Enum TensorMapFormat

pub enum TensorMapFormat {
    Tiled {
        tile_size: Vec<u32>,
    },
    Im2col {
        pixel_box_lower_corner: Vec<i32>,
        pixel_box_upper_corner: Vec<i32>,
        channels_per_pixel: u32,
        pixels_per_column: u32,
    },
    Im2colWide {
        pixel_box_lower_corner_width: i32,
        pixel_box_upper_corner_width: i32,
        channels_per_pixel: u32,
        pixels_per_column: u32,
    },
}

Format of [TensorMap]

Variants

Tiled

Simple tiling

Fields

tile_size: Vec<u32>

Tile size that’s loaded from memory in each copy operation. Must have rank elements. In matmul, for example, this might be batch x m x k, or whatever the stage size is. If a dimension isn’t present in the tile, it should just be set to 1.

For CUDA, this must be a power of two and <= 256 on each dimension.

Im2col

Im2col indexing. Loads a “column” (not the same column as im2col) of pixels into shared memory, with a certain offset (kernel position). The corners are the bounds to load pixels from at offset 0, so the top left corner of the kernel. The offset is added to the corner offsets, so a (-1, -1) corner will stop the bounding box at (1, 1) for kernel offset (2, 2).

Fields

pixel_box_lower_corner: Vec<i32>

Pixel box lower corner. This is the logical upper left corner in the input tensor, when offset is 0. The length of this value should equal the spatial dimensions of the input tensor (i.e. h, w for an NHWC tensor). Should normally be set to -padding.

pixel_box_upper_corner: Vec<i32>

Pixel box top corner. This is the logical lower right corner in the input tensor, when offset is 0. The length of this value should equal the spatial dimensions of the input tensor (i.e. h, w for an NHWC tensor). Should normally be set to padding - kernel_size - 1 (where kernel_size accounts for dilation). This is not equal to padding, it’s equal to the bounding box for the top left corner of the kernel.

channels_per_pixel: u32

Channels to load per pixel, should be a multiple or divisor of the matmul tile size. This is not the total number of channels in the tensor, but only the number loaded in each load. Must be <= 256 and aligned to 16 bytes.

pixels_per_column: u32

Pixels per column, equivalent to the m/n dimension of each tile in the matrix multiplication. i.e. NHW for a 4D tensor. Must be <= 256 and aligned to 16 bytes

Im2colWide

Wide im2col

Fields

pixel_box_lower_corner_width: i32

Pixel box lower corner width. TODO: How does this work?

pixel_box_upper_corner_width: i32

Pixel box upper corner width. TODO: How does this work?

channels_per_pixel: u32

Channels per pixel

pixels_per_column: u32

Pixels per column

Trait Implementations

impl Clone for TensorMapFormat

fn clone(&self) -> TensorMapFormat

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for TensorMapFormat

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for TensorMapFormat

fn deserialize<__D>( __deserializer: __D, ) -> Result<TensorMapFormat, <__D as Deserializer<'de>>::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Hash for TensorMapFormat

fn hash<__H>(&self, state: &mut __H)

where __H: Hasher,

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for TensorMapFormat

fn eq(&self, other: &TensorMapFormat) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Serialize for TensorMapFormat

fn serialize<__S>( &self, __serializer: __S, ) -> Result<<__S as Serializer>::Ok, <__S as Serializer>::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl Eq for TensorMapFormat

impl StructuralPartialEq for TensorMapFormat