#![allow(dead_code, unused_imports)] macro_rules! convert_fn { (fn $name:ident($var:ident : $vartype:ty) -> $restype:ty { if feature("f16c") { $f16c:expr } else { $fallback:expr }}) => { #[inline] pub(crate) fn $name($var: $vartype) -> $restype { // Use CPU feature detection if using std #[cfg(all( feature = "use-intrinsics", feature = "std", any(target_arch = "x86", target_arch = "x86_64"), not(target_feature = "f16c") ))] { if is_x86_feature_detected!("f16c") { $f16c } else { $fallback } } // Use intrinsics directly when a compile target or using no_std #[cfg(all( feature = "use-intrinsics", any(target_arch = "x86", target_arch = "x86_64"), target_feature = "f16c" ))] { $f16c } // Fallback to software #[cfg(any( not(feature = "use-intrinsics"), not(any(target_arch = "x86", target_arch = "x86_64")), all(not(feature = "std"), not(target_feature = "f16c")) ))] { $fallback } } }; } convert_fn! { fn f32_to_f16(f: f32) -> u16 { if feature("f16c") { unsafe { x86::f32_to_f16_x86_f16c(f) } } else { f32_to_f16_fallback(f) } } } convert_fn! { fn f64_to_f16(f: f64) -> u16 { if feature("f16c") { unsafe { x86::f32_to_f16_x86_f16c(f as f32) } } else { f64_to_f16_fallback(f) } } } convert_fn! { fn f16_to_f32(i: u16) -> f32 { if feature("f16c") { unsafe { x86::f16_to_f32_x86_f16c(i) } } else { f16_to_f32_fallback(i) } } } convert_fn! { fn f16_to_f64(i: u16) -> f64 { if feature("f16c") { unsafe { x86::f16_to_f32_x86_f16c(i) as f64 } } else { f16_to_f64_fallback(i) } } } // TODO: While SIMD versions are faster, further improvements can be made by doing runtime feature // detection once at beginning of convert slice method, rather than per chunk convert_fn! { fn f32x4_to_f16x4(f: &[f32]) -> [u16; 4] { if feature("f16c") { unsafe { x86::f32x4_to_f16x4_x86_f16c(f) } } else { f32x4_to_f16x4_fallback(f) } } } convert_fn! { fn f16x4_to_f32x4(i: &[u16]) -> [f32; 4] { if feature("f16c") { unsafe { x86::f16x4_to_f32x4_x86_f16c(i) } } else { f16x4_to_f32x4_fallback(i) } } } convert_fn! { fn f64x4_to_f16x4(f: &[f64]) -> [u16; 4] { if feature("f16c") { unsafe { x86::f64x4_to_f16x4_x86_f16c(f) } } else { f64x4_to_f16x4_fallback(f) } } } convert_fn! { fn f16x4_to_f64x4(i: &[u16]) -> [f64; 4] { if feature("f16c") { unsafe { x86::f16x4_to_f64x4_x86_f16c(i) } } else { f16x4_to_f64x4_fallback(i) } } } /////////////// Fallbacks //////////////// // In the below functions, round to nearest, with ties to even. // Let us call the most significant bit that will be shifted out the round_bit. // // Round up if either // a) Removed part > tie. // (mantissa & round_bit) != 0 && (mantissa & (round_bit - 1)) != 0 // b) Removed part == tie, and retained part is odd. // (mantissa & round_bit) != 0 && (mantissa & (2 * round_bit)) != 0 // (If removed part == tie and retained part is even, do not round up.) // These two conditions can be combined into one: // (mantissa & round_bit) != 0 && (mantissa & ((round_bit - 1) | (2 * round_bit))) != 0 // which can be simplified into // (mantissa & round_bit) != 0 && (mantissa & (3 * round_bit - 1)) != 0 fn f32_to_f16_fallback(value: f32) -> u16 { // Convert to raw bytes let x = value.to_bits(); // Extract IEEE754 components let sign = x & 0x8000_0000u32; let exp = x & 0x7F80_0000u32; let man = x & 0x007F_FFFFu32; // Check for all exponent bits being set, which is Infinity or NaN if exp == 0x7F80_0000u32 { // Set mantissa MSB for NaN (and also keep shifted mantissa bits) let nan_bit = if man == 0 { 0 } else { 0x0200u32 }; return ((sign >> 16) | 0x7C00u32 | nan_bit | (man >> 13)) as u16; } // The number is normalized, start assembling half precision version let half_sign = sign >> 16; // Unbias the exponent, then bias for half precision let unbiased_exp = ((exp >> 23) as i32) - 127; let half_exp = unbiased_exp + 15; // Check for exponent overflow, return +infinity if half_exp >= 0x1F { return (half_sign | 0x7C00u32) as u16; } // Check for underflow if half_exp <= 0 { // Check mantissa for what we can do if 14 - half_exp > 24 { // No rounding possibility, so this is a full underflow, return signed zero return half_sign as u16; } // Don't forget about hidden leading mantissa bit when assembling mantissa let man = man | 0x0080_0000u32; let mut half_man = man >> (14 - half_exp); // Check for rounding (see comment above functions) let round_bit = 1 << (13 - half_exp); if (man & round_bit) != 0 && (man & (3 * round_bit - 1)) != 0 { half_man += 1; } // No exponent for subnormals return (half_sign | half_man) as u16; } // Rebias the exponent let half_exp = (half_exp as u32) << 10; let half_man = man >> 13; // Check for rounding (see comment above functions) let round_bit = 0x0000_1000u32; if (man & round_bit) != 0 && (man & (3 * round_bit - 1)) != 0 { // Round it ((half_sign | half_exp | half_man) + 1) as u16 } else { (half_sign | half_exp | half_man) as u16 } } fn f64_to_f16_fallback(value: f64) -> u16 { // Convert to raw bytes, truncating the last 32-bits of mantissa; that precision will always // be lost on half-precision. let val = value.to_bits(); let x = (val >> 32) as u32; // Extract IEEE754 components let sign = x & 0x8000_0000u32; let exp = x & 0x7FF0_0000u32; let man = x & 0x000F_FFFFu32; // Check for all exponent bits being set, which is Infinity or NaN if exp == 0x7FF0_0000u32 { // Set mantissa MSB for NaN (and also keep shifted mantissa bits). // We also have to check the last 32 bits. let nan_bit = if man == 0 && (val as u32 == 0) { 0 } else { 0x0200u32 }; return ((sign >> 16) | 0x7C00u32 | nan_bit | (man >> 10)) as u16; } // The number is normalized, start assembling half precision version let half_sign = sign >> 16; // Unbias the exponent, then bias for half precision let unbiased_exp = ((exp >> 20) as i64) - 1023; let half_exp = unbiased_exp + 15; // Check for exponent overflow, return +infinity if half_exp >= 0x1F { return (half_sign | 0x7C00u32) as u16; } // Check for underflow if half_exp <= 0 { // Check mantissa for what we can do if 10 - half_exp > 21 { // No rounding possibility, so this is a full underflow, return signed zero return half_sign as u16; } // Don't forget about hidden leading mantissa bit when assembling mantissa let man = man | 0x0010_0000u32; let mut half_man = man >> (11 - half_exp); // Check for rounding (see comment above functions) let round_bit = 1 << (10 - half_exp); if (man & round_bit) != 0 && (man & (3 * round_bit - 1)) != 0 { half_man += 1; } // No exponent for subnormals return (half_sign | half_man) as u16; } // Rebias the exponent let half_exp = (half_exp as u32) << 10; let half_man = man >> 10; // Check for rounding (see comment above functions) let round_bit = 0x0000_0200u32; if (man & round_bit) != 0 && (man & (3 * round_bit - 1)) != 0 { // Round it ((half_sign | half_exp | half_man) + 1) as u16 } else { (half_sign | half_exp | half_man) as u16 } } fn f16_to_f32_fallback(i: u16) -> f32 { // Check for signed zero if i & 0x7FFFu16 == 0 { return f32::from_bits((i as u32) << 16); } let half_sign = (i & 0x8000u16) as u32; let half_exp = (i & 0x7C00u16) as u32; let half_man = (i & 0x03FFu16) as u32; // Check for an infinity or NaN when all exponent bits set if half_exp == 0x7C00u32 { // Check for signed infinity if mantissa is zero if half_man == 0 { return f32::from_bits((half_sign << 16) | 0x7F80_0000u32); } else { // NaN, keep current mantissa but also set most significiant mantissa bit return f32::from_bits((half_sign << 16) | 0x7FC0_0000u32 | (half_man << 13)); } } // Calculate single-precision components with adjusted exponent let sign = half_sign << 16; // Unbias exponent let unbiased_exp = ((half_exp as i32) >> 10) - 15; // Check for subnormals, which will be normalized by adjusting exponent if half_exp == 0 { // Calculate how much to adjust the exponent by let e = (half_man as u16).leading_zeros() - 6; // Rebias and adjust exponent let exp = (127 - 15 - e) << 23; let man = (half_man << (14 + e)) & 0x7F_FF_FFu32; return f32::from_bits(sign | exp | man); } // Rebias exponent for a normalized normal let exp = ((unbiased_exp + 127) as u32) << 23; let man = (half_man & 0x03FFu32) << 13; f32::from_bits(sign | exp | man) } fn f16_to_f64_fallback(i: u16) -> f64 { // Check for signed zero if i & 0x7FFFu16 == 0 { return f64::from_bits((i as u64) << 48); } let half_sign = (i & 0x8000u16) as u64; let half_exp = (i & 0x7C00u16) as u64; let half_man = (i & 0x03FFu16) as u64; // Check for an infinity or NaN when all exponent bits set if half_exp == 0x7C00u64 { // Check for signed infinity if mantissa is zero if half_man == 0 { return f64::from_bits((half_sign << 48) | 0x7FF0_0000_0000_0000u64); } else { // NaN, keep current mantissa but also set most significiant mantissa bit return f64::from_bits((half_sign << 48) | 0x7FF8_0000_0000_0000u64 | (half_man << 42)); } } // Calculate double-precision components with adjusted exponent let sign = half_sign << 48; // Unbias exponent let unbiased_exp = ((half_exp as i64) >> 10) - 15; // Check for subnormals, which will be normalized by adjusting exponent if half_exp == 0 { // Calculate how much to adjust the exponent by let e = (half_man as u16).leading_zeros() - 6; // Rebias and adjust exponent let exp = ((1023 - 15 - e) as u64) << 52; let man = (half_man << (43 + e)) & 0xF_FFFF_FFFF_FFFFu64; return f64::from_bits(sign | exp | man); } // Rebias exponent for a normalized normal let exp = ((unbiased_exp + 1023) as u64) << 52; let man = (half_man & 0x03FFu64) << 42; f64::from_bits(sign | exp | man) } #[inline] fn f16x4_to_f32x4_fallback(v: &[u16]) -> [f32; 4] { debug_assert!(v.len() >= 4); [ f16_to_f32_fallback(v[0]), f16_to_f32_fallback(v[1]), f16_to_f32_fallback(v[2]), f16_to_f32_fallback(v[3]), ] } #[inline] fn f32x4_to_f16x4_fallback(v: &[f32]) -> [u16; 4] { debug_assert!(v.len() >= 4); [ f32_to_f16_fallback(v[0]), f32_to_f16_fallback(v[1]), f32_to_f16_fallback(v[2]), f32_to_f16_fallback(v[3]), ] } #[inline] fn f16x4_to_f64x4_fallback(v: &[u16]) -> [f64; 4] { debug_assert!(v.len() >= 4); [ f16_to_f64_fallback(v[0]), f16_to_f64_fallback(v[1]), f16_to_f64_fallback(v[2]), f16_to_f64_fallback(v[3]), ] } #[inline] fn f64x4_to_f16x4_fallback(v: &[f64]) -> [u16; 4] { debug_assert!(v.len() >= 4); [ f64_to_f16_fallback(v[0]), f64_to_f16_fallback(v[1]), f64_to_f16_fallback(v[2]), f64_to_f16_fallback(v[3]), ] } /////////////// x86/x86_64 f16c //////////////// #[cfg(all( feature = "use-intrinsics", any(target_arch = "x86", target_arch = "x86_64") ))] mod x86 { use core::{mem::MaybeUninit, ptr}; #[cfg(target_arch = "x86")] use core::arch::x86::{__m128, __m128i, _mm_cvtph_ps, _mm_cvtps_ph, _MM_FROUND_TO_NEAREST_INT}; #[cfg(target_arch = "x86_64")] use core::arch::x86_64::{ __m128, __m128i, _mm_cvtph_ps, _mm_cvtps_ph, _MM_FROUND_TO_NEAREST_INT, }; #[target_feature(enable = "f16c")] #[inline] pub(super) unsafe fn f16_to_f32_x86_f16c(i: u16) -> f32 { let mut vec = MaybeUninit::<__m128i>::zeroed(); vec.as_mut_ptr().cast::().write(i); let retval = _mm_cvtph_ps(vec.assume_init()); *(&retval as *const __m128).cast() } #[target_feature(enable = "f16c")] #[inline] pub(super) unsafe fn f32_to_f16_x86_f16c(f: f32) -> u16 { let mut vec = MaybeUninit::<__m128>::zeroed(); vec.as_mut_ptr().cast::().write(f); let retval = _mm_cvtps_ph(vec.assume_init(), _MM_FROUND_TO_NEAREST_INT); *(&retval as *const __m128i).cast() } #[target_feature(enable = "f16c")] #[inline] pub(super) unsafe fn f16x4_to_f32x4_x86_f16c(v: &[u16]) -> [f32; 4] { debug_assert!(v.len() >= 4); let mut vec = MaybeUninit::<__m128i>::zeroed(); ptr::copy_nonoverlapping(v.as_ptr(), vec.as_mut_ptr().cast(), 4); let retval = _mm_cvtph_ps(vec.assume_init()); *(&retval as *const __m128).cast() } #[target_feature(enable = "f16c")] #[inline] pub(super) unsafe fn f32x4_to_f16x4_x86_f16c(v: &[f32]) -> [u16; 4] { debug_assert!(v.len() >= 4); let mut vec = MaybeUninit::<__m128>::uninit(); ptr::copy_nonoverlapping(v.as_ptr(), vec.as_mut_ptr().cast(), 4); let retval = _mm_cvtps_ph(vec.assume_init(), _MM_FROUND_TO_NEAREST_INT); *(&retval as *const __m128i).cast() } #[target_feature(enable = "f16c")] #[inline] pub(super) unsafe fn f16x4_to_f64x4_x86_f16c(v: &[u16]) -> [f64; 4] { debug_assert!(v.len() >= 4); let mut vec = MaybeUninit::<__m128i>::zeroed(); ptr::copy_nonoverlapping(v.as_ptr(), vec.as_mut_ptr().cast(), 4); let retval = _mm_cvtph_ps(vec.assume_init()); let array = *(&retval as *const __m128).cast::<[f32; 4]>(); // Let compiler vectorize this regular cast for now. // TODO: investigate auto-detecting sse2/avx convert features [ array[0] as f64, array[1] as f64, array[2] as f64, array[3] as f64, ] } #[target_feature(enable = "f16c")] #[inline] pub(super) unsafe fn f64x4_to_f16x4_x86_f16c(v: &[f64]) -> [u16; 4] { debug_assert!(v.len() >= 4); // Let compiler vectorize this regular cast for now. // TODO: investigate auto-detecting sse2/avx convert features let v = [v[0] as f32, v[1] as f32, v[2] as f32, v[3] as f32]; let mut vec = MaybeUninit::<__m128>::uninit(); ptr::copy_nonoverlapping(v.as_ptr(), vec.as_mut_ptr().cast(), 4); let retval = _mm_cvtps_ph(vec.assume_init(), _MM_FROUND_TO_NEAREST_INT); *(&retval as *const __m128i).cast() } }