[Refactor] Refactor the directory of csrc

[zhouzaida]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Refactor the directory mmcv/ops/csrc to flexibly support more backends.

Follow commands have been successfully tested

MMCV_WITH_OPS=1 pip install -e .
pytest tests
MMCV_WITH_OPS=1 MMCV_WITH_ORT=1 pip install -e .
pytest tests
MMCV_WITH_OPS=1 MMCV_WITH_TRT=1 pip install -e .
pytest tests

Modification

• mmcv/ops/csrc and setup.py are modified
• mmcv/ops/csrc/README.md is added for explaining the structure of csrc and how to add new ops
• docs/compatibility.md and docs_zh_CN/compatibility.md are added for recording compatibility

.
├── common
│   ├── box_iou_rotated_utils.hpp
│   ├── parrots_cpp_helper.hpp
│   ├── parrots_cuda_helper.hpp
│   ├── pytorch_cpp_helper.hpp
│   └── pytorch_cuda_helper.hpp
│   ├── cuda
│   │   ├── bbox_overlaps_cuda_kernel.cuh
│   │   ├── border_align_cuda_kernel.cuh
│   │   ├── box_iou_rotated_cuda.cuh
│   │   ├── carafe_cuda_kernel.cuh
│   │   ├── carafe_naive_cuda_kernel.cuh
│   │   ├── cc_attention_cuda_kernel.cuh
│   │   ├── common_cuda_helper.hpp
│   │   ├── deform_conv_cuda_kernel.cuh
│   │   ├── deform_roi_pool_cuda_kernel.cuh
│   │   ├── masked_conv2d_cuda_kernel.cuh
│   │   ├── modulated_deform_conv_cuda_kernel.cuh
│   │   ├── ms_deform_attn_cuda_kernel.cuh
│   │   ├── nms_cuda_kernel.cuh
│   │   ├── nms_rotated_cuda.cuh
│   │   ├── parrots_cudawarpfunction.cuh
│   │   ├── psamask_cuda_kernel.cuh
│   │   ├── roi_align_cuda_kernel.cuh
│   │   ├── roi_align_rotated_cuda_kernel.cuh
│   │   ├── roi_pool_cuda_kernel.cuh
│   │   ├── sigmoid_focal_loss_cuda_kernel.cuh
│   │   ├── softmax_focal_loss_cuda_kernel.cuh
│   │   ├── sync_bn_cuda_kernel.cuh
│   │   └── tin_shift_cuda_kernel.cuh
├── onnxruntime
│   ├── corner_pool.h
│   ├── grid_sample.h
│   ├── nms.h
│   ├── onnxruntime_register.h
│   ├── onnxruntime_session_options_config_keys.h
│   ├── ort_mmcv_utils.h
│   ├── reduce_ops.h
│   ├── roi_align.h
│   ├── roi_align_rotated.h
│   └── soft_nms.h
│   ├── cpu
│   │   ├── corner_pool.cpp
│   │   ├── gridSample.cpp
│   │   ├── nms.cpp
│   │   ├── onnxruntime_register.cpp
│   │   ├── reduce_ops.cpp
│   │   ├── roi_align.cpp
│   │   ├── roi_align_rotated.cpp
│   │   └── soft_nms.cpp
├── parrots
│   ├── bbox_overlaps.cpp
│   ├── bbox_overlaps_parrots.cpp
│   ├── bbox_overlaps_pytorch.h
│   ├── border_align.cpp
│   ├── border_align_parrots.cpp
│   ├── border_align_pytorch.h
│   ├── box_iou_rotated.cpp
│   ├── box_iou_rotated_cpu.cpp
│   ├── box_iou_rotated_parrots.cpp
│   ├── box_iou_rotated_pytorch.h
│   ├── carafe.cpp
│   ├── carafe_naive.cpp
│   ├── carafe_naive_parrots.cpp
│   ├── carafe_naive_pytorch.h
│   ├── carafe_parrots.cpp
│   ├── carafe_pytorch.h
│   ├── cc_attention.cpp
│   ├── cc_attention_parrots.cpp
│   ├── cc_attention_pytorch.h
│   ├── corner_pool.cpp
│   ├── corner_pool_parrots.cpp
│   ├── corner_pool_pytorch.h
│   ├── deform_conv.cpp
│   ├── deform_conv_parrots.cpp
│   ├── deform_conv_pytorch.h
│   ├── deform_roi_pool.cpp
│   ├── deform_roi_pool_parrots.cpp
│   ├── deform_roi_pool_pytorch.h
│   ├── focal_loss.cpp
│   ├── focal_loss_parrots.cpp
│   ├── focal_loss_pytorch.h
│   ├── fused_bias_leakyrelu.cpp
│   ├── fused_bias_parrots.cpp
│   ├── masked_conv2d.cpp
│   ├── masked_conv2d_parrots.cpp
│   ├── masked_conv2d_pytorch.h
│   ├── modulated_deform_conv.cpp
│   ├── modulated_deform_conv_parrots.cpp
│   ├── modulated_deform_conv_pytorch.h
│   ├── ms_deform_attn.cpp
│   ├── ms_deform_attn_parrots.cpp
│   ├── nms.cpp
│   ├── nms_parrots.cpp
│   ├── nms_pytorch.h
│   ├── nms_rotated.cpp
│   ├── nms_rotated_cpu.cpp
│   ├── psamask.cpp
│   ├── psamask_parrots.cpp
│   ├── psamask_pytorch.h
│   ├── roi_align.cpp
│   ├── roi_align_cpu.cpp
│   ├── roi_align_parrots.cpp
│   ├── roi_align_pytorch.h
│   ├── roi_align_rotated.cpp
│   ├── roi_align_rotated_cpu.cpp
│   ├── roi_align_rotated_parrots.cpp
│   ├── roi_align_rotated_pytorch.h
│   ├── roi_pool.cpp
│   ├── roi_pool_parrots.cpp
│   ├── roi_pool_pytorch.h
│   ├── sync_bn.cpp
│   ├── sync_bn_parrots.cpp
│   ├── sync_bn_pytorch.h
│   ├── tin_shift.cpp
│   ├── tin_shift_parrots.cpp
│   ├── tin_shift_pytorch.h
│   ├── upfirdn2d.cpp
│   └── upfirdn2d_parrots.cpp
├── pytorch
│   ├── bbox_overlaps.cpp
│   ├── border_align.cpp
│   ├── box_iou_rotated.cpp
│   ├── box_iou_rotated_cpu.cpp
│   ├── carafe.cpp
│   ├── carafe_naive.cpp
│   ├── cc_attention.cpp
│   ├── contour_expand.cpp
│   ├── corner_pool.cpp
│   ├── deform_conv.cpp
│   ├── deform_roi_pool.cpp
│   ├── focal_loss.cpp
│   ├── fused_bias_leakyrelu.cpp
│   ├── info.cpp
│   ├── masked_conv2d.cpp
│   ├── modulated_deform_conv.cpp
│   ├── ms_deform_attn.cpp
│   ├── nms.cpp
│   ├── nms_rotated.cpp
│   ├── nms_rotated_cpu.cpp
│   ├── pixel_group.cpp
│   ├── psamask.cpp
│   ├── pybind.cpp
│   ├── roi_align.cpp
│   ├── roi_align_cpu.cpp
│   ├── roi_align_rotated.cpp
│   ├── roi_align_rotated_cpu.cpp
│   ├── roi_pool.cpp
│   ├── sync_bn.cpp
│   ├── tin_shift.cpp
│   └── upfirdn2d.cpp
│   ├── cuda
│   │   ├── bbox_overlaps_cuda.cu
│   │   ├── border_align_cuda.cu
│   │   ├── box_iou_rotated_cuda.cu
│   │   ├── carafe_cuda.cu
│   │   ├── carafe_naive_cuda.cu
│   │   ├── cc_attention_cuda.cu
│   │   ├── deform_conv_cuda.cu
│   │   ├── deform_roi_pool_cuda.cu
│   │   ├── focal_loss_cuda.cu
│   │   ├── fused_bias_leakyrelu_cuda.cu
│   │   ├── masked_conv2d_cuda.cu
│   │   ├── modulated_deform_conv_cuda.cu
│   │   ├── ms_deform_attn_cuda.cu
│   │   ├── nms_cuda.cu
│   │   ├── nms_rotated_cuda.cu
│   │   ├── psamask_cuda.cu
│   │   ├── roi_align_cuda.cu
│   │   ├── roi_align_rotated_cuda.cu
│   │   ├── roi_pool_cuda.cu
│   │   ├── sync_bn_cuda.cu
│   │   ├── tin_shift_cuda.cu
│   │   └── upfirdn2d_kernel.cu
└── tensorrt
    ├── plugins
    │   ├── trt_corner_pool.cpp
    │   ├── trt_corner_pool_kernel.cu
    │   ├── trt_cuda_helper.cu
    │   ├── trt_cummaxmin.cpp
    │   ├── trt_cummaxmin_kernel.cu
    │   ├── trt_deform_conv.cpp
    │   ├── trt_deform_conv_kernel.cu
    │   ├── trt_grid_sampler.cpp
    │   ├── trt_grid_sampler_kernel.cu
    │   ├── trt_instance_norm.cpp
    │   ├── trt_modulated_deform_conv.cpp
    │   ├── trt_modulated_deform_conv_kernel.cu
    │   ├── trt_nms.cpp
    │   ├── trt_nms_kernel.cu
    │   ├── trt_plugin.cpp
    │   ├── trt_roi_align.cpp
    │   ├── trt_roi_align_kernel.cu
    │   ├── trt_scatternd.cpp
    │   └── trt_scatternd_kernel.cu
    ├── trt_corner_pool.hpp
    ├── trt_cuda_helper.cuh
    ├── trt_cummaxmin.hpp
    ├── trt_deform_conv.hpp
    ├── trt_grid_sampler.hpp
    ├── trt_instance_norm.hpp
    ├── trt_modulated_deform_conv.hpp
    ├── trt_nms.hpp
    ├── trt_plugin_helper.hpp
    ├── trt_plugin.hpp
    ├── trt_roi_align.hpp
    ├── trt_scatternd.hpp
    └── trt_serialize.hpp

9 directories, 200 files

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repos?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
3. If the modification has potential influence on downstream projects, this PR should be tested with downstream projects, like MMDet or MMCls.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[codecov]

Codecov Report

Merging #1206 (84c02a0) into master (94a677d) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1206   +/-   ##
=======================================
  Coverage   68.28%   68.28%           
=======================================
  Files         160      160           
  Lines       10603    10603           
  Branches     1937     1937           
=======================================
  Hits         7240     7240           
  Misses       2979     2979           
  Partials      384      384           

Flag	Coverage Δ
unittests	68.28% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 94a677d...84c02a0. Read the comment docs.

[ZwwWayne]

I suggest we start to maintain compatibility documentation as in https://mmdetection.readthedocs.io/en/latest/compatibility.html to record these refactoring changes. Though they do not affect the usage in API.

[zhouzaida]

I suggest we start to maintain compatibility documentation as in https://mmdetection.readthedocs.io/en/latest/compatibility.html to record these refactoring changes. Though they do not affect the usage in API.

OK

[grimoire]

Could we rename HIP_DIFF to MMCV_WITH_HIP, just like other mmcv flag?