Overview

1. is written in ansi C, and uses standard  MPI and the LAPACK & ScaLAPACK libraries (either C or Fortran) for efficiency and portability.
2. is communication-intensive, and so works most efficiently when the matrices are dense on the processors.
3. performs all calculations in core in 8-byte precision, but stores all intermediate data products on disc as unformatted 4-byte float binary files.
4. stores all matrices on disc in full column-major form, but only guarantees the lower triangle of any symmetric matrix to be correct.
5. does not have a makefile since parallel systems differ so widely (even on the same hardware at different sites).
6. should be compiled with linkage to the local MPI, LAPACK & ScaLAPACK (and hence BLAS/BLACS/PBLAS) libraries, as well as the madcap.h header file and madtool.o object file.

The MADCAP header file includes machine-specific pre-processor instructions to make the code machine-independent. These cover

• REQUIRED datatype definitions for 4-byte integers, 4-byte floats and 8-byte floats to ensure that variables have the required sizes.
• OPTIONAL redefinition of malloc to allow for the use of more efficient but non-standard memory allocation.
• OPTIONAL redefinition of ScaLAPACK C-type function calls to meet local standards.
• REQUIRED definition of the function mad_char() used to pass character variables to ScaLAPACK function calls in the locally required manner.

The MADCAP tool code includes I/O and error-handling functions for

1. converting input/output between a 4-byte float binary file and an 8-byte float variable.
2. scattering/gathering matrices between disc and processors according to ScaLAPACK block cyclic distribution.
3. error-checking and distributed error-handling.