Alexandre Eichenberger
Alexandre Eichenberger
@adriansion I ran an example of your test like this: ``` python ~/onnx-mlir/test/perf/ParseBenchmarks.py --compare PerfGemm_Benchmark_1650564009 PerfGemm_Benchmark_1650564009 Pct --verbose ``` and got this ``` # arg = PerfGemm_Benchmark_1650564009 "BM_MatmulSquare/512" "BM_MatmulSquare/1024" 39...
As discussed in our meetings, a naive solution would simply force down inf to, say 30. We won't crash anymore, and the one between 30-inf will return wrong values as...
@negiyas any insight on what is happening?
> Please look at the following outputs of the both cases. Good to see that you are finding inefficiencies that can be further optimized in the `?`=`1` case. Please try...
I don't think we can expect our users to setup this option. Can we set it up to a higher value? * if no changes occurs at iteration n, do...
@negiyas please do, it should be an easy one. Thanks for the investigation and finding a fix. And link the PR so we know when to close this one.
The problem with simd disabling is that it changes a lot of how the code is generated. Are you running on z? If so, you could also omit the -mcpu=z14...
I should probably investigate myself, as I wrote the code. Will try to look at it tomorrow.
I am currently looking into this issue, no firm result yet. I cannot reproduce this issue on x86/linux, but can on z14 without simd code being actually generated as I...