Cycles: support animated object scale in motion blur.

This was disabled previously due to CUDA compiler bugs, see T32900.

Differential Revision: https://developer.blender.org/D2937
This commit is contained in:
Stefan Werner 2017-12-02 03:28:13 +01:00 committed by Brecht Van Lommel
parent ac2ebf9c54
commit 25b794a39d
8 changed files with 58 additions and 71 deletions

View File

@ -53,7 +53,7 @@ ccl_device_inline AttributeDescriptor attribute_not_found()
ccl_device_inline uint object_attribute_map_offset(KernelGlobals *kg, int object)
{
int offset = object*OBJECT_SIZE + 11;
int offset = object*OBJECT_SIZE + 15;
float4 f = kernel_tex_fetch(__objects, offset);
return __float_as_uint(f.y);
}

View File

@ -28,11 +28,12 @@ CCL_NAMESPACE_BEGIN
enum ObjectTransform {
OBJECT_TRANSFORM = 0,
OBJECT_TRANSFORM_MOTION_PRE = 0,
OBJECT_INVERSE_TRANSFORM = 4,
OBJECT_TRANSFORM_MOTION_POST = 4,
OBJECT_PROPERTIES = 8,
OBJECT_DUPLI = 9
OBJECT_TRANSFORM_MOTION_PRE = 0,
OBJECT_TRANSFORM_MOTION_MID = 4,
OBJECT_TRANSFORM_MOTION_POST = 8,
OBJECT_PROPERTIES = 12,
OBJECT_DUPLI = 13
};
enum ObjectVectorTransform {
@ -90,19 +91,24 @@ ccl_device_inline Transform object_fetch_vector_transform(KernelGlobals *kg, int
#ifdef __OBJECT_MOTION__
ccl_device_inline Transform object_fetch_transform_motion(KernelGlobals *kg, int object, float time)
{
DecompMotionTransform motion;
MotionTransform motion;
int offset = object*OBJECT_SIZE + (int)OBJECT_TRANSFORM_MOTION_PRE;
motion.mid.x = kernel_tex_fetch(__objects, offset + 0);
motion.mid.y = kernel_tex_fetch(__objects, offset + 1);
motion.mid.z = kernel_tex_fetch(__objects, offset + 2);
motion.mid.w = kernel_tex_fetch(__objects, offset + 3);
motion.pre.x = kernel_tex_fetch(__objects, offset + 0);
motion.pre.y = kernel_tex_fetch(__objects, offset + 1);
motion.pre.z = kernel_tex_fetch(__objects, offset + 2);
motion.pre.w = kernel_tex_fetch(__objects, offset + 3);
motion.pre_x = kernel_tex_fetch(__objects, offset + 4);
motion.pre_y = kernel_tex_fetch(__objects, offset + 5);
motion.post_x = kernel_tex_fetch(__objects, offset + 6);
motion.post_y = kernel_tex_fetch(__objects, offset + 7);
motion.mid.x = kernel_tex_fetch(__objects, offset + 4);
motion.mid.y = kernel_tex_fetch(__objects, offset + 5);
motion.mid.z = kernel_tex_fetch(__objects, offset + 6);
motion.mid.w = kernel_tex_fetch(__objects, offset + 7);
motion.post.x = kernel_tex_fetch(__objects, offset + 8);
motion.post.y = kernel_tex_fetch(__objects, offset + 9);
motion.post.z = kernel_tex_fetch(__objects, offset + 10);
motion.post.w = kernel_tex_fetch(__objects, offset + 11);
Transform tfm;
transform_motion_interpolate(&tfm, &motion, time);
@ -328,7 +334,7 @@ ccl_device_inline uint object_patch_map_offset(KernelGlobals *kg, int object)
if(object == OBJECT_NONE)
return 0;
int offset = object*OBJECT_SIZE + 11;
int offset = object*OBJECT_SIZE + 15;
float4 f = kernel_tex_fetch(__objects, offset);
return __float_as_uint(f.x);
}

View File

@ -95,11 +95,11 @@ ccl_device void camera_sample_perspective(KernelGlobals *kg, float raster_x, flo
# ifdef __KERNEL_OPENCL__
const MotionTransform tfm = kernel_data.cam.motion;
transform_motion_interpolate(&cameratoworld,
((const DecompMotionTransform*)&tfm),
&tfm,
ray->time);
# else
transform_motion_interpolate(&cameratoworld,
((const DecompMotionTransform*)&kernel_data.cam.motion),
&kernel_data.cam.motion,
ray->time);
# endif
}
@ -207,11 +207,11 @@ ccl_device void camera_sample_orthographic(KernelGlobals *kg, float raster_x, fl
# ifdef __KERNEL_OPENCL__
const MotionTransform tfm = kernel_data.cam.motion;
transform_motion_interpolate(&cameratoworld,
(const DecompMotionTransform*)&tfm,
&tfm,
ray->time);
# else
transform_motion_interpolate(&cameratoworld,
(const DecompMotionTransform*)&kernel_data.cam.motion,
&kernel_data.cam.motion,
ray->time);
# endif
}
@ -285,11 +285,11 @@ ccl_device_inline void camera_sample_panorama(KernelGlobals *kg,
# ifdef __KERNEL_OPENCL__
const MotionTransform tfm = kernel_data.cam.motion;
transform_motion_interpolate(&cameratoworld,
(const DecompMotionTransform*)&tfm,
&tfm,
ray->time);
# else
transform_motion_interpolate(&cameratoworld,
(const DecompMotionTransform*)&kernel_data.cam.motion,
&kernel_data.cam.motion,
ray->time);
# endif
}

View File

@ -35,7 +35,7 @@
CCL_NAMESPACE_BEGIN
/* Constants */
#define OBJECT_SIZE 12
#define OBJECT_SIZE 16
#define OBJECT_VECTOR_SIZE 6
#define LIGHT_SIZE 11
#define FILTER_TABLE_SIZE 1024

View File

@ -353,7 +353,7 @@ void Camera::device_update(Device *device, DeviceScene *dscene, Scene *scene)
#ifdef __CAMERA_MOTION__
else if(need_motion == Scene::MOTION_BLUR) {
if(use_motion) {
transform_motion_decompose((DecompMotionTransform*)&kcam->motion, &motion, &matrix);
transform_motion_decompose(&kcam->motion, &motion, &matrix);
kcam->have_motion = 1;
}
if(use_perspective_motion) {

View File

@ -97,7 +97,7 @@ void Object::compute_bounds(bool motion_blur)
mtfm.post = tfm;
}
DecompMotionTransform decomp;
MotionTransform decomp;
transform_motion_decompose(&decomp, &mtfm, &tfm);
bounds = BoundBox::empty;
@ -365,7 +365,7 @@ void ObjectManager::device_update_object_transform(UpdateObejctTransformState *s
/* OBJECT_INVERSE_TRANSFORM */
memcpy(&objects[offset+4], &itfm, sizeof(float4)*3);
/* OBJECT_PROPERTIES */
objects[offset+8] = make_float4(surface_area, pass_id, random_number, __int_as_float(particle_index));
objects[offset+12] = make_float4(surface_area, pass_id, random_number, __int_as_float(particle_index));
if(mesh->use_motion_blur) {
state->have_motion = true;
@ -402,10 +402,10 @@ void ObjectManager::device_update_object_transform(UpdateObejctTransformState *s
else if(state->need_motion == Scene::MOTION_BLUR) {
if(ob->use_motion) {
/* decompose transformations for interpolation. */
DecompMotionTransform decomp;
MotionTransform decomp;
transform_motion_decompose(&decomp, &ob->motion, &ob->tfm);
memcpy(&objects[offset], &decomp, sizeof(float4)*8);
memcpy(&objects[offset], &decomp, sizeof(float4)*12);
flag |= SD_OBJECT_MOTION;
state->have_motion = true;
}
@ -418,9 +418,9 @@ void ObjectManager::device_update_object_transform(UpdateObejctTransformState *s
int numverts = mesh->verts.size();
int numkeys = mesh->curve_keys.size();
objects[offset+9] = make_float4(ob->dupli_generated[0], ob->dupli_generated[1], ob->dupli_generated[2], __int_as_float(numkeys));
objects[offset+10] = make_float4(ob->dupli_uv[0], ob->dupli_uv[1], __int_as_float(numsteps), __int_as_float(numverts));
objects[offset+11] = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
objects[offset+13] = make_float4(ob->dupli_generated[0], ob->dupli_generated[1], ob->dupli_generated[2], __int_as_float(numkeys));
objects[offset+14] = make_float4(ob->dupli_uv[0], ob->dupli_uv[1], __int_as_float(numsteps), __int_as_float(numverts));
objects[offset+15] = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
/* Object flag. */
if(ob->use_holdout) {
@ -655,7 +655,7 @@ void ObjectManager::device_update_mesh_offsets(Device *, DeviceScene *dscene, Sc
foreach(Object *object, scene->objects) {
Mesh* mesh = object->mesh;
int offset = object_index*OBJECT_SIZE + 11;
int offset = object_index*OBJECT_SIZE + 15;
if(mesh->patch_table) {
uint patch_map_offset = 2*(mesh->patch_table_offset + mesh->patch_table->total_size() -

View File

@ -247,30 +247,18 @@ static void transform_decompose(Transform *decomp, const Transform *tfm)
decomp->w = make_float4(scale.y.z, scale.z.x, scale.z.y, scale.z.z);
}
void transform_motion_decompose(DecompMotionTransform *decomp, const MotionTransform *motion, const Transform *mid)
void transform_motion_decompose(MotionTransform *decomp, const MotionTransform *motion, const Transform *mid)
{
Transform pre, post;
transform_decompose(&pre, &motion->pre);
transform_decompose(&decomp->pre, &motion->pre);
transform_decompose(&decomp->mid, mid);
transform_decompose(&post, &motion->post);
transform_decompose(&decomp->post, &motion->post);
/* ensure rotation around shortest angle, negated quaternions are the same
* but this means we don't have to do the check in quat_interpolate */
if(dot(decomp->mid.x, post.x) < 0.0f)
if(dot(decomp->pre.x, decomp->mid.x) < 0.0f)
decomp->pre.x = -decomp->pre.x;
if(dot(decomp->mid.x, decomp->post.x) < 0.0f)
decomp->mid.x = -decomp->mid.x;
if(dot(pre.x, decomp->mid.x) < 0.0f)
pre.x = -pre.x;
/* drop scale of pre/post */
pre.y.w = decomp->mid.y.w;
post.y.w = decomp->mid.y.w;
/* store translation/rotation part of pre/post */
decomp->pre_x = pre.x;
decomp->pre_y = pre.y;
decomp->post_x = post.x;
decomp->post_y = post.y;
}
Transform transform_from_viewplane(BoundBox2D& viewplane)

View File

@ -39,9 +39,7 @@ typedef struct Transform {
/* transform decomposed in rotation/translation/scale. we use the same data
* structure as Transform, and tightly pack decomposition into it. first the
* rotation (4), then translation (3), then 3x3 scale matrix (9).
*
* For the DecompMotionTransform we drop scale from pre/post. */
* rotation (4), then translation (3), then 3x3 scale matrix (9). */
typedef struct ccl_may_alias MotionTransform {
Transform pre;
@ -49,12 +47,6 @@ typedef struct ccl_may_alias MotionTransform {
Transform post;
} MotionTransform;
typedef struct DecompMotionTransform {
Transform mid;
float4 pre_x, pre_y;
float4 post_x, post_y;
} DecompMotionTransform;
typedef struct PerspectiveMotionTransform {
Transform pre;
Transform post;
@ -466,7 +458,7 @@ ccl_device_inline void transform_compose(Transform *tfm, const Transform *decomp
/* Disabled for now, need arc-length parametrization for constant speed motion.
* #define CURVED_MOTION_INTERPOLATE */
ccl_device void transform_motion_interpolate(Transform *tfm, const DecompMotionTransform *motion, float t)
ccl_device void transform_motion_interpolate(Transform *tfm, const MotionTransform *motion, float t)
{
/* possible optimization: is it worth it adding a check to skip scaling?
* it's probably quite uncommon to have scaling objects. or can we skip
@ -475,9 +467,9 @@ ccl_device void transform_motion_interpolate(Transform *tfm, const DecompMotionT
#ifdef CURVED_MOTION_INTERPOLATE
/* 3 point bezier curve interpolation for position */
float3 Ppre = float4_to_float3(motion->pre_y);
float3 Ppre = float4_to_float3(motion->pre.y);
float3 Pmid = float4_to_float3(motion->mid.y);
float3 Ppost = float4_to_float3(motion->post_y);
float3 Ppost = float4_to_float3(motion->post.y);
float3 Pcontrol = 2.0f*Pmid - 0.5f*(Ppre + Ppost);
float3 P = Ppre*t*t + Pcontrol*2.0f*t*(1.0f - t) + Ppost*(1.0f - t)*(1.0f - t);
@ -491,27 +483,28 @@ ccl_device void transform_motion_interpolate(Transform *tfm, const DecompMotionT
if(t < 0.5f) {
t *= 2.0f;
decomp.x = quat_interpolate(motion->pre_x, motion->mid.x, t);
decomp.x = quat_interpolate(motion->pre.x, motion->mid.x, t);
#ifdef CURVED_MOTION_INTERPOLATE
decomp.y.w = (1.0f - t)*motion->pre_y.w + t*motion->mid.y.w;
decomp.y.w = (1.0f - t)*motion->pre.y.w + t*motion->mid.y.w;
#else
decomp.y = (1.0f - t)*motion->pre_y + t*motion->mid.y;
decomp.y = (1.0f - t)*motion->pre.y + t*motion->mid.y;
#endif
decomp.z = (1.0f - t)*motion->pre.z + t*motion->mid.z;
decomp.w = (1.0f - t)*motion->pre.w + t*motion->mid.w;
}
else {
t = (t - 0.5f)*2.0f;
decomp.x = quat_interpolate(motion->mid.x, motion->post_x, t);
decomp.x = quat_interpolate(motion->mid.x, motion->post.x, t);
#ifdef CURVED_MOTION_INTERPOLATE
decomp.y.w = (1.0f - t)*motion->mid.y.w + t*motion->post_y.w;
decomp.y.w = (1.0f - t)*motion->mid.y.w + t*motion->post.y.w;
#else
decomp.y = (1.0f - t)*motion->mid.y + t*motion->post_y;
decomp.y = (1.0f - t)*motion->mid.y + t*motion->post.y;
#endif
decomp.z = (1.0f - t)*motion->mid.z + t*motion->post.z;
decomp.w = (1.0f - t)*motion->mid.w + t*motion->post.w;
}
decomp.z = motion->mid.z;
decomp.w = motion->mid.w;
/* compose rotation, translation, scale into matrix */
transform_compose(tfm, &decomp);
}
@ -526,7 +519,7 @@ ccl_device_inline bool operator==(const MotionTransform& A, const MotionTransfor
}
float4 transform_to_quat(const Transform& tfm);
void transform_motion_decompose(DecompMotionTransform *decomp, const MotionTransform *motion, const Transform *mid);
void transform_motion_decompose(MotionTransform *decomp, const MotionTransform *motion, const Transform *mid);
Transform transform_from_viewplane(BoundBox2D& viewplane);
#endif