Cycles: revert async CUDA changes, these are giving too much trouble still.

Fixes T40027. This means we get more CPU usage again when using multiple CUDA,
but the impact on performance is too big a problem with the current code.

1 files changed, 3 insertions, 43 deletions

diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp
index 68955211146..b19f5e22769 100644
--- a/intern/cycles/device/device_cuda.cpp
+++ b/intern/cycles/device/device_cuda.cpp
@@ -41,14 +41,11 @@ public:
 	CUdevice cuDevice;
 	CUcontext cuContext;
 	CUmodule cuModule;
-	CUstream cuStream;
-	CUevent tileDone;
 	map<device_ptr, bool> tex_interp_map;
 	int cuDevId;
 	int cuDevArchitecture;
 	bool first_error;
 	bool use_texture_storage;
-	unsigned int target_update_frequency;
 
 	struct PixelMem {
 		GLuint cuPBO;
@@ -180,8 +177,6 @@ public:
 		first_error = true;
 		background = background_;
 		use_texture_storage = true;
-		/* we try an update / sync every 1000 ms */
-		target_update_frequency = 1000;
 
 		cuDevId = info.num;
 		cuDevice = 0;
@@ -212,9 +207,6 @@ public:
 		if(cuda_error_(result, "cuCtxCreate"))
 			return;
 
-		cuda_assert(cuStreamCreate(&cuStream, 0));
-		cuda_assert(cuEventCreate(&tileDone, 0x1));
-
 		int major, minor;
 		cuDeviceComputeCapability(&major, &minor, cuDevId);
 		cuDevArchitecture = major*100 + minor*10;
@@ -231,8 +223,6 @@ public:
 	{
 		task_pool.stop();
 
-		cuda_assert(cuEventDestroy(tileDone));
-		cuda_assert(cuStreamDestroy(cuStream));
 		cuda_assert(cuCtxDestroy(cuContext));
 	}
 
@@ -665,15 +655,9 @@ public:
 
 		cuda_assert(cuFuncSetCacheConfig(cuPathTrace, CU_FUNC_CACHE_PREFER_L1));
 		cuda_assert(cuFuncSetBlockShape(cuPathTrace, xthreads, ythreads, 1));
+		cuda_assert(cuLaunchGrid(cuPathTrace, xblocks, yblocks));
 
-		if(info.display_device) {
-			/* don't use async for device used for display, locks up UI too much */
-			cuda_assert(cuLaunchGrid(cuPathTrace, xblocks, yblocks));
-			cuda_assert(cuCtxSynchronize());
-		}
-		else {
-			cuda_assert(cuLaunchGridAsync(cuPathTrace, xblocks, yblocks, cuStream));
-		}
+		cuda_assert(cuCtxSynchronize());
 
 		cuda_pop_context();
 	}
@@ -1024,10 +1008,6 @@ public:
 				int start_sample = tile.start_sample;
 				int end_sample = tile.start_sample + tile.num_samples;
 
-				boost::posix_time::ptime start_time(boost::posix_time::microsec_clock::local_time());
-				boost::posix_time::ptime last_time = start_time;
-				int sync_sample = 10;
-
 				for(int sample = start_sample; sample < end_sample; sample++) {
 					if (task->get_cancel()) {
 						if(task->need_finish_queue == false)
@@ -1037,28 +1017,8 @@ public:
 					path_trace(tile, sample, branched);
 
 					tile.sample = sample + 1;
-					task->update_progress(tile);
 
-					if(!info.display_device && sample == sync_sample) {
-						cuda_push_context();
-						cuda_assert(cuEventRecord(tileDone, cuStream));
-						cuda_assert(cuEventSynchronize(tileDone));
-
-						/* Do some time keeping to find out if we need to sync less */
-						boost::posix_time::ptime current_time(boost::posix_time::microsec_clock::local_time());
-						boost::posix_time::time_duration sample_duration = current_time - last_time;
-
-						long msec = sample_duration.total_milliseconds();
-						float scaling_factor = (float)target_update_frequency / (float)msec;
-
-						/* sync at earliest next sample and probably later */
-						sync_sample = (sample + 1) + sync_sample * (int)ceil(scaling_factor);
-
-						sync_sample = min(end_sample - 1, sync_sample); // make sure we sync the last sample always
-
-						last_time = current_time;
-						cuda_pop_context();
-					}
+					task->update_progress(tile);
 				}
 
 				task->release_tile(tile);