L1 cache sharing method for GPU

Abstract

The invention discloses an L1 cache sharing method for a GPU (Graphics Processing Unit), which comprises the following steps: S11, judging whether a local memory access request is empty or not, if so, executing S21, and if not, executing S12; s12, taking out the request to access the L1 cache; s13, judging whether the target is hit or not, if yes, returning data, and if not, executing S14; s14, judging whether the program is a storage-intensive program or not, if so, sending the request to other SM and executing S15, and if not, sending the request to L2 cache; s15, judging whether a cache data block needs to be replaced or not, and if yes, sending a data block replacement request to other SM; s21, judging whether the far-end memory access request is empty or not, and if not, executing S22; s22, taking out the request to access the L1 cache; s23, judging whether the request is hit or not, if yes, returning data, and executing S24, and if not, sending the request to the L2 cache, and executing S24; and S24, judging whether the far-end data request is empty or not, and if not, storing the data block needing to be replaced into the L1 cache. According to the invention, the operation of the storage-intensive program and the use of the L1 cache on the SM of the operation calculation-intensive program can be realized.

Description

technical field [0001] The invention relates to the technical field of GPUs, in particular to an L1 cache sharing method for GPUs. Background technique [0002] Graphics Processing Unit (GPU) is a microprocessor used to do image and graphics-related calculations. GPU is widely used in cloud computing platforms and data centers because of its powerful computing capabilities, providing users with the calculation. Compared with a single-task GPU that only runs one task on the GPU, a multi-task GPU can run multiple tasks on the GPU at the same time, which can effectively improve resource utilization. Specifically, a multi-task GPU can simultaneously run a computation-intensive program and a storage-intensive program on one GPU, and the computation resources and storage resources on the GPU can be fully utilized at the same time. [0003] At present, the spatial multitasking method is mainly used to realize the GPU to run multiple tasks at the same time. Specifically, in the sp...

AI Technical Summary

Problems solved by technology

[0004] However, although spatially multitasking GPUs that run computation-intensive programs and memory-intensive programs at the same time can effectively improve the overall resource utilization of the system, running different programs on different SMs will cause resources on the SMs, especially the L1 cache (Level 1 cache, L1 Cache) Resource utilization is unbalanced, which affects the further improvement of multi-tasking GPU performance

Specifically, for SMs running memory-intensive programs, such programs will generate a large number of memory access requests, resulting in excessive use of L1 cache resources, high L1 cache failure rate, and invalid requests through the on-chip interconnection network It is sent to the L2 cache (secondary cache, L2 Cache) and the storage system, which will bring a large memory access overhead; for SMs running computationally intensive programs, there are very few memory access requests for such programs, Causes L1 cache resources to be underutilized

Images