search

JeffersonLab / qphix

QCD for Intel Xeon Phi and Xeon processors
http://jeffersonlab.github.io/qphix/
Other
13 stars 11 forks source link

AVX2 Correctness #14

Open bjoo opened 9 years ago

bjoo commented 9 years ago

I am re-filing this for Mario:

Mario:

"The single and double precision tests succeed but the half precision tests currently fail with

(x,y,z,t)=(0,0,0,0) site=0 spin=0 color=0 Diff = ( -65534.99 , -65532.15 ) chi = ( 65534 , 65534 ) qdp++ =( -0.9914258 , 1.84756 )

...and so forth for all sites. Maybe there is a conversion from half prec. back to single prec. missing before calling the qdp_unpacker?"

bjoo commented 9 years ago

Comment from Dhiraj:

About the correctness of half precision code, I think there is some bug in AVX2 code generator (inst_sp_avx8.cc).

martin-ueding commented 7 years ago

Is this still tested? If so, then it's fixed. If it is not tested, we should definitely add it to the tests!