[QST] Epilogue Reduction

[jeromeku]

What is your question? I'm looking to define a GEMM that does the following (in pseudocode):

D = AB + C
F = norm(D, axis=1)
return F

That is, the epilogue should a) compute the column-wise 2-norm of D and b) store F to global, no need to store D. (2-norm being the sqrt of the sum of squares along axis=1).

What's the most appropriate epilogue type for this pattern specific for Ampere?

• It seems like EVT would fit this well -- are there examples for this NOT for stream-k? Always get compilation errors when trying to instantiate an EVT device GEMM for Ampere (see #1459).
• Would it be possible to adapt a convolution epilogue that does a full reduction across the channel dimension? The GEMM epilogue reduction unit tests only do partial reductions (understandably so).

[github-actions[bot]]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.