DP and SP validation tests on AMD

[tks2004]

Hi,

When I try to run the validation tests of PAPI_DP_OPS and PAPI_SP_OPS on AMD- EPYC7601, the operation count generated is 0. While the tests works fine on Intel Sklylake processor. Am I missing something here, PAPI_AVAIL displays those two hardware counters are available.

[Vince Weaver]

This is a known issue with these events on EPYC. The floating point instructions generated by our tests somehow aren't counted by the events.

We've reported this issue to AMD and they said they'd look into it, but that was a while ago and we have not gotten a response yet.

[Daniel Barry]

I am unable to reproduce this issue on the AMD-EPYC7301 (Zen) architecture, where the operation counts are precisely what we expect per the validation tests.

Testing the PAPI_SP_OPS event.

Testing a loop with 0 floating point (100 times):
Expected: 0
High: 0 Low: 0 Average: 0
( note, a small value above 0 may be expected due
to overhead and interrupt noise, among other reasons)
Average Error = -nan%

Testing a matrix multiply with 2000000000 single-precision FP operations (3 times)
Result 4225831394628142628864.000000
Expected: 2000000000
High: 2000000000 Low: 2000000000 Average: 2000000000
( note, a small value above 2000000000 may be expected due
to overhead and interrupt noise, among other reasons)
Average Error = 0.00%

Testing a matrix multiply with 0 double-precision FP operations (3 times)
Result 4306628287873107361792.000000
Expected: 0
High: 0 Low: 0 Average: 0
( note, a small value above 0 may be expected due
to overhead and interrupt noise, among other reasons)
Average Error = -nan%

PASSED

and

Testing the PAPI_DP_OPS event.

Testing a loop with 0 floating point (100 times):
Expected: 0
High: 0 Low: 0 Average: 0
( note, a small value above 0 may be expected due
to overhead and interrupt noise, among other reasons)
Average Error = -nan%

Testing a matrix multiply with 0 single-precision FP operations (3 times)
Result 4225831394628142628864.000000
Expected: 0
High: 0 Low: 0 Average: 0
( note, a small value above 0 may be expected due
to overhead and interrupt noise, among other reasons)
Average Error = -nan%

Testing a matrix multiply with 2000000000 double-precision FP operations (3 times)
Result 4306628287873107361792.000000
Expected: 2000000000
High: 2000000000 Low: 2000000000 Average: 2000000000
( note, a small value above 2000000000 may be expected due
to overhead and interrupt noise, among other reasons)
Average Error = 0.00%

PASSED

‌

[Daniel Barry]

Unable to reproduce this issue.