mirrors
/
mmcv
mirror of https://github.com/open-mmlab/mmcv.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

[Feature] Add NPU adapter for Deformable RoIPool (#2481) 

* add npu adapter for deformroipool

* cleancode

* Update deform_roi_pool.cpp

* Update deform_roi_pool.cpp

* Update deform_roi_pool.cpp

* Update deform_roi_pool.cpp

* Update deform_roi_pool.cpp

* Update deform_roi_pool.cpp

* Update deform_roi_pool.cpp
pull/2460/head^2
jayggh 2022-12-14 15:09:46 +08:00 committed by GitHub
parent 4c51afce2a
commit 46eb9ec5d0
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 71 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
docs/en/understand_mmcv/ops.md
View File

@ -19,7 +19,7 @@ We implement common ops used in detection, segmentation, etc.
| CornerPool | | √ | | | |
| Correlation | | √ | | | |
| Deformable Convolution v1/v2 | √ | √ | | | |
| Deformable RoIPool | | √ | √ | | |
| Deformable RoIPool | | √ | √ | | √ |
| DiffIoURotated | | √ | | | |
| DynamicScatter | | √ | | | |
| FurthestPointSample | | √ | | | |

2
docs/zh_cn/understand_mmcv/ops.md
View File

@ -19,7 +19,7 @@ MMCV 提供了检测、分割等任务中常用的算子
| CornerPool | | √ | | | |
| Correlation | | √ | | | |
| Deformable Convolution v1/v2 | √ | √ | | | |
| Deformable RoIPool | | √ | √ | | |
| Deformable RoIPool | | √ | √ | | √ |
| DiffIoURotated | | √ | | | |
| DynamicScatter | | √ | | | |
| FurthestPointSample | | √ | | | |

63
mmcv/ops/csrc/pytorch/npu/deform_roi_pool.cpp 100644
View File

@ -0,0 +1,63 @@
#include "pytorch_npu_helper.hpp"
using namespace NPU_NAME_SPACE;
using namespace std;
void deform_roi_pool_forward_impl(Tensor input, Tensor rois, Tensor offset,
Tensor output, int pooled_height,
int pooled_width, float spatial_scale,
int sampling_ratio, float gamma);
void deform_roi_pool_backward_impl(Tensor grad_output, Tensor input,
Tensor rois, Tensor offset,
Tensor grad_input, Tensor grad_offset,
int pooled_height, int pooled_width,
float spatial_scale, int sampling_ratio,
float gamma);
void deform_roi_pool_forward_npu(Tensor input, Tensor rois, Tensor offset,
Tensor output, int pooled_height,
int pooled_width, float spatial_scale,
int sampling_ratio, float gamma) {
c10::SmallVector<int64_t, 2> output_sizes = {pooled_height, pooled_width};
at::IntArrayRef output_size = at::IntArrayRef(output_sizes);
int64_t sampling_ratio_ = (int64_t)sampling_ratio;
OpCommand cmd;
cmd.Name("DeformableRoiPool")
.Input(input)
.Input(rois)
.Input(offset)
.Output(output)
.Attr("spatial_scale", spatial_scale)
.Attr("output_size", output_size)
.Attr("sampling_ratio", sampling_ratio_)
.Attr("gamma", gamma)
.Run();
}
void deform_roi_pool_backward_npu(Tensor grad_output, Tensor input, Tensor rois,
Tensor offset, Tensor grad_input,
Tensor grad_offset, int pooled_height,
int pooled_width, float spatial_scale,
int sampling_ratio, float gamma) {
c10::SmallVector<int64_t, 2> output_sizes = {pooled_height, pooled_width};
at::IntArrayRef output_size = at::IntArrayRef(output_sizes);
int64_t sampling_ratio_ = (int64_t)sampling_ratio;
OpCommand cmd;
cmd.Name("DeformableRoiPoolGrad")
.Input(grad_input)
.Input(input)
.Input(rois)
.Input(offset)
.Output(grad_output)
.Output(grad_offset)
.Attr("output_size", output_size)
.Attr("spatial_scale", spatial_scale)
.Attr("sample_ratio", sampling_ratio_)
.Attr("gamma", gamma)
.Run();
}
REGISTER_NPU_IMPL(deform_roi_pool_forward_impl, deform_roi_pool_forward_npu);
REGISTER_NPU_IMPL(deform_roi_pool_backward_impl, deform_roi_pool_backward_npu);

8
tests/test_ops/test_deform_roi_pool.py
View File

@ -5,7 +5,7 @@ import numpy as np
import pytest
import torch
from mmcv.utils import IS_CUDA_AVAILABLE, IS_MLU_AVAILABLE
from mmcv.utils import IS_CUDA_AVAILABLE, IS_MLU_AVAILABLE, IS_NPU_AVAILABLE
_USING_PARROTS = True
try:
@ -133,7 +133,11 @@ class TestDeformRoIPool:
pytest.param(
'mlu',
marks=pytest.mark.skipif(
not IS_MLU_AVAILABLE, reason='requires MLU support'))
not IS_MLU_AVAILABLE, reason='requires MLU support')),
pytest.param(
'npu',
marks=pytest.mark.skipif(
not IS_NPU_AVAILABLE, reason='requires NPU support'))
])
@pytest.mark.parametrize('dtype', [
torch.float,