Skip to content

Support building on CPUs with AVX but not AVX2 #3040

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
greenrazer merged 1 commit into from
Jul 31, 2025
Merged

Support building on CPUs with AVX but not AVX2 #3040

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 10 additions & 10 deletions candle-core/src/cpu/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,10 +60,10 @@ trait CpuBF16<const ARR: usize> {
use half::{bf16, f16};

#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
pub mod avx;
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
pub use avx::{CurrentCpu, CurrentCpuBF16, CurrentCpuF16};

#[cfg(target_arch = "wasm32")]
Expand All @@ -82,7 +82,7 @@ pub use neon::CurrentCpu;

#[cfg(any(
target_feature = "neon",
target_feature = "avx",
target_feature = "avx2",
target_feature = "simd128"
))]
#[inline(always)]
Expand Down Expand Up @@ -112,7 +112,7 @@ pub(crate) unsafe fn vec_dot_f32(a_row: *const f32, b_row: *const f32, c: *mut f

#[cfg(not(any(
target_feature = "neon",
target_feature = "avx",
target_feature = "avx2",
target_feature = "simd128"
)))]
#[inline(always)]
Expand All @@ -125,7 +125,7 @@ pub(crate) unsafe fn vec_dot_f32(a_row: *const f32, b_row: *const f32, c: *mut f

#[cfg(any(
target_feature = "neon",
target_feature = "avx",
target_feature = "avx2",
target_feature = "simd128"
))]
#[inline(always)]
Expand All @@ -152,7 +152,7 @@ pub(crate) unsafe fn vec_sum(row: *const f32, b: *mut f32, k: usize) {

#[cfg(not(any(
target_feature = "neon",
target_feature = "avx",
target_feature = "avx2",
target_feature = "simd128"
)))]
#[inline(always)]
Expand All @@ -163,7 +163,7 @@ pub(crate) unsafe fn vec_sum(row: *const f32, b: *mut f32, k: usize) {
}
}

#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
#[inline(always)]
pub(crate) unsafe fn vec_dot_f16(a_row: *const f16, b_row: *const f16, c: *mut f32, k: usize) {
let mut sumf = 0.0f32;
Expand Down Expand Up @@ -191,7 +191,7 @@ pub(crate) unsafe fn vec_dot_f16(a_row: *const f16, b_row: *const f16, c: *mut f
*c = sumf;
}

#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
#[inline(always)]
pub(crate) unsafe fn vec_dot_bf16(a_row: *const bf16, b_row: *const bf16, c: *mut f32, k: usize) {
let mut sumf = 0.0f32;
Expand Down Expand Up @@ -219,7 +219,7 @@ pub(crate) unsafe fn vec_dot_bf16(a_row: *const bf16, b_row: *const bf16, c: *mu
*c = sumf;
}

#[cfg(not(target_feature = "avx"))]
#[cfg(not(target_feature = "avx2"))]
#[inline(always)]
pub(crate) unsafe fn vec_dot_f16(a_row: *const f16, b_row: *const f16, c: *mut f32, k: usize) {
// leftovers
Expand All @@ -230,7 +230,7 @@ pub(crate) unsafe fn vec_dot_f16(a_row: *const f16, b_row: *const f16, c: *mut f
*c = sum;
}

#[cfg(not(target_feature = "avx"))]
#[cfg(not(target_feature = "avx2"))]
#[inline(always)]
pub(crate) unsafe fn vec_dot_bf16(a_row: *const bf16, b_row: *const bf16, c: *mut f32, k: usize) {
// leftovers
Expand Down
16 changes: 8 additions & 8 deletions candle-core/src/quantized/k_quants.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -222,7 +222,7 @@ impl GgmlType for BlockQ4_0 {
// https://github.com/ggerganov/llama.cpp/blob/b5ffb2849d23afe73647f68eec7b68187af09be6/ggml.c#L2361C10-L2361C122
#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
return super::avx::vec_dot_q4_0_q8_0(n, xs, ys);

#[cfg(target_feature = "neon")]
Expand Down Expand Up @@ -616,7 +616,7 @@ impl GgmlType for BlockQ8_0 {

#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
return super::avx::vec_dot_q8_0_q8_0(n, xs, ys);

#[cfg(target_feature = "neon")]
Expand Down Expand Up @@ -702,7 +702,7 @@ impl GgmlType for BlockQ2K {

#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
return super::avx::vec_dot_q2k_q8k(n, xs, ys);

#[cfg(target_feature = "neon")]
Expand Down Expand Up @@ -878,7 +878,7 @@ impl GgmlType for BlockQ3K {

#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
return super::avx::vec_dot_q3k_q8k(n, xs, ys);

#[cfg(target_feature = "neon")]
Expand Down Expand Up @@ -1156,7 +1156,7 @@ impl GgmlType for BlockQ4K {

#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
return super::avx::vec_dot_q4k_q8k(n, xs, ys);

#[cfg(target_feature = "neon")]
Expand Down Expand Up @@ -1349,7 +1349,7 @@ impl GgmlType for BlockQ5K {

#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
return super::avx::vec_dot_q5k_q8k(n, xs, ys);

#[cfg(target_feature = "neon")]
Expand Down Expand Up @@ -1570,7 +1570,7 @@ impl GgmlType for BlockQ6K {

#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
return super::avx::vec_dot_q6k_q8k(n, xs, ys);

#[cfg(target_feature = "neon")]
Expand Down Expand Up @@ -1753,7 +1753,7 @@ impl GgmlType for BlockQ8K {

#[allow(unreachable_code)]
fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
return super::avx::vec_dot_q8k_q8k(n, xs, ys);

#[cfg(target_feature = "neon")]
Expand Down
2 changes: 1 addition & 1 deletion candle-core/src/quantized/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ use crate::{Context, CpuStorage, DType, Device, Result, Shape, Storage, Tensor};
use k_quants::*;
use std::borrow::Cow;

#[cfg(target_feature = "avx")]
#[cfg(target_feature = "avx2")]
pub mod avx;
mod dummy_cuda;
mod dummy_metal;
Expand Down
2 changes: 1 addition & 1 deletion candle-core/src/utils.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,7 @@ pub fn metal_is_available() -> bool {
}

pub fn with_avx() -> bool {
cfg!(target_feature = "avx")
cfg!(target_feature = "avx2")
}

pub fn with_neon() -> bool {
Expand Down
Loading