comments
Jeffrey Mark Siskind
(24 Apr 2020 18:59 UTC)
|
Re: comments
Jeffrey Mark Siskind
(24 Apr 2020 19:53 UTC)
|
Re: comments
Bradley Lucier
(17 May 2020 21:40 UTC)
|
Re: comments
Jeffrey Mark Siskind
(24 Apr 2020 19:54 UTC)
|
Re: comments
John Cowan
(24 Apr 2020 21:13 UTC)
|
Re: comments Bradley Lucier (25 Apr 2020 23:34 UTC)
|
Re: comments
Bradley Lucier
(26 Apr 2020 00:09 UTC)
|
Re: comments
John Cowan
(26 Apr 2020 03:46 UTC)
|
Re: comments
Bradley Lucier
(28 Apr 2020 20:03 UTC)
|
Re: comments
Bradley Lucier
(26 Apr 2020 22:11 UTC)
|
Re: comments Bradley Lucier 25 Apr 2020 23:34 UTC
On 4/24/20 3:54 PM, Jeffrey Mark Siskind wrote: > One thing that came to mind is that the SRFI does not support half-floats > which are popular with GPUs. Jeff: Thanks for your input. I've been thinking about your comments and what changes might be made in response to them. I don't think this SRFI can be extended to cover the functionality of (py)Torch or Scorch, but it should not be designed in such a way as to preclude, or make difficult, supporting those libraries. The current draft has this language about u16-storage-class, etc.: ============================================================== Each of these could be defined simply as generic-storage-class, but it is assumed that implementations with homogeneous vectors will give definitions that either save space, avoid boxing, etc., for the specialized arrays. ============================================================== Thus, these specialized storage classes are seen as possible optimized implementations, rather than requiring, e.g., adjacent 16-bit unsigned integers as storage. I think now that each of these global variables should have a value either of a storage class that implements "packed" vectors of the appropriate type, or #f if that storage class does not have an implementation. I don't see how to support half floats generically in Scheme, or specifically in Gambit, for which the sample implementation is written, so I propose to (define f16-storage-class #f) and perhaps (define f8-storage-class #f) If someone wants to "pun" u16 or u8 vectors as f16 or f8 vectors to pass to a library, then they can do it. I'll respond to your other comments in another email. Brad