For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

conv2d_update_wo_tile

def conv2d_update_wo_tile[micro_kernel_height: Int, micro_kernel_width: Int, simd_size: Int, filter_packed: Bool, effected_by_padding: Bool, has_residual: Bool, last_c_tile: Bool, output_dt: DType, input_dt: DType, filter_dt: DType, elementwise_epilogue: Optional[def[rank: Int](coords: IndexList[rank], f_size: Int) capturing thin -> None] = None](output: Pointer[Scalar[output_dt], address_space=output.address_space, _safe=False], input: Pointer[Scalar[input_dt], address_space=input.address_space, _safe=False], filter: Pointer[Scalar[filter_dt], address_space=filter.address_space, _safe=False], first_c_tile: Bool, c_tile_size: Int, f_tile_offset: Int, f_tile_size: Int, conv_shape: ConvShape[Int(2)], n: Int, howo: IndexList[Int(2)])

Updates one micro tile of the 2D convolution output for a given (c, f) tile, accumulating over the R x S filter window and optionally applying an elementwise epilogue on the last C tile.

Parameters:

micro_kernel_height (Int): Number of output points along the WO dimension covered by the micro tile in register tiling.
micro_kernel_width (Int): Number of SIMD registers assigned to the F dimension per output point.
simd_size (Int): Number of elements in a SIMD register.
filter_packed (Bool): True when the filter is prepacked in FRSCf layout for grouped convolution.
effected_by_padding (Bool): True when the tile may touch padded input regions, requiring per-point bounds checks.
has_residual (Bool): True when F is not a multiple of simd_size. The residual elements are loaded and padded with zero to fit a simd vector.
last_c_tile (Bool): True when this is the last C tile, enabling the elementwise epilogue after accumulation.
output_dt (DType): Element type of the output tensor.
input_dt (DType): Element type of the input tensor.
filter_dt (DType): Element type of the filter tensor.
elementwise_epilogue (Optional[def[rank: Int](coords: IndexList[rank], f_size: Int) capturing thin -> None]): Optional elementwise function applied to the output after the last channel tile (defaults to None).

Args:

output (Pointer[Scalar[output_dt], address_space=output.address_space, _safe=False]): Pointer to the start of the output micro tile at (n, howo[0], howo[1], f_tile_offset).
input (Pointer[Scalar[input_dt], address_space=input.address_space, _safe=False]): Pointer to the first input element of the micro tile before padding adjustment.
filter (Pointer[Scalar[filter_dt], address_space=filter.address_space, _safe=False]): Pointer to the first filter coefficient in the filter window for the current (c, f) tile.
first_c_tile (Bool): True when this is the first C tile, zero-initializing the accumulator instead of loading from the output.
c_tile_size (Int): Number of input channels in the current C tile.
f_tile_offset (Int): Offset of the current tile along the F (output channel) dimension.
f_tile_size (Int): Number of output channels in the current F tile.
conv_shape (ConvShape[Int(2)]): Convolution dimension description for the 2D convolution.
n (Int): Batch index of the current input image.
howo (IndexList[Int(2)]): Output spatial coordinates (ho, wo) of the tile's first output point, with the micro tile spanning micro_kernel_height consecutive wo values starting at howo[1].