RC5 magma_dsytrd (actually, dlartd)

[jeremiahpalmer]

Hello! I think I may have found an error in RC5 dlatrd.cpp.

In RC3, in src/dlatrd.cpp, I find this statement:

Code: Select all

#define cublasDsymv magmablas_dsymv

However, in RC5, I don't find that statement.

Here is the output I get from running testing_dsytrd as is:

Code: Select all

device 0: Tesla C2070, 1147.0 MHz clock, 5375.2 MB memory
device 1: Tesla C2050, 1147.0 MHz clock, 2687.4 MB memory

Usage:
  testing_dsytrd -L|U -N 1024



  N    CPU GFlop/s    GPU GFlop/s   |A-QHQ'|/N|A|  |I-QQ'|/N
=============================================================
 1024    20.42         18.77
 2048    17.97         17.40
 3072    14.48         20.33
 4032    13.04         21.02
 5184    12.48         23.11
 6016    12.21         23.78
 7040    11.60         24.11
 8064    10.96         23.02

If I put that define statement in the dlatrd.cpp code, I get these results:

Code: Select all

device 0: Tesla C2070, 1147.0 MHz clock, 5375.2 MB memory
device 1: Tesla C2050, 1147.0 MHz clock, 2687.4 MB memory

Usage:
  testing_dsytrd -L|U -N 1024



  N    CPU GFlop/s    GPU GFlop/s   |A-QHQ'|/N|A|  |I-QQ'|/N
=============================================================
 1024    20.06         18.55
 2048    17.50         13.80
 3072    13.71         20.43
 4032    12.44         25.11
 5184    11.93         29.98
 6016    11.85         32.53
 7040    11.46         34.81
 8064    10.99         35.80

Thanks,
Jeremiah

[Stan Tomov]

Hello,
Thanks for this note. We had a problem with the magma_dsymv kernel and had temporary removed the redefinition of cublasDsymv (thus ending up using cublasDsymv) and forgot to put it back. As your result also shows cublaDsymv is slower. We will go through the release and see which CUBLAS we want to redefine to use MAGMA BLAS.
Stan

[jeremiahpalmer]

Thanks for the explanation. So, is the magma_dsymv that is used in that definition correct and stable?

[lennet]

Yes it's correct)Thanks! I managed to get it right on point.