Use header files:

- Create header files for kernel implementation and remove definitions from vision_*.h files.
- Eliminate unnecessary headers and ensure all cpp include their headers.

torchvision/csrc/autocast.h

@@ -1,5 +1,7 @@
 #pragma once
 
+// TODO: Delete this file once none of the methods use it
+
 #if defined(WITH_CUDA) || defined(WITH_HIP)
 #include <ATen/autocast_mode.h>
 #endif

torchvision/csrc/cpu/deform_conv2d_cpu.cpp

@@ -66,13 +66,7 @@
 // modified from
 // https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda.cpp
 
-#include <ATen/ATen.h>
-#include <ATen/TensorUtils.h>
-#include <TH/TH.h>
-
-#include <cmath>
-#include <iostream>
-#include <tuple>
+#include "deform_conv2d_cpu.h"
 
 namespace {
 

torchvision/csrc/cpu/deform_conv2d_cpu.h

@@ -0,0 +1,39 @@
+#pragma once
+
+#include <ATen/ATen.h>
+#include "../macros.h"
+
+VISION_API at::Tensor deform_conv2d_forward_cpu(
+    const at::Tensor& input_param,
+    const at::Tensor& weight_param,
+    const at::Tensor& offset_param,
+    const at::Tensor& mask_param,
+    const at::Tensor& bias_param,
+    int64_t stride_h,
+    int64_t stride_w,
+    int64_t pad_h,
+    int64_t pad_w,
+    int64_t dil_h,
+    int64_t dil_w,
+    int64_t n_weight_grps,
+    int64_t n_offset_grps,
+    bool use_mask);
+
+VISION_API std::
+    tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
+    deform_conv2d_backward_cpu(
+        const at::Tensor& grad_out_param,
+        const at::Tensor& input_param,
+        const at::Tensor& weight_param,
+        const at::Tensor& offset_param,
+        const at::Tensor& mask_param,
+        const at::Tensor& bias_param,
+        int64_t stride_h,
+        int64_t stride_w,
+        int64_t pad_h,
+        int64_t pad_w,
+        int64_t dil_h,
+        int64_t dil_w,
+        int64_t n_weight_grps,
+        int64_t n_offset_grps,
+        bool use_mask);

torchvision/csrc/cpu/vision_cpu.h

@@ -2,40 +2,7 @@
 #include <torch/extension.h>
 #include "../macros.h"
 
-VISION_API at::Tensor deform_conv2d_forward_cpu(
-    const at::Tensor& input_param,
-    const at::Tensor& weight_param,
-    const at::Tensor& offset_param,
-    const at::Tensor& mask_param,
-    const at::Tensor& bias_param,
-    int64_t stride_h,
-    int64_t stride_w,
-    int64_t pad_h,
-    int64_t pad_w,
-    int64_t dil_h,
-    int64_t dil_w,
-    int64_t n_weight_grps,
-    int64_t n_offset_grps,
-    bool use_mask);
-
-VISION_API std::
-    tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
-    deform_conv2d_backward_cpu(
-        const at::Tensor& grad_out_param,
-        const at::Tensor& input_param,
-        const at::Tensor& weight_param,
-        const at::Tensor& offset_param,
-        const at::Tensor& mask_param,
-        const at::Tensor& bias_param,
-        int64_t stride_h,
-        int64_t stride_w,
-        int64_t pad_h,
-        int64_t pad_w,
-        int64_t dil_h,
-        int64_t dil_w,
-        int64_t n_weight_grps,
-        int64_t n_offset_grps,
-        bool use_mask);
+// TODO: Delete this file once all the methods are gone
 
 VISION_API at::Tensor nms_cpu(
     const at::Tensor& dets,

torchvision/csrc/cuda/deform_conv2d_cuda.cu

@@ -67,17 +67,13 @@
 // https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda.cpp
 
 #include <ATen/ATen.h>
-#include <ATen/TensorUtils.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 #include <THC/THCAtomics.cuh>
 
+#include "deform_conv2d_cuda.h"
 #include "cuda_helpers.h"
 
-#include <cmath>
-#include <iostream>
-#include <tuple>
-
 namespace {
 
 const int kMaxParallelImgs = 32;

torchvision/csrc/cuda/deform_conv2d_cuda.h

@@ -0,0 +1,39 @@
+#pragma once
+
+#include <ATen/ATen.h>
+#include "../macros.h"
+
+VISION_API at::Tensor deform_conv2d_forward_cuda(
+    const at::Tensor& input_param,
+    const at::Tensor& weight_param,
+    const at::Tensor& offset_param,
+    const at::Tensor& mask_param,
+    const at::Tensor& bias_param,
+    int64_t stride_h,
+    int64_t stride_w,
+    int64_t pad_h,
+    int64_t pad_w,
+    int64_t dil_h,
+    int64_t dil_w,
+    int64_t n_weight_grps,
+    int64_t n_offset_grps,
+    bool use_mask);
+
+VISION_API std::
+    tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
+    deform_conv2d_backward_cuda(
+        const at::Tensor& grad_out_param,
+        const at::Tensor& input_param,
+        const at::Tensor& weight_param,
+        const at::Tensor& offset_param,
+        const at::Tensor& mask_param,
+        const at::Tensor& bias_param,
+        int64_t stride_h,
+        int64_t stride_w,
+        int64_t pad_h,
+        int64_t pad_w,
+        int64_t dil_h,
+        int64_t dil_w,
+        int64_t n_weight_grps,
+        int64_t n_offset_grps,
+        bool use_mask);

torchvision/csrc/cuda/vision_cuda.h

@@ -2,40 +2,7 @@
 #include <torch/extension.h>
 #include "../macros.h"
 
-VISION_API at::Tensor deform_conv2d_forward_cuda(
-    const at::Tensor& input_param,
-    const at::Tensor& weight_param,
-    const at::Tensor& offset_param,
-    const at::Tensor& mask_param,
-    const at::Tensor& bias_param,
-    int64_t stride_h,
-    int64_t stride_w,
-    int64_t pad_h,
-    int64_t pad_w,
-    int64_t dil_h,
-    int64_t dil_w,
-    int64_t n_weight_grps,
-    int64_t n_offset_grps,
-    bool use_mask);
-
-VISION_API std::
-    tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
-    deform_conv2d_backward_cuda(
-        const at::Tensor& grad_out_param,
-        const at::Tensor& input_param,
-        const at::Tensor& weight_param,
-        const at::Tensor& offset_param,
-        const at::Tensor& mask_param,
-        const at::Tensor& bias_param,
-        int64_t stride_h,
-        int64_t stride_w,
-        int64_t pad_h,
-        int64_t pad_w,
-        int64_t dil_h,
-        int64_t dil_w,
-        int64_t n_weight_grps,
-        int64_t n_offset_grps,
-        bool use_mask);
+// TODO: Delete this file once all the methods are gone
 
 VISION_API at::Tensor nms_cuda(
     const at::Tensor& dets,

torchvision/csrc/deform_conv2d.cpp

@@ -1,14 +1,8 @@
-#pragma once
+#include "deform_conv2d.h"
+#include <torch/extension.h>
 
-#include "cpu/vision_cpu.h"
-
-#ifdef WITH_CUDA
-#include "autocast.h"
-#include "cuda/vision_cuda.h"
-#endif
-#ifdef WITH_HIP
-#include "autocast.h"
-#include "hip/vision_cuda.h"
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+#include <ATen/autocast_mode.h>
 #endif
 
 namespace {

torchvision/csrc/deform_conv2d.h

@@ -0,0 +1,64 @@
+#pragma once
+
+#include "cpu/deform_conv2d_cpu.h"
+
+#ifdef WITH_CUDA
+#include "cuda/deform_conv2d_cuda.h"
+#endif
+#ifdef WITH_HIP
+#include "hip/deform_conv2d_cuda.h"
+#endif
+
+// Autocast Registration
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+at::Tensor deform_conv2d_autocast(
+    const at::Tensor& input,
+    const at::Tensor& weight,
+    const at::Tensor& offset,
+    const at::Tensor& mask,
+    const at::Tensor& bias,
+    int64_t stride_h,
+    int64_t stride_w,
+    int64_t pad_h,
+    int64_t pad_w,
+    int64_t dilation_h,
+    int64_t dilation_w,
+    int64_t groups,
+    int64_t offset_groups,
+    bool use_mask);
+#endif
+
+// Autograd Registration
+at::Tensor deform_conv2d_autograd(
+    const at::Tensor& input,
+    const at::Tensor& weight,
+    const at::Tensor& offset,
+    const at::Tensor& mask,
+    const at::Tensor& bias,
+    int64_t stride_h,
+    int64_t stride_w,
+    int64_t pad_h,
+    int64_t pad_w,
+    int64_t dilation_h,
+    int64_t dilation_w,
+    int64_t groups,
+    int64_t offset_groups,
+    bool use_mask);
+
+std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
+deform_conv2d_backward_autograd(
+    const at::Tensor& grad,
+    const at::Tensor& input,
+    const at::Tensor& weight,
+    const at::Tensor& offset,
+    const at::Tensor& mask,
+    const at::Tensor& bias,
+    int64_t stride_h,
+    int64_t stride_w,
+    int64_t pad_h,
+    int64_t pad_w,
+    int64_t dilation_h,
+    int64_t dilation_w,
+    int64_t groups,
+    int64_t offset_groups,
+    bool use_mask);