Cycles: OpenCL kernel split

This commit contains all the work related on the AMD megakernel split work
which was mainly done by Varun Sundar, George Kyriazis and Lenny Wang, plus
some help from Sergey Sharybin, Martijn Berger, Thomas Dinges and likely
someone else which we're forgetting to mention.

Currently only AMD cards are enabled for the new split kernel, but it is
possible to force split opencl kernel to be used by setting the following
environment variable: CYCLES_OPENCL_SPLIT_KERNEL_TEST=1.

Not all the features are supported yet, and that being said no motion blur,
camera blur, SSS and volumetrics for now. Also transparent shadows are
disabled on AMD device because of some compiler bug.

This kernel is also only implements regular path tracing and supporting
branched one will take a bit. Branched path tracing is exposed to the
interface still, which is a bit misleading and will be hidden there soon.

More feature will be enabled once they're ported to the split kernel and
tested.

Neither regular CPU nor CUDA has any difference, they're generating the
same exact code, which means no regressions/improvements there.

Based on the research paper:

  https://research.nvidia.com/sites/default/files/publications/laine2013hpg_paper.pdf

Here's the documentation:

  https://docs.google.com/document/d/1LuXW-CV-sVJkQaEGZlMJ86jZ8FmoPfecaMdR-oiWbUY/edit

Design discussion of the patch:

  https://developer.blender.org/T44197

Differential Revision: https://developer.blender.org/D1200

6 files changed, 114 insertions, 79 deletions

diff --git a/intern/cycles/kernel/geom/geom_attribute.h b/intern/cycles/kernel/geom/geom_attribute.h
index 9ac16e86085..c7364e9edac 100644
--- a/intern/cycles/kernel/geom/geom_attribute.h
+++ b/intern/cycles/kernel/geom/geom_attribute.h
@@ -29,13 +29,13 @@ CCL_NAMESPACE_BEGIN
 
 ccl_device_inline int find_attribute(KernelGlobals *kg, const ShaderData *sd, uint id, AttributeElement *elem)
 {
-	if(sd->object == PRIM_NONE)
+	if(ccl_fetch(sd, object) == PRIM_NONE)
 		return (int)ATTR_STD_NOT_FOUND;
 
 	/* for SVM, find attribute by unique id */
-	uint attr_offset = sd->object*kernel_data.bvh.attributes_map_stride;
+	uint attr_offset = ccl_fetch(sd, object)*kernel_data.bvh.attributes_map_stride;
 #ifdef __HAIR__
-	attr_offset = (sd->type & PRIMITIVE_ALL_CURVE)? attr_offset + ATTR_PRIM_CURVE: attr_offset;
+	attr_offset = (ccl_fetch(sd, type) & PRIMITIVE_ALL_CURVE)? attr_offset + ATTR_PRIM_CURVE: attr_offset;
 #endif
 	uint4 attr_map = kernel_tex_fetch(__attributes_map, attr_offset);
 	
@@ -49,7 +49,7 @@ ccl_device_inline int find_attribute(KernelGlobals *kg, const ShaderData *sd, ui
 
 	*elem = (AttributeElement)attr_map.y;
 	
-	if(sd->prim == PRIM_NONE && (AttributeElement)attr_map.y != ATTR_ELEMENT_MESH)
+	if(ccl_fetch(sd, prim) == PRIM_NONE && (AttributeElement)attr_map.y != ATTR_ELEMENT_MESH)
 		return ATTR_STD_NOT_FOUND;
 
 	/* return result */
diff --git a/intern/cycles/kernel/geom/geom_bvh.h b/intern/cycles/kernel/geom/geom_bvh.h
index 2e8e27c709d..3d0d406dd0b 100644
--- a/intern/cycles/kernel/geom/geom_bvh.h
+++ b/intern/cycles/kernel/geom/geom_bvh.h
@@ -447,6 +447,7 @@ ccl_device_inline float3 ray_offset(float3 P, float3 Ng)
 #endif
 }
 
+#if defined(__SHADOW_RECORD_ALL__) || defined (__VOLUME_RECORD_ALL__)
 /* ToDo: Move to another file? */
 ccl_device int intersections_compare(const void *a, const void *b)
 {
@@ -460,6 +461,7 @@ ccl_device int intersections_compare(const void *a, const void *b)
 	else
 		return 0;
 }
+#endif
 
 CCL_NAMESPACE_END
 
diff --git a/intern/cycles/kernel/geom/geom_motion_triangle.h b/intern/cycles/kernel/geom/geom_motion_triangle.h
index a5a25f4a9ae..4ea9e4714c4 100644
--- a/intern/cycles/kernel/geom/geom_motion_triangle.h
+++ b/intern/cycles/kernel/geom/geom_motion_triangle.h
@@ -236,25 +236,25 @@ ccl_device_inline float3 motion_triangle_refine_subsurface(KernelGlobals *kg, Sh
 ccl_device_noinline void motion_triangle_shader_setup(KernelGlobals *kg, ShaderData *sd, const Intersection *isect, const Ray *ray, bool subsurface)
 {
 	/* get shader */
-	sd->shader = kernel_tex_fetch(__tri_shader, sd->prim);
+	ccl_fetch(sd, shader) = kernel_tex_fetch(__tri_shader, ccl_fetch(sd, prim));
 
 	/* get motion info */
 	int numsteps, numverts;
-	object_motion_info(kg, sd->object, &numsteps, &numverts, NULL);
+	object_motion_info(kg, ccl_fetch(sd, object), &numsteps, &numverts, NULL);
 
 	/* figure out which steps we need to fetch and their interpolation factor */
 	int maxstep = numsteps*2;
-	int step = min((int)(sd->time*maxstep), maxstep-1);
-	float t = sd->time*maxstep - step;
+	int step = min((int)(ccl_fetch(sd, time)*maxstep), maxstep-1);
+	float t = ccl_fetch(sd, time)*maxstep - step;
 
 	/* find attribute */
 	AttributeElement elem;
-	int offset = find_attribute_motion(kg, sd->object, ATTR_STD_MOTION_VERTEX_POSITION, &elem);
+	int offset = find_attribute_motion(kg, ccl_fetch(sd, object), ATTR_STD_MOTION_VERTEX_POSITION, &elem);
 	kernel_assert(offset != ATTR_STD_NOT_FOUND);
 
 	/* fetch vertex coordinates */
 	float3 verts[3], next_verts[3];
-	float3 tri_vindex = float4_to_float3(kernel_tex_fetch(__tri_vindex, sd->prim));
+	float3 tri_vindex = float4_to_float3(kernel_tex_fetch(__tri_vindex, ccl_fetch(sd, prim)));
 
 	motion_triangle_verts_for_step(kg, tri_vindex, offset, numverts, numsteps, step, verts);
 	motion_triangle_verts_for_step(kg, tri_vindex, offset, numverts, numsteps, step+1, next_verts);
@@ -268,33 +268,33 @@ ccl_device_noinline void motion_triangle_shader_setup(KernelGlobals *kg, ShaderD
 #ifdef __SUBSURFACE__
 	if(!subsurface)
 #endif
-		sd->P = motion_triangle_refine(kg, sd, isect, ray, verts);
+		ccl_fetch(sd, P) = motion_triangle_refine(kg, sd, isect, ray, verts);
 #ifdef __SUBSURFACE__
 	else
-		sd->P = motion_triangle_refine_subsurface(kg, sd, isect, ray, verts);
+		ccl_fetch(sd, P) = motion_triangle_refine_subsurface(kg, sd, isect, ray, verts);
 #endif
 
 	/* compute face normal */
 	float3 Ng;
-	if(sd->flag & SD_NEGATIVE_SCALE_APPLIED)
+	if(ccl_fetch(sd, flag) & SD_NEGATIVE_SCALE_APPLIED)
 		Ng = normalize(cross(verts[2] - verts[0], verts[1] - verts[0]));
 	else
 		Ng = normalize(cross(verts[1] - verts[0], verts[2] - verts[0]));
 
-	sd->Ng = Ng;
-	sd->N = Ng;
+	ccl_fetch(sd, Ng) = Ng;
+	ccl_fetch(sd, N) = Ng;
 
 	/* compute derivatives of P w.r.t. uv */
 #ifdef __DPDU__
-	sd->dPdu = (verts[0] - verts[2]);
-	sd->dPdv = (verts[1] - verts[2]);
+	ccl_fetch(sd, dPdu) = (verts[0] - verts[2]);
+	ccl_fetch(sd, dPdv) = (verts[1] - verts[2]);
 #endif
 
 	/* compute smooth normal */
-	if(sd->shader & SHADER_SMOOTH_NORMAL) {
+	if(ccl_fetch(sd, shader) & SHADER_SMOOTH_NORMAL) {
 		/* find attribute */
 		AttributeElement elem;
-		int offset = find_attribute_motion(kg, sd->object, ATTR_STD_MOTION_VERTEX_NORMAL, &elem);
+		int offset = find_attribute_motion(kg, ccl_fetch(sd, object), ATTR_STD_MOTION_VERTEX_NORMAL, &elem);
 		kernel_assert(offset != ATTR_STD_NOT_FOUND);
 
 		/* fetch vertex coordinates */
@@ -308,10 +308,10 @@ ccl_device_noinline void motion_triangle_shader_setup(KernelGlobals *kg, ShaderD
 		normals[2] = (1.0f - t)*normals[2] + t*next_normals[2];
 
 		/* interpolate between vertices */
-		float u = sd->u;
-		float v = sd->v;
+		float u = ccl_fetch(sd, u);
+		float v = ccl_fetch(sd, v);
 		float w = 1.0f - u - v;
-		sd->N = (u*normals[0] + v*normals[1] + w*normals[2]);
+		ccl_fetch(sd, N) = (u*normals[0] + v*normals[1] + w*normals[2]);
 	}
 }
 
diff --git a/intern/cycles/kernel/geom/geom_object.h b/intern/cycles/kernel/geom/geom_object.h
index 7df71010232..40cbca243a7 100644
--- a/intern/cycles/kernel/geom/geom_object.h
+++ b/intern/cycles/kernel/geom/geom_object.h
@@ -123,9 +123,9 @@ ccl_device_inline Transform object_fetch_transform_motion_test(KernelGlobals *kg
 ccl_device_inline void object_position_transform(KernelGlobals *kg, const ShaderData *sd, float3 *P)
 {
 #ifdef __OBJECT_MOTION__
-	*P = transform_point(&sd->ob_tfm, *P);
+	*P = transform_point(&ccl_fetch(sd, ob_tfm), *P);
 #else
-	Transform tfm = object_fetch_transform(kg, sd->object, OBJECT_TRANSFORM);
+	Transform tfm = object_fetch_transform(kg, ccl_fetch(sd, object), OBJECT_TRANSFORM);
 	*P = transform_point(&tfm, *P);
 #endif
 }
@@ -135,9 +135,9 @@ ccl_device_inline void object_position_transform(KernelGlobals *kg, const Shader
 ccl_device_inline void object_inverse_position_transform(KernelGlobals *kg, const ShaderData *sd, float3 *P)
 {
 #ifdef __OBJECT_MOTION__
-	*P = transform_point(&sd->ob_itfm, *P);
+	*P = transform_point(&ccl_fetch(sd, ob_itfm), *P);
 #else
-	Transform tfm = object_fetch_transform(kg, sd->object, OBJECT_INVERSE_TRANSFORM);
+	Transform tfm = object_fetch_transform(kg, ccl_fetch(sd, object), OBJECT_INVERSE_TRANSFORM);
 	*P = transform_point(&tfm, *P);
 #endif
 }
@@ -147,9 +147,9 @@ ccl_device_inline void object_inverse_position_transform(KernelGlobals *kg, cons
 ccl_device_inline void object_inverse_normal_transform(KernelGlobals *kg, const ShaderData *sd, float3 *N)
 {
 #ifdef __OBJECT_MOTION__
-	*N = normalize(transform_direction_transposed(&sd->ob_tfm, *N));
+	*N = normalize(transform_direction_transposed(&ccl_fetch(sd, ob_tfm), *N));
 #else
-	Transform tfm = object_fetch_transform(kg, sd->object, OBJECT_TRANSFORM);
+	Transform tfm = object_fetch_transform(kg, ccl_fetch(sd, object), OBJECT_TRANSFORM);
 	*N = normalize(transform_direction_transposed(&tfm, *N));
 #endif
 }
@@ -159,9 +159,9 @@ ccl_device_inline void object_inverse_normal_transform(KernelGlobals *kg, const
 ccl_device_inline void object_normal_transform(KernelGlobals *kg, const ShaderData *sd, float3 *N)
 {
 #ifdef __OBJECT_MOTION__
-	*N = normalize(transform_direction_transposed(&sd->ob_itfm, *N));
+	*N = normalize(transform_direction_transposed(&ccl_fetch(sd, ob_itfm), *N));
 #else
-	Transform tfm = object_fetch_transform(kg, sd->object, OBJECT_INVERSE_TRANSFORM);
+	Transform tfm = object_fetch_transform(kg, ccl_fetch(sd, object), OBJECT_INVERSE_TRANSFORM);
 	*N = normalize(transform_direction_transposed(&tfm, *N));
 #endif
 }
@@ -171,9 +171,9 @@ ccl_device_inline void object_normal_transform(KernelGlobals *kg, const ShaderDa
 ccl_device_inline void object_dir_transform(KernelGlobals *kg, const ShaderData *sd, float3 *D)
 {
 #ifdef __OBJECT_MOTION__
-	*D = transform_direction(&sd->ob_tfm, *D);
+	*D = transform_direction(&ccl_fetch(sd, ob_tfm), *D);
 #else
-	Transform tfm = object_fetch_transform(kg, sd->object, OBJECT_TRANSFORM);
+	Transform tfm = object_fetch_transform(kg, ccl_fetch(sd, object), OBJECT_TRANSFORM);
 	*D = transform_direction(&tfm, *D);
 #endif
 }
@@ -183,9 +183,9 @@ ccl_device_inline void object_dir_transform(KernelGlobals *kg, const ShaderData
 ccl_device_inline void object_inverse_dir_transform(KernelGlobals *kg, const ShaderData *sd, float3 *D)
 {
 #ifdef __OBJECT_MOTION__
-	*D = transform_direction(&sd->ob_itfm, *D);
+	*D = transform_direction(&ccl_fetch(sd, ob_itfm), *D);
 #else
-	Transform tfm = object_fetch_transform(kg, sd->object, OBJECT_INVERSE_TRANSFORM);
+	Transform tfm = object_fetch_transform(kg, ccl_fetch(sd, object), OBJECT_INVERSE_TRANSFORM);
 	*D = transform_direction(&tfm, *D);
 #endif
 }
@@ -194,13 +194,13 @@ ccl_device_inline void object_inverse_dir_transform(KernelGlobals *kg, const Sha
 
 ccl_device_inline float3 object_location(KernelGlobals *kg, const ShaderData *sd)
 {
-	if(sd->object == OBJECT_NONE)
+	if(ccl_fetch(sd, object) == OBJECT_NONE)
 		return make_float3(0.0f, 0.0f, 0.0f);
 
 #ifdef __OBJECT_MOTION__
-	return make_float3(sd->ob_tfm.x.w, sd->ob_tfm.y.w, sd->ob_tfm.z.w);
+	return make_float3(ccl_fetch(sd, ob_tfm).x.w, ccl_fetch(sd, ob_tfm).y.w, ccl_fetch(sd, ob_tfm).z.w);
 #else
-	Transform tfm = object_fetch_transform(kg, sd->object, OBJECT_TRANSFORM);
+	Transform tfm = object_fetch_transform(kg, ccl_fetch(sd, object), OBJECT_TRANSFORM);
 	return make_float3(tfm.x.w, tfm.y.w, tfm.z.w);
 #endif
 }
@@ -296,7 +296,7 @@ ccl_device_inline void object_motion_info(KernelGlobals *kg, int object, int *nu
 
 ccl_device int shader_pass_id(KernelGlobals *kg, const ShaderData *sd)
 {
-	return kernel_tex_fetch(__shader_flag, (sd->shader & SHADER_MASK)*2 + 1);
+	return kernel_tex_fetch(__shader_flag, (ccl_fetch(sd, shader) & SHADER_MASK)*2 + 1);
 }
 
 /* Particle data from which object was instanced */
@@ -377,7 +377,7 @@ ccl_device_inline float3 bvh_inverse_direction(float3 dir)
 
 /* Transform ray into object space to enter static object in BVH */
 
-ccl_device_inline void bvh_instance_push(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *dir, float3 *idir, float *t)
+ccl_device_inline void bvh_instance_push(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *dir, float3 *idir, ccl_addr_space float *t)
 {
 	Transform tfm = object_fetch_transform(kg, object, OBJECT_INVERSE_TRANSFORM);
 
@@ -425,7 +425,7 @@ ccl_device_inline void qbvh_instance_push(KernelGlobals *kg,
 
 /* Transorm ray to exit static object in BVH */
 
-ccl_device_inline void bvh_instance_pop(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *dir, float3 *idir, float *t)
+ccl_device_inline void bvh_instance_pop(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *dir, float3 *idir, ccl_addr_space float *t)
 {
 	if(*t != FLT_MAX) {
 		Transform tfm = object_fetch_transform(kg, object, OBJECT_TRANSFORM);
@@ -520,5 +520,38 @@ ccl_device_inline void bvh_instance_motion_pop_factor(KernelGlobals *kg, int obj
 
 #endif
 
+/* TODO(sergey): This is only for until we've got OpenCL 2.0
+ * on all devices we consider supported. It'll be replaced with
+ * generic address space.
+ */
+
+#ifdef __KERNEL_OPENCL__
+ccl_device_inline void object_dir_transform_addrspace(KernelGlobals *kg,
+                                                      const ShaderData *sd,
+                                                      ccl_addr_space float3 *D)
+{
+	float3 private_D = *D;
+	object_dir_transform(kg, sd, &private_D);
+	*D = private_D;
+}
+
+ccl_device_inline void object_normal_transform_addrspace(KernelGlobals *kg,
+                                                         const ShaderData *sd,
+                                                         ccl_addr_space float3 *N)
+{
+	float3 private_N = *N;
+	object_dir_transform(kg, sd, &private_N);
+	*N = private_N;
+}
+#endif
+
+#ifndef __KERNEL_OPENCL__
+#  define object_dir_transform_auto object_dir_transform
+#  define object_normal_transform_auto object_normal_transform
+#else
+#  define object_dir_transform_auto object_dir_transform_addrspace
+#  define object_normal_transform_auto object_normal_transform_addrspace
+#endif
+
 CCL_NAMESPACE_END
 
diff --git a/intern/cycles/kernel/geom/geom_primitive.h b/intern/cycles/kernel/geom/geom_primitive.h
index d2543c5943e..30f12d32355 100644
--- a/intern/cycles/kernel/geom/geom_primitive.h
+++ b/intern/cycles/kernel/geom/geom_primitive.h
@@ -25,16 +25,16 @@ CCL_NAMESPACE_BEGIN
 
 ccl_device float primitive_attribute_float(KernelGlobals *kg, const ShaderData *sd, AttributeElement elem, int offset, float *dx, float *dy)
 {
-	if(sd->type & PRIMITIVE_ALL_TRIANGLE) {
+	if(ccl_fetch(sd, type) & PRIMITIVE_ALL_TRIANGLE) {
 		return triangle_attribute_float(kg, sd, elem, offset, dx, dy);
 	}
 #ifdef __HAIR__
-	else if(sd->type & PRIMITIVE_ALL_CURVE) {
+	else if(ccl_fetch(sd, type) & PRIMITIVE_ALL_CURVE) {
 		return curve_attribute_float(kg, sd, elem, offset, dx, dy);
 	}
 #endif
 #ifdef __VOLUME__
-	else if(sd->object != OBJECT_NONE && elem == ATTR_ELEMENT_VOXEL) {
+	else if(ccl_fetch(sd, object) != OBJECT_NONE && elem == ATTR_ELEMENT_VOXEL) {
 		return volume_attribute_float(kg, sd, elem, offset, dx, dy);
 	}
 #endif
@@ -47,16 +47,16 @@ ccl_device float primitive_attribute_float(KernelGlobals *kg, const ShaderData *
 
 ccl_device float3 primitive_attribute_float3(KernelGlobals *kg, const ShaderData *sd, AttributeElement elem, int offset, float3 *dx, float3 *dy)
 {
-	if(sd->type & PRIMITIVE_ALL_TRIANGLE) {
+	if(ccl_fetch(sd, type) & PRIMITIVE_ALL_TRIANGLE) {
 		return triangle_attribute_float3(kg, sd, elem, offset, dx, dy);
 	}
 #ifdef __HAIR__
-	else if(sd->type & PRIMITIVE_ALL_CURVE) {
+	else if(ccl_fetch(sd, type) & PRIMITIVE_ALL_CURVE) {
 		return curve_attribute_float3(kg, sd, elem, offset, dx, dy);
 	}
 #endif
 #ifdef __VOLUME__
-	else if(sd->object != OBJECT_NONE && elem == ATTR_ELEMENT_VOXEL) {
+	else if(ccl_fetch(sd, object) != OBJECT_NONE && elem == ATTR_ELEMENT_VOXEL) {
 		return volume_attribute_float3(kg, sd, elem, offset, dx, dy);
 	}
 #endif
@@ -108,9 +108,9 @@ ccl_device bool primitive_ptex(KernelGlobals *kg, ShaderData *sd, float2 *uv, in
 ccl_device float3 primitive_tangent(KernelGlobals *kg, ShaderData *sd)
 {
 #ifdef __HAIR__
-	if(sd->type & PRIMITIVE_ALL_CURVE)
+	if(ccl_fetch(sd, type) & PRIMITIVE_ALL_CURVE)
 #ifdef __DPDU__
-		return normalize(sd->dPdu);
+		return normalize(ccl_fetch(sd, dPdu));
 #else
 		return make_float3(0.0f, 0.0f, 0.0f);
 #endif
@@ -124,12 +124,12 @@ ccl_device float3 primitive_tangent(KernelGlobals *kg, ShaderData *sd)
 		float3 data = primitive_attribute_float3(kg, sd, attr_elem, attr_offset, NULL, NULL);
 		data = make_float3(-(data.y - 0.5f), (data.x - 0.5f), 0.0f);
 		object_normal_transform(kg, sd, &data);
-		return cross(sd->N, normalize(cross(data, sd->N)));
+		return cross(ccl_fetch(sd, N), normalize(cross(data, ccl_fetch(sd, N))));
 	}
 	else {
 		/* otherwise use surface derivatives */
 #ifdef __DPDU__
-		return normalize(sd->dPdu);
+		return normalize(ccl_fetch(sd, dPdu));
 #else
 		return make_float3(0.0f, 0.0f, 0.0f);
 #endif
@@ -144,16 +144,16 @@ ccl_device float4 primitive_motion_vector(KernelGlobals *kg, ShaderData *sd)
 	float3 center;
 
 #ifdef __HAIR__
-	bool is_curve_primitive = sd->type & PRIMITIVE_ALL_CURVE;
+	bool is_curve_primitive = ccl_fetch(sd, type) & PRIMITIVE_ALL_CURVE;
 	if(is_curve_primitive) {
 		center = curve_motion_center_location(kg, sd);
 
-		if(!(sd->flag & SD_TRANSFORM_APPLIED))
+		if(!(ccl_fetch(sd, flag) & SD_TRANSFORM_APPLIED))
 			object_position_transform(kg, sd, &center);
 	}
 	else
 #endif
-		center = sd->P;
+		center = ccl_fetch(sd, P);
 
 	float3 motion_pre = center, motion_post = center;
 
@@ -164,16 +164,16 @@ ccl_device float4 primitive_motion_vector(KernelGlobals *kg, ShaderData *sd)
 	if(offset != ATTR_STD_NOT_FOUND) {
 		/* get motion info */
 		int numverts, numkeys;
-		object_motion_info(kg, sd->object, NULL, &numverts, &numkeys);
+		object_motion_info(kg, ccl_fetch(sd, object), NULL, &numverts, &numkeys);
 
 		/* lookup attributes */
-		int offset_next = (sd->type & PRIMITIVE_ALL_TRIANGLE)? offset + numverts: offset + numkeys;
+		int offset_next = (ccl_fetch(sd, type) & PRIMITIVE_ALL_TRIANGLE)? offset + numverts: offset + numkeys;
 
 		motion_pre = primitive_attribute_float3(kg, sd, elem, offset, NULL, NULL);
 		motion_post = primitive_attribute_float3(kg, sd, elem, offset_next, NULL, NULL);
 
 #ifdef __HAIR__
-		if(is_curve_primitive && (sd->flag & SD_OBJECT_HAS_VERTEX_MOTION) == 0) {
+		if(is_curve_primitive && (ccl_fetch(sd, flag) & SD_OBJECT_HAS_VERTEX_MOTION) == 0) {
 			object_position_transform(kg, sd, &motion_pre);
 			object_position_transform(kg, sd, &motion_post);
 		}
@@ -184,10 +184,10 @@ ccl_device float4 primitive_motion_vector(KernelGlobals *kg, ShaderData *sd)
 	 * transformation was set match the world/object space of motion_pre/post */
 	Transform tfm;
 	
-	tfm = object_fetch_vector_transform(kg, sd->object, OBJECT_VECTOR_MOTION_PRE);
+	tfm = object_fetch_vector_transform(kg, ccl_fetch(sd, object), OBJECT_VECTOR_MOTION_PRE);
 	motion_pre = transform_point(&tfm, motion_pre);
 
-	tfm = object_fetch_vector_transform(kg, sd->object, OBJECT_VECTOR_MOTION_POST);
+	tfm = object_fetch_vector_transform(kg, ccl_fetch(sd, object), OBJECT_VECTOR_MOTION_POST);
 	motion_post = transform_point(&tfm, motion_post);
 
 	float3 motion_center;
diff --git a/intern/cycles/kernel/geom/geom_triangle.h b/intern/cycles/kernel/geom/geom_triangle.h
index dd3928682e3..995dfac5b09 100644
--- a/intern/cycles/kernel/geom/geom_triangle.h
+++ b/intern/cycles/kernel/geom/geom_triangle.h
@@ -27,14 +27,14 @@ CCL_NAMESPACE_BEGIN
 ccl_device_inline float3 triangle_normal(KernelGlobals *kg, ShaderData *sd)
 {
 	/* load triangle vertices */
-	float4 tri_vindex = kernel_tex_fetch(__tri_vindex, sd->prim);
+	float4 tri_vindex = kernel_tex_fetch(__tri_vindex, ccl_fetch(sd, prim));
 
 	float3 v0 = float4_to_float3(kernel_tex_fetch(__tri_verts, __float_as_int(tri_vindex.x)));
 	float3 v1 = float4_to_float3(kernel_tex_fetch(__tri_verts, __float_as_int(tri_vindex.y)));
 	float3 v2 = float4_to_float3(kernel_tex_fetch(__tri_verts, __float_as_int(tri_vindex.z)));
 	
 	/* return normal */
-	if(sd->flag & SD_NEGATIVE_SCALE_APPLIED)
+	if(ccl_fetch(sd, flag) & SD_NEGATIVE_SCALE_APPLIED)
 		return normalize(cross(v2 - v0, v1 - v0));
 	else
 		return normalize(cross(v1 - v0, v2 - v0));
@@ -94,7 +94,7 @@ ccl_device_inline float3 triangle_smooth_normal(KernelGlobals *kg, int prim, flo
 
 /* Ray differentials on triangle */
 
-ccl_device_inline void triangle_dPdudv(KernelGlobals *kg, int prim, float3 *dPdu, float3 *dPdv)
+ccl_device_inline void triangle_dPdudv(KernelGlobals *kg, int prim, ccl_addr_space float3 *dPdu, ccl_addr_space float3 *dPdv)
 {
 	/* fetch triangle vertex coordinates */
 	float4 tri_vindex = kernel_tex_fetch(__tri_vindex, prim);
@@ -116,34 +116,34 @@ ccl_device float triangle_attribute_float(KernelGlobals *kg, const ShaderData *s
 		if(dx) *dx = 0.0f;
 		if(dy) *dy = 0.0f;
 
-		return kernel_tex_fetch(__attributes_float, offset + sd->prim);
+		return kernel_tex_fetch(__attributes_float, offset + ccl_fetch(sd, prim));
 	}
 	else if(elem == ATTR_ELEMENT_VERTEX || elem == ATTR_ELEMENT_VERTEX_MOTION) {
-		float4 tri_vindex = kernel_tex_fetch(__tri_vindex, sd->prim);
+		float4 tri_vindex = kernel_tex_fetch(__tri_vindex, ccl_fetch(sd, prim));
 
 		float f0 = kernel_tex_fetch(__attributes_float, offset + __float_as_int(tri_vindex.x));
 		float f1 = kernel_tex_fetch(__attributes_float, offset + __float_as_int(tri_vindex.y));
 		float f2 = kernel_tex_fetch(__attributes_float, offset + __float_as_int(tri_vindex.z));
 
 #ifdef __RAY_DIFFERENTIALS__
-		if(dx) *dx = sd->du.dx*f0 + sd->dv.dx*f1 - (sd->du.dx + sd->dv.dx)*f2;
-		if(dy) *dy = sd->du.dy*f0 + sd->dv.dy*f1 - (sd->du.dy + sd->dv.dy)*f2;
+		if(dx) *dx = ccl_fetch(sd, du).dx*f0 + ccl_fetch(sd, dv).dx*f1 - (ccl_fetch(sd, du).dx + ccl_fetch(sd, dv).dx)*f2;
+		if(dy) *dy = ccl_fetch(sd, du).dy*f0 + ccl_fetch(sd, dv).dy*f1 - (ccl_fetch(sd, du).dy + ccl_fetch(sd, dv).dy)*f2;
 #endif
 
-		return sd->u*f0 + sd->v*f1 + (1.0f - sd->u - sd->v)*f2;
+		return ccl_fetch(sd, u)*f0 + ccl_fetch(sd, v)*f1 + (1.0f - ccl_fetch(sd, u) - ccl_fetch(sd, v))*f2;
 	}
 	else if(elem == ATTR_ELEMENT_CORNER) {
-		int tri = offset + sd->prim*3;
+		int tri = offset + ccl_fetch(sd, prim)*3;
 		float f0 = kernel_tex_fetch(__attributes_float, tri + 0);
 		float f1 = kernel_tex_fetch(__attributes_float, tri + 1);
 		float f2 = kernel_tex_fetch(__attributes_float, tri + 2);
 
 #ifdef __RAY_DIFFERENTIALS__
-		if(dx) *dx = sd->du.dx*f0 + sd->dv.dx*f1 - (sd->du.dx + sd->dv.dx)*f2;
-		if(dy) *dy = sd->du.dy*f0 + sd->dv.dy*f1 - (sd->du.dy + sd->dv.dy)*f2;
+		if(dx) *dx = ccl_fetch(sd, du).dx*f0 + ccl_fetch(sd, dv).dx*f1 - (ccl_fetch(sd, du).dx + ccl_fetch(sd, dv).dx)*f2;
+		if(dy) *dy = ccl_fetch(sd, du).dy*f0 + ccl_fetch(sd, dv).dy*f1 - (ccl_fetch(sd, du).dy + ccl_fetch(sd, dv).dy)*f2;
 #endif
 
-		return sd->u*f0 + sd->v*f1 + (1.0f - sd->u - sd->v)*f2;
+		return ccl_fetch(sd, u)*f0 + ccl_fetch(sd, v)*f1 + (1.0f - ccl_fetch(sd, u) - ccl_fetch(sd, v))*f2;
 	}
 	else {
 		if(dx) *dx = 0.0f;
@@ -159,24 +159,24 @@ ccl_device float3 triangle_attribute_float3(KernelGlobals *kg, const ShaderData
 		if(dx) *dx = make_float3(0.0f, 0.0f, 0.0f);
 		if(dy) *dy = make_float3(0.0f, 0.0f, 0.0f);
 
-		return float4_to_float3(kernel_tex_fetch(__attributes_float3, offset + sd->prim));
+		return float4_to_float3(kernel_tex_fetch(__attributes_float3, offset + ccl_fetch(sd, prim)));
 	}
 	else if(elem == ATTR_ELEMENT_VERTEX || elem == ATTR_ELEMENT_VERTEX_MOTION) {
-		float4 tri_vindex = kernel_tex_fetch(__tri_vindex, sd->prim);
+		float4 tri_vindex = kernel_tex_fetch(__tri_vindex, ccl_fetch(sd, prim));
 
 		float3 f0 = float4_to_float3(kernel_tex_fetch(__attributes_float3, offset + __float_as_int(tri_vindex.x)));
 		float3 f1 = float4_to_float3(kernel_tex_fetch(__attributes_float3, offset + __float_as_int(tri_vindex.y)));
 		float3 f2 = float4_to_float3(kernel_tex_fetch(__attributes_float3, offset + __float_as_int(tri_vindex.z)));
 
 #ifdef __RAY_DIFFERENTIALS__
-		if(dx) *dx = sd->du.dx*f0 + sd->dv.dx*f1 - (sd->du.dx + sd->dv.dx)*f2;
-		if(dy) *dy = sd->du.dy*f0 + sd->dv.dy*f1 - (sd->du.dy + sd->dv.dy)*f2;
+		if(dx) *dx = ccl_fetch(sd, du).dx*f0 + ccl_fetch(sd, dv).dx*f1 - (ccl_fetch(sd, du).dx + ccl_fetch(sd, dv).dx)*f2;
+		if(dy) *dy = ccl_fetch(sd, du).dy*f0 + ccl_fetch(sd, dv).dy*f1 - (ccl_fetch(sd, du).dy + ccl_fetch(sd, dv).dy)*f2;
 #endif
 
-		return sd->u*f0 + sd->v*f1 + (1.0f - sd->u - sd->v)*f2;
+		return ccl_fetch(sd, u)*f0 + ccl_fetch(sd, v)*f1 + (1.0f - ccl_fetch(sd, u) - ccl_fetch(sd, v))*f2;
 	}
 	else if(elem == ATTR_ELEMENT_CORNER || elem == ATTR_ELEMENT_CORNER_BYTE) {
-		int tri = offset + sd->prim*3;
+		int tri = offset + ccl_fetch(sd, prim)*3;
 		float3 f0, f1, f2;
 
 		if(elem == ATTR_ELEMENT_CORNER) {
@@ -191,11 +191,11 @@ ccl_device float3 triangle_attribute_float3(KernelGlobals *kg, const ShaderData
 		}
 
 #ifdef __RAY_DIFFERENTIALS__
-		if(dx) *dx = sd->du.dx*f0 + sd->dv.dx*f1 - (sd->du.dx + sd->dv.dx)*f2;
-		if(dy) *dy = sd->du.dy*f0 + sd->dv.dy*f1 - (sd->du.dy + sd->dv.dy)*f2;
+		if(dx) *dx = ccl_fetch(sd, du).dx*f0 + ccl_fetch(sd, dv).dx*f1 - (ccl_fetch(sd, du).dx + ccl_fetch(sd, dv).dx)*f2;
+		if(dy) *dy = ccl_fetch(sd, du).dy*f0 + ccl_fetch(sd, dv).dy*f1 - (ccl_fetch(sd, du).dy + ccl_fetch(sd, dv).dy)*f2;
 #endif
 
-		return sd->u*f0 + sd->v*f1 + (1.0f - sd->u - sd->v)*f2;
+		return ccl_fetch(sd, u)*f0 + ccl_fetch(sd, v)*f1 + (1.0f - ccl_fetch(sd, u) - ccl_fetch(sd, v))*f2;
 	}
 	else {
 		if(dx) *dx = make_float3(0.0f, 0.0f, 0.0f);