chore: import upstream snapshot with attribution
Docker Image CI / build-ubuntu2004 (push) Waiting to run
Docker Image CI / build-ubuntu2004 (push) Waiting to run
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1,28 @@
|
||||
#
|
||||
# SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
#
|
||||
|
||||
add_plugin_source(
|
||||
efficientNMSInference.cu
|
||||
efficientNMSInference.cuh
|
||||
efficientNMSInference.h
|
||||
efficientNMSParameters.h
|
||||
efficientNMSPlugin.cpp
|
||||
efficientNMSPlugin.h
|
||||
)
|
||||
|
||||
add_subdirectory(tftrt)
|
||||
|
||||
@@ -0,0 +1,127 @@
|
||||
#
|
||||
# SPDX-FileCopyrightText: Copyright (c) 2022-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
#
|
||||
---
|
||||
name: EfficientNMS_TRT
|
||||
interface: "IPluginV2DynamicExt"
|
||||
versions:
|
||||
"1":
|
||||
inputs:
|
||||
- boxes
|
||||
- scores
|
||||
- anchors
|
||||
outputs:
|
||||
- num_detections
|
||||
- detection_boxes
|
||||
- detection_scores
|
||||
- detection_classes
|
||||
attributes:
|
||||
- score_threshold
|
||||
- iou_threshold
|
||||
- max_output_boxes
|
||||
- background_class
|
||||
- score_activation
|
||||
- class_agnostic
|
||||
- box_coding
|
||||
attribute_types:
|
||||
score_threshold: float32
|
||||
iou_threshold: float32
|
||||
max_output_boxes: int32
|
||||
background_class: int32
|
||||
score_activation: int32
|
||||
class_agnostic: int32
|
||||
box_coding: int32
|
||||
attribute_length:
|
||||
score_threshold: 1
|
||||
iou_threshold: 1
|
||||
max_output_boxes: 1
|
||||
background_class: 1
|
||||
score_activation: 1
|
||||
class_agnostic: 1
|
||||
box_coding: 1
|
||||
attribute_options:
|
||||
score_threshold:
|
||||
min: "=0"
|
||||
max: "=pinf"
|
||||
iou_threshold:
|
||||
min: "0"
|
||||
max: "=pinf"
|
||||
max_output_boxes:
|
||||
min: "0"
|
||||
max: "=pinf"
|
||||
background_class:
|
||||
min: "=ninf"
|
||||
max: "=pinf"
|
||||
score_activation:
|
||||
- 0
|
||||
- 1
|
||||
class_agnostic:
|
||||
- 0
|
||||
- 1
|
||||
box_coding:
|
||||
- 0
|
||||
- 1
|
||||
attributes_required:
|
||||
- score_threshold
|
||||
- iou_threshold
|
||||
- max_output_boxes
|
||||
- background_class
|
||||
- score_activation
|
||||
- box_coding
|
||||
golden_io_path: "plugin/EfficientNMSPlugin_PluginGoldenIO.json"
|
||||
abs_tol: 1e-5
|
||||
rel_tol: 1e-5
|
||||
configs:
|
||||
config1:
|
||||
input_types:
|
||||
boxes: float32
|
||||
scores: float32
|
||||
attribute_options:
|
||||
"background_class":
|
||||
value: -1
|
||||
shape: "1"
|
||||
"score_activation":
|
||||
value: 0
|
||||
shape: "1"
|
||||
"class_agnostic":
|
||||
value: 0
|
||||
shape: "1"
|
||||
"box_coding":
|
||||
value: 0
|
||||
shape: "1"
|
||||
output_types:
|
||||
num_detections: int32
|
||||
detection_boxes: float32
|
||||
class_agnostic:
|
||||
input_types:
|
||||
boxes: float32
|
||||
scores: float32
|
||||
attribute_options:
|
||||
"background_class":
|
||||
value: -1
|
||||
shape: "1"
|
||||
"score_activation":
|
||||
value: 0
|
||||
shape: "1"
|
||||
"class_agnostic":
|
||||
value: 1
|
||||
shape: "1"
|
||||
"box_coding":
|
||||
value: 0
|
||||
shape: "1"
|
||||
output_types:
|
||||
num_detections: int32
|
||||
detection_boxes: float32
|
||||
@@ -0,0 +1,183 @@
|
||||
# Efficient NMS Plugin [DEPRECATED]
|
||||
|
||||
**This plugin is deprecated since TensorRT 10.12 and will be removed in a future release. This includes the TF-TRT variants (EfficientNMS_Explicit_TF_TRT and EfficientNMS_Implicit_TF_TRT). No alternatives are planned to be provided.**
|
||||
|
||||
#### Table of Contents
|
||||
- [Description](#description)
|
||||
- [Structure](#structure)
|
||||
* [Inputs](#inputs)
|
||||
* [Dynamic Shape Support](#dynamic-shape-support)
|
||||
* [Box Coding Type](#box-coding-type)
|
||||
* [Outputs](#outputs)
|
||||
* [Parameters](#parameters)
|
||||
- [Algorithm](#algorithm)
|
||||
* [Process Description](#process-description)
|
||||
* [Performance Tuning](#performance-tuning)
|
||||
* [Additional Resources](#additional-resources)
|
||||
- [License](#license)
|
||||
|
||||
## Description
|
||||
|
||||
> NOTE: `EfficientNMSONNXPlugin` (`EfficientNMS_ONNX_TRT`) was removed in TensorRT 10.16. Its functionality has been superseded by the `INMSLayer`. `EfficientNMS_TRT` is deprecated since TensorRT 10.12; `INMSLayer` can be used as appropriate to replace its functionality.
|
||||
|
||||
This TensorRT plugin implements an efficient algorithm to perform Non Maximum Suppression for object detection networks.
|
||||
|
||||
This plugin is primarily intended for using with EfficientDet on TensorRT, as this network is particularly sensitive to the latencies introduced by slower NMS implementations. However, the plugin is generic enough that it will work correctly for other detections architectures, such as SSD or FasterRCNN.
|
||||
|
||||
## Structure
|
||||
|
||||
### Inputs
|
||||
|
||||
The plugin has two modes of operation, depending on the given input data. The plugin will automatically detect which mode to operate as, depending on the number of inputs it receives, as follows:
|
||||
|
||||
1. **Standard NMS Mode:** Only two input tensors are given, (i) the bounding box coordinates and (ii) the corresponding classification scores for each box.
|
||||
|
||||
2. **Fused Box Decoder Mode:** Three input tensors are given, (i) the raw localization predictions for each box originating directly from the localization head of the network, (ii) the corresponding classification scores originating from the classification head of the network, and (iii) the default anchor box coordinates usually hardcoded as constant tensors in the network.
|
||||
|
||||
Most object detection networks work by generating raw predictions from a "localization head" which adjust the coordinates of standard non-learned anchor coordinates to produce a tighter fitting bounding box. This process is called "box decoding", and it usually involves a large number of element-wise operations to transform the anchors to final box coordinates. As this can involve exponential operations on a large number of anchors, it can be computationally expensive, so this plugin gives the option of fusing the box decoder within the NMS operation which can be done in a far more efficient manner, resulting in lower latency for the network.
|
||||
|
||||
#### Boxes Input
|
||||
> **Input Shape:** `[batch_size, number_boxes, 4]` or `[batch_size, number_boxes, number_classes, 4]`
|
||||
>
|
||||
> **Data Type:** `float32` or `float16`
|
||||
|
||||
The boxes input can have 3 dimensions in case a single box prediction is produced for all classes (such as in EfficientDet or SSD), or 4 dimensions when separate box predictions are generated for each class (such as in FasterRCNN), in which case `number_classes` >= 1 and must match the number of classes in the scores input. The final dimension represents the four coordinates that define the bounding box prediction.
|
||||
|
||||
For *Standard NMS* mode, this tensor should contain the final box coordinates for each predicted detection. For *Fused Box Decoder* mode, this tensor should have the raw localization predictions. In either case, this data is given as `4` coordinates which makes up the final shape dimension.
|
||||
|
||||
#### Scores Input
|
||||
> **Input Shape:** `[batch_size, number_boxes, number_classes]`
|
||||
>
|
||||
> **Data Type:** `float32` or `float16`
|
||||
|
||||
The scores input has `number_classes` elements with the predicted scores for each candidate class for each of the `number_boxes` anchor boxes.
|
||||
|
||||
Usually, the score values will have passed through a sigmoid activation function before reaching the NMS operation. However, as an optimization, the pre-sigmoid raw scores can also be provided to the NMS plugin to reduce overall network latency. If raw scores are given, enable the `score_activation` parameter so they are processed accordingly.
|
||||
|
||||
#### Anchors Input (Optional)
|
||||
> **Input Shape:** `[1, number_boxes, 4]` or `[batch_size, number_boxes, 4]`
|
||||
>
|
||||
> **Data Type:** `float32` or `float16`
|
||||
|
||||
Only used in *Fused Box Decoder* mode. It is much more efficient to perform the box decoding within this plugin. In this case, the boxes input will be treated as the raw localization head box corrections, and this third input should contain the default anchor/prior box coordinates.
|
||||
|
||||
When used, the input must have 3 dimensions, where the first one may be either `1` in case anchors are constant for all images in a batch, or `batch_size` in case each image has different anchors -- such as in the box refinement NMS of FasterRCNN's second stage.
|
||||
|
||||
### Dynamic Shape Support
|
||||
|
||||
Most input shape dimensions, namely `batch_size`, `number_boxes`, and `number_classes`, for all inputs can be defined dynamically at runtime if the TensorRT engine is built with dynamic input shapes. However, once defined, these dimensions must match across all tensors that use them (e.g. the same `number_boxes` dimension must be given for both boxes and scores, etc.)
|
||||
|
||||
### Box Coding Type
|
||||
Different object detection networks represent their box coordinate system differently. The two types supported by this plugin are:
|
||||
|
||||
1. **BoxCorners:** The four coordinates represent `[x1, y1, x2, y2]` values, where each x,y pair defines the top-left and bottom-right corners of a bounding box.
|
||||
2. **BoxCenterSize:** The four coordinates represent `[x, y, w, h]` values, where the x,y pair define the box center location, and the w,h pair define its width and height.
|
||||
|
||||
Note that for NMS purposes, horizontal and vertical coordinates are fully interchangeable. TensorFlow-trained networks, for example, often uses vertical-first coordinates such as `[y1, x1, y2, x2]`, but this coordinate system will work equally well under the BoxCorner coding. Similarly, `[y, x, h, w]` will be properly covered by the BoxCornerSize coding.
|
||||
|
||||
In *Fused Box Decoder* mode, the boxes and anchor tensors should both use the same coding.
|
||||
|
||||
### Outputs
|
||||
|
||||
The following four output tensors are generated:
|
||||
|
||||
- **num_detections:**
|
||||
This is a `[batch_size, 1]` tensor of data type `int32`. The last dimension is a scalar indicating the number of valid detections per batch image. It can be less than `max_output_boxes`. Only the top `num_detections[i]` entries in `nms_boxes[i]`, `nms_scores[i]` and `nms_classes[i]` are valid.
|
||||
|
||||
- **detection_boxes:**
|
||||
This is a `[batch_size, max_output_boxes, 4]` tensor of data type `float32` or `float16`, containing the coordinates of non-max suppressed boxes. The output coordinates will always be in BoxCorner format, regardless of the input code type.
|
||||
|
||||
- **detection_scores:**
|
||||
This is a `[batch_size, max_output_boxes]` tensor of data type `float32` or `float16`, containing the scores for the boxes.
|
||||
|
||||
- **detection_classes:**
|
||||
This is a `[batch_size, max_output_boxes]` tensor of data type `int32`, containing the classes for the boxes.
|
||||
|
||||
### Parameters
|
||||
|
||||
| Type | Parameter | Description
|
||||
|----------|--------------------------|--------------------------------------------------------
|
||||
|`float` |`score_threshold` * |The scalar threshold for score (low scoring boxes are removed).
|
||||
|`float` |`iou_threshold` |The scalar threshold for IOU (additional boxes that have high IOU overlap with previously selected boxes are removed).
|
||||
|`int` |`max_output_boxes` |The maximum number of detections to output per image.
|
||||
|`int` |`background_class` |The label ID for the background class. If there is no background class, set it to `-1`.
|
||||
|`bool` |`score_activation` * |Set to true to apply sigmoid activation to the confidence scores during NMS operation.
|
||||
|`bool` |`class_agnostic` |Set to true to do class-independent NMS; otherwise, boxes of different classes would be considered separately during NMS.
|
||||
|`int` |`box_coding` |Coding type used for boxes (and anchors if applicable), 0 = BoxCorner, 1 = BoxCenterSize.
|
||||
|
||||
Parameters marked with a `*` have a non-negligible effect on runtime latency. See the [Performance Tuning](#performance-tuning) section below for more details on how to set them optimally.
|
||||
|
||||
## Algorithm
|
||||
|
||||
### Process Description
|
||||
|
||||
The NMS algorithm in this plugin first filters the scores below the given `scoreThreshold`. This subset of scores is then sorted, and their corresponding boxes are then further filtered out by removing boxes that overlap each other with an IOU above the given `iouThreshold`.
|
||||
|
||||
The algorithm launcher and its relevant CUDA kernels are all defined in the `efficientNMSInference.cu` file.
|
||||
|
||||
Specifically, the NMS algorithm does the following:
|
||||
|
||||
- The scores are filtered with the `score_threshold` parameter to reject any scores below the score threshold, while maintaining indexing to cross-reference these scores to their corresponding box coordinates. This is done with the `EfficientNMSFilter` CUDA kernel.
|
||||
|
||||
- If too many elements are kept, due to a very low (or zero) score threshold, the filter operation can become a bottleneck due to the atomic operations involved. To mitigate this, a fallback kernel `EfficientNMSDenseIndex` is used instead which passes all the score elements densely packed and indexed. This method is heuristically selected only if the score threshold is less than 0.007.
|
||||
|
||||
- The selected scores that remain after filtering are sorted in descending order. The indexing is carefully handled to still maintain score to box relationships after sorting.
|
||||
|
||||
- After sorting, the highest 4096 scores are processed by the `EfficientNMS` CUDA kernel. This algorithm uses the index data maintained throughout the previous steps to find the boxes corresponding to the remaining scores. If the fused box decoder is being used, decoding will happen until this stage, where only the top scoring boxes need to be decoded.
|
||||
|
||||
- The NMS kernel uses an efficient filtering algorithm that largely reduces the number of IOU overlap cross-checks between box pairs. The boxes that survive the IOU filtering finally pass through to the output results. At this stage, the sigmoid activation is applied to only the final remaining scores, if `score_activation` is enabled, thereby greatly reducing the amount of sigmoid calculations required otherwise.
|
||||
|
||||
### Performance Tuning
|
||||
|
||||
The plugin implements a very efficient NMS algorithm which largely reduces the latency of this operation in comparison to other NMS plugins. However, there are certain considerations that can help to better fine tune its performance:
|
||||
|
||||
#### Choosing the Score Threshold
|
||||
|
||||
The algorithm is highly sensitive to the selected `score_threshold` parameter. With a higher threshold, fewer elements need to be processed and so the algorithm runs much faster. Therefore, it's beneficial to always select the highest possible score threshold that fulfills the application requirements. Threshold values lower than approximately 0.01 may cause substantially higher latency.
|
||||
|
||||
#### Using Sigmoid Activation
|
||||
|
||||
Depending on network configuration, it is usually more efficient to provide raw scores (pre-sigmoid) to the NMS plugin scores input, and enable the `score_activation` parameter. Doing so applies a sigmoid activation only to the last `max_output_boxes` selected scores, instead of all the predicted scores, largely reducing the computational cost.
|
||||
|
||||
#### Class Independent NMS
|
||||
|
||||
Some object detection networks/architectures like YOLO series need to use class-independent NMS operations. If `class_agnostic` is enabled, class-independent NMS is performed; otherwise, different classes would do NMS separately.
|
||||
|
||||
#### Using the Fused Box Decoder
|
||||
|
||||
When using networks with many anchors, such as EfficientDet or SSD, it may be more efficient to do box decoding within the NMS plugin. For this, pass the raw box predictions as the boxes input, and the default anchor coordinates as the optional third input to the plugin.
|
||||
|
||||
### Additional Resources
|
||||
|
||||
The following resources provide a deeper understanding of the NMS algorithm:
|
||||
|
||||
#### Networks
|
||||
- [EfficientDet](https://arxiv.org/abs/1911.09070)
|
||||
- [SSD: Single Shot MultiBox Detector](https://arxiv.org/abs/1512.02325)
|
||||
- [Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks](https://arxiv.org/abs/1506.01497)
|
||||
- [Mask R-CNN](https://arxiv.org/abs/1703.06870)
|
||||
|
||||
|
||||
#### Documentation
|
||||
- [NMS algorithm](https://www.coursera.org/lecture/convolutional-neural-networks/non-max-suppression-dvrjH)
|
||||
- [NonMaxSuppression ONNX Op](https://github.com/onnx/onnx/blob/master/docs/Operators.md#NonMaxSuppression)
|
||||
|
||||
## License
|
||||
|
||||
For terms and conditions for use, reproduction, and distribution, see the [TensorRT Software License Agreement](https://docs.nvidia.com/deeplearning/sdk/tensorrt-sla/index.html)
|
||||
documentation.
|
||||
|
||||
## Changelog
|
||||
|
||||
March 2026
|
||||
Remove `EfficientNMS_ONNX_TRT` plugin. Use `INMSLayer` instead.
|
||||
|
||||
May 2025
|
||||
Add deprecation note for the `EfficientNMS_TRT` plugin.
|
||||
|
||||
June 2023
|
||||
Add deprecation note for the `EfficientNMSONNXPlugin` plugin.
|
||||
|
||||
## Known issues
|
||||
|
||||
There are no known issues in this plugin.
|
||||
@@ -0,0 +1,725 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#include "common/bboxUtils.h"
|
||||
#include "cub/cub.cuh"
|
||||
#include "cuda_runtime_api.h"
|
||||
|
||||
#include "efficientNMSInference.cuh"
|
||||
#include "efficientNMSInference.h"
|
||||
|
||||
#define NMS_TILES 5
|
||||
|
||||
using namespace nvinfer1;
|
||||
using namespace nvinfer1::plugin;
|
||||
|
||||
template <typename T>
|
||||
__device__ float IOU(EfficientNMSParameters param, BoxCorner<T> box1, BoxCorner<T> box2)
|
||||
{
|
||||
// Regardless of the selected box coding, IOU is always performed in BoxCorner coding.
|
||||
// The boxes are copied so that they can be reordered without affecting the originals.
|
||||
BoxCorner<T> b1 = box1;
|
||||
BoxCorner<T> b2 = box2;
|
||||
b1.reorder();
|
||||
b2.reorder();
|
||||
float intersectArea = BoxCorner<T>::intersect(b1, b2).area();
|
||||
if (intersectArea <= 0.f)
|
||||
{
|
||||
return 0.f;
|
||||
}
|
||||
float unionArea = b1.area() + b2.area() - intersectArea;
|
||||
if (unionArea <= 0.f)
|
||||
{
|
||||
return 0.f;
|
||||
}
|
||||
return intersectArea / unionArea;
|
||||
}
|
||||
|
||||
template <typename T, typename Tb>
|
||||
__device__ BoxCorner<T> DecodeBoxes(EfficientNMSParameters param, int boxIdx, int anchorIdx,
|
||||
const Tb* __restrict__ boxesInput, const Tb* __restrict__ anchorsInput)
|
||||
{
|
||||
// The inputs will be in the selected coding format, as well as the decoding function. But the decoded box
|
||||
// will always be returned as BoxCorner.
|
||||
Tb box = boxesInput[boxIdx];
|
||||
if (!param.boxDecoder)
|
||||
{
|
||||
return BoxCorner<T>(box);
|
||||
}
|
||||
Tb anchor = anchorsInput[anchorIdx];
|
||||
box.reorder();
|
||||
anchor.reorder();
|
||||
return BoxCorner<T>(box.decode(anchor));
|
||||
}
|
||||
|
||||
template <typename T, typename Tb>
|
||||
__device__ void MapNMSData(EfficientNMSParameters param, int idx, int imageIdx, const Tb* __restrict__ boxesInput,
|
||||
const Tb* __restrict__ anchorsInput, const int* __restrict__ topClassData, const int* __restrict__ topAnchorsData,
|
||||
const int* __restrict__ topNumData, const T* __restrict__ sortedScoresData, const int* __restrict__ sortedIndexData,
|
||||
T& scoreMap, int& classMap, BoxCorner<T>& boxMap, int& boxIdxMap)
|
||||
{
|
||||
// idx: Holds the NMS box index, within the current batch.
|
||||
// idxSort: Holds the batched NMS box index, which indexes the (filtered, but sorted) score buffer.
|
||||
// scoreMap: Holds the score that corresponds to the indexed box being processed by NMS.
|
||||
if (idx >= topNumData[imageIdx])
|
||||
{
|
||||
return;
|
||||
}
|
||||
int idxSort = imageIdx * param.numScoreElements + idx;
|
||||
scoreMap = sortedScoresData[idxSort];
|
||||
|
||||
// idxMap: Holds the re-mapped index, which indexes the (filtered, but unsorted) buffers.
|
||||
// classMap: Holds the class that corresponds to the idx'th sorted score being processed by NMS.
|
||||
// anchorMap: Holds the anchor that corresponds to the idx'th sorted score being processed by NMS.
|
||||
int idxMap = imageIdx * param.numScoreElements + sortedIndexData[idxSort];
|
||||
classMap = topClassData[idxMap];
|
||||
int anchorMap = topAnchorsData[idxMap];
|
||||
|
||||
// boxIdxMap: Holds the re-re-mapped index, which indexes the (unfiltered, and unsorted) boxes input buffer.
|
||||
boxIdxMap = -1;
|
||||
if (param.shareLocation) // Shape of boxesInput: [batchSize, numAnchors, 1, 4]
|
||||
{
|
||||
boxIdxMap = imageIdx * param.numAnchors + anchorMap;
|
||||
}
|
||||
else // Shape of boxesInput: [batchSize, numAnchors, numClasses, 4]
|
||||
{
|
||||
int batchOffset = imageIdx * param.numAnchors * param.numClasses;
|
||||
int anchorOffset = anchorMap * param.numClasses;
|
||||
boxIdxMap = batchOffset + anchorOffset + classMap;
|
||||
}
|
||||
// anchorIdxMap: Holds the re-re-mapped index, which indexes the (unfiltered, and unsorted) anchors input buffer.
|
||||
int anchorIdxMap = -1;
|
||||
if (param.shareAnchors) // Shape of anchorsInput: [1, numAnchors, 4]
|
||||
{
|
||||
anchorIdxMap = anchorMap;
|
||||
}
|
||||
else // Shape of anchorsInput: [batchSize, numAnchors, 4]
|
||||
{
|
||||
anchorIdxMap = imageIdx * param.numAnchors + anchorMap;
|
||||
}
|
||||
// boxMap: Holds the box that corresponds to the idx'th sorted score being processed by NMS.
|
||||
boxMap = DecodeBoxes<T, Tb>(param, boxIdxMap, anchorIdxMap, boxesInput, anchorsInput);
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
__device__ void WriteNMSResult(EfficientNMSParameters param, int* __restrict__ numDetectionsOutput,
|
||||
T* __restrict__ nmsScoresOutput, int* __restrict__ nmsClassesOutput, BoxCorner<T>* __restrict__ nmsBoxesOutput,
|
||||
T threadScore, int threadClass, BoxCorner<T> threadBox, int imageIdx, unsigned int resultsCounter)
|
||||
{
|
||||
int outputIdx = imageIdx * param.numOutputBoxes + resultsCounter - 1;
|
||||
if (param.scoreSigmoid)
|
||||
{
|
||||
nmsScoresOutput[outputIdx] = sigmoid_mp(threadScore);
|
||||
}
|
||||
else if (param.scoreBits > 0)
|
||||
{
|
||||
nmsScoresOutput[outputIdx] = add_mp(threadScore, (T) -1);
|
||||
}
|
||||
else
|
||||
{
|
||||
nmsScoresOutput[outputIdx] = threadScore;
|
||||
}
|
||||
nmsClassesOutput[outputIdx] = threadClass;
|
||||
if (param.clipBoxes)
|
||||
{
|
||||
nmsBoxesOutput[outputIdx] = threadBox.clip((T) 0, (T) 1);
|
||||
}
|
||||
else
|
||||
{
|
||||
nmsBoxesOutput[outputIdx] = threadBox;
|
||||
}
|
||||
numDetectionsOutput[imageIdx] = resultsCounter;
|
||||
}
|
||||
|
||||
__device__ void WriteONNXResult(EfficientNMSParameters param, int* outputIndexData, int* __restrict__ nmsIndicesOutput,
|
||||
int imageIdx, int threadClass, int boxIdxMap)
|
||||
{
|
||||
int index = boxIdxMap % param.numAnchors;
|
||||
int idx = atomicAdd((unsigned int*) &outputIndexData[0], 1);
|
||||
nmsIndicesOutput[idx * 3 + 0] = imageIdx;
|
||||
nmsIndicesOutput[idx * 3 + 1] = threadClass;
|
||||
nmsIndicesOutput[idx * 3 + 2] = index;
|
||||
}
|
||||
|
||||
__global__ void PadONNXResult(EfficientNMSParameters param, int* outputIndexData, int* __restrict__ nmsIndicesOutput)
|
||||
{
|
||||
if (threadIdx.x > 0)
|
||||
{
|
||||
return;
|
||||
}
|
||||
int pidx = outputIndexData[0] - 1;
|
||||
if (pidx < 0)
|
||||
{
|
||||
return;
|
||||
}
|
||||
for (int idx = pidx + 1; idx < param.batchSize * param.numOutputBoxes; idx++)
|
||||
{
|
||||
nmsIndicesOutput[idx * 3 + 0] = nmsIndicesOutput[pidx * 3 + 0];
|
||||
nmsIndicesOutput[idx * 3 + 1] = nmsIndicesOutput[pidx * 3 + 1];
|
||||
nmsIndicesOutput[idx * 3 + 2] = nmsIndicesOutput[pidx * 3 + 2];
|
||||
}
|
||||
}
|
||||
|
||||
template <typename T, typename Tb>
|
||||
__global__ void EfficientNMS(EfficientNMSParameters param, const int* topNumData, int* outputIndexData,
|
||||
int* outputClassData, const int* sortedIndexData, const T* __restrict__ sortedScoresData,
|
||||
const int* __restrict__ topClassData, const int* __restrict__ topAnchorsData, const Tb* __restrict__ boxesInput,
|
||||
const Tb* __restrict__ anchorsInput, int* __restrict__ numDetectionsOutput, T* __restrict__ nmsScoresOutput,
|
||||
int* __restrict__ nmsClassesOutput, int* __restrict__ nmsIndicesOutput, BoxCorner<T>* __restrict__ nmsBoxesOutput)
|
||||
{
|
||||
unsigned int thread = threadIdx.x;
|
||||
unsigned int imageIdx = blockIdx.y;
|
||||
unsigned int tileSize = blockDim.x;
|
||||
if (imageIdx >= param.batchSize)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
int numSelectedBoxes = min(topNumData[imageIdx], param.numSelectedBoxes);
|
||||
int numTiles = (numSelectedBoxes + tileSize - 1) / tileSize;
|
||||
if (thread >= numSelectedBoxes)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
__shared__ int blockState;
|
||||
__shared__ unsigned int resultsCounter;
|
||||
if (thread == 0)
|
||||
{
|
||||
blockState = 0;
|
||||
resultsCounter = 0;
|
||||
}
|
||||
|
||||
int threadState[NMS_TILES];
|
||||
unsigned int boxIdx[NMS_TILES];
|
||||
T threadScore[NMS_TILES];
|
||||
int threadClass[NMS_TILES];
|
||||
BoxCorner<T> threadBox[NMS_TILES];
|
||||
int boxIdxMap[NMS_TILES];
|
||||
for (int tile = 0; tile < numTiles; tile++)
|
||||
{
|
||||
threadState[tile] = 0;
|
||||
boxIdx[tile] = thread + tile * blockDim.x;
|
||||
MapNMSData<T, Tb>(param, boxIdx[tile], imageIdx, boxesInput, anchorsInput, topClassData, topAnchorsData,
|
||||
topNumData, sortedScoresData, sortedIndexData, threadScore[tile], threadClass[tile], threadBox[tile],
|
||||
boxIdxMap[tile]);
|
||||
}
|
||||
|
||||
// Iterate through all boxes to NMS against.
|
||||
for (int i = 0; i < numSelectedBoxes; i++)
|
||||
{
|
||||
int tile = i / tileSize;
|
||||
|
||||
if (boxIdx[tile] == i)
|
||||
{
|
||||
// Iteration lead thread, figure out what the other threads should do,
|
||||
// this will be signaled via the blockState shared variable.
|
||||
if (threadState[tile] == -1)
|
||||
{
|
||||
// Thread already dead, this box was already dropped in a previous iteration,
|
||||
// because it had a large IOU overlap with another lead thread previously, so
|
||||
// it would never be kept anyway, therefore it can safely be skip all IOU operations
|
||||
// in this iteration.
|
||||
blockState = -1; // -1 => Signal all threads to skip iteration
|
||||
}
|
||||
else if (threadState[tile] == 0)
|
||||
{
|
||||
// As this box will be kept, this is a good place to find what index in the results buffer it
|
||||
// should have, as this allows to perform an early loop exit if there are enough results.
|
||||
if (resultsCounter >= param.numOutputBoxes)
|
||||
{
|
||||
blockState = -2; // -2 => Signal all threads to do an early loop exit.
|
||||
}
|
||||
else
|
||||
{
|
||||
// Thread is still alive, because it has not had a large enough IOU overlap with
|
||||
// any other kept box previously. Therefore, this box will be kept for sure. However,
|
||||
// we need to check against all other subsequent boxes from this position onward,
|
||||
// to see how those other boxes will behave in future iterations.
|
||||
blockState = 1; // +1 => Signal all (higher index) threads to calculate IOU against this box
|
||||
threadState[tile] = 1; // +1 => Mark this box's thread to be kept and written out to results
|
||||
|
||||
// If the numOutputBoxesPerClass check is enabled, write the result only if the limit for this
|
||||
// class on this image has not been reached yet. Other than (possibly) skipping the write, this
|
||||
// won't affect anything else in the NMS threading.
|
||||
bool write = true;
|
||||
if (param.numOutputBoxesPerClass >= 0)
|
||||
{
|
||||
int classCounterIdx = imageIdx * param.numClasses + threadClass[tile];
|
||||
write = (outputClassData[classCounterIdx] < param.numOutputBoxesPerClass);
|
||||
outputClassData[classCounterIdx]++;
|
||||
}
|
||||
if (write)
|
||||
{
|
||||
// This branch is visited by one thread per iteration, so it's safe to do non-atomic increments.
|
||||
resultsCounter++;
|
||||
if (param.outputONNXIndices)
|
||||
{
|
||||
WriteONNXResult(
|
||||
param, outputIndexData, nmsIndicesOutput, imageIdx, threadClass[tile], boxIdxMap[tile]);
|
||||
}
|
||||
else
|
||||
{
|
||||
WriteNMSResult<T>(param, numDetectionsOutput, nmsScoresOutput, nmsClassesOutput,
|
||||
nmsBoxesOutput, threadScore[tile], threadClass[tile], threadBox[tile], imageIdx,
|
||||
resultsCounter);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
// This state should never be reached, but just in case...
|
||||
blockState = 0; // 0 => Signal all threads to not do any updates, nothing happens.
|
||||
}
|
||||
}
|
||||
|
||||
__syncthreads();
|
||||
|
||||
if (blockState == -2)
|
||||
{
|
||||
// This is the signal to exit from the loop.
|
||||
return;
|
||||
}
|
||||
|
||||
if (blockState == -1)
|
||||
{
|
||||
// This is the signal for all threads to just skip this iteration, as no IOU's need to be checked.
|
||||
continue;
|
||||
}
|
||||
|
||||
// Grab a box and class to test the current box against. The test box corresponds to iteration i,
|
||||
// therefore it will have a lower index than the current thread box, and will therefore have a higher score
|
||||
// than the current box because it's located "before" in the sorted score list.
|
||||
T testScore;
|
||||
int testClass;
|
||||
BoxCorner<T> testBox;
|
||||
int testBoxIdxMap;
|
||||
MapNMSData<T, Tb>(param, i, imageIdx, boxesInput, anchorsInput, topClassData, topAnchorsData, topNumData,
|
||||
sortedScoresData, sortedIndexData, testScore, testClass, testBox, testBoxIdxMap);
|
||||
|
||||
for (int tile = 0; tile < numTiles; tile++)
|
||||
{
|
||||
bool ignoreClass = true;
|
||||
if (!param.classAgnostic)
|
||||
{
|
||||
ignoreClass = threadClass[tile] == testClass;
|
||||
}
|
||||
|
||||
// IOU
|
||||
if (boxIdx[tile] > i && // Make sure two different boxes are being tested, and that it's a higher index;
|
||||
boxIdx[tile] < numSelectedBoxes && // Make sure the box is within numSelectedBoxes;
|
||||
blockState == 1 && // Signal that allows IOU checks to be performed;
|
||||
threadState[tile] == 0 && // Make sure this box hasn't been either dropped or kept already;
|
||||
ignoreClass && // Compare only boxes of matching classes when classAgnostic is false;
|
||||
lte_mp(threadScore[tile], testScore) && // Make sure the sorting order of scores is as expected;
|
||||
IOU<T>(param, threadBox[tile], testBox) >= param.iouThreshold) // And... IOU overlap.
|
||||
{
|
||||
// Current box overlaps with the box tested in this iteration, this box will be skipped.
|
||||
threadState[tile] = -1; // -1 => Mark this box's thread to be dropped.
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
cudaError_t EfficientNMSLauncher(EfficientNMSParameters& param, int* topNumData, int* outputIndexData,
|
||||
int* outputClassData, int* sortedIndexData, T* sortedScoresData, int* topClassData, int* topAnchorsData,
|
||||
const void* boxesInput, const void* anchorsInput, int* numDetectionsOutput, T* nmsScoresOutput,
|
||||
int* nmsClassesOutput, int* nmsIndicesOutput, void* nmsBoxesOutput, cudaStream_t stream)
|
||||
{
|
||||
unsigned int tileSize = param.numSelectedBoxes / NMS_TILES;
|
||||
if (param.numSelectedBoxes <= 512)
|
||||
{
|
||||
tileSize = 512;
|
||||
}
|
||||
if (param.numSelectedBoxes <= 256)
|
||||
{
|
||||
tileSize = 256;
|
||||
}
|
||||
|
||||
const dim3 blockSize = {tileSize, 1, 1};
|
||||
const dim3 gridSize = {1, (unsigned int) param.batchSize, 1};
|
||||
|
||||
if (param.boxCoding == 0)
|
||||
{
|
||||
EfficientNMS<T, BoxCorner<T>><<<gridSize, blockSize, 0, stream>>>(param, topNumData, outputIndexData,
|
||||
outputClassData, sortedIndexData, sortedScoresData, topClassData, topAnchorsData,
|
||||
(BoxCorner<T>*) boxesInput, (BoxCorner<T>*) anchorsInput, numDetectionsOutput, nmsScoresOutput,
|
||||
nmsClassesOutput, nmsIndicesOutput, (BoxCorner<T>*) nmsBoxesOutput);
|
||||
}
|
||||
else if (param.boxCoding == 1)
|
||||
{
|
||||
// Note that nmsBoxesOutput is always coded as BoxCorner<T>, regardless of the input coding type.
|
||||
EfficientNMS<T, BoxCenterSize<T>><<<gridSize, blockSize, 0, stream>>>(param, topNumData, outputIndexData,
|
||||
outputClassData, sortedIndexData, sortedScoresData, topClassData, topAnchorsData,
|
||||
(BoxCenterSize<T>*) boxesInput, (BoxCenterSize<T>*) anchorsInput, numDetectionsOutput, nmsScoresOutput,
|
||||
nmsClassesOutput, nmsIndicesOutput, (BoxCorner<T>*) nmsBoxesOutput);
|
||||
}
|
||||
|
||||
if (param.outputONNXIndices)
|
||||
{
|
||||
PadONNXResult<<<1, 1, 0, stream>>>(param, outputIndexData, nmsIndicesOutput);
|
||||
}
|
||||
|
||||
return cudaGetLastError();
|
||||
}
|
||||
|
||||
__global__ void EfficientNMSFilterSegments(EfficientNMSParameters param, const int* __restrict__ topNumData,
|
||||
int* __restrict__ topOffsetsStartData, int* __restrict__ topOffsetsEndData)
|
||||
{
|
||||
int imageIdx = threadIdx.x;
|
||||
if (imageIdx > param.batchSize)
|
||||
{
|
||||
return;
|
||||
}
|
||||
topOffsetsStartData[imageIdx] = imageIdx * param.numScoreElements;
|
||||
topOffsetsEndData[imageIdx] = imageIdx * param.numScoreElements + topNumData[imageIdx];
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
__global__ void EfficientNMSFilter(EfficientNMSParameters param, const T* __restrict__ scoresInput,
|
||||
int* __restrict__ topNumData, int* __restrict__ topIndexData, int* __restrict__ topAnchorsData,
|
||||
T* __restrict__ topScoresData, int* __restrict__ topClassData)
|
||||
{
|
||||
int elementIdx = blockDim.x * blockIdx.x + threadIdx.x;
|
||||
int imageIdx = blockDim.y * blockIdx.y + threadIdx.y;
|
||||
|
||||
// Boundary Conditions
|
||||
if (elementIdx >= param.numScoreElements || imageIdx >= param.batchSize)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
// Shape of scoresInput: [batchSize, numAnchors, numClasses]
|
||||
int scoresInputIdx = imageIdx * param.numScoreElements + elementIdx;
|
||||
|
||||
// For each class, check its corresponding score if it crosses the threshold, and if so select this anchor,
|
||||
// and keep track of the maximum score and the corresponding (argmax) class id
|
||||
T score = scoresInput[scoresInputIdx];
|
||||
if (gte_mp(score, (T) param.scoreThreshold))
|
||||
{
|
||||
// Unpack the class and anchor index from the element index
|
||||
int classIdx = elementIdx % param.numClasses;
|
||||
int anchorIdx = elementIdx / param.numClasses;
|
||||
|
||||
// If this is a background class, ignore it.
|
||||
if (classIdx == param.backgroundClass)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
// Use an atomic to find an open slot where to write the selected anchor data.
|
||||
if (topNumData[imageIdx] >= param.numScoreElements)
|
||||
{
|
||||
return;
|
||||
}
|
||||
int selectedIdx = atomicAdd((unsigned int*) &topNumData[imageIdx], 1);
|
||||
if (selectedIdx >= param.numScoreElements)
|
||||
{
|
||||
topNumData[imageIdx] = param.numScoreElements;
|
||||
return;
|
||||
}
|
||||
|
||||
// Shape of topScoresData / topClassData: [batchSize, numScoreElements]
|
||||
int topIdx = imageIdx * param.numScoreElements + selectedIdx;
|
||||
|
||||
if (param.scoreBits > 0)
|
||||
{
|
||||
score = add_mp(score, (T) 1);
|
||||
if (gt_mp(score, (T) (2.f - 1.f / 1024.f)))
|
||||
{
|
||||
// Ensure the incremented score fits in the mantissa without changing the exponent
|
||||
score = (2.f - 1.f / 1024.f);
|
||||
}
|
||||
}
|
||||
|
||||
topIndexData[topIdx] = selectedIdx;
|
||||
topAnchorsData[topIdx] = anchorIdx;
|
||||
topScoresData[topIdx] = score;
|
||||
topClassData[topIdx] = classIdx;
|
||||
}
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
__global__ void EfficientNMSDenseIndex(EfficientNMSParameters param, int* __restrict__ topNumData,
|
||||
int* __restrict__ topIndexData, int* __restrict__ topAnchorsData, int* __restrict__ topOffsetsStartData,
|
||||
int* __restrict__ topOffsetsEndData, T* __restrict__ topScoresData, int* __restrict__ topClassData)
|
||||
{
|
||||
int elementIdx = blockDim.x * blockIdx.x + threadIdx.x;
|
||||
int imageIdx = blockDim.y * blockIdx.y + threadIdx.y;
|
||||
|
||||
if (elementIdx >= param.numScoreElements || imageIdx >= param.batchSize)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
int dataIdx = imageIdx * param.numScoreElements + elementIdx;
|
||||
int anchorIdx = elementIdx / param.numClasses;
|
||||
int classIdx = elementIdx % param.numClasses;
|
||||
if (param.scoreBits > 0)
|
||||
{
|
||||
T score = topScoresData[dataIdx];
|
||||
if (lt_mp(score, (T) param.scoreThreshold))
|
||||
{
|
||||
score = (T) 1;
|
||||
}
|
||||
else if (classIdx == param.backgroundClass)
|
||||
{
|
||||
score = (T) 1;
|
||||
}
|
||||
else
|
||||
{
|
||||
score = add_mp(score, (T) 1);
|
||||
if (gt_mp(score, (T) (2.f - 1.f / 1024.f)))
|
||||
{
|
||||
// Ensure the incremented score fits in the mantissa without changing the exponent
|
||||
score = (2.f - 1.f / 1024.f);
|
||||
}
|
||||
}
|
||||
topScoresData[dataIdx] = score;
|
||||
}
|
||||
else
|
||||
{
|
||||
T score = topScoresData[dataIdx];
|
||||
if (lt_mp(score, (T) param.scoreThreshold))
|
||||
{
|
||||
topScoresData[dataIdx] = -(1 << 15);
|
||||
}
|
||||
else if (classIdx == param.backgroundClass)
|
||||
{
|
||||
topScoresData[dataIdx] = -(1 << 15);
|
||||
}
|
||||
}
|
||||
|
||||
topIndexData[dataIdx] = elementIdx;
|
||||
topAnchorsData[dataIdx] = anchorIdx;
|
||||
topClassData[dataIdx] = classIdx;
|
||||
|
||||
if (elementIdx == 0)
|
||||
{
|
||||
// Saturate counters
|
||||
topNumData[imageIdx] = param.numScoreElements;
|
||||
topOffsetsStartData[imageIdx] = imageIdx * param.numScoreElements;
|
||||
topOffsetsEndData[imageIdx] = (imageIdx + 1) * param.numScoreElements;
|
||||
}
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
cudaError_t EfficientNMSFilterLauncher(EfficientNMSParameters& param, const T* scoresInput, int* topNumData,
|
||||
int* topIndexData, int* topAnchorsData, int* topOffsetsStartData, int* topOffsetsEndData, T* topScoresData,
|
||||
int* topClassData, cudaStream_t stream)
|
||||
{
|
||||
const unsigned int elementsPerBlock = 512;
|
||||
const unsigned int imagesPerBlock = 1;
|
||||
const unsigned int elementBlocks = (param.numScoreElements + elementsPerBlock - 1) / elementsPerBlock;
|
||||
const unsigned int imageBlocks = (param.batchSize + imagesPerBlock - 1) / imagesPerBlock;
|
||||
const dim3 blockSize = {elementsPerBlock, imagesPerBlock, 1};
|
||||
const dim3 gridSize = {elementBlocks, imageBlocks, 1};
|
||||
|
||||
float kernelSelectThreshold = 0.007f;
|
||||
if (param.scoreSigmoid)
|
||||
{
|
||||
// Inverse Sigmoid
|
||||
if (param.scoreThreshold <= 0.f)
|
||||
{
|
||||
param.scoreThreshold = -(1 << 15);
|
||||
}
|
||||
else
|
||||
{
|
||||
param.scoreThreshold = logf(param.scoreThreshold / (1.f - param.scoreThreshold));
|
||||
}
|
||||
kernelSelectThreshold = logf(kernelSelectThreshold / (1.f - kernelSelectThreshold));
|
||||
// Disable Score Bits Optimization
|
||||
param.scoreBits = -1;
|
||||
}
|
||||
|
||||
if (param.scoreThreshold < kernelSelectThreshold)
|
||||
{
|
||||
// A full copy of the buffer is necessary because sorting will scramble the input data otherwise.
|
||||
PLUGIN_CHECK_CUDA(cudaMemcpyAsync(topScoresData, scoresInput,
|
||||
param.batchSize * param.numScoreElements * sizeof(T), cudaMemcpyDeviceToDevice, stream));
|
||||
|
||||
EfficientNMSDenseIndex<T><<<gridSize, blockSize, 0, stream>>>(param, topNumData, topIndexData, topAnchorsData,
|
||||
topOffsetsStartData, topOffsetsEndData, topScoresData, topClassData);
|
||||
}
|
||||
else
|
||||
{
|
||||
EfficientNMSFilter<T><<<gridSize, blockSize, 0, stream>>>(
|
||||
param, scoresInput, topNumData, topIndexData, topAnchorsData, topScoresData, topClassData);
|
||||
|
||||
EfficientNMSFilterSegments<<<1, param.batchSize, 0, stream>>>(
|
||||
param, topNumData, topOffsetsStartData, topOffsetsEndData);
|
||||
}
|
||||
|
||||
return cudaGetLastError();
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
size_t EfficientNMSSortWorkspaceSize(int batchSize, int numScoreElements)
|
||||
{
|
||||
size_t sortedWorkspaceSize = 0;
|
||||
cub::DoubleBuffer<T> keysDB(nullptr, nullptr);
|
||||
cub::DoubleBuffer<int> valuesDB(nullptr, nullptr);
|
||||
cub::DeviceSegmentedRadixSort::SortPairsDescending(nullptr, sortedWorkspaceSize, keysDB, valuesDB,
|
||||
numScoreElements, batchSize, (const int*) nullptr, (const int*) nullptr);
|
||||
return sortedWorkspaceSize;
|
||||
}
|
||||
|
||||
size_t EfficientNMSWorkspaceSize(int batchSize, int numScoreElements, int numClasses, DataType datatype)
|
||||
{
|
||||
size_t total = 0;
|
||||
const size_t align = 256;
|
||||
// Counters
|
||||
// 3 for Filtering
|
||||
// 1 for Output Indexing
|
||||
// C for Max per Class Limiting
|
||||
size_t size = (3 + 1 + numClasses) * batchSize * sizeof(int);
|
||||
total += size + (size % align ? align - (size % align) : 0);
|
||||
// Int Buffers
|
||||
for (int i = 0; i < 4; i++)
|
||||
{
|
||||
size = batchSize * numScoreElements * sizeof(int);
|
||||
total += size + (size % align ? align - (size % align) : 0);
|
||||
}
|
||||
// Float Buffers
|
||||
for (int i = 0; i < 2; i++)
|
||||
{
|
||||
size = batchSize * numScoreElements * dataTypeSize(datatype);
|
||||
total += size + (size % align ? align - (size % align) : 0);
|
||||
}
|
||||
// Sort Workspace
|
||||
if (datatype == DataType::kHALF)
|
||||
{
|
||||
size = EfficientNMSSortWorkspaceSize<__half>(batchSize, numScoreElements);
|
||||
total += size + (size % align ? align - (size % align) : 0);
|
||||
}
|
||||
else if (datatype == DataType::kFLOAT)
|
||||
{
|
||||
size = EfficientNMSSortWorkspaceSize<float>(batchSize, numScoreElements);
|
||||
total += size + (size % align ? align - (size % align) : 0);
|
||||
}
|
||||
|
||||
return total;
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
T* EfficientNMSWorkspace(void* workspace, size_t& offset, size_t elements)
|
||||
{
|
||||
T* buffer = (T*) ((size_t) workspace + offset);
|
||||
size_t align = 256;
|
||||
size_t size = elements * sizeof(T);
|
||||
size_t sizeAligned = size + (size % align ? align - (size % align) : 0);
|
||||
offset += sizeAligned;
|
||||
return buffer;
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
pluginStatus_t EfficientNMSDispatch(EfficientNMSParameters param, const void* boxesInput, const void* scoresInput,
|
||||
const void* anchorsInput, void* numDetectionsOutput, void* nmsBoxesOutput, void* nmsScoresOutput,
|
||||
void* nmsClassesOutput, void* nmsIndicesOutput, void* workspace, cudaStream_t stream)
|
||||
{
|
||||
// Clear Outputs (not all elements will get overwritten by the kernels, so safer to clear everything out)
|
||||
if (param.outputONNXIndices)
|
||||
{
|
||||
CSC(cudaMemsetAsync(nmsIndicesOutput, 0xFF, param.batchSize * param.numOutputBoxes * 3 * sizeof(int), stream), STATUS_FAILURE);
|
||||
}
|
||||
else
|
||||
{
|
||||
CSC(cudaMemsetAsync(numDetectionsOutput, 0x00, param.batchSize * sizeof(int), stream), STATUS_FAILURE);
|
||||
CSC(cudaMemsetAsync(nmsScoresOutput, 0x00, param.batchSize * param.numOutputBoxes * sizeof(T), stream), STATUS_FAILURE);
|
||||
CSC(cudaMemsetAsync(nmsBoxesOutput, 0x00, param.batchSize * param.numOutputBoxes * 4 * sizeof(T), stream), STATUS_FAILURE);
|
||||
CSC(cudaMemsetAsync(nmsClassesOutput, 0x00, param.batchSize * param.numOutputBoxes * sizeof(int), stream), STATUS_FAILURE);
|
||||
}
|
||||
|
||||
// Empty Inputs
|
||||
if (param.numScoreElements < 1)
|
||||
{
|
||||
return STATUS_SUCCESS;
|
||||
}
|
||||
|
||||
// Counters Workspace
|
||||
size_t workspaceOffset = 0;
|
||||
int countersTotalSize = (3 + 1 + param.numClasses) * param.batchSize;
|
||||
int* topNumData = EfficientNMSWorkspace<int>(workspace, workspaceOffset, countersTotalSize);
|
||||
int* topOffsetsStartData = topNumData + param.batchSize;
|
||||
int* topOffsetsEndData = topNumData + 2 * param.batchSize;
|
||||
int* outputIndexData = topNumData + 3 * param.batchSize;
|
||||
int* outputClassData = topNumData + 4 * param.batchSize;
|
||||
CSC(cudaMemsetAsync(topNumData, 0x00, countersTotalSize * sizeof(int), stream), STATUS_FAILURE);
|
||||
cudaError_t status = cudaGetLastError();
|
||||
CSC(status, STATUS_FAILURE);
|
||||
|
||||
// Other Buffers Workspace
|
||||
int* topIndexData
|
||||
= EfficientNMSWorkspace<int>(workspace, workspaceOffset, param.batchSize * param.numScoreElements);
|
||||
int* topClassData
|
||||
= EfficientNMSWorkspace<int>(workspace, workspaceOffset, param.batchSize * param.numScoreElements);
|
||||
int* topAnchorsData
|
||||
= EfficientNMSWorkspace<int>(workspace, workspaceOffset, param.batchSize * param.numScoreElements);
|
||||
int* sortedIndexData
|
||||
= EfficientNMSWorkspace<int>(workspace, workspaceOffset, param.batchSize * param.numScoreElements);
|
||||
T* topScoresData = EfficientNMSWorkspace<T>(workspace, workspaceOffset, param.batchSize * param.numScoreElements);
|
||||
T* sortedScoresData
|
||||
= EfficientNMSWorkspace<T>(workspace, workspaceOffset, param.batchSize * param.numScoreElements);
|
||||
size_t sortedWorkspaceSize = EfficientNMSSortWorkspaceSize<T>(param.batchSize, param.numScoreElements);
|
||||
char* sortedWorkspaceData = EfficientNMSWorkspace<char>(workspace, workspaceOffset, sortedWorkspaceSize);
|
||||
cub::DoubleBuffer<T> scoresDB(topScoresData, sortedScoresData);
|
||||
cub::DoubleBuffer<int> indexDB(topIndexData, sortedIndexData);
|
||||
|
||||
// Kernels
|
||||
status = EfficientNMSFilterLauncher<T>(param, (T*) scoresInput, topNumData, topIndexData, topAnchorsData,
|
||||
topOffsetsStartData, topOffsetsEndData, topScoresData, topClassData, stream);
|
||||
CSC(status, STATUS_FAILURE);
|
||||
|
||||
status = cub::DeviceSegmentedRadixSort::SortPairsDescending(sortedWorkspaceData, sortedWorkspaceSize, scoresDB,
|
||||
indexDB, param.batchSize * param.numScoreElements, param.batchSize, topOffsetsStartData, topOffsetsEndData,
|
||||
param.scoreBits > 0 ? (10 - param.scoreBits) : 0, param.scoreBits > 0 ? 10 : sizeof(T) * 8, stream);
|
||||
CSC(status, STATUS_FAILURE);
|
||||
|
||||
status = EfficientNMSLauncher<T>(param, topNumData, outputIndexData, outputClassData, indexDB.Current(),
|
||||
scoresDB.Current(), topClassData, topAnchorsData, boxesInput, anchorsInput, (int*) numDetectionsOutput,
|
||||
(T*) nmsScoresOutput, (int*) nmsClassesOutput, (int*) nmsIndicesOutput, nmsBoxesOutput, stream);
|
||||
CSC(status, STATUS_FAILURE);
|
||||
|
||||
return STATUS_SUCCESS;
|
||||
}
|
||||
|
||||
pluginStatus_t EfficientNMSInference(EfficientNMSParameters param, const void* boxesInput, const void* scoresInput,
|
||||
const void* anchorsInput, void* numDetectionsOutput, void* nmsBoxesOutput, void* nmsScoresOutput,
|
||||
void* nmsClassesOutput, void* nmsIndicesOutput, void* workspace, cudaStream_t stream)
|
||||
{
|
||||
if (param.datatype == DataType::kFLOAT)
|
||||
{
|
||||
param.scoreBits = -1;
|
||||
return EfficientNMSDispatch<float>(param, boxesInput, scoresInput, anchorsInput, numDetectionsOutput,
|
||||
nmsBoxesOutput, nmsScoresOutput, nmsClassesOutput, nmsIndicesOutput, workspace, stream);
|
||||
}
|
||||
else if (param.datatype == DataType::kHALF)
|
||||
{
|
||||
if (param.scoreBits <= 0 || param.scoreBits > 10)
|
||||
{
|
||||
param.scoreBits = -1;
|
||||
}
|
||||
return EfficientNMSDispatch<__half>(param, boxesInput, scoresInput, anchorsInput, numDetectionsOutput,
|
||||
nmsBoxesOutput, nmsScoresOutput, nmsClassesOutput, nmsIndicesOutput, workspace, stream);
|
||||
}
|
||||
else
|
||||
{
|
||||
return STATUS_NOT_SUPPORTED;
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,261 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#ifndef TRT_EFFICIENT_NMS_INFERENCE_CUH
|
||||
#define TRT_EFFICIENT_NMS_INFERENCE_CUH
|
||||
|
||||
#include <cuda_fp16.h>
|
||||
|
||||
// FP32 Intrinsics
|
||||
|
||||
float __device__ __inline__ exp_mp(const float a)
|
||||
{
|
||||
return __expf(a);
|
||||
}
|
||||
float __device__ __inline__ sigmoid_mp(const float a)
|
||||
{
|
||||
return __frcp_rn(__fadd_rn(1.f, __expf(-a)));
|
||||
}
|
||||
float __device__ __inline__ add_mp(const float a, const float b)
|
||||
{
|
||||
return __fadd_rn(a, b);
|
||||
}
|
||||
float __device__ __inline__ sub_mp(const float a, const float b)
|
||||
{
|
||||
return __fsub_rn(a, b);
|
||||
}
|
||||
float __device__ __inline__ mul_mp(const float a, const float b)
|
||||
{
|
||||
return __fmul_rn(a, b);
|
||||
}
|
||||
bool __device__ __inline__ gt_mp(const float a, const float b)
|
||||
{
|
||||
return a > b;
|
||||
}
|
||||
bool __device__ __inline__ lt_mp(const float a, const float b)
|
||||
{
|
||||
return a < b;
|
||||
}
|
||||
bool __device__ __inline__ lte_mp(const float a, const float b)
|
||||
{
|
||||
return a <= b;
|
||||
}
|
||||
bool __device__ __inline__ gte_mp(const float a, const float b)
|
||||
{
|
||||
return a >= b;
|
||||
}
|
||||
|
||||
#if __CUDA_ARCH__ >= 530
|
||||
|
||||
// FP16 Intrinsics
|
||||
|
||||
__half __device__ __inline__ exp_mp(const __half a)
|
||||
{
|
||||
return hexp(a);
|
||||
}
|
||||
__half __device__ __inline__ sigmoid_mp(const __half a)
|
||||
{
|
||||
return hrcp(__hadd((__half) 1, hexp(__hneg(a))));
|
||||
}
|
||||
__half __device__ __inline__ add_mp(const __half a, const __half b)
|
||||
{
|
||||
return __hadd(a, b);
|
||||
}
|
||||
__half __device__ __inline__ sub_mp(const __half a, const __half b)
|
||||
{
|
||||
return __hsub(a, b);
|
||||
}
|
||||
__half __device__ __inline__ mul_mp(const __half a, const __half b)
|
||||
{
|
||||
return __hmul(a, b);
|
||||
}
|
||||
bool __device__ __inline__ gt_mp(const __half a, const __half b)
|
||||
{
|
||||
return __hgt(a, b);
|
||||
}
|
||||
bool __device__ __inline__ lt_mp(const __half a, const __half b)
|
||||
{
|
||||
return __hlt(a, b);
|
||||
}
|
||||
bool __device__ __inline__ lte_mp(const __half a, const __half b)
|
||||
{
|
||||
return __hle(a, b);
|
||||
}
|
||||
bool __device__ __inline__ gte_mp(const __half a, const __half b)
|
||||
{
|
||||
return __hge(a, b);
|
||||
}
|
||||
|
||||
#else
|
||||
|
||||
// FP16 Fallbacks on older architectures that lack support
|
||||
|
||||
__half __device__ __inline__ exp_mp(const __half a)
|
||||
{
|
||||
return __float2half(exp_mp(__half2float(a)));
|
||||
}
|
||||
__half __device__ __inline__ sigmoid_mp(const __half a)
|
||||
{
|
||||
return __float2half(sigmoid_mp(__half2float(a)));
|
||||
}
|
||||
__half __device__ __inline__ add_mp(const __half a, const __half b)
|
||||
{
|
||||
return __float2half(add_mp(__half2float(a), __half2float(b)));
|
||||
}
|
||||
__half __device__ __inline__ sub_mp(const __half a, const __half b)
|
||||
{
|
||||
return __float2half(sub_mp(__half2float(a), __half2float(b)));
|
||||
}
|
||||
__half __device__ __inline__ mul_mp(const __half a, const __half b)
|
||||
{
|
||||
return __float2half(mul_mp(__half2float(a), __half2float(b)));
|
||||
}
|
||||
bool __device__ __inline__ gt_mp(const __half a, const __half b)
|
||||
{
|
||||
return __float2half(gt_mp(__half2float(a), __half2float(b)));
|
||||
}
|
||||
bool __device__ __inline__ lt_mp(const __half a, const __half b)
|
||||
{
|
||||
return __float2half(lt_mp(__half2float(a), __half2float(b)));
|
||||
}
|
||||
bool __device__ __inline__ lte_mp(const __half a, const __half b)
|
||||
{
|
||||
return __float2half(lte_mp(__half2float(a), __half2float(b)));
|
||||
}
|
||||
bool __device__ __inline__ gte_mp(const __half a, const __half b)
|
||||
{
|
||||
return __float2half(gte_mp(__half2float(a), __half2float(b)));
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
template <typename T>
|
||||
struct __align__(4 * sizeof(T)) BoxCorner;
|
||||
|
||||
template <typename T>
|
||||
struct __align__(4 * sizeof(T)) BoxCenterSize;
|
||||
|
||||
template <typename T>
|
||||
struct __align__(4 * sizeof(T)) BoxCorner
|
||||
{
|
||||
// For NMS/IOU purposes, YXYX coding is identical to XYXY
|
||||
T y1, x1, y2, x2;
|
||||
|
||||
__device__ void reorder()
|
||||
{
|
||||
if (gt_mp(y1, y2))
|
||||
{
|
||||
// Swap values, so y1 < y2
|
||||
y1 = sub_mp(y1, y2);
|
||||
y2 = add_mp(y1, y2);
|
||||
y1 = sub_mp(y2, y1);
|
||||
}
|
||||
if (gt_mp(x1, x2))
|
||||
{
|
||||
// Swap values, so x1 < x2
|
||||
x1 = sub_mp(x1, x2);
|
||||
x2 = add_mp(x1, x2);
|
||||
x1 = sub_mp(x2, x1);
|
||||
}
|
||||
}
|
||||
|
||||
__device__ BoxCorner<T> clip(T low, T high) const
|
||||
{
|
||||
return {lt_mp(y1, low) ? low : (gt_mp(y1, high) ? high : y1),
|
||||
lt_mp(x1, low) ? low : (gt_mp(x1, high) ? high : x1), lt_mp(y2, low) ? low : (gt_mp(y2, high) ? high : y2),
|
||||
lt_mp(x2, low) ? low : (gt_mp(x2, high) ? high : x2)};
|
||||
}
|
||||
|
||||
__device__ BoxCorner<T> decode(BoxCorner<T> anchor) const
|
||||
{
|
||||
return {add_mp(y1, anchor.y1), add_mp(x1, anchor.x1), add_mp(y2, anchor.y2), add_mp(x2, anchor.x2)};
|
||||
}
|
||||
|
||||
__device__ float area() const
|
||||
{
|
||||
T w = sub_mp(x2, x1);
|
||||
T h = sub_mp(y2, y1);
|
||||
if (lte_mp(h, (T) 0))
|
||||
{
|
||||
return 0;
|
||||
}
|
||||
if (lte_mp(w, (T) 0))
|
||||
{
|
||||
return 0;
|
||||
}
|
||||
return (float) h * (float) w;
|
||||
}
|
||||
|
||||
__device__ operator BoxCenterSize<T>() const
|
||||
{
|
||||
T w = sub_mp(x2, x1);
|
||||
T h = sub_mp(y2, y1);
|
||||
return BoxCenterSize<T>{add_mp(y1, mul_mp((T) 0.5, h)), add_mp(x1, mul_mp((T) 0.5, w)), h, w};
|
||||
}
|
||||
|
||||
__device__ static BoxCorner<T> intersect(BoxCorner<T> a, BoxCorner<T> b)
|
||||
{
|
||||
return {gt_mp(a.y1, b.y1) ? a.y1 : b.y1, gt_mp(a.x1, b.x1) ? a.x1 : b.x1, lt_mp(a.y2, b.y2) ? a.y2 : b.y2,
|
||||
lt_mp(a.x2, b.x2) ? a.x2 : b.x2};
|
||||
}
|
||||
};
|
||||
|
||||
template <typename T>
|
||||
struct __align__(4 * sizeof(T)) BoxCenterSize
|
||||
{
|
||||
// For NMS/IOU purposes, YXHW coding is identical to XYWH
|
||||
T y, x, h, w;
|
||||
|
||||
__device__ void reorder() {}
|
||||
|
||||
__device__ BoxCenterSize<T> clip(T low, T high) const
|
||||
{
|
||||
return BoxCenterSize<T>(BoxCorner<T>(*this).clip(low, high));
|
||||
}
|
||||
|
||||
__device__ BoxCenterSize<T> decode(BoxCenterSize<T> anchor) const
|
||||
{
|
||||
return {add_mp(mul_mp(y, anchor.h), anchor.y), add_mp(mul_mp(x, anchor.w), anchor.x),
|
||||
mul_mp(anchor.h, exp_mp(h)), mul_mp(anchor.w, exp_mp(w))};
|
||||
}
|
||||
|
||||
__device__ float area() const
|
||||
{
|
||||
if (h <= (T) 0)
|
||||
{
|
||||
return 0;
|
||||
}
|
||||
if (w <= (T) 0)
|
||||
{
|
||||
return 0;
|
||||
}
|
||||
return (float) h * (float) w;
|
||||
}
|
||||
|
||||
__device__ operator BoxCorner<T>() const
|
||||
{
|
||||
T h2 = mul_mp(h, (T) 0.5);
|
||||
T w2 = mul_mp(w, (T) 0.5);
|
||||
return BoxCorner<T>{sub_mp(y, h2), sub_mp(x, w2), add_mp(y, h2), add_mp(x, w2)};
|
||||
}
|
||||
__device__ static BoxCenterSize<T> intersect(BoxCenterSize<T> a, BoxCenterSize<T> b)
|
||||
{
|
||||
return BoxCenterSize<T>(BoxCorner<T>::intersect(BoxCorner<T>(a), BoxCorner<T>(b)));
|
||||
}
|
||||
};
|
||||
|
||||
#endif
|
||||
@@ -0,0 +1,32 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#ifndef TRT_EFFICIENT_NMS_INFERENCE_H
|
||||
#define TRT_EFFICIENT_NMS_INFERENCE_H
|
||||
|
||||
#include "common/plugin.h"
|
||||
|
||||
#include "efficientNMSParameters.h"
|
||||
|
||||
size_t EfficientNMSWorkspaceSize(
|
||||
int32_t batchSize, int32_t numScoreElements, int32_t numClasses, nvinfer1::DataType datatype);
|
||||
|
||||
pluginStatus_t EfficientNMSInference(nvinfer1::plugin::EfficientNMSParameters param, void const* boxesInput,
|
||||
void const* scoresInput, void const* anchorsInput, void* numDetectionsOutput, void* nmsBoxesOutput,
|
||||
void* nmsScoresOutput, void* nmsClassesOutput, void* nmsIndicesOutput, void* workspace, cudaStream_t stream);
|
||||
|
||||
#endif
|
||||
@@ -0,0 +1,63 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#ifndef TRT_EFFICIENT_NMS_PARAMETERS_H
|
||||
#define TRT_EFFICIENT_NMS_PARAMETERS_H
|
||||
|
||||
#include "common/plugin.h"
|
||||
|
||||
namespace nvinfer1
|
||||
{
|
||||
namespace plugin
|
||||
{
|
||||
|
||||
struct EfficientNMSParameters
|
||||
{
|
||||
// Related to NMS Options
|
||||
float iouThreshold = 0.5F;
|
||||
float scoreThreshold = 0.5F;
|
||||
int32_t numOutputBoxes = 100;
|
||||
int32_t numOutputBoxesPerClass = -1;
|
||||
bool padOutputBoxesPerClass = false;
|
||||
int32_t backgroundClass = -1;
|
||||
bool scoreSigmoid = false;
|
||||
bool clipBoxes = false;
|
||||
int32_t boxCoding = 0;
|
||||
bool classAgnostic = false;
|
||||
|
||||
// Related to NMS Internals
|
||||
int32_t numSelectedBoxes = 4096;
|
||||
int32_t scoreBits = -1;
|
||||
bool outputONNXIndices = false;
|
||||
|
||||
// Related to Tensor Configuration
|
||||
// (These are set by the various plugin configuration methods, no need to define them during plugin creation.)
|
||||
int32_t batchSize = -1;
|
||||
int32_t numClasses = 1;
|
||||
int32_t numBoxElements = -1;
|
||||
int32_t numScoreElements = -1;
|
||||
int32_t numAnchors = -1;
|
||||
bool shareLocation = true;
|
||||
bool shareAnchors = true;
|
||||
bool boxDecoder = false;
|
||||
nvinfer1::DataType datatype = nvinfer1::DataType::kFLOAT;
|
||||
};
|
||||
|
||||
} // namespace plugin
|
||||
} // namespace nvinfer1
|
||||
|
||||
#endif
|
||||
@@ -0,0 +1,530 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#include "efficientNMSPlugin.h"
|
||||
#include "efficientNMSInference.h"
|
||||
|
||||
#include <memory>
|
||||
#include <string_view>
|
||||
|
||||
using namespace nvinfer1;
|
||||
using nvinfer1::plugin::EfficientNMSPlugin;
|
||||
using nvinfer1::plugin::EfficientNMSParameters;
|
||||
using nvinfer1::plugin::EfficientNMSPluginCreator;
|
||||
namespace
|
||||
{
|
||||
using namespace std::string_view_literals;
|
||||
char const* const kEFFICIENT_NMS_PLUGIN_VERSION{"1"};
|
||||
char const* const kEFFICIENT_NMS_PLUGIN_NAME{"EfficientNMS_TRT"};
|
||||
} // namespace
|
||||
|
||||
EfficientNMSPlugin::EfficientNMSPlugin(EfficientNMSParameters param)
|
||||
: mParam(std::move(param))
|
||||
{
|
||||
}
|
||||
|
||||
EfficientNMSPlugin::EfficientNMSPlugin(void const* data, size_t length)
|
||||
{
|
||||
deserialize(static_cast<int8_t const*>(data), length);
|
||||
}
|
||||
|
||||
void EfficientNMSPlugin::deserialize(int8_t const* data, size_t length)
|
||||
{
|
||||
auto const* d{data};
|
||||
mParam = read<EfficientNMSParameters>(d);
|
||||
PLUGIN_VALIDATE(d == data + length);
|
||||
}
|
||||
|
||||
char const* EfficientNMSPlugin::getPluginType() const noexcept
|
||||
{
|
||||
return kEFFICIENT_NMS_PLUGIN_NAME;
|
||||
}
|
||||
|
||||
char const* EfficientNMSPlugin::getPluginVersion() const noexcept
|
||||
{
|
||||
return kEFFICIENT_NMS_PLUGIN_VERSION;
|
||||
}
|
||||
|
||||
int32_t EfficientNMSPlugin::getNbOutputs() const noexcept
|
||||
{
|
||||
if (mParam.outputONNXIndices)
|
||||
{
|
||||
// ONNX NonMaxSuppression Compatibility
|
||||
return 1;
|
||||
}
|
||||
|
||||
// Standard Plugin Implementation
|
||||
return 4;
|
||||
}
|
||||
|
||||
int32_t EfficientNMSPlugin::initialize() noexcept
|
||||
{
|
||||
if (!initialized)
|
||||
{
|
||||
int32_t device;
|
||||
CSC(cudaGetDevice(&device), STATUS_FAILURE);
|
||||
struct cudaDeviceProp properties;
|
||||
CSC(cudaGetDeviceProperties(&properties, device), STATUS_FAILURE);
|
||||
if (properties.regsPerBlock >= 65536)
|
||||
{
|
||||
// Most Devices
|
||||
mParam.numSelectedBoxes = 5000;
|
||||
}
|
||||
else
|
||||
{
|
||||
// Jetson TX1/TX2
|
||||
mParam.numSelectedBoxes = 2000;
|
||||
}
|
||||
initialized = true;
|
||||
}
|
||||
return STATUS_SUCCESS;
|
||||
}
|
||||
|
||||
void EfficientNMSPlugin::terminate() noexcept {}
|
||||
|
||||
size_t EfficientNMSPlugin::getSerializationSize() const noexcept
|
||||
{
|
||||
return sizeof(EfficientNMSParameters);
|
||||
}
|
||||
|
||||
void EfficientNMSPlugin::serialize(void* buffer) const noexcept
|
||||
{
|
||||
char *d = reinterpret_cast<char*>(buffer), *a = d;
|
||||
write(d, mParam);
|
||||
PLUGIN_ASSERT(d == a + getSerializationSize());
|
||||
}
|
||||
|
||||
void EfficientNMSPlugin::destroy() noexcept
|
||||
{
|
||||
delete this;
|
||||
}
|
||||
|
||||
void EfficientNMSPlugin::setPluginNamespace(char const* pluginNamespace) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
mNamespace = pluginNamespace;
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
}
|
||||
|
||||
char const* EfficientNMSPlugin::getPluginNamespace() const noexcept
|
||||
{
|
||||
return mNamespace.c_str();
|
||||
}
|
||||
|
||||
nvinfer1::DataType EfficientNMSPlugin::getOutputDataType(
|
||||
int32_t index, nvinfer1::DataType const* inputTypes, int32_t nbInputs) const noexcept
|
||||
{
|
||||
if (mParam.outputONNXIndices)
|
||||
{
|
||||
// ONNX NMS uses an integer output
|
||||
return nvinfer1::DataType::kINT32;
|
||||
}
|
||||
|
||||
// On standard NMS, num_detections and detection_classes use integer outputs
|
||||
if (index == 0 || index == 3)
|
||||
{
|
||||
return nvinfer1::DataType::kINT32;
|
||||
}
|
||||
// All others should use the same datatype as the input
|
||||
return inputTypes[0];
|
||||
}
|
||||
|
||||
IPluginV2DynamicExt* EfficientNMSPlugin::clone() const noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
auto plugin = std::make_unique<EfficientNMSPlugin>(mParam);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
DimsExprs EfficientNMSPlugin::getOutputDimensions(
|
||||
int32_t outputIndex, DimsExprs const* inputs, int32_t nbInputs, IExprBuilder& exprBuilder) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
DimsExprs out_dim;
|
||||
|
||||
// When pad per class is set, the output size may need to be reduced:
|
||||
// i.e.: outputBoxes = min(outputBoxes, outputBoxesPerClass * numClasses)
|
||||
// As the number of classes may not be static, numOutputBoxes must be a dynamic
|
||||
// expression. The corresponding parameter can not be set at this time, so the
|
||||
// value will be calculated again in configurePlugin() and the param overwritten.
|
||||
IDimensionExpr const* numOutputBoxes = exprBuilder.constant(mParam.numOutputBoxes);
|
||||
if (mParam.padOutputBoxesPerClass && mParam.numOutputBoxesPerClass > 0)
|
||||
{
|
||||
IDimensionExpr const* numOutputBoxesPerClass = exprBuilder.constant(mParam.numOutputBoxesPerClass);
|
||||
IDimensionExpr const* numClasses = inputs[1].d[2];
|
||||
numOutputBoxes = exprBuilder.operation(DimensionOperation::kMIN, *numOutputBoxes,
|
||||
*exprBuilder.operation(DimensionOperation::kPROD, *numOutputBoxesPerClass, *numClasses));
|
||||
}
|
||||
|
||||
if (mParam.outputONNXIndices)
|
||||
{
|
||||
// ONNX NMS
|
||||
PLUGIN_ASSERT(outputIndex == 0);
|
||||
|
||||
// detection_indices
|
||||
out_dim.nbDims = 2;
|
||||
out_dim.d[0] = exprBuilder.operation(DimensionOperation::kPROD, *inputs[0].d[0], *numOutputBoxes);
|
||||
out_dim.d[1] = exprBuilder.constant(3);
|
||||
}
|
||||
else
|
||||
{
|
||||
// Standard NMS
|
||||
PLUGIN_ASSERT(outputIndex >= 0 && outputIndex <= 3);
|
||||
|
||||
// num_detections
|
||||
if (outputIndex == 0)
|
||||
{
|
||||
out_dim.nbDims = 2;
|
||||
out_dim.d[0] = inputs[0].d[0];
|
||||
out_dim.d[1] = exprBuilder.constant(1);
|
||||
}
|
||||
// detection_boxes
|
||||
else if (outputIndex == 1)
|
||||
{
|
||||
out_dim.nbDims = 3;
|
||||
out_dim.d[0] = inputs[0].d[0];
|
||||
out_dim.d[1] = numOutputBoxes;
|
||||
out_dim.d[2] = exprBuilder.constant(4);
|
||||
}
|
||||
// detection_scores: outputIndex == 2
|
||||
// detection_classes: outputIndex == 3
|
||||
else if (outputIndex == 2 || outputIndex == 3)
|
||||
{
|
||||
out_dim.nbDims = 2;
|
||||
out_dim.d[0] = inputs[0].d[0];
|
||||
out_dim.d[1] = numOutputBoxes;
|
||||
}
|
||||
}
|
||||
|
||||
return out_dim;
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return DimsExprs{};
|
||||
}
|
||||
|
||||
bool EfficientNMSPlugin::supportsFormatCombination(
|
||||
int32_t pos, PluginTensorDesc const* inOut, int32_t nbInputs, int32_t nbOutputs) noexcept
|
||||
{
|
||||
if (inOut[pos].format != PluginFormat::kLINEAR)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if (mParam.outputONNXIndices)
|
||||
{
|
||||
PLUGIN_ASSERT(nbInputs == 2);
|
||||
PLUGIN_ASSERT(nbOutputs == 1);
|
||||
|
||||
// detection_indices output: int32_t
|
||||
if (pos == 2)
|
||||
{
|
||||
return inOut[pos].type == DataType::kINT32;
|
||||
}
|
||||
|
||||
// boxes and scores input: fp32 or fp16
|
||||
return (inOut[pos].type == DataType::kHALF || inOut[pos].type == DataType::kFLOAT)
|
||||
&& (inOut[0].type == inOut[pos].type);
|
||||
}
|
||||
|
||||
PLUGIN_ASSERT(nbInputs == 2 || nbInputs == 3);
|
||||
PLUGIN_ASSERT(nbOutputs == 4);
|
||||
if (nbInputs == 2)
|
||||
{
|
||||
PLUGIN_ASSERT(0 <= pos && pos <= 5);
|
||||
}
|
||||
if (nbInputs == 3)
|
||||
{
|
||||
PLUGIN_ASSERT(0 <= pos && pos <= 6);
|
||||
}
|
||||
|
||||
// num_detections and detection_classes output: int32_t
|
||||
int32_t const posOut = pos - nbInputs;
|
||||
if (posOut == 0 || posOut == 3)
|
||||
{
|
||||
return inOut[pos].type == DataType::kINT32 && inOut[pos].format == PluginFormat::kLINEAR;
|
||||
}
|
||||
|
||||
// all other inputs/outputs: fp32 or fp16
|
||||
return (inOut[pos].type == DataType::kHALF || inOut[pos].type == DataType::kFLOAT)
|
||||
&& (inOut[0].type == inOut[pos].type);
|
||||
}
|
||||
|
||||
void EfficientNMSPlugin::configurePlugin(
|
||||
DynamicPluginTensorDesc const* in, int32_t nbInputs, DynamicPluginTensorDesc const* out, int32_t nbOutputs) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
if (mParam.outputONNXIndices)
|
||||
{
|
||||
// Accepts two inputs
|
||||
// [0] boxes, [1] scores
|
||||
PLUGIN_ASSERT(nbInputs == 2);
|
||||
PLUGIN_ASSERT(nbOutputs == 1);
|
||||
}
|
||||
else
|
||||
{
|
||||
// Accepts two or three inputs
|
||||
// If two inputs: [0] boxes, [1] scores
|
||||
// If three inputs: [0] boxes, [1] scores, [2] anchors
|
||||
PLUGIN_ASSERT(nbInputs == 2 || nbInputs == 3);
|
||||
PLUGIN_ASSERT(nbOutputs == 4);
|
||||
}
|
||||
mParam.datatype = in[0].desc.type;
|
||||
|
||||
// Shape of scores input should be
|
||||
// [batch_size, num_boxes, num_classes] or [batch_size, num_boxes, num_classes, 1]
|
||||
PLUGIN_ASSERT(in[1].desc.dims.nbDims == 3 || (in[1].desc.dims.nbDims == 4 && in[1].desc.dims.d[3] == 1));
|
||||
mParam.numScoreElements = in[1].desc.dims.d[1] * in[1].desc.dims.d[2];
|
||||
mParam.numClasses = in[1].desc.dims.d[2];
|
||||
|
||||
// When pad per class is set, the total output boxes size may need to be reduced.
|
||||
// This operation is also done in getOutputDimension(), but for dynamic shapes, the
|
||||
// numOutputBoxes param can't be set until the number of classes is fully known here.
|
||||
if (mParam.padOutputBoxesPerClass && mParam.numOutputBoxesPerClass > 0)
|
||||
{
|
||||
if (mParam.numOutputBoxesPerClass * mParam.numClasses < mParam.numOutputBoxes)
|
||||
{
|
||||
mParam.numOutputBoxes = mParam.numOutputBoxesPerClass * mParam.numClasses;
|
||||
}
|
||||
}
|
||||
|
||||
// Shape of boxes input should be
|
||||
// [batch_size, num_boxes, 4] or [batch_size, num_boxes, 1, 4] or [batch_size, num_boxes, num_classes, 4]
|
||||
PLUGIN_ASSERT(in[0].desc.dims.nbDims == 3 || in[0].desc.dims.nbDims == 4);
|
||||
if (in[0].desc.dims.nbDims == 3)
|
||||
{
|
||||
PLUGIN_ASSERT(in[0].desc.dims.d[2] == 4);
|
||||
mParam.shareLocation = true;
|
||||
mParam.numBoxElements = in[0].desc.dims.d[1] * in[0].desc.dims.d[2];
|
||||
}
|
||||
else
|
||||
{
|
||||
mParam.shareLocation = (in[0].desc.dims.d[2] == 1);
|
||||
PLUGIN_ASSERT(in[0].desc.dims.d[2] == mParam.numClasses || mParam.shareLocation);
|
||||
PLUGIN_ASSERT(in[0].desc.dims.d[3] == 4);
|
||||
mParam.numBoxElements = in[0].desc.dims.d[1] * in[0].desc.dims.d[2] * in[0].desc.dims.d[3];
|
||||
}
|
||||
mParam.numAnchors = in[0].desc.dims.d[1];
|
||||
|
||||
if (nbInputs == 2)
|
||||
{
|
||||
// Only two inputs are used, disable the fused box decoder
|
||||
mParam.boxDecoder = false;
|
||||
}
|
||||
if (nbInputs == 3)
|
||||
{
|
||||
// All three inputs are used, enable the box decoder
|
||||
// Shape of anchors input should be
|
||||
// Constant shape: [1, numAnchors, 4] or [batch_size, numAnchors, 4]
|
||||
PLUGIN_ASSERT(in[2].desc.dims.nbDims == 3);
|
||||
mParam.boxDecoder = true;
|
||||
mParam.shareAnchors = (in[2].desc.dims.d[0] == 1);
|
||||
}
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
}
|
||||
|
||||
size_t EfficientNMSPlugin::getWorkspaceSize(
|
||||
PluginTensorDesc const* inputs, int32_t nbInputs, PluginTensorDesc const* outputs, int32_t nbOutputs) const noexcept
|
||||
{
|
||||
int32_t batchSize = inputs[1].dims.d[0];
|
||||
int32_t numScoreElements = inputs[1].dims.d[1] * inputs[1].dims.d[2];
|
||||
int32_t numClasses = inputs[1].dims.d[2];
|
||||
return EfficientNMSWorkspaceSize(batchSize, numScoreElements, numClasses, mParam.datatype);
|
||||
}
|
||||
|
||||
int32_t EfficientNMSPlugin::enqueue(PluginTensorDesc const* inputDesc, PluginTensorDesc const* /* outputDesc */,
|
||||
void const* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
PLUGIN_VALIDATE(inputDesc != nullptr && inputs != nullptr && outputs != nullptr && workspace != nullptr);
|
||||
|
||||
mParam.batchSize = inputDesc[0].dims.d[0];
|
||||
|
||||
if (mParam.outputONNXIndices)
|
||||
{
|
||||
// ONNX NonMaxSuppression Op Support
|
||||
void const* const boxesInput = inputs[0];
|
||||
void const* const scoresInput = inputs[1];
|
||||
|
||||
void* nmsIndicesOutput = outputs[0];
|
||||
|
||||
return EfficientNMSInference(mParam, boxesInput, scoresInput, nullptr, nullptr, nullptr, nullptr, nullptr,
|
||||
nmsIndicesOutput, workspace, stream);
|
||||
}
|
||||
|
||||
// Standard NMS Operation
|
||||
void const* const boxesInput = inputs[0];
|
||||
void const* const scoresInput = inputs[1];
|
||||
void const* const anchorsInput = mParam.boxDecoder ? inputs[2] : nullptr;
|
||||
|
||||
void* numDetectionsOutput = outputs[0];
|
||||
void* nmsBoxesOutput = outputs[1];
|
||||
void* nmsScoresOutput = outputs[2];
|
||||
void* nmsClassesOutput = outputs[3];
|
||||
|
||||
return EfficientNMSInference(mParam, boxesInput, scoresInput, anchorsInput, numDetectionsOutput, nmsBoxesOutput,
|
||||
nmsScoresOutput, nmsClassesOutput, nullptr, workspace, stream);
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return -1;
|
||||
}
|
||||
|
||||
// Standard NMS Plugin Operation
|
||||
|
||||
EfficientNMSPluginCreator::EfficientNMSPluginCreator()
|
||||
: mParam{}
|
||||
{
|
||||
mPluginAttributes.clear();
|
||||
mPluginAttributes.emplace_back(PluginField("score_threshold", nullptr, PluginFieldType::kFLOAT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("iou_threshold", nullptr, PluginFieldType::kFLOAT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("max_output_boxes", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("background_class", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("score_activation", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("class_agnostic", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("box_coding", nullptr, PluginFieldType::kINT32, 1));
|
||||
mFC.nbFields = mPluginAttributes.size();
|
||||
mFC.fields = mPluginAttributes.data();
|
||||
}
|
||||
|
||||
char const* EfficientNMSPluginCreator::getPluginName() const noexcept
|
||||
{
|
||||
return kEFFICIENT_NMS_PLUGIN_NAME;
|
||||
}
|
||||
|
||||
char const* EfficientNMSPluginCreator::getPluginVersion() const noexcept
|
||||
{
|
||||
return kEFFICIENT_NMS_PLUGIN_VERSION;
|
||||
}
|
||||
|
||||
PluginFieldCollection const* EfficientNMSPluginCreator::getFieldNames() noexcept
|
||||
{
|
||||
return &mFC;
|
||||
}
|
||||
|
||||
IPluginV2DynamicExt* EfficientNMSPluginCreator::createPlugin(char const* name, PluginFieldCollection const* fc) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
PLUGIN_VALIDATE(fc != nullptr);
|
||||
PluginField const* fields = fc->fields;
|
||||
PLUGIN_VALIDATE(fields != nullptr);
|
||||
plugin::validateRequiredAttributesExist({"score_threshold", "iou_threshold", "max_output_boxes",
|
||||
"background_class", "score_activation", "box_coding"},
|
||||
fc);
|
||||
for (int32_t i{0}; i < fc->nbFields; ++i)
|
||||
{
|
||||
std::string_view const attrName = fields[i].name;
|
||||
if (attrName == "score_threshold"sv)
|
||||
{
|
||||
PLUGIN_VALIDATE(fields[i].type == PluginFieldType::kFLOAT32);
|
||||
auto const scoreThreshold = *(static_cast<float const*>(fields[i].data));
|
||||
PLUGIN_VALIDATE(scoreThreshold >= 0.0F);
|
||||
mParam.scoreThreshold = scoreThreshold;
|
||||
}
|
||||
if (attrName == "iou_threshold"sv)
|
||||
{
|
||||
PLUGIN_VALIDATE(fields[i].type == PluginFieldType::kFLOAT32);
|
||||
auto const iouThreshold = *(static_cast<float const*>(fields[i].data));
|
||||
PLUGIN_VALIDATE(iouThreshold > 0.0F);
|
||||
mParam.iouThreshold = iouThreshold;
|
||||
}
|
||||
if (attrName == "max_output_boxes"sv)
|
||||
{
|
||||
PLUGIN_VALIDATE(fields[i].type == PluginFieldType::kINT32);
|
||||
auto const numOutputBoxes = *(static_cast<int32_t const*>(fields[i].data));
|
||||
PLUGIN_VALIDATE(numOutputBoxes > 0);
|
||||
mParam.numOutputBoxes = numOutputBoxes;
|
||||
}
|
||||
if (attrName == "background_class"sv)
|
||||
{
|
||||
PLUGIN_VALIDATE(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.backgroundClass = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "score_activation"sv)
|
||||
{
|
||||
auto const scoreSigmoid = *(static_cast<int32_t const*>(fields[i].data));
|
||||
PLUGIN_VALIDATE(scoreSigmoid == 0 || scoreSigmoid == 1);
|
||||
mParam.scoreSigmoid = static_cast<bool>(scoreSigmoid);
|
||||
}
|
||||
if (attrName == "class_agnostic"sv)
|
||||
{
|
||||
auto const classAgnostic = *(static_cast<int32_t const*>(fields[i].data));
|
||||
PLUGIN_VALIDATE(classAgnostic == 0 || classAgnostic == 1);
|
||||
mParam.classAgnostic = static_cast<bool>(classAgnostic);
|
||||
}
|
||||
if (attrName == "box_coding"sv)
|
||||
{
|
||||
PLUGIN_VALIDATE(fields[i].type == PluginFieldType::kINT32);
|
||||
auto const boxCoding = *(static_cast<int32_t const*>(fields[i].data));
|
||||
PLUGIN_VALIDATE(boxCoding == 0 || boxCoding == 1);
|
||||
mParam.boxCoding = boxCoding;
|
||||
}
|
||||
}
|
||||
|
||||
auto plugin = std::make_unique<EfficientNMSPlugin>(mParam);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
IPluginV2DynamicExt* EfficientNMSPluginCreator::deserializePlugin(
|
||||
char const* name, void const* serialData, size_t serialLength) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
// This object will be deleted when the network is destroyed, which will
|
||||
// call EfficientNMSPlugin::destroy()
|
||||
auto plugin = std::make_unique<EfficientNMSPlugin>(serialData, serialLength);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
@@ -0,0 +1,102 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
#ifndef TRT_EFFICIENT_NMS_PLUGIN_H
|
||||
#define TRT_EFFICIENT_NMS_PLUGIN_H
|
||||
|
||||
#include <vector>
|
||||
|
||||
#include "common/plugin.h"
|
||||
#include "efficientNMSPlugin/efficientNMSParameters.h"
|
||||
|
||||
namespace nvinfer1
|
||||
{
|
||||
namespace plugin
|
||||
{
|
||||
|
||||
class TRT_DEPRECATED_BECAUSE("Deprecated since TRT 10.12. Use INMSLayer instead.") EfficientNMSPlugin
|
||||
: public IPluginV2DynamicExt
|
||||
{
|
||||
public:
|
||||
explicit EfficientNMSPlugin(EfficientNMSParameters param);
|
||||
EfficientNMSPlugin(void const* data, size_t length);
|
||||
~EfficientNMSPlugin() override = default;
|
||||
|
||||
// IPluginV2 methods
|
||||
char const* getPluginType() const noexcept override;
|
||||
char const* getPluginVersion() const noexcept override;
|
||||
int32_t getNbOutputs() const noexcept override;
|
||||
int32_t initialize() noexcept override;
|
||||
void terminate() noexcept override;
|
||||
size_t getSerializationSize() const noexcept override;
|
||||
void serialize(void* buffer) const noexcept override;
|
||||
void destroy() noexcept override;
|
||||
void setPluginNamespace(char const* libNamespace) noexcept override;
|
||||
char const* getPluginNamespace() const noexcept override;
|
||||
|
||||
// IPluginV2Ext methods
|
||||
nvinfer1::DataType getOutputDataType(
|
||||
int32_t index, nvinfer1::DataType const* inputType, int32_t nbInputs) const noexcept override;
|
||||
|
||||
// IPluginV2DynamicExt methods
|
||||
IPluginV2DynamicExt* clone() const noexcept override;
|
||||
DimsExprs getOutputDimensions(
|
||||
int32_t outputIndex, DimsExprs const* inputs, int32_t nbInputs, IExprBuilder& exprBuilder) noexcept override;
|
||||
bool supportsFormatCombination(
|
||||
int32_t pos, PluginTensorDesc const* inOut, int32_t nbInputs, int32_t nbOutputs) noexcept override;
|
||||
void configurePlugin(DynamicPluginTensorDesc const* in, int32_t nbInputs, DynamicPluginTensorDesc const* out,
|
||||
int32_t nbOutputs) noexcept override;
|
||||
size_t getWorkspaceSize(PluginTensorDesc const* inputs, int32_t nbInputs, PluginTensorDesc const* outputs,
|
||||
int32_t nbOutputs) const noexcept override;
|
||||
int32_t enqueue(PluginTensorDesc const* inputDesc, PluginTensorDesc const* outputDesc, void const* const* inputs,
|
||||
void* const* outputs, void* workspace, cudaStream_t stream) noexcept override;
|
||||
|
||||
protected:
|
||||
EfficientNMSParameters mParam{};
|
||||
bool initialized{false};
|
||||
std::string mNamespace;
|
||||
|
||||
private:
|
||||
void deserialize(int8_t const* data, size_t length);
|
||||
};
|
||||
|
||||
// Standard NMS Plugin Operation
|
||||
class TRT_DEPRECATED_BECAUSE("Deprecated since TRT 10.12. Use INMSLayer instead.") EfficientNMSPluginCreator
|
||||
: public nvinfer1::pluginInternal::BaseCreator
|
||||
{
|
||||
public:
|
||||
EfficientNMSPluginCreator();
|
||||
~EfficientNMSPluginCreator() override = default;
|
||||
|
||||
char const* getPluginName() const noexcept override;
|
||||
char const* getPluginVersion() const noexcept override;
|
||||
PluginFieldCollection const* getFieldNames() noexcept override;
|
||||
|
||||
IPluginV2DynamicExt* createPlugin(char const* name, PluginFieldCollection const* fc) noexcept override;
|
||||
IPluginV2DynamicExt* deserializePlugin(
|
||||
char const* name, void const* serialData, size_t serialLength) noexcept override;
|
||||
|
||||
protected:
|
||||
PluginFieldCollection mFC;
|
||||
EfficientNMSParameters mParam;
|
||||
std::vector<PluginField> mPluginAttributes;
|
||||
std::string mPluginName;
|
||||
};
|
||||
|
||||
} // namespace plugin
|
||||
} // namespace nvinfer1
|
||||
|
||||
#endif // TRT_EFFICIENT_NMS_PLUGIN_H
|
||||
@@ -0,0 +1,23 @@
|
||||
#
|
||||
# SPDX-FileCopyrightText: Copyright (c) 1993-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
#
|
||||
|
||||
add_plugin_source(
|
||||
efficientNMSExplicitTFTRTPlugin.cpp
|
||||
efficientNMSExplicitTFTRTPlugin.h
|
||||
efficientNMSImplicitTFTRTPlugin.cpp
|
||||
efficientNMSImplicitTFTRTPlugin.h
|
||||
)
|
||||
@@ -0,0 +1,171 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#include "efficientNMSExplicitTFTRTPlugin.h"
|
||||
#include "efficientNMSPlugin/efficientNMSInference.h"
|
||||
|
||||
#include <memory>
|
||||
#include <string_view>
|
||||
|
||||
// This plugin provides CombinedNMS op compatibility for TF-TRT in Explicit Batch
|
||||
// and Dymamic Shape modes
|
||||
|
||||
using namespace nvinfer1;
|
||||
using namespace nvinfer1::plugin;
|
||||
using nvinfer1::plugin::EfficientNMSExplicitTFTRTPlugin;
|
||||
using nvinfer1::plugin::EfficientNMSExplicitTFTRTPluginCreator;
|
||||
|
||||
namespace
|
||||
{
|
||||
const char* EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_VERSION{"1"};
|
||||
const char* EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_NAME{"EfficientNMS_Explicit_TF_TRT"};
|
||||
} // namespace
|
||||
|
||||
EfficientNMSExplicitTFTRTPlugin::EfficientNMSExplicitTFTRTPlugin(EfficientNMSParameters param)
|
||||
: EfficientNMSPlugin(std::move(param))
|
||||
{
|
||||
}
|
||||
|
||||
EfficientNMSExplicitTFTRTPlugin::EfficientNMSExplicitTFTRTPlugin(const void* data, size_t length)
|
||||
: EfficientNMSPlugin(data, length)
|
||||
{
|
||||
}
|
||||
|
||||
const char* EfficientNMSExplicitTFTRTPlugin::getPluginType() const noexcept
|
||||
{
|
||||
return EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_NAME;
|
||||
}
|
||||
|
||||
const char* EfficientNMSExplicitTFTRTPlugin::getPluginVersion() const noexcept
|
||||
{
|
||||
return EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_VERSION;
|
||||
}
|
||||
|
||||
IPluginV2DynamicExt* EfficientNMSExplicitTFTRTPlugin::clone() const noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
auto plugin = std::make_unique<EfficientNMSExplicitTFTRTPlugin>(mParam);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
EfficientNMSExplicitTFTRTPluginCreator::EfficientNMSExplicitTFTRTPluginCreator()
|
||||
: mParam{}
|
||||
{
|
||||
mPluginAttributes.clear();
|
||||
mPluginAttributes.emplace_back(PluginField("max_output_size_per_class", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("max_total_size", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("iou_threshold", nullptr, PluginFieldType::kFLOAT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("score_threshold", nullptr, PluginFieldType::kFLOAT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("pad_per_class", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("clip_boxes", nullptr, PluginFieldType::kINT32, 1));
|
||||
mFC.nbFields = mPluginAttributes.size();
|
||||
mFC.fields = mPluginAttributes.data();
|
||||
}
|
||||
|
||||
const char* EfficientNMSExplicitTFTRTPluginCreator::getPluginName() const noexcept
|
||||
{
|
||||
return EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_NAME;
|
||||
}
|
||||
|
||||
const char* EfficientNMSExplicitTFTRTPluginCreator::getPluginVersion() const noexcept
|
||||
{
|
||||
return EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_VERSION;
|
||||
}
|
||||
|
||||
const PluginFieldCollection* EfficientNMSExplicitTFTRTPluginCreator::getFieldNames() noexcept
|
||||
{
|
||||
return &mFC;
|
||||
}
|
||||
|
||||
IPluginV2DynamicExt* EfficientNMSExplicitTFTRTPluginCreator::createPlugin(
|
||||
char const* name, PluginFieldCollection const* fc) noexcept
|
||||
{
|
||||
using namespace std::string_view_literals;
|
||||
try
|
||||
{
|
||||
PluginField const* fields = fc->fields;
|
||||
for (int32_t i = 0; i < fc->nbFields; ++i)
|
||||
{
|
||||
std::string_view const attrName = fields[i].name;
|
||||
if (attrName == "max_output_size_per_class"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.numOutputBoxesPerClass = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "max_total_size"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.numOutputBoxes = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "iou_threshold"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kFLOAT32);
|
||||
mParam.iouThreshold = *(static_cast<float const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "score_threshold"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kFLOAT32);
|
||||
mParam.scoreThreshold = *(static_cast<float const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "pad_per_class"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.padOutputBoxesPerClass = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "clip_boxes"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.clipBoxes = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
}
|
||||
|
||||
auto plugin = std::make_unique<EfficientNMSExplicitTFTRTPlugin>(mParam);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
IPluginV2DynamicExt* EfficientNMSExplicitTFTRTPluginCreator::deserializePlugin(
|
||||
const char* name, const void* serialData, size_t serialLength) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
// This object will be deleted when the network is destroyed, which will
|
||||
// call EfficientNMSPlugin::destroy()
|
||||
auto plugin = std::make_unique<EfficientNMSExplicitTFTRTPlugin>(serialData, serialLength);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
@@ -0,0 +1,73 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
#ifndef TRT_EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_H
|
||||
#define TRT_EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_H
|
||||
|
||||
#include <vector>
|
||||
|
||||
#include "common/plugin.h"
|
||||
#include "efficientNMSPlugin/efficientNMSParameters.h"
|
||||
#include "efficientNMSPlugin/efficientNMSPlugin.h"
|
||||
|
||||
// This plugin provides CombinedNMS op compatibility for TF-TRT in Explicit Batch
|
||||
// and Dymamic Shape modes
|
||||
|
||||
namespace nvinfer1
|
||||
{
|
||||
namespace plugin
|
||||
{
|
||||
|
||||
class TRT_DEPRECATED_BECAUSE("TF-TRT integration is obsolete. No alternatives planned.") EfficientNMSExplicitTFTRTPlugin
|
||||
: public EfficientNMSPlugin
|
||||
{
|
||||
public:
|
||||
explicit EfficientNMSExplicitTFTRTPlugin(EfficientNMSParameters param);
|
||||
EfficientNMSExplicitTFTRTPlugin(const void* data, size_t length);
|
||||
~EfficientNMSExplicitTFTRTPlugin() override = default;
|
||||
const char* getPluginType() const noexcept override;
|
||||
const char* getPluginVersion() const noexcept override;
|
||||
IPluginV2DynamicExt* clone() const noexcept override;
|
||||
};
|
||||
|
||||
// TF-TRT CombinedNMS Op Compatibility
|
||||
class TRT_DEPRECATED_BECAUSE(
|
||||
"TF-TRT integration is obsolete. No alternatives planned.") EfficientNMSExplicitTFTRTPluginCreator
|
||||
: public nvinfer1::pluginInternal::BaseCreator
|
||||
{
|
||||
public:
|
||||
EfficientNMSExplicitTFTRTPluginCreator();
|
||||
~EfficientNMSExplicitTFTRTPluginCreator() override = default;
|
||||
|
||||
const char* getPluginName() const noexcept override;
|
||||
const char* getPluginVersion() const noexcept override;
|
||||
const PluginFieldCollection* getFieldNames() noexcept override;
|
||||
|
||||
IPluginV2DynamicExt* createPlugin(const char* name, const PluginFieldCollection* fc) noexcept override;
|
||||
IPluginV2DynamicExt* deserializePlugin(
|
||||
const char* name, const void* serialData, size_t serialLength) noexcept override;
|
||||
|
||||
protected:
|
||||
PluginFieldCollection mFC;
|
||||
EfficientNMSParameters mParam;
|
||||
std::vector<PluginField> mPluginAttributes;
|
||||
std::string mPluginName;
|
||||
};
|
||||
|
||||
} // namespace plugin
|
||||
} // namespace nvinfer1
|
||||
|
||||
#endif // TRT_EFFICIENT_NMS_EXPLICIT_TFTRT_PLUGIN_H
|
||||
@@ -0,0 +1,394 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#include "efficientNMSImplicitTFTRTPlugin.h"
|
||||
#include "efficientNMSPlugin/efficientNMSInference.h"
|
||||
|
||||
#include <memory>
|
||||
#include <string_view>
|
||||
|
||||
// This plugin provides CombinedNMS op compatibility for TF-TRT in Implicit Batch
|
||||
// mode for legacy back-compatibilty
|
||||
|
||||
using namespace nvinfer1;
|
||||
using namespace nvinfer1::plugin;
|
||||
using nvinfer1::plugin::EfficientNMSParameters;
|
||||
using nvinfer1::plugin::EfficientNMSImplicitTFTRTPlugin;
|
||||
using nvinfer1::plugin::EfficientNMSImplicitTFTRTPluginCreator;
|
||||
|
||||
namespace
|
||||
{
|
||||
const char* EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_VERSION{"1"};
|
||||
const char* EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_NAME{"EfficientNMS_Implicit_TF_TRT"};
|
||||
} // namespace
|
||||
|
||||
EfficientNMSImplicitTFTRTPlugin::EfficientNMSImplicitTFTRTPlugin(EfficientNMSParameters param)
|
||||
: mParam(std::move(param))
|
||||
{
|
||||
}
|
||||
|
||||
EfficientNMSImplicitTFTRTPlugin::EfficientNMSImplicitTFTRTPlugin(void const* data, size_t length)
|
||||
{
|
||||
deserialize(static_cast<int8_t const*>(data), length);
|
||||
}
|
||||
|
||||
void EfficientNMSImplicitTFTRTPlugin::deserialize(int8_t const* data, size_t length)
|
||||
{
|
||||
auto const* d{data};
|
||||
mParam = read<EfficientNMSParameters>(d);
|
||||
PLUGIN_ASSERT(d == data + length);
|
||||
}
|
||||
|
||||
const char* EfficientNMSImplicitTFTRTPlugin::getPluginType() const noexcept
|
||||
{
|
||||
return EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_NAME;
|
||||
}
|
||||
|
||||
const char* EfficientNMSImplicitTFTRTPlugin::getPluginVersion() const noexcept
|
||||
{
|
||||
return EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_VERSION;
|
||||
}
|
||||
|
||||
int32_t EfficientNMSImplicitTFTRTPlugin::getNbOutputs() const noexcept
|
||||
{
|
||||
return 4;
|
||||
}
|
||||
|
||||
int32_t EfficientNMSImplicitTFTRTPlugin::initialize() noexcept
|
||||
{
|
||||
return STATUS_SUCCESS;
|
||||
}
|
||||
|
||||
void EfficientNMSImplicitTFTRTPlugin::terminate() noexcept {}
|
||||
|
||||
size_t EfficientNMSImplicitTFTRTPlugin::getSerializationSize() const noexcept
|
||||
{
|
||||
return sizeof(EfficientNMSParameters);
|
||||
}
|
||||
|
||||
void EfficientNMSImplicitTFTRTPlugin::serialize(void* buffer) const noexcept
|
||||
{
|
||||
char *d = reinterpret_cast<char*>(buffer), *a = d;
|
||||
write(d, mParam);
|
||||
PLUGIN_ASSERT(d == a + getSerializationSize());
|
||||
}
|
||||
|
||||
void EfficientNMSImplicitTFTRTPlugin::destroy() noexcept
|
||||
{
|
||||
delete this;
|
||||
}
|
||||
|
||||
void EfficientNMSImplicitTFTRTPlugin::setPluginNamespace(const char* pluginNamespace) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
mNamespace = pluginNamespace;
|
||||
}
|
||||
catch (const std::exception& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
}
|
||||
|
||||
const char* EfficientNMSImplicitTFTRTPlugin::getPluginNamespace() const noexcept
|
||||
{
|
||||
return mNamespace.c_str();
|
||||
}
|
||||
|
||||
Dims EfficientNMSImplicitTFTRTPlugin::getOutputDimensions(
|
||||
int32_t outputIndex, const Dims* inputs, int32_t nbInputs) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
Dims outDim{};
|
||||
|
||||
// When pad per class is set, the output size may need to be reduced:
|
||||
// i.e.: outputBoxes = min(outputBoxes, outputBoxesPerClass * numClasses)
|
||||
PLUGIN_ASSERT(inputs[1].nbDims == 2);
|
||||
if (mParam.padOutputBoxesPerClass && mParam.numOutputBoxesPerClass > 0)
|
||||
{
|
||||
const int32_t numClasses = inputs[1].d[1];
|
||||
if (mParam.numOutputBoxesPerClass * numClasses < mParam.numOutputBoxes)
|
||||
{
|
||||
mParam.numOutputBoxes = mParam.numOutputBoxesPerClass * numClasses;
|
||||
}
|
||||
}
|
||||
|
||||
// Standard NMS
|
||||
PLUGIN_ASSERT(outputIndex >= 0 && outputIndex <= 3);
|
||||
|
||||
// num_detections
|
||||
if (outputIndex == 0)
|
||||
{
|
||||
outDim.nbDims = 0;
|
||||
outDim.d[0] = 0;
|
||||
}
|
||||
// detection_boxes
|
||||
else if (outputIndex == 1)
|
||||
{
|
||||
outDim.nbDims = 2;
|
||||
outDim.d[0] = mParam.numOutputBoxes;
|
||||
outDim.d[1] = 4;
|
||||
}
|
||||
// detection_scores: outputIndex == 2
|
||||
// detection_classes: outputIndex == 3
|
||||
else if (outputIndex == 2 || outputIndex == 3)
|
||||
{
|
||||
outDim.nbDims = 1;
|
||||
outDim.d[0] = mParam.numOutputBoxes;
|
||||
}
|
||||
|
||||
return outDim;
|
||||
}
|
||||
catch (const std::exception& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return Dims{};
|
||||
}
|
||||
|
||||
size_t EfficientNMSImplicitTFTRTPlugin::getWorkspaceSize(int32_t maxBatchSize) const noexcept
|
||||
{
|
||||
return EfficientNMSWorkspaceSize(maxBatchSize, mParam.numScoreElements, mParam.numClasses, mParam.datatype);
|
||||
}
|
||||
|
||||
int32_t EfficientNMSImplicitTFTRTPlugin::enqueue(int32_t batchSize, void const* const* inputs,
|
||||
EfficientNMSImplicitTFTRTOutputsDataType outputs, void* workspace, cudaStream_t stream) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
mParam.batchSize = batchSize;
|
||||
|
||||
void const* const boxesInput = inputs[0];
|
||||
void const* const scoresInput = inputs[1];
|
||||
void const* const anchorsInput = nullptr;
|
||||
|
||||
void* numDetectionsOutput = outputs[0];
|
||||
void* nmsBoxesOutput = outputs[1];
|
||||
void* nmsScoresOutput = outputs[2];
|
||||
void* nmsClassesOutput = outputs[3];
|
||||
|
||||
return EfficientNMSInference(mParam, boxesInput, scoresInput, anchorsInput, numDetectionsOutput, nmsBoxesOutput,
|
||||
nmsScoresOutput, nmsClassesOutput, nullptr, workspace, stream);
|
||||
}
|
||||
catch (const std::exception& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return -1;
|
||||
}
|
||||
|
||||
DataType EfficientNMSImplicitTFTRTPlugin::getOutputDataType(
|
||||
int32_t index, DataType const* inputTypes, int32_t nbInputs) const noexcept
|
||||
{
|
||||
// num_detections and detection_classes use integer outputs
|
||||
if (index == 0 || index == 3)
|
||||
{
|
||||
return DataType::kINT32;
|
||||
}
|
||||
// All others should use the same datatype as the input
|
||||
return inputTypes[0];
|
||||
}
|
||||
|
||||
IPluginV2IOExt* EfficientNMSImplicitTFTRTPlugin::clone() const noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
auto plugin = std::make_unique<EfficientNMSImplicitTFTRTPlugin>(mParam);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
bool EfficientNMSImplicitTFTRTPlugin::supportsFormatCombination(
|
||||
int32_t pos, PluginTensorDesc const* inOut, int32_t nbInputs, int32_t nbOutputs) const noexcept
|
||||
{
|
||||
if (inOut[pos].format != PluginFormat::kLINEAR)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
PLUGIN_ASSERT(nbInputs == 2);
|
||||
PLUGIN_ASSERT(nbOutputs == 4);
|
||||
if (nbInputs == 2)
|
||||
{
|
||||
PLUGIN_ASSERT(0 <= pos && pos <= 5);
|
||||
}
|
||||
|
||||
// num_detections and detection_classes output: int32_t
|
||||
const int32_t posOut = pos - nbInputs;
|
||||
if (posOut == 0 || posOut == 3)
|
||||
{
|
||||
return inOut[pos].type == DataType::kINT32 && inOut[pos].format == PluginFormat::kLINEAR;
|
||||
}
|
||||
|
||||
// all other inputs/outputs: fp32 or fp16
|
||||
return (inOut[pos].type == DataType::kHALF || inOut[pos].type == DataType::kFLOAT)
|
||||
&& (inOut[0].type == inOut[pos].type);
|
||||
}
|
||||
|
||||
void EfficientNMSImplicitTFTRTPlugin::configurePlugin(
|
||||
const PluginTensorDesc* in, int32_t nbInputs, const PluginTensorDesc* out, int32_t nbOutputs) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
// Inputs: [0] boxes, [1] scores
|
||||
PLUGIN_ASSERT(nbInputs == 2);
|
||||
PLUGIN_ASSERT(nbOutputs == 4);
|
||||
mParam.datatype = in[0].type;
|
||||
|
||||
// Shape of scores input should be
|
||||
// [batch_size, num_boxes, num_classes] or [batch_size, num_boxes,
|
||||
// num_classes, 1]
|
||||
PLUGIN_ASSERT(in[1].dims.nbDims == 2 || (in[1].dims.nbDims == 3 && in[1].dims.d[2] == 1));
|
||||
mParam.numScoreElements = in[1].dims.d[0] * in[1].dims.d[1];
|
||||
mParam.numClasses = in[1].dims.d[1];
|
||||
|
||||
// Shape of boxes input should be
|
||||
// [batch_size, num_boxes, 4] or [batch_size, num_boxes, 1, 4] or [batch_size,
|
||||
// num_boxes, num_classes, 4]
|
||||
PLUGIN_ASSERT(in[0].dims.nbDims == 2 || in[0].dims.nbDims == 3);
|
||||
if (in[0].dims.nbDims == 2)
|
||||
{
|
||||
PLUGIN_ASSERT(in[0].dims.d[1] == 4);
|
||||
mParam.shareLocation = true;
|
||||
mParam.numBoxElements = in[0].dims.d[0] * in[0].dims.d[1];
|
||||
}
|
||||
else
|
||||
{
|
||||
mParam.shareLocation = (in[0].dims.d[1] == 1);
|
||||
PLUGIN_ASSERT(in[0].dims.d[1] == mParam.numClasses || mParam.shareLocation);
|
||||
PLUGIN_ASSERT(in[0].dims.d[2] == 4);
|
||||
mParam.numBoxElements = in[0].dims.d[0] * in[0].dims.d[1] * in[0].dims.d[2];
|
||||
}
|
||||
mParam.numAnchors = in[0].dims.d[0];
|
||||
|
||||
if (nbInputs == 2)
|
||||
{
|
||||
mParam.boxDecoder = false;
|
||||
}
|
||||
}
|
||||
catch (const std::exception& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
}
|
||||
|
||||
EfficientNMSImplicitTFTRTPluginCreator::EfficientNMSImplicitTFTRTPluginCreator()
|
||||
: mParam{}
|
||||
{
|
||||
mPluginAttributes.clear();
|
||||
mPluginAttributes.emplace_back(PluginField("max_output_size_per_class", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("max_total_size", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("iou_threshold", nullptr, PluginFieldType::kFLOAT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("score_threshold", nullptr, PluginFieldType::kFLOAT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("pad_per_class", nullptr, PluginFieldType::kINT32, 1));
|
||||
mPluginAttributes.emplace_back(PluginField("clip_boxes", nullptr, PluginFieldType::kINT32, 1));
|
||||
mFC.nbFields = mPluginAttributes.size();
|
||||
mFC.fields = mPluginAttributes.data();
|
||||
}
|
||||
|
||||
const char* EfficientNMSImplicitTFTRTPluginCreator::getPluginName() const noexcept
|
||||
{
|
||||
return EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_NAME;
|
||||
}
|
||||
|
||||
const char* EfficientNMSImplicitTFTRTPluginCreator::getPluginVersion() const noexcept
|
||||
{
|
||||
return EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_VERSION;
|
||||
}
|
||||
|
||||
const PluginFieldCollection* EfficientNMSImplicitTFTRTPluginCreator::getFieldNames() noexcept
|
||||
{
|
||||
return &mFC;
|
||||
}
|
||||
|
||||
IPluginV2IOExt* EfficientNMSImplicitTFTRTPluginCreator::createPlugin(
|
||||
char const* name, PluginFieldCollection const* fc) noexcept
|
||||
{
|
||||
using namespace std::string_view_literals;
|
||||
try
|
||||
{
|
||||
PluginField const* fields = fc->fields;
|
||||
for (int32_t i = 0; i < fc->nbFields; ++i)
|
||||
{
|
||||
std::string_view const attrName = fields[i].name;
|
||||
if (attrName == "max_output_size_per_class"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.numOutputBoxesPerClass = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "max_total_size"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.numOutputBoxes = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "iou_threshold"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kFLOAT32);
|
||||
mParam.iouThreshold = *(static_cast<float const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "score_threshold"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kFLOAT32);
|
||||
mParam.scoreThreshold = *(static_cast<float const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "pad_per_class"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.padOutputBoxesPerClass = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
if (attrName == "clip_boxes"sv)
|
||||
{
|
||||
PLUGIN_ASSERT(fields[i].type == PluginFieldType::kINT32);
|
||||
mParam.clipBoxes = *(static_cast<int32_t const*>(fields[i].data));
|
||||
}
|
||||
}
|
||||
|
||||
auto plugin = std::make_unique<EfficientNMSImplicitTFTRTPlugin>(mParam);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
IPluginV2IOExt* EfficientNMSImplicitTFTRTPluginCreator::deserializePlugin(
|
||||
const char* name, const void* serialData, size_t serialLength) noexcept
|
||||
{
|
||||
try
|
||||
{
|
||||
// This object will be deleted when the network is destroyed, which will
|
||||
// call EfficientNMSImplicitTFTRTPlugin::destroy()
|
||||
auto plugin = std::make_unique<EfficientNMSImplicitTFTRTPlugin>(serialData, serialLength);
|
||||
plugin->setPluginNamespace(mNamespace.c_str());
|
||||
return plugin.release();
|
||||
}
|
||||
catch (std::exception const& e)
|
||||
{
|
||||
caughtError(e);
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
@@ -0,0 +1,111 @@
|
||||
/*
|
||||
* SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
#ifndef TRT_EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_H
|
||||
#define TRT_EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_H
|
||||
|
||||
#include <vector>
|
||||
|
||||
#include "common/plugin.h"
|
||||
#include "efficientNMSPlugin/efficientNMSParameters.h"
|
||||
|
||||
// This plugin provides CombinedNMS op compatibility for TF-TRT in Implicit Batch
|
||||
// mode for legacy back-compatibilty
|
||||
|
||||
namespace nvinfer1
|
||||
{
|
||||
namespace plugin
|
||||
{
|
||||
|
||||
#if NV_TENSORRT_MAJOR >= 8
|
||||
using EfficientNMSImplicitTFTRTOutputsDataType = void* const*;
|
||||
#else
|
||||
using EfficientNMSImplicitTFTRTOutputsDataType = void**;
|
||||
#endif
|
||||
|
||||
// TF-TRT CombinedNMS Op Compatibility, for Legacy Implicit Batch Mode
|
||||
class TRT_DEPRECATED_BECAUSE("TF-TRT integration is obsolete. No alternatives planned.") EfficientNMSImplicitTFTRTPlugin
|
||||
: public nvinfer1::IPluginV2IOExt
|
||||
{
|
||||
public:
|
||||
explicit EfficientNMSImplicitTFTRTPlugin(EfficientNMSParameters param);
|
||||
EfficientNMSImplicitTFTRTPlugin(const void* data, size_t length);
|
||||
~EfficientNMSImplicitTFTRTPlugin() override = default;
|
||||
|
||||
// IPluginV2 methods
|
||||
const char* getPluginType() const noexcept override;
|
||||
const char* getPluginVersion() const noexcept override;
|
||||
int32_t getNbOutputs() const noexcept override;
|
||||
int32_t initialize() noexcept override;
|
||||
void terminate() noexcept override;
|
||||
size_t getSerializationSize() const noexcept override;
|
||||
void serialize(void* buffer) const noexcept override;
|
||||
void destroy() noexcept override;
|
||||
void setPluginNamespace(const char* libNamespace) noexcept override;
|
||||
const char* getPluginNamespace() const noexcept override;
|
||||
|
||||
nvinfer1::Dims getOutputDimensions(
|
||||
int32_t outputIndex, const nvinfer1::Dims* inputs, int32_t nbInputs) noexcept override;
|
||||
size_t getWorkspaceSize(int32_t maxBatchSize) const noexcept override;
|
||||
int32_t enqueue(int32_t batchSize, void const* const* inputs, EfficientNMSImplicitTFTRTOutputsDataType outputs,
|
||||
void* workspace, cudaStream_t stream) noexcept override;
|
||||
|
||||
// IPluginV2Ext methods
|
||||
nvinfer1::DataType getOutputDataType(
|
||||
int32_t index, nvinfer1::DataType const* inputType, int32_t nbInputs) const noexcept override;
|
||||
nvinfer1::IPluginV2IOExt* clone() const noexcept override;
|
||||
|
||||
// IPluginV2IOExt methods
|
||||
bool supportsFormatCombination(int32_t pos, nvinfer1::PluginTensorDesc const* inOut, int32_t nbInputs,
|
||||
int32_t nbOutputs) const noexcept override;
|
||||
void configurePlugin(const nvinfer1::PluginTensorDesc* in, int32_t nbInputs, const nvinfer1::PluginTensorDesc* out,
|
||||
int32_t nbOutputs) noexcept override;
|
||||
|
||||
protected:
|
||||
void deserialize(int8_t const* data, size_t length);
|
||||
|
||||
EfficientNMSParameters mParam{};
|
||||
std::string mNamespace;
|
||||
};
|
||||
|
||||
class TRT_DEPRECATED_BECAUSE(
|
||||
"TF-TRT integration is obsolete. No alternatives planned.") EfficientNMSImplicitTFTRTPluginCreator
|
||||
: public nvinfer1::pluginInternal::BaseCreator
|
||||
{
|
||||
public:
|
||||
EfficientNMSImplicitTFTRTPluginCreator();
|
||||
~EfficientNMSImplicitTFTRTPluginCreator() override = default;
|
||||
|
||||
const char* getPluginName() const noexcept override;
|
||||
const char* getPluginVersion() const noexcept override;
|
||||
const nvinfer1::PluginFieldCollection* getFieldNames() noexcept override;
|
||||
|
||||
nvinfer1::IPluginV2IOExt* createPlugin(
|
||||
const char* name, const nvinfer1::PluginFieldCollection* fc) noexcept override;
|
||||
nvinfer1::IPluginV2IOExt* deserializePlugin(
|
||||
const char* name, const void* serialData, size_t serialLength) noexcept override;
|
||||
|
||||
protected:
|
||||
nvinfer1::PluginFieldCollection mFC;
|
||||
EfficientNMSParameters mParam;
|
||||
std::vector<nvinfer1::PluginField> mPluginAttributes;
|
||||
std::string mPluginName;
|
||||
};
|
||||
|
||||
} // namespace plugin
|
||||
} // namespace nvinfer1
|
||||
|
||||
#endif // TRT_EFFICIENT_NMS_IMPLICIT_TFTRT_PLUGIN_H
|
||||
Reference in New Issue
Block a user