Skip to content

Commit

Permalink
[NPU] add NPU ops of stack and unstack, test=develop (#34084)
Browse files Browse the repository at this point in the history
  • Loading branch information
@qili93
qili93 authored Jul 12, 2021
1 parent 2dde0eb commit 0b20b76
Show file tree
Hide file tree
Showing 4 changed files with 404 additions and 120 deletions.
95 changes: 42 additions & 53 deletions paddle/fluid/operators/stack_op_npu.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -12,15 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#ifdef PADDLE_WITH_ASCEND_CL
#include <memory>
#include <string>
#include <vector>

#include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/operators/npu_op_runner.h"
#include "paddle/fluid/operators/stack_op.h"
#include "paddle/fluid/operators/unsqueeze_op.h"
#include "paddle/fluid/operators/npu_op_runner.h"

namespace paddle {
namespace operators {
Expand All @@ -32,64 +25,56 @@ class StackNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto x = ctx.MultiInput<Tensor>("X");
int32_t N = x.size();
auto* y = ctx.Output<Tensor>("Y");
int axis = ctx.Attr<int>("axis");
if (axis < 0) axis += (x[0]->dims().size() + 1);
int num = static_cast<int>(x.size());

PADDLE_ENFORCE_GT(
N, 0, platform::errors::InvalidArgument("number of input Tensor <= 0"));
PADDLE_ENFORCE_GT(num, 0, platform::errors::InvalidArgument(
"number of input Tensor <= 0"));

auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();

std::vector<paddle::framework::Tensor> x_list;
for (int i = 0; i < N; i++) {
for (int i = 0; i < num; i++) {
x_list.push_back(*x[i]);
}
y->mutable_data<T>(ctx.GetPlace());

int axis = ctx.Attr<int>("axis");
const auto& runner =
NpuOpRunner("Pack", {x_list}, {*y}, {{"axis", axis}, {"N", num}});
runner.Run(stream);
}
};

if (axis < 0) {
axis = axis + x_list[0].dims().size() + 1;
}
auto* out = ctx.Output<Tensor>("Y");
template <typename DeviceContext, typename T>
class StackGradNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
auto dx = ctx.MultiOutput<Tensor>(framework::GradVarName("X"));
int axis = ctx.Attr<int>("axis");
if (axis < 0) axis += dy->dims().size();
int num = dy->dims()[axis];

auto place = ctx.GetPlace();
PADDLE_ENFORCE_GT(num, 0, platform::errors::InvalidArgument(
"number of input Tensor <= 0"));

auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();

out->mutable_data<T>(place);

if (axis != 0) {
auto x_dim = x_list[0].dims();
std::vector<int> vec_dim_tmp;
vec_dim_tmp.push_back(N);
for (auto i = 0; i < x_dim.size(); ++i) {
vec_dim_tmp.push_back(x_dim[i]);
}

Tensor tmp_stack(out->type());
tmp_stack.Resize(framework::make_ddim(vec_dim_tmp));
tmp_stack.mutable_data<T>(ctx.GetPlace());

const auto& runner =
NpuOpRunner("Pack", {x_list}, {tmp_stack}, {{"axis", 0}, {"N", N}});
runner.Run(stream);

std::vector<int64_t> vec_trans;
for (auto i = 1; i <= x_dim.size(); ++i) {
vec_trans.push_back(i);
if (i == axis) {
vec_trans.push_back(0);
}
}

const auto& runner_trans_final =
NpuOpRunner("TransposeD", {tmp_stack}, {*out}, {{"perm", vec_trans}});
runner_trans_final.Run(stream);

} else {
const auto& runner =
NpuOpRunner("Pack", {x_list}, {*out}, {{"axis", axis}, {"N", N}});
runner.Run(stream);
std::vector<paddle::framework::Tensor> dx_list;
for (int i = 0; i < num; i++) {
dx[i]->mutable_data<T>(ctx.GetPlace());
dx_list.push_back(*dx[i]);
}

const auto& runner =
NpuOpRunner("Unpack", {*dy}, {dx_list}, {{"axis", axis}, {"num", num}});
runner.Run(stream);
}
};

Expand All @@ -103,4 +88,8 @@ REGISTER_OP_NPU_KERNEL(
ops::StackNPUKernel<paddle::platform::NPUDeviceContext,
paddle::platform::float16>);

#endif
REGISTER_OP_NPU_KERNEL(
stack_grad,
ops::StackGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
ops::StackGradNPUKernel<paddle::platform::NPUDeviceContext,
paddle::platform::float16>);
85 changes: 85 additions & 0 deletions paddle/fluid/operators/unstack_op_npu.cc
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,85 @@
/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/fluid/operators/unstack_op.h"
#include "paddle/fluid/operators/npu_op_runner.h"

namespace paddle {
namespace operators {

template <typename DeviceContext, typename T>
class UnStackNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
auto *dy = ctx.Input<Tensor>("X");
auto dx = ctx.MultiOutput<Tensor>("Y");
int axis = ctx.Attr<int>("axis");
if (axis < 0) axis += dy->dims().size();
int num = dy->dims()[axis];

auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();

std::vector<paddle::framework::Tensor> dx_list;
for (int i = 0; i < num; i++) {
dx[i]->mutable_data<T>(ctx.GetPlace());
dx_list.push_back(*dx[i]);
}

const auto &runner =
NpuOpRunner("Unpack", {*dy}, {dx_list}, {{"axis", axis}, {"num", num}});
runner.Run(stream);
}
};

template <typename DeviceContext, typename T>
class UnStackGradNPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
auto x = ctx.MultiInput<Tensor>(framework::GradVarName("Y"));
auto *y = ctx.Output<Tensor>(framework::GradVarName("X"));
int axis = ctx.Attr<int>("axis");
if (axis < 0) axis += (x[0]->dims().size() + 1);
int num = static_cast<int>(x.size());

auto stream =
ctx.template device_context<paddle::platform::NPUDeviceContext>()
.stream();

std::vector<paddle::framework::Tensor> x_list;
for (int i = 0; i < num; i++) {
x_list.push_back(*x[i]);
}
y->mutable_data<T>(ctx.GetPlace());

const auto &runner =
NpuOpRunner("Pack", {x_list}, {*y}, {{"axis", axis}, {"N", num}});
runner.Run(stream);
}
};

} // namespace operators
} // namespace paddle

namespace plat = paddle::platform;
namespace ops = paddle::operators;

REGISTER_OP_NPU_KERNEL(
unstack, ops::UnStackNPUKernel<plat::NPUDeviceContext, float>,
ops::UnStackNPUKernel<plat::NPUDeviceContext, plat::float16>);

REGISTER_OP_NPU_KERNEL(
unstack_grad, ops::UnStackGradNPUKernel<plat::NPUDeviceContext, float>,
ops::UnStackGradNPUKernel<plat::NPUDeviceContext, plat::float16>);
Loading

0 comments on commit 0b20b76

Please sign in to comment.