[NPU] Support npu op expand_v2 and expand_v2_grad#34764
[NPU] Support npu op expand_v2 and expand_v2_grad#34764qili93 merged 10 commits intoPaddlePaddle:developfrom
Conversation
|
Thanks for your contribution! |
|
|
||
| #include "paddle/fluid/operators/expand_v2_op.h" | ||
| #include <memory> | ||
| #include <string> |
| namespace paddle { | ||
| namespace operators { | ||
|
|
||
| inline std::vector<int> get_expand_shape_npu( |
There was a problem hiding this comment.
参考 https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/expand_op.h 中 get_expand_times 的写法，直接修改 expand_v2_op.h中get_expand_shape的方法，增加platform::is_npu_place的功能实现即可，不需要重写整个函数。
There was a problem hiding this comment.
已在expand_v2_op.h中添加platform::is_npu_place，expand_v2_op_npu.cc中取消get_expand_shape_npu函数
| "The %uth element of 'shape' for expand_v2_npu op must be " | ||
| "greater than 0, but the value given is %d.", | ||
| i, expand_shape[i])); | ||
| } |
There was a problem hiding this comment.
需要结合TestExpandV2OpRank4进行代码调整。
There was a problem hiding this comment.
expand_v2_op_npu.cc中添加对expand_shape的再构造，考虑三种情况 1) expand_shape比input_dim维度要大的情况 2) expand_shape[i]维度大于0的情况 3) expand_shape[i]等于-1的情况
| # def init_data(self): | ||
| # self.ori_shape = (2, 4, 5, 7) | ||
| # self.shape = (-1, -1, -1, -1) | ||
| # self.expand_times = (1, 1, 1, 1) |
There was a problem hiding this comment.
确认下，在CPU/CUDA代码中，对于self.shape = (-1, -1, -1, -1)的情况时如何处理的？需要看下expand_v2_op.h的CPU计算逻辑，应该最后可以得到一个正值的out_dims的结果，之后看下是否可以通过NPU的ExpandD或者BoradCastD等算子进行实现。
There was a problem hiding this comment.
已添加三个shape中有-1情况的单测用例，单测已通过
|
|
||
| /* 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. */ |
There was a problem hiding this comment.
Plz check the copyright format.
There was a problem hiding this comment.
reformat the copyright done.
|
|
||
| REGISTER_OP_NPU_KERNEL( | ||
| expand_v2_grad, | ||
| ops::ExpandV2NPUGradKernel<paddle::platform::NPUDeviceContext, float>, |
There was a problem hiding this comment.
Seems missing fp16 kernel for expand_grad
There was a problem hiding this comment.
add fp16 kernel support for expand_grad
| for (auto i = 0; i < reduce_ndim; ++i) { | ||
| axes.push_back(i); | ||
| } | ||
| Tensor* tmp_dout = const_cast<Tensor*>(dout); |
There was a problem hiding this comment.
Avoid const_cast by using SharedDataWith, see dropout_op_npu.cc:L61 for details.
4 participants
PR types
New features
PR changes
OPs
Describe
Add NPU kernel for expand_v2 and expand_v2_grad
