core(OpenCL): optimize convertTo() with CV_16F (convertFp16() replacement) #24918
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
99ba03c to
b9b3860
Compare
2 tasks
fengyuentau
reviewed
Jan 25, 2024
| DUMP_CONFIG_PROPERTY("cv_ocl_current_maxMemAllocSize", device.maxMemAllocSize()); | ||
|
|
||
| const char* doubleSupportStr = device.doubleFPConfig() > 0 ? "Yes" : "No"; | ||
| const char* doubleSupportStr = device.hasFP64() ? "Yes" : "No"; |
Member
There was a problem hiding this comment.
So what is going to happen with doubleFPConfig and halfFPConfig. Are they deprecating as well?
Contributor
Author
There was a problem hiding this comment.
No, they are still needed if we want to compute with proper inf/nans support.
Member
There was a problem hiding this comment.
I checked doubleFPConfig in the whole opencv project and it is basically used like this,
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0they are still needed if we want to compute with proper inf/nans support
Did I miss anything here? Or it is in the user code instead?
Contributor
Author
There was a problem hiding this comment.
Usage is not correct.
Member
There was a problem hiding this comment.
What is the correct way? All these code is wrong?
Contributor
Author
There was a problem hiding this comment.
All are subject for revising.
fengyuentau
approved these changes
Jan 26, 2024
vpisarev
approved these changes
Jan 26, 2024
asmorkalov
pushed a commit
that referenced
this pull request
Jan 26, 2024
…16s_usage DNN: avoid CV_16S usage for FP16 #24892 **Merge after**: #24918 TODO: - [x] measure performance changes - [x] optimize convertTo for OpenCL: #24918 12700K iGPU: |Name of Test|0|1|1 vs 0 (x-factor)| |---|:-:|:-:|:-:| |AlexNet::DNNTestNetwork::OCV/OCL_FP16|7.441|7.480|0.99| |CRNN::DNNTestNetwork::OCV/OCL_FP16|10.776|10.736|1.00| |DenseNet_121::DNNTestNetwork::OCV/OCL_FP16|52.762|52.833|1.00| |EAST_text_detection::DNNTestNetwork::OCV/OCL_FP16|60.694|60.721|1.00| |EfficientNet::DNNTestNetwork::OCV/OCL_FP16|33.373|33.173|1.01| |FastNeuralStyle_eccv16::DNNTestNetwork::OCV/OCL_FP16|81.840|81.724|1.00| |GoogLeNet::DNNTestNetwork::OCV/OCL_FP16|20.965|20.927|1.00| |Inception_5h::DNNTestNetwork::OCV/OCL_FP16|22.204|22.173|1.00| |Inception_v2_SSD_TensorFlow::DNNTestNetwork::OCV/OCL_FP16|47.115|47.460|0.99| |MPHand::DNNTestNetwork::OCV/OCL_FP16|6.760|6.670|1.01| |MPPalm::DNNTestNetwork::OCV/OCL_FP16|10.188|10.171|1.00| |MPPose::DNNTestNetwork::OCV/OCL_FP16|12.510|12.561|1.00| |MobileNet_SSD_Caffe::DNNTestNetwork::OCV/OCL_FP16|17.290|17.072|1.01| |MobileNet_SSD_v1_TensorFlow::DNNTestNetwork::OCV/OCL_FP16|19.473|19.306|1.01| |MobileNet_SSD_v2_TensorFlow::DNNTestNetwork::OCV/OCL_FP16|22.874|23.404|0.98| |OpenFace::DNNTestNetwork::OCV/OCL_FP16|9.568|9.517|1.01| |OpenPose_pose_mpi_faster_4_stages::DNNTestNetwork::OCV/OCL_FP16|539.899|539.845|1.00| |PPHumanSeg::DNNTestNetwork::OCV/OCL_FP16|18.015|18.769|0.96| |PPOCRv3::DNNTestNetwork::OCV/OCL_FP16|63.122|63.540|0.99| |ResNet_50::DNNTestNetwork::OCV/OCL_FP16|34.947|34.925|1.00| |SFace::DNNTestNetwork::OCV/OCL_FP16|10.249|10.206|1.00| |SSD::DNNTestNetwork::OCV/OCL_FP16|213.068|213.108|1.00| |SqueezeNet_v1_1::DNNTestNetwork::OCV/OCL_FP16|4.867|4.878|1.00| |VIT_B_32::DNNTestNetwork::OCV/OCL_FP16|200.563|190.788|1.05| |VitTrack::DNNTestNetwork::OCV/OCL_FP16|7.528|7.173|1.05| |YOLOX::DNNTestNetwork::OCV/OCL_FP16|132.858|132.701|1.00| |YOLOv3::DNNTestNetwork::OCV/OCL_FP16|209.559|208.809|1.00| |YOLOv4::DNNTestNetwork::OCV/OCL_FP16|221.357|220.924|1.00| |YOLOv4_tiny::DNNTestNetwork::OCV/OCL_FP16|24.446|24.382|1.00| |YOLOv5::DNNTestNetwork::OCV/OCL_FP16|43.922|44.080|1.00| |YOLOv8::DNNTestNetwork::OCV/OCL_FP16|64.159|63.842|1.00| |YuNet::DNNTestNetwork::OCV/OCL_FP16|10.177|10.231|0.99| |opencv_face_detector::DNNTestNetwork::OCV/OCL_FP16|15.121|15.445|0.98| Co-authored-by: Alexander Alekhin <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
relates #24909
relates #24917
relates #24892
Performance changes:
Patched version is not worse than convertFp16 and convertTo baseline (except MatUMat 32->16, baseline uses CPU code+dst buffer map).
There are still gaps against noOpenCL(CPU only) mode due to T-API implementation issues (unnecessary synchronization).
convertTo-baseline can't compile OpenCL kernel for FP16 properly - FIXED.
dGPU has much more power, so results are x16-17 better than single cpu core.
Patched version is not worse than convertFp16 and convertTo baseline.
There are still gaps against noOpenCL(CPU only) mode due to T-API implementation issues (unnecessary synchronization) and required memory transfers.