Skip to content

Implement reduce_product() #361

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
Shnatsel merged 9 commits into from
Sep 3, 2026
Merged

Implement reduce_product() #361

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
ac7181b
implement reduce_product
Shnatsel Aug 30, 2026
0a28d38
Add tests for reduce_product
Shnatsel Aug 30, 2026
1542be3
Optimize WASM implementation of reduce_product
Shnatsel Aug 30, 2026
f7a62ac
Specialize u8/i8/u16/i16 paths for up to 15% boost for 8-bit and up t…
Shnatsel Aug 30, 2026
ae45898
Polish the docs a bit
Shnatsel Aug 30, 2026
cb674ee
Fix build after fiddling with docs
Shnatsel Aug 30, 2026
2f6d79e
Fix the documented number of roundings
Shnatsel Sep 3, 2026
0bf528a
Fix doc string error, thank you Laurenz
Shnatsel Sep 3, 2026
03d4e05
Merge branch 'main' into reduce-product
Shnatsel Sep 3, 2026
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
119 changes: 119 additions & 0 deletions fearless_simd/src/generated/avx2.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -328,6 +328,18 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_f32x4(self, a: f32x4<Self>) -> f32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, a: f32x4<Avx2>) -> f32 {
let a: __m128 = a.into();
let adjacent = _mm_mul_ps(a, _mm_shuffle_ps::<0b10_11_00_01>(a, a));
_mm_cvtss_f32(_mm_mul_ss(adjacent, _mm_movehl_ps(adjacent, adjacent)))
}
);
kernel(self, a)
}
#[inline(always)]
fn max_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -1006,6 +1018,25 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i8x16(self, a: i8x16<Self>) -> i8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, a: i8x16<Avx2>) -> i8 {
let value: __m128i = a.into();
let product = _mm_mullo_epi16(
_mm_unpacklo_epi8(value, value),
_mm_unpackhi_epi8(value, value),
);
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [i8; 16usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -1588,6 +1619,25 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u8x16(self, a: u8x16<Self>) -> u8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, a: u8x16<Avx2>) -> u8 {
let value: __m128i = a.into();
let product = _mm_mullo_epi16(
_mm_unpacklo_epi8(value, value),
_mm_unpackhi_epi8(value, value),
);
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [u8; 16usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -2288,6 +2338,21 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i16x8(self, a: i16x8<Self>) -> i16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, a: i16x8<Avx2>) -> i16 {
let product: __m128i = a.into();
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [i16; 8usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -2821,6 +2886,21 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u16x8(self, a: u16x8<Self>) -> u16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, a: u16x8<Avx2>) -> u16 {
let product: __m128i = a.into();
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [u16; 8usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -3572,6 +3652,15 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i32x4(self, a: i32x4<Self>) -> i32 {
let mul_level_0: [i32; 2usize] = [
a[0usize].wrapping_mul(a[1usize]),
a[2usize].wrapping_mul(a[3usize]),
];
let mul_level_1: [i32; 1usize] = [mul_level_0[0usize].wrapping_mul(mul_level_0[1usize])];
mul_level_1[0]
}
#[inline(always)]
fn max_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -4087,6 +4176,15 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u32x4(self, a: u32x4<Self>) -> u32 {
let mul_level_0: [u32; 2usize] = [
a[0usize].wrapping_mul(a[1usize]),
a[2usize].wrapping_mul(a[3usize]),
];
let mul_level_1: [u32; 1usize] = [mul_level_0[0usize].wrapping_mul(mul_level_0[1usize])];
mul_level_1[0]
}
#[inline(always)]
fn max_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -4763,6 +4861,17 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_f64x2(self, a: f64x2<Self>) -> f64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, a: f64x2<Avx2>) -> f64 {
let a: __m128d = a.into();
_mm_cvtsd_f64(_mm_mul_sd(a, _mm_unpackhi_pd(a, a)))
}
);
kernel(self, a)
}
#[inline(always)]
fn max_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -5406,6 +5515,11 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i64x2(self, a: i64x2<Self>) -> i64 {
let mul_level_0: [i64; 1usize] = [a[0usize].wrapping_mul(a[1usize])];
mul_level_0[0]
}
#[inline(always)]
fn max_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
[
i64::max(a[0usize], b[0usize]),
Expand Down Expand Up @@ -5889,6 +6003,11 @@ impl Simd for Avx2 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u64x2(self, a: u64x2<Self>) -> u64 {
let mul_level_0: [u64; 1usize] = [a[0usize].wrapping_mul(a[1usize])];
mul_level_0[0]
}
#[inline(always)]
fn max_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
[
u64::max(a[0usize], b[0usize]),
Expand Down
119 changes: 119 additions & 0 deletions fearless_simd/src/generated/avx512.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -586,6 +586,18 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_f32x4(self, a: f32x4<Self>) -> f32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx512, a: f32x4<Avx512>) -> f32 {
let a: __m128 = a.into();
let adjacent = _mm_mul_ps(a, _mm_shuffle_ps::<0b10_11_00_01>(a, a));
_mm_cvtss_f32(_mm_mul_ss(adjacent, _mm_movehl_ps(adjacent, adjacent)))
}
);
kernel(self, a)
}
#[inline(always)]
fn max_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -1208,6 +1220,25 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i8x16(self, a: i8x16<Self>) -> i8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx512, a: i8x16<Avx512>) -> i8 {
let value: __m128i = a.into();
let product = _mm_mullo_epi16(
_mm_unpacklo_epi8(value, value),
_mm_unpackhi_epi8(value, value),
);
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [i8; 16usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -1776,6 +1807,25 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u8x16(self, a: u8x16<Self>) -> u8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx512, a: u8x16<Avx512>) -> u8 {
let value: __m128i = a.into();
let product = _mm_mullo_epi16(
_mm_unpacklo_epi8(value, value),
_mm_unpackhi_epi8(value, value),
);
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [u8; 16usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -2381,6 +2431,21 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i16x8(self, a: i16x8<Self>) -> i16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx512, a: i16x8<Avx512>) -> i16 {
let product: __m128i = a.into();
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [i16; 8usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -2882,6 +2947,21 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u16x8(self, a: u16x8<Self>) -> u16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx512, a: u16x8<Avx512>) -> u16 {
let product: __m128i = a.into();
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [u16; 8usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -3517,6 +3597,15 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i32x4(self, a: i32x4<Self>) -> i32 {
let mul_level_0: [i32; 2usize] = [
a[0usize].wrapping_mul(a[1usize]),
a[2usize].wrapping_mul(a[3usize]),
];
let mul_level_1: [i32; 1usize] = [mul_level_0[0usize].wrapping_mul(mul_level_0[1usize])];
mul_level_1[0]
}
#[inline(always)]
fn max_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -4012,6 +4101,15 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u32x4(self, a: u32x4<Self>) -> u32 {
let mul_level_0: [u32; 2usize] = [
a[0usize].wrapping_mul(a[1usize]),
a[2usize].wrapping_mul(a[3usize]),
];
let mul_level_1: [u32; 1usize] = [mul_level_0[0usize].wrapping_mul(mul_level_0[1usize])];
mul_level_1[0]
}
#[inline(always)]
fn max_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -4611,6 +4709,17 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_f64x2(self, a: f64x2<Self>) -> f64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx512, a: f64x2<Avx512>) -> f64 {
let a: __m128d = a.into();
_mm_cvtsd_f64(_mm_mul_sd(a, _mm_unpackhi_pd(a, a)))
}
);
kernel(self, a)
}
#[inline(always)]
fn max_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -5179,6 +5288,11 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i64x2(self, a: i64x2<Self>) -> i64 {
let mul_level_0: [i64; 1usize] = [a[0usize].wrapping_mul(a[1usize])];
mul_level_0[0]
}
#[inline(always)]
fn max_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
Expand Down Expand Up @@ -5639,6 +5753,11 @@ impl Simd for Avx512 {
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u64x2(self, a: u64x2<Self>) -> u64 {
let mul_level_0: [u64; 1usize] = [a[0usize].wrapping_mul(a[1usize])];
mul_level_0[0]
}
#[inline(always)]
fn max_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
Expand Down
Loading
Loading