[Bf-blender-cvs] [18da79f] master: Cycles CUDA: only do async execution for GPUs not used for display.
Brecht Van Lommel
noreply at git.blender.org
Thu Apr 17 12:08:28 CEST 2014
Commit: 18da79f471f4aa15df895d253b03c9c3600411cf
Author: Brecht Van Lommel
Date: Thu Apr 17 11:59:05 2014 +0200
https://developer.blender.org/rB18da79f471f4aa15df895d253b03c9c3600411cf
Cycles CUDA: only do async execution for GPUs not used for display.
Otherwise devices used for display will lock up the UI too much. This means
you might still get 100% CPU for the display device, but for others CPU usage
should be low still.
The check to see if a device is used for display may not be entirely reliable,
it checks if there is a watchdog timeout on the device, but I'm not entirely
sure that always exists for display devices or is disabled for non-display
devices, though some tools like cuda-gdb seem to make the same assumption.
Ref T39559
===================================================================
M intern/cycles/device/device_cuda.cpp
===================================================================
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp
index a47d056..9739717 100644
--- a/intern/cycles/device/device_cuda.cpp
+++ b/intern/cycles/device/device_cuda.cpp
@@ -666,6 +666,11 @@ public:
cuda_assert(cuFuncSetBlockShape(cuPathTrace, xthreads, ythreads, 1))
cuda_assert(cuLaunchGridAsync(cuPathTrace, xblocks, yblocks, cuStream))
+ if(info.display_device) {
+ /* don't use async for device used for display, locks up UI too much */
+ cuda_assert(cuStreamSynchronize(cuStream))
+ }
+
cuda_pop_context();
}
@@ -995,7 +1000,6 @@ public:
bool branched = task->integrator_branched;
-
/* keep rendering tiles until done */
while(task->acquire_tile(this, tile)) {
int start_sample = tile.start_sample;
@@ -1016,9 +1020,9 @@ public:
tile.sample = sample + 1;
task->update_progress(tile);
- if(sample == sync_sample){
+ if(!info.display_device && sample == sync_sample) {
cuda_push_context();
- cuda_assert(cuEventRecord(tileDone, cuStream ))
+ cuda_assert(cuEventRecord(tileDone, cuStream))
cuda_assert(cuEventSynchronize(tileDone))
/* Do some time keeping to find out if we need to sync less */
More information about the Bf-blender-cvs
mailing list