Forcing ATLAS to use your GEMM
If ATLAS detects you are on a platform where a contributed full GEMM is
superior to ATLAS's large-case GEMM, ATLAS will automatically handle
the details of making ATLAS call the user-contributed routine. If,
however, you wish to force ATLAS to use your GEMM (for instance, you
are testing your code before contribution, or just want to utilize ATLAS
for complete BLAS coverage with your GEMM), you should take
the following steps, after creating the appropriate subdirectory and
API as previously described:
- Edit your Make.inc file:
- Change UMMdir to point to your full GEMM's architecture
subdirectory
- Add -DUSERGEMM to the CDEFS macro.
- In ATLAS/src/blas/gemm, touch ATL_gemmXX.c and
ATL_AgemmXX.c to force recompilation
- In BLDdir/src/blas/gemm/ type make lib
- In BLDdir/include/, issue
- rm ?Xover.h atlas_cacheedge.h
- touch altas_cacheedge.h sXover.h dXover.h cXover.h zXover.h
- In BLDdir/tune/blas/gemm/, issue:
- rm res/?Xover.h res/atlas_cacheedge.h
- make res/atlas_cacheedge.h
- make res/sXover.h pre=s
- make res/dXover.h pre=d
- make res/cXover.h pre=c
- make res/zXover.h pre=z
Clint Whaley
2012-07-10