fused_fftconv2d_bhl_chunked#

fused_fftconv2d_bhl_chunked(
x,
kernel,
shortcut=None,
chunk_size=None,
fft_size=None,
)#

Channel-chunked fused 2D FFT convolution, BHL layout.

Processes channels in groups of chunk_size to cap the kernel’s working set. The fused kernel’s spatial tile is bounded at 64x64, so its per-call footprint is already small; chunking mainly helps at very large H.

Parameters:

  • x (Tensor) – Input tensor [B, H, X, Y].

  • kernel (Tensor) – Kernel tensor [1|B, H, Kx, Ky].

  • shortcut (Tensor | None) – Optional per-channel scale [H].

  • chunk_size (int | None) – Channels per chunk (default 128).

  • fft_size (int | None) – Optional explicit FFT tile size.

Returns:

Output tensor [B, H, X, Y] in x.dtype.

Return type:

Tensor