Open broskoTT opened 1 month ago
Investigate perf drawback on a path of writing to device. @joelsmithTT created a proposal for TLBWindows which should improve the perf here, although there is already tt::Writer added to umd which essentially does this, but there is still room for improved design: https://docs.google.com/document/d/1uvvLeV1-wT9masV8T2yQpX8xe-1bh5uFIlM-N4dh3S8/edit#heading=h.x1vn32gnpys7
I think that write_* interfaces could also be improved, there are many unordered_maps which degrade perf on each write. Metal folks also mentioned this issue here: https://docs.google.com/presentation/d/1kR7BswtHLS9J6dda2illQ_FV4IRJ2I28/edit#slide=id.p1
Investigate perf drawback on a path of writing to device. @joelsmithTT created a proposal for TLBWindows which should improve the perf here, although there is already tt::Writer added to umd which essentially does this, but there is still room for improved design: https://docs.google.com/document/d/1uvvLeV1-wT9masV8T2yQpX8xe-1bh5uFIlM-N4dh3S8/edit#heading=h.x1vn32gnpys7
I think that write_* interfaces could also be improved, there are many unordered_maps which degrade perf on each write. Metal folks also mentioned this issue here: https://docs.google.com/presentation/d/1kR7BswtHLS9J6dda2illQ_FV4IRJ2I28/edit#slide=id.p1