I installed binary distribution of rusticl implementation of OpenCL on Dragon Q6A, but its performance is very poor, i.e. up to 4x slower than Qualcomm OpenCL implementation. I tried to build Qualcomm OpenCL SDK on Ubuntu 24.04 running on Q6A, but I failed. It seems that the SDK is designed to be built on a host platform. Is there a binary available somewhere or another method to get a performant OpenCL on Q6A?
Incidentally, I just found at https://forum.radxa.com/t/opencl-on-q900/29862 one option, which is a binary package qcom-adreno-cl-dev. Still not ideal performance, but some improvement. The data I’m using is produced by clpeak benchmark GitHub - krrishnarraj/clpeak: A synthetic micro-benchmark that measures the peak achievable performance of GPU compute devices · GitHub. Here is Adreno 643 performance (half / float GFLOPS):
Android: 1600 / 1247 (screenshot seen somewhere)
Ubuntu 24.04 rusticl 25.2.8-0 backend: 405 / 553
Ubuntu 24.04 Qualcomm 0838.5 backend: 943 / 650
Ubuntu 24.04 Vulkan 25.2.8 backend: 1096 / 748