An Efficient Sparse-Dense Matrix Multiplication on a Multicore System

摘要

Deep Neural Network (DNN) is currently widely used in various applications, such as speech recognition, computer vision, etc. The computation kernel of DNN-based applications is large sparse-dense matrix multiplication. As the performance of existing methods and software libraries for sparse matrix multiplication is not as good as expected, real-time recognition process has not been achieved yet. Therefore, we propose a novel sparse matrix storage format, block-based CSR (compressed storage format) and COO (coordinate format), called BCSR&BCOO, and a thread-scalable computing kernel for sparse-dense matrix multiplication, called BSpMM. We evaluate the performance of our proposed data structure and computing kernel in a real application in DNN-based online speech recognition. The experimental results demonstrate up to 4x speedup over Intel MKL on a typical CPU-based multicore system. Significant improvement in FLOPS is observed as well.