February 16, 2023, 2:51pm
I saw pull request to tflite opencl from PowerVR SWE. It was not accepted because uses not public builtin functions: __builtin_IMG_dotF16_m2k4n4_accF32_bt.
11:23AM - 06 Jun 22 UTC
- Adds powervr gpu_info
- On the new powervr gpu ge
… neration, uses the fp16 matmul builtin for conv2d
- Uses better workgroup size and block size
We see 1.3 ~ 1.55 times speedup of inference time on IMG BXM-8-256 using our internal NN models.
benchmark_model option: --use_gpu=true --gpu_backend=cl --gpu_precision_loss_allowed=true
First question: can code from this CL used internally inside Google?
Second: do you have plans to have public documentation for this builtin or make some extension that can be used?
Welcome to our forum!
Sorry about that PR. We are looking into documenting (or making an extension as you suggested) so we can get this functionality accepted.
Yes, Google can use this code as it was approved for public release.
Let me know if you have any other question. Hopefully we will update the PR shortly.
February 20, 2023, 8:27am
Added PowerVR gpu info using your PR.