New Nvidia driver adds 16bit float support to OpenCL [pdf]
us.download.nvidia.comCan someone explain why they update to Clang 7, that was released 19 September 2018?
It's very very common for software like shader compilers or AOT compilers for other languages to be based on old clang/llvm forks.
It doesn't seem it's new. The release notes for the 511.79 driver from February 2022 say the same thing:
https://us.download.nvidia.com/Windows/511.79/511.79-win11-w...
It was some kind of opt-in option (which I wasn't even aware of), now it's on by default / rolled out to everyone.
It seems they are finally enable the 16bit float extension in their OpenCL implementation.
Is this the IEEE float16 type or the bfloat16 type? It doesn't say
IEEE
any idea why this took them so long?
Because they have a near monopoly with Cuda and get away with dragging their heels due to lack of market pressure (everyone loves to lock themselves into Cuda and then complain about GPU prices), despite having been on the OpenCL committee (just like Apple with Metal conflict of interest).
Anyway, I'm very glad to see it and will be using it immediately.
Also because none of the competition were/are serious about OpenCL either.
AMD still doesn't have OpenCL 3.0 support and their implementation of previous versions was far far less stable than CUDA.
I can't find a definite source on this, but afaik none of the official OpenCL implementations have ever fully supported mixed CPU-GPU code the way CUDA does.
AMD's OpenCL support on GPU is overall excellent in my experience (2x commercial apps and lots of hobby code), but I tend to mostly use low level OpenCL 1.1 stuff, which I find sufficient.
Also Intel GPUs are actually incredibly competent with OpenCL if you give them wide enough NDrange, and somehow try to look past lack of any fp64 support at all :/
On top of that even the "do no evil" Google, never supported OpenCL on Android, pushing instead their own dialect, Renderscript.
Yes, there are some custom Android deployments that have a libopencl.so kind of thing, it is used by the OEMs themselves, and never exposed as official Android API.
This is why I am investing in AMD. They can only improve!
CUDA itself only got bumped from LLVM 5 to LLVM 7 in CUDA 11.2 (https://developer.nvidia.com/blog/boosting-productivity-and-...).
LLVM 7 opt-in for OpenCL happened some time later (available since r510). What changes now is that LLVM 7 is the new default.