14:05 MrCooper: DemiMarie: DRI drivers are user-space, did you mean DRM drivers?
14:06 DemiMarie: MrCooper: yes
14:26 DemiMarie: Assuming the platform has the needed infrastructure, such as allowing the driver to map the needed PCI BARs and be notified of interrupts.
14:30 pac85[m]: Doesn't a drm driver need to know about processes memory maps so it can map things to the GPU? Isn't exposing this kind of stuff to user space a bit "too much"? (As in, you'd get better security by leaving it in the kernel)
17:27 penguin42:is getting clGetEventProfilingInfo failed: PROFILING_INFO_NOT_AVAILABLE on Mesa with my Radeon, clinfo says 'Queue Properties...Profiling Yes' but also says 'Profiling timer resolution 0ns' any ideas?
17:28 penguin42: that's 'OpenCL 1.1 Mesa 23.0.3' on Fedora 38 with Radeon RX 550 (Polaris12)
18:05 TjNiko: hello
18:41 DemiMarie: <pac85[m]> "Doesn't a drm driver need to..." <- My understanding is that in a capability-based microkernel system, the client process would pass the driver a capability that gives the driver the necessary access to the relevant portions of the client’s address space. The driver would use this capability and its own capabilities to map this memory into the GPU’s I/O address space via the IOMMU.
18:44 DemiMarie: MrCooper pac85: Assume that the driver has access to APIs that allow it to do its job, but which are generally going to be different than the Linux kernel’s own APIs. This means that either the driver would need to be ported to use the new APIs or a compatibility layer would need to be written.
20:02 pac85[m]: <DemiMarie> "MrCooper pac85: Assume that..." <- Sorry but I'm really no expert when it comes to those things, hence the question marks. I can only assume that this would be significantly different than things are currently and therefore really hard. Also I think in some platforms the GPU shares the same va space of the CPU anyway so this is just not possible? I suppose those are things somebody else here can answer. GPU kernel drivers really
20:02 pac85[m]: seem like an extention of the kernel itself, moving them to user space sounds a lot like moving core parts of the kernel to user space.
21:08 penguin42: hmm rocm I can get profiling working