Skip to content

[NPU] add conv_op_npu and test #34055

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
ronny1996 merged 4 commits into from
Aug 24, 2021
Merged

[NPU] add conv_op_npu and test #34055

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3322c01
add conv_op_npu and test
ronny1996 Jul 9, 2021
c4f8cdc
add more tests
ronny1996 Jul 13, 2021
2f535ab
clean headers & support fp16
ronny1996 Jul 15, 2021
6d719df
update
ronny1996 Aug 20, 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
167 changes: 167 additions & 0 deletions paddle/fluid/operators/conv_op_npu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -126,6 +126,169 @@ class DepthwiseConvNPUKernel : public framework::OpKernel<T> {
}
};

template <typename T>
class NPUConvOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto& dev_ctx = ctx.template device_context<platform::NPUDeviceContext>();
const Tensor* input = ctx.Input<Tensor>("Input");
auto* filter = ctx.Input<Tensor>("Filter");
auto* output = ctx.Output<Tensor>("Output");
output->mutable_data<T>(ctx.GetPlace());
const std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
int groups = ctx.Attr<int>("groups");
const std::string padding_algorithm =
ctx.Attr<std::string>("padding_algorithm");
const std::string data_format = ctx.Attr<std::string>("data_format");

const bool channel_last = data_format == "NHWC";

// update padding and dilation
auto in_dims = input->dims();
auto filter_dims = filter->dims();
framework::DDim in_data_dims;
framework::DDim filter_data_dims;

if (channel_last) {
in_data_dims = framework::slice_ddim(in_dims, 1, in_dims.size() - 1);
} else {
in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
}
filter_data_dims = framework::slice_ddim(filter_dims, 2, in_dims.size());

std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm,
in_data_dims, strides, ksize);

std::vector<int> strides_vec(4, 1);
std::vector<int> dilations_vec(4, 1);

Tensor input_tensor, output_tensor;
input_tensor.ShareDataWith(*input);
output_tensor.ShareDataWith(*output);
if (channel_last) {
input_tensor.set_layout(DataLayout::kNHWC);
output_tensor.set_layout(DataLayout::kNHWC);
strides_vec[1] = strides[0];
strides_vec[2] = strides[1];
dilations_vec[1] = dilations[0];
dilations_vec[2] = dilations[1];
} else {
strides_vec[2] = strides[0];
strides_vec[3] = strides[1];
dilations_vec[2] = dilations[0];
dilations_vec[3] = dilations[1];
}

const auto& runner =
NpuOpRunner("Conv2D", {input_tensor, *filter}, {output_tensor},
{{"strides", strides_vec},
{"pads", paddings},
{"dilations", dilations_vec},
{"groups", groups},
{"data_format", data_format}});
runner.Run(dev_ctx.stream());
}
};

template <typename T>
class NPUConvGradOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto& dev_ctx = ctx.template device_context<platform::NPUDeviceContext>();

auto input = ctx.Input<Tensor>("Input");
auto filter = ctx.Input<Tensor>("Filter");
auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
auto input_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
auto filter_grad = ctx.Output<Tensor>(framework::GradVarName("Filter"));

const std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
int groups = ctx.Attr<int>("groups");
const std::string padding_algorithm =
ctx.Attr<std::string>("padding_algorithm");
const std::string data_format = ctx.Attr<std::string>("data_format");

const bool channel_last = data_format == "NHWC";

// update padding and dilation
auto in_dims = input->dims();
auto filter_dims = filter->dims();
framework::DDim in_data_dims;
framework::DDim filter_data_dims;

if (channel_last) {
in_data_dims = framework::slice_ddim(in_dims, 1, in_dims.size() - 1);
} else {
in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
}
filter_data_dims = framework::slice_ddim(filter_dims, 2, in_dims.size());

std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm,
in_data_dims, strides, ksize);

std::vector<int> strides_vec(4, 1);
std::vector<int> dilations_vec(4, 1);

Tensor input_tensor, output_grad_tensor;
input_tensor.ShareDataWith(*input);
output_grad_tensor.ShareDataWith(*output_grad);
if (channel_last) {
input_tensor.set_layout(DataLayout::kNHWC);
output_grad_tensor.set_layout(DataLayout::kNHWC);
strides_vec[1] = strides[0];
strides_vec[2] = strides[1];
dilations_vec[1] = dilations[0];
dilations_vec[2] = dilations[1];
} else {
strides_vec[2] = strides[0];
strides_vec[3] = strides[1];
dilations_vec[2] = dilations[0];
dilations_vec[3] = dilations[1];
}

if (filter_grad) {
filter_grad->mutable_data<T>(ctx.GetPlace());
std::vector<int> filter_shape_vec =
framework::vectorize<int>(filter->dims());

const auto& runner = NpuOpRunner(
"Conv2DBackpropFilterD", {input_tensor, output_grad_tensor},
{*filter_grad}, {{"filter_size", filter_shape_vec},
{"strides", strides_vec},
{"pads", paddings},
{"dilations", dilations_vec},
{"groups", groups},
{"data_format", data_format}});
runner.Run(dev_ctx.stream());
}
if (input_grad) {
input_grad->mutable_data<T>(ctx.GetPlace());
std::vector<int> input_shape_vec =
framework::vectorize<int>(input->dims());

Tensor input_grad_tensor;
input_grad_tensor.ShareDataWith(*input_grad);
if (channel_last) {
input_grad_tensor.set_layout(DataLayout::kNHWC);
}
const auto& runner =
NpuOpRunner("Conv2DBackpropInputD", {*filter, output_grad_tensor},
{input_grad_tensor}, {{"input_size", input_shape_vec},
{"strides", strides_vec},
{"pads", paddings},
{"dilations", dilations_vec},
{"groups", groups},
{"data_format", data_format}});
runner.Run(dev_ctx.stream());
}
}
};
} // namespace operators
} // namespace paddle

Expand All @@ -135,3 +298,7 @@ REGISTER_OP_NPU_KERNEL(
depthwise_conv2d,
ops::DepthwiseConvNPUKernel<paddle::platform::NPUDeviceContext,
paddle::platform::float16>);
REGISTER_OP_NPU_KERNEL(conv2d, ops::NPUConvOpKernel<float>,
ops::NPUConvOpKernel<paddle::platform::float16>);
REGISTER_OP_NPU_KERNEL(conv2d_grad, ops::NPUConvGradOpKernel<float>,
ops::NPUConvGradOpKernel<paddle::platform::float16>);
Loading