arXiv subjects

Roman Snytsar

Publications and source records attributed to Roman Snytsar.

5 recordsLinked to original sources

arXiv2023

Accelerating Machine Learning Primitives on Commodity Hardware

Sliding Window Sum algorithms have been successfully used for training and inference of Deep Neural Networks. We have shown before how both pooling and convolution 1-D primitives could be expressed as sliding sums and evaluated by the compute kernels with a shared structure. In this paper, we present an extensive study of the Sliding Window convolution technique as a more efficient alternative to the commonly used General Matrix Multiplication (GEMM) based convolution in Deep Neural Networks (DNNs). The Sliding Window technique addresses the memory bloating problem and demonstrates a significant speedup in 2-D convolution. We explore the performance of this technique on a range of implementations, including custom kernels for specific filter sizes. Our results suggest that the Sliding Window computation kernels can outperform GEMM-based convolution on a CPU and even on dedicated hardware accelerators. This could promote a wider adoption of AI on low-power and low-memory devices without the need for specialized hardware. We also discuss the compatibility of model compression methods and optimized network architectures with the Sliding Window technique, encouraging further research in these areas.

cs.LG↗

arXiv2023

Sliding Window Sum Algorithms for Deep Neural Networks

Sliding window sums are widely used for string indexing, hashing and time series analysis. We have developed a family of the generic vectorized sliding sum algorithms that provide speedup of O(P/w) for window size $w$ and number of processors P. For a sum with a commutative operator the speedup is improved to O(P/log(w)). Even more important, our algorithms exhibit efficient memory access patterns. In this paper we study the application of the sliding sum algorithms to the training and inference of the Deep Neural Networks. We demonstrate how both pooling and convolution primitives could be expressed as sliding sums and evaluated by the compute kernels with the shared structure. We show that the sliding sum convolution kernels are more efficient than the commonly used GEMM kernels on the CPU, and could even outperform their GPU counterparts.

cs.LG↗

arXiv2019

Parallel approach to sliding window sums

Sliding window sums are widely used in bioinformatics applications, including sequence assembly, k-mer generation, hashing and compression. New vector algorithms which utilize the advanced vector extension (AVX) instructions available on modern processors, or the parallel compute units on GPUs and FPGAs, would provide a significant performance boost for the bioinformatics applications. We develop a generic vectorized sliding sum algorithm with speedup for window size w and number of processors P is O(P/w) for a generic sliding sum. For a sum with commutative operator the speedup is improved to O(P/log(w)). When applied to the genomic application of minimizer based k-mer table generation using AVX instructions, we obtain a speedup of over 5X.

cs.DS↗

arXiv2019

De(con)struction of the lazy-F loop: improving performance of Smith Waterman alignment

Striped variation of the Smith-Waterman algorithm is known as extremely efficient and easily adaptable for the SIMD architectures. However, the potential for improvement has not been exhausted yet. The popular Lazy-F loop heuristic requires additional memory access operations, and the worst-case performance of the loop could be as bad as the nonvectorized version. We demonstrate the progression of the lazy-F loop transformations that improve the loop performance, and ultimately eliminate the loop completely. Our algorithm achieves the best asymptotic performance of all scan-based SW algorithms O(n/p+log(p)), and is very efficient in practice.

cs.DS↗

arXiv2019

Vectorized Character Counting for Faster Pattern Matching

Many modern sequence alignment tools implement fast string matching using the space efficient data structure called FM-index. The succinct nature of this data structure presents unique challenges for the algorithm designers. In this paper, we explore the opportunities for parallelization of the exact and inexact matches and present an efficient SIMD solution for the Occ portion of the algorithm. Our implementation computes all eight Occ values required for the inexact match algorithm step in a single pass. We showcase the algorithm performance in a multi-core genome aligner and discuss effects of the memory prefetch.

cs.DS↗