Deep Learning on GPU - Adreno 530
My need is to run some deep learning on the GPU. (Specifically Fast Depth.) Model works fine, just not on the GPU - and I need performance.
Chad nicely commented before about options. I've played with that information and also played with some more. Looking for feedback (like 'try harder') or ideas. Options/situation:
- ARM Compute Library (existing modalAI example.) Works fine, but ACL uses their 'graph' language for neural nets. Googling finds no way to import existing models from, say Tensorflow.
- ARM NN (companion-esque library to ACL.) Looks fine, imports Tensorflow. Except documentation is adamant that OpenCL only works with ARM Mali GPU.
- OpenCV (dnn calling OpenCL.) I honestly haven't tried this, but the 'web' seems to show using the GPU is slower than the CPU (https://github.com/opencv/opencv/wiki/DNN-Efficiency)
- tvm (tvm.ai) Still trying to get the OpenCL to build. Stay tuned.
- Tensorflow built to use OpenCL (using SYKL). Feels like a lot of work and the end result, as they say will not be good performance.
- OpenVINO - seems to be (entirely?) Intel centric
- MACE - seems to be of potential. Again, stay tuned. (My colleagues tell me I'm getting to the nether regions of the web now...)
Am I missing something obvious here?
Thanks for any/all ideas! p