Research On Heterogeneous Multi-Core Processor For Cell Image Processing

With the wide application of cell image processing in clinical medical diagnosis,the real-time requirements of cell image processing are getting higher and higher.Since isomorphic multi-core processors have limited computational speed when processing cell images,it is important to study heterogeneous multi-core processors with different acceleration instructions for cell image processing.A multi-core processor and cell image processing algorithm was designed to design a heterogeneous multi-core processor for cell image processing.The processor has eight cores,each of which is a two-stage pipeline processor.The zeroth kernel sets the input buffer as the buffer for the cell image,and the eight cores access the shared data memory pool with four SRAM banks through the crossbar.The crossbar preferentially encodes data with a low core address.When two cores access the same SRAM bank at the same time,the lower-address kernel preferentially accesses it.The multi-core processor adopts a producer-consumer parallel programming model with no lock structure to realize the cooperation of multiple cores for storing and reading data.Algorithms for cell image processing use K-nearest neighbor smoothing filter,full-scale histogram grayscale stretching,Gaussian blur,USM sharpening,OTSU and connected domain labeling algorithms,and design Absort instructions,Inireg instructions,Sumreg instructions in different kernels.The Sum4 instruction,the Abs instruction,the Mulsub instruction,and the Mulpow instruction accelerate the algorithm,which improves the kernel speed of the K-nearest neighbor smoothing algorithm by 2.66 times,and runs the full-scale histogram gray-scale stretching and the speed of the Gaussian fuzzy algorithm kernel.At 1.21times,the speed of running the USM sharpening algorithm kernel increased by 1.36 times,and the speed of running the OTSU kernel increased by 1.17 times.The heterogeneous multi-core processor is simulated and integrated under the UMC110nm process.The image processing speed is 203.11 frames per second when each core runs the corresponding algorithm,which is 10.83 times faster than the single-core structure.The highest clock frequency is 136.98.MHz with a total area of 19.0 mm~2.