====================================================================== BSOF/I README FILE

Structured Orthogonal Inversion Codes for Block p-Cyclic Matrices

---

This is a set of pure C99 codes for block structured factorization and inversion (BSOF/I) of block p-cyclic matrices on CPU+GPU platforms.

• To COMPILE BSOF/I codes, create the settings.make file to indicate where CUDA, CPU BLAS, MAGMA, and LAPACK are installed on your system. Example is given in zwolf.make, showing how to link to MKL, CuBLAS and MAGMA on zWolf platform in UC Davis. See also dirac.make for Dirac at NERSC.

• To TUNE CPU+GPU codes, you can modify the parameters inserted in file include/bsofi_params.h. The default values are tuned for zWolf. Parameters KAPPA_* control workload distribution between host and device. Parameters *_SWITCH_* specify when to switch between CPU and hybrid codes.

• To TEST BSOF/I, compile codes and run test.

  Note that on some systems the experimental code must be linked against sequential MKL. If you notice excessively poor performance, change make.inc to link against sequential MKL.

• To CALL BSOF/I subroutines in own codes, see examples ftest_hybrid.f90, ftest_cpu.f90 (Fortran) and ctest.c (C/C++).
  The general scheme for Fortran is shown below

    	integer  :: nnb, lw
    	integer  ::handler(10)
    	
    	! Initialize device and allocate GPU memory
    	call hybridXinitH(n, nb, nnb, handler)
    	
    	! Create working space on CPU
    	lw = hybridXbsofiLWorkH(n, nb, nnb, -1)
    	allocate(W1(lw))
    	
    	! Inversion
    	call hybridXbsoftriH(handler, n, nb, A, nnb, tauBsofi, W1, lw, info)
    	call hybridXbsoiH(handler, n, nb, A, nnb, tauBsofi, W1, lw, info)
    	
    	! Shutdown device and deallocate GPU memory
    	call hybridXfinalizeH(handler)`
  

For more INFORMATION, please refer to the technical report:

• Sergiy Gogolenko and Zhaojun Bai [A Structured Orthogonal Inversion of Block p-Cyclic Matrices on Multicores with GPU Accelerators] (http://www.cs.ucdavis.edu/research/tech-reports/2012/CSE-2013-78.pdf), Tech. rep. No. CSE-2013-78, 2013

mailto: [email protected]