database/pgvecto.rs
1
0
Fork 0
mirror of https://github.com/tensorchord/pgvecto.rs.git synced 2025-08-01 06:46:52 +03:00
Code Activity

C code tests & avx512f f16 implement (#183)

Browse Source 
* test: add tests for c code

Signed-off-by: usamoi <usamoi@outlook.com>

* fix: relax EPSILON for tests

Signed-off-by: usamoi <usamoi@outlook.com>

---------

Signed-off-by: usamoi <usamoi@outlook.com>
This commit is contained in:
Usamoi
2023-12-15 16:00:08 +08:00
committed by GitHub
parent 2869fbd44c
commit c50912e87d
13 changed files with 276 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
crates/service/Cargo.toml
View File

@ -16,7 +16,7 @@ bincode.workspace = true
half.workspace = true
num-traits.workspace = true
c = { path = "../c" }
std_detect = { git = "https://github.com/tensorchord/stdarch.git", branch = "avx512fp16" }
detect = { path = "../detect" }
rand = "0.8.5"
crc32fast = "1.3.2"
crossbeam = "0.8.2"
@ -32,7 +32,6 @@ arc-swap = "1.6.0"
bytemuck = { version = "1.14.0", features = ["extern_crate_alloc"] }
serde_with = "3.4.0"
multiversion = "0.7.3"
ctor = "0.2.6"
[target.'cfg(target_os = "macos")'.dependencies]
ulock-sys = "0.1.0"

36
crates/service/src/prelude/global/f16.rs
View File

@ -22,7 +22,7 @@ pub fn cosine(lhs: &[F16], rhs: &[F16]) -> F32 {
xy / (x2 * y2).sqrt()
}
#[cfg(target_arch = "x86_64")]
if crate::utils::detect::x86_64::detect_avx512fp16() {
if detect::x86_64::detect_avx512fp16() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {
@ -30,7 +30,15 @@ pub fn cosine(lhs: &[F16], rhs: &[F16]) -> F32 {
}
}
#[cfg(target_arch = "x86_64")]
if crate::utils::detect::x86_64::detect_v3() {
if detect::x86_64::detect_v4() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {
return c::v_f16_cosine_v4(lhs.as_ptr().cast(), rhs.as_ptr().cast(), n).into();
}
}
#[cfg(target_arch = "x86_64")]
if detect::x86_64::detect_v3() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {
@ -58,7 +66,7 @@ pub fn dot(lhs: &[F16], rhs: &[F16]) -> F32 {
xy
}
#[cfg(target_arch = "x86_64")]
if crate::utils::detect::x86_64::detect_avx512fp16() {
if detect::x86_64::detect_avx512fp16() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {
@ -66,7 +74,15 @@ pub fn dot(lhs: &[F16], rhs: &[F16]) -> F32 {
}
}
#[cfg(target_arch = "x86_64")]
if crate::utils::detect::x86_64::detect_v3() {
if detect::x86_64::detect_v4() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {
return c::v_f16_dot_v4(lhs.as_ptr().cast(), rhs.as_ptr().cast(), n).into();
}
}
#[cfg(target_arch = "x86_64")]
if detect::x86_64::detect_v3() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {
@ -95,7 +111,7 @@ pub fn sl2(lhs: &[F16], rhs: &[F16]) -> F32 {
d2
}
#[cfg(target_arch = "x86_64")]
if crate::utils::detect::x86_64::detect_avx512fp16() {
if detect::x86_64::detect_avx512fp16() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {
@ -103,7 +119,15 @@ pub fn sl2(lhs: &[F16], rhs: &[F16]) -> F32 {
}
}
#[cfg(target_arch = "x86_64")]
if crate::utils::detect::x86_64::detect_v3() {
if detect::x86_64::detect_v4() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {
return c::v_f16_sl2_v4(lhs.as_ptr().cast(), rhs.as_ptr().cast(), n).into();
}
}
#[cfg(target_arch = "x86_64")]
if detect::x86_64::detect_v3() {
assert!(lhs.len() == rhs.len());
let n = lhs.len();
unsafe {

1
crates/service/src/utils/detect.rs
View File

@ -1 +0,0 @@
pub mod x86_64;

85
crates/service/src/utils/detect/x86_64.rs
View File

@ -1,85 +0,0 @@
#![cfg(target_arch = "x86_64")]
use std::sync::atomic::{AtomicBool, Ordering};
static ATOMIC_AVX512FP16: AtomicBool = AtomicBool::new(false);
pub fn test_avx512fp16() -> bool {
std_detect::is_x86_feature_detected!("avx512fp16") && test_v4()
}
#[ctor::ctor]
fn ctor_avx512fp16() {
ATOMIC_AVX512FP16.store(test_avx512fp16(), Ordering::Relaxed);
}
pub fn detect_avx512fp16() -> bool {
ATOMIC_AVX512FP16.load(Ordering::Relaxed)
}
static ATOMIC_V4: AtomicBool = AtomicBool::new(false);
pub fn test_v4() -> bool {
std_detect::is_x86_feature_detected!("avx512bw")
&& std_detect::is_x86_feature_detected!("avx512cd")
&& std_detect::is_x86_feature_detected!("avx512dq")
&& std_detect::is_x86_feature_detected!("avx512f")
&& std_detect::is_x86_feature_detected!("avx512vl")
&& test_v3()
}
#[ctor::ctor]
fn ctor_v4() {
ATOMIC_V4.store(test_v4(), Ordering::Relaxed);
}
pub fn _detect_v4() -> bool {
ATOMIC_V4.load(Ordering::Relaxed)
}
static ATOMIC_V3: AtomicBool = AtomicBool::new(false);
pub fn test_v3() -> bool {
std_detect::is_x86_feature_detected!("avx")
&& std_detect::is_x86_feature_detected!("avx2")
&& std_detect::is_x86_feature_detected!("bmi1")
&& std_detect::is_x86_feature_detected!("bmi2")
&& std_detect::is_x86_feature_detected!("f16c")
&& std_detect::is_x86_feature_detected!("fma")
&& std_detect::is_x86_feature_detected!("lzcnt")
&& std_detect::is_x86_feature_detected!("movbe")
&& std_detect::is_x86_feature_detected!("xsave")
&& test_v2()
}
#[ctor::ctor]
fn ctor_v3() {
ATOMIC_V3.store(test_v3(), Ordering::Relaxed);
}
pub fn detect_v3() -> bool {
ATOMIC_V3.load(Ordering::Relaxed)
}
static ATOMIC_V2: AtomicBool = AtomicBool::new(false);
pub fn test_v2() -> bool {
std_detect::is_x86_feature_detected!("cmpxchg16b")
&& std_detect::is_x86_feature_detected!("fxsr")
&& std_detect::is_x86_feature_detected!("popcnt")
&& std_detect::is_x86_feature_detected!("sse")
&& std_detect::is_x86_feature_detected!("sse2")
&& std_detect::is_x86_feature_detected!("sse3")
&& std_detect::is_x86_feature_detected!("sse4.1")
&& std_detect::is_x86_feature_detected!("sse4.2")
&& std_detect::is_x86_feature_detected!("ssse3")
}
#[ctor::ctor]
fn ctor_v2() {
ATOMIC_V2.store(test_v2(), Ordering::Relaxed);
}
pub fn _detect_v2() -> bool {
ATOMIC_V2.load(Ordering::Relaxed)
}

1
crates/service/src/utils/mod.rs
View File

@ -1,6 +1,5 @@
pub mod cells;
pub mod clean;
pub mod detect;
pub mod dir_ops;
pub mod file_atomic;
pub mod file_wal;