Restructuring C++ project: (#3146)

Summary:
* Reduce unnecessary header inclusions in models and io.

* Move autocast to separate folder and hide autograd implementation in an anonymous namespace.

* Moving files in subfolders.

Reviewed By: fmassa

Differential Revision: D25461523

fbshipit-source-id: 756eeb6848aacaa474de4825ed4c1045d17e2cea

torchvision/csrc/deform_conv2d.cpp → torchvision/csrc/ops/deform_conv2d.cpp

@@ -3,10 +3,6 @@
 #include <torch/autograd.h>
 #include <torch/types.h>
 
-#if defined(WITH_CUDA) || defined(WITH_HIP)
-#include <ATen/autocast_mode.h>
-#endif
-
 namespace vision {
 namespace ops {
 
@@ -45,46 +41,6 @@ at::Tensor deform_conv2d(
       use_mask);
 }
 
-#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) {
-  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
-  return deform_conv2d(
-             at::autocast::cached_cast(at::kFloat, input),
-             at::autocast::cached_cast(at::kFloat, weight),
-             at::autocast::cached_cast(at::kFloat, offset),
-             at::autocast::cached_cast(at::kFloat, mask),
-             at::autocast::cached_cast(at::kFloat, bias),
-             stride_h,
-             stride_w,
-             pad_h,
-             pad_w,
-             dilation_h,
-             dilation_w,
-             groups,
-             offset_groups,
-             use_mask)
-      .to(input.scalar_type());
-}
-
-TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
-  m.impl("deform_conv2d", deform_conv2d_autocast);
-}
-#endif
-
 std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
 _deform_conv2d_backward(
     const at::Tensor& grad,
@@ -307,8 +263,6 @@ class DeformConv2dBackwardFunction
   }
 };
 
-} // namespace
-
 at::Tensor deform_conv2d_autograd(
     const at::Tensor& input,
     const at::Tensor& weight,
@@ -378,6 +332,8 @@ deform_conv2d_backward_autograd(
   return std::make_tuple(result[0], result[1], result[2], result[3], result[4]);
 }
 
+} // namespace
+
 TORCH_LIBRARY_IMPL(torchvision, Autograd, m) {
   m.impl("deform_conv2d", deform_conv2d_autograd);
   m.impl("_deform_conv2d_backward", deform_conv2d_backward_autograd);

torchvision/csrc/deform_conv2d.h → torchvision/csrc/ops/deform_conv2d.h

 #pragma once
 
 #include <ATen/ATen.h>
-#include "macros.h"
+#include "../macros.h"
 
 namespace vision {
 namespace ops {

torchvision/csrc/new_empty_tensor_op.h → torchvision/csrc/ops/new_empty_tensor_op.h

 #pragma once
 
 #include <ATen/ATen.h>
-#include "macros.h"
+#include "../macros.h"
 
 namespace vision {
 namespace ops {

torchvision/csrc/nms.cpp → torchvision/csrc/ops/nms.cpp

@@ -3,10 +3,6 @@
 #include <torch/autograd.h>
 #include <torch/types.h>
 
-#if defined(WITH_CUDA) || defined(WITH_HIP)
-#include <ATen/autocast_mode.h>
-#endif
-
 namespace vision {
 namespace ops {
 
@@ -20,23 +16,6 @@ at::Tensor nms(
   return op.call(dets, scores, iou_threshold);
 }
 
-#if defined(WITH_CUDA) || defined(WITH_HIP)
-at::Tensor nms_autocast(
-    const at::Tensor& dets,
-    const at::Tensor& scores,
-    double iou_threshold) {
-  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
-  return nms(
-      at::autocast::cached_cast(at::kFloat, dets),
-      at::autocast::cached_cast(at::kFloat, scores),
-      iou_threshold);
-}
-
-TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
-  m.impl("nms", nms_autocast);
-}
-#endif
-
 TORCH_LIBRARY_FRAGMENT(torchvision, m) {
   m.def("nms(Tensor dets, Tensor scores, float iou_threshold) -> Tensor");
 }

torchvision/csrc/nms.h → torchvision/csrc/ops/nms.h

 #pragma once
 
 #include <ATen/ATen.h>
-#include "macros.h"
+#include "../macros.h"
 
 namespace vision {
 namespace ops {

torchvision/csrc/ps_roi_align.cpp → torchvision/csrc/ops/ps_roi_align.cpp

@@ -3,10 +3,6 @@
 #include <torch/autograd.h>
 #include <torch/types.h>
 
-#if defined(WITH_CUDA) || defined(WITH_HIP)
-#include <ATen/autocast_mode.h>
-#endif
-
 namespace vision {
 namespace ops {
 
@@ -24,33 +20,6 @@ std::tuple<at::Tensor, at::Tensor> ps_roi_align(
       input, rois, spatial_scale, pooled_height, pooled_width, sampling_ratio);
 }
 
-#if defined(WITH_CUDA) || defined(WITH_HIP)
-std::tuple<at::Tensor, at::Tensor> ps_roi_align_autocast(
-    const at::Tensor& input,
-    const at::Tensor& rois,
-    double spatial_scale,
-    int64_t pooled_height,
-    int64_t pooled_width,
-    int64_t sampling_ratio) {
-  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
-  auto result = ps_roi_align(
-      at::autocast::cached_cast(at::kFloat, input),
-      at::autocast::cached_cast(at::kFloat, rois),
-      spatial_scale,
-      pooled_height,
-      pooled_width,
-      sampling_ratio);
-
-  return std::make_tuple(
-      std::get<0>(result).to(input.scalar_type()),
-      std::get<1>(result).to(input.scalar_type()));
-}
-
-TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
-  m.impl("ps_roi_align", ps_roi_align_autocast);
-}
-#endif
-
 at::Tensor _ps_roi_align_backward(
     const at::Tensor& grad,
     const at::Tensor& rois,
@@ -194,8 +163,6 @@ class PSROIAlignBackwardFunction
   }
 };
 
-} // namespace
-
 std::tuple<at::Tensor, at::Tensor> ps_roi_align_autograd(
     const at::Tensor& input,
     const at::Tensor& rois,
@@ -235,6 +202,8 @@ at::Tensor ps_roi_align_backward_autograd(
       width)[0];
 }
 
+} // namespace
+
 TORCH_LIBRARY_IMPL(torchvision, Autograd, m) {
   m.impl("ps_roi_align", ps_roi_align_autograd);
   m.impl("_ps_roi_align_backward", ps_roi_align_backward_autograd);

torchvision/csrc/ps_roi_align.h → torchvision/csrc/ops/ps_roi_align.h

 #pragma once
 
 #include <ATen/ATen.h>
-#include "macros.h"
+#include "../macros.h"
 
 namespace vision {
 namespace ops {

torchvision/csrc/ps_roi_pool.cpp → torchvision/csrc/ops/ps_roi_pool.cpp

@@ -3,10 +3,6 @@
 #include <torch/autograd.h>
 #include <torch/types.h>
 
-#if defined(WITH_CUDA) || defined(WITH_HIP)
-#include <ATen/autocast_mode.h>
-#endif
-
 namespace vision {
 namespace ops {
 
@@ -22,31 +18,6 @@ std::tuple<at::Tensor, at::Tensor> ps_roi_pool(
   return op.call(input, rois, spatial_scale, pooled_height, pooled_width);
 }
 
-#if defined(WITH_CUDA) || defined(WITH_HIP)
-std::tuple<at::Tensor, at::Tensor> ps_roi_pool_autocast(
-    const at::Tensor& input,
-    const at::Tensor& rois,
-    double spatial_scale,
-    int64_t pooled_height,
-    int64_t pooled_width) {
-  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
-  auto result = ps_roi_pool(
-      at::autocast::cached_cast(at::kFloat, input),
-      at::autocast::cached_cast(at::kFloat, rois),
-      spatial_scale,
-      pooled_height,
-      pooled_width);
-
-  return std::make_tuple(
-      std::get<0>(result).to(input.scalar_type()),
-      std::get<1>(result).to(input.scalar_type()));
-}
-
-TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
-  m.impl("ps_roi_pool", ps_roi_pool_autocast);
-}
-#endif
-
 at::Tensor _ps_roi_pool_backward(
     const at::Tensor& grad,
     const at::Tensor& rois,
@@ -176,8 +147,6 @@ class PSROIPoolBackwardFunction
   }
 };
 
-} // namespace
-
 std::tuple<at::Tensor, at::Tensor> ps_roi_pool_autograd(
     const at::Tensor& input,
     const at::Tensor& rois,
@@ -214,6 +183,8 @@ at::Tensor ps_roi_pool_backward_autograd(
       width)[0];
 }
 
+} // namespace
+
 TORCH_LIBRARY_IMPL(torchvision, Autograd, m) {
   m.impl("ps_roi_pool", ps_roi_pool_autograd);
   m.impl("_ps_roi_pool_backward", ps_roi_pool_backward_autograd);

torchvision/csrc/ps_roi_pool.h → torchvision/csrc/ops/ps_roi_pool.h

 #pragma once
 
 #include <ATen/ATen.h>
-#include "macros.h"
+#include "../macros.h"
 
 namespace vision {
 namespace ops {