acts-project / traccc

Demonstrator tracking chain on accelerators
Mozilla Public License 2.0
29 stars 50 forks source link

x86-64-v2, main branch (2024.10.24.) #751

Closed krasznaa closed 1 month ago

krasznaa commented 1 month ago

Use -march=x86-64-v2 for the build when appropriate. While still allowing builds to use something more aggressive, like:

CXXFLAGS="-march=native" cmake ...

This was triggered by recent discussions about vectorization. So I thought we might as well use the same code for this that the ATLAS offline build uses. :thinking:

https://gitlab.cern.ch/atlas/atlasexternals/-/blob/main/Build/AtlasCMake/modules/AtlasCompilerSettings.cmake?ref_type=heads#L103

stephenswat commented 1 month ago

Also, x86-64-v3 has been around for well over a decade now, so to me it makes sense to target that. :stuck_out_tongue:

krasznaa commented 1 month ago

For v3: We could, but at the moment we only use v2 on the grid with the ATLAS code. That's why I went with v2 out of the box. Again, one can set something more aggressive in local tests. But for the ATLAS offline code we've seen practically no improvement with more aggressive settings.

For SYCL: Do you have a suggestion? 😕 Right now, building SYCL code is still sort of an expert task. Asking people to set up, among other things, the $SYCLFLAGS environment variable in a non-trivial way to make the build do exactly what they want. That by itself is a complicated enough procedure that it would be hard to tweak the settings safely from the project's CMake code.

sonarcloud[bot] commented 1 month ago

Quality Gate Passed Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarCloud

stephenswat commented 1 month ago

For SYCL: Do you have a suggestion? 😕 Right now, building SYCL code is still sort of an expert task. Asking people to set up, among other things, the $SYCLFLAGS environment variable in a non-trivial way to make the build do exactly what they want. That by itself is a complicated enough procedure that it would be hard to tweak the settings safely from the project's CMake code.

I don't think it needs to be so complicated; if we slap a -match=whatever flag on the SYCL flags we'll get architectural support for at least the host code, and hopefully also for whatever device code is compiled with x86_64 SYCL targets.