Optimizing strassen matrix multiply on GPUs

Many core systems are basically designed for applications having large data parallelism. Strassen Matrix Multiply (MM) can be formulated as a depth first (DFS) traversal of a recursion tree where all cores work in parallel on computing each of the NxN sub-matrices that reduces storage at the detriment of large data motion to gather and aggregate the results. We propose Strassen and Winograd algorithms (S-MM and W-MM) based on three optimizations: a set of basic algebra functions to reduce overhead, invoking efficient library (CUBLAS 5.5), and parameter-tuning of parametric kernel to improve resource occupancy. On GPUs, W-MM and S-MM with one recursion level outperform CUBLAS 5.5 Library with up to twice as faster for large arrays satisfying N>=2048 and N>=3072, respectively. Compared to NVIDIA SDK library, S-MM and W-MM achieved a speedup between 20x to 80x for the above arrays. The proposed approach can be used to enhance the performance of CUBLAS and MKL libraries.