| 1 |
/* |
| 2 |
* Copyright (c) 2023. |
| 3 |
* |
| 4 |
* This software is free software; |
| 5 |
* |
| 6 |
* You can redistribute it or modify it under terms of the MIT, Apache License or Zlib license |
| 7 |
*/ |
| 8 |
|
| 9 |
#![cfg(all(feature = "neon", target_arch = "aarch64"))] |
| 10 |
// TODO can this be extended to armv7 |
| 11 |
|
| 12 |
//! This module provides unsafe ways to do some things |
| 13 |
#![allow(clippy::wildcard_imports)] |
| 14 |
|
| 15 |
use core::arch::aarch64::*; |
| 16 |
use core::ops::{Add, AddAssign, BitOr, BitOrAssign, Mul, MulAssign, Sub}; |
| 17 |
|
| 18 |
pub type VecType = int32x4x2_t; |
| 19 |
|
| 20 |
pub unsafe fn loadu(src: *const i32) -> VecType { |
| 21 |
vld1q_s32_x2(src as *const _) |
| 22 |
} |
| 23 |
|
| 24 |
/// An abstraction of an AVX ymm register that |
| 25 |
///allows some things to not look ugly |
| 26 |
#[derive(Clone, Copy)] |
| 27 |
pub struct YmmRegister { |
| 28 |
/// An AVX register |
| 29 |
pub(crate) mm256: VecType |
| 30 |
} |
| 31 |
|
| 32 |
impl YmmRegister { |
| 33 |
#[inline] |
| 34 |
pub unsafe fn load(src: *const i32) -> Self { |
| 35 |
loadu(src).into() |
| 36 |
} |
| 37 |
|
| 38 |
#[inline] |
| 39 |
pub fn map2(self, other: Self, f: impl Fn(int32x4_t, int32x4_t) -> int32x4_t) -> Self { |
| 40 |
let m0 = f(self.mm256.0, other.mm256.0); |
| 41 |
let m1 = f(self.mm256.1, other.mm256.1); |
| 42 |
|
| 43 |
YmmRegister { |
| 44 |
mm256: int32x4x2_t(m0, m1) |
| 45 |
} |
| 46 |
} |
| 47 |
|
| 48 |
#[inline] |
| 49 |
pub fn all_zero(self) -> bool { |
| 50 |
unsafe { |
| 51 |
let both = vorrq_s32(self.mm256.0, self.mm256.1); |
| 52 |
let both_unsigned = vreinterpretq_u32_s32(both); |
| 53 |
0 == vmaxvq_u32(both_unsigned) |
| 54 |
} |
| 55 |
} |
| 56 |
|
| 57 |
#[inline] |
| 58 |
pub fn const_shl<const N: i32>(self) -> Self { |
| 59 |
// Ensure that we logically shift left |
| 60 |
unsafe { |
| 61 |
let m0 = vreinterpretq_s32_u32(vshlq_n_u32::<N>(vreinterpretq_u32_s32(self.mm256.0))); |
| 62 |
let m1 = vreinterpretq_s32_u32(vshlq_n_u32::<N>(vreinterpretq_u32_s32(self.mm256.1))); |
| 63 |
|
| 64 |
YmmRegister { |
| 65 |
mm256: int32x4x2_t(m0, m1) |
| 66 |
} |
| 67 |
} |
| 68 |
} |
| 69 |
|
| 70 |
#[inline] |
| 71 |
pub fn const_shra<const N: i32>(self) -> Self { |
| 72 |
unsafe { |
| 73 |
let i0 = vshrq_n_s32::<N>(self.mm256.0); |
| 74 |
let i1 = vshrq_n_s32::<N>(self.mm256.1); |
| 75 |
|
| 76 |
YmmRegister { |
| 77 |
mm256: int32x4x2_t(i0, i1) |
| 78 |
} |
| 79 |
} |
| 80 |
} |
| 81 |
} |
| 82 |
|
| 83 |
impl<T> Add<T> for YmmRegister |
| 84 |
where |
| 85 |
T: Into<Self> |
| 86 |
{ |
| 87 |
type Output = YmmRegister; |
| 88 |
|
| 89 |
#[inline] |
| 90 |
fn add(self, rhs: T) -> Self::Output { |
| 91 |
let rhs = rhs.into(); |
| 92 |
unsafe { self.map2(rhs, |a, b| vaddq_s32(a, b)) } |
| 93 |
} |
| 94 |
} |
| 95 |
|
| 96 |
impl<T> Sub<T> for YmmRegister |
| 97 |
where |
| 98 |
T: Into<Self> |
| 99 |
{ |
| 100 |
type Output = YmmRegister; |
| 101 |
|
| 102 |
#[inline] |
| 103 |
fn sub(self, rhs: T) -> Self::Output { |
| 104 |
let rhs = rhs.into(); |
| 105 |
unsafe { self.map2(rhs, |a, b| vsubq_s32(a, b)) } |
| 106 |
} |
| 107 |
} |
| 108 |
|
| 109 |
impl<T> AddAssign<T> for YmmRegister |
| 110 |
where |
| 111 |
T: Into<Self> |
| 112 |
{ |
| 113 |
#[inline] |
| 114 |
fn add_assign(&mut self, rhs: T) { |
| 115 |
let rhs: Self = rhs.into(); |
| 116 |
*self = *self + rhs; |
| 117 |
} |
| 118 |
} |
| 119 |
|
| 120 |
impl<T> Mul<T> for YmmRegister |
| 121 |
where |
| 122 |
T: Into<Self> |
| 123 |
{ |
| 124 |
type Output = YmmRegister; |
| 125 |
|
| 126 |
#[inline] |
| 127 |
fn mul(self, rhs: T) -> Self::Output { |
| 128 |
let rhs = rhs.into(); |
| 129 |
unsafe { self.map2(rhs, |a, b| vmulq_s32(a, b)) } |
| 130 |
} |
| 131 |
} |
| 132 |
|
| 133 |
impl<T> MulAssign<T> for YmmRegister |
| 134 |
where |
| 135 |
T: Into<Self> |
| 136 |
{ |
| 137 |
#[inline] |
| 138 |
fn mul_assign(&mut self, rhs: T) { |
| 139 |
let rhs: Self = rhs.into(); |
| 140 |
*self = *self * rhs; |
| 141 |
} |
| 142 |
} |
| 143 |
|
| 144 |
impl<T> BitOr<T> for YmmRegister |
| 145 |
where |
| 146 |
T: Into<Self> |
| 147 |
{ |
| 148 |
type Output = YmmRegister; |
| 149 |
|
| 150 |
#[inline] |
| 151 |
fn bitor(self, rhs: T) -> Self::Output { |
| 152 |
let rhs = rhs.into(); |
| 153 |
unsafe { self.map2(rhs, |a, b| vorrq_s32(a, b)) } |
| 154 |
} |
| 155 |
} |
| 156 |
|
| 157 |
impl<T> BitOrAssign<T> for YmmRegister |
| 158 |
where |
| 159 |
T: Into<Self> |
| 160 |
{ |
| 161 |
#[inline] |
| 162 |
fn bitor_assign(&mut self, rhs: T) { |
| 163 |
let rhs: Self = rhs.into(); |
| 164 |
*self = *self | rhs; |
| 165 |
} |
| 166 |
} |
| 167 |
|
| 168 |
impl From<i32> for YmmRegister { |
| 169 |
#[inline] |
| 170 |
fn from(val: i32) -> Self { |
| 171 |
unsafe { |
| 172 |
let dup = vdupq_n_s32(val); |
| 173 |
|
| 174 |
YmmRegister { |
| 175 |
mm256: int32x4x2_t(dup, dup) |
| 176 |
} |
| 177 |
} |
| 178 |
} |
| 179 |
} |
| 180 |
|
| 181 |
impl From<VecType> for YmmRegister { |
| 182 |
#[inline] |
| 183 |
fn from(mm256: VecType) -> Self { |
| 184 |
YmmRegister { mm256 } |
| 185 |
} |
| 186 |
} |
| 187 |
|
| 188 |
#[allow(clippy::too_many_arguments)] |
| 189 |
#[inline] |
| 190 |
unsafe fn transpose4( |
| 191 |
v0: &mut int32x4_t, v1: &mut int32x4_t, v2: &mut int32x4_t, v3: &mut int32x4_t |
| 192 |
) { |
| 193 |
let w0 = vtrnq_s32( |
| 194 |
vreinterpretq_s32_s64(vtrn1q_s64( |
| 195 |
vreinterpretq_s64_s32(*v0), |
| 196 |
vreinterpretq_s64_s32(*v2) |
| 197 |
)), |
| 198 |
vreinterpretq_s32_s64(vtrn1q_s64( |
| 199 |
vreinterpretq_s64_s32(*v1), |
| 200 |
vreinterpretq_s64_s32(*v3) |
| 201 |
)) |
| 202 |
); |
| 203 |
let w1 = vtrnq_s32( |
| 204 |
vreinterpretq_s32_s64(vtrn2q_s64( |
| 205 |
vreinterpretq_s64_s32(*v0), |
| 206 |
vreinterpretq_s64_s32(*v2) |
| 207 |
)), |
| 208 |
vreinterpretq_s32_s64(vtrn2q_s64( |
| 209 |
vreinterpretq_s64_s32(*v1), |
| 210 |
vreinterpretq_s64_s32(*v3) |
| 211 |
)) |
| 212 |
); |
| 213 |
|
| 214 |
*v0 = w0.0; |
| 215 |
*v1 = w0.1; |
| 216 |
*v2 = w1.0; |
| 217 |
*v3 = w1.1; |
| 218 |
} |
| 219 |
|
| 220 |
/// Transpose an array of 8 by 8 i32 |
| 221 |
/// Arm has dedicated interleave/transpose instructions |
| 222 |
/// we: |
| 223 |
/// 1. Transpose the upper left and lower right quadrants |
| 224 |
/// 2. Swap and transpose the upper right and lower left quadrants |
| 225 |
#[allow(clippy::too_many_arguments)] |
| 226 |
#[inline] |
| 227 |
pub unsafe fn transpose( |
| 228 |
v0: &mut YmmRegister, v1: &mut YmmRegister, v2: &mut YmmRegister, v3: &mut YmmRegister, |
| 229 |
v4: &mut YmmRegister, v5: &mut YmmRegister, v6: &mut YmmRegister, v7: &mut YmmRegister |
| 230 |
) { |
| 231 |
use core::mem::swap; |
| 232 |
|
| 233 |
let ul0 = &mut v0.mm256.0; |
| 234 |
let ul1 = &mut v1.mm256.0; |
| 235 |
let ul2 = &mut v2.mm256.0; |
| 236 |
let ul3 = &mut v3.mm256.0; |
| 237 |
|
| 238 |
let ur0 = &mut v0.mm256.1; |
| 239 |
let ur1 = &mut v1.mm256.1; |
| 240 |
let ur2 = &mut v2.mm256.1; |
| 241 |
let ur3 = &mut v3.mm256.1; |
| 242 |
|
| 243 |
let ll0 = &mut v4.mm256.0; |
| 244 |
let ll1 = &mut v5.mm256.0; |
| 245 |
let ll2 = &mut v6.mm256.0; |
| 246 |
let ll3 = &mut v7.mm256.0; |
| 247 |
|
| 248 |
let lr0 = &mut v4.mm256.1; |
| 249 |
let lr1 = &mut v5.mm256.1; |
| 250 |
let lr2 = &mut v6.mm256.1; |
| 251 |
let lr3 = &mut v7.mm256.1; |
| 252 |
|
| 253 |
swap(ur0, ll0); |
| 254 |
swap(ur1, ll1); |
| 255 |
swap(ur2, ll2); |
| 256 |
swap(ur3, ll3); |
| 257 |
|
| 258 |
transpose4(ul0, ul1, ul2, ul3); |
| 259 |
|
| 260 |
transpose4(ur0, ur1, ur2, ur3); |
| 261 |
|
| 262 |
transpose4(ll0, ll1, ll2, ll3); |
| 263 |
|
| 264 |
transpose4(lr0, lr1, lr2, lr3); |
| 265 |
} |
| 266 |
|
| 267 |
#[cfg(test)] |
| 268 |
mod tests { |
| 269 |
use super::*; |
| 270 |
|
| 271 |
#[test] |
| 272 |
fn test_transpose() { |
| 273 |
fn get_val(i: usize, j: usize) -> i32 { |
| 274 |
((i * 8) / (j + 1)) as i32 |
| 275 |
} |
| 276 |
unsafe { |
| 277 |
let mut vals: [i32; 8 * 8] = [0; 8 * 8]; |
| 278 |
|
| 279 |
for i in 0..8 { |
| 280 |
for j in 0..8 { |
| 281 |
// some order-dependent value of i and j |
| 282 |
let value = get_val(i, j); |
| 283 |
vals[i * 8 + j] = value; |
| 284 |
} |
| 285 |
} |
| 286 |
|
| 287 |
let mut regs: [YmmRegister; 8] = core::mem::transmute(vals); |
| 288 |
let mut reg0 = regs[0]; |
| 289 |
let mut reg1 = regs[1]; |
| 290 |
let mut reg2 = regs[2]; |
| 291 |
let mut reg3 = regs[3]; |
| 292 |
let mut reg4 = regs[4]; |
| 293 |
let mut reg5 = regs[5]; |
| 294 |
let mut reg6 = regs[6]; |
| 295 |
let mut reg7 = regs[7]; |
| 296 |
|
| 297 |
transpose( |
| 298 |
&mut reg0, &mut reg1, &mut reg2, &mut reg3, &mut reg4, &mut reg5, &mut reg6, |
| 299 |
&mut reg7 |
| 300 |
); |
| 301 |
|
| 302 |
regs[0] = reg0; |
| 303 |
regs[1] = reg1; |
| 304 |
regs[2] = reg2; |
| 305 |
regs[3] = reg3; |
| 306 |
regs[4] = reg4; |
| 307 |
regs[5] = reg5; |
| 308 |
regs[6] = reg6; |
| 309 |
regs[7] = reg7; |
| 310 |
|
| 311 |
let vals_from_reg: [i32; 8 * 8] = core::mem::transmute(regs); |
| 312 |
|
| 313 |
for i in 0..8 { |
| 314 |
for j in 0..i { |
| 315 |
let orig = vals[i * 8 + j]; |
| 316 |
vals[i * 8 + j] = vals[j * 8 + i]; |
| 317 |
vals[j * 8 + i] = orig; |
| 318 |
} |
| 319 |
} |
| 320 |
|
| 321 |
for i in 0..8 { |
| 322 |
for j in 0..8 { |
| 323 |
assert_eq!(vals[j * 8 + i], get_val(i, j)); |
| 324 |
assert_eq!(vals_from_reg[j * 8 + i], get_val(i, j)); |
| 325 |
} |
| 326 |
} |
| 327 |
|
| 328 |
assert_eq!(vals, vals_from_reg); |
| 329 |
} |
| 330 |
} |
| 331 |
} |