Optimizing Deep Learning RNN Topologies on Intel Architecture

Abstract

Recurrent neural network (RNN) models have been found to be well suited for processing temporal data. In this work, we present an optimized implementation of vanilla RNN cell and its two popular variants: LSTM and GRU for Intel Xeon architecture. Typical implementations of these RNN cells employ one or two large matrix multiplication (GEMM) calls and then apply the element-wise operations (sigmoid/tanh) onto the GEMM results. While this approach is easy to implement by exploiting vendor-optimized GEMM library calls, the data reuse relies on how GEMMs are parallelized and is sub-optimal for GEMM sizes stemming from small minibatch. Also, the element-wise operations are exposed as a bandwidth-bound kernel after the GEMM which is typically a compute-bound kernel. To address this discrepancy, we implemented a parallel blocked matrix GEMM in order to (a) achieve load balance, (b) maximize weight matrix reuse, (c) fuse the element-wise operations after partial GEMM blocks are computed and while they are hot in cache. Additionally, we bring the time step loop in our cell to further increase the weight reuse and amortize the overhead to transform the weights into blocked layout. The results show that our implementation is generally faster than Intel MKL-DNN library implementations, e.g. for RNN, forward pass is up to ~3× faster whereas the backward/weight update pass is up to ~5× faster. Furthermore, we investigate high-performance implementations of sigmoid and tanh activation functions that achieve various levels of accuracy. These implementations rely on minimax polynomial approximations, rational polynomials, Taylor expansions and exponential approximation techniques. Our vectorized implementations can be flexibly integrated into deep learning computations with different accuracy requirements without compromising performance; in fact, these are able to outperform vectorized and reduced accuracy vendor-optimized (Intel SVML) libraries by 1.6–2.6× while speep up over GNU libm is close to two orders of magnitude. All our experiments are conducted on Intel’s latest CascadeLake architecture.