Mojo struct
ConvDirectNHWC
struct ConvDirectNHWC[input_mut: Bool, filter_mut: Bool, //, input_rank: Int, filter_rank: Int, output_rank: Int, input_origin: Origin[input_mut], filter_origin: Origin[filter_mut], output_origin: MutableOrigin, input_shape: DimList, filter_shape: DimList, output_shape: DimList, input_type: DType, filter_type: DType, output_type: DType, filter_packed: Bool, conv_attr: ConvInfoStatic[(input_rank + -2)], elementwise_epilogue: OptionalReg[fn[Int](coords: IndexList[$0], f_size: Int) capturing -> None] = OptionalReg[fn[Int](coords: IndexList[$0], f_size: Int) capturing -> None]({:i1 0, 1})]
Implement the outer loops for direct convolution. Collapse N, HO, WO into one dimension n_ho_wo. Tile n_ho_wo, C, and F. The tile factor for C and F are chosen by a heuristic prioritizing C. n_ho_wo is tiled by micro kernel's height.
If n_ho_wo is large enough to spill LLC, we may need to tile n_ho_wo as the outer most loop with a factor fit in LLC.
Assume F is divisible at least by simd_size.
Fields
- output (
NDBuffer[output_type, output_rank, output_origin, output_shape]): - input (
NDBuffer[input_type, input_rank, input_origin, input_shape]): - filter (
NDBuffer[filter_type, filter_rank, filter_origin, filter_shape]): - conv_shape (
ConvShape[(input_rank + -2)]): - partition (
ConvPartition): - cf_tile_size (
IndexList[2]):
Implemented traits
AnyType,
Copyable,
ExplicitlyCopyable,
Movable,
UnknownDestructibility
Aliases
packed_and_fully_static
alias packed_and_fully_static = filter_packed if filter_shape.all_known[::Int]() if output_shape.all_known[::Int,::Int]() if input_shape.all_known[::Int,::Int]() if conv_attr.all_known() else conv_attr.all_known() else input_shape.all_known[::Int,::Int]() if conv_attr.all_known() else conv_attr.all_known() else output_shape.all_known[::Int,::Int]() if input_shape.all_known[::Int,::Int]() if conv_attr.all_known() else conv_attr.all_known() else input_shape.all_known[::Int,::Int]() if conv_attr.all_known() else conv_attr.all_known() else filter_shape.all_known[::Int]() if output_shape.all_known[::Int,::Int]() if input_shape.all_known[::Int,::Int]() if conv_attr.all_known() else conv_attr.all_known() else input_shape.all_known[::Int,::Int]() if conv_attr.all_known() else conv_attr.all_known() else output_shape.all_known[::Int,::Int]() if input_shape.all_known[::Int,::Int]() if conv_attr.all_known() else conv_attr.all_known() else input_shape.all_known[::Int,::Int]() if conv_attr.all_known() else conv_attr.all_known()
Methods
run
static run(output: NDBuffer[output_type, output_rank, output_origin, output_shape], input: NDBuffer[input_type, input_rank, input_origin, input_shape], filter: NDBuffer[filter_type, filter_rank, filter_origin, filter_shape], conv_shape: ConvShape[(input_rank + -2)])
is_new_c_accum
is_new_c_accum(self, c_idx: Int) -> Bool
update_output_tile_no_padding
update_output_tile_no_padding[micro_kernel_height: Int, micro_kernel_width: Int, c_fully_cached: Bool, 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, output_flat_coord: Int)
output_space_flat_loop
output_space_flat_loop[micro_kernel_f_size: 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)
output_space_loop
output_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)
output_space_loop_1d
output_space_loop_1d[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: UnsafePointer[SIMD[output_dt, 1]], input: UnsafePointer[SIMD[input_dt, 1]], filter: UnsafePointer[SIMD[filter_dt, 1]], 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)
output_space_loop_2d
output_space_loop_2d[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: UnsafePointer[SIMD[output_dt, 1]], input: UnsafePointer[SIMD[input_dt, 1]], filter: UnsafePointer[SIMD[filter_dt, 1]], 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)
output_space_loop_3d
output_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: UnsafePointer[SIMD[output_dt, 1]], input: UnsafePointer[SIMD[input_dt, 1]], filter: UnsafePointer[SIMD[filter_dt, 1]], 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)
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!