Cycles: Always use guarded allocator of vectors

We don't have vectors re-allocation happening multiple times from inside
a loop anymore, so we can safely switch to a memory guarded allocator for
vectors and keep track on the memory usage at various stages of rendering.

Additionally, when building from inside Blender repository, Cycles will
use Blender's guarded allocator, so actual memory usage will be displayed
in the Space Info header.

There are couple of tricky aspects of the patch:

- TaskScheduler::exit() now explicitly frees memory used by `threads`.
  This is needed because `threads` is a static member which destructor
  isn't getting called on Blender's exit which caused memory leak print
  to happen.

  This shouldn't give any measurable speed issues, reallocation of that
  vector is only one of fewzillion other allocations happening during
  synchronization.

- Use regular guarded malloc (not aligned one). No idea why it was
  made to be aligned in the first place. Perhaps some corner case tests
  or so. Vector was never expected to be aligned anyway. Let's see if
  we'll have actual bugs with this.

Reviewers: dingto, lukasstockner97, juicyfruit, brecht

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D1774

intern/cycles/CMakeLists.txt

@@ -210,6 +210,7 @@ endif()
 # Subdirectories
 
 if(WITH_CYCLES_BLENDER)
+	add_definitions(-DWITH_BLENDER_GUARDEDALLOC)
 	add_subdirectory(blender)
 endif()
 

intern/cycles/blender/addon/__init__.py

@@ -88,10 +88,17 @@ class CyclesRender(bpy.types.RenderEngine):
             self.report({'ERROR'}, "OSL support disabled in this build.")
 
 
+def engine_exit():
+    engine.exit()
+
+
 def register():
     from . import ui
     from . import properties
     from . import presets
+    import atexit
+
+    atexit.register(engine_exit)
 
     engine.init()
 
@@ -107,6 +114,7 @@ def unregister():
     from . import ui
     from . import properties
     from . import presets
+    import atexit
 
     bpy.app.handlers.version_update.remove(version_update.do_versions)
 
@@ -114,3 +122,6 @@ def unregister():
     properties.unregister()
     presets.unregister()
     bpy.utils.unregister_module(__name__)
+
+    atexit.unregister(engine_exit)
+    engine_exit()

intern/cycles/blender/addon/engine.py

@@ -74,6 +74,10 @@ def init():
     _cycles.init(path, user_path, bpy.app.background)
 
 
+def exit():
+    import _cycles
+    _cycles.exit()
+
 def create(engine, data, scene, region=None, v3d=None, rv3d=None, preview_osl=False):
     import bpy
     import _cycles

intern/cycles/blender/blender_python.cpp

@@ -39,6 +39,10 @@ CCL_NAMESPACE_BEGIN
 
 namespace {
 
+/* Device list stored static (used by compute_device_list()). */
+static ccl::vector<CCLDeviceInfo> device_list;
+static ccl::DeviceType device_type = DEVICE_NONE;
+
 /* Flag describing whether debug flags were synchronized from scene. */
 bool debug_flags_set = false;
 
@@ -172,6 +176,16 @@ static PyObject *init_func(PyObject * /*self*/, PyObject *args)
 	Py_RETURN_NONE;
 }
 
+
+static PyObject *exit_func(PyObject * /*self*/, PyObject * /*args*/)
+{
+	ShaderManager::free_memory();
+	TaskScheduler::free_memory();
+	Device::free_memory();
+	device_list.free_memory();
+	Py_RETURN_NONE;
+}
+
 static PyObject *create_func(PyObject * /*self*/, PyObject *args)
 {
 	PyObject *pyengine, *pyuserpref, *pydata, *pyscene, *pyregion, *pyv3d, *pyrv3d;
@@ -616,6 +630,7 @@ static PyObject *debug_flags_reset_func(PyObject * /*self*/, PyObject * /*args*/
 
 static PyMethodDef methods[] = {
 	{"init", init_func, METH_VARARGS, ""},
+	{"exit", exit_func, METH_VARARGS, ""},
 	{"create", create_func, METH_VARARGS, ""},
 	{"free", free_func, METH_O, ""},
 	{"render", render_func, METH_O, ""},
@@ -648,10 +663,6 @@ static struct PyModuleDef module = {
 
 static CCLDeviceInfo *compute_device_list(DeviceType type)
 {
-	/* device list stored static */
-	static ccl::vector<CCLDeviceInfo> device_list;
-	static ccl::DeviceType device_type = DEVICE_NONE;
-
 	/* create device list if it's not already done */
 	if(type != device_type) {
 		ccl::vector<DeviceInfo>& devices = ccl::Device::available_devices();

intern/cycles/device/device.cpp

@@ -34,6 +34,8 @@ CCL_NAMESPACE_BEGIN
 
 bool Device::need_types_update = true;
 bool Device::need_devices_update = true;
+vector<DeviceType> Device::types;
+vector<DeviceInfo> Device::devices;
 
 /* Device Requested Features */
 
@@ -280,8 +282,6 @@ string Device::string_from_type(DeviceType type)
 
 vector<DeviceType>& Device::available_types()
 {
-	static vector<DeviceType> types;
-
 	if(need_types_update) {
 		types.clear();
 		types.push_back(DEVICE_CPU);
@@ -311,8 +311,6 @@ vector<DeviceType>& Device::available_types()
 
 vector<DeviceInfo>& Device::available_devices()
 {
-	static vector<DeviceInfo> devices;
-
 	if(need_devices_update) {
 		devices.clear();
 #ifdef WITH_CUDA
@@ -368,4 +366,10 @@ void Device::tag_update()
 	need_devices_update = true;
 }
 
+void Device::free_memory()
+{
+	types.free_memory();
+	devices.free_memory();
+}
+
 CCL_NAMESPACE_END

intern/cycles/device/device.h

@@ -269,9 +269,12 @@ public:
 	/* Tag devices lists for update. */
 	static void tag_update();
 
+	static void free_memory();
 private:
 	/* Indicted whether device types and devices lists were initialized. */
 	static bool need_types_update, need_devices_update;
+	static vector<DeviceType> types;
+	static vector<DeviceInfo> devices;
 };
 
 CCL_NAMESPACE_END

intern/cycles/render/scene.cpp

@@ -34,13 +34,10 @@
 #include "tables.h"
 
 #include "util_foreach.h"
+#include "util_guarded_allocator.h"
+#include "util_logging.h"
 #include "util_progress.h"
 
-#ifdef WITH_CYCLES_DEBUG
-#  include "util_guarded_allocator.h"
-#  include "util_logging.h"
-#endif
-
 CCL_NAMESPACE_BEGIN
 
 Scene::Scene(const SceneParams& params_, const DeviceInfo& device_info_)
@@ -245,11 +242,9 @@ void Scene::device_update(Device *device_, Progress& progress)
 		device->const_copy_to("__data", &dscene.data, sizeof(dscene.data));
 	}
 
-#ifdef WITH_CYCLES_DEBUG
 	VLOG(1) << "System memory statistics after full device sync:\n"
 	        << "  Usage: " << util_guarded_get_mem_used() << "\n"
 	        << "  Peak: " << util_guarded_get_mem_peak();
-#endif
 }
 
 Scene::MotionType Scene::need_motion(bool advanced_shading)

intern/cycles/render/shader.cpp

@@ -531,5 +531,10 @@ void ShaderManager::get_requested_features(Scene *scene,
 	}
 }
 
+void ShaderManager::free_memory()
+{
+	beckmann_table.free_memory();
+}
+
 CCL_NAMESPACE_END
 

intern/cycles/render/shader.h

@@ -169,6 +169,8 @@ public:
 	void get_requested_features(Scene *scene,
 	                            DeviceRequestedFeatures *requested_features);
 
+	static void free_memory();
+
 protected:
 	ShaderManager();
 

intern/cycles/util/CMakeLists.txt

@@ -40,8 +40,10 @@ set(SRC_HEADERS
 	util_atomic.h
 	util_boundbox.h
 	util_debug.h
+	util_guarded_allocator.cpp
 	util_foreach.h
 	util_function.h
+	util_guarded_allocator.h
 	util_half.h
 	util_hash.h
 	util_image.h
@@ -77,15 +79,6 @@ set(SRC_HEADERS
 	util_xml.h
 )
 
-if(WITH_CYCLES_DEBUG)
-	list(APPEND SRC
-		util_guarded_allocator.cpp
-	)
-	list(APPEND SRC_HEADERS
-		util_guarded_allocator.h
-	)
-endif()
-
 include_directories(${INC})
 include_directories(SYSTEM ${INC_SYS})
 

intern/cycles/util/util_guarded_allocator.h

@@ -17,17 +17,10 @@
 #ifndef __UTIL_GUARDED_ALLOCATOR_H__
 #define __UTIL_GUARDED_ALLOCATOR_H__
 
-/* Define this in order to use Blender's guarded allocator to keep
- * track of allocated buffers, their sizes and peak memory usage.
- *
- * This is usually a bad level call, but it's really handy to keep
- * track of overall peak memory consumption during the scene
- * synchronization step.
- */
-#undef WITH_BLENDER_GUARDEDALLOC
-
+#include <cstddef>
 #include <memory>
 
+#include "util_debug.h"
 #include "util_types.h"
 
 #ifdef WITH_BLENDER_GUARDEDALLOC
@@ -42,39 +35,85 @@ void util_guarded_mem_free(size_t n);
 
 /* Guarded allocator for the use with STL. */
 template <typename T>
-class GuardedAllocator : public std::allocator<T> {
+class GuardedAllocator {
 public:
-	template<typename _Tp1>
-	struct rebind {
-		typedef GuardedAllocator<_Tp1> other;
-	};
+	typedef size_t size_type;
+	typedef ptrdiff_t difference_type;
+	typedef T *pointer;
+	typedef const T *const_pointer;
+	typedef T& reference;
+	typedef const T& const_reference;
+	typedef T value_type;
+
+	GuardedAllocator() {}
+	GuardedAllocator(const GuardedAllocator&) {}
 
 	T *allocate(size_t n, const void *hint = 0)
 	{
 		util_guarded_mem_alloc(n * sizeof(T));
-#ifdef WITH_BLENDER_GUARDEDALLOC
 		(void)hint;
-		return (T*)MEM_mallocN_aligned(n * sizeof(T), 16, "Cycles Alloc");
+#ifdef WITH_BLENDER_GUARDEDALLOC
+		if(n == 0) {
+			return NULL;
+		}
+		return (T*)MEM_mallocN(n * sizeof(T), "Cycles Alloc");
 #else
-		return std::allocator<T>::allocate(n, hint);
+		return (T*)malloc(n * sizeof(T));
 #endif
 	}
 
 	void deallocate(T *p, size_t n)
 	{
 		util_guarded_mem_free(n * sizeof(T));
+		if(p != NULL) {
 #ifdef WITH_BLENDER_GUARDEDALLOC
-		MEM_freeN((void*)p);
+			MEM_freeN(p);
 #else
-		std::allocator<T>::deallocate(p, n);
+			free(p);
 #endif
+		}
+	}
+
+	T *address(T& x) const
+	{
+		return &x;
+	}
+
+	const T *address(const T& x) const
+	{
+		return &x;
+	}
+
+	GuardedAllocator<T>& operator=(const GuardedAllocator&)
+	{
+		return *this;
+	}
+
+	void construct(T *p, const T& val)
+	{
+		new ((T *)p) T(val);
+	}
+
+	void destroy(T *p)
+	{
+		p->~T();
+	}
+
+	size_t max_size() const
+	{
+		return size_t(-1);
 	}
 
-	GuardedAllocator() : std::allocator<T>() {  }
-	GuardedAllocator(const GuardedAllocator &a) : std::allocator<T>(a) { }
 	template <class U>
-	GuardedAllocator(const GuardedAllocator<U> &a) : std::allocator<T>(a) { }
-	~GuardedAllocator() { }
+	struct rebind {
+		typedef GuardedAllocator<U> other;
+	};
+
+	template <class U>
+	GuardedAllocator(const GuardedAllocator<U>&) {}
+
+	template <class U>
+	GuardedAllocator& operator=(const GuardedAllocator<U>&) { return *this; }
 };
 
 /* Get memory usage and peak from the guarded STL allocator. */

intern/cycles/util/util_task.cpp

@@ -219,6 +219,12 @@ void TaskScheduler::exit()
 	}
 }
 
+void TaskScheduler::free_memory()
+{
+	assert(users == 0);
+	threads.free_memory();
+}
+
 bool TaskScheduler::thread_wait_pop(Entry& entry)
 {
 	thread_scoped_lock queue_lock(queue_mutex);

intern/cycles/util/util_task.h

@@ -91,6 +91,7 @@ class TaskScheduler
 public:
 	static void init(int num_threads = 0);
 	static void exit();
+	static void free_memory();
 
 	/* number of threads that can work on task */
 	static int num_threads() { return threads.size(); }

intern/cycles/util/util_vector.h

@@ -24,30 +24,21 @@
 #include <vector>
 
 #include "util_aligned_malloc.h"
+#include "util_guarded_allocator.h"
 #include "util_types.h"
 
-#ifdef WITH_CYCLES_DEBUG
-#  include "util_guarded_allocator.h"
-#endif
-
 CCL_NAMESPACE_BEGIN
 
 /* Vector
  *
  * Own subclass-ed vestion of std::vector. Subclass is needed because:
  *
- * - When building with WITH_CYCLES_DEBUG we need to use own allocator which
- *   keeps track of used/peak memory.
+ * - Use own allocator which keeps track of used/peak memory.
  *
  * - Have method to ensure capacity is re-set to 0.
  */
 template<typename value_type,
-#ifdef WITH_CYCLES_DEBUG
-         typename allocator_type = GuardedAllocator<value_type>
-#else
-         typename allocator_type = std::allocator<value_type>
-#endif
-        >
+         typename allocator_type = GuardedAllocator<value_type> >
 class vector : public std::vector<value_type, allocator_type>
 {
 public: