On Tue, 2014-07-22 at 19:40 -0700, Kenneth Graunke wrote: > On Tuesday, July 22, 2014 11:38:05 AM Ilia Mirkin wrote: > > On Tue, Jul 22, 2014 at 11:25 AM, Samuel Iglesias Gonsálvez > > <sigles...@igalia.com> wrote: > > > Hello, > > > > > > I have a doubt related to Streamed Vertex Buffer Write message and its > > > message descriptor in SandyBridge. > > > > > > Reading about Stream Output Primitives Written > > > (snb_ihd_os_vol2_part1.pdf, pag 175), it says the following: > > > > > > "Whenever a GS thread outputs a DataPort Streamed Vertex Buffer Write > > > (SVBWrite) message with the Increment Num Prims Written bit set, the > > > SO_NUM_PRIMS_WRITTEN register will be incremented." > > > > > > According to SNB's spec [0], all the bits in the SVBWrite message > > > descriptor are ignored and I have not found the definition of this bit > > > in the documentation. Maybe I'm running the wrong pdfgrep command :-/ > > > > > > After doing some Google searches, I found that in other doc [1] it's > > > mentioned the "Increment Num Prims Written" bit inside the SVBWrite > > > message descriptor is not ignored for SNB (at least, it doesn't say that > > > explicitly). > > > > > > Do you know where is the "Increment Num Prims Written" bit defined in > > > SNB's doc? > > > > Not 100% sure if this is what you're looking for, but take a look at > > > > https://01.org/linuxgraphics/sites/default/files/documentation/snb_ihd_os_vol4_part2.pdf > > > > Page 23, 2.4.3.1 - SONumPrimsWritten and/or Page 29 - 2.4.4.1 > > NumGSPrimsGenerated and perhaps some of the other fields in the > > FF_SYNC message. > > > > Or wait until someone who knows what they're talking about replies :) > > > > -ilia > > What Ilia said :) > > It looks like SO_PRIM_STORAGE_NEEDED is updated by FF_SYNC messages, and > SO_NUM_PRIMS_WRITTEN is incremented by the URB_WRITE message (not SVB writes). > > Every GS thread terminates with a URB write message - apparently, there's a > field in the URB write message header (not descriptor) which determines how > much the counter should be incremented by. > > It looks like you /used/ to increment both counters via SVBWrite messages on > GM45 (SNB PRM, Volume 2, Part 1, Page 57), but that changed.
Thanks for the info :) Sam
signature.asc
Description: This is a digitally signed message part
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
