Precision–Energy–Throughput Scaling of Generic Matrix Multiplication and Convolution Kernels via Linear Projections

摘要: Generic matrix multiplication (GEMM) and con- volution (CONV)/cross-correlation kernels often constitute the bulk of the compute- memory-intensive processing within image/audio recognition matching systems. We propose a novel method to scale energy throughput of GEMM CONV for such error-tolerant multimedia applications by adjusting precision computation. Our technique employs linear projections input or signal data during top-level GEMM blocking and reordering. The kernel then uses projected inputs results are accumulated to form final outputs. Throughput scaling takes place changing number computed by each kernel, which in turn produces approximate results, i.e., changes performed Results derived from a voltage- frequency-scaled ARM Cortex A15 processor running face music-matching algorithms demonstrate that proposed approach allows for a 280%–440% increase throug hput 75%– 80% decrease consumption against optimized GEMM without any impact on obtained recognition or accuracy. Even higher gains can be obtained, if one is willing tolerate some reduction the accuracy applications