TrilinearDevoxelizeKernel.h File Reference

(1e7b174 (Tue Jan 7 15:20:18 2025 -0800))

Go to the source code of this file.

Functions
void	TrilinearDevoxelize (int b, int c, int n, int r, int r2, int r3, bool is_training, const float *coords, const float *feat, int *inds, float *wgts, float *outs)
void	TrilinearDevoxelizeGrad (int b, int c, int n, int r3, const int *inds, const float *wgts, const float *grad_y, float *grad_x)

Function Documentation

TrilinearDevoxelize()

void TrilinearDevoxelize	(	int	b,
		int	c,
		int	n,
		int	r,
		int	r2,
		int	r3,
		bool	is_training,
		const float *	coords,
		const float *	feat,
		int *	inds,
		float *	wgts,
		float *	outs
	)

This function performs trilinear devoxelization operation. It computes aggregated features from the voxel grid for each point passed in the input.

Parameters

b	The batch size.
c	Feature dimension of voxel grid.
n	Number of points per batch.
r	Resolution of the grid.
r2	r squared.
r3	r cubed.
is_training	Whether model is in training phase.
coords	Array with the point positions. The shape is [b, 3, n]
feat	Array with the voxel grid. The shape is [b, c, r, r, r]
inds	The voxel coordinates of point cube [b, 8, n]
wgts	weight for trilinear interpolation [b, 8, n]
outs	Outputs, FloatTensor[b, c, n]

TrilinearDevoxelizeGrad()

void TrilinearDevoxelizeGrad	(	int	b,
		int	c,
		int	n,
		int	r3,
		const int *	inds,
		const float *	wgts,
		const float *	grad_y,
		float *	grad_x
	)

This function computes gradient for trilinear devoxelization op. It computes gradient for the input voxelgrid.

Parameters

b	The batch size.
c	Feature dimension of voxel grid.
n	Number of points per batch.
r3	resolution cubed.
inds	The voxel coordinates of point cube [b, 8, n]
wgts	weight for trilinear interpolation [b, 8, n]
grad_y	The gradient passed from top.
grad_x	The computed gradient for voxelgrid.