mboehm7 commented on pull request #1385: URL: https://github.com/apache/systemds/pull/1385#issuecomment-915533612
Thanks for staring the discussion, a few additional points to consider (besides separating it from the compression-related changed): * we need to check for know dims, otherwise the rewrite also triggers if hi.getDim2() is unknown (-1), and as we don't consider the other way around, it would never come back during recompilation * The overall benefit comes from reduced allocation in the cbind output (while incurring more floating point operations in the matrix multiply in some kernels that avoid branches in the inner loop). Accordingly, we cannot just look at `hi.getDim1() > hi.getDim2() * 2`. Instead given a m-x-n matrix X and n-x-k matrix Y, we have to compare the following: (a): m*k + m*(k+1) versus (b) n*k + m*(k+1), which reduces to comparing m and n, not m and (k+1), isn't it? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
