|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include <torch/extension.h> |
|
#include <vector> |
|
|
|
std::vector<at::Tensor> dynamicconv_cuda_forward( |
|
at::Tensor input, |
|
at::Tensor filters, |
|
int padding_l); |
|
|
|
std::vector<at::Tensor> dynamicconv_cuda_backward( |
|
at::Tensor gradOutput, |
|
int padding_l, |
|
at::Tensor input, |
|
at::Tensor filters); |
|
|
|
|
|
#define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") |
|
#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") |
|
#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) |
|
|
|
std::vector<at::Tensor> dynamicconv_forward( |
|
at::Tensor input, |
|
at::Tensor filters, |
|
int padding_l) { |
|
|
|
CHECK_INPUT(input); |
|
CHECK_INPUT(filters); |
|
|
|
return dynamicconv_cuda_forward(input, filters, |
|
padding_l); |
|
} |
|
|
|
std::vector<at::Tensor> dynamicconv_backward( |
|
at::Tensor gradOutput, |
|
int padding_l, |
|
at::Tensor input, |
|
at::Tensor filters) { |
|
|
|
CHECK_INPUT(gradOutput); |
|
CHECK_INPUT(input); |
|
CHECK_INPUT(filters); |
|
|
|
return dynamicconv_cuda_backward(gradOutput, padding_l, |
|
input, filters); |
|
} |
|
|
|
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { |
|
m.def("forward", &dynamicconv_forward, "dynamicconv forward (CUDA)"); |
|
m.def("backward", &dynamicconv_backward, "dynamicconv backward (CUDA)"); |
|
} |
|
|