Wavelet transform is an important branch as a new theory in the history of mathematics. It is a signal analyzing and processing method in time-frequency domain, and has an excellent local character in time domain and frequency domain. Wavelet transform has led to many new wavelets applications in various fields such as signal detection, signal processing and image processing, and the hardware realization of wavelet transform algorithms receives more and more recognition. This paper introduces the principle of fast wavelet transform and the realizing method in DSP. The DSP instructions such as parallel multiply-accumulate instruction, circular addressing and repeating operation in TMS320C3X of, which are suitable to realize the wavelet algorithms, are described in detail. A framework of assemble language program for the calculation of scale expanding value is given out and the problem which should be paid attention to in the fast wavelet algorithms using DSP is discussed specially.
A significant trend in machine learning is sparsifying the training of neural networks to reduce the amount of computation required. Algorithms like Sub-LInear Deep learning Engine (SLIDE) [2] use locality-sensitive hashing (LSH) to create sparsity. These sparse training algorithms were originally developed on multi-threaded multicore CPUs. However, they are not well-studied and optimized for accelerator platforms such as GPUs and reconfigurable dataflow architectures (RDAs). In this paper, we study the different variants of the SLIDE algorithm and investigate accuracy-performance tradeoffs on CPU, GPU, and RDAs. The implementation targeting RDA outperforms the GPU by 7.5×. The performance on a limited-memory RDA is improved further by proposing a smart caching algorithm, which is 2 × faster than the baseline RDA. Furthermore, we are able to achieve another 2 × performance by putting all of the weights on-chip using an RDA with enough memory. We believe our work will pave the road for the future development of both algorithm and hardware architecture for sparse training.