Automatic Loop Tuning and Memory Management for Stencil Computations

The Texas Instruments C66x Digital Signal Processor (DSP) is an embedded processor technology that is targeted at real time signal processing. It is also developed with a high potential to become the new generation of coprocessor technology for high performance embedded computing. Of particular interest is its performance for stencil computations, such as those found in signal processing and computer vision tasks. A stencil is a loop in which the output value is updated at each position of an array by taking a weighted function of its neighbors. Efficiently mapping stencil-based kernels to the C66x device presents two challenges. The first one is how to efficiently optimize loops in order to facilitate the usage of Single Instruction Multiple Data (SIMD) instructions. On this architecture, like most others, SIMD instructions are not directly generated by the compiler. The second problem is how to manage on-chip memory in a way that minimizes off-chip memory access. Although this could theoretically be achieved by using a highly associative cache, the high rate of data reuse in stencil loops causes a high conflict miss rate. One way to solve this problem is to configure the on-chip memory as a program controlled scratchpad. It allows user to buffer a 2D block of data and minimizes the off-chip data access. For this dissertation, we have accomplished two goals: (1) Develop a methodology for optimization of arbitrary 2D stencils that fully utilize SIMD instructions through microachitecture-aware loop unrolling. (2) Deliver an easy-to-use scratchpad buffer management system and use it to improve the memory efficiency for 2D stencils. We show in the results and analysis section that our stencil compiler is able to achieve up to 2x speed up compared with the code generated by the industrial standard compiler developed by Texas Instruments, and our memory management system is able to achieve up to 10x speed up compared with cache