00:03 karolherbst: jenatali_: from a quick look though, what's the point of work_group_offset?
00:04 karolherbst: although I guess calculating at runtime would come with a drawback.. let me see
00:08 karolherbst: mhh.. I will play around with it I guess and see how all of that works with clover
04:43 jenatali_: Not sure how you could calculate it at runtime if you have to kick off two dispatches for a single kernel enqueue. Each dispatch would have its own 0-based indices, but you need to maintain a continuous set of work group IDs
05:50 imirkin: mareko: if there's a coherent vbo, do you kick off the batch immediately? i'm seeing some weirdness with the glVertex stuff you added to use coherent mappings, and it all gets "fixed" if i just kick off the draw immediately
06:35 imirkin: mareko: nevermind. figured out what's going on.
06:36 imirkin: not correctly handling glBufferData (equivalent) on persistent buffers. not sure that it can happen with normal GL usage, given that it has to go through glBufferStorage.
06:37 imirkin: however you end up calling Driver.BufferData() on a coherent buffer from vbo_exec_vtx_map
08:34 bbrezillon: robher: as pointed out by alyssa, looks like the "BOs should be unmapped before being marked purgeable" constraint was known => https://gitlab.freedesktop.org/mesa/mesa/-/commit/d0ec5d38f6cff2e61d9e526423c0a016cc830d14
08:35 bbrezillon: I still don't understand what drm_vma_node_unmap()
08:35 bbrezillon: is about if it's not unmapping things
16:02 mareko: imirkin: Driver.BufferData is called for glBufferStorage too
16:32 imirkin: mareko: yeah, but only once
16:32 imirkin: mareko: i was excluding persistent resources from the invalidate_resource() logic
16:32 imirkin: i removed that, and it fixed everything right up
16:33 imirkin: i.e. make it reallocate the data backing, which will change the mapping, which is why i was excluding it before