Some voxels out of bounds

[m-schuetz]

• When mapping float coordinates to quantized voxel coordinates, the coordinate should be clamped to [0, gridSize - 1]
• Code below occasionally maps coordinates to exactly gridSize if a sample is roughly equal to the bounding box maximum.
• Occurs in first-come and random sampling strategies

CudaLOD/modules/simlod/sampling_cuda_nonprogressive/voxelize_sampleselect_first_blockwise.cu

Lines 328 to 330 in cfaa4dc

	int ix = float(bitgrid_size) * (point.x - child->min.x) / childSize.x;
	int iy = float(bitgrid_size) * (point.y - child->min.y) / childSize.y;
	int iz = float(bitgrid_size) * (point.z - child->min.z) / childSize.z;

Should be clamped like this:

int ix = clamp(fGridSize * (point.x - node->min.x) / boxSize.x, 0.0f, fGridSize - 1.0f);
int iy = clamp(fGridSize * (point.y - node->min.y) / boxSize.y, 0.0f, fGridSize - 1.0f);
int iz = clamp(fGridSize * (point.z - node->min.z) / boxSize.z, 0.0f, fGridSize - 1.0f);

Should probably also be done in similar sections in average and weighted strategies? e.g.

CudaLOD/modules/simlod/sampling_cuda_nonprogressive/voxelize_neighborhood_blockwise.cu

Lines 175 to 177 in cfaa4dc

	float fx = fGridSize * (point.x - node->min.x) / boxSize.x;
	float fy = fGridSize * (point.y - node->min.y) / boxSize.y;
	float fz = fGridSize * (point.z - node->min.z) / boxSize.z;

[m-schuetz]

There is also a small amount of voxels in the first processed inner node that are overwritten with other data due to mistakenly overlapping buffer allocations.

At the end of the voxelization, the allocator offset should be updated to after the last allocated voxel, like first-come and random sampling already do:

CudaLOD/modules/simlod/sampling_cuda_nonprogressive/voxelize_sampleselect_first_blockwise.cu

Line 390 in cfaa4dc

allocator.offset = globalAllocatorOffset;

This update of the allocator is missing in the average and weighted average strategies, at this place:

CudaLOD/modules/simlod/sampling_cuda_nonprogressive/voxelize_neighborhood_blockwise.cu

Line 473 in cfaa4dc

The problem arises because kernel3 directly afterwards allocates some memory to compute stats, but because the allocator wasn't updated, that memory ends up pointing to the voxels of the first processed node:

CudaLOD/modules/simlod/sampling_cuda_nonprogressive/kernel.cu

Lines 303 to 315 in cfaa4dc

	uint32_t& stat_num_leaves = *allocator.alloc<uint32_t*>(4);
	uint32_t& stat_num_inner = *allocator.alloc<uint32_t*>(4);
	uint64_t& stat_sum_leaves = *allocator.alloc<uint64_t*>(8);
	uint64_t& stat_sum_inner = *allocator.alloc<uint64_t*>(8);
	uint64_t& stat_max_leaves = *allocator.alloc<uint64_t*>(8);
	uint64_t& stat_max_inner = *allocator.alloc<uint64_t*>(8);
	uint64_t& stat_min_leaves = *allocator.alloc<uint64_t*>(8);
	uint64_t& stat_min_inner = *allocator.alloc<uint64_t*>(8);
	uint64_t* pointsAtLevel = allocator.alloc<uint64_t*>(50 * 8);
	uint64_t* voxelsAtLevel = allocator.alloc<uint64_t*>(50 * 8);
	uint64_t* innerAtLevel = allocator.alloc<uint64_t*>(50 * 8);
	uint64_t* leavesAtLevel = allocator.alloc<uint64_t*>(50 * 8);

And this is what it looks like. Should be about 100-150 voxels that end up garbage due do the amount of mem that computing stats allocates.