Cycles: Add sample-based runtime profiler that measures time spent in various parts of the CPU kernel

This commit adds a sample-based profiler that runs during CPU rendering and collects statistics on time spent in different parts of the kernel (ray intersection, shader evaluation etc.) as well as time spent per material and object.

The results are currently not exposed in the user interface or per Python yet, to see the stats on the console pass the "--cycles-print-stats" argument to Cycles (e.g. "./blender -- --cycles-print-stats").

Unfortunately, there is no clear way to extend this functionality to CUDA or OpenCL, so it is CPU-only for now.

Reviewers: brecht, sergey, swerner

Reviewed By: brecht, swerner

Differential Revision: https://developer.blender.org/D3892

5 files changed, 32 insertions, 1 deletions

diff --git a/intern/cycles/device/device.cpp b/intern/cycles/device/device.cpp
index 7e20bb449c3..54ffd4bc4df 100644
--- a/intern/cycles/device/device.cpp
+++ b/intern/cycles/device/device.cpp
@@ -362,6 +362,7 @@ DeviceInfo Device::get_multi_device(const vector<DeviceInfo>& subdevices, int th
 	info.has_half_images = true;
 	info.has_volume_decoupled = true;
 	info.has_osl = true;
+	info.has_profiling = true;
 
 	foreach(const DeviceInfo &device, subdevices) {
 		/* Ensure CPU device does not slow down GPU. */
@@ -396,6 +397,7 @@ DeviceInfo Device::get_multi_device(const vector<DeviceInfo>& subdevices, int th
 		info.has_half_images &= device.has_half_images;
 		info.has_volume_decoupled &= device.has_volume_decoupled;
 		info.has_osl &= device.has_osl;
+		info.has_profiling &= device.has_profiling;
 	}
 
 	return info;
diff --git a/intern/cycles/device/device.h b/intern/cycles/device/device.h
index f3fb338e638..071f61a7566 100644
--- a/intern/cycles/device/device.h
+++ b/intern/cycles/device/device.h
@@ -60,6 +60,7 @@ public:
 	bool has_volume_decoupled;      /* Decoupled volume shading. */
 	bool has_osl;                   /* Support Open Shading Language. */
 	bool use_split_kernel;          /* Use split or mega kernel. */
+	bool has_profiling;             /* Supports runtime collection of profiling info. */
 	int cpu_threads;
 	vector<DeviceInfo> multi_devices;
 
@@ -75,6 +76,7 @@ public:
 		has_volume_decoupled = false;
 		has_osl = false;
 		use_split_kernel = false;
+		has_profiling = false;
 	}
 
 	bool operator==(const DeviceInfo &info) {
diff --git a/intern/cycles/device/device_cpu.cpp b/intern/cycles/device/device_cpu.cpp
index 36fe9bfc92b..f0a6fd6e3f4 100644
--- a/intern/cycles/device/device_cpu.cpp
+++ b/intern/cycles/device/device_cpu.cpp
@@ -477,6 +477,8 @@ public:
 	bool denoising_non_local_means(device_ptr image_ptr, device_ptr guide_ptr, device_ptr variance_ptr, device_ptr out_ptr,
 	                               DenoisingTask *task)
 	{
+		ProfilingHelper profiling(task->profiler, PROFILING_DENOISING_NON_LOCAL_MEANS);
+
 		int4 rect = task->rect;
 		int   r   = task->nlm_state.r;
 		int   f   = task->nlm_state.f;
@@ -529,6 +531,8 @@ public:
 
 	bool denoising_construct_transform(DenoisingTask *task)
 	{
+		ProfilingHelper profiling(task->profiler, PROFILING_DENOISING_CONSTRUCT_TRANSFORM);
+
 		for(int y = 0; y < task->filter_area.w; y++) {
 			for(int x = 0; x < task->filter_area.z; x++) {
 				filter_construct_transform_kernel()((float*) task->buffer.mem.device_pointer,
@@ -551,6 +555,8 @@ public:
 	                           device_ptr output_ptr,
 	                           DenoisingTask *task)
 	{
+		ProfilingHelper profiling(task->profiler, PROFILING_DENOISING_RECONSTRUCT);
+
 		mem_zero(task->storage.XtWX);
 		mem_zero(task->storage.XtWY);
 
@@ -609,8 +615,10 @@ public:
 
 	bool denoising_combine_halves(device_ptr a_ptr, device_ptr b_ptr,
 	                              device_ptr mean_ptr, device_ptr variance_ptr,
-	                              int r, int4 rect, DenoisingTask * /*task*/)
+	                              int r, int4 rect, DenoisingTask *task)
 	{
+		ProfilingHelper profiling(task->profiler, PROFILING_DENOISING_COMBINE_HALVES);
+
 		for(int y = rect.y; y < rect.w; y++) {
 			for(int x = rect.x; x < rect.z; x++) {
 				filter_combine_halves_kernel()(x, y,
@@ -629,6 +637,8 @@ public:
 	                             device_ptr sample_variance_ptr, device_ptr sv_variance_ptr,
 	                             device_ptr buffer_variance_ptr, DenoisingTask *task)
 	{
+		ProfilingHelper profiling(task->profiler, PROFILING_DENOISING_DIVIDE_SHADOW);
+
 		for(int y = task->rect.y; y < task->rect.w; y++) {
 			for(int x = task->rect.x; x < task->rect.z; x++) {
 				filter_divide_shadow_kernel()(task->render_buffer.samples,
@@ -653,6 +663,8 @@ public:
 	                           device_ptr variance_ptr,
 	                           DenoisingTask *task)
 	{
+		ProfilingHelper profiling(task->profiler, PROFILING_DENOISING_GET_FEATURE);
+
 		for(int y = task->rect.y; y < task->rect.w; y++) {
 			for(int x = task->rect.x; x < task->rect.z; x++) {
 				filter_get_feature_kernel()(task->render_buffer.samples,
@@ -676,6 +688,8 @@ public:
 	                               device_ptr output_ptr,
 	                               DenoisingTask *task)
 	{
+		ProfilingHelper profiling(task->profiler, PROFILING_DENOISING_DETECT_OUTLIERS);
+
 		for(int y = task->rect.y; y < task->rect.w; y++) {
 			for(int x = task->rect.x; x < task->rect.z; x++) {
 				filter_detect_outliers_kernel()(x, y,
@@ -735,6 +749,8 @@ public:
 
 	void denoise(DenoisingTask& denoising, RenderTile &tile)
 	{
+		ProfilingHelper profiling(denoising.profiler, PROFILING_DENOISING);
+
 		tile.sample = tile.start_sample + tile.num_samples;
 
 		denoising.functions.construct_transform = function_bind(&CPUDevice::denoising_construct_transform, this, &denoising);
@@ -765,6 +781,8 @@ public:
 
 		KernelGlobals *kg = new ((void*) kgbuffer.device_pointer) KernelGlobals(thread_kernel_globals_init());
 
+		stats.profiler.add_state(&kg->profiler);
+
 		CPUSplitKernel *split_kernel = NULL;
 		if(use_split_kernel) {
 			split_kernel = new CPUSplitKernel(this);
@@ -778,6 +796,7 @@ public:
 
 		RenderTile tile;
 		DenoisingTask denoising(this, task);
+		denoising.profiler = &kg->profiler;
 
 		while(task.acquire_tile(this, tile)) {
 			if(tile.task == RenderTile::PATH_TRACE) {
@@ -802,6 +821,8 @@ public:
 			}
 		}
 
+		stats.profiler.remove_state(&kg->profiler);
+
 		thread_kernel_globals_free((KernelGlobals*)kgbuffer.device_pointer);
 		kg->~KernelGlobals();
 		kgbuffer.free();
@@ -1061,6 +1082,7 @@ void device_cpu_info(vector<DeviceInfo>& devices)
 	info.has_volume_decoupled = true;
 	info.has_osl = true;
 	info.has_half_images = true;
+	info.has_profiling = true;
 
 	devices.insert(devices.begin(), info);
 }
diff --git a/intern/cycles/device/device_denoising.cpp b/intern/cycles/device/device_denoising.cpp
index 78c65a3d22d..433cbd3c265 100644
--- a/intern/cycles/device/device_denoising.cpp
+++ b/intern/cycles/device/device_denoising.cpp
@@ -22,6 +22,7 @@ CCL_NAMESPACE_BEGIN
 
 DenoisingTask::DenoisingTask(Device *device, const DeviceTask &task)
 : tile_info_mem(device, "denoising tile info mem", MEM_READ_WRITE),
+  profiler(NULL),
   storage(device),
   buffer(device),
   device(device)
diff --git a/intern/cycles/device/device_denoising.h b/intern/cycles/device/device_denoising.h
index 8e0666d0e59..beae60c220f 100644
--- a/intern/cycles/device/device_denoising.h
+++ b/intern/cycles/device/device_denoising.h
@@ -23,6 +23,8 @@
 
 #include "kernel/filter/filter_defines.h"
 
+#include "util/util_profiling.h"
+
 CCL_NAMESPACE_BEGIN
 
 class DenoisingTask {
@@ -51,6 +53,8 @@ public:
 	TileInfo *tile_info;
 	device_vector<int> tile_info_mem;
 
+	ProfilingState *profiler;
+
 	int4 rect;
 	int4 filter_area;