FP16 support with OpenCL

Did run clpeak tool to check capability of GPU on IMX8QM and discovered that OpenCL does not support half precision operations. That means twice slower than it is capable of .

Is there any way to enable it?
Do you plan to add support for it?

@oostap, please post new questions about Apalis iMX8QM in our confidential space of the community only:
We will look at your question and get back with feedback soon.


The GPU drivers are provided by NXP in binary form only. So we have no possibility to add any feature to OpenCL. If NXP releases newer version of the driver then we will integrate those binary drops into our BSP. What exactly will be part of such new releases we do not know.

NXP describes some of its OpenCL offering in the i.MX_Graphics_User’s_Guide available in the BSP release bundle. Check the L4.9.123_2.3.0_8MM_GA_LINUX_DOC tarball under the Supporting Information. Unfortunately they don’t specify if their implementation does support half-precision, they only state that in the standard support is optional.

Please also check this post on the NXP community.