I don't see the cl_khr_fp16 extension in the latest Broadwell OpenCL driver under Win10.
Does it exist in a Skylake Gen9 IGP OpenCL driver under Win10?
I see that it has been added to the Beignet Linux driver for BDW, SKL and beyond.
Any chance this extension will appear in the Windows driver?
I hope so since fp16x2 support is showing up on other GPUs and is great fit for certain tasks.
Link Copied
Hi Allan,
Standalone driver for Linux https://software.intel.com/en-us/articles/opencl-drivers#latest_linux_driver already supports it on Broadwell.
Skylake support for Linux should be coming by some time this year.
Product team informed me that they enabled support on Windows as well, which means that it should show up in a production driver in a couple of months.
Thanks Robert!
Also, I noticed that the Linux driver appears to support 512 work-item threadgroups. I haven't yet tested it.
BDW reports a 256 item limit on the x.y.z.4404 Win10 driver.
Being able to execute a 448 item threadgroup (7*8*8) that covers an entire BDW/SKL subslice would boost the performance of some kernels.
You can add that to my wishlist. :)
Yes it is, and from what I've seen the use of it should result in ~50% perf increase. Your mileage may vary, of course.
Any idea if GEN8 (Broadwell, etc.) IGPs will ever receive cl_khr_fp16 support?
Today's 20.19.15.4501 driver doesn't show it:
In my tests see cl_khr_fp16 available as below:
I will check for more info and get back to you if there are any updates.
Just wanted to let you know that missing support for 5th Generation Core in Windows is filed as a bug.
Is there any plan to support cl_khr_fp16 on BDW on Windows?
I'm checking with the dev team and will get back to you as soon as I can.
Unfortunately, no. Active development has wrapped up for BDW and development work is now on a different driver branch. However, fp16 should be there for SKL and forward.
For more complete information about compiler optimizations, see our Optimization Notice.