[libre-riscv-dev] whole stack of vulkan llvm spirv stuff

[Luke Kenneth Casson Leighton]

On Sunday, September 15, 2019, Jacob Lifshay <programmerjake at gmail.com>
wrote:


> As far as I know, AMDGPU is the only LLVM backend that implements
> vectorization in the LLVM-IR to machine-code translation portion.
> Everything else can translate already-vectorized operations to SIMD
> instructions, but relies on LLVM's basic-block and loop vectorizers
> (which are not sufficient for Kazan) to actually generate the vector
> operations.
>
>
Which were lost by the lack of indicators you mentioned.

Blech.

I say "blech", I mean "for a first iteration it is suboptimal but
functional with minimal work".

With extra work to be done further down the line.

Allow me to explain why this matters. We need to be able to meet a
"success" criteria of the NLNet project, 100% completion, before we can
apply for further funds.

More than that, we need to do so quickly before they allocate all of the
EUR 5M to other projects.

With 50% of the budget allocated to Kazan Milestones, the hardware side is
under pressure.

If we can reduce the amount of work a LOT to something that "at least
works", then a second phase of optimising can take place *with a new EUR
50k budget*.

So I am less concerned about performance for an initial version, however I
would also be unhappy if the only way we could get good performance was by
THROWING OUT the work we did in this first phase.

L.



-- 
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68