[Pten]Refactor the Elementwise_add Kernel #37043
Merged
Conversation
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
|
Thanks for your contribution! |
chenwhql
reviewed
Nov 10, 2021
| int axis = PackTensorsIntoVector<T>(ctx, &ins, &outs); | ||
| LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>( | ||
| cuda_ctx, ins, &outs, axis, AddFunctor<T>()); | ||
| auto* x = ctx.Input<framework::LoDTensor>("X"); |
There was a problem hiding this comment.
这个函数是不是可以移除了，直接复用下面paddle/fluid/operators/elementwise/elementwise_add_op.h中的kernel，在注册时，给cuda注册一份就可以
| @@ -20,6 +20,13 @@ limitations under the License. */ | |||
| #include "paddle/fluid/operators/math/blas.h" | |||
| #include "paddle/fluid/operators/math/math_function.h" | |||
|
|
|||
| #include "paddle/fluid/framework/pten_utils.h" | |||
|
|
|||
| // only can include the headers in paddle/top/api dirs | |||
There was a problem hiding this comment.
注释目录需要更改，现在应该是paddle/pten/include
| @@ -29,6 +29,11 @@ limitations under the License. */ | |||
| #include "paddle/fluid/platform/gpu_info.h" | |||
| #include "paddle/fluid/platform/transform.h" | |||
|
|
|||
| // only can include the headers in paddle/top/api dirs | |||
| (*post) *= x_dims[i]; | ||
| } | ||
| pten::general::get_mid_dims(x_dims, y_dims, axis, pre, n, post, | ||
| is_run_common_broadcast); | ||
| } | ||
|
|
||
| inline int GetElementwiseIndex(const int *x_dims_array, const int max_dim, |
| } | ||
| std::vector<const pten::DenseTensor *> pt_inputs; | ||
| std::vector<pten::DenseTensor *> pt_outputs; | ||
| // *_tmp for cache DenseTensor |
There was a problem hiding this comment.
可以加TODO注释说明是DenseTensor暂不支持拷贝构造导致需要这样写？后续会优化？
| @@ -3,3 +3,5 @@ cc_library(linalg_cpu SRCS linalg.cc DEPS dense_tensor kernel_context kernel_fac | |||
| cc_library(creation_cpu SRCS creation.cc DEPS dense_tensor kernel_context kernel_factory eigen_function) | |||
| cc_library(utils_cpu SRCS utils.cc DEPS dense_tensor kernel_context kernel_factory memory convert_utils) | |||
| cc_library(manipulation_cpu SRCS manipulation.cc DEPS dense_tensor kernel_context kernel_factory utils_cpu unary) | |||
| cc_library(nn_cpu SRCS nn.cc DEPS dense_tensor kernel_context kernel_factory blas eigen_function) | |||
| add_subdirectory(funcs) | |||
There was a problem hiding this comment.
可以将原来的functions改为funcs，是不是不在每个目录下新建funcs比较好
| @@ -4,10 +4,12 @@ if(WITH_GPU) | |||
| nv_library(creation_cuda SRCS creation.cu DEPS eigen_function dense_tensor kernel_context kernel_factory) | |||
| nv_library(utils_cuda SRCS utils.cu DEPS dense_tensor kernel_context kernel_factory memory convert_utils) | |||
| nv_library(manipulation_cuda SRCS manipulation.cu DEPS dense_tensor kernel_context kernel_factory utils_cuda unary) | |||
| elseif(WITH_ROCM) | |||
| nv_library(nn_cuda SRCS nn.cu DEPS dense_tensor kernel_context kernel_factory eigen_function) | |||
| elseif(WITH_ROCM) | |||
| int64_t post_; | ||
| }; | ||
|
|
||
| // #if defined(__NVCC__) || defined(__HIPCC__) |
| namespace general { | ||
|
|
||
| using DDim = paddle::framework::DDim; | ||
| using CPUContext = paddle::platform::CPUDeviceContext; |
There was a problem hiding this comment.
这里是cpu和gpu的公共方法，所以放在里general下边
|
|
||
| #pragma once | ||
|
|
||
| #include "paddle/pten/kernels/cuda/funcs/elementwise/elementwise_broadcast.cu.h" |
There was a problem hiding this comment.
同上，要不在functions下面新增cpu和cuda目录吧，不在kernel层的设备目录里新增子目录了，后续这里也还要调整，另外如果都在cuda目录下，是不是文件后缀中的cu也可以去掉了
zyfncg
reviewed
Nov 10, 2021
paddle/pten/api/lib/nn.cc
Outdated
| namespace paddle { | ||
| namespace experimental { | ||
|
|
||
| Tensor elementwise_add(const Tensor& x, const Tensor& y, int axis) { |
chenwhql
approved these changes
Nov 11, 2021
| @@ -2020,7 +1856,8 @@ void FusedElemwiseAndActComputeWithBroadcast( | |||
| axis = (y_dim.size() == 0) ? x_dim.size() : axis; | |||
|
|
|||
| int pre, n, post, is_run_common_broadcast; | |||
| get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post, &is_run_common_broadcast); | |||
| pten::general::get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post, | |||
There was a problem hiding this comment.
函数命名建议按照code style统一，采用驼峰式命名，这里原先就不太规范，可在后续PR更改
Avin0323
approved these changes
Nov 11, 2021
raindrops2sea
approved these changes
Nov 11, 2021
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR types
Others
PR changes
OPs
Describe
相关背景
需要迁移elementwise系列kernel到新的Tensor计算库下，其中elementwise_add是迁移的第一个该系列kernel。该PR主要迁移了elementwise_add计算逻辑以及大部分elementwise系列kernel依赖的基本组件，涉及到了较多的文件，构成了后续其他elementwise迁移的基础。
迁移主要逻辑
1，elementwise基础组件迁移
2，elementwise_add kernel迁移