Cycles Denoising: Refactor denoiser tile handling

This deduplicates the calls for tile (un)mapping and allows to have a target buffer that is different from the source buffer (needed for baking and animation denoising).

intern/cycles/device/device_cpu.cpp

@@ -635,7 +635,7 @@ public:
 				                              (float*) buffer_variance_ptr,
 				                              &task->rect.x,
 				                              task->render_buffer.pass_stride,
-				                              task->render_buffer.denoising_data_offset);
+				                              task->render_buffer.offset);
 			}
 		}
 		return true;
@@ -658,7 +658,7 @@ public:
 				                            (float*) variance_ptr,
 				                            &task->rect.x,
 				                            task->render_buffer.pass_stride,
-				                            task->render_buffer.denoising_data_offset);
+				                            task->render_buffer.offset);
 			}
 		}
 		return true;
@@ -711,7 +711,7 @@ public:
 		}
 	}
 
-	void denoise(DeviceTask &task, DenoisingTask& denoising, RenderTile &tile)
+	void denoise(DenoisingTask& denoising, RenderTile &tile)
 	{
 		tile.sample = tile.start_sample + tile.num_samples;
 
@@ -727,16 +727,7 @@ public:
 		denoising.filter_area = make_int4(tile.x, tile.y, tile.w, tile.h);
 		denoising.render_buffer.samples = tile.sample;
 
-		RenderTile rtiles[9];
+		denoising.run_denoising(&tile);
-		rtiles[4] = tile;
-		task.map_neighbor_tiles(rtiles, this);
-		denoising.tiles_from_rendertiles(rtiles);
-
-		denoising.run_denoising();
-
-		task.unmap_neighbor_tiles(rtiles, this);
-
-		task.update_progress(&tile, tile.w*tile.h);
 	}
 
 	void thread_render(DeviceTask& task)
@@ -777,7 +768,9 @@ public:
 				}
 			}
 			else if(tile.task == RenderTile::DENOISE) {
-				denoise(task, denoising, tile);
+				denoise(denoising, tile);
+
+				task.update_progress(&tile, tile.w*tile.h);
 			}
 
 			task.release_tile(tile);

intern/cycles/device/device_cuda.cpp

@@ -1542,7 +1542,7 @@ public:
 		                &buffer_variance_ptr,
 		                &task->rect,
 		                &task->render_buffer.pass_stride,
-		                &task->render_buffer.denoising_data_offset};
+		                &task->render_buffer.offset};
 		CUDA_LAUNCH_KERNEL(cuFilterDivideShadow, args);
 		cuda_assert(cuCtxSynchronize());
 
@@ -1575,7 +1575,7 @@ public:
 		                &variance_ptr,
 		                &task->rect,
 		                &task->render_buffer.pass_stride,
-		                &task->render_buffer.denoising_data_offset};
+		                &task->render_buffer.offset};
 		CUDA_LAUNCH_KERNEL(cuFilterGetFeature, args);
 		cuda_assert(cuCtxSynchronize());
 
@@ -1613,7 +1613,7 @@ public:
 		return !have_error();
 	}
 
-	void denoise(RenderTile &rtile, DenoisingTask& denoising, const DeviceTask &task)
+	void denoise(RenderTile &rtile, DenoisingTask& denoising)
 	{
 		denoising.functions.construct_transform = function_bind(&CUDADevice::denoising_construct_transform, this, &denoising);
 		denoising.functions.reconstruct = function_bind(&CUDADevice::denoising_reconstruct, this, _1, _2, _3, &denoising);
@@ -1627,14 +1627,7 @@ public:
 		denoising.filter_area = make_int4(rtile.x, rtile.y, rtile.w, rtile.h);
 		denoising.render_buffer.samples = rtile.sample;
 
-		RenderTile rtiles[9];
+		denoising.run_denoising(&rtile);
-		rtiles[4] = rtile;
-		task.map_neighbor_tiles(rtiles, this);
-		denoising.tiles_from_rendertiles(rtiles);
-
-		denoising.run_denoising();
-
-		task.unmap_neighbor_tiles(rtiles, this);
 	}
 
 	void path_trace(DeviceTask& task, RenderTile& rtile, device_vector<WorkTile>& work_tiles)
@@ -2087,7 +2080,7 @@ public:
 				else if(tile.task == RenderTile::DENOISE) {
 					tile.sample = tile.start_sample + tile.num_samples;
 
-					denoise(tile, denoising, *task);
+					denoise(tile, denoising);
 
 					task->update_progress(&tile, tile.w*tile.h);
 				}

intern/cycles/device/device_denoising.cpp

@@ -36,8 +36,13 @@ DenoisingTask::DenoisingTask(Device *device, const DeviceTask &task)
 	}
 
 	render_buffer.pass_stride = task.pass_stride;
-	render_buffer.denoising_data_offset = task.pass_denoising_data;
+	render_buffer.offset = task.pass_denoising_data;
-	render_buffer.denoising_clean_offset = task.pass_denoising_clean;
+
+	target_buffer.pass_stride = task.pass_stride;
+	target_buffer.denoising_clean_offset = task.pass_denoising_clean;
+
+	functions.map_neighbor_tiles = function_bind(task.map_neighbor_tiles, _1, device);
+	functions.unmap_neighbor_tiles = function_bind(task.unmap_neighbor_tiles, _1, device);
 }
 
 DenoisingTask::~DenoisingTask()
@@ -53,8 +58,7 @@ DenoisingTask::~DenoisingTask()
 	tiles_mem.free();
 }
 
-
+void DenoisingTask::set_render_buffer(RenderTile *rtiles)
-void DenoisingTask::tiles_from_rendertiles(RenderTile *rtiles)
 {
 	tiles = (TilesInfo*) tiles_mem.alloc(sizeof(TilesInfo)/sizeof(int));
 
@@ -73,9 +77,9 @@ void DenoisingTask::tiles_from_rendertiles(RenderTile *rtiles)
 	tiles->y[2] = rtiles[7].y;
 	tiles->y[3] = rtiles[7].y + rtiles[7].h;
 
-	render_buffer.offset = rtiles[4].offset;
+	target_buffer.offset = rtiles[9].offset;
-	render_buffer.stride = rtiles[4].stride;
+	target_buffer.stride = rtiles[9].stride;
-	render_buffer.ptr    = rtiles[4].buffer;
+	target_buffer.ptr    = rtiles[9].buffer;
 
 	functions.set_tiles(buffers);
 }
@@ -228,21 +232,26 @@ void DenoisingTask::reconstruct()
 	storage.XtWY.alloc_to_device(storage.w*storage.h*XTWY_SIZE, false);
 
 	reconstruction_state.filter_window = rect_from_shape(filter_area.x-rect.x, filter_area.y-rect.y, storage.w, storage.h);
-	int tile_coordinate_offset = filter_area.y*render_buffer.stride + filter_area.x;
+	int tile_coordinate_offset = filter_area.y*target_buffer.stride + filter_area.x;
-	reconstruction_state.buffer_params = make_int4(render_buffer.offset + tile_coordinate_offset,
+	reconstruction_state.buffer_params = make_int4(target_buffer.offset + tile_coordinate_offset,
-	                                               render_buffer.stride,
+	                                               target_buffer.stride,
-	                                               render_buffer.pass_stride,
+	                                               target_buffer.pass_stride,
-	                                               render_buffer.denoising_clean_offset);
+	                                               target_buffer.denoising_clean_offset);
 	reconstruction_state.source_w = rect.z-rect.x;
 	reconstruction_state.source_h = rect.w-rect.y;
 
 	device_sub_ptr color_ptr    (buffer.mem,  8*buffer.pass_stride, 3*buffer.pass_stride);
 	device_sub_ptr color_var_ptr(buffer.mem, 11*buffer.pass_stride, 3*buffer.pass_stride);
-	functions.reconstruct(*color_ptr, *color_var_ptr, render_buffer.ptr);
+	functions.reconstruct(*color_ptr, *color_var_ptr, target_buffer.ptr);
 }
 
-void DenoisingTask::run_denoising()
+void DenoisingTask::run_denoising(RenderTile *tile)
 {
+	RenderTile rtiles[10];
+	rtiles[4] = *tile;
+	functions.map_neighbor_tiles(rtiles);
+	set_render_buffer(rtiles);
+
 	setup_denoising_buffer();
 
 	prefilter_shadowing();
@@ -252,6 +261,7 @@ void DenoisingTask::run_denoising()
 	construct_transform();
 	reconstruct();
 
+	functions.unmap_neighbor_tiles(rtiles);
 }
 
 CCL_NAMESPACE_END

intern/cycles/device/device_denoising.h

@@ -32,20 +32,24 @@ public:
 	float nlm_k_2;
 	float pca_threshold;
 
-	/* Pointer and parameters of the RenderBuffers. */
+	/* Parameters of the RenderBuffers. */
 	struct RenderBuffers {
-		int denoising_data_offset;
-		int denoising_clean_offset;
-		int pass_stride;
 		int offset;
-		int stride;
+		int pass_stride;
-		device_ptr ptr;
 		int samples;
 	} render_buffer;
 
+	/* Pointer and parameters of the target buffer. */
+	struct TargetBuffer {
+		int offset;
+		int stride;
+		int pass_stride;
+		int denoising_clean_offset;
+		device_ptr ptr;
+	} target_buffer;
+
 	TilesInfo *tiles;
 	device_vector<int> tiles_mem;
-	void tiles_from_rendertiles(RenderTile *rtiles);
 
 	int4 rect;
 	int4 filter_area;
@@ -86,6 +90,8 @@ public:
 		              device_ptr output_ptr
 		              )> detect_outliers;
 		function<bool(device_ptr*)> set_tiles;
+		function<void(RenderTile *rtiles)> map_neighbor_tiles;
+		function<void(RenderTile *rtiles)> unmap_neighbor_tiles;
 	} functions;
 
 	/* Stores state of the current Reconstruction operation,
@@ -141,7 +147,7 @@ public:
 	DenoisingTask(Device *device, const DeviceTask &task);
 	~DenoisingTask();
 
-	void run_denoising();
+	void run_denoising(RenderTile *tile);
 
 	struct DenoiseBuffers {
 		int pass_stride;

intern/cycles/device/opencl/opencl.h

@@ -362,7 +362,7 @@ public:
 	void film_convert(DeviceTask& task, device_ptr buffer, device_ptr rgba_byte, device_ptr rgba_half);
 	void shader(DeviceTask& task);
 
-	void denoise(RenderTile& tile, DenoisingTask& denoising, const DeviceTask& task);
+	void denoise(RenderTile& tile, DenoisingTask& denoising);
 
 	class OpenCLDeviceTask : public DeviceTask {
 	public:

intern/cycles/device/opencl/opencl_base.cpp

@@ -991,7 +991,7 @@ bool OpenCLDeviceBase::denoising_divide_shadow(device_ptr a_ptr,
 	                buffer_variance_mem,
 	                task->rect,
 	                task->render_buffer.pass_stride,
-	                task->render_buffer.denoising_data_offset);
+	                task->render_buffer.offset);
 	enqueue_kernel(ckFilterDivideShadow,
 	               task->rect.z-task->rect.x,
 	               task->rect.w-task->rect.y);
@@ -1021,7 +1021,7 @@ bool OpenCLDeviceBase::denoising_get_feature(int mean_offset,
 	                variance_mem,
 	                task->rect,
 	                task->render_buffer.pass_stride,
-	                task->render_buffer.denoising_data_offset);
+	                task->render_buffer.offset);
 	enqueue_kernel(ckFilterGetFeature,
 	               task->rect.z-task->rect.x,
 	               task->rect.w-task->rect.y);
@@ -1076,7 +1076,7 @@ bool OpenCLDeviceBase::denoising_set_tiles(device_ptr *buffers,
 	return true;
 }
 
-void OpenCLDeviceBase::denoise(RenderTile &rtile, DenoisingTask& denoising, const DeviceTask &task)
+void OpenCLDeviceBase::denoise(RenderTile &rtile, DenoisingTask& denoising)
 {
 	denoising.functions.set_tiles = function_bind(&OpenCLDeviceBase::denoising_set_tiles, this, _1, &denoising);
 	denoising.functions.construct_transform = function_bind(&OpenCLDeviceBase::denoising_construct_transform, this, &denoising);
@@ -1090,14 +1090,7 @@ void OpenCLDeviceBase::denoise(RenderTile &rtile, DenoisingTask& denoising, cons
 	denoising.filter_area = make_int4(rtile.x, rtile.y, rtile.w, rtile.h);
 	denoising.render_buffer.samples = rtile.sample;
 
-	RenderTile rtiles[9];
+	denoising.run_denoising(&rtile);
-	rtiles[4] = rtile;
-	task.map_neighbor_tiles(rtiles, this);
-	denoising.tiles_from_rendertiles(rtiles);
-
-	denoising.run_denoising();
-
-	task.unmap_neighbor_tiles(rtiles, this);
 }
 
 void OpenCLDeviceBase::shader(DeviceTask& task)

intern/cycles/device/opencl/opencl_mega.cpp

@@ -141,7 +141,7 @@ public:
 				}
 				else if(tile.task == RenderTile::DENOISE) {
 					tile.sample = tile.start_sample + tile.num_samples;
-					denoise(tile, denoising, *task);
+					denoise(tile, denoising);
 					task->update_progress(&tile, tile.w*tile.h);
 				}
 

intern/cycles/device/opencl/opencl_split.cpp

@@ -159,7 +159,7 @@ public:
 				}
 				else if(tile.task == RenderTile::DENOISE) {
 					tile.sample = tile.start_sample + tile.num_samples;
-					denoise(tile, denoising, *task);
+					denoise(tile, denoising);
 					task->update_progress(&tile, tile.w*tile.h);
 				}
 

intern/cycles/render/session.cpp

@@ -502,6 +502,9 @@ void Session::map_neighbor_tiles(RenderTile *tiles, Device *tile_device)
 
 	assert(tiles[4].buffers);
 	device->map_neighbor_tiles(tile_device, tiles);
+
+	/* The denoised result is written back to the original tile. */
+	tiles[9] = tiles[4];
 }
 
 void Session::unmap_neighbor_tiles(RenderTile *tiles, Device *tile_device)