_output);
phi::MetaTensor meta_out(kernel_out);

phi::ConvInferMeta(MakeMetaTensor(*input_input),
MakeMetaTensor(*input_filter),
strides,
paddings,
paddding_algorithm,
groups,
dilations,
data_format,
use_addto,
workspace_size_MB,
exhaustive_search,
&meta_out);

using kernel_signature = void (*)(const platform::DeviceContext&,
const phi::DenseTensor&,
const phi::DenseTensor&,
const std::vector<int>&,
const std::vector<int>&,
const std::string&,
int,
const std::vector<int>&,
const std::string&,
bool,
int,
bool,
phi::DenseTensor*);
auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();

{
(*kernel_fn)(*dev_ctx,
*input_input,
*input_filter,
strides,
paddings,
paddding_algorithm,
groups,
dilations,
data_format,
use_addto,
workspace_size_MB,
exhaustive_search,
kernel_out);
}

return api_output;
}

std::vector<std::vector<Tensor>> conv2d_grad_impl(
const Tensor& input,
const Tensor& filter,
const Tensor& out_grad,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
int groups,
const std::vector<int>& dilations,
const std::string& data_format,
bool use_addto,
int workspace_size_MB,
bool exhaustive_search) {
Backend kernel_backend = Backend::UNDEFINED;
DataLayout kernel_layout = DataLayout::UNDEFINED;
DataType kernel_data_type = DataType::UNDEFINED;

if (kernel_backend == Backend::UNDEFINED ||
kernel_layout == DataLayout::UNDEFINED ||
kernel_data_type == DataType::UNDEFINED) {
auto kernel_key_set = ParseKernelKeyByInputArgs(input, filter, out_grad);
auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
if (kernel_backend == Backend::UNDEFINED) {
kernel_backend = kernel_key.backend();
}
if (kernel_layout == DataLayout::UNDEFINED) {
kernel_layout = kernel_key.layout();
}
if (kernel_data_type == DataType::UNDEFINED) {
kernel_data_type = kernel_key.dtype();
}
}

VLOG(6) << "conv2d_grad API kernel key: [" << kernel_backend << ", "
<< kernel_layout << ", " << kernel_data_type << "]";
const auto& kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
"conv2d_grad", {kernel_backend, kernel_layout, kernel_data_type}, true);
VLOG(6) << "conv2d_grad API kernel: " << kernel;

auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);

phi::TensorArgDef args0 = kernel.InputAt(0);
phi::TensorArgDef args1 = kernel.InputAt(1);
phi::TensorArgDef args2 = kernel.InputAt(2);
if (kernel_backend == Backend::GPU) {
args0.backend = Backend::GPU;
args1.backend = Backend::GPU;
args2.backend = Backend::GPU;
}

auto input_input = PrepareData(input, args0, {});
auto input_filter = PrepareData(filter, args1, {});
auto input_out_grad = PrepareData(out_grad, args2, {});

std::vector<std::vector<Tensor>> api_output(2);
api_output[0].emplace_back();
auto kernel_out_0 = SetKernelOutput(kernel_backend, &api_output[0][0]);
api_output[1].emplace_back();
auto kernel_out_1 = SetKernelOutput(kernel_backend, &api_output[1][0]);
phi::MetaTensor meta_out_0(kernel_out_0);
phi::MetaTensor meta_out_1(kernel_out_1);

phi::GeneralBinaryGradInferMeta(MakeMetaTensor(*input_input),
MakeMetaTensor(*input_filter),
&meta_out_0,
&meta_out_1);

using kernel_signature = void (*)(const platform::DeviceContext&,
const phi::DenseTensor&,
const phi::DenseTensor&,
const phi::DenseTensor&,
const std::vector<int>&,
const std::vector<int>&,
const std::string&,
int,
const std::vector<int>&,
const std::string&,
bool,
int,
bool,
phi::DenseTensor*,
phi::DenseTensor*);
auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();

{
(*kernel_fn)(*dev_ctx,
*input_input,
*input_filter,
*input_out_grad,
strides,
paddings,
paddding_algorithm,
groups,
dilations,
data_format,
use_addto,
workspace_size_MB,
exhaustive_search,
kernel_out_0,
kernel_out_1);
}

return api_output;
}

Tensor copy_to_impl(const Tensor& x, Place place, bool blocking) {
auto kernel_key_set = ParseKernelKeyByInputArgs(x);
kernel_key_set.backend_set =
Expand Down Expand Up @@ -61,6 +268,7 @@ Tensor copy_to_impl(const Tensor& x, Place place, bool blocking) {
phi::DenseTensor*);

auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();

(*kernel_fn)(*dev_ctx, *dense_x, place, blocking, kernel_out);

return out;
Expand Down
26 changes: 26 additions & 0 deletions paddle/phi/api/lib/api_custom_impl.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,32 @@ namespace experimental {

////////////////// Forward api impls //////////////////////

Tensor conv2d_impl(const Tensor& input,
const Tensor& filter,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
int groups,
const std::vector<int>& dilations,
const std::string& data_format,
bool use_addto,
int workspace_size_MB,
bool exhaustive_search);

std::vector<std::vector<Tensor>> conv2d_grad_impl(
const Tensor& input,
const Tensor& filter,
const Tensor& out_grad,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
int groups,
const std::vector<int>& dilations,
const std::string& data_format,
bool use_addto,
int workspace_size_MB,
bool exhaustive_search);

Tensor copy_to_impl(const Tensor& x, Place place, bool blocking);

std::vector<Tensor> split_impl(const Tensor& x,
Expand Down
6 changes: 3 additions & 3 deletions paddle/phi/kernels/conv_grad_kernel.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,9 +20,9 @@ namespace phi {

template <typename T, typename Context>
void ConvGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const DenseTensor& input,
const DenseTensor& filter,
const DenseTensor& out_grad,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
Expand All @@ -37,9 +37,9 @@ void ConvGradKernel(const Context& dev_ctx,

template <typename T, typename Context>
void Conv3DGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const DenseTensor& input,
const DenseTensor& filter,
const DenseTensor& out_grad,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
Expand All @@ -54,9 +54,9 @@ void Conv3DGradKernel(const Context& dev_ctx,

template <typename T, typename Context>
void DepthwiseConvGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const DenseTensor& input,
const DenseTensor& filter,
const DenseTensor& out_grad,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
Expand Down
8 changes: 4 additions & 4 deletions paddle/phi/kernels/cpu/conv_grad_kernel.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,9 +22,9 @@ namespace phi {

template <typename T, typename Context>
void DepthwiseConvGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const DenseTensor& input,
const DenseTensor& filter,
const DenseTensor& out_grad,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
Expand All @@ -38,9 +38,9 @@ void DepthwiseConvGradKernel(const Context& dev_ctx,
DenseTensor* input_grad,
DenseTensor* filter_grad) {
ConvGradKernel<T>(dev_ctx,
out_grad,
input,
filter,
out_grad,
strides,
paddings,
paddding_algorithm,
Expand All @@ -56,9 +56,9 @@ void DepthwiseConvGradKernel(const Context& dev_ctx,

template <typename T, typename Context>
void Conv3DGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const DenseTensor& input,
const DenseTensor& filter,
const DenseTensor& out_grad,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
Expand All @@ -71,9 +71,9 @@ void Conv3DGradKernel(const Context& dev_ctx,
DenseTensor* input_grad,
DenseTensor* filter_grad) {
ConvGradKernel<T>(dev_ctx,
out_grad,
input,
filter,
out_grad,
strides,
paddings,
paddding_algorithm,
Expand Down
4 changes: 2 additions & 2 deletions paddle/phi/kernels/gpu/conv_grad_kernel.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,9 +22,9 @@ namespace phi {

template <typename T, typename Context>
void Conv3DGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const DenseTensor& input,
const DenseTensor& filter,
const DenseTensor& out_grad,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& paddding_algorithm,
Expand All @@ -37,9 +37,9 @@ void Conv3DGradKernel(const Context& dev_ctx,
DenseTensor* input_grad,
DenseTensor* filter_grad) {
ConvGradKernel<T>(dev_ctx,
out_grad,
input,
filter,
out_grad,
strides,
paddings,
paddding_algorithm,
Expand Down
2 changes: 1 addition & 1 deletion paddle/phi/kernels/gpu/depthwise_conv_grad_kernel.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -24,9 +24,9 @@ namespace phi {

template <typename T, typename Context>
void DepthwiseConvGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const DenseTensor& input,
const DenseTensor& filter,
const DenseTensor& out_grad,
const std::vector<int>& strides_t,
const std::vector<int>& paddings_t,
const std::string& padding_algorithm,
Expand Down
Loading

0 comments on commit 7ed7c6c

Please sign in to comment.