Per file C++ Operator registration (#3135)

Summary: * Moving deform_conv2d op registration. * Moving nms op registration. * Moving new_empty_tensor op registration. * Moving ps_roi_align op registration. * Moving ps_roi_pool op registration. * Moving roi_align op registration. * Moving roi_pool op registration. * Restoring headers for forward/backward and fixing styles. * Restoring the test hack on windows. * Stricter header inclusion. Reviewed By: fmassa Differential Revision: D25460675 fbshipit-source-id: 8b7c65320d41c9c0f43815d156957c1e40aef106
pytorch · Dec 10, 2020 · b5b2a2c · b5b2a2c
1 parent 4774807
commit b5b2a2c
Show file tree

Hide file tree

Showing 40 changed files with 306 additions and 804 deletions.
diff --git a/test/tracing/frcnn/test_frcnn_tracing.cpp b/test/tracing/frcnn/test_frcnn_tracing.cpp
@@ -1,13 +1,13 @@
 #include <ATen/ATen.h>
 #include <torch/script.h>
 #include <torch/torch.h>
-#include <torchvision/roi_align.h>
 #include <torchvision/nms.h>
+#include <torchvision/roi_align.h>
 
 #ifdef _WIN32
 // Windows only
 // This is necessary until operators are automatically registered on include
-static auto _nms = &vision::ops::nms_cpu;
+static auto _nms = &vision::ops::nms;
 #endif
 
 int main() {

diff --git a/torchvision/csrc/cpu/deform_conv2d_kernel.cpp b/torchvision/csrc/cpu/deform_conv2d_kernel.cpp
@@ -66,7 +66,8 @@
 // modified from
 // https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda.cpp
 
-#include "deform_conv2d_kernel.h"
+#include <ATen/ATen.h>
+#include <torch/library.h>
 
 namespace vision {
 namespace ops {
@@ -852,9 +853,7 @@ at::Tensor backward_gradient_parameters(
   return grad_weight;
 }
 
-} // namespace
-
-at::Tensor deform_conv2d_forward_cpu(
+at::Tensor deform_conv2d_forward_kernel(
     const at::Tensor& input,
     const at::Tensor& weight,
     const at::Tensor& offset,
@@ -1070,7 +1069,7 @@ at::Tensor deform_conv2d_forward_cpu(
 }
 
 std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
-deform_conv2d_backward_cpu(
+deform_conv2d_backward_kernel(
     const at::Tensor& grad_out,
     const at::Tensor& input,
     const at::Tensor& weight,
@@ -1141,5 +1140,12 @@ deform_conv2d_backward_cpu(
       grad_input, grad_weight, grad_offset, grad_mask, grad_bias);
 }
 
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, CPU, m) {
+  m.impl("deform_conv2d", deform_conv2d_forward_kernel);
+  m.impl("_deform_conv2d_backward", deform_conv2d_backward_kernel);
+}
+
 } // namespace ops
 } // namespace vision
diff --git a/torchvision/csrc/cpu/deform_conv2d_kernel.h b/torchvision/csrc/cpu/deform_conv2d_kernel.h
diff --git a/torchvision/csrc/cpu/nms_kernel.cpp b/torchvision/csrc/cpu/nms_kernel.cpp
@@ -1,4 +1,5 @@
-#include "nms_kernel.h"
+#include <ATen/ATen.h>
+#include <torch/library.h>
 
 namespace vision {
 namespace ops {
@@ -74,9 +75,7 @@ at::Tensor nms_kernel_impl(
   return keep_t.narrow(/*dim=*/0, /*start=*/0, /*length=*/num_to_keep);
 }
 
-} // namespace
-
-at::Tensor nms_cpu(
+at::Tensor nms_kernel(
     const at::Tensor& dets,
     const at::Tensor& scores,
     double iou_threshold) {
@@ -101,11 +100,17 @@ at::Tensor nms_cpu(
 
   auto result = at::empty({0}, dets.options());
 
-  AT_DISPATCH_FLOATING_TYPES(dets.scalar_type(), "nms_cpu", [&] {
+  AT_DISPATCH_FLOATING_TYPES(dets.scalar_type(), "nms_kernel", [&] {
     result = nms_kernel_impl<scalar_t>(dets, scores, iou_threshold);
   });
   return result;
 }
 
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, CPU, m) {
+  m.impl("nms", nms_kernel);
+}
+
 } // namespace ops
 } // namespace vision
diff --git a/torchvision/csrc/cpu/nms_kernel.h b/torchvision/csrc/cpu/nms_kernel.h
diff --git a/torchvision/csrc/cpu/ps_roi_align_kernel.cpp b/torchvision/csrc/cpu/ps_roi_align_kernel.cpp
@@ -1,4 +1,5 @@
-#include "ps_roi_align_kernel.h"
+#include <ATen/ATen.h>
+#include <torch/library.h>
 
 namespace vision {
 namespace ops {
@@ -301,9 +302,7 @@ void ps_roi_align_backward_kernel_impl(
   }
 }
 
-} // namespace
-
-std::tuple<at::Tensor, at::Tensor> ps_roi_align_forward_cpu(
+std::tuple<at::Tensor, at::Tensor> ps_roi_align_forward_kernel(
     const at::Tensor& input,
     const at::Tensor& rois,
     double spatial_scale,
@@ -318,7 +317,7 @@ std::tuple<at::Tensor, at::Tensor> ps_roi_align_forward_cpu(
 
   at::TensorArg input_t{input, "input", 1}, rois_t{rois, "rois", 2};
 
-  at::CheckedFrom c = "ps_roi_align_forward_cpu";
+  at::CheckedFrom c = "ps_roi_align_forward_kernel";
   at::checkAllSameType(c, {input_t, rois_t});
 
   int num_rois = rois.size(0);
@@ -343,7 +342,7 @@ std::tuple<at::Tensor, at::Tensor> ps_roi_align_forward_cpu(
 
   auto input_ = input.contiguous(), rois_ = rois.contiguous();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
-      input.scalar_type(), "ps_roi_align_forward_cpu", [&] {
+      input.scalar_type(), "ps_roi_align_forward_kernel", [&] {
         ps_roi_align_forward_kernel_impl<scalar_t>(
             output_size,
             input_.data_ptr<scalar_t>(),
@@ -362,7 +361,7 @@ std::tuple<at::Tensor, at::Tensor> ps_roi_align_forward_cpu(
   return std::make_tuple(output, channel_mapping);
 }
 
-at::Tensor ps_roi_align_backward_cpu(
+at::Tensor ps_roi_align_backward_kernel(
     const at::Tensor& grad,
     const at::Tensor& rois,
     const at::Tensor& channel_mapping,
@@ -384,7 +383,7 @@ at::Tensor ps_roi_align_backward_cpu(
   at::TensorArg grad_t{grad, "grad", 1}, rois_t{rois, "rois", 2},
       channel_mapping_t{channel_mapping, "channel_mapping", 3};
 
-  at::CheckedFrom c = "ps_roi_align_backward_cpu";
+  at::CheckedFrom c = "ps_roi_align_backward_kernel";
   at::checkAllSameType(c, {grad_t, rois_t});
 
   auto num_rois = rois.size(0);
@@ -400,7 +399,7 @@ at::Tensor ps_roi_align_backward_cpu(
 
   auto grad_ = grad.contiguous(), rois_ = rois.contiguous();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
-      grad.scalar_type(), "ps_roi_align_backward_cpu", [&] {
+      grad.scalar_type(), "ps_roi_align_backward_kernel", [&] {
         ps_roi_align_backward_kernel_impl<scalar_t>(
             grad.numel(),
             grad_.data_ptr<scalar_t>(),
@@ -420,5 +419,12 @@ at::Tensor ps_roi_align_backward_cpu(
   return grad_input;
 }
 
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, CPU, m) {
+  m.impl("ps_roi_align", ps_roi_align_forward_kernel);
+  m.impl("_ps_roi_align_backward", ps_roi_align_backward_kernel);
+}
+
 } // namespace ops
 } // namespace vision
diff --git a/torchvision/csrc/cpu/ps_roi_align_kernel.h b/torchvision/csrc/cpu/ps_roi_align_kernel.h
diff --git a/torchvision/csrc/cpu/ps_roi_pool_kernel.cpp b/torchvision/csrc/cpu/ps_roi_pool_kernel.cpp
@@ -1,4 +1,5 @@
-#include "ps_roi_pool_kernel.h"
+#include <ATen/ATen.h>
+#include <torch/library.h>
 
 namespace vision {
 namespace ops {
@@ -145,9 +146,7 @@ void ps_roi_pool_backward_kernel_impl(
   }
 }
 
-} // namespace
-
-std::tuple<at::Tensor, at::Tensor> ps_roi_pool_forward_cpu(
+std::tuple<at::Tensor, at::Tensor> ps_roi_pool_forward_kernel(
     const at::Tensor& input,
     const at::Tensor& rois,
     double spatial_scale,
@@ -161,7 +160,7 @@ std::tuple<at::Tensor, at::Tensor> ps_roi_pool_forward_cpu(
 
   at::TensorArg input_t{input, "input", 1}, rois_t{rois, "rois", 2};
 
-  at::CheckedFrom c = "ps_roi_pool_forward_cpu";
+  at::CheckedFrom c = "ps_roi_pool_forward_kernel";
   at::checkAllSameType(c, {input_t, rois_t});
 
   int num_rois = rois.size(0);
@@ -186,7 +185,7 @@ std::tuple<at::Tensor, at::Tensor> ps_roi_pool_forward_cpu(
 
   auto input_ = input.contiguous(), rois_ = rois.contiguous();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
-      input.scalar_type(), "ps_roi_pool_forward_cpu", [&] {
+      input.scalar_type(), "ps_roi_pool_forward_kernel", [&] {
         ps_roi_pool_forward_kernel_impl<scalar_t>(
             input_.data_ptr<scalar_t>(),
             spatial_scale,
@@ -204,7 +203,7 @@ std::tuple<at::Tensor, at::Tensor> ps_roi_pool_forward_cpu(
   return std::make_tuple(output, channel_mapping);
 }
 
-at::Tensor ps_roi_pool_backward_cpu(
+at::Tensor ps_roi_pool_backward_kernel(
     const at::Tensor& grad,
     const at::Tensor& rois,
     const at::Tensor& channel_mapping,
@@ -225,7 +224,7 @@ at::Tensor ps_roi_pool_backward_cpu(
   at::TensorArg grad_t{grad, "grad", 1}, rois_t{rois, "rois", 2},
       channel_mapping_t{channel_mapping, "channel_mapping", 3};
 
-  at::CheckedFrom c = "ps_roi_pool_backward_cpu";
+  at::CheckedFrom c = "ps_roi_pool_backward_kernel";
   at::checkAllSameType(c, {grad_t, rois_t});
 
   auto num_rois = rois.size(0);
@@ -241,7 +240,7 @@ at::Tensor ps_roi_pool_backward_cpu(
 
   auto grad_ = grad.contiguous(), rois_ = rois.contiguous();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
-      grad.scalar_type(), "ps_roi_pool_backward_cpu", [&] {
+      grad.scalar_type(), "ps_roi_pool_backward_kernel", [&] {
         ps_roi_pool_backward_kernel_impl<scalar_t>(
             grad_.data_ptr<scalar_t>(),
             channel_mapping.data_ptr<int>(),
@@ -259,5 +258,12 @@ at::Tensor ps_roi_pool_backward_cpu(
   return grad_input;
 }
 
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, CPU, m) {
+  m.impl("ps_roi_pool", ps_roi_pool_forward_kernel);
+  m.impl("_ps_roi_pool_backward", ps_roi_pool_backward_kernel);
+}
+
 } // namespace ops
 } // namespace vision
diff --git a/torchvision/csrc/cpu/ps_roi_pool_kernel.h b/torchvision/csrc/cpu/ps_roi_pool_kernel.h