gather, scatter with gpu support, passed python test #4483
Conversation
gather scatter gpu
paddle/operators/gather.cu.h
Outdated
| * return: output tensor | ||
| */ | ||
| template <typename T> | ||
| void GPUTGather(const Place& place, const Tensor* src, const Tensor* index, |
There was a problem hiding this comment.
GPUTGather name is not clear, and can we merge GPUGather GPUTGather together?
And we'd better change the parameter Place of GPUTGather to parameter CUDADeviceContext. CUDADeviceContext has a CUDA stream, we should launch a CUDA kernel on specific CUDA stream.
There was a problem hiding this comment.
Great idea. Working on it now...
They were using float64 for FP32 kernel before.
|
Please merge latest develop branch first |
paddle/operators/gather.h
Outdated
| void Gather(const platform::Place& place, const paddle::framework::Tensor* src, | ||
| const paddle::framework::Tensor* index, | ||
| paddle::framework::Tensor* output) { | ||
| void CPUGather(const platform::Place& place, |
There was a problem hiding this comment.
It's better unify the parameter of CPUGather and GPUGather, the first parameter should be DeviceContext
void CPUGather(const platform::DeviceContext& ctx...)
paddle/operators/scatter.h
Outdated
| */ | ||
| template <typename T> | ||
| void ScatterUpdate(const platform::Place& place, | ||
| void ScatterAssign(const platform::Place& place, |
paddle/operators/gather.cu.h
Outdated
| * return: output tensor | ||
| */ | ||
| template <typename T> | ||
| void GPUGather(const platform::DeviceContext& ctx, const Tensor* src, |
There was a problem hiding this comment.
If the parameter is input, we should take const T&;
If the parameter is output, we should take T*;
If the parameter is both input and output, we should take T*;
Please refer to https://google.github.io/styleguide/cppguide.html#Reference_Arguments
| template <typename T> | ||
| void GPUGather(const platform::DeviceContext& ctx, const Tensor* src, | ||
| const Tensor* index, Tensor* output) { | ||
| // PADDLE_ENFORCE(platform::is_gpu_place(place)); |
There was a problem hiding this comment.
PADDLE_ENFORCE(platform::is_gpu_place(place));
paddle/operators/scatter.cu.h
Outdated
| * return: output tensor | ||
| */ | ||
| template <typename T> | ||
| void GPUScatterAssign(const platform::DeviceContext& ctx, |
There was a problem hiding this comment.
The same as CPUGather, input should be const T&
3 participants
No description provided.