2 * Adapted from code Copyright 2009-2010 NVIDIA Corporation,
3 * and code copyright 2009-2012 Intel Corporation
5 * Modifications Copyright 2011-2014, Blender Foundation.
7 * Licensed under the Apache License, Version 2.0 (the "License");
8 * you may not use this file except in compliance with the License.
9 * You may obtain a copy of the License at
11 * http://www.apache.org/licenses/LICENSE-2.0
13 * Unless required by applicable law or agreed to in writing, software
14 * distributed under the License is distributed on an "AS IS" BASIS,
15 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16 * See the License for the specific language governing permissions and
17 * limitations under the License.
20 /* This is a template BVH traversal function for volumes, where
21 * various features can be enabled/disabled. This way we can compile optimized
22 * versions for each case without new features slowing things down.
24 * BVH_INSTANCING: object instancing
25 * BVH_HAIR: hair curve rendering
26 * BVH_MOTION: motion blur rendering
30 ccl_device uint BVH_FUNCTION_FULL_NAME(QBVH)(KernelGlobals *kg,
32 Intersection *isect_array,
36 * - Test if pushing distance on the stack helps.
37 * - Likely and unlikely for if() statements.
38 * - Test restrict attribute for pointers.
41 /* Traversal stack in CUDA thread-local memory. */
42 QBVHStackItem traversalStack[BVH_QSTACK_SIZE];
43 traversalStack[0].addr = ENTRYPOINT_SENTINEL;
45 /* Traversal variables in registers. */
47 int nodeAddr = kernel_data.bvh.root;
49 /* Ray parameters in registers. */
50 const float tmax = ray->t;
52 float3 dir = bvh_clamp_direction(ray->D);
53 float3 idir = bvh_inverse_direction(dir);
54 int object = OBJECT_NONE;
57 const uint visibility = PATH_RAY_ALL_VISIBILITY;
59 #if BVH_FEATURE(BVH_MOTION)
64 isect_array->t = tmax;
66 #ifndef __KERNEL_SSE41__
72 #if BVH_FEATURE(BVH_INSTANCING)
73 int num_hits_in_instance = 0;
76 ssef tnear(0.0f), tfar(isect_t);
77 sse3f idir4(ssef(idir.x), ssef(idir.y), ssef(idir.z));
79 #ifdef __KERNEL_AVX2__
80 float3 P_idir = P*idir;
81 sse3f P_idir4 = sse3f(P_idir.x, P_idir.y, P_idir.z);
83 sse3f org = sse3f(ssef(P.x), ssef(P.y), ssef(P.z));
86 /* Offsets to select the side that becomes the lower or upper bound. */
87 int near_x, near_y, near_z;
88 int far_x, far_y, far_z;
90 if(idir.x >= 0.0f) { near_x = 0; far_x = 1; } else { near_x = 1; far_x = 0; }
91 if(idir.y >= 0.0f) { near_y = 2; far_y = 3; } else { near_y = 3; far_y = 2; }
92 if(idir.z >= 0.0f) { near_z = 4; far_z = 5; } else { near_z = 5; far_z = 4; }
94 IsectPrecalc isect_precalc;
95 triangle_intersect_precalc(dir, &isect_precalc);
100 /* Traverse internal nodes. */
101 while(nodeAddr >= 0 && nodeAddr != ENTRYPOINT_SENTINEL) {
103 int traverseChild = qbvh_node_intersect(kg,
106 #ifdef __KERNEL_AVX2__
112 near_x, near_y, near_z,
117 if(traverseChild != 0) {
118 float4 cnodes = kernel_tex_fetch(__bvh_nodes, nodeAddr*BVH_QNODE_SIZE+6);
120 /* One child is hit, continue with that child. */
121 int r = __bscf(traverseChild);
122 if(traverseChild == 0) {
123 nodeAddr = __float_as_int(cnodes[r]);
127 /* Two children are hit, push far child, and continue with
130 int c0 = __float_as_int(cnodes[r]);
131 float d0 = ((float*)&dist)[r];
132 r = __bscf(traverseChild);
133 int c1 = __float_as_int(cnodes[r]);
134 float d1 = ((float*)&dist)[r];
135 if(traverseChild == 0) {
139 kernel_assert(stackPtr < BVH_QSTACK_SIZE);
140 traversalStack[stackPtr].addr = c0;
141 traversalStack[stackPtr].dist = d0;
147 kernel_assert(stackPtr < BVH_QSTACK_SIZE);
148 traversalStack[stackPtr].addr = c1;
149 traversalStack[stackPtr].dist = d1;
154 /* Here starts the slow path for 3 or 4 hit children. We push
155 * all nodes onto the stack to sort them there.
158 kernel_assert(stackPtr < BVH_QSTACK_SIZE);
159 traversalStack[stackPtr].addr = c1;
160 traversalStack[stackPtr].dist = d1;
162 kernel_assert(stackPtr < BVH_QSTACK_SIZE);
163 traversalStack[stackPtr].addr = c0;
164 traversalStack[stackPtr].dist = d0;
166 /* Three children are hit, push all onto stack and sort 3
167 * stack items, continue with closest child.
169 r = __bscf(traverseChild);
170 int c2 = __float_as_int(cnodes[r]);
171 float d2 = ((float*)&dist)[r];
172 if(traverseChild == 0) {
174 kernel_assert(stackPtr < BVH_QSTACK_SIZE);
175 traversalStack[stackPtr].addr = c2;
176 traversalStack[stackPtr].dist = d2;
177 qbvh_stack_sort(&traversalStack[stackPtr],
178 &traversalStack[stackPtr - 1],
179 &traversalStack[stackPtr - 2]);
180 nodeAddr = traversalStack[stackPtr].addr;
185 /* Four children are hit, push all onto stack and sort 4
186 * stack items, continue with closest child.
188 r = __bscf(traverseChild);
189 int c3 = __float_as_int(cnodes[r]);
190 float d3 = ((float*)&dist)[r];
192 kernel_assert(stackPtr < BVH_QSTACK_SIZE);
193 traversalStack[stackPtr].addr = c3;
194 traversalStack[stackPtr].dist = d3;
196 kernel_assert(stackPtr < BVH_QSTACK_SIZE);
197 traversalStack[stackPtr].addr = c2;
198 traversalStack[stackPtr].dist = d2;
199 qbvh_stack_sort(&traversalStack[stackPtr],
200 &traversalStack[stackPtr - 1],
201 &traversalStack[stackPtr - 2],
202 &traversalStack[stackPtr - 3]);
205 nodeAddr = traversalStack[stackPtr].addr;
209 /* If node is leaf, fetch triangle list. */
211 float4 leaf = kernel_tex_fetch(__bvh_leaf_nodes, (-nodeAddr-1)*BVH_QNODE_LEAF_SIZE);
212 int primAddr = __float_as_int(leaf.x);
214 #if BVH_FEATURE(BVH_INSTANCING)
217 int primAddr2 = __float_as_int(leaf.y);
218 const uint type = __float_as_int(leaf.w);
219 const uint p_type = type & PRIMITIVE_ALL;
223 nodeAddr = traversalStack[stackPtr].addr;
226 /* Primitive intersection. */
228 case PRIMITIVE_TRIANGLE: {
229 for(; primAddr < primAddr2; primAddr++) {
230 kernel_assert(kernel_tex_fetch(__prim_type, primAddr) == type);
231 /* Only primitives from volume object. */
232 uint tri_object = (object == OBJECT_NONE)? kernel_tex_fetch(__prim_object, primAddr): object;
233 int object_flag = kernel_tex_fetch(__object_flag, tri_object);
234 if((object_flag & SD_OBJECT_HAS_VOLUME) == 0) {
237 /* Intersect ray against primitive. */
238 hit = triangle_intersect(kg, &isect_precalc, isect_array, P, visibility, object, primAddr);
240 /* Move on to next entry in intersections array. */
243 #if BVH_FEATURE(BVH_INSTANCING)
244 num_hits_in_instance++;
246 isect_array->t = isect_t;
247 if(num_hits == max_hits) {
248 #if BVH_FEATURE(BVH_INSTANCING)
249 # if BVH_FEATURE(BVH_MOTION)
250 float t_fac = 1.0f / len(transform_direction(&ob_itfm, dir));
252 Transform itfm = object_fetch_transform(kg, object, OBJECT_INVERSE_TRANSFORM);
253 float t_fac = 1.0f / len(transform_direction(&itfm, dir));
255 for(int i = 0; i < num_hits_in_instance; i++) {
256 (isect_array-i-1)->t *= t_fac;
258 #endif /* BVH_FEATURE(BVH_INSTANCING) */
265 #if BVH_FEATURE(BVH_MOTION)
266 case PRIMITIVE_MOTION_TRIANGLE: {
267 for(; primAddr < primAddr2; primAddr++) {
268 kernel_assert(kernel_tex_fetch(__prim_type, primAddr) == type);
269 /* Only primitives from volume object. */
270 uint tri_object = (object == OBJECT_NONE)? kernel_tex_fetch(__prim_object, primAddr): object;
271 int object_flag = kernel_tex_fetch(__object_flag, tri_object);
272 if((object_flag & SD_OBJECT_HAS_VOLUME) == 0) {
275 /* Intersect ray against primitive. */
276 hit = motion_triangle_intersect(kg, isect_array, P, dir, ray->time, visibility, object, primAddr);
278 /* Move on to next entry in intersections array. */
281 # if BVH_FEATURE(BVH_INSTANCING)
282 num_hits_in_instance++;
284 isect_array->t = isect_t;
285 if(num_hits == max_hits) {
286 # if BVH_FEATURE(BVH_INSTANCING)
287 # if BVH_FEATURE(BVH_MOTION)
288 float t_fac = 1.0f / len(transform_direction(&ob_itfm, dir));
290 Transform itfm = object_fetch_transform(kg, object, OBJECT_INVERSE_TRANSFORM);
291 float t_fac = 1.0f / len(transform_direction(&itfm, dir));
293 for(int i = 0; i < num_hits_in_instance; i++) {
294 (isect_array-i-1)->t *= t_fac;
296 # endif /* BVH_FEATURE(BVH_INSTANCING) */
304 #if BVH_FEATURE(BVH_HAIR)
305 case PRIMITIVE_CURVE:
306 case PRIMITIVE_MOTION_CURVE: {
307 for(; primAddr < primAddr2; primAddr++) {
308 kernel_assert(kernel_tex_fetch(__prim_type, primAddr) == type);
309 /* Only primitives from volume object. */
310 uint tri_object = (object == OBJECT_NONE)? kernel_tex_fetch(__prim_object, primAddr): object;
311 int object_flag = kernel_tex_fetch(__object_flag, tri_object);
312 if((object_flag & SD_OBJECT_HAS_VOLUME) == 0) {
315 /* Intersect ray against primitive. */
316 if(kernel_data.curve.curveflags & CURVE_KN_INTERPOLATE)
317 hit = bvh_cardinal_curve_intersect(kg, isect_array, P, dir, visibility, object, primAddr, ray->time, type, NULL, 0, 0);
319 hit = bvh_curve_intersect(kg, isect_array, P, dir, visibility, object, primAddr, ray->time, type, NULL, 0, 0);
321 /* Move on to next entry in intersections array. */
324 # if BVH_FEATURE(BVH_INSTANCING)
325 num_hits_in_instance++;
327 isect_array->t = isect_t;
328 if(num_hits == max_hits) {
329 # if BVH_FEATURE(BVH_INSTANCING)
330 # if BVH_FEATURE(BVH_MOTION)
331 float t_fac = 1.0f / len(transform_direction(&ob_itfm, dir));
333 Transform itfm = object_fetch_transform(kg, object, OBJECT_INVERSE_TRANSFORM);
334 float t_fac = 1.0f / len(transform_direction(&itfm, dir));
336 for(int i = 0; i < num_hits_in_instance; i++) {
337 (isect_array-i-1)->t *= t_fac;
339 # endif /* BVH_FEATURE(BVH_INSTANCING) */
346 #endif /* BVH_HAIR */
349 #if BVH_FEATURE(BVH_INSTANCING)
352 object = kernel_tex_fetch(__prim_object, -primAddr-1);
353 int object_flag = kernel_tex_fetch(__object_flag, object);
355 if(object_flag & SD_OBJECT_HAS_VOLUME) {
357 # if BVH_FEATURE(BVH_MOTION)
358 bvh_instance_motion_push(kg, object, ray, &P, &dir, &idir, &isect_t, &ob_itfm);
360 bvh_instance_push(kg, object, ray, &P, &dir, &idir, &isect_t);
363 if(idir.x >= 0.0f) { near_x = 0; far_x = 1; } else { near_x = 1; far_x = 0; }
364 if(idir.y >= 0.0f) { near_y = 2; far_y = 3; } else { near_y = 3; far_y = 2; }
365 if(idir.z >= 0.0f) { near_z = 4; far_z = 5; } else { near_z = 5; far_z = 4; }
366 tfar = ssef(isect_t);
367 idir4 = sse3f(ssef(idir.x), ssef(idir.y), ssef(idir.z));
368 # ifdef __KERNEL_AVX2__
370 P_idir4 = sse3f(P_idir.x, P_idir.y, P_idir.z);
372 org = sse3f(ssef(P.x), ssef(P.y), ssef(P.z));
374 triangle_intersect_precalc(dir, &isect_precalc);
375 num_hits_in_instance = 0;
376 isect_array->t = isect_t;
379 kernel_assert(stackPtr < BVH_QSTACK_SIZE);
380 traversalStack[stackPtr].addr = ENTRYPOINT_SENTINEL;
382 nodeAddr = kernel_tex_fetch(__object_node, object);
386 object = OBJECT_NONE;
387 nodeAddr = traversalStack[stackPtr].addr;
392 #endif /* FEATURE(BVH_INSTANCING) */
393 } while(nodeAddr != ENTRYPOINT_SENTINEL);
395 #if BVH_FEATURE(BVH_INSTANCING)
397 kernel_assert(object != OBJECT_NONE);
400 if(num_hits_in_instance) {
402 # if BVH_FEATURE(BVH_MOTION)
403 bvh_instance_motion_pop_factor(kg, object, ray, &P, &dir, &idir, &t_fac, &ob_itfm);
405 bvh_instance_pop_factor(kg, object, ray, &P, &dir, &idir, &t_fac);
407 triangle_intersect_precalc(dir, &isect_precalc);
408 /* Scale isect->t to adjust for instancing. */
409 for(int i = 0; i < num_hits_in_instance; i++) {
410 (isect_array-i-1)->t *= t_fac;
414 float ignore_t = FLT_MAX;
415 # if BVH_FEATURE(BVH_MOTION)
416 bvh_instance_motion_pop(kg, object, ray, &P, &dir, &idir, &ignore_t, &ob_itfm);
418 bvh_instance_pop(kg, object, ray, &P, &dir, &idir, &ignore_t);
420 triangle_intersect_precalc(dir, &isect_precalc);
423 if(idir.x >= 0.0f) { near_x = 0; far_x = 1; } else { near_x = 1; far_x = 0; }
424 if(idir.y >= 0.0f) { near_y = 2; far_y = 3; } else { near_y = 3; far_y = 2; }
425 if(idir.z >= 0.0f) { near_z = 4; far_z = 5; } else { near_z = 5; far_z = 4; }
426 tfar = ssef(isect_t);
427 idir4 = sse3f(ssef(idir.x), ssef(idir.y), ssef(idir.z));
428 # ifdef __KERNEL_AVX2__
430 P_idir4 = sse3f(P_idir.x, P_idir.y, P_idir.z);
432 org = sse3f(ssef(P.x), ssef(P.y), ssef(P.z));
434 triangle_intersect_precalc(dir, &isect_precalc);
436 isect_array->t = isect_t;
438 object = OBJECT_NONE;
439 nodeAddr = traversalStack[stackPtr].addr;
442 #endif /* FEATURE(BVH_INSTANCING) */
443 } while(nodeAddr != ENTRYPOINT_SENTINEL);