GSoC 2019 | OpenCV | Adding a CUDA backend to the DNN module

A0_NOTICE.md

DISCLAIMER

This gist documents the Google Summer of Code project. It is not updated and hence does not indicate current status of the CUDA backend.

For updated details, please see this gist.

D0_Summary.md

Allow the OpenCV's DNN module to work with GPUs

Student: Yashas Samaga B L

Mentor: Davis King

Project Link: https://summerofcode.withgoogle.com/projects/#6021087400296448

Relevant PRs:

• add cuDNN dependency and setup build for cuda4dnn
• add CUDA backend to the DNN module

Introduction

The OpenCV’s DNN module has a blazing fast inference capability on CPUs. It supports performing inference on GPUs using OpenCL but lacks a CUDA backend. NVIDIA’s GPUs support OpenCL, but their capabilities are limited by OpenCL.

This project adds a new CUDA backend that can perform lightning fast inference on NVIDIA GPUs.

How to use?

Build

The CUDA backend requires CUDA Toolkit and cuDNN (min: 7.5.0) to be installed on the system. The CMake scripts will automatically detect the dependencies when the following options are set:

• WITH_CUDA
• WITH_CUDNN

The CUDA backend is enabled by setting the following option:

• OPENCV_DNN_CUDA

After building, run [build dir]/bin/opencv_test_dnn and [build dir]/bin/opencv_perf_dnn.

Usage

The project adds the following new backends and targets to the existing list.

Backend	Target
DNN_BACKEND_CUDA	DNN_TARGET_CUDA
DNN_BACKEND_CUDA	DNN_TARGET_CUDA_FP16

D1_SupportMatrix.md

Support Matrix

The CUDA backend uses OpenCV's CPU backend as a fallback for unsupported layers and partially supported layers with unsupported configurations.

Blip	Meaning
✔️	fully supported
🔵	partially supported
❌	unsupported

Layer	Status
Activations	✔️
Batch Normalization	✔️
Blank Layer	✔️
Concat Layer	✔️
Const Layer	✔️
Convolution 2d	✔️
Convolution 3d	✔️
Crop and resize	❌
Crop Layer	✔️
Detection Output Layer	❌
Deconvolution 2d	🔵 (most configurations supported)
Deconvolution 3d	🔵 (most configurations supported)
Elementwise Layers	✔️
Eltwise Layer	✔️
Flatten Layer	✔️
Fully Connected Layer	✔️
Input Layer	❌
Interp Layer	✔️
Local Response Normalization	✔️
Max Unpooling 2d	✔️
Max Unpooling 3d	✔️
MVN Layer	❌
Normalize Layer	🔵 (L1 and L2 supported)
Padding Layer	✔️
Permute Layer	✔️
Pooling 2d	🔵 (max and average supported)
Pooling 3d	🔵 (max and average supported)
Prior Box Layer	✔️
Proposal Layer	❌
Region Layer	✔️
Reorg Layer	✔️
Reshape Layer	✔️
Resize Layer	✔️
Scale Layer	✔️
Shift Layer	✔️
Shuffle Channel Layer	✔️
Slice Layer	✔️
Softmax Layer	✔️
Split Layer	✔️
LSTM Layer	❌

D2_Backend_Comparision.md

OCV CPU vs IE CPU vs CUDA

CPU: i7 7700HQ

GPU: NVIDIA GTX 1050 Mobile

CPU BLAS Library: MKL 2019.0.4

CUDA Version: 10.1

cuDNN: 7.6.2

Warmup Runs: 3 (forward pass is performed three times before benchmarks)

Benchmark Runs: 10 (the average of ten forward passes is reported)

Test Code: https://gist.github.com/YashasSamaga/71157cf0c3768c497e5e70fb95435596

Batch Size = 1

Model	CUDA FP32	Inference Engine CPU	OpenCV CPU
GoogLeNet	7.2447ms	10.4981ms	17.9176ms
DenseNet121	12.6324ms	19.1823ms	48.0628ms
EAST Text Detection	18.8281ms	49.0508ms	88.9429ms
ENet	11.5014ms	Exception	62.5854ms
FastNeuralStyle StaryNight	27.498ms	178.309ms	160.359ms
Inception 5h	7.8546ms	22.2789ms	20.3255ms
Inception v2 FasterRCNN	112.736ms	Exception	374.26ms
MobileNet SSD	58.4751ms	9.2896ms	27.3061ms
OpenCV Face Detector	6.9831ms	8.3981ms	17.6683ms
OpenPose Pose MPI	160.561ms	509.446ms	838.161ms
Resnet 50	11.3603ms	28.1529ms	50.2752ms
SqueezeNet	2.4084ms	3.2918ms	5.476ms
VGG16 SSD	70.4117ms	249.725ms	360.207ms
Yolo v3	57.9822ms	214.629ms	296.806ms
Yolo v2	51.5784ms	193.453ms	260.19ms

Batch Size = 10

Model	CUDA FP32	Inference Engine CPU	OpenCV CPU
GoogLeNet	35.7556ms	108.946ms	225.928ms
DenseNet121	74.9241ms	295.105ms	650.924ms
EAST Text Detection	149.58ms	536.946ms	1273.93ms
FastNeuralStyle StaryNight	283.173ms	1966.5ms	2175.3ms
Inception 5h	36.6225ms	180.429ms	233.276ms
MobileNet SSD	277.753ms	111.872ms	316.063ms
OpenCV Face Detector	52.4366ms	95.7866ms	202.657ms
OpenPose Pose MPI	628.617ms	5650.05ms	10683.5ms
Resnet 50	74.283ms	230.817ms	541.308ms
SqueezeNet	15.8144ms	35.4915ms	69.4122ms
VGG16 SSD	594.286ms	2796.23ms	4661.51ms
Yolo v3	488.704ms	2419.8ms	4209.74ms
Yolo v2	491.414ms	2185.47ms	3788.34ms

D3_CUDA_vs_CPU.md

OpenCV CUDA vs OpenCV CPU

CPU: 2x Intel Xeon E5-2640 v4

GPU: 1x NVIDIA GTX 1080 Ti (11 GB)

CPU BLAS Library: OpenBLAS 0.2.20

CUDA Version: 10.0

cuDNN: 7.6.2

Warmup Runs: 3 (forward pass is performed three times before benchmarks)

Benchmark Runs: 10 (the average of ten forward passes is reported)

Test Code: https://gist.github.com/YashasSamaga/71157cf0c3768c497e5e70fb95435596

Backend Comparision

Batch Size = 1

Model	CUDA FP32	OpenCV CPU
GoogLeNet	4.8824ms	14.2981ms
DenseNet121	6.4555ms	57.8244ms
EAST Text Detection	5.901ms	67.4301ms
ENet	4.5979ms	30.2767ms
FastNeuralStyle StaryNight	5.3193ms	51.3313ms
Inception 5h	4.9487ms	16.0048ms
Inception v2 FasterRCNN	82.0298ms	179.245ms
MobileNet SSD	70.9177ms	23.9348ms
OpenCV Face Detector	4.9288ms	15.4205ms
OpenPose Pose MPI	30.5954ms	246.747ms
Resnet 50	4.5968ms	45.1153ms
SqueezeNet	1.0888ms	3.6492ms
VGG16 SSD	23.5926ms	194.976ms
Yolo v3	18.0002ms	141.861ms
Yolo v2	12.1279ms	111.642ms

Batch Size = 10

Model	CUDA FP32	OpenCV CPU
GoogLeNet	10.149ms	75.9591ms
DenseNet121	20.269ms	312.426ms
EAST Text Detection	32.1556ms	402.16ms
FastNeuralStyle StaryNight	49.1025ms	461.095ms
Inception 5h	9.9721ms	67.9308ms
MobileNet SSD	96.2898ms	110.783ms
OpenCV Face Detector	22.7501ms	77.8742ms
OpenPose Pose MPI	118.858ms	2321.89ms
Resnet 50	18.4139ms	229.599ms
SqueezeNet	4.4893ms	22.3049ms
VGG16 SSD	194.181ms	1319.67ms
Yolo v3	122.603ms	1044.11ms
Yolo v2	104.072ms	819.177ms

Batch Size = 128

Model	CUDA FP32	OpenCV CPU
GoogLeNet	90.3755ms	775.769ms
DenseNet121	199.516ms	3536.38ms
EAST Text Detection	376.458ms	7685.72ms
FastNeuralStyle StaryNight	801.778ms	6607.15ms
Inception 5h	93.4188ms	771.575ms
MobileNet SSD	1028.93ms	1110.37ms
OpenCV Face Detector	276.992ms	977.997ms
OpenPose Pose MPI	1279.26ms	32159.3ms
Resnet 50	200.789ms	1719.92ms
SqueezeNet	55.6244ms	255.397ms
VGG16 SSD	2969.05ms	17201ms
Yolo v3	1564.78ms	13699.2ms
Yolo v2	1362.84ms	11254.9ms

Images processed per second (CUDA FP32)

Model	batch size = 1	batch size = 10	batch size = 128
GoogLeNet	204	985	1416
DenseNet121	154	493	641
EAST Text Detection	169	311	340
ENet	217	Not Applicable	Not Applicable
FastNeuralStyle StaryNight	188	204	160
Inception 5h	202	1002	1370
Inception v2 FasterRCNN	12	Not Aplicable	Not Applicable
MobileNet SSD	14	104	124
OpenCV Face Detector	202	440	462
OpenPose Pose MPI	33	84	100
Resnet 50	217	540	637
SqueezeNet	918	2228	2301
VGG16 SSD	42	52	43
Yolo v3	55	82	81
Yolo v2	82	96	93

D4_OpenCV_vs_TensorFlow.md

OpenCV CUDA vs TensorFlow

GPU: NVIDIA GTX 1080 Ti (11 GB)

Batch of 1

Model	OpenCV CUDA	TensorFlow
ResNet-50	4.5968ms	7.1163ms
EAST Text Detection	5.901ms	8.6890ms

Batch of 10

Model	OpenCV CUDA	TensorFlow
ResNet-50	18.4139ms	22.3665ms
EAST Text Detection	32.1556ms	39.4857ms

Batch of 128

Model	OpenCV CUDA	TensorFlow
ResNet-50	200.789ms	216.3923ms
EAST Text Detection	376.458ms	421.8292ms

The benchmarks I posted are for MobileNetSSD_deploy.prototxt/MobileNetSSD_deploy.caffemodel which you can find here.

MobileNet is slow with the CUDA backend because of depthwise convolutions. The CUDA backend fully relies on cuDNN for convolutions and cuDNN is very bad at depthwise convolutions.

Has this bad performance of depthwise conv fixed for cuDNN now? I just tested a DNN model with depthwise conv on a Jetson Nano B01 with CUDA 10.2 & cuDNN 8.2.1. Results turned out to be slower than the SoC of Jetson Nano (arm-based, 4-core, 1.5GHz), and much slower than the SoC on Raspberry Pi 4B (4-core, 1.5 GHz).

Also tested a model w/o depthwise conv, which seems to be meeting expectations (~10X times faster than the Jetson Nano SoC).

Do you have an academic paper for these results?
I would like to cite.

@Algabri No, there is no paper that is specific to this project.

@YashasSamaga , Thanks for your reply.