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

1 files changed, 7 insertions, 3 deletions

diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp
index a47d0561e95..9739717df0f 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 */