Cycles: BVH build time optimizations:
* BVH building was multithreaded. Not all building is multithreaded, packing
and the initial bounding/splitting is still single threaded, but recursive
splitting is, which was the main bottleneck.
* Object splitting now uses binning rather than sorting of all elements, using
code from the Embree raytracer from Intel.
http://software.intel.com/en-us/articles/embree-photo-realistic-ray-tracing-kernels/
* Other small changes to avoid allocations, pack memory more tightly, avoid
some unnecessary operations, ...
These optimizations do not work yet when Spatial Splits are enabled, for that
more work is needed. There's also other optimizations still needed, in
particular for the case of many low poly objects, the packing step and node
memory allocation.
BVH raytracing time should remain about the same, but BVH build time should be
significantly reduced, test here show speedup of about 5x to 10x on a dual core
and 5x to 25x on an 8-core machine, depending on the scene.
2012-04-24 11:52:58 +00:00
|
|
|
/*
|
|
|
|
* Adapted from code copyright 2009-2011 Intel Corporation
|
|
|
|
* Modifications Copyright 2012, Blender Foundation.
|
|
|
|
*
|
|
|
|
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
|
|
* you may not use this file except in compliance with the License.
|
|
|
|
* You may obtain a copy of the License at
|
|
|
|
*
|
|
|
|
* http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
*
|
|
|
|
* Unless required by applicable law or agreed to in writing, software
|
|
|
|
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
|
|
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
|
|
* See the License for the specific language governing permissions and
|
|
|
|
* limitations under the License.
|
|
|
|
*/
|
|
|
|
|
2012-04-24 17:56:27 +00:00
|
|
|
#include <stdlib.h>
|
|
|
|
|
Cycles: BVH build time optimizations:
* BVH building was multithreaded. Not all building is multithreaded, packing
and the initial bounding/splitting is still single threaded, but recursive
splitting is, which was the main bottleneck.
* Object splitting now uses binning rather than sorting of all elements, using
code from the Embree raytracer from Intel.
http://software.intel.com/en-us/articles/embree-photo-realistic-ray-tracing-kernels/
* Other small changes to avoid allocations, pack memory more tightly, avoid
some unnecessary operations, ...
These optimizations do not work yet when Spatial Splits are enabled, for that
more work is needed. There's also other optimizations still needed, in
particular for the case of many low poly objects, the packing step and node
memory allocation.
BVH raytracing time should remain about the same, but BVH build time should be
significantly reduced, test here show speedup of about 5x to 10x on a dual core
and 5x to 25x on an 8-core machine, depending on the scene.
2012-04-24 11:52:58 +00:00
|
|
|
#include "bvh_binning.h"
|
|
|
|
|
|
|
|
#include "util_algorithm.h"
|
|
|
|
#include "util_boundbox.h"
|
|
|
|
#include "util_types.h"
|
|
|
|
|
|
|
|
CCL_NAMESPACE_BEGIN
|
|
|
|
|
|
|
|
/* SSE replacements */
|
|
|
|
|
|
|
|
__forceinline void prefetch_L1 (const void* ptr) { }
|
|
|
|
__forceinline void prefetch_L2 (const void* ptr) { }
|
|
|
|
__forceinline void prefetch_L3 (const void* ptr) { }
|
|
|
|
__forceinline void prefetch_NTA(const void* ptr) { }
|
|
|
|
|
|
|
|
template<size_t src> __forceinline float extract(const int4& b)
|
|
|
|
{ return b[src]; }
|
|
|
|
template<size_t dst> __forceinline const float4 insert(const float4& a, const float b)
|
|
|
|
{ float4 r = a; r[dst] = b; return r; }
|
|
|
|
|
|
|
|
__forceinline int get_best_dimension(const float4& bestSAH)
|
|
|
|
{
|
|
|
|
// return (int)__bsf(movemask(reduce_min(bestSAH) == bestSAH));
|
|
|
|
|
|
|
|
float minSAH = min(bestSAH.x, min(bestSAH.y, bestSAH.z));
|
|
|
|
|
|
|
|
if(bestSAH.x == minSAH) return 0;
|
|
|
|
else if(bestSAH.y == minSAH) return 1;
|
|
|
|
else return 2;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* BVH Object Binning */
|
|
|
|
|
|
|
|
BVHObjectBinning::BVHObjectBinning(const BVHRange& job, BVHReference *prims)
|
|
|
|
: BVHRange(job), splitSAH(FLT_MAX), dim(0), pos(0)
|
|
|
|
{
|
|
|
|
/* compute number of bins to use and precompute scaling factor for binning */
|
|
|
|
num_bins = min(size_t(MAX_BINS), size_t(4.0f + 0.05f*size()));
|
|
|
|
scale = rcp(cent_bounds().size()) * make_float3((float)num_bins);
|
|
|
|
|
|
|
|
/* initialize binning counter and bounds */
|
|
|
|
BoundBox bin_bounds[MAX_BINS][4]; /* bounds for every bin in every dimension */
|
|
|
|
int4 bin_count[MAX_BINS]; /* number of primitives mapped to bin */
|
|
|
|
|
|
|
|
for(size_t i = 0; i < num_bins; i++) {
|
|
|
|
bin_count[i] = make_int4(0);
|
|
|
|
bin_bounds[i][0] = bin_bounds[i][1] = bin_bounds[i][2] = BoundBox::empty;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* map geometry to bins, unrolled once */
|
|
|
|
{
|
|
|
|
ssize_t i;
|
|
|
|
|
|
|
|
for(i = 0; i < ssize_t(size()) - 1; i += 2) {
|
|
|
|
prefetch_L2(&prims[start() + i + 8]);
|
|
|
|
|
|
|
|
/* map even and odd primitive to bin */
|
|
|
|
BVHReference prim0 = prims[start() + i + 0];
|
|
|
|
BVHReference prim1 = prims[start() + i + 1];
|
|
|
|
|
|
|
|
int4 bin0 = get_bin(prim0.bounds());
|
|
|
|
int4 bin1 = get_bin(prim1.bounds());
|
|
|
|
|
|
|
|
/* increase bounds for bins for even primitive */
|
|
|
|
int b00 = extract<0>(bin0); bin_count[b00][0]++; bin_bounds[b00][0].grow(prim0.bounds());
|
|
|
|
int b01 = extract<1>(bin0); bin_count[b01][1]++; bin_bounds[b01][1].grow(prim0.bounds());
|
|
|
|
int b02 = extract<2>(bin0); bin_count[b02][2]++; bin_bounds[b02][2].grow(prim0.bounds());
|
|
|
|
|
|
|
|
/* increase bounds of bins for odd primitive */
|
|
|
|
int b10 = extract<0>(bin1); bin_count[b10][0]++; bin_bounds[b10][0].grow(prim1.bounds());
|
|
|
|
int b11 = extract<1>(bin1); bin_count[b11][1]++; bin_bounds[b11][1].grow(prim1.bounds());
|
|
|
|
int b12 = extract<2>(bin1); bin_count[b12][2]++; bin_bounds[b12][2].grow(prim1.bounds());
|
|
|
|
}
|
|
|
|
|
|
|
|
/* for uneven number of primitives */
|
|
|
|
if(i < ssize_t(size())) {
|
|
|
|
/* map primitive to bin */
|
|
|
|
BVHReference prim0 = prims[start() + i];
|
|
|
|
int4 bin0 = get_bin(prim0.bounds());
|
|
|
|
|
|
|
|
/* increase bounds of bins */
|
|
|
|
int b00 = extract<0>(bin0); bin_count[b00][0]++; bin_bounds[b00][0].grow(prim0.bounds());
|
|
|
|
int b01 = extract<1>(bin0); bin_count[b01][1]++; bin_bounds[b01][1].grow(prim0.bounds());
|
|
|
|
int b02 = extract<2>(bin0); bin_count[b02][2]++; bin_bounds[b02][2].grow(prim0.bounds());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* sweep from right to left and compute parallel prefix of merged bounds */
|
|
|
|
float4 r_area[MAX_BINS]; /* area of bounds of primitives on the right */
|
|
|
|
float4 r_count[MAX_BINS]; /* number of primitives on the right */
|
|
|
|
int4 count = make_int4(0);
|
|
|
|
|
|
|
|
BoundBox bx = BoundBox::empty;
|
|
|
|
BoundBox by = BoundBox::empty;
|
|
|
|
BoundBox bz = BoundBox::empty;
|
|
|
|
|
|
|
|
for(size_t i = num_bins - 1; i > 0; i--) {
|
|
|
|
count = count + bin_count[i];
|
|
|
|
r_count[i] = blocks(count);
|
|
|
|
|
|
|
|
bx = merge(bx,bin_bounds[i][0]); r_area[i][0] = bx.half_area();
|
|
|
|
by = merge(by,bin_bounds[i][1]); r_area[i][1] = by.half_area();
|
|
|
|
bz = merge(bz,bin_bounds[i][2]); r_area[i][2] = bz.half_area();
|
2012-05-04 16:38:11 +00:00
|
|
|
r_area[i][3] = r_area[i][2];
|
Cycles: BVH build time optimizations:
* BVH building was multithreaded. Not all building is multithreaded, packing
and the initial bounding/splitting is still single threaded, but recursive
splitting is, which was the main bottleneck.
* Object splitting now uses binning rather than sorting of all elements, using
code from the Embree raytracer from Intel.
http://software.intel.com/en-us/articles/embree-photo-realistic-ray-tracing-kernels/
* Other small changes to avoid allocations, pack memory more tightly, avoid
some unnecessary operations, ...
These optimizations do not work yet when Spatial Splits are enabled, for that
more work is needed. There's also other optimizations still needed, in
particular for the case of many low poly objects, the packing step and node
memory allocation.
BVH raytracing time should remain about the same, but BVH build time should be
significantly reduced, test here show speedup of about 5x to 10x on a dual core
and 5x to 25x on an 8-core machine, depending on the scene.
2012-04-24 11:52:58 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* sweep from left to right and compute SAH */
|
|
|
|
int4 ii = make_int4(1);
|
|
|
|
float4 bestSAH = make_float4(FLT_MAX);
|
|
|
|
int4 bestSplit = make_int4(-1);
|
|
|
|
|
|
|
|
count = make_int4(0);
|
|
|
|
|
|
|
|
bx = BoundBox::empty;
|
|
|
|
by = BoundBox::empty;
|
|
|
|
bz = BoundBox::empty;
|
|
|
|
|
|
|
|
for(size_t i = 1; i < num_bins; i++, ii += make_int4(1)) {
|
|
|
|
count = count + bin_count[i-1];
|
|
|
|
|
|
|
|
bx = merge(bx,bin_bounds[i-1][0]); float Ax = bx.half_area();
|
|
|
|
by = merge(by,bin_bounds[i-1][1]); float Ay = by.half_area();
|
|
|
|
bz = merge(bz,bin_bounds[i-1][2]); float Az = bz.half_area();
|
|
|
|
|
|
|
|
float4 lCount = blocks(count);
|
|
|
|
float4 lArea = make_float4(Ax,Ay,Az,Az);
|
|
|
|
float4 sah = lArea*lCount + r_area[i]*r_count[i];
|
|
|
|
|
|
|
|
bestSplit = select(sah < bestSAH,ii,bestSplit);
|
|
|
|
bestSAH = min(sah,bestSAH);
|
|
|
|
}
|
|
|
|
|
|
|
|
int4 mask = float3_to_float4(cent_bounds().size()) <= make_float4(0.0f);
|
|
|
|
bestSAH = insert<3>(select(mask, make_float4(FLT_MAX), bestSAH), FLT_MAX);
|
|
|
|
|
|
|
|
/* find best dimension */
|
|
|
|
dim = get_best_dimension(bestSAH);
|
|
|
|
splitSAH = bestSAH[dim];
|
|
|
|
pos = bestSplit[dim];
|
|
|
|
leafSAH = bounds().half_area() * blocks(size());
|
|
|
|
}
|
|
|
|
|
|
|
|
void BVHObjectBinning::split(BVHReference* prims, BVHObjectBinning& left_o, BVHObjectBinning& right_o) const
|
|
|
|
{
|
|
|
|
size_t N = size();
|
|
|
|
|
|
|
|
BoundBox lgeom_bounds = BoundBox::empty;
|
|
|
|
BoundBox rgeom_bounds = BoundBox::empty;
|
|
|
|
BoundBox lcent_bounds = BoundBox::empty;
|
|
|
|
BoundBox rcent_bounds = BoundBox::empty;
|
|
|
|
|
|
|
|
ssize_t l = 0, r = N-1;
|
|
|
|
|
|
|
|
while(l <= r) {
|
|
|
|
prefetch_L2(&prims[start() + l + 8]);
|
|
|
|
prefetch_L2(&prims[start() + r - 8]);
|
|
|
|
|
|
|
|
BVHReference prim = prims[start() + l];
|
|
|
|
float3 center = prim.bounds().center2();
|
|
|
|
|
|
|
|
if(get_bin(center)[dim] < pos) {
|
|
|
|
lgeom_bounds.grow(prim.bounds());
|
|
|
|
lcent_bounds.grow(center);
|
|
|
|
l++;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
rgeom_bounds.grow(prim.bounds());
|
|
|
|
rcent_bounds.grow(center);
|
|
|
|
swap(prims[start()+l],prims[start()+r]);
|
|
|
|
r--;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* finish */
|
|
|
|
if(l != 0 && N-1-r != 0) {
|
|
|
|
right_o = BVHObjectBinning(BVHRange(rgeom_bounds, rcent_bounds, start() + l, N-1-r), prims);
|
|
|
|
left_o = BVHObjectBinning(BVHRange(lgeom_bounds, lcent_bounds, start(), l), prims);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* object medium split if we did not make progress, can happen when all
|
|
|
|
primitives have same centroid */
|
|
|
|
lgeom_bounds = BoundBox::empty;
|
|
|
|
rgeom_bounds = BoundBox::empty;
|
|
|
|
lcent_bounds = BoundBox::empty;
|
|
|
|
rcent_bounds = BoundBox::empty;
|
|
|
|
|
|
|
|
for(size_t i = 0; i < N/2; i++) {
|
|
|
|
lgeom_bounds.grow(prims[start()+i].bounds());
|
|
|
|
lcent_bounds.grow(prims[start()+i].bounds().center2());
|
|
|
|
}
|
|
|
|
|
|
|
|
for(size_t i = N/2; i < N; i++) {
|
|
|
|
rgeom_bounds.grow(prims[start()+i].bounds());
|
|
|
|
rcent_bounds.grow(prims[start()+i].bounds().center2());
|
|
|
|
}
|
|
|
|
|
|
|
|
right_o = BVHObjectBinning(BVHRange(rgeom_bounds, rcent_bounds, start() + N/2, N/2 + N%2), prims);
|
|
|
|
left_o = BVHObjectBinning(BVHRange(lgeom_bounds, lcent_bounds, start(), N/2), prims);
|
|
|
|
}
|
|
|
|
|
|
|
|
CCL_NAMESPACE_END
|
|
|
|
|