Re: Enabling ROCm on Everything
On 2023-03-21 06:17, Cordell Bloor wrote:
> One possible split would be on the GFX architecture major version. There
> would be binary packages for librocsparse0-gfx8, librocsparse0-gfx9,
> librocsparse0-gfx10, and librocsparse0-gfx11 with each providing
> librocsparse0. The GFX9 grouping would be pretty large with six
> architectures, but that's still within acceptable limits. If need be, it
> could be split into gfx9-gcn (gfx900, gfx904, gfx906) and gfx9-cnda
> (gfx908, gfx90a, gfx940).
Something that just came to my mind: why don't we (eventually) split the
packages by RDNA1/2/3 resp. CDNA1/2/3 resp. "other"[1]?
Advantages:
* My gut feels that would leave packages small enough for 2GiB to not
be a problem
* Reduced maintenance overhead (both for us, and for ftp-master)
* Correct package trivially found by users (as compared to gfx____)
I think the last point is key because honestly, the gfx____ nomenclature
is totally opaque unless you're somebody who works close enough to
hardware.[2] On the other hand, RDNA1/2/3 and CDNA1/2/3 have widespread
use and understanding.
Best,
Christian
[1] If such things exist. Custom silicon, or whatever
[2] Thanks to LLMs, GPU-accelerated is more popular than ever, however
most users aren't close to the hardware. At least from what I see.
Reply to: