substrate-data 0.3.1

One data model, text syntax, binary form, packed frames and schema system — for any program's files, saves and wire messages.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32

//! Numbers bit-packed in lanes: what a frame's list of ints is made of.
//!
//! A list of `n` numbers `width` bits each (1..=32) is `n / GROUP` groups,
//! then the rest packed one after another (little-endian bits). A group is
//! [`GROUP`] numbers in [`LANES`] lanes of 32-bit words: number `j` goes to
//! lane `j % LANES`, the `j / LANES`-th in it; a lane's numbers are packed
//! one after another into its words, and word `k` of lane `l` is the group's
//! `k * LANES + l`-th. A row of `LANES` numbers is shifted and merged as one
//! 256-bit vector, so packing and unpacking run a group at a time — with the
//! best instruction set the processor has ([`Backend`]: SSE4.2, AVX2, AVX-512,
//! NEON), chosen once at start; the same bytes whichever. Numbers wider than 32 bits go one after another, no groups.
//!
//! Number `i` is found in O(1) ([`get`]), so views read in place.

#[cfg(all(target_arch = "x86_64", target_endian = "little"))]
mod avx2;
#[cfg(all(target_arch = "x86_64", target_endian = "little"))]
mod avx512;
mod backend;
mod dispatch;
mod kernel;
#[cfg(all(target_arch = "aarch64", target_endian = "little"))]
mod neon;
mod packing;
mod portable;
mod row;
#[cfg(all(target_arch = "x86_64", target_endian = "little"))]
mod sse42;

pub use backend::Backend;
pub use packing::{GROUP, GROUP_MAX_BITS, LANES, get, group_bytes, pack, packed_bytes, unpack};
use packing::{ROWS, WORD_BYTES, mask};