| 1 |
use core::{mem::MaybeUninit, ptr}; |
| 2 |
|
| 3 |
#[cfg(target_arch = "x86")] |
| 4 |
use core::arch::x86::{ |
| 5 |
__m128, __m128i, __m256, _mm256_cvtph_ps, _mm256_cvtps_ph, _mm_cvtph_ps, |
| 6 |
_MM_FROUND_TO_NEAREST_INT, |
| 7 |
}; |
| 8 |
#[cfg(target_arch = "x86_64")] |
| 9 |
use core::arch::x86_64::{ |
| 10 |
__m128, __m128i, __m256, _mm256_cvtph_ps, _mm256_cvtps_ph, _mm_cvtph_ps, _mm_cvtps_ph, |
| 11 |
_MM_FROUND_TO_NEAREST_INT, |
| 12 |
}; |
| 13 |
|
| 14 |
#[cfg(target_arch = "x86")] |
| 15 |
use core::arch::x86::_mm_cvtps_ph; |
| 16 |
|
| 17 |
use super::convert_chunked_slice_8; |
| 18 |
|
| 19 |
/////////////// x86/x86_64 f16c //////////////// |
| 20 |
|
| 21 |
#[target_feature(enable = "f16c")] |
| 22 |
#[inline] |
| 23 |
pub(super) unsafe fn f16_to_f32_x86_f16c(i: u16) -> f32 { |
| 24 |
let mut vec = MaybeUninit::<__m128i>::zeroed(); |
| 25 |
vec.as_mut_ptr().cast::<u16>().write(i); |
| 26 |
let retval = _mm_cvtph_ps(vec.assume_init()); |
| 27 |
*(&retval as *const __m128).cast() |
| 28 |
} |
| 29 |
|
| 30 |
#[target_feature(enable = "f16c")] |
| 31 |
#[inline] |
| 32 |
pub(super) unsafe fn f32_to_f16_x86_f16c(f: f32) -> u16 { |
| 33 |
let mut vec = MaybeUninit::<__m128>::zeroed(); |
| 34 |
vec.as_mut_ptr().cast::<f32>().write(f); |
| 35 |
let retval = _mm_cvtps_ph(vec.assume_init(), _MM_FROUND_TO_NEAREST_INT); |
| 36 |
*(&retval as *const __m128i).cast() |
| 37 |
} |
| 38 |
|
| 39 |
#[target_feature(enable = "f16c")] |
| 40 |
#[inline] |
| 41 |
pub(super) unsafe fn f16x4_to_f32x4_x86_f16c(v: &[u16; 4]) -> [f32; 4] { |
| 42 |
let mut vec = MaybeUninit::<__m128i>::zeroed(); |
| 43 |
ptr::copy_nonoverlapping(v.as_ptr(), vec.as_mut_ptr().cast(), 4); |
| 44 |
let retval = _mm_cvtph_ps(vec.assume_init()); |
| 45 |
*(&retval as *const __m128).cast() |
| 46 |
} |
| 47 |
|
| 48 |
#[target_feature(enable = "f16c")] |
| 49 |
#[inline] |
| 50 |
pub(super) unsafe fn f32x4_to_f16x4_x86_f16c(v: &[f32; 4]) -> [u16; 4] { |
| 51 |
let mut vec = MaybeUninit::<__m128>::uninit(); |
| 52 |
ptr::copy_nonoverlapping(v.as_ptr(), vec.as_mut_ptr().cast(), 4); |
| 53 |
let retval = _mm_cvtps_ph(vec.assume_init(), _MM_FROUND_TO_NEAREST_INT); |
| 54 |
*(&retval as *const __m128i).cast() |
| 55 |
} |
| 56 |
|
| 57 |
#[target_feature(enable = "f16c")] |
| 58 |
#[inline] |
| 59 |
pub(super) unsafe fn f16x4_to_f64x4_x86_f16c(v: &[u16; 4]) -> [f64; 4] { |
| 60 |
let array = f16x4_to_f32x4_x86_f16c(v); |
| 61 |
// Let compiler vectorize this regular cast for now. |
| 62 |
// TODO: investigate auto-detecting sse2/avx convert features |
| 63 |
[ |
| 64 |
array[0] as f64, |
| 65 |
array[1] as f64, |
| 66 |
array[2] as f64, |
| 67 |
array[3] as f64, |
| 68 |
] |
| 69 |
} |
| 70 |
|
| 71 |
#[target_feature(enable = "f16c")] |
| 72 |
#[inline] |
| 73 |
pub(super) unsafe fn f64x4_to_f16x4_x86_f16c(v: &[f64; 4]) -> [u16; 4] { |
| 74 |
// Let compiler vectorize this regular cast for now. |
| 75 |
// TODO: investigate auto-detecting sse2/avx convert features |
| 76 |
let v = [v[0] as f32, v[1] as f32, v[2] as f32, v[3] as f32]; |
| 77 |
f32x4_to_f16x4_x86_f16c(&v) |
| 78 |
} |
| 79 |
|
| 80 |
#[target_feature(enable = "f16c")] |
| 81 |
#[inline] |
| 82 |
pub(super) unsafe fn f16x8_to_f32x8_x86_f16c(v: &[u16; 8]) -> [f32; 8] { |
| 83 |
let mut vec = MaybeUninit::<__m128i>::zeroed(); |
| 84 |
ptr::copy_nonoverlapping(v.as_ptr(), vec.as_mut_ptr().cast(), 8); |
| 85 |
let retval = _mm256_cvtph_ps(vec.assume_init()); |
| 86 |
*(&retval as *const __m256).cast() |
| 87 |
} |
| 88 |
|
| 89 |
#[target_feature(enable = "f16c")] |
| 90 |
#[inline] |
| 91 |
pub(super) unsafe fn f32x8_to_f16x8_x86_f16c(v: &[f32; 8]) -> [u16; 8] { |
| 92 |
let mut vec = MaybeUninit::<__m256>::uninit(); |
| 93 |
ptr::copy_nonoverlapping(v.as_ptr(), vec.as_mut_ptr().cast(), 8); |
| 94 |
let retval = _mm256_cvtps_ph(vec.assume_init(), _MM_FROUND_TO_NEAREST_INT); |
| 95 |
*(&retval as *const __m128i).cast() |
| 96 |
} |
| 97 |
|
| 98 |
#[target_feature(enable = "f16c")] |
| 99 |
#[inline] |
| 100 |
pub(super) unsafe fn f16x8_to_f64x8_x86_f16c(v: &[u16; 8]) -> [f64; 8] { |
| 101 |
let array = f16x8_to_f32x8_x86_f16c(v); |
| 102 |
// Let compiler vectorize this regular cast for now. |
| 103 |
// TODO: investigate auto-detecting sse2/avx convert features |
| 104 |
[ |
| 105 |
array[0] as f64, |
| 106 |
array[1] as f64, |
| 107 |
array[2] as f64, |
| 108 |
array[3] as f64, |
| 109 |
array[4] as f64, |
| 110 |
array[5] as f64, |
| 111 |
array[6] as f64, |
| 112 |
array[7] as f64, |
| 113 |
] |
| 114 |
} |
| 115 |
|
| 116 |
#[target_feature(enable = "f16c")] |
| 117 |
#[inline] |
| 118 |
pub(super) unsafe fn f64x8_to_f16x8_x86_f16c(v: &[f64; 8]) -> [u16; 8] { |
| 119 |
// Let compiler vectorize this regular cast for now. |
| 120 |
// TODO: investigate auto-detecting sse2/avx convert features |
| 121 |
let v = [ |
| 122 |
v[0] as f32, |
| 123 |
v[1] as f32, |
| 124 |
v[2] as f32, |
| 125 |
v[3] as f32, |
| 126 |
v[4] as f32, |
| 127 |
v[5] as f32, |
| 128 |
v[6] as f32, |
| 129 |
v[7] as f32, |
| 130 |
]; |
| 131 |
f32x8_to_f16x8_x86_f16c(&v) |
| 132 |
} |