RADV Will Now Enable "Sisched" For The Talos Principle, Boosting Frame Rates
The RADV Mesa Radeon Vulkan driver will now enable the sisched optimization automatically when running The Talos Principle in order to boost performance.
Valve driver developer Samuel Pitoiset pushed out the patches recently to use sisched by default for Talos Principle. The SISCHED feature for RADV and RadeonSI is about using the LLVM SI Machine Instruction Scheduler.
For RadeonSI and RADV, this feature can be flipped on via the R600_DEBUG=sisched for the OpenGL driver or RADV_PERFTEST=sisched for RADV environment variables while now within the driver code it's being enabled if the application's name matches that of The Talos Principle. This isn't being done for now in a DriConf/drirc-type configuration but is hard-coded within the driver.
Samuel commented on the patch as for doing the hard-coded approach, "keep it simple and improve later if it's really needed."
He has found using this instruction scheduler helps The Talos Principle frame-rates by about 4%. In the event you don't want this optimization, Samuel also ended up adding a new "nosisched" option too.
This work is queued in the current Mesa 17.4 code on Git master.
Valve driver developer Samuel Pitoiset pushed out the patches recently to use sisched by default for Talos Principle. The SISCHED feature for RADV and RadeonSI is about using the LLVM SI Machine Instruction Scheduler.
For RadeonSI and RADV, this feature can be flipped on via the R600_DEBUG=sisched for the OpenGL driver or RADV_PERFTEST=sisched for RADV environment variables while now within the driver code it's being enabled if the application's name matches that of The Talos Principle. This isn't being done for now in a DriConf/drirc-type configuration but is hard-coded within the driver.
Samuel commented on the patch as for doing the hard-coded approach, "keep it simple and improve later if it's really needed."
He has found using this instruction scheduler helps The Talos Principle frame-rates by about 4%. In the event you don't want this optimization, Samuel also ended up adding a new "nosisched" option too.
This work is queued in the current Mesa 17.4 code on Git master.
6 Comments