Implementing Level-3 BLAS Routines in OpenCL on Different Processing Units

Report Date: Written Language: Any Other Identifying Information of this Report: Distribution Statement: Supplementary Notes: The University of Aizu Aizu-Wakamatsu Fukushima 965-8580 Japan 10/22/2014 English First Issue: 10 copies Kazuya Matsumoto, Naohito Nakasato, and Stanislav Sedukhin Implementing Level-3 BLAS Routines in OpenCL on Different Processing Units Level-3 BLAS, GPU, multi-core CPU, many-core processor, OpenCL, performance porting, auto-tuning This paper presents an implementation of different matrix-matrix multiplication routines in OpenCL. We utilize the high-performance GEMM (GEneral Matrix-Matrix Multiply) implementation from our previous work for the present implementation of other matrix-matrix multiply routines in Level-3 BLAS (Basic Linear Algebra Subprograms). The other routines include SYMM (Symmetric MatrixMatrix Multiply), SYRK (Symmetric Rank-K Update), SYR2K (Symmetric Rank-2K Update), and TRMM (Triangular Matrix-Matrix Multiply). A key in our approach is to copy given matrix data by copying OpenCL kernels into a form such that a high-performance GEMM kernel can be utilized for computation. We use a previously developed auto-tuning system for the highly optimized copying kernels as well as for GEMM kernel. The performance evaluation of our implementation is conducted on four different GPUs (AMD Radeon R9 290X, FirePro W9100, Radeon HD 7970, and NVIDIA GeForce GTX Titan), a many-core processor (Intel Xeon Phi 5110P), and a multi-core processor (Core i7 3960X). The evaluation results show that the tuning on the copying kernels is effective and contributes to develop high-performance BLAS3 routines. Distributed Parallel Processing Laboratory