[PATCH] D38417: [test-suite] Adding HACCKernels app

[Hal Finkel via Phabricator via llvm-commits]

hfinkel added inline comments.


================
Comment at: MultiSource/Benchmarks/DOE-ProxyApps-C++/HACCKernels/Makefile:4
+PROG     = HACCKernels
+FP_TOLERANCE = 0.00001
+CXXFLAGS = -ffast-math -DVERIFICATION_OUTPUT_ONLY=ON
----------------
spatel wrote:
> That's not big enough? We're seeing a 0.001 difference in the output string. Is that wiggle acceptable for this program? How high can we go before we decide the result is bogus?
> 
> You can reproduce this locally if you have at least a Sandybridge to test on. Ie, if you specify -march=nehalem (or nothing), you should see "689.584" in the output, but if you specify -march=sandybridge, you should see "689.585". 
> 
> I don't know what underlying transforms cause that difference, but that seems like a reasonable error for -ffast-math.
> 
> That's not big enough? We're seeing a 0.001 difference in the output string. Is that wiggle acceptable for this program? How high can we go before we decide the result is bogus?

Why not? It's a relative tolerance.  (We have FP_ABSTOLERANCE for setting the absolute tolerance).

In any case, the changes generally come from the inverse sqrt approximation. I think we should set this to not much more than needed for observed divergence.

Brian, were you able to confirm that this tolerance is sufficient?


https://reviews.llvm.org/D38417