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

Mojo function

create_tensor_tile_im2col

def create_tensor_tile_im2col[dtype: DType, tile_shape: IndexList[Int(2)], swizzle_mode: TensorMapSwizzle = TensorMapSwizzle.SWIZZLE_NONE, *, __tile_shape: IndexList[Int(2)] = tile_shape, __desc_shape: IndexList[Int(2)] = _im2col_desc_shape[dtype, tile_shape, swizzle_mode]()](ctx: DeviceContext, tensor: LayoutTensor[dtype, address_space=tensor.address_space, element_layout=tensor.element_layout, layout_int_type=tensor.layout_int_type, linear_idx_type=tensor.linear_idx_type, masked=tensor.masked, alignment=tensor.alignment], lower_corner_h: Int, lower_corner_w: Int, upper_corner_h: Int, upper_corner_w: Int, out_height: Int, out_width: Int, filter_h: Int, filter_w: Int) -> TMATensorTileIm2col[dtype, Int(2), __tile_shape, __desc_shape]

Creates a TMA tensor tile with im2col transformation for 2D convolution.

This factory function creates a TMA descriptor that performs hardware im2col transformation during loads. The descriptor encodes the convolution geometry and the TMA hardware computes addresses on-the-fly.

For im2col TMA, each transaction loads one output pixel with multiple channels. This follows CUTLASS's approach where:

pixels_per_column = 1 (one pixel per TMA transaction)
channels_per_pixel = min(K_tile, swizzle_width) (contiguous channels)

Note: For stride=1, dilation=1 convolution with padding (following CUTLASS convention):

lower_corner_h = -pad_h
lower_corner_w = -pad_w
upper_corner_h = pad_h - (filter_h - 1)
upper_corner_w = pad_w - (filter_w - 1)

The filter offsets passed to the PTX instruction range from 0 to (filter_size - 1) and are added to lower_corner to compute actual input coordinates.

Parameters:

dtype (DType): The data type of tensor elements.
tile_shape (IndexList[Int(2)]): Shape [M_tile, K_tile] for the GEMM tile.
- M_tile: Number of output pixels (batch * H_out * W_out slice).
- K_tile: Number of channels (C_in * R * S slice for filter).
swizzle_mode (TensorMapSwizzle): Memory swizzling pattern.
__tile_shape (IndexList[Int(2)]): Internal parameter for the tile shape.
__desc_shape (IndexList[Int(2)]): Internal parameter for the descriptor shape.

Args:

ctx (DeviceContext): The CUDA device context.
tensor (LayoutTensor[dtype, address_space=tensor.address_space, element_layout=tensor.element_layout, layout_int_type=tensor.layout_int_type, linear_idx_type=tensor.linear_idx_type, masked=tensor.masked, alignment=tensor.alignment]): The 4D activation tensor in NHWC layout.
lower_corner_h (Int): Lower corner offset for height (negative for padding).
lower_corner_w (Int): Lower corner offset for width (negative for padding).
upper_corner_h (Int): Upper corner offset for height.
upper_corner_w (Int): Upper corner offset for width.
out_height (Int): Output height (H_out) for M coordinate decomposition.
out_width (Int): Output width (W_out) for M coordinate decomposition.
filter_h (Int): Filter height (R) for K coordinate decomposition.
filter_w (Int): Filter width (S) for K coordinate decomposition.

Returns:

TMATensorTileIm2col[dtype, Int(2), __tile_shape, __desc_shape]: A TMATensorTileIm2col configured for im2col loads.

Raises:

Error if TMA descriptor creation fails.

def create_tensor_tile_im2col[dtype: DType, tile_shape: IndexList[Int(2)], swizzle_mode: TensorMapSwizzle = TensorMapSwizzle.SWIZZLE_NONE, *, __tile_shape: IndexList[Int(2)] = tile_shape, __desc_shape: IndexList[Int(2)] = _im2col_desc_shape[dtype, tile_shape, swizzle_mode]()](ctx: DeviceContext, tensor: TileTensor[dtype, Storage=tensor.Storage, address_space=tensor.address_space, linear_idx_type=tensor.linear_idx_type], lower_corner_h: Int, lower_corner_w: Int, upper_corner_h: Int, upper_corner_w: Int, out_height: Int, out_width: Int, filter_h: Int, filter_w: Int) -> TMATensorTileIm2col[dtype, Int(2), __tile_shape, __desc_shape]