-
Notifications
You must be signed in to change notification settings - Fork 5.6k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge branch 'develop' into opt/aligned_vector
- Loading branch information
Showing
51 changed files
with
6,615 additions
and
676 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,120 @@ | ||
| /* 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/gather_nd_op.h" | ||
| #include "paddle/fluid/operators/npu_op_runner.h" | ||
|
|
||
| namespace paddle { | ||
| namespace operators { | ||
|
|
||
| template <typename DeviceContext, typename T> | ||
| class GatherNdNPUKernel : public framework::OpKernel<T> { | ||
| public: | ||
| void Compute(const framework::ExecutionContext &ctx) const override { | ||
| auto *x = ctx.Input<Tensor>("X"); | ||
| auto *index = ctx.Input<Tensor>("Index"); | ||
| auto *out = ctx.Output<Tensor>("Out"); | ||
|
|
||
| out->template mutable_data<T>(ctx.GetPlace()); | ||
|
|
||
| if (x->numel() == 0) return; | ||
|
|
||
| if (index->numel() == 0) { | ||
| framework::TensorCopy(*x, ctx.GetPlace(), ctx.device_context(), out); | ||
| return; | ||
| } | ||
|
|
||
| const auto &index_type = index->type(); | ||
| bool index_type_match = index_type == framework::proto::VarType::INT32 || | ||
| index_type == framework::proto::VarType::INT64; | ||
| PADDLE_ENFORCE_EQ(index_type_match, true, | ||
| platform::errors::InvalidArgument( | ||
| "Index holds the wrong type, it holds [%s]," | ||
| "but desires to be [%s] or [%s]", | ||
| paddle::framework::DataTypeToString(index_type), | ||
| paddle::framework::DataTypeToString( | ||
| framework::proto::VarType::INT32), | ||
| paddle::framework::DataTypeToString( | ||
| framework::proto::VarType::INT64))); | ||
|
|
||
| const auto &runner = NpuOpRunner("GatherNd", {*x, *index}, {*out}, {}); | ||
| auto stream = | ||
| ctx.template device_context<paddle::platform::NPUDeviceContext>() | ||
| .stream(); | ||
| runner.Run(stream); | ||
| } | ||
| }; | ||
|
|
||
| template <typename DeviceContext, typename T> | ||
| class GatherNdGradNPUKernel : public framework::OpKernel<T> { | ||
| public: | ||
| void Compute(const framework::ExecutionContext &ctx) const override { | ||
| auto *index = ctx.Input<Tensor>("Index"); | ||
| auto *x = ctx.Input<Tensor>("X"); | ||
| auto *dout = ctx.Input<Tensor>(framework::GradVarName("Out")); | ||
| auto *dx = ctx.Output<Tensor>(framework::GradVarName("X")); | ||
| auto *p = dx->mutable_data<T>(ctx.GetPlace()); | ||
|
|
||
| if (dx->numel() == 0) return; | ||
|
|
||
| if (index->numel() == 0) { | ||
| framework::TensorCopy(*dout, ctx.GetPlace(), ctx.device_context(), dx); | ||
| return; | ||
| } | ||
|
|
||
| framework::Tensor tmp_tensor(index->type()); | ||
| framework::Tensor tmp_tensor2(dout->type()); | ||
| const auto index_dims = index->dims(); | ||
| if (index_dims.size() == 1) { | ||
| tmp_tensor.ShareDataWith(*index); | ||
| std::vector<int64_t> new_dim = {1, index_dims[0]}; | ||
| tmp_tensor.Resize(framework::make_ddim(new_dim)); | ||
| index = &tmp_tensor; | ||
|
|
||
| tmp_tensor2.ShareDataWith(*dout); | ||
| std::vector<int64_t> new_dim2{1}; | ||
| for (int i = index->numel(); i < x->dims().size(); i++) { | ||
| new_dim2.push_back(x->dims()[i]); | ||
| } | ||
| tmp_tensor2.Resize(framework::make_ddim(new_dim2)); | ||
| dout = &tmp_tensor2; | ||
| } | ||
|
|
||
| auto stream = | ||
| ctx.template device_context<paddle::platform::NPUDeviceContext>() | ||
| .stream(); | ||
|
|
||
| platform::NPUMemsetAsync(static_cast<void *>(p), 0, dx->numel() * sizeof(T), | ||
| stream); | ||
|
|
||
| const auto &runner_scatter = NpuOpRunner( | ||
| "ScatterNdAdd", {*dx, *index, *dout}, {*dx}, {{"use_locking", false}}); | ||
| runner_scatter.Run(stream); | ||
| } | ||
| }; | ||
|
|
||
| } // namespace operators | ||
| } // namespace paddle | ||
|
|
||
| namespace ops = paddle::operators; | ||
| REGISTER_OP_NPU_KERNEL( | ||
| gather_nd, ops::GatherNdNPUKernel<paddle::platform::NPUDeviceContext, | ||
| paddle::platform::float16>, | ||
| ops::GatherNdNPUKernel<paddle::platform::NPUDeviceContext, float>); | ||
|
|
||
| REGISTER_OP_NPU_KERNEL( | ||
| gather_nd_grad, | ||
| ops::GatherNdGradNPUKernel<paddle::platform::NPUDeviceContext, | ||
| paddle::platform::float16>, | ||
| ops::GatherNdGradNPUKernel<paddle::platform::NPUDeviceContext, float>); |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,108 @@ | ||
| // 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/label_smooth_op.h" | ||
| #include "paddle/fluid/operators/npu_op_runner.h" | ||
|
|
||
| namespace paddle { | ||
| namespace operators { | ||
|
|
||
| using Tensor = framework::Tensor; | ||
| using LoDTensor = framework::LoDTensor; | ||
|
|
||
| template <typename T> | ||
| void LabelSmoothMuls(const platform::Place& place, const aclrtStream& stream, | ||
| const Tensor* in, float val, Tensor* out) { | ||
| out->mutable_data<T>(in->dims(), place); | ||
| const auto& runner = NpuOpRunner("Muls", {*in}, {*out}, {{"value", val}}); | ||
| runner.Run(stream); | ||
| } | ||
|
|
||
| template <typename T> | ||
| void LabelSmoothAdds(const platform::Place& place, const aclrtStream& stream, | ||
| const Tensor* in, float val, Tensor* out) { | ||
| out->mutable_data<T>(in->dims(), place); | ||
| const auto& runner = NpuOpRunner("Adds", {*in}, {*out}, {{"value", val}}); | ||
| runner.Run(stream); | ||
| } | ||
|
|
||
| template <typename T> | ||
| void LabelSmoothAddBroadCast(const platform::Place& place, | ||
| const aclrtStream& stream, const Tensor* in1, | ||
| const Tensor* in2, Tensor* out) { | ||
| out->mutable_data<T>(place); | ||
| const auto& runner = NpuOpRunner("AddV2", {*in1, *in2}, {*out}, {}); | ||
| runner.Run(stream); | ||
| } | ||
|
|
||
| template <typename T> | ||
| class LabelSmoothNPUKernel : public framework::OpKernel<T> { | ||
| public: | ||
| void Compute(const framework::ExecutionContext& ctx) const override { | ||
| auto* out_t = ctx.Output<LoDTensor>("Out"); | ||
| auto* in_t = ctx.Input<LoDTensor>("X"); | ||
| auto* dist_t = ctx.Input<Tensor>("PriorDist"); | ||
| auto epsilon = ctx.Attr<float>("epsilon"); | ||
|
|
||
| auto label_dim = in_t->dims()[in_t->dims().size() - 1]; | ||
| auto place = ctx.GetPlace(); | ||
|
|
||
| auto stream = | ||
| ctx.template device_context<paddle::platform::NPUDeviceContext>() | ||
| .stream(); | ||
|
|
||
| if (dist_t) { | ||
| Tensor tmp; | ||
| Tensor dist; | ||
| Tensor tmp2; | ||
| LabelSmoothMuls<T>(place, stream, in_t, (1 - epsilon), &tmp); | ||
| LabelSmoothMuls<T>(place, stream, dist_t, epsilon, &tmp2); | ||
| tmp2.Resize({1, label_dim}); | ||
| LabelSmoothAddBroadCast<T>(place, stream, &tmp, &tmp2, out_t); | ||
| } else { | ||
| Tensor tmp; | ||
| LabelSmoothMuls<T>(place, stream, in_t, (1 - epsilon), &tmp); | ||
| LabelSmoothAdds<T>(place, stream, &tmp, (epsilon / label_dim), out_t); | ||
| } | ||
| } | ||
| }; | ||
|
|
||
| template <typename T> | ||
| class LabelSmoothGradNPUKernel : public framework::OpKernel<T> { | ||
| public: | ||
| void Compute(const framework::ExecutionContext& ctx) const override { | ||
| auto* d_out_t = ctx.Input<framework::Tensor>(framework::GradVarName("Out")); | ||
| auto* d_in_t = ctx.Output<framework::Tensor>(framework::GradVarName("X")); | ||
| auto epsilon = ctx.Attr<float>("epsilon"); | ||
|
|
||
| auto place = ctx.GetPlace(); | ||
|
|
||
| auto stream = | ||
| ctx.template device_context<paddle::platform::NPUDeviceContext>() | ||
| .stream(); | ||
|
|
||
| LabelSmoothMuls<T>(place, stream, d_out_t, 1 - epsilon, d_in_t); | ||
| } | ||
| }; | ||
|
|
||
| } // namespace operators | ||
| } // namespace paddle | ||
|
|
||
| namespace ops = paddle::operators; | ||
| namespace plat = paddle::platform; | ||
|
|
||
| REGISTER_OP_NPU_KERNEL(label_smooth, ops::LabelSmoothNPUKernel<float>, | ||
| ops::LabelSmoothNPUKernel<plat::float16>); | ||
| REGISTER_OP_NPU_KERNEL(label_smooth_grad, ops::LabelSmoothGradNPUKernel<float>, | ||
| ops::LabelSmoothGradNPUKernel<plat::float16>); |
Oops, something went wrong.
0dfc8f7There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
✅Congratulation! Your pull request passed all required CI. You could ask reviewer(s) to approve and merge. 🎉