Saagar Jha

I keep hearing about sufficiently smart compilers but have we considered working on getting rid of sufficiently dumb compilers
“What If You Wanted to Go To Heaven” meme template. The first panel is unchanged and says “What if you”. The second panel has been completely replaced with a screenshot of an attention pipeline from https://www.modular.com/blog/software-pipelining-for-gpu-kernels-part-1-the-pipeline-problem, with text “wanted to write a deep pipeline”. The third panel is almost unchanged, except the text that used to say “But god said” now has an Nvidia logo on top of “god”. The last picture has the text replaced with a screenshot of ptxas output: ptxas info : (C7517) warpgroup.wait is injected in around line 52 by compiler to allow use of registers defined by GMMA in function 'mma' ptxas info : (C7514) Potential Performance Loss: wgmma.mma_async instructions are serialized due to non wgmma instructions reading accumulator registers of a wgmma between start and end of the pipeline stage in the function 'mma'