00:00dboyan: I've been messing with nvenc recently. The blob's usage of "unified virtual memory" made my life really hard.
00:01imirkin: dboyan: yeah, i've been reading the logs... try a non-uvm blob version :)
00:02imirkin: although... valgrind should detect such mappings? not sure.
00:02imirkin: perhaps it doesn't right now
00:02imirkin: [the mmt tool, that is]
00:02imirkin: anyways, i assume it'll work largely like vdec
00:02imirkin: i.e. you feed it pointers to buffers
00:02imirkin: and then tell it to encode
00:02imirkin: and then it encodes.
00:04dboyan: mmt doesn't detect it. I'm manually decoding command stream for now.
00:05imirkin: if you're decoding it, mmt does detect it :p
00:05imirkin: maybe you mean demmt doesn't show it?
00:05imirkin: anyways, i'd recommend writing a simple sample app that takes a small frame of data and feeds it in, and recording its output
00:05imirkin: and then searching for those bytes in the trace
00:06dboyan: The bad thing is demmt cannot properly detect IB.
00:06imirkin: something like i did for decoding - https://github.com/imirkin/re-vp2 -- look at the h264_player.c -- that used to be a much simpler application at first
00:06dboyan: And uvm just hides frame data transfer
00:06imirkin: well, you can force it to find a IB by hacking it
00:07imirkin: and then this was my proof-of-concept for actually decoding stuff: https://github.com/imirkin/re-vp2/blob/master/decode_frame.c
00:08dboyan: a non uvm version still runnable on my system is 340, but my nvenc demos (official and my own) fails to start as promised
00:08dboyan: I've seen your re-vp2 repo
00:09dboyan: Trying to learn from that
00:11imirkin: it's obviously for a totally different thing, but i think it's the same basic approach
00:11imirkin: the problem is that tracing something like ffmpeg, etc -- it'll be using the api in a complex way
00:11imirkin: and doing a ton of stuff that is tangential to what you're aiming for
00:11imirkin: having a controlled environment is very beneficial
00:13dboyan: Yeah, I use nvidia's encode sdk, that much more controllable, encoding yuv to h264 bitstream
00:14imirkin: ok good
23:46karolherbst: okay nice, my clk_update code works without issues now :)
23:48karolherbst: except that NvClkMode parameter causes troubles now :(
23:50karolherbst: uhh, race condition
23:56karolherbst: skeggsb: any idea how nouveau could get a NULL pointer deref here? https://github.com/karolherbst/nouveau/blob/master_4.10/drm/nouveau/nvkm/engine/fifo/gk104.c#L653
23:59imirkin_: karolherbst: i sent a patch which should fix that LODQ issue - https://patchwork.freedesktop.org/patch/142173/