Remember that what ATLAS is doing is substituting your GEMM for its own large-case GEMM. However, ATLAS's large-case GEMM is still compiled in the library, it is just not being used. The following code will call ATLAS's large case code for all precisions:
#include "atlas_misc.h"
#include "atlas_lvl3.h"
int Mjoin(Mjoin(ATL_U,PRE),usergemm)
(const enum ATLAS_TRANS TA, const enum ATLAS_TRANS TB,
const int M, const int N, const int K, const SCALAR alpha,
const TYPE *A, const int lda, const TYPE *B, const int ldb,
const SCALAR beta, TYPE *C, const int ldc)
{
int ierr;
if (N >= M)
{
if ( ierr = Mjoin(PATL,mmJIK)(TA, TB, M, N, K, alpha, A, lda, B, ldb,
beta, C, ldc) )
ierr = Mjoin(PATL,mmIJK)(TA, TB, M, N, K, alpha, A, lda, B, ldb,
beta, C, ldc);
}
else
{
if ( ierr = Mjoin(PATL,mmIJK)(TA, TB, M, N, K, alpha, A, lda, B, ldb,
beta, C, ldc) )
ierr = Mjoin(PATL,mmJIK)(TA, TB, M, N, K, alpha, A, lda, B, ldb,
beta, C, ldc);
}
return (ierr);
}
So, you can use the above code to have ATLAS call it's normal routines for those precisions you do not supply, and call your own otherwise. In order to help understand this process, ATLAS includes a ``user-supplied'' GEMM that simply calls ATLAS's own GEMM for all precisions, in ATLAS/src/blas/gemm/UMMEXAMPLE. If, for instance, you have a single precision real GEMM but nothing else, you can take this directory as a starting point, adding your own commands for single precision real in the Makefile, etc, and leaving everything else alone.
In order to do this, you should do the following: