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

accumulate_wo_tile

def accumulate_wo_tile[micro_kernel_height: Int, micro_kernel_width: Int, simd_size: Int, partial_load: Bool, output_dt: DType, input_dt: DType, filter_dt: DType](c_tile_size: Int, output: Pointer[Scalar[output_dt], _safe=False], output_stride: Int, input: Pointer[Scalar[input_dt], _safe=False], input_stride: Int, filter: Pointer[Scalar[filter_dt], _safe=False], filter_stride: Int, partial_load_size: Int)

Accumulates one width tile of the transposed convolution into the output buffer.

Loads the current output tile into an accumulator, multiplies and accumulates the input and packed filter contributions across the channel tile, then stores the result back to the output buffer.

Parameters:

micro_kernel_height (Int): Height of the micro kernel in the output spatial dimension.
micro_kernel_width (Int): Width of the micro kernel along the output channel dimension.
simd_size (Int): SIMD vector width for the output element type.
partial_load (Bool): Whether the tile contains a residual not aligned to the SIMD width.
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.

Args:

c_tile_size (Int): Size of the channel tile to accumulate.
output (Pointer[Scalar[output_dt], _safe=False]): Pointer to the output tile to accumulate into.
output_stride (Int): Stride between consecutive output rows.
input (Pointer[Scalar[input_dt], _safe=False]): Pointer to the input tile to read from.
input_stride (Int): Stride between consecutive input rows.
filter (Pointer[Scalar[filter_dt], _safe=False]): Pointer to the packed filter tile to read from.
filter_stride (Int): Stride of the filter in the feature dimension.
partial_load_size (Int): Number of valid elements when the tile is a partial load.