<html> <head> <base href="https://bugs.freedesktop.org/" /> </head> <body> <div> <a class="bz_bug_link bz_status_NEW " title="NEW - Add FP64 support to the i965 shader backends" href="https://bugs.freedesktop.org/show_bug.cgi?id=92760#c33">Comment # 33</a> on <a class="bz_bug_link bz_status_NEW " title="NEW - Add FP64 support to the i965 shader backends" href="https://bugs.freedesktop.org/show_bug.cgi?id=92760">bug 92760</a> from <a class="email" href="mailto:cwabbott0@gmail.com" title="Connor Abbott <cwabbott0@gmail.com>"> Connor Abbott</a> <pre>(In reply to Samuel Iglesias from <a href="show_bug.cgi?id=92760#c32">comment #32</a>) > I found an issue related to the opt_algebraic passes after we lowered lrp > instruction when it has double-based types as arguments (lrp instruction > does not support doubles in BDW). > > Currently, the lowering of double ops run as part the optimization loop in > nir_optimize(). In that lowering function, I converted the lrp to its > equivalent mathematical function when we have doubles as arguments... which > is reverted by opt_algebraic() in next iteration of the optimization loop :-) > > My idea is: remove my lrp's lowering pass, add a new flag in > nir_shader_compiler_options called 'lower_lrp_doubles' and modify > nir_opt_algebraic.py to run its lrp's lowering when the arguments are > doubles, if that flag is enabled. Any driver that does not support lrp with > double arguments, just enable that nir option in its configuration. > > I would like to know your opinion about this solution before implementing > it. Do you have other solutions in mind? > > Thanks! That seems like a fine solution -- we already do similar things in a number of different places in opt_algebraic. Just keep in mind that you'll have to disable the lrp optimization in opt_algebraic for doubles when that option is enabled to avoid running into the same problem.</pre> </div> <hr> You are receiving this mail because: <ul> <li>You are the QA Contact for the bug.</li> </ul> </body> </html>