I am looking for an implementation of the OpenCL language that supports the cl_khr_fp16 extension. To my knowledge, no publicly available implementation supports this at present.
views:
56answers:
1
+1
A:
I think you're right, no implementation supports this at this time.
But since cl_khr_fp16 ext is part of v1.1 (AFAIK) and v1.1 is coming very soon, you shouldn't have to wait too much.
Roadmap from AMD:
And NVIDIA already have pre-released v1.1 drivers:
Stringer Bell
2010-07-30 13:34:37
From http://www.khronos.org/files/opencl-1-1-quick-reference-card.pdf , the half scalar and half vector types are still optional extensions in 1.1 for more than 'storage only' use.
grrussel
2010-08-01 19:39:34
Oic. I know that PTX supports `.fp16` but again 'storage only'. From NV doc: *The .f16 floating-point type is allowed only in conversions to and from .f32 and .f64 types. All floating-point instructions operate only on .f32 and .f64 types.)*. So it looks like current NVIDIA gen don't support fp16 ops. This also seems to be the case for current ATI gen (looking at CAL there is no mention of half-precision float at all). Maybe next-gen will support it?
Stringer Bell
2010-08-01 22:47:30