Design And Implementation Of High-performance DMA Transmission Modes For Scientific Computation On GPDSP

Scientific computation is an important field of the modern computing science research, and it involves the algorithm research and the design of micro-processors. GPDSP processor is a multi-core processor designed by our university, and it has the advantages of both general CPU and DSP.Through analyzing the essence of HPL, the matrix update arithmetic is the main factor that decides the HPL executing efficiency, and the matrix update arithmetic is done by the GEMM algorithm. The GEMM algorithm based on the thought of GEPP-GEBP is able to achieve the highest efficiency among many GEMM algorithms.This text designed the DMA's special transmission modes through the analyzing of GPDSP's Architecture characteristics and GEMM algorithm scheme. DMA's special transmission modes include DMA Matrix Transposition Transmission, DMA Segment Transmission, DMA Inter-core Synchronizing Transmission, and DMA Block-segment Transmission which is designed on basis of DMA Segment Transmission.DMA Matrix Transposition Transmission is able to transmit the 2D data from the source memory to the destination memory, and make the 2D data transposited during this transmission. It can promote the Matrix Multiplication's arithmetic efficiency greatly. Through test, its transmission efficiency can be 1.56 times higher than the traditional matrix transposition transimission.The general thought of GEMM algorithm is to divide the block data into many parts and these divided parts are transmitted into the core to be computed, and then the computed results transmitted to the outside memory to be synchronized. Therefore, this text designed DMA Segment Transmission, DMA Inter-core Synchronizing Transmission and DMA Block-segment Transmission. The DMA Segment Transmission can transmit the data from outside to inside of cores quickly, and DMA Inter-core Synchronizing Transmission can tansmit the data of many cores to the outside memory, and moreover, DMA Block-segment Transmission can hide the data transmitting time effectively. According to the test on the simulation and verification platform of Cadence Company's NC-VERILOG, the special transmission modes can achieve higher performance than the traditional transmission modes. The speed of DMA Segment Transmission is 1.24 times higher than that of the traditional transmission. And DMA Block-segment Transmission can make the time of GEMM algorithm reduce at least 3000 cycles. The average speed of DMA Inter-core Synchronizing Transmission is 2.56 times the traditional transmission's speed.Through enough verification and experiment test, the DMA special transmission modes satisfy the function need, and could promote HPL's executing efficiency.