Quentin Colombet
Quentin Colombet
Thanks @dcaballe for the heads-up. Taking a look.
> The idea would be to extend the combiner so that vsext can have multiple uses that we know are going to be turned into a vwmul as well. I'm...
> I'm also wondering if there would be a way to pre-compute if an op can fold before actually folding it. In that way, we would be certain that everything...
> vsext.vf4 extends the macc operands to i32 whereas vsext.vf2 extends them to i16. That means that we use 1/2 less registers (lower IMUL) Ah makes sense. I missed that...
Posted a slightly improved version of the #2 patch at https://reviews.llvm.org/D133739 to get the conversation started with Craig.
Forgot to post the update here. I've refactored the code SDISel to make the folding decisions centralized https://reviews.llvm.org/D134703. Then I've updated https://reviews.llvm.org/D133739 for a slightly better term solution. Long term,...
You're right, when I run the program with `sudo`, the limit is 1024. I can collect what I need if I do: `sudo sh -c "ulimit -n && "`
Thanks @pzread for the workaround.
Do you have a profile handy to see if we can narrow down the issue to a few dispatches without having the run/reproduce the full thing?
What are the steps to run with TFLite?