[Pten] Add reduce mean kernel, replace with mean API #37559
Conversation
|
Thanks for your contribution! |
|
|
||
| // call new kernel | ||
| pten::Mean<T>(dev_ctx, *pt_x.get(), pt_out.get()); | ||
| y.device(place) = X.mean(); |
There was a problem hiding this comment.
注意check下kernel注册的写法也要恢复到和原先一样
| @@ -0,0 +1,962 @@ | |||
| // Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. | |||
There was a problem hiding this comment.
已修改命名为 reduce_cuda_impl.h
| PD_DLL_DECL Tensor mean(const Tensor& x, | ||
| const std::vector<int64_t>& axis, | ||
| bool keep_dim); |
There was a problem hiding this comment.
后续PR 再加默认值的接口
| PD_DLL_DECL Tensor sum(const Tensor& x, | ||
| const std::vector<int64_t>& axis, | ||
| DataType dtype, | ||
| bool keep_dim); | ||
|
|
There was a problem hiding this comment.
后续PR 再加默认值的接口
paddle/pten/api/lib/math.cc
Outdated
| kernel_context.EmplaceBackAttr(dense_x->dtype()); | ||
| kernel_context.EmplaceBackAttr(out_dtype); | ||
|
|
||
| // 4. InferShape |
| if (x.dtype() == pten::DataType::BOOL || x.dtype() == pten::DataType::INT32 || | ||
| x.dtype() == pten::DataType::INT64) { | ||
| dtype = pten::DataType::INT64; | ||
| } |
There was a problem hiding this comment.
在InferMeta里看到有类似的执行逻辑，这里的逻辑能否仅放在InferMeta或者kernel中处理？如果要代码自动生成的话这类情况可能还需要单独配置，会增加配置项的复杂性
There was a problem hiding this comment.
后续PR 优化这里
| const std::vector<int64_t>& axis, | ||
| bool keep_dim) { | ||
| bool reduce_all = true; | ||
| std::set<int64_t> dims_set(axis.begin(), axis.end()); |
There was a problem hiding this comment.
原来的kernel 就是用的set，为保持一致，没用unordered_set
| void Reduce(const CUDAContext& dev_ctx, | ||
| const DenseTensor& x, | ||
| bool reduce_all, | ||
| std::vector<int64_t> dims, |
There was a problem hiding this comment.
why not const vector<int64&>& or DDim ?
There was a problem hiding this comment.
改为了const vector<int64&>&
| // eg: x_dim = [2, 4, 6] origin_reduce_dims = [0, 1] | ||
| // --SetReduceDim--> x_dim = [8,6], reduce_dim = [0], left_dim = [1] | ||
| void SetReduceDim() { | ||
| std::set<int64_t> reduce_set; |
There was a problem hiding this comment.
原来的kernel 就是用的set，为保持一致，没用unordered_set
| } | ||
|
|
||
| std::vector<int64_t> reduce_dim_temp(reduce_set.begin(), reduce_set.end()); | ||
| std::sort(reduce_dim_temp.begin(), reduce_dim_temp.end()); |
There was a problem hiding this comment.
应该可以，后续再和原作者确认下，提PR修复
| void Reduce(const DeviceContext& dev_ctx, | ||
| const DenseTensor& x, | ||
| bool reduce_all, | ||
| std::vector<int64_t> dims, |
| pten::DenseTensor* out, | ||
| const std::vector<int64_t>& axis) { |
There was a problem hiding this comment.
该函数参数顺序是与原来定义保持一致的
MingMingShangTian
changed the title
…7559) * add pten reduce kernel * add reduce_sum kernel * update attribute args and order * make out dtype undefined * fix empty input error * merge develop branch * rename sum as reduce function * rename sum as reduce function * fix reducekernelImpl args error * add reduce cuda kernel * modify dims type to const & * remove unsed log * fix reduce_all out eigen function error * remove unused codes * add the missing sum api define and testcase * merge develop branch * fix sum test axis value error * replace pten mean kernel with reduce_mean * revcover meam cuda to original implement
PR types
New features
PR changes
Others
Describe