arnaucube/poulpy
1
0
Fork 0
mirror of https://github.com/arnaucube/poulpy.git synced 2026-02-10 13:16:44 +01:00
Code Issues Packages Projects Releases Wiki Activity

Changes to build on macos (#92)

Browse Source 
* update spqlios-airthmetic to latest e8aed63

* finally, builds on macos

* clippy

---------

Co-authored-by: Pro7ech <jeanphilippe.bossuat@gmail.com>
This commit is contained in:
Janmajayamall
2025-09-25 14:39:27 +02:00
committed by GitHub
parent af5d8a308f
commit 4da790ea6a
18 changed files with 127 additions and 154 deletions
Download Patch File Download Diff File Expand all files Collapse all files

171
poulpy-backend/benches/fft.rs
View File

@@ -32,47 +32,6 @@ pub fn bench_fft_ref(c: &mut Criterion) {
group.finish();
}
pub fn bench_fft_avx2_fma(c: &mut Criterion) {
let group_name: String = "fft_avx2_fma".to_string();
let mut group = c.benchmark_group(group_name);
#[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
#[target_feature(enable = "avx2,fma")]
fn runner(m: usize) -> impl FnMut() {
let mut values: Vec<f64> = vec![0f64; m << 1];
let scale = 1.0f64 / (2 * m) as f64;
values
.iter_mut()
.enumerate()
.for_each(|(i, x)| *x = (i + 1) as f64 * scale);
let table: ReimFFTTable<f64> = ReimFFTTable::<f64>::new(m);
move || {
use poulpy_backend::cpu_fft64_avx::ReimFFTAvx;
ReimFFTAvx::reim_dft_execute(&table, &mut values);
black_box(());
}
}
if std::is_x86_feature_detected!("avx2") {
for log_m in [9, 10, 11, 12, 13, 14, 15] {
let id: BenchmarkId = BenchmarkId::from_parameter(format!("n: {}", 2 << log_m));
unsafe {
let mut runner = runner(1 << log_m);
group.bench_with_input(id, &(), |b, _| b.iter(&mut runner));
}
}
} else {
eprintln!("skipping: CPU lacks avx2");
return;
}
group.finish();
}
pub fn bench_fft_spqlios(c: &mut Criterion) {
let group_name: String = "fft_spqlios".to_string();
@@ -136,47 +95,6 @@ pub fn bench_ifft_ref(c: &mut Criterion) {
group.finish();
}
pub fn bench_ifft_avx2_fma(c: &mut Criterion) {
let group_name: String = "ifft_avx2_fma".to_string();
let mut group = c.benchmark_group(group_name);
#[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
#[target_feature(enable = "avx2,fma")]
fn runner(m: usize) -> impl FnMut() {
let mut values: Vec<f64> = vec![0f64; m << 1];
let scale = 1.0f64 / (2 * m) as f64;
values
.iter_mut()
.enumerate()
.for_each(|(i, x)| *x = (i + 1) as f64 * scale);
let table: ReimIFFTTable<f64> = ReimIFFTTable::<f64>::new(m);
move || {
use poulpy_backend::cpu_fft64_avx::ReimIFFTAvx;
ReimIFFTAvx::reim_dft_execute(&table, &mut values);
black_box(());
}
}
if std::is_x86_feature_detected!("avx2") {
for log_m in [9, 10, 11, 12, 13, 14, 15] {
let id: BenchmarkId = BenchmarkId::from_parameter(format!("n: {}", 2 << log_m));
unsafe {
let mut runner = runner(1 << log_m);
group.bench_with_input(id, &(), |b, _| b.iter(&mut runner));
}
}
} else {
eprintln!("skipping: CPU lacks avx2");
return;
}
group.finish();
}
pub fn bench_ifft_spqlios(c: &mut Criterion) {
let group_name: String = "ifft_spqlios".to_string();
@@ -212,13 +130,98 @@ pub fn bench_ifft_spqlios(c: &mut Criterion) {
group.finish();
}
#[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
mod x86 {
use super::*;
#[allow(dead_code)]
pub fn bench_ifft_avx2_fma(c: &mut Criterion) {
let group_name: String = "ifft_avx2_fma".to_string();
let mut group = c.benchmark_group(group_name);
if std::is_x86_feature_detected!("avx2") {
fn runner(m: usize) -> impl FnMut() {
let mut values: Vec<f64> = vec![0f64; m << 1];
let scale = 1.0f64 / (2 * m) as f64;
values
.iter_mut()
.enumerate()
.for_each(|(i, x)| *x = (i + 1) as f64 * scale);
let table: ReimIFFTTable<f64> = ReimIFFTTable::<f64>::new(m);
move || {
use poulpy_backend::cpu_fft64_avx::ReimIFFTAvx;
ReimIFFTAvx::reim_dft_execute(&table, &mut values);
black_box(());
}
}
for log_m in [9, 10, 11, 12, 13, 14, 15] {
let id: BenchmarkId = BenchmarkId::from_parameter(format!("n: {}", 2 << log_m));
let mut runner = runner(1 << log_m);
group.bench_with_input(id, &(), |b, _| b.iter(&mut runner));
}
} else {
eprintln!("skipping: CPU lacks avx2");
return;
}
group.finish();
}
#[allow(dead_code)]
pub fn bench_fft_avx2_fma(c: &mut Criterion) {
let group_name: String = "fft_avx2_fma".to_string();
let mut group = c.benchmark_group(group_name);
if std::is_x86_feature_detected!("avx2") {
fn runner(m: usize) -> impl FnMut() {
let mut values: Vec<f64> = vec![0f64; m << 1];
let scale = 1.0f64 / (2 * m) as f64;
values
.iter_mut()
.enumerate()
.for_each(|(i, x)| *x = (i + 1) as f64 * scale);
let table: ReimFFTTable<f64> = ReimFFTTable::<f64>::new(m);
move || {
use poulpy_backend::cpu_fft64_avx::ReimFFTAvx;
ReimFFTAvx::reim_dft_execute(&table, &mut values);
black_box(());
}
}
for log_m in [9, 10, 11, 12, 13, 14, 15] {
let id: BenchmarkId = BenchmarkId::from_parameter(format!("n: {}", 2 << log_m));
let mut runner = runner(1 << log_m);
group.bench_with_input(id, &(), |b, _| b.iter(&mut runner));
}
} else {
eprintln!("skipping: CPU lacks avx2");
return;
}
group.finish();
}
criterion_group!(benches_x86, bench_fft_avx2_fma, bench_ifft_avx2_fma,);
criterion_main!(benches_x86);
}
criterion_group!(
benches,
bench_fft_ref,
bench_fft_avx2_fma,
bench_fft_spqlios,
bench_ifft_ref,
bench_ifft_avx2_fma,
bench_ifft_spqlios
);
criterion_main!(benches);

17
poulpy-backend/benches/vmp.rs
View File

@@ -1,6 +1,6 @@
// poulpy-backend/benches/vec_znx_add.rs
use criterion::{Criterion, criterion_group, criterion_main};
use poulpy_backend::{FFT64Avx, FFT64Ref, FFT64Spqlios};
use poulpy_backend::{FFT64Ref, FFT64Spqlios};
use poulpy_hal::bench_suite::vmp::bench_vmp_apply_dft_to_dft;
fn bench_vmp_apply_dft_to_dft_cpu_spqlios_fft64(c: &mut Criterion) {
@@ -11,14 +11,23 @@ fn bench_vmp_apply_dft_to_dft_cpu_ref_fft64(c: &mut Criterion) {
bench_vmp_apply_dft_to_dft::<FFT64Ref>(c, "cpu_ref::fft64");
}
fn bench_vmp_apply_dft_to_dft_cpu_avx_fft64(c: &mut Criterion) {
bench_vmp_apply_dft_to_dft::<FFT64Avx>(c, "cpu_avx::fft64");
#[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
mod x86 {
use super::*;
use poulpy_backend::FFT64Avx;
#[allow(dead_code)]
fn bench_vmp_apply_dft_to_dft_cpu_avx_fft64(c: &mut Criterion) {
bench_vmp_apply_dft_to_dft::<FFT64Avx>(c, "cpu_avx::fft64");
}
criterion_group!(benches_x86, bench_vmp_apply_dft_to_dft_cpu_avx_fft64,);
criterion_main!(benches_x86);
}
criterion_group!(
benches,
bench_vmp_apply_dft_to_dft_cpu_spqlios_fft64,
bench_vmp_apply_dft_to_dft_cpu_ref_fft64,
bench_vmp_apply_dft_to_dft_cpu_avx_fft64,
);
criterion_main!(benches);