Contributing a complete GEMM implementation

This feature has been temporarily disabled in 3.8, though it may be re-enabled in the 3.9 series if there is user demand. This section is therefore kept around solely historical purposes, and will need to be updated if the feature is added back in.

Contributing an L1 kernel is the prefered method of user contribution for Level 3 BLAS speedup, but it is not the only one supported by ATLAS. ATLAS also allows a user to contribute a complete system-specific GEMM implementation. This method of contribution is far less desirable than kernel contribution, and thus the standards of acceptance are correspondingly higher.

When only a kernel is contributed, it is only used when timings indicate it is superior to the best ATLAS-supplied routine for a given architecture. Because kernel routines are called in known ways by the ATLAS infrastructure, the timer can be made to accurately reflect typical usage. A full GEMM, which is to all intents called directly by the user, has no ``typical'' usage, and the timer is thus not able to ensure that the user's full GEMM is superior to that supplied by ATLAS in a system-independent way, even if the additional installation time required to choose amoung full GEMM implementations were allowed. Thus, full GEMM implementations will be used only when ATLAS's configuration detects a known architecture where the ATLAS team has certified the full GEMM to be significantly better than ATLAS's native GEMM, across the entire spectrum of problem shapes and sizes (with the exception of those shapes and sizes handled by ATLAS's non-copy code, as explained below).

As explained in Section 2.1, ATLAS has both a small-case matmul, which does not copy the user's input operands, and a large-case code that does. The user contributed GEMM replaces ATLAS's large-case GEMM, and then timings are used as normal to determine the crossover points at which the contributed GEMM outperforms ATLAS's small-case code.