Cycles: reduce memory usage of instanced objects by about 40%, as long as the

motion vector pass is not enabled.

6 files changed, 47 insertions, 20 deletions

diff --git a/intern/cycles/kernel/kernel_object.h b/intern/cycles/kernel/kernel_object.h
index 2b38544e527..40aa4753daa 100644
--- a/intern/cycles/kernel/kernel_object.h
+++ b/intern/cycles/kernel/kernel_object.h
@@ -20,11 +20,16 @@ CCL_NAMESPACE_BEGIN
 
 enum ObjectTransform {
 	OBJECT_TRANSFORM = 0,
-	OBJECT_INVERSE_TRANSFORM = 3,
-	OBJECT_PROPERTIES = 6,
-	OBJECT_TRANSFORM_MOTION_PRE = 8,
-	OBJECT_TRANSFORM_MOTION_POST = 12,
-	OBJECT_DUPLI = 16
+	OBJECT_TRANSFORM_MOTION_PRE = 0,
+	OBJECT_INVERSE_TRANSFORM = 4,
+	OBJECT_TRANSFORM_MOTION_POST = 4,
+	OBJECT_PROPERTIES = 8,
+	OBJECT_DUPLI = 9
+};
+
+enum ObjectVectorTransform {
+	OBJECT_VECTOR_MOTION_PRE = 0,
+	OBJECT_VECTOR_MOTION_POST = 3
 };
 
 __device_inline Transform object_fetch_transform(KernelGlobals *kg, int object, enum ObjectTransform type)
@@ -40,6 +45,19 @@ __device_inline Transform object_fetch_transform(KernelGlobals *kg, int object,
 	return tfm;
 }
 
+__device_inline Transform object_fetch_vector_transform(KernelGlobals *kg, int object, enum ObjectVectorTransform type)
+{
+	int offset = object*OBJECT_VECTOR_SIZE + (int)type;
+
+	Transform tfm;
+	tfm.x = kernel_tex_fetch(__objects_vector, offset + 0);
+	tfm.y = kernel_tex_fetch(__objects_vector, offset + 1);
+	tfm.z = kernel_tex_fetch(__objects_vector, offset + 2);
+	tfm.w = make_float4(0.0f, 0.0f, 0.0f, 1.0f);
+
+	return tfm;
+}
+
 #ifdef __OBJECT_MOTION__
 __device_inline Transform object_fetch_transform_motion(KernelGlobals *kg, int object, float time)
 {
diff --git a/intern/cycles/kernel/kernel_textures.h b/intern/cycles/kernel/kernel_textures.h
index 4855a948c6e..29f6b3f072c 100644
--- a/intern/cycles/kernel/kernel_textures.h
+++ b/intern/cycles/kernel/kernel_textures.h
@@ -34,6 +34,7 @@ KERNEL_TEX(uint, texture_uint, __object_node)
 
 /* objects */
 KERNEL_TEX(float4, texture_float4, __objects)
+KERNEL_TEX(float4, texture_float4, __objects_vector)
 
 /* triangles */
 KERNEL_TEX(float4, texture_float4, __tri_normal)
diff --git a/intern/cycles/kernel/kernel_triangle.h b/intern/cycles/kernel/kernel_triangle.h
index 0db447289c8..570ae52d6c2 100644
--- a/intern/cycles/kernel/kernel_triangle.h
+++ b/intern/cycles/kernel/kernel_triangle.h
@@ -209,10 +209,10 @@ __device float4 triangle_motion_vector(KernelGlobals *kg, ShaderData *sd)
 	 * transformation was set match the world/object space of motion_pre/post */
 	Transform tfm;
 	
-	tfm = object_fetch_transform(kg, sd->object, OBJECT_TRANSFORM_MOTION_PRE);
+	tfm = object_fetch_vector_transform(kg, sd->object, OBJECT_VECTOR_MOTION_PRE);
 	motion_pre = transform_point(&tfm, motion_pre);
 
-	tfm = object_fetch_transform(kg, sd->object, OBJECT_TRANSFORM_MOTION_POST);
+	tfm = object_fetch_vector_transform(kg, sd->object, OBJECT_VECTOR_MOTION_POST);
 	motion_post = transform_point(&tfm, motion_post);
 
 	float3 P;
diff --git a/intern/cycles/kernel/kernel_types.h b/intern/cycles/kernel/kernel_types.h
index a0673f55681..d11b96503d9 100644
--- a/intern/cycles/kernel/kernel_types.h
+++ b/intern/cycles/kernel/kernel_types.h
@@ -29,7 +29,8 @@
 CCL_NAMESPACE_BEGIN
 
 /* constants */
-#define OBJECT_SIZE 		18
+#define OBJECT_SIZE 		11
+#define OBJECT_VECTOR_SIZE	6
 #define LIGHT_SIZE			4
 #define FILTER_TABLE_SIZE	256
 #define RAMP_TABLE_SIZE		256
diff --git a/intern/cycles/render/object.cpp b/intern/cycles/render/object.cpp
index bd9f16d64ef..5df8e8c1368 100644
--- a/intern/cycles/render/object.cpp
+++ b/intern/cycles/render/object.cpp
@@ -150,12 +150,17 @@ ObjectManager::~ObjectManager()
 
 void ObjectManager::device_update_transforms(Device *device, DeviceScene *dscene, Scene *scene, uint *object_flag, Progress& progress)
 {
-	float4 *objects = dscene->objects.resize(OBJECT_SIZE*scene->objects.size());
+	float4 *objects;
+	float4 *objects_vector = NULL;
 	int i = 0;
 	map<Mesh*, float> surface_area_map;
 	Scene::MotionType need_motion = scene->need_motion(device->info.advanced_shading);
 	bool have_motion = false;
 
+	objects = dscene->objects.resize(OBJECT_SIZE*scene->objects.size());
+	if(need_motion == Scene::MOTION_PASS)
+		objects_vector = dscene->objects_vector.resize(OBJECT_VECTOR_SIZE*scene->objects.size());
+
 	foreach(Object *ob, scene->objects) {
 		Mesh *mesh = ob->mesh;
 		uint flag = 0;
@@ -205,8 +210,8 @@ void ObjectManager::device_update_transforms(Device *device, DeviceScene *dscene
 		int offset = i*OBJECT_SIZE;
 
 		memcpy(&objects[offset], &tfm, sizeof(float4)*3);
-		memcpy(&objects[offset+3], &itfm, sizeof(float4)*3);
-		objects[offset+6] = make_float4(surface_area, pass_id, random_number, __int_as_float(ob->particle_id));
+		memcpy(&objects[offset+4], &itfm, sizeof(float4)*3);
+		objects[offset+8] = make_float4(surface_area, pass_id, random_number, __int_as_float(ob->particle_id));
 
 		if(need_motion == Scene::MOTION_PASS) {
 			/* motion transformations, is world/object space depending if mesh
@@ -220,8 +225,8 @@ void ObjectManager::device_update_transforms(Device *device, DeviceScene *dscene
 			if(!mesh->attributes.find(ATTR_STD_MOTION_POST))
 				mtfm_post = mtfm_post * itfm;
 
-			memcpy(&objects[offset+8], &mtfm_pre, sizeof(float4)*4);
-			memcpy(&objects[offset+12], &mtfm_post, sizeof(float4)*4);
+			memcpy(&objects_vector[i*OBJECT_VECTOR_SIZE+0], &mtfm_pre, sizeof(float4)*3);
+			memcpy(&objects_vector[i*OBJECT_VECTOR_SIZE+3], &mtfm_post, sizeof(float4)*3);
 		}
 #ifdef __OBJECT_MOTION__
 		else if(need_motion == Scene::MOTION_BLUR) {
@@ -230,20 +235,16 @@ void ObjectManager::device_update_transforms(Device *device, DeviceScene *dscene
 				DecompMotionTransform decomp;
 
 				transform_motion_decompose(&decomp, &ob->motion, &ob->tfm);
-				memcpy(&objects[offset+8], &decomp, sizeof(float4)*8);
+				memcpy(&objects[offset], &decomp, sizeof(float4)*8);
 				flag |= SD_OBJECT_MOTION;
 				have_motion = true;
 			}
-			else {
-				float4 no_motion = make_float4(FLT_MAX);
-				memcpy(&objects[offset+8], &no_motion, sizeof(float4)*8);
-			}
 		}
 #endif
 
 		/* dupli object coords */
-		objects[offset+16] = make_float4(ob->dupli_generated[0], ob->dupli_generated[1], ob->dupli_generated[2], 0.0f);
-		objects[offset+17] = make_float4(ob->dupli_uv[0], ob->dupli_uv[1], 0.0f, 0.0f);
+		objects[offset+9] = make_float4(ob->dupli_generated[0], ob->dupli_generated[1], ob->dupli_generated[2], 0.0f);
+		objects[offset+10] = make_float4(ob->dupli_uv[0], ob->dupli_uv[1], 0.0f, 0.0f);
 
 		/* object flag */
 		if(ob->use_holdout)
@@ -256,6 +257,8 @@ void ObjectManager::device_update_transforms(Device *device, DeviceScene *dscene
 	}
 
 	device->tex_alloc("__objects", dscene->objects);
+	if(need_motion == Scene::MOTION_PASS)
+		device->tex_alloc("__objects_vector", dscene->objects_vector);
 
 	dscene->data.bvh.have_motion = have_motion;
 }
@@ -297,6 +300,9 @@ void ObjectManager::device_free(Device *device, DeviceScene *dscene)
 	device->tex_free(dscene->objects);
 	dscene->objects.clear();
 
+	device->tex_free(dscene->objects_vector);
+	dscene->objects_vector.clear();
+
 	device->tex_free(dscene->object_flag);
 	dscene->object_flag.clear();
 }
diff --git a/intern/cycles/render/scene.h b/intern/cycles/render/scene.h
index 92ef692b4b9..ebe932e40e7 100644
--- a/intern/cycles/render/scene.h
+++ b/intern/cycles/render/scene.h
@@ -74,6 +74,7 @@ public:
 
 	/* objects */
 	device_vector<float4> objects;
+	device_vector<float4> objects_vector;
 
 	/* attributes */
 	device_vector<uint4> attributes_map;