[Libre-soc-bugs] [Bug 386] Evaluate Vallium instead of Kazan

bugzilla-daemon at libre-soc.org bugzilla-daemon at libre-soc.org
Tue Jun 16 12:25:47 BST 2020


--- Comment #3 from Luke Kenneth Casson Leighton <lkcl at lkcl.net> ---
(In reply to cand from comment #2)
> I understand the issues with Airlie's work, but if (assumed) Vallium is
> relatively complete by Oct, can we afford not to use it? 

it will need significant evaluation.  the usual mistake is to optimise
vector intrinsics out early in the SPIR-V passes.

this is one of the reasons why we cannot start from SwiftShader because
the *fundamental* assumption is that it is going to be software-only
therefore why bother sustaining vectors, subvectors, predication and
swizzle through to the final IR pass?

> Getting Kazan to
> the same point would take months and EUR.

there's a fundamental reason for that: jacob is doing a proper shader
compiler and is taking into consideration the vector and predication
intrinsics right from the start.

if those have not been sustained right the way through from the SPIR-V
through to when they go actually into the LLVM IR conversion phase to
assembler, they have to be *REINTRODUCED* using a massive complex costly
and timeconsuming "auto-vectorisation" phase.

> llvmpipe is only single-threaded in some parts,

jacob explained last year - we'll have to wait until U.S. timezone for
confirmation - these "some parts" are what killed gallium3 performance
and would also severely hamper performance for us.

bottom line is: it's extremely unlikely to be suitable.

You are receiving this mail because:
You are on the CC list for the bug.

More information about the libre-soc-bugs mailing list