mirror of
https://github.com/arnaucube/poulpy.git
synced 2026-02-10 13:16:44 +01:00
* added cross_basek_normalization * updated method signatures to take layouts * fixed cross-base normalization fix #91 fix #93
265 lines
7.3 KiB
Rust
265 lines
7.3 KiB
Rust
use std::hash::{DefaultHasher, Hasher};
|
|
|
|
use rand::seq::SliceRandom;
|
|
use rand_core::RngCore;
|
|
use rand_distr::{Distribution, weighted::WeightedIndex};
|
|
|
|
use crate::{
|
|
alloc_aligned,
|
|
layouts::{
|
|
Data, DataMut, DataRef, DataView, DataViewMut, DigestU64, FillUniform, ReaderFrom, ToOwnedDeep, VecZnx, WriterTo,
|
|
ZnxInfos, ZnxSliceSize, ZnxView, ZnxViewMut, ZnxZero,
|
|
},
|
|
source::Source,
|
|
};
|
|
|
|
#[repr(C)]
|
|
#[derive(PartialEq, Eq, Debug, Clone, Hash)]
|
|
pub struct ScalarZnx<D: Data> {
|
|
pub data: D,
|
|
pub n: usize,
|
|
pub cols: usize,
|
|
}
|
|
|
|
impl<D: DataRef> DigestU64 for ScalarZnx<D> {
|
|
fn digest_u64(&self) -> u64 {
|
|
let mut h: DefaultHasher = DefaultHasher::new();
|
|
h.write(self.data.as_ref());
|
|
h.write_usize(self.n);
|
|
h.write_usize(self.cols);
|
|
h.finish()
|
|
}
|
|
}
|
|
|
|
impl<D: DataRef> ToOwnedDeep for ScalarZnx<D> {
|
|
type Owned = ScalarZnx<Vec<u8>>;
|
|
fn to_owned_deep(&self) -> Self::Owned {
|
|
ScalarZnx {
|
|
data: self.data.as_ref().to_vec(),
|
|
n: self.n,
|
|
cols: self.cols,
|
|
}
|
|
}
|
|
}
|
|
|
|
impl<D: Data> ZnxInfos for ScalarZnx<D> {
|
|
fn cols(&self) -> usize {
|
|
self.cols
|
|
}
|
|
|
|
fn rows(&self) -> usize {
|
|
1
|
|
}
|
|
|
|
fn n(&self) -> usize {
|
|
self.n
|
|
}
|
|
|
|
fn size(&self) -> usize {
|
|
1
|
|
}
|
|
}
|
|
|
|
impl<D: Data> ZnxSliceSize for ScalarZnx<D> {
|
|
fn sl(&self) -> usize {
|
|
self.n()
|
|
}
|
|
}
|
|
|
|
impl<D: Data> DataView for ScalarZnx<D> {
|
|
type D = D;
|
|
fn data(&self) -> &Self::D {
|
|
&self.data
|
|
}
|
|
}
|
|
|
|
impl<D: Data> DataViewMut for ScalarZnx<D> {
|
|
fn data_mut(&mut self) -> &mut Self::D {
|
|
&mut self.data
|
|
}
|
|
}
|
|
|
|
impl<D: DataRef> ZnxView for ScalarZnx<D> {
|
|
type Scalar = i64;
|
|
}
|
|
|
|
impl<D: DataMut> ScalarZnx<D> {
|
|
pub fn fill_ternary_prob(&mut self, col: usize, prob: f64, source: &mut Source) {
|
|
let choices: [i64; 3] = [-1, 0, 1];
|
|
let weights: [f64; 3] = [prob / 2.0, 1.0 - prob, prob / 2.0];
|
|
let dist: WeightedIndex<f64> = WeightedIndex::new(weights).unwrap();
|
|
self.at_mut(col, 0)
|
|
.iter_mut()
|
|
.for_each(|x: &mut i64| *x = choices[dist.sample(source)]);
|
|
}
|
|
|
|
pub fn fill_ternary_hw(&mut self, col: usize, hw: usize, source: &mut Source) {
|
|
assert!(hw <= self.n());
|
|
self.at_mut(col, 0)[..hw]
|
|
.iter_mut()
|
|
.for_each(|x: &mut i64| *x = (((source.next_u32() & 1) as i64) << 1) - 1);
|
|
self.at_mut(col, 0).shuffle(source);
|
|
}
|
|
|
|
pub fn fill_binary_prob(&mut self, col: usize, prob: f64, source: &mut Source) {
|
|
let choices: [i64; 2] = [0, 1];
|
|
let weights: [f64; 2] = [1.0 - prob, prob];
|
|
let dist: WeightedIndex<f64> = WeightedIndex::new(weights).unwrap();
|
|
self.at_mut(col, 0)
|
|
.iter_mut()
|
|
.for_each(|x: &mut i64| *x = choices[dist.sample(source)]);
|
|
}
|
|
|
|
pub fn fill_binary_hw(&mut self, col: usize, hw: usize, source: &mut Source) {
|
|
assert!(hw <= self.n());
|
|
self.at_mut(col, 0)[..hw]
|
|
.iter_mut()
|
|
.for_each(|x: &mut i64| *x = (source.next_u32() & 1) as i64);
|
|
self.at_mut(col, 0).shuffle(source);
|
|
}
|
|
|
|
pub fn fill_binary_block(&mut self, col: usize, block_size: usize, source: &mut Source) {
|
|
assert!(self.n().is_multiple_of(block_size));
|
|
let max_idx: u64 = (block_size + 1) as u64;
|
|
let mask_idx: u64 = (1 << ((u64::BITS - max_idx.leading_zeros()) as u64)) - 1;
|
|
for block in self.at_mut(col, 0).chunks_mut(block_size) {
|
|
let idx: usize = source.next_u64n(max_idx, mask_idx) as usize;
|
|
if idx != block_size {
|
|
block[idx] = 1;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
impl ScalarZnx<Vec<u8>> {
|
|
pub fn alloc_bytes(n: usize, cols: usize) -> usize {
|
|
n * cols * size_of::<i64>()
|
|
}
|
|
|
|
pub fn alloc(n: usize, cols: usize) -> Self {
|
|
let data: Vec<u8> = alloc_aligned::<u8>(Self::alloc_bytes(n, cols));
|
|
Self { data, n, cols }
|
|
}
|
|
|
|
pub fn from_bytes(n: usize, cols: usize, bytes: impl Into<Vec<u8>>) -> Self {
|
|
let data: Vec<u8> = bytes.into();
|
|
assert!(data.len() == Self::alloc_bytes(n, cols));
|
|
Self { data, n, cols }
|
|
}
|
|
}
|
|
|
|
impl<D: DataMut> ZnxZero for ScalarZnx<D> {
|
|
fn zero(&mut self) {
|
|
self.raw_mut().fill(0)
|
|
}
|
|
fn zero_at(&mut self, i: usize, j: usize) {
|
|
self.at_mut(i, j).fill(0);
|
|
}
|
|
}
|
|
|
|
impl<D: DataMut> FillUniform for ScalarZnx<D> {
|
|
fn fill_uniform(&mut self, log_bound: usize, source: &mut Source) {
|
|
match log_bound {
|
|
64 => source.fill_bytes(self.data.as_mut()),
|
|
0 => panic!("invalid log_bound, cannot be zero"),
|
|
_ => {
|
|
let mask: u64 = (1u64 << log_bound) - 1;
|
|
for x in self.raw_mut().iter_mut() {
|
|
let r = source.next_u64() & mask;
|
|
*x = ((r << (64 - log_bound)) as i64) >> (64 - log_bound);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
pub type ScalarZnxOwned = ScalarZnx<Vec<u8>>;
|
|
|
|
impl<D: Data> ScalarZnx<D> {
|
|
pub fn from_data(data: D, n: usize, cols: usize) -> Self {
|
|
Self { data, n, cols }
|
|
}
|
|
}
|
|
|
|
pub trait ScalarZnxToRef {
|
|
fn to_ref(&self) -> ScalarZnx<&[u8]>;
|
|
}
|
|
|
|
impl<D: DataRef> ScalarZnxToRef for ScalarZnx<D> {
|
|
fn to_ref(&self) -> ScalarZnx<&[u8]> {
|
|
ScalarZnx {
|
|
data: self.data.as_ref(),
|
|
n: self.n,
|
|
cols: self.cols,
|
|
}
|
|
}
|
|
}
|
|
|
|
pub trait ScalarZnxToMut {
|
|
fn to_mut(&mut self) -> ScalarZnx<&mut [u8]>;
|
|
}
|
|
|
|
impl<D: DataMut> ScalarZnxToMut for ScalarZnx<D> {
|
|
fn to_mut(&mut self) -> ScalarZnx<&mut [u8]> {
|
|
ScalarZnx {
|
|
data: self.data.as_mut(),
|
|
n: self.n,
|
|
cols: self.cols,
|
|
}
|
|
}
|
|
}
|
|
|
|
impl<D: DataRef> ScalarZnx<D> {
|
|
pub fn as_vec_znx(&self) -> VecZnx<&[u8]> {
|
|
VecZnx {
|
|
data: self.data.as_ref(),
|
|
n: self.n,
|
|
cols: self.cols,
|
|
size: 1,
|
|
max_size: 1,
|
|
}
|
|
}
|
|
}
|
|
|
|
impl<D: DataMut> ScalarZnx<D> {
|
|
pub fn as_vec_znx_mut(&mut self) -> VecZnx<&mut [u8]> {
|
|
VecZnx {
|
|
data: self.data.as_mut(),
|
|
n: self.n,
|
|
cols: self.cols,
|
|
size: 1,
|
|
max_size: 1,
|
|
}
|
|
}
|
|
}
|
|
|
|
use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt};
|
|
|
|
impl<D: DataMut> ReaderFrom for ScalarZnx<D> {
|
|
fn read_from<R: std::io::Read>(&mut self, reader: &mut R) -> std::io::Result<()> {
|
|
self.n = reader.read_u64::<LittleEndian>()? as usize;
|
|
self.cols = reader.read_u64::<LittleEndian>()? as usize;
|
|
let len: usize = reader.read_u64::<LittleEndian>()? as usize;
|
|
let buf: &mut [u8] = self.data.as_mut();
|
|
if buf.len() != len {
|
|
return Err(std::io::Error::new(
|
|
std::io::ErrorKind::UnexpectedEof,
|
|
format!("self.data.len()={} != read len={}", buf.len(), len),
|
|
));
|
|
}
|
|
reader.read_exact(&mut buf[..len])?;
|
|
Ok(())
|
|
}
|
|
}
|
|
|
|
impl<D: DataRef> WriterTo for ScalarZnx<D> {
|
|
fn write_to<W: std::io::Write>(&self, writer: &mut W) -> std::io::Result<()> {
|
|
writer.write_u64::<LittleEndian>(self.n as u64)?;
|
|
writer.write_u64::<LittleEndian>(self.cols as u64)?;
|
|
let buf: &[u8] = self.data.as_ref();
|
|
writer.write_u64::<LittleEndian>(buf.len() as u64)?;
|
|
writer.write_all(buf)?;
|
|
Ok(())
|
|
}
|
|
}
|