[NPU] Support op kernel for Fill constant batch size like op #34721
Conversation
|
Thanks for your contribution! |
| limitations under the License. */ | ||
|
|
||
| #include <memory> | ||
| #include <string> |
| out->mutable_data(platform::CPUPlace(), data_type); | ||
| functor(reinterpret_cast<const platform::CPUDeviceContext &>(dev_ctx), | ||
| out, static_cast<T>(value)); | ||
| } else { |
There was a problem hiding this comment.
参考fill_constant_batch_size_like_op.h，这里应该是以下两个条件同时成立的情况下才走CPU
bool cpu_place = force_cpu || ctx.GetPlace() == platform::CPUPlace();
| } else if (str_value == "-inf") { | ||
| value = static_cast<T>(-std::numeric_limits<double>::infinity()); | ||
| } else if (str_value == "nan") { | ||
| value = static_cast<T>(std::numeric_limits<double>::quiet_NaN()); |
There was a problem hiding this comment.
参考fill_constant_batch_size_like_op.h，不支持inf或者nan的输入，可以把这里特殊处理去掉。
| auto odims = out->dims(); | ||
| auto idims = input->dims(); | ||
| int output_dim_idx = ctx.Attr<int>("output_dim_idx"); | ||
| odims[output_dim_idx] = static_cast<int>(idims[0]); |
There was a problem hiding this comment.
能解释一下为什么这里的计算公式与以下不一致呢？
odims[output_dim_idx] = static_cast(in->lod().back().size()) - 1;
There was a problem hiding this comment.
这里要支持Lodtensor是吗，我默认是用的tensor，所以直接取得第0维
There was a problem hiding this comment.
如果是Tensor的情况下，那应该是 odims[output_dim_idx] = idims[input_dim_idx]，参考 batch_size_like.h里面的OpMaker的定义。
output_dim[output_dim_idx] = ctx->GetInputDim("Input")[input_dim_idx];
… fill_constant_batch_size_like_op_npu
3 participants
PR types
Others
PR changes
OPs
Describe
[NPU] Support op kernel for Fill constant batch size like op
op unittest result:
GLOG: