strassen | Strassen matrix multiplication on heterogenous hardware

There is no contradiction to apply this programming model on M1 processor. However for developer like me (C/C++), Apple does not provide any C++-API about their scheduler (grand central), Metal, and Neural Engine. It is a bit sad, because these technologies are developed in C/C++ but not open to the public. Indeed the key operation of deep learning is SGEMM, however the wait of legacy and BLAS/LAPACK has privileged contiguous buffer for matrix representation. Inside Tensor Flow the heterogenous computing is performed on node which represent operation coming from layer, this node can be executed on devices. I can imagine TF working on tile matrices where every nodes proposed heterogenous solver. Well Every vendor (will) proposes heterogenous microchip. The last example of Apple is the most significant. After the question of the programming model and granularity is open. I consider the management of the memory the most important, specially in this kind of algorithm (Strassen). If the memory is not correctly managed get good performance is impossible. In the CUDA-sdk it is well demonstrated that the memory management is the most important (stride, etc ..). Indeed I could but it is very time consuming, and it will be more an academic work, not the purpose of this note.