Integrate Caliper

[michaelmckinsey1]

Caliper Integration for RAJAPerf

This work is a variant of #254

• This PR is a feature.
• It does the following:
  • Implements caliper annotations into RAJAPerf.
  • Adds new Adiak metadata fields.
• This feature is at the request of the PAVE team.

This is dependent on #333 #335

[michaelmckinsey1]

Ignore the tuning commits, they will go away once #333 is merged merged

[michaelmckinsey1]

I put CALI_START after and CALI_STOP before the existing timers. Is this acceptable? I assume if someone turns caliper on they would be looking at those timings so they should be more accurate and in the opposite case caliper would be off so it shouldn't impact the original timers that much. Not sure how much impact this has.

[rhornung67]

I think that choice makes most sense.

[rhornung67]

If Caliper is off, do the macros become no-ops with a simple run time conditional evaluation?

[michaelmckinsey1]

Yes. If RAJAPerf is not built with caliper, RAJA_PERFSUITE_USE_CALIPER=OFF, the macros will be defined as here, so the only operation will be the #if defined(RAJA_PERFSUITE_USE_CALIPER).

[michaelmckinsey1]

fc53a26 For the doc files I changed as little wording as possible for now. It may be necessary to have stronger wording in regards to running 1 variant & 1 tuning at a time, since multiple tunings at a time makes less sense as we removed the variant.tuning node level from the caliper annotations. It also may be a good idea to enforce (in the code) single variant/tuning or file splitting when caliper is turned on.

[rhornung67]

This doesn't look right to me. Is 'vstr' the "variant" name? If so, then it seems like it should be set by calling getVariantName(), which takes a VariantID enum value.

Am I misunderstanding this?

Also, it feels like we can set up some special utility methods to avoid having to re-parse strings to extract substrings repeatedly.

[michaelmckinsey1]

We specficially moved away from getVariantName() to a static string, RAJAPerf, so different runs will not have different top-level node names. This fits the standard expected by our tools. This information should instead be collected in the metadata, as shown here.

[rhornung67]

OK. I'm still confused about this. I can ask you about it later sometime.

[michaelmckinsey1]

fc53a26 For the doc files I changed as little wording as possible for now. It may be necessary to have stronger wording in regards to running 1 variant & 1 tuning at a time, since multiple tunings at a time makes less sense as we removed the variant.tuning node level from the caliper annotations. It also may be a good idea to enforce (in the code) single variant/tuning or file splitting when caliper is turned on.

Addressed in c0197bf. Generate a single cali file per variant, tuning combo.

[michaelmckinsey1]

New commits on #352