`ptwm_core::transforms::source`

Source op — graph entry node that describes the raw input tensor.

rsmod

ptwm_core::transforms::source

crates/ptwm-core/src/transforms/source.rs:1

mod source

`Source` op — graph entry node that describes the raw input tensor.

rsfn

ptwm_core::transforms::source::element_width_for

crates/ptwm-core/src/transforms/source.rs:18

pub fn element_width_for(dtype_code: u16) -> ElementWidth

Element width inferred from a chain-internal dtype code. The chain's Source params use a different code space than `Dtype::from_code` (the public registry); see `_CHAIN_DTYPE` in `python/ptwm/preprocessing/_chains.py`. This is the one owner of the mapping: `compressor::source_descriptor_for` calls it rather than restating it.

rsfn

ptwm_core::transforms::source::is_nibble_packed_dtype

crates/ptwm-core/src/transforms/source.rs:39

pub fn is_nibble_packed_dtype(dtype_code: u16) -> bool

Whether a chain-internal dtype stores two values per byte.

Kept separate from element width on purpose: nibble width says how wide an element is, this says that elements share a byte. Codecs that read packed nibbles (`PerGroupCodebook`) require both, and transforms that cannot handle sharing (`BurrowsWheeler`, `IndexPack`) reject on this one.

rsconst

ptwm_core::transforms::source::MAX_SOURCE_DIMS

crates/ptwm-core/src/transforms/source.rs:76

pub const MAX_SOURCE_DIMS

Maximum number of dimensions the wire format supports. Encoded as a single `u8` length prefix in `write_params`.

rsfn

ptwm_core::transforms::source::read_source_params

crates/ptwm-core/src/transforms/source.rs:177

pub fn read_source_params(buf: &[u8]) -> Result<(Source, usize), PtwmCoreError>

Parse `Source` params from a raw byte buffer. Returns `(Source, bytes_consumed)`.

rstype

ptwm_core::transforms::source::Source

crates/ptwm-core/src/transforms/source.rs:69

pub struct Source

Mandatory entry node of every PPG chain. Takes no inputs (the runtime injects raw tensor bytes externally) and emits one `Role::Raw` plane whose descriptor follows from the tensor `shape` and `dtype_code`.

### forward / inverse semantics The runtime wires tensor bytes into the graph, so `forward` expects a **single synthetic input** carrying the raw tensor bytes (the runtime wraps them in a `Plane` before calling). It validates the descriptor and forwards the plane unchanged. `inverse` is symmetric: takes one plane, returns it unchanged.