[Libre-soc-dev] video assembler

Luke Kenneth Casson Leighton lkcl at lkcl.net
Tue May 11 18:40:16 BST 2021


On Tuesday, May 11, 2021, Lauri Kasanen <cand at gmx.com> wrote:

>
>
> That's not what I meant. I meant for the simulator, as per the email
> thread.


oh right, ok.


> The simulator should report executed cycles and executed
> sub-cycles separately.


cando.  it will be pretty easy to report some counters.  also hmm maybe
some perfcounters, showing number of ops executed.

btw i am serious about the example i gave, that the code, i expect it
really should not change depending on hardware capability.

more to the point: if it *does* need changing (other than because we added
a new instruction, that is) this should be taken as a sign that we've done
something wrong.

Cray-style Vectors are an anti-pattern: all of the "hard work" that
normally has to go into SIMD optimisation, all of that is completely out
the window with Cray Vectors.

the most extreme example i have been able to find on this so far is:

* strncpy for RVV, 13 instructions (some compressed)
* strncpy for VSX, a staggering 250.

https://patchwork.ozlabs.org/project/glibc/patch/20200929152103.18564-1-rzinsly@linux.ibm.com/

l.



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


More information about the Libre-soc-dev mailing list