pub fn parse_stream_parallel<N>(
input: &[u8],
sink: &N,
) -> Result<(), Error<N::Error>>Expand description
Read an ɴsɪ stream on every core, applying it to sink.
For a sink that accepts calls from many threads at once – a
renderer’s context. The resulting scene is the one parse_stream
produces, but the calls arrive in a different order:
- every
Create, concurrently, in stream order per handle; - then every
SetAttributeandSetAttributeAtTime, concurrently, in stream order per handle, and everyConnect, concurrently, in stream order per destination attribute.
Delete, DeleteAttribute, Disconnect, Evaluate and
RenderControl are barriers: everything before one is applied
before it, it is applied alone, and nothing after it is applied
before it. So an interactive stream – edits, a RenderControl,
more edits – keeps its meaning.
§What it buys
Measured on 20 000 meshes, 5.3 MiB, 16 threads: into a sink that does
nothing, 166 MiB/s becomes about 320. Into 3Delight 2.9, only
1.15-1.4x: its context serialises calls internally – 20 000
SetAttributes take 11.8 ms from one thread and 21.4 ms from
sixteen – so what is gained there is parsing overlapped with the
renderer, not calls made side by side. A sink that takes calls
concurrently gains the most.
§Differences from parse_stream
- A reference to a node created later succeeds. Every
Createin a segment runs first, so aSetAttributeon a node the stream creates further down finds it, where the sequential parser fails. - On failure, which statements were applied is unspecified.
Statements that do not depend on the failing one may have been
applied, before or after it in the stream. The error returned is
the earliest failure found, and a stream with a single error
reports the same error as
parse_stream.
§Errors
As parse_stream.