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

get_conv_tile_shape

def get_conv_tile_shape[dtype: DType](c: Int, filter_window_size: Int, micro_kernel_width: Int) -> IndexList[Int(2)]

Compute the (c, f) tile shape in L2.

Assume NHWC layout, the tile shape is (R, S, c_tile, f_tile). R and S are by default fully covered. The heuristic tried to block in C as much as possible. If C is small, it would start to block F.

Parameters:

dtype (DType): Element type of the convolution operands, used to derive the target SIMD width and L2 tile size.

Args:

c (Int): Number of input channels, bounding the C tile size.
filter_window_size (Int): Flattened filter window size (the product of the spatial filter dimensions, for example R * S for 2D), divided into the L2 tile budget to size the C and F tiles.
micro_kernel_width (Int): Micro-kernel width in SIMD lanes; multiplied by the SIMD width to give the micro-kernel size in the F dimension, which sets the F tile granularity.

Returns:

IndexList[Int(2)]