JuliaGPU/CUDA.jl

Summing a vector is faster than summing a multidimensional array

Open

#2,764 创建于 2025年4月28日

在 GitHub 查看
 (1 评论) (0 反应) (0 负责人)Julia (1,408 star) (274 fork)batch import
cuda arraygood first issueperformance

描述

I am not sure if this is an expected behavior, but it seems that sum(vec(A)) could be much faster than sum(A) when A is a multi-dimensional array. Please see https://discourse.julialang.org/t/summing-a-vector-is-faster-than-summing-a-multi-dimensional-array-of-the-same-length-using-cuda/116711 for an MWE. Although this post is from about a year ago, I can reproduce matching results with CUDA v5.7.3.

贡献者指南

技术栈
无
领域
performance
议题类型
bug
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
half day
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
fresh
清晰度议题是否清楚说明期望改动、验收标准和下一步。
clear
前置要求
Julia programmingCUDA basicsArray operations
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
35
研究方向
Investigate the `sum` function implementations in CUDA.jl, particularly for multidimensional arrays versus flattened vectors. Check the linked Discourse post for the MWE. The performance difference may be due to kernel launch overhead or memory access patterns. Look at `src/kernels/reductions.jl` or similar files. Compare with `sum` on the CPU to isolate the GPU specific issue.