sxyu/svox

PlenOctrees construction + rendering PyTorch CUDA extension

Website: https://alexyu.net

Discord: @doriath0

svox's Issues

How to transform the sampling point?

Hi,
At first, thanks for the great packages!

I'd like to render a image following curved rays (not straight rays)
It seems I should handle the .cu file to implement this.

svox/svox/csrc/rt_kernel.cu

Lines 799 to 819 in 0f300d7

	template <typename scalar_t>
	__global__ void render_image_kernel(
	PackedTreeSpec<scalar_t> tree,
	PackedCameraSpec<scalar_t> cam,
	RenderOptions opt,
	torch::PackedTensorAccessor32<scalar_t, 3, torch::RestrictPtrTraits>
	out) {
	CUDA_GET_THREAD_ID(tid, cam.width * cam.height);
	int iy = tid / cam.width, ix = tid % cam.width;
	scalar_t dir[3], origin[3];
	cam2world_ray(ix, iy, dir, origin, cam);
	scalar_t vdir[3] = {dir[0], dir[1], dir[2]};
	maybe_world2ndc(opt, dir, origin);
	transform_coord<scalar_t>(origin, tree.offset, tree.scaling);
	trace_ray<scalar_t>(
	tree,
	SingleRaySpec<scalar_t>{origin, dir, vdir},
	opt,
	out[iy][ix]);
	}

svox/svox/csrc/rt_kernel.cu

Lines 775 to 796 in 0f300d7

	template <typename scalar_t>
	__host__ __device__ __inline__ static void maybe_world2ndc(
	RenderOptions& __restrict__ opt,
	scalar_t* __restrict__ dir,
	scalar_t* __restrict__ cen, scalar_t near = 1.f) {
	if (opt.ndc_width < 0)
	return;
	scalar_t t = -(near + cen[2]) / dir[2];
	for (int i = 0; i < 3; ++i) {
	cen[i] = cen[i] + t * dir[i];
	}
	dir[0] = -((2 * opt.ndc_focal) / opt.ndc_width) * (dir[0] / dir[2] - cen[0] / cen[2]);
	dir[1] = -((2 * opt.ndc_focal) / opt.ndc_height) * (dir[1] / dir[2] - cen[1] / cen[2]);
	dir[2] = -2 * near / cen[2];
	cen[0] = -((2 * opt.ndc_focal) / opt.ndc_width) * (cen[0] / cen[2]);
	cen[1] = -((2 * opt.ndc_focal) / opt.ndc_height) * (cen[1] / cen[2]);
	cen[2] = 1 + 2 * near / cen[2];
	_normalize(dir);
	}

Unfortunately, I barely handle the CUDA.
Is there any way to manipulate the ray (or sampling points)?
Any glitch or niche is also fine. :)

Thank you for your time!

How to use cpu to render?

Hello, may I ask how to use svox to render an image on a cpu? It seems that svox does not support cpu rendering at this moment.

Inconsistent Querying Results of 'cpu' and 'cuda' in N3Tree

K = 16
tree = svox.N3Tree(data_dim=4,
                   radius=K,
                   center=[K, K, K],
                   data_format='RGBA',
                   device='cpu') # or with 'cuda'
print(tree[20,20,20].corners)

In 'cpu' mode it outputs [16., 16., 16.] as expected, however in 'cuda' mode, it outputs [16., 0., 0.]

svox.csrc

CUDA extension svox.csrc could not be loaded! Operations will be slow.
Please do not import svox in the SVOX source directory.
"Operations will be slow.\n"

I installed svox with "pip install svox", but there will be such an error. Is there any solution? thank you very much!

Unable to load N3Tree while on running server listening on certain port

While I was trying to run the renderer command using svox to load the trained tree.npz on the server, which is keep listening on certain port.
The loading command seems to never return.
Wonder what might be the cause of the problem.

The low volume rendering seems not to work.

发生异常: RuntimeError (note: full exception trace is shown but execution is paused at: _run_module_as_main)
The size of tensor a (3) must match the size of tensor b (640000) at non-singleton dimension 1

UserWarning: Using slow volume rendering, should only be used for debugging
warn("Using slow volume rendering, should only be used for debugging")

out_rgb += light_intensity. * self.background_brightness