Skip to content
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

Jump to bottom

aarch64: marginally improve performance of the aes-ctr part of aes-gcm #20

Merged
merged 3 commits into from
Oct 27, 2024
Merged

aarch64: marginally improve performance of the aes-ctr part of aes-gcm #20

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9b02e3b
Prepare to specialise aes_gcm
ctz Oct 27, 2024
e2514e9
aarch64: by-8 AES-CTR
ctz Oct 27, 2024
6db3ff3
aarch64: centralise prefetch asm & use in aes
ctz Oct 27, 2024
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
259 changes: 243 additions & 16 deletions graviola/src/low/aarch64/aes.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
// cf. the x86_64 version, on which this one is based.

use crate::low;
use crate::low::aarch64::cpu;
use core::arch::aarch64::*;

pub(crate) enum AesKey {
Expand Down Expand Up @@ -38,18 +39,102 @@ impl AesKey {
}
}

pub(crate) fn round_keys(&self) -> (uint8x16_t, &[uint8x16_t], uint8x16_t) {
match self {
Self::Aes128(a128) => (
a128.round_keys[0],
&a128.round_keys[1..10],
a128.round_keys[10],
),
Self::Aes256(a256) => (
a256.round_keys[0],
&a256.round_keys[1..14],
a256.round_keys[14],
),
pub(crate) fn ctr(&self, initial_counter: &[u8; 16], cipher_inout: &mut [u8]) {
// SAFETY: this crate requires the `aes` & `neon` cpu features
unsafe { self._ctr(initial_counter, cipher_inout) }
}

#[target_feature(enable = "aes,neon")]
unsafe fn _ctr(&self, initial_counter: &[u8; 16], cipher_inout: &mut [u8]) {
// counter and inc are big endian, so must be vrev32q_u8'd before use
let counter = vld1q_u8(initial_counter.as_ptr().cast());
let mut counter = vreinterpretq_u32_u8(vrev32q_u8(counter));

let inc = vsetq_lane_u8(1, vdupq_n_u8(0), 15);
let inc = vreinterpretq_u32_u8(vrev32q_u8(inc));

let mut by8 = cipher_inout.chunks_exact_mut(128);

for cipher8 in by8.by_ref() {
cpu::prefetch_rw(cipher8.as_ptr());
counter = vaddq_u32(counter, inc);
let b0 = vrev32q_u8(vreinterpretq_u8_u32(counter));
counter = vaddq_u32(counter, inc);
let b1 = vrev32q_u8(vreinterpretq_u8_u32(counter));
counter = vaddq_u32(counter, inc);
let b2 = vrev32q_u8(vreinterpretq_u8_u32(counter));
counter = vaddq_u32(counter, inc);
let b3 = vrev32q_u8(vreinterpretq_u8_u32(counter));
counter = vaddq_u32(counter, inc);
let b4 = vrev32q_u8(vreinterpretq_u8_u32(counter));
counter = vaddq_u32(counter, inc);
let b5 = vrev32q_u8(vreinterpretq_u8_u32(counter));
counter = vaddq_u32(counter, inc);
let b6 = vrev32q_u8(vreinterpretq_u8_u32(counter));
counter = vaddq_u32(counter, inc);
let b7 = vrev32q_u8(vreinterpretq_u8_u32(counter));

let (b0, b1, b2, b3, b4, b5, b6, b7) = match self {
Self::Aes128(a128) => {
_aes128_8_blocks(&a128.round_keys, b0, b1, b2, b3, b4, b5, b6, b7)
}
Self::Aes256(a256) => {
_aes256_8_blocks(&a256.round_keys, b0, b1, b2, b3, b4, b5, b6, b7)
}
};

let b0 = veorq_u8(vld1q_u8(cipher8.as_ptr().add(0).cast()), b0);
let b1 = veorq_u8(vld1q_u8(cipher8.as_ptr().add(16).cast()), b1);
let b2 = veorq_u8(vld1q_u8(cipher8.as_ptr().add(32).cast()), b2);
let b3 = veorq_u8(vld1q_u8(cipher8.as_ptr().add(48).cast()), b3);
let b4 = veorq_u8(vld1q_u8(cipher8.as_ptr().add(64).cast()), b4);
let b5 = veorq_u8(vld1q_u8(cipher8.as_ptr().add(80).cast()), b5);
let b6 = veorq_u8(vld1q_u8(cipher8.as_ptr().add(96).cast()), b6);
let b7 = veorq_u8(vld1q_u8(cipher8.as_ptr().add(112).cast()), b7);

vst1q_u8(cipher8.as_mut_ptr().add(0).cast(), b0);
vst1q_u8(cipher8.as_mut_ptr().add(16).cast(), b1);
vst1q_u8(cipher8.as_mut_ptr().add(32).cast(), b2);
vst1q_u8(cipher8.as_mut_ptr().add(48).cast(), b3);
vst1q_u8(cipher8.as_mut_ptr().add(64).cast(), b4);
vst1q_u8(cipher8.as_mut_ptr().add(80).cast(), b5);
vst1q_u8(cipher8.as_mut_ptr().add(96).cast(), b6);
vst1q_u8(cipher8.as_mut_ptr().add(112).cast(), b7);
}

let mut singles = by8.into_remainder().chunks_exact_mut(16);

for cipher in singles.by_ref() {
counter = vaddq_u32(counter, inc);
let block = vrev32q_u8(vreinterpretq_u8_u32(counter));

let block = match self {
Self::Aes128(a128) => _aes128_block(&a128.round_keys, block),
Self::Aes256(a256) => _aes256_block(&a256.round_keys, block),
};
let block = veorq_u8(vld1q_u8(cipher.as_ptr().cast()), block);
vst1q_u8(cipher.as_mut_ptr().cast(), block);
}

let cipher_inout = singles.into_remainder();
if !cipher_inout.is_empty() {
let mut cipher = [0u8; 16];
let len = cipher_inout.len();
debug_assert!(len < 16);
cipher[..len].copy_from_slice(cipher_inout);

counter = vaddq_u32(counter, inc);
let block = vrev32q_u8(vreinterpretq_u8_u32(counter));

let block = match self {
Self::Aes128(a128) => _aes128_block(&a128.round_keys, block),
Self::Aes256(a256) => _aes256_block(&a256.round_keys, block),
};

let block = veorq_u8(vld1q_u8(cipher.as_ptr().cast()), block);
vst1q_u8(cipher.as_mut_ptr().cast(), block);

cipher_inout.copy_from_slice(&cipher[..len]);
}
}
}
Expand Down Expand Up @@ -179,6 +264,13 @@ const RCON: [u32; 10] = [0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x1b, 0
#[target_feature(enable = "aes")]
unsafe fn aes128_block(round_keys: &[uint8x16_t; 11], block_inout: &mut [u8]) {
let block = vld1q_u8(block_inout.as_ptr() as *const _);
let block = _aes128_block(round_keys, block);
vst1q_u8(block_inout.as_mut_ptr() as *mut _, block);
}

#[target_feature(enable = "aes")]
#[inline]
unsafe fn _aes128_block(round_keys: &[uint8x16_t; 11], block: uint8x16_t) -> uint8x16_t {
let block = vaeseq_u8(block, round_keys[0]);
let block = vaesmcq_u8(block);
let block = vaeseq_u8(block, round_keys[1]);
Expand All @@ -198,13 +290,93 @@ unsafe fn aes128_block(round_keys: &[uint8x16_t; 11], block_inout: &mut [u8]) {
let block = vaeseq_u8(block, round_keys[8]);
let block = vaesmcq_u8(block);
let block = vaeseq_u8(block, round_keys[9]);
let block = veorq_u8(block, round_keys[10]);
vst1q_u8(block_inout.as_mut_ptr() as *mut _, block);
veorq_u8(block, round_keys[10])
}

macro_rules! round_8 {
($b0:ident, $b1:ident, $b2:ident, $b3:ident, $b4:ident, $b5:ident, $b6:ident, $b7:ident, $rk:expr) => {
let rk = $rk;
$b0 = vaeseq_u8($b0, rk);
$b0 = vaesmcq_u8($b0);
$b1 = vaeseq_u8($b1, rk);
$b1 = vaesmcq_u8($b1);
$b2 = vaeseq_u8($b2, rk);
$b2 = vaesmcq_u8($b2);
$b3 = vaeseq_u8($b3, rk);
$b3 = vaesmcq_u8($b3);
$b4 = vaeseq_u8($b4, rk);
$b4 = vaesmcq_u8($b4);
$b5 = vaeseq_u8($b5, rk);
$b5 = vaesmcq_u8($b5);
$b6 = vaeseq_u8($b6, rk);
$b6 = vaesmcq_u8($b6);
$b7 = vaeseq_u8($b7, rk);
$b7 = vaesmcq_u8($b7);
};
}

#[target_feature(enable = "aes")]
#[inline]
unsafe fn _aes128_8_blocks(
round_keys: &[uint8x16_t; 11],
mut b0: uint8x16_t,
mut b1: uint8x16_t,
mut b2: uint8x16_t,
mut b3: uint8x16_t,
mut b4: uint8x16_t,
mut b5: uint8x16_t,
mut b6: uint8x16_t,
mut b7: uint8x16_t,
) -> (
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
) {
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[0]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[1]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[2]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[3]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[4]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[5]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[6]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[7]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[8]);

let b0 = vaeseq_u8(b0, round_keys[9]);
let b1 = vaeseq_u8(b1, round_keys[9]);
let b2 = vaeseq_u8(b2, round_keys[9]);
let b3 = vaeseq_u8(b3, round_keys[9]);
let b4 = vaeseq_u8(b4, round_keys[9]);
let b5 = vaeseq_u8(b5, round_keys[9]);
let b6 = vaeseq_u8(b6, round_keys[9]);
let b7 = vaeseq_u8(b7, round_keys[9]);
(
veorq_u8(b0, round_keys[10]),
veorq_u8(b1, round_keys[10]),
veorq_u8(b2, round_keys[10]),
veorq_u8(b3, round_keys[10]),
veorq_u8(b4, round_keys[10]),
veorq_u8(b5, round_keys[10]),
veorq_u8(b6, round_keys[10]),
veorq_u8(b7, round_keys[10]),
)
}

#[target_feature(enable = "aes")]
unsafe fn aes256_block(round_keys: &[uint8x16_t; 15], block_inout: &mut [u8]) {
let block = vld1q_u8(block_inout.as_ptr() as *const _);
let block = _aes256_block(round_keys, block);
vst1q_u8(block_inout.as_mut_ptr() as *mut _, block);
}

#[target_feature(enable = "aes")]
#[inline]
unsafe fn _aes256_block(round_keys: &[uint8x16_t; 15], block: uint8x16_t) -> uint8x16_t {
let block = vaeseq_u8(block, round_keys[0]);
let block = vaesmcq_u8(block);
let block = vaeseq_u8(block, round_keys[1]);
Expand Down Expand Up @@ -232,8 +404,63 @@ unsafe fn aes256_block(round_keys: &[uint8x16_t; 15], block_inout: &mut [u8]) {
let block = vaeseq_u8(block, round_keys[12]);
let block = vaesmcq_u8(block);
let block = vaeseq_u8(block, round_keys[13]);
let block = veorq_u8(block, round_keys[14]);
vst1q_u8(block_inout.as_mut_ptr() as *mut _, block);
veorq_u8(block, round_keys[14])
}

#[target_feature(enable = "aes")]
#[inline]
unsafe fn _aes256_8_blocks(
round_keys: &[uint8x16_t; 15],
mut b0: uint8x16_t,
mut b1: uint8x16_t,
mut b2: uint8x16_t,
mut b3: uint8x16_t,
mut b4: uint8x16_t,
mut b5: uint8x16_t,
mut b6: uint8x16_t,
mut b7: uint8x16_t,
) -> (
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
uint8x16_t,
) {
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[0]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[1]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[2]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[3]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[4]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[5]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[6]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[7]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[8]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[9]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[10]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[11]);
round_8!(b0, b1, b2, b3, b4, b5, b6, b7, round_keys[12]);

let b0 = vaeseq_u8(b0, round_keys[13]);
let b1 = vaeseq_u8(b1, round_keys[13]);
let b2 = vaeseq_u8(b2, round_keys[13]);
let b3 = vaeseq_u8(b3, round_keys[13]);
let b4 = vaeseq_u8(b4, round_keys[13]);
let b5 = vaeseq_u8(b5, round_keys[13]);
let b6 = vaeseq_u8(b6, round_keys[13]);
let b7 = vaeseq_u8(b7, round_keys[13]);
(
veorq_u8(b0, round_keys[14]),
veorq_u8(b1, round_keys[14]),
veorq_u8(b2, round_keys[14]),
veorq_u8(b3, round_keys[14]),
veorq_u8(b4, round_keys[14]),
veorq_u8(b5, round_keys[14]),
veorq_u8(b6, round_keys[14]),
veorq_u8(b7, round_keys[14]),
)
}

#[cfg(test)]
Expand Down
29 changes: 29 additions & 0 deletions graviola/src/low/aarch64/aes_gcm.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,29 @@
// Written for Graviola by Joe Birr-Pixton, 2024.
// SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT-0

use crate::low::ghash::Ghash;
use crate::low::AesKey;

pub(crate) fn encrypt(
key: &AesKey,
ghash: &mut Ghash<'_>,
initial_counter: &[u8; 16],
aad: &[u8],
cipher_inout: &mut [u8],
) {
ghash.add(aad);
key.ctr(initial_counter, cipher_inout);
ghash.add(cipher_inout);
}

pub(crate) fn decrypt(
key: &AesKey,
ghash: &mut Ghash<'_>,
initial_counter: &[u8; 16],
aad: &[u8],
cipher_inout: &mut [u8],
) {
ghash.add(aad);
ghash.add(cipher_inout);
key.ctr(initial_counter, cipher_inout);
}
24 changes: 24 additions & 0 deletions graviola/src/low/aarch64/cpu.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,3 +113,27 @@ mod dit {
}
}
}

/// Read-only prefetch hint.
pub(in crate::low) fn prefetch_ro<T>(ptr: *const T) {
// SAFETY: inline assembly
unsafe {
core::arch::asm!(
"prfm pldl1strm, [{ptr}]",
ptr = in(reg) ptr,
options(readonly, nostack)
);
}
}

/// Read-write prefetch hint.
pub(in crate::low) fn prefetch_rw<T>(ptr: *const T) {
// SAFETY: inline assembly
unsafe {
core::arch::asm!(
"prfm pstl1keep, [{ptr}]",
ptr = in(reg) ptr,
options(readonly, nostack)
);
}
}
1 change: 1 addition & 0 deletions graviola/src/low/aarch64/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,7 @@
// SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT-0

pub(crate) mod aes;
pub(crate) mod aes_gcm;
pub(crate) mod bignum_add;
pub(crate) mod bignum_add_p256;
pub(crate) mod bignum_add_p384;
Expand Down
11 changes: 2 additions & 9 deletions graviola/src/low/aarch64/sha256.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
// Written for Graviola by Joe Birr-Pixton, 2024.
// SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT-0

use crate::low::aarch64::cpu;
use core::arch::aarch64::*;

pub(crate) fn sha256_compress_blocks(state: &mut [u32; 8], blocks: &[u8]) {
Expand Down Expand Up @@ -60,7 +61,7 @@ unsafe fn sha256(state: &mut [u32; 8], blocks: &[u8]) {
let state1_prev = state1;

// prefetch next block
prefetch(block.as_ptr().add(64));
cpu::prefetch_ro(block.as_ptr().add(64));

let msg0 = vld1q_u32(block[0..].as_ptr() as *const _);
let msg1 = vld1q_u32(block[16..].as_ptr() as *const _);
Expand Down Expand Up @@ -100,14 +101,6 @@ unsafe fn sha256(state: &mut [u32; 8], blocks: &[u8]) {
vst1q_u32(state[4..8].as_mut_ptr(), state1);
}

unsafe fn prefetch<T>(ptr: *const T) {
core::arch::asm!(
"prfm pldl1strm, [{ptr}]",
ptr = in(reg) ptr,
options(readonly, nostack)
);
}

#[repr(align(16))]
struct Aligned([u32; 64]);

Expand Down
Loading