S7280 - CLBlast: A Tuned BLAS Library for Faster Deep Learning

We'll demonstrate how to accelerate dense linear algebra computations using CLBlast, an open-source OpenCL BLAS library providing optimized routines for a wide variety of devices. It is targeted at deep learning training and inference and thus provides a fast matrix-multiplication routine (GEMM) to accelerate the convolutional layers: the computational heart of all deep-learning frameworks (TensorFlow, Caffe, etc.). CLBlast has three main advantages over other BLAS libraries: 1) it can be explicitly tuned for specific matrix-sizes and hardware platforms, 2) it runs on less common devices (and it is fast), such as embedded and low-power GPUs, and 3) it can perform operations in half-precision FP16 format, saving precious bandwidth, time, and power.