What's the main idea for irregular-shaped GEMM?

[zhuangbility111]

trafficstars

Hi! Your job is pretty good! I have a question about irregular-shaped GEMM: How to implement the irregular-shaped GEMM more efficiently? What's the main idea?

[AnonymousYWL]

Hello! we overlap packing and computing. Efficient handling of edge cases And use more efficient parallelization methods.

[zhuangbility111]

I have read your SC paper. It's a great job! I have a question. How do you implement the parallel packB? It seems that it needs some barriers in parallel packB when the parallel mode is 2d, but I don't see any synchronizations in your code.

[AnonymousYWL]

Thank you for your attention!  The packing of B is private to each thread, and this process overlaps with the calculation. Therefore, we do not need to synchronize this operation.

------------------ 原始邮件 ------------------ 发件人: "AnonymousYWL/LibShalom" @.>; 发送时间: 2021年11月22日(星期一) 中午1:18 @.>; @.@.>; 主题: Re: [AnonymousYWL/LibShalom] What's the main idea for irregular-shaped GEMM? (Issue #1)

I have read your SC paper. It's a great job! I have a question. How do you implement the parallel packB? It seems that it needs some barriers in parallel packB when the parallel mode is 2d, but I don't see any synchronizations in your code.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

[zhuangbility111]

So you mean that since the overhead of packing B can be amortized by the calculation, even though private packing B will cause redundant packing B, it dosen't decrease the overall performance?

[AnonymousYWL]

Yes, this may be improved in the future.