Skip to content

Optimize interleaved loads on x86 #307

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

Draft
Shnatsel wants to merge 1 commit into
base: main
Choose a base branch
from
+56 −83
Draft

Optimize interleaved loads on x86 #307

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
93 changes: 51 additions & 42 deletions fearless_simd/src/generated/avx2.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4732,23 +4732,26 @@ impl Simd for Avx2 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, src: &[f64; 8usize]) -> [f64x2<Avx2>; 4usize] {
let (chunks, []) = src.as_chunks::<4>() else {
let (chunks, []) = src.as_chunks::<2usize>() else {
unreachable!()
};
let v0: __m256d =
crate::transmute::checked_transmute_copy::<[f64; 4], __m256d>(&chunks[0]);
let v1: __m256d =
crate::transmute::checked_transmute_copy::<[f64; 4], __m256d>(&chunks[1]);
let lo = _mm256_unpacklo_pd(v0, v1);
let hi = _mm256_unpackhi_pd(v0, v1);
let out0 = _mm256_permute2f128_pd::<0x20>(lo, hi);
let out1 = _mm256_permute2f128_pd::<0x31>(lo, hi);
let outputs: [__m128d; 4] = crate::transmute::checked_transmute_copy(&[out0, out1]);
let v0: __m128d =
crate::transmute::checked_transmute_copy::<[f64; 2usize], __m128d>(&chunks[0]);
let v1: __m128d =
crate::transmute::checked_transmute_copy::<[f64; 2usize], __m128d>(&chunks[1]);
let v2: __m128d =
crate::transmute::checked_transmute_copy::<[f64; 2usize], __m128d>(&chunks[2]);
let v3: __m128d =
crate::transmute::checked_transmute_copy::<[f64; 2usize], __m128d>(&chunks[3]);
let out0 = _mm_unpacklo_pd(v0, v2);
let out1 = _mm_unpackhi_pd(v0, v2);
let out2 = _mm_unpacklo_pd(v1, v3);
let out3 = _mm_unpackhi_pd(v1, v3);
[
outputs[0].simd_into(token),
outputs[1].simd_into(token),
outputs[2].simd_into(token),
outputs[3].simd_into(token),
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
Expand Down Expand Up @@ -5153,23 +5156,26 @@ impl Simd for Avx2 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, src: &[i64; 8usize]) -> [i64x2<Avx2>; 4usize] {
let (chunks, []) = src.as_chunks::<4>() else {
let (chunks, []) = src.as_chunks::<2usize>() else {
unreachable!()
};
let v0: __m256i =
crate::transmute::checked_transmute_copy::<[i64; 4], __m256i>(&chunks[0]);
let v1: __m256i =
crate::transmute::checked_transmute_copy::<[i64; 4], __m256i>(&chunks[1]);
let lo = _mm256_unpacklo_epi64(v0, v1);
let hi = _mm256_unpackhi_epi64(v0, v1);
let out0 = _mm256_permute2x128_si256::<0x20>(lo, hi);
let out1 = _mm256_permute2x128_si256::<0x31>(lo, hi);
let outputs: [__m128i; 4] = crate::transmute::checked_transmute_copy(&[out0, out1]);
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[i64; 2usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[i64; 2usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[i64; 2usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[i64; 2usize], __m128i>(&chunks[3]);
let out0 = _mm_unpacklo_epi64(v0, v2);
let out1 = _mm_unpackhi_epi64(v0, v2);
let out2 = _mm_unpacklo_epi64(v1, v3);
let out3 = _mm_unpackhi_epi64(v1, v3);
[
outputs[0].simd_into(token),
outputs[1].simd_into(token),
outputs[2].simd_into(token),
outputs[3].simd_into(token),
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
Expand Down Expand Up @@ -5561,23 +5567,26 @@ impl Simd for Avx2 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Avx2, src: &[u64; 8usize]) -> [u64x2<Avx2>; 4usize] {
let (chunks, []) = src.as_chunks::<4>() else {
let (chunks, []) = src.as_chunks::<2usize>() else {
unreachable!()
};
let v0: __m256i =
crate::transmute::checked_transmute_copy::<[u64; 4], __m256i>(&chunks[0]);
let v1: __m256i =
crate::transmute::checked_transmute_copy::<[u64; 4], __m256i>(&chunks[1]);
let lo = _mm256_unpacklo_epi64(v0, v1);
let hi = _mm256_unpackhi_epi64(v0, v1);
let out0 = _mm256_permute2x128_si256::<0x20>(lo, hi);
let out1 = _mm256_permute2x128_si256::<0x31>(lo, hi);
let outputs: [__m128i; 4] = crate::transmute::checked_transmute_copy(&[out0, out1]);
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[u64; 2usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[u64; 2usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[u64; 2usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[u64; 2usize], __m128i>(&chunks[3]);
let out0 = _mm_unpacklo_epi64(v0, v2);
let out1 = _mm_unpackhi_epi64(v0, v2);
let out2 = _mm_unpacklo_epi64(v1, v3);
let out3 = _mm_unpackhi_epi64(v1, v3);
[
outputs[0].simd_into(token),
outputs[1].simd_into(token),
outputs[2].simd_into(token),
outputs[3].simd_into(token),
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
Expand Down
46 changes: 5 additions & 41 deletions fearless_simd_gen/src/mk_x86.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3252,58 +3252,22 @@ impl X86 {
if *self == Self::Sse2 && matches!(vec_ty.scalar_bits, 8 | 16) {
return fallback_method(op, vec_ty);
}
// SSE4.2 intrinsics are used on both SSE4.2 and AVX2,
// because AVX2 cross-lane shuffles have higher latency than this SSE4.2 formulation,
// and for load operations latency matters a lot more than throughput
// since data processing cannot start until the data is loaded.
match vec_ty.scalar_bits {
64 | 32 | 16 | 8 => {
let avx2_64 = *self == Self::Avx2 && vec_ty.scalar_bits == 64;
let block_len = if avx2_64 {
4
} else {
block_size as usize / vec_ty.scalar_bits
};
let block_len = block_size as usize / vec_ty.scalar_bits;
let block_ty = VecType::new(vec_ty.scalar, vec_ty.scalar_bits, block_len);
let scalar_ty = block_ty.scalar.rust(block_ty.scalar_bits);
let native_ty = self.arch_ty(&block_ty);
let native_block_ty = self.arch_ty(vec_ty);
let vec_32 = block_ty.reinterpret(block_ty.scalar, 32);
let unpacklo_32 = simple_sign_unaware_intrinsic("unpacklo", &vec_32);
let unpackhi_32 = simple_sign_unaware_intrinsic("unpackhi", &vec_32);
let vec_64 = block_ty.reinterpret(block_ty.scalar, 64);
let unpacklo_64 = simple_sign_unaware_intrinsic("unpacklo", &vec_64);
let unpackhi_64 = simple_sign_unaware_intrinsic("unpackhi", &vec_64);
let permute_128 = intrinsic_ident(
match vec_ty.scalar {
ScalarType::Float => "permute2f128",
_ => "permute2x128",
},
coarse_type(&block_ty),
256,
);

if avx2_64 {
return self.kernel_method(op, vec_ty, |token| {
quote! {
let (chunks, []) = src.as_chunks::<4>() else {
unreachable!()
};
let v0: #native_ty = crate::transmute::checked_transmute_copy::<[#scalar_ty; 4], #native_ty>(&chunks[0]);
let v1: #native_ty = crate::transmute::checked_transmute_copy::<[#scalar_ty; 4], #native_ty>(&chunks[1]);

let lo = #unpacklo_64(v0, v1); // [0,4,2,6]
let hi = #unpackhi_64(v0, v1); // [1,5,3,7]
let out0 = #permute_128::<0x20>(lo, hi); // [0,4,1,5]
let out1 = #permute_128::<0x31>(lo, hi); // [2,6,3,7]
let outputs: [#native_block_ty; 4] =
crate::transmute::checked_transmute_copy(&[out0, out1]);

[
outputs[0].simd_into(#token),
outputs[1].simd_into(#token),
outputs[2].simd_into(#token),
outputs[3].simd_into(#token),
]
}
});
}

let init_shuffle = match vec_ty.scalar_bits {
16 => Some(quote! {
Expand Down
Loading