Optimizing GPGPU Kernel Summation for Performance and Energy Efficiency

摘要

Kernel summation is a widely used computational kernel that involves matrix-matrix multiplication (GEMM) and matrix-vector multiplication (GEMV) computational primitives. The parallelism exhibited in kernel summation suggests performance improvement when running on GPGPU. State of the art GPU solutions apply cuBLAS library but cannot exploit much of the data locality because intermediate results are written back to main memory in between key operations. This paper presents an optimized implementation that yields better performance and high energy efficiency. Our contributions are fusing all steps of kernel summation into the matrix multiplication code structure and optimizing memory access ordering to make good use of shared memory and cache hierarchy. We decompose the kernel summation problem into individual tasks with few dependencies and strike a balance between finer grained parallelism and reduced data replication. Based on hardware characteristics, we map threads to matrix elements in an interleaved way, and reposition matrix elements to avoid shared memory load and store bank conflicts. We also apply double buffering to hide memory access latency. We analyze both performance and energy benefits of our fused kernel summation compared with the implementation based on cuBLAS. We show that in low dimensions our approach achieves a speedup of up to 1.8X, and saves up to 33% of total energy in all tested problem sizes.