Skip to content

Filter 2D Sync Task

The Filter 2D sync task constructs and applies a radial frequency-domain filter to complex (CF32) images on the GPU. The task builds a per-frame complex filter (CUFFT callback) from the provided radii and smoothing parameters, shifts corners (FFT quadrant swap), and uses an FFT-based pipeline (forward FFT → multiply by filter → inverse FFT) to filter each input frame.

Supported filter behaviors (depending on how r_inner, r_outer, s_inner, and s_outer are selected):

  • Low-pass: choose r_outer small and r_inner == 0 to preserve low spatial frequencies and attenuate high frequencies.
  • High-pass: choose r_inner large and r_outer small (or zero) so low frequencies are suppressed and high frequencies preserved.
  • Band-pass: choose r_inner < r_outer to keep a ring of spatial frequencies between the inner and outer radii.
  • Band-stop: choose r_inner < r_outer and invert interpretation (the implemented filter produces a ring-shaped pass region; parameter choices can emulate stop behavior by post-processing).

The filter value at each frequency location is computed from r_inner, r_outer and their corresponding smoothing widths s_inner/s_outer. The smoothing parameters produce cosine transitions over the smoothing width so the filter transitions smoothly between 0 and 1 rather than abruptly.

Warning

Very large radius or smoothing values relative to the image size may produce degenerate filters. Ensure your radii/smoothing parameters are chosen with respect to the frame dimensions to obtain the intended frequency response.

Inputs

This task consumes a single input tensor.

The input requirements (as enforced by Filter2DFactory::infer) are:

  • The tensor must be 3D with shape (B, H, W) (batch, height, width).
  • The dtype must be complex32 (CF32).
  • The tensor must reside in device (GPU) memory.

Outputs

This task produces one output tensor with the same shape and dtype as the input (the filter is applied in-place relative to shape/dtype but the implementation may use intermediate buffers).

The memory location of the output tensor is device memory.

Inplace

This task supports in-place operation for the tensor shapes used (the infer result returns an in-place pair), however the implementation constructs FFT plans and may use device buffers; the net effect is that the output has the same shape and dtype as the input.

Ownership

This task does not own any inputs or outputs.

Settings

Example:

1
2
3
4
5
6
{
    "r_inner": 10,
    "r_outer": 100,
    "s_inner": 5,
    "s_outer": 10
}
Property Pattern Type Deprecated Definition Title/Description
+ r_inner No integer No - Inner radius in pixels (non-negative). Controls the inner boundary of the spatial frequency filter.
+ r_outer No integer No - Outer radius in pixels (non-negative). Controls the outer boundary of the spatial frequency filter.
+ s_inner No integer No - Inner slope (smoothness) in pixels (non-negative). Defines the smoothing width around the inner radius.
+ s_outer No integer No - Outer slope (smoothness) in pixels (non-negative). Defines the smoothing width around the outer radius.