View page source

warp.tile_fft#

warp.tile_fft(
inout: Tile[Vector[Float, Literal[2]], tuple[int, ...]],
) → None#

  • Kernel

  • Differentiable

Compute the forward FFT along the last dimension of an N-D tile of data.

This function cooperatively computes the forward FFT on a tile of data inplace. All leading dimensions are treated as independent batch dimensions. The tile must have at least two dimensions.

The transform is unnormalized, meaning that applying tile_fft() followed by tile_ifft() will scale the data by N, where N is the FFT size (the last dimension of the tile). Normalization is left to the user to perform as needed.

Supported datatypes are:

  • vec2f, vec2d

Parameters:

inout – The input/output tile.

Notes

Supported FFT sizes by backend:

  • CPU: Any size. Non-power-of-two sizes are capped at 4096; larger non-power-of-two sizes raise ValueError.

  • GPU with libmathdx: Any size. This is the default when Warp is built with libmathdx.

  • GPU without libmathdx (or enable_mathdx_fft=False): Power-of-two sizes only, and the FFT size must be divisible by block_dim. Other sizes raise ValueError. Slower than the libmathdx path.

See warp.config.enable_mathdx_fft to control GPU backend selection.