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 struct

ConvTransposedPacked

struct ConvTransposedPacked[input_linear_idx_type: DType, input_storage: TensorStorage, filter_linear_idx_type: DType, filter_storage: TensorStorage, output_linear_idx_type: DType, output_storage: TensorStorage, InputLayoutType: TensorLayout, FilterLayoutType: TensorLayout, OutputLayoutType: TensorLayout, conv_attr_rank: Int, //, input_origin: ImmOrigin, filter_origin: ImmOrigin, output_origin: MutOrigin, input_type: DType, filter_type: DType, output_type: DType, conv_attr: ConvInfoStatic[conv_attr_rank], elementwise_epilogue: Optional[def[rank: Int](coords: IndexList[rank], f_size: Int) capturing thin -> None] = None]

Holds the packed tensors and partition state for one task of a direct transposed convolution.

Encapsulates the output, input, and packed filter tile tensors along with the convolution shape and the partition assigned to a single parallel task, and drives the tiled batch, channel, and feature loops that accumulate the result.

Parameters

• ​input_linear_idx_type (DType): Linear index dtype of the input tensor (inferred).
• ​input_storage (TensorStorage): Storage backing the input tile tensor (inferred).
• ​filter_linear_idx_type (DType): Linear index dtype of the filter tensor (inferred).
• ​filter_storage (TensorStorage): Storage backing the filter tile tensor (inferred).
• ​output_linear_idx_type (DType): Linear index dtype of the output tensor (inferred).
• ​output_storage (TensorStorage): Storage backing the output tile tensor (inferred).
• ​InputLayoutType (TensorLayout): Compile-time layout of the input tensor (inferred).
• ​FilterLayoutType (TensorLayout): Compile-time layout of the filter tensor (inferred).
• ​OutputLayoutType (TensorLayout): Compile-time layout of the output tensor (inferred).
• ​conv_attr_rank (Int): Number of spatial dimensions in the convolution (inferred).
• ​input_origin (ImmOrigin): Immutable memory origin of the input tile tensor.
• ​filter_origin (ImmOrigin): Immutable memory origin of the filter tile tensor.
• ​output_origin (MutOrigin): Mutable memory origin of the output tile tensor.
• ​input_type (DType): Element type of the input tensor.
• ​filter_type (DType): Element type of the filter tensor.
• ​output_type (DType): Element type of the output tensor.
• ​conv_attr (ConvInfoStatic[conv_attr_rank]): Static convolution attributes carrying strides, pads, dilations, and group count.
• ​elementwise_epilogue (Optional[def[rank: Int](coords: IndexList[rank], f_size: Int) capturing thin -> None]): Optional elementwise epilogue applied to the output after accumulation (defaults to None).

Fields

• ​output (TileTensor[output_type, OutputLayoutType, output_origin, Storage=output_storage, linear_idx_type=output_linear_idx_type]):
• ​input (TileTensor[input_type, InputLayoutType, input_origin, Storage=input_storage, linear_idx_type=input_linear_idx_type]):
• ​filter (TileTensor[filter_type, FilterLayoutType, filter_origin, Storage=filter_storage, linear_idx_type=filter_linear_idx_type]):
• ​conv_shape (ConvShape[conv_attr_rank]):
• ​partition (ConvPartition):
• ​cf_tile_size (IndexList[Int(2)]):

Implemented traits

AnyType, Copyable, Deinitable, ImplicitlyCopyable, Movable

Methods

run

static def run(output: TileTensor[output_type, OutputLayoutType, output_origin, Storage=output_storage, linear_idx_type=output_linear_idx_type], input: TileTensor[input_type, InputLayoutType, input_origin, Storage=input_storage, linear_idx_type=input_linear_idx_type], filter: TileTensor[filter_type, FilterLayoutType, filter_origin, Storage=filter_storage, linear_idx_type=filter_linear_idx_type], conv_shape: ConvShape[conv_attr_rank], ctx: Optional[DeviceContext] = None)

input_space_loop

def input_space_loop[micro_kernel_height: Int, micro_kernel_width: Int, has_residual: Bool, last_c_tile: Bool](self, n: Int, f_tile_offset: Int, f_tile_size: Int, c_tile_offset: Int, c_tile_size: Int)

input_space_loop_2d

def input_space_loop_2d[output_dt: DType, input_dt: DType, filter_dt: DType, //, micro_kernel_height: Int, micro_kernel_width: Int, has_residual: Bool, last_c_tile: Bool](self, output: Pointer[Scalar[output_dt]], input: Pointer[Scalar[input_dt]], filter: Pointer[Scalar[filter_dt]], n: Int, first_c_tile_in_group: Bool, c_tile_size: Int, f_tile_offset: Int, f_tile_size: Int, left_pad_impact_end: Int, right_pad_impact_start: Int)

input_space_loop_3d

def input_space_loop_3d[micro_kernel_height: Int, micro_kernel_width: Int, has_residual: Bool, last_c_tile: Bool, output_dt: DType, input_dt: DType, filter_dt: DType](self, output: Pointer[Scalar[output_dt]], input: Pointer[Scalar[input_dt]], filter: Pointer[Scalar[filter_dt]], n: Int, first_c_tile_in_group: Bool, c_tile_size: Int, f_tile_offset: Int, f_tile_size: Int, left_pad_impact_end: Int, right_pad_impact_start: Int)

apply_epilogue

def apply_epilogue(self, n: Int, g: Int)