Skip to content

Cone-beam: huge memory consumption when padding != "zeros"

!484 (merged) added support for padding modes other than zeros.

Cone-beam reconstruction acts on 3D data. Generally speaking the full data volume (+ reconstruction) cannot fit the GPU memory, so reconstruction is done by slabs (with an overlap to account for beam divergence).

It looks like astra re-allocates memory, even when using astra.data3d.GPULink. So, doing

cone_data --[FDK_CUDA]--> reconstructed_volume

requires 2X the data volume and 2X the reconstruction volume.

When doing padding by nabu:

cone_data --[nabu:CudaPadding]--> padded_cone_data --[astra:FDK_CUDA]--> reconstructed_volume

it requires (cone_data, padded_cone_data*2, reconstructed_volume*2).

That's huge! Needs to investigate why memory is duplicated.

Update

BP3D_CUDA (used by FDK_CUDA and also with custom filtering) will use textures (so CudaArray) which will duplicate the input data.
Worse, astra expects a sinogram layout in contiguous C-Order (whereas we work on projections layout). So the memory consumption is 3 * input_data + vol_data.

Found a couple of interesting discussion along the way:

Edited by Pierre Paleo