To support warp-level short-circuiting of evaluation of gaussians on distant pixels, we want spatially compact pixel regions to be contiguous in memory. We might do this by ordering the flattened pixel arrays in "super-pixel" order, where a super-pixel is a block of 4x8 or 8x8 (some multiple of the warp size).
This will introduce new padding elements along each dimension, which can be indicated with ierr = 0. Some care will be needed in unpacking the flattened, padded residual images back into their original sizes and shapes.
To support warp-level short-circuiting of evaluation of gaussians on distant pixels, we want spatially compact pixel regions to be contiguous in memory. We might do this by ordering the flattened pixel arrays in "super-pixel" order, where a super-pixel is a block of 4x8 or 8x8 (some multiple of the warp size).
This will introduce new padding elements along each dimension, which can be indicated with
ierr = 0. Some care will be needed in unpacking the flattened, padded residual images back into their original sizes and shapes.