Abstract:
The execution time, resource and energy costs of deep learning applications become much more important as their popularity grows. The Constant Matrix Multi plication has been studied for a long time and takes place in deep learning applications. Reducing the computation cost of those applications is a highly active research topic. The weights are pruned or quantized while satisfying the desired accuracy requirement. The pruned matrices are compressed into one-dimensional arrays without data loss. Matrix multiplication is performed by processing those arrays without decompression. Processing one-dimensional arrays to perform matrix multiplication is deployed on vari ous hardware platforms that employ Central Processing Unit, Graphics Processor Unit and Field-Programmable Gate Array. The deployments can also be supported with common subexpression elimination methods to reduce the number of multiplications, additions and storage size. However, the state-of-the-art methods do not scale well for the large constant matrices as they reach hours for extracting common subexpressions in a 200 × 200 matrix. In this thesis, a random search-based common subexpression elimination method is constructed to reduce the run-time of the algorithm. The algo rithm produces an adder tree for a 1000 × 1000 matrix in a minute. The Compressed Sparse Row format is extended to build a one-dimensional compression notation for the proposed method. Simulations for a single-core embedded system show that the latency is reduced by 80% for a given 100×100 matrix compared to the state-of-the- art methods. The storage size of the sparse matrices is also reduced by more than half in the experiments compared to the Compressed Sparse Row format.
