14:55 karolherbst: imirkin: do you know if how the GPU reacts if we have a window into constant memory in global memory and access this range from a shader?
14:56 karolherbst: do we get the benefits of the cached data or not?
14:57 karolherbst: ohh.. apparently that only works with local and shared memory only :/
14:57 karolherbst: sad
16:18 karolherbst: fun.. I can't dump shaders via mmt tracing anymore :/ uff, what's wrong this time
16:25 HdkR: karolherbst: Still want a PC to see if it works for CI?
16:27 karolherbst: huh?
16:28 karolherbst: I am more interested in actually dumbing the kernel nvidia generates
16:28 HdkR: Throwing you hardware for CI remember? :)
16:28 karolherbst: ohhh
16:28 karolherbst: that thing
16:31 HdkR: Just trying to send you hardware :P
16:40 karolherbst: ehh "Warning: noted but unhandled ioctl 0x30000001 with no size/direction hints."
16:41 karolherbst: HdkR: I first have to wire up my jetson for CI :D
16:41 HdkR: karolherbst: I'm fine with you having a pile of hardware before that :P
16:46 karolherbst: HdkR: wait until christmas and send everything at once :D
16:47 HdkR: karolherbst: https://pcpartpicker.com/list/qcwKZf Plan was to dump this pile of stuff on you right away and then if it works, send you more of it
16:48 karolherbst: HdkR: I mean... if you have time you could also just wire it up to the gitlab CI stuff :D
16:48 HdkR: I don't have the space :P
16:50 HdkR: I can't even keep my 2990WX system once I upgrade it since i don't have room to stuff it somewhere
16:50 karolherbst: uff
16:51 HdkR: Accept my hardware Karol :D
17:48 karolherbst: mwk: mind if we have a git repository with all official nvidia header files we are allowed to share? Like the UVM thing.
17:48 karolherbst: I'd would like to track all the versions for uvm support
17:48 karolherbst: need it in valgrind _and_ demmt sadly :/
17:50 karolherbst: the UVM ones are MIT licensed or something
17:51 karolherbst: so it would be really cool to just use those
17:51 karolherbst: ahhh, now I have my kernels :)
17:54 karolherbst: imirkin: ohhh, OpenCL global: "ldg e b32 $r2 ncg[$r2]" vs OpenCL constant: "ldg e ci b32 $r2 ncg[$r2]"
17:55 karolherbst: is that our CG?
17:55 karolherbst: although I think that ci is something we don't use
17:59 karolherbst: ci == cs.. oh well
18:21 mwk: karolherbst: uh, sure
18:21 mwk: no idea why you need my permission though
18:21 karolherbst: I was mainly wondering if you have a better idea
18:22 mwk: esp. since I'm done with nvidia stuff for the foreseeable future
18:22 karolherbst: ohh, I see
18:22 karolherbst: that bad though?
18:22 mwk: tbh I'd just stuff it n envytools
18:22 karolherbst: I need it in valgrind
18:22 karolherbst: https://github.com/karolherbst/valgrind/commit/e29d6ef2b3de297f20c7f52756f3de50ad9461ba
18:22 karolherbst: because you know, nvidia is dumb
18:23 karolherbst: and the uvm ioctl have no size information on their own
18:23 mwk: hehe right
18:23 karolherbst: no idea how I want to handle versioning in valgrind though
18:23 karolherbst: but I suspect those ioctl sizes _could_ change
18:26 mwk: could just duplicate the headers in two repos, not like they change a lot
18:27 karolherbst: the anoying part is just extracting those headers though
18:27 karolherbst: and maybe even "diff" them or something.. although I guess we could just download all the versions and extract everything
18:28 karolherbst: and always match 1:1
18:28 karolherbst: or maybe we can always use the latest version... I actually never checked that
21:02 karolherbst: imirkin: any luck with the depth32f_stencil8 test fix?
23:14 imirkin: karolherbst: didn't get to it. need to find a better test.
23:14 imirkin: s/find/write