Ascii benchmarks #3016
Ascii benchmarks #3016
Conversation
| { | ||
| [Params( | ||
| 6, // non-vectorized code path | ||
| 128)] // vectorized code path |
There was a problem hiding this comment.
Should there be a size for Vector128/256 too?
The logic for Vec256 and Vec128 is actually the same, just different vector sizes.
There was a problem hiding this comment.
Should there be a size for Vector128/256 too?
You mean the smallest size that gets vectorized? Since we cover both byte and char inputs it would require adding at least few more cases, which would increase the time it takes to run the benchmarks (it's 6.5 minute currently). I would prefer to avoid doing that.
There was a problem hiding this comment.
On x64 with AVX available, these benchmarks examine the scalar path and the Vector256 path, whilst the Vector128 path won't be hit.
I don't know if this is a problem. Benchmark get slower by adding a size for that path, but coverage gets higher.
I would prefer to avoid doing that.
So do I.
|
@DrewScoggins Could you please run the tool that generates the html files for https://pvscmdupload.blob.core.windows.net/reports/allTestHistory/TestHistoryIndexIndex.html so the new benchmarks are included? My motivation is showing the gains from dotnet/runtime#85926 and dotnet/runtime#85266 cc @sblom |
|
@adamsitnik To have the new HTMLs generated is bit elaborate and requires some time on our side. In the meantime here is the new benchmarks in a graph in ADX (you can change the Windows.10.Amd64.19H1.Tiger.Perf and x64 in case you want different CPU/OS). Let me know if you need something more. |
|
@cincuranet thank you! please consider me unblocked |
4 participants
benchmarks for dotnet/runtime#85926
cc @gfoidl @GrabYourPitchforks