Cycles: delay CUDA and OpenCL initialization to avoid driver crashes.

We've had many reported crashes on Windows where we suspect there is a
corrupted OpenCL driver. The purpose here is to keep Blender generally
usable in such cases.

Now it always shows None / CUDA / OpenCL in the preferences, and only when
selecting one will it reveal if there are any GPUs available. This should
avoid crashes when opening the preferences or on startup.

Differential Revision: https://developer.blender.org/D4265

6 files changed, 178 insertions, 115 deletions

diff --git a/intern/cycles/app/cycles_standalone.cpp b/intern/cycles/app/cycles_standalone.cpp
index ff2e0156b1a..9c899a38e7b 100644
--- a/intern/cycles/app/cycles_standalone.cpp
+++ b/intern/cycles/app/cycles_standalone.cpp
@@ -363,13 +363,8 @@ static void options_parse(int argc, const char **argv)
 	string devicename = "CPU";
 	bool list = false;
 
-	vector<DeviceType>& types = Device::available_types();
-
-	/* TODO(sergey): Here's a feedback loop happens: on the one hand we want
-	 * the device list to be printed in help message, on the other hand logging
-	 * is not initialized yet so we wouldn't have debug log happening in the
-	 * device initialization.
-	 */
+	/* List devices for which support is compiled in. */
+	vector<DeviceType> types = Device::available_types();
 	foreach(DeviceType type, types) {
 		if(device_names != "")
 			device_names += ", ";
@@ -421,7 +416,7 @@ static void options_parse(int argc, const char **argv)
 	}
 
 	if(list) {
-		vector<DeviceInfo>& devices = Device::available_devices();
+		vector<DeviceInfo> devices = Device::available_devices();
 		printf("Devices:\n");
 
 		foreach(DeviceInfo& info, devices) {
@@ -456,15 +451,12 @@ static void options_parse(int argc, const char **argv)
 
 	/* find matching device */
 	DeviceType device_type = Device::type_from_string(devicename.c_str());
-	vector<DeviceInfo>& devices = Device::available_devices();
-	bool device_available = false;
+	vector<DeviceInfo> devices = Device::available_devices(DEVICE_MASK(device_type));
 
-	foreach(DeviceInfo& device, devices) {
-		if(device_type == device.type) {
-			options.session_params.device = device;
-			device_available = true;
-			break;
-		}
+	bool device_available = false;
+	if (!devices.empty()) {
+		options.session_params.device = devices.front();
+		device_available = true;
 	}
 
 	/* handle invalid configurations */
diff --git a/intern/cycles/blender/addon/properties.py b/intern/cycles/blender/addon/properties.py
index 9d00a3f75b8..23ab1cf6a30 100644
--- a/intern/cycles/blender/addon/properties.py
+++ b/intern/cycles/blender/addon/properties.py
@@ -1478,7 +1478,7 @@ class CyclesPreferences(bpy.types.AddonPreferences):
     def get_devices(self):
         import _cycles
         # Layout of the device tuples: (Name, Type, Persistent ID)
-        device_list = _cycles.available_devices()
+        device_list = _cycles.available_devices(self.compute_device_type)
         # Make sure device entries are up to date and not referenced before
         # we know we don't add new devices. This way we guarantee to not
         # hold pointers to a resized array.
@@ -1502,7 +1502,7 @@ class CyclesPreferences(bpy.types.AddonPreferences):
 
     def get_num_gpu_devices(self):
         import _cycles
-        device_list = _cycles.available_devices()
+        device_list = _cycles.available_devices(self.compute_device_type)
         num = 0
         for device in device_list:
             if device[1] != self.compute_device_type:
@@ -1515,26 +1515,32 @@ class CyclesPreferences(bpy.types.AddonPreferences):
     def has_active_device(self):
         return self.get_num_gpu_devices() > 0
 
-    def draw_impl(self, layout, context):
-        available_device_types = self.get_device_types(context)
-        layout.label(text="Cycles Compute Device:")
-        if len(available_device_types) == 1:
-            layout.label(text="No compatible GPUs found", icon='INFO')
+    def _draw_devices(self, layout, device_type, devices):
+        box = layout.box()
+
+        found_device = False
+        for device in devices:
+            if device.type == device_type:
+                found_device = True
+                break
+
+        if not found_device:
+            box.label(text="No compatible GPUs found", icon='INFO')
             return
-        layout.row().prop(self, "compute_device_type", expand=True)
 
-        cuda_devices, opencl_devices = self.get_devices()
-        row = layout.row()
+        for device in devices:
+            box.prop(device, "use", text=device.name)
 
-        if self.compute_device_type == 'CUDA' and cuda_devices:
-            box = row.box()
-            for device in cuda_devices:
-                box.prop(device, "use", text=device.name)
+    def draw_impl(self, layout, context):
+        row = layout.row()
+        row.prop(self, "compute_device_type", expand=True)
 
-        if self.compute_device_type == 'OPENCL' and opencl_devices:
-            box = row.box()
-            for device in opencl_devices:
-                box.prop(device, "use", text=device.name)
+        cuda_devices, opencl_devices = self.get_devices()
+        row = layout.row()
+        if self.compute_device_type == 'CUDA':
+            self._draw_devices(row, 'CUDA', cuda_devices)
+        elif self.compute_device_type == 'OPENCL':
+            self._draw_devices(row, 'OPENCL', opencl_devices)
 
     def draw(self, context):
         self.draw_impl(self.layout, context)
diff --git a/intern/cycles/blender/blender_python.cpp b/intern/cycles/blender/blender_python.cpp
index d9ff95b2578..9dd0cd4c0bc 100644
--- a/intern/cycles/blender/blender_python.cpp
+++ b/intern/cycles/blender/blender_python.cpp
@@ -384,9 +384,18 @@ static PyObject *sync_func(PyObject * /*self*/, PyObject *value)
 	Py_RETURN_NONE;
 }
 
-static PyObject *available_devices_func(PyObject * /*self*/, PyObject * /*args*/)
+static PyObject *available_devices_func(PyObject * /*self*/, PyObject * args)
 {
-	vector<DeviceInfo>& devices = Device::available_devices();
+	const char *type_name;
+	if(!PyArg_ParseTuple(args, "s", &type_name)) {
+		return NULL;
+	}
+
+	DeviceType type = Device::type_from_string(type_name);
+	uint mask = (type == DEVICE_NONE) ? DEVICE_MASK_ALL : DEVICE_MASK(type);
+	mask |= DEVICE_MASK_CPU;
+
+	vector<DeviceInfo> devices = Device::available_devices(mask);
 	PyObject *ret = PyTuple_New(devices.size());
 
 	for(size_t i = 0; i < devices.size(); i++) {
@@ -742,11 +751,11 @@ static PyObject *enable_print_stats_func(PyObject * /*self*/, PyObject * /*args*
 
 static PyObject *get_device_types_func(PyObject * /*self*/, PyObject * /*args*/)
 {
-	vector<DeviceInfo>& devices = Device::available_devices();
+	vector<DeviceType> device_types = Device::available_types();
 	bool has_cuda = false, has_opencl = false;
-	for(int i = 0; i < devices.size(); i++) {
-		has_cuda   |= (devices[i].type == DEVICE_CUDA);
-		has_opencl |= (devices[i].type == DEVICE_OPENCL);
+	foreach(DeviceType device_type, device_types) {
+		has_cuda   |= (device_type == DEVICE_CUDA);
+		has_opencl |= (device_type == DEVICE_OPENCL);
 	}
 	PyObject *list = PyTuple_New(2);
 	PyTuple_SET_ITEM(list, 0, PyBool_FromLong(has_cuda));
@@ -768,7 +777,7 @@ static PyMethodDef methods[] = {
 	{"osl_update_node", osl_update_node_func, METH_VARARGS, ""},
 	{"osl_compile", osl_compile_func, METH_VARARGS, ""},
 #endif
-	{"available_devices", available_devices_func, METH_NOARGS, ""},
+	{"available_devices", available_devices_func, METH_VARARGS, ""},
 	{"system_info", system_info_func, METH_NOARGS, ""},
 #ifdef WITH_OPENCL
 	{"opencl_disable", opencl_disable_func, METH_NOARGS, ""},
diff --git a/intern/cycles/blender/blender_sync.cpp b/intern/cycles/blender/blender_sync.cpp
index 2b73c41fb19..703fcc2078b 100644
--- a/intern/cycles/blender/blender_sync.cpp
+++ b/intern/cycles/blender/blender_sync.cpp
@@ -766,24 +766,18 @@ SessionParams BlenderSync::get_session_params(BL::RenderEngine& b_engine,
 	/* Background */
 	params.background = background;
 
-	/* device type */
-	vector<DeviceInfo>& devices = Device::available_devices();
-
-	/* device default CPU */
-	foreach(DeviceInfo& device, devices) {
-		if(device.type == DEVICE_CPU) {
-			params.device = device;
-			break;
-		}
-	}
+	/* Default to CPU device. */
+	params.device = Device::available_devices(DEVICE_MASK_CPU).front();
 
 	if(get_enum(cscene, "device") == 2) {
-		/* find network device */
-		foreach(DeviceInfo& info, devices)
-			if(info.type == DEVICE_NETWORK)
-				params.device = info;
+		/* Find network device. */
+		vector<DeviceInfo> devices = Device::available_devices(DEVICE_MASK_NETWORK);
+		if(!devices.empty()) {
+			params.device = devices.front();
+		}
 	}
 	else if(get_enum(cscene, "device") == 1) {
+		/* Find cycles preferences. */
 		PointerRNA b_preferences;
 
 		BL::UserPreferences::addons_iterator b_addon_iter;
@@ -794,6 +788,7 @@ SessionParams BlenderSync::get_session_params(BL::RenderEngine& b_engine,
 			}
 		}
 
+		/* Test if we are using GPU devices. */
 		enum ComputeDevice {
 			COMPUTE_DEVICE_CPU = 0,
 			COMPUTE_DEVICE_CUDA = 1,
@@ -807,15 +802,20 @@ SessionParams BlenderSync::get_session_params(BL::RenderEngine& b_engine,
 		                                                       COMPUTE_DEVICE_CPU);
 
 		if(compute_device != COMPUTE_DEVICE_CPU) {
+			/* Query GPU devices with matching types. */
+			uint mask = DEVICE_MASK_CPU;
+			if(compute_device == COMPUTE_DEVICE_CUDA) {
+				mask |= DEVICE_MASK_CUDA;
+			}
+			else if(compute_device == COMPUTE_DEVICE_OPENCL) {
+				mask |= DEVICE_MASK_OPENCL;
+			}
+			vector<DeviceInfo> devices = Device::available_devices(mask);
+
+			/* Match device preferences and available devices. */
 			vector<DeviceInfo> used_devices;
 			RNA_BEGIN(&b_preferences, device, "devices") {
-				ComputeDevice device_type = (ComputeDevice)get_enum(device,
-				                                                    "type",
-				                                                    COMPUTE_DEVICE_NUM,
-				                                                    COMPUTE_DEVICE_CPU);
-
-				if(get_boolean(device, "use") &&
-				   (device_type == compute_device || device_type == COMPUTE_DEVICE_CPU)) {
+				if(get_boolean(device, "use")) {
 					string id = get_string(device, "id");
 					foreach(DeviceInfo& info, devices) {
 						if(info.id == id) {
@@ -826,10 +826,7 @@ SessionParams BlenderSync::get_session_params(BL::RenderEngine& b_engine,
 				}
 			} RNA_END;
 
-			if(used_devices.size() == 1) {
-				params.device = used_devices[0];
-			}
-			else if(used_devices.size() > 1) {
+			if(!used_devices.empty()) {
 				params.device = Device::get_multi_device(used_devices,
 				                                         params.threads,
 				                                         params.background);
diff --git a/intern/cycles/device/device.cpp b/intern/cycles/device/device.cpp
index c2d1512492c..317e62b2f69 100644
--- a/intern/cycles/device/device.cpp
+++ b/intern/cycles/device/device.cpp
@@ -36,8 +36,11 @@ CCL_NAMESPACE_BEGIN
 bool Device::need_types_update = true;
 bool Device::need_devices_update = true;
 thread_mutex Device::device_mutex;
-vector<DeviceType> Device::types;
-vector<DeviceInfo> Device::devices;
+vector<DeviceInfo> Device::opencl_devices;
+vector<DeviceInfo> Device::cuda_devices;
+vector<DeviceInfo> Device::cpu_devices;
+vector<DeviceInfo> Device::network_devices;
+uint Device::devices_initialized_mask = 0;
 
 /* Device Requested Features */
 
@@ -279,70 +282,108 @@ string Device::string_from_type(DeviceType type)
 	return "";
 }
 
-vector<DeviceType>& Device::available_types()
+vector<DeviceType> Device::available_types()
 {
-	thread_scoped_lock lock(device_mutex);
-	if(need_types_update) {
-		types.clear();
-		types.push_back(DEVICE_CPU);
+	vector<DeviceType> types;
+	types.push_back(DEVICE_CPU);
 #ifdef WITH_CUDA
-		if(device_cuda_init()) {
-			types.push_back(DEVICE_CUDA);
-		}
+	types.push_back(DEVICE_CUDA);
 #endif
 #ifdef WITH_OPENCL
-		if(device_opencl_init()) {
-			types.push_back(DEVICE_OPENCL);
-		}
+	types.push_back(DEVICE_OPENCL);
 #endif
 #ifdef WITH_NETWORK
-		types.push_back(DEVICE_NETWORK);
+	types.push_back(DEVICE_NETWORK);
 #endif
-		need_types_update = false;
-	}
 	return types;
 }
 
-vector<DeviceInfo>& Device::available_devices()
+vector<DeviceInfo> Device::available_devices(uint mask)
 {
+	/* Lazy initialize devices. On some platforms OpenCL or CUDA drivers can
+	 * be broken and cause crashes when only trying to get device info, so
+	 * we don't want to do any initialization until the user chooses to. */
 	thread_scoped_lock lock(device_mutex);
-	if(need_devices_update) {
-		devices.clear();
+	vector<DeviceInfo> devices;
+
 #ifdef WITH_OPENCL
-		if(device_opencl_init()) {
-			device_opencl_info(devices);
+	if(mask & DEVICE_MASK_OPENCL) {
+		if(!(devices_initialized_mask & DEVICE_MASK_OPENCL)) {
+			if(device_opencl_init()) {
+				device_opencl_info(opencl_devices);
+			}
+			devices_initialized_mask |= DEVICE_MASK_OPENCL;
 		}
+		foreach(DeviceInfo& info, opencl_devices) {
+			devices.push_back(info);
+		}
+	}
 #endif
+
 #ifdef WITH_CUDA
-		if(device_cuda_init()) {
-			device_cuda_info(devices);
+	if(mask & DEVICE_MASK_CUDA) {
+		if(!(devices_initialized_mask & DEVICE_MASK_CUDA)) {
+			if(device_cuda_init()) {
+				device_cuda_info(cuda_devices);
+			}
+			devices_initialized_mask |= DEVICE_MASK_CUDA;
 		}
+		foreach(DeviceInfo& info, cuda_devices) {
+			devices.push_back(info);
+		}
+	}
 #endif
-		device_cpu_info(devices);
+
+	if(mask & DEVICE_MASK_CPU) {
+		if(!(devices_initialized_mask & DEVICE_MASK_CPU)) {
+			device_cpu_info(cpu_devices);
+			devices_initialized_mask |= DEVICE_MASK_CPU;
+		}
+		foreach(DeviceInfo& info, cpu_devices) {
+			devices.push_back(info);
+		}
+	}
+
 #ifdef WITH_NETWORK
-		device_network_info(devices);
-#endif
-		need_devices_update = false;
+	if(mask & DEVICE_MASK_NETWORK) {
+		if(!(devices_initialized_mask & DEVICE_MASK_NETWORK)) {
+			device_network_info(network_devices);
+			devices_initialized_mask |= DEVICE_MASK_NETWORK;
+		}
+		foreach(DeviceInfo& info, network_devices) {
+			devices.push_back(info);
+		}
 	}
+#endif
+
 	return devices;
 }
 
-string Device::device_capabilities()
+string Device::device_capabilities(uint mask)
 {
-	string capabilities = "CPU device capabilities: ";
-	capabilities += device_cpu_capabilities() + "\n";
+	thread_scoped_lock lock(device_mutex);
+	string capabilities = "";
+
+	if(mask & DEVICE_MASK_CPU) {
+		capabilities += "\nCPU device capabilities: ";
+		capabilities += device_cpu_capabilities() + "\n";
+	}
 
 #ifdef WITH_OPENCL
-	if(device_opencl_init()) {
-		capabilities += "\nOpenCL device capabilities:\n";
-		capabilities += device_opencl_capabilities();
+	if(mask & DEVICE_MASK_OPENCL) {
+		if(device_opencl_init()) {
+			capabilities += "\nOpenCL device capabilities:\n";
+			capabilities += device_opencl_capabilities();
+		}
 	}
 #endif
 
 #ifdef WITH_CUDA
-	if(device_cuda_init()) {
-		capabilities += "\nCUDA device capabilities:\n";
-		capabilities += device_cuda_capabilities();
+	if(mask & DEVICE_MASK_CUDA) {
+		if(device_cuda_init()) {
+			capabilities += "\nCUDA device capabilities:\n";
+			capabilities += device_cuda_capabilities();
+		}
 	}
 #endif
 
@@ -351,7 +392,12 @@ string Device::device_capabilities()
 
 DeviceInfo Device::get_multi_device(const vector<DeviceInfo>& subdevices, int threads, bool background)
 {
-	assert(subdevices.size() > 1);
+	assert(subdevices.size() > 0);
+
+	if(subdevices.size() == 1) {
+		/* No multi device needed. */
+		return subdevices.front();
+	}
 
 	DeviceInfo info;
 	info.type = DEVICE_MULTI;
@@ -405,16 +451,16 @@ DeviceInfo Device::get_multi_device(const vector<DeviceInfo>& subdevices, int th
 
 void Device::tag_update()
 {
-	need_types_update = true;
-	need_devices_update = true;
+	free_memory();
 }
 
 void Device::free_memory()
 {
-	need_types_update = true;
-	need_devices_update = true;
-	types.free_memory();
-	devices.free_memory();
+	devices_initialized_mask = 0;
+	cuda_devices.clear();
+	opencl_devices.clear();
+	cpu_devices.clear();
+	network_devices.clear();
 }
 
 CCL_NAMESPACE_END
diff --git a/intern/cycles/device/device.h b/intern/cycles/device/device.h
index 55c39188210..082f9f758a8 100644
--- a/intern/cycles/device/device.h
+++ b/intern/cycles/device/device.h
@@ -40,7 +40,7 @@ class RenderTile;
 /* Device Types */
 
 enum DeviceType {
-	DEVICE_NONE,
+	DEVICE_NONE = 0,
 	DEVICE_CPU,
 	DEVICE_OPENCL,
 	DEVICE_CUDA,
@@ -48,6 +48,16 @@ enum DeviceType {
 	DEVICE_MULTI
 };
 
+enum DeviceTypeMask {
+	DEVICE_MASK_CPU = (1 << DEVICE_CPU),
+	DEVICE_MASK_OPENCL = (1 << DEVICE_OPENCL),
+	DEVICE_MASK_CUDA = (1 << DEVICE_CUDA),
+	DEVICE_MASK_NETWORK = (1 << DEVICE_NETWORK),
+	DEVICE_MASK_ALL = ~0
+};
+
+#define DEVICE_MASK(type) (DeviceTypeMask)(1 << type)
+
 class DeviceInfo {
 public:
 	DeviceType type;
@@ -328,9 +338,9 @@ public:
 
 	static DeviceType type_from_string(const char *name);
 	static string string_from_type(DeviceType type);
-	static vector<DeviceType>& available_types();
-	static vector<DeviceInfo>& available_devices();
-	static string device_capabilities();
+	static vector<DeviceType> available_types();
+	static vector<DeviceInfo> available_devices(uint device_type_mask = DEVICE_MASK_ALL);
+	static string device_capabilities(uint device_type_mask = DEVICE_MASK_ALL);
 	static DeviceInfo get_multi_device(const vector<DeviceInfo>& subdevices,
 	                                   int threads,
 	                                   bool background);
@@ -357,8 +367,11 @@ private:
 	/* Indicted whether device types and devices lists were initialized. */
 	static bool need_types_update, need_devices_update;
 	static thread_mutex device_mutex;
-	static vector<DeviceType> types;
-	static vector<DeviceInfo> devices;
+	static vector<DeviceInfo> cuda_devices;
+	static vector<DeviceInfo> opencl_devices;
+	static vector<DeviceInfo> cpu_devices;
+	static vector<DeviceInfo> network_devices;
+	static uint devices_initialized_mask;
 };
 
 CCL_NAMESPACE_END