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

Add fast type conversion methods where possible #117

Open
wants to merge 5 commits into
base: main
Choose a base branch
from
Open

Add fast type conversion methods where possible #117

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8a89d31
Add fast type conversion methods where possible
shssoichiro Oct 27, 2022
7bee32d
Use wide round functions to handle no_std case
shssoichiro Oct 27, 2022
6cc8681
Make round/trunc_int functions call to new functions
shssoichiro Oct 28, 2022
fedd7cd
Add since property to deprecation notices
shssoichiro Oct 28, 2022
f5c6956
Fix fallback compiler errors
shssoichiro Oct 28, 2022
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
214 changes: 154 additions & 60 deletions src/f32x4_.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,13 +4,13 @@ pick! {
if #[cfg(target_feature="sse")] {
#[derive(Default, Clone, Copy, PartialEq)]
#[repr(C, align(16))]
pub struct f32x4 { sse: m128 }
pub struct f32x4 { pub(crate) sse: m128 }
} else if #[cfg(target_feature="simd128")] {
use core::arch::wasm32::*;

#[derive(Clone, Copy)]
#[repr(transparent)]
pub struct f32x4 { simd: v128 }
pub struct f32x4 { pub(crate) simd: v128 }

impl Default for f32x4 {
fn default() -> Self {
Expand All @@ -26,7 +26,7 @@ pick! {
} else {
#[derive(Default, Clone, Copy, PartialEq)]
#[repr(C, align(16))]
pub struct f32x4 { arr: [f32;4] }
pub struct f32x4 { pub(crate) arr: [f32;4] }
}
}

Expand Down Expand Up @@ -664,85 +664,39 @@ impl f32x4 {
/// values you get implementation defined behavior.
#[inline]
#[must_use]
#[deprecated(since = "0.7.6", note = "use `to_i32x4_round_fast` instead")]
pub fn fast_round_int(self) -> i32x4 {
pick! {
if #[cfg(target_feature="sse2")] {
cast(convert_to_i32_m128i_from_m128(self.sse))
} else {
self.round_int()
}
}
self.to_i32x4_round_fast()
}

/// Rounds each lane into an integer. This saturates out of range values and
/// turns NaNs into 0. Use `fast_round_int` for a faster implementation that
/// doesn't handle out of range values or NaNs.
#[inline]
#[must_use]
#[deprecated(since = "0.7.6", note = "use `to_i32x4_round` instead")]
pub fn round_int(self) -> i32x4 {
pick! {
if #[cfg(target_feature="sse2")] {
// Based on: https://github.com/v8/v8/blob/210987a552a2bf2a854b0baa9588a5959ff3979d/src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h#L489-L504
let non_nan_mask = self.cmp_eq(self);
let non_nan = self & non_nan_mask;
let flip_to_max: i32x4 = cast(self.cmp_ge(Self::splat(2147483648.0)));
let cast: i32x4 = cast(convert_to_i32_m128i_from_m128(non_nan.sse));
flip_to_max ^ cast
} else if #[cfg(target_feature="simd128")] {
cast(Self { simd: i32x4_trunc_sat_f32x4(f32x4_nearest(self.simd)) })
} else {
let rounded: [f32; 4] = cast(self.round());
cast([
rounded[0] as i32,
rounded[1] as i32,
rounded[2] as i32,
rounded[3] as i32,
])
}
}
self.to_i32x4_round()
}

/// Truncates each lane into an integer. This is a faster implementation than
/// `trunc_int`, but it doesn't handle out of range values or NaNs. For those
/// values you get implementation defined behavior.
#[inline]
#[must_use]
#[deprecated(since = "0.7.6", note = "use `to_i32x4_truncate_fast` instead")]
pub fn fast_trunc_int(self) -> i32x4 {
pick! {
if #[cfg(target_feature="sse2")] {
cast(truncate_m128_to_m128i(self.sse))
} else {
self.trunc_int()
}
}
self.to_i32x4_truncate_fast()
}

/// Truncates each lane into an integer. This saturates out of range values
/// and turns NaNs into 0. Use `fast_trunc_int` for a faster implementation
/// that doesn't handle out of range values or NaNs.
#[inline]
#[must_use]
#[deprecated(since = "0.7.6", note = "use `to_i32x4_truncate` instead")]
pub fn trunc_int(self) -> i32x4 {
pick! {
if #[cfg(target_feature="sse2")] {
// Based on: https://github.com/v8/v8/blob/210987a552a2bf2a854b0baa9588a5959ff3979d/src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h#L489-L504
let non_nan_mask = self.cmp_eq(self);
let non_nan = self & non_nan_mask;
let flip_to_max: i32x4 = cast(self.cmp_ge(Self::splat(2147483648.0)));
let cast: i32x4 = cast(truncate_m128_to_m128i(non_nan.sse));
flip_to_max ^ cast
} else if #[cfg(target_feature="simd128")] {
cast(Self { simd: i32x4_trunc_sat_f32x4(self.simd) })
} else {
let n: [f32;4] = cast(self);
cast([
n[0] as i32,
n[1] as i32,
n[2] as i32,
n[3] as i32,
])
}
}
self.to_i32x4_truncate()
}
#[inline]
#[must_use]
Expand Down Expand Up @@ -1037,7 +991,7 @@ impl f32x4 {

// Find quadrant
let y = (xa * TWO_OVER_PI).round();
let q: i32x4 = y.round_int();
let q: i32x4 = y.to_i32x4_round();

let x = y.mul_neg_add(DP3F, y.mul_neg_add(DP2F, y.mul_neg_add(DP1F, xa)));

Expand Down Expand Up @@ -1417,7 +1371,7 @@ impl f32x4 {
);

let ee = e1 + e2 + e3;
let ei = cast::<_, i32x4>(ee.round_int());
let ei = cast::<_, i32x4>(ee.to_i32x4_round());
let ej = cast::<_, i32x4>(ei + (cast::<_, i32x4>(z) >> 23));

let overflow = cast::<_, f32x4>(ej.cmp_gt(i32x4::splat(0x0FF)))
Expand Down Expand Up @@ -1451,7 +1405,7 @@ impl f32x4 {
// Y into an integer
let yi = y.cmp_eq(y.round());
// Is y odd?
let y_odd = cast::<_, i32x4>(y.round_int() << 31).round_float();
let y_odd = cast::<_, i32x4>(y.to_i32x4_round() << 31).round_float();

let z1 =
yi.blend(z | y_odd, self.cmp_eq(Self::ZERO).blend(z, Self::nan_pow()));
Expand Down Expand Up @@ -1484,4 +1438,144 @@ impl f32x4 {
pub fn as_array_ref(&self) -> &[f32; 4] {
cast_ref(self)
}

/// Converts the first two f32 elements within this struct to f64 elements.
///
/// The remaining elements are discarded.
#[inline]
#[must_use]
pub fn to_f64x2(self) -> f64x2 {
pick! {
if #[cfg(target_feature="sse2")] {
f64x2 { sse: convert_to_m128d_from_lower2_m128(self.sse) }
} else {
let arr = self.to_array();
f64x2::new([
f64::from(arr[0]),
f64::from(arr[1]),
])
}
}
}

/// Converts the f32 elements within this struct to f64 elements.
#[inline]
#[must_use]
pub fn to_f64x4(self) -> f64x4 {
pick! {
if #[cfg(target_feature="avx")] {
f64x4 { avx: convert_to_m256d_from_m128(self.sse) }
} else {
let arr = self.to_array();
f64x4::new([
f64::from(arr[0]),
f64::from(arr[1]),
f64::from(arr[2]),
f64::from(arr[3]),
])
}
}
}

/// Converts the f32 elements within this struct to i32 elements.
///
/// The decimal portions of the values are truncated.
///
/// This is a faster implementation than `to_i32x4_truncate`,
/// but it doesn't handle out of range values or NaNs. For those
/// values you get implementation defined behavior.
#[inline]
#[must_use]
pub fn to_i32x4_truncate_fast(self) -> i32x4 {
pick! {
if #[cfg(target_feature="sse2")] {
i32x4 { sse: truncate_m128_to_m128i(self.sse) }
} else {
self.to_i32x4_truncate()
}
}
}

/// Converts the f32 elements within this struct to i32 elements.
///
/// The decimal portions of the values are truncated.
///
/// This saturates out of range values and turns NaNs into 0.
/// Use `to_i32x4_truncate_fast` for a faster implementation
/// that doesn't handle out of range values or NaNs.
#[inline]
#[must_use]
pub fn to_i32x4_truncate(self) -> i32x4 {
pick! {
if #[cfg(target_feature="sse2")] {
// Based on: https://github.com/v8/v8/blob/210987a552a2bf2a854b0baa9588a5959ff3979d/src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h#L489-L504
let non_nan_mask = self.cmp_eq(self);
let non_nan = self & non_nan_mask;
let flip_to_max: i32x4 = cast(self.cmp_ge(Self::splat(2147483648.0)));
let cast: i32x4 = cast(truncate_m128_to_m128i(non_nan.sse));
flip_to_max ^ cast
} else if #[cfg(target_feature="simd128")] {
cast(Self { simd: i32x4_trunc_sat_f32x4(self.simd) })
} else {
let n: [f32;4] = cast(self);
cast([
n[0] as i32,
n[1] as i32,
n[2] as i32,
n[3] as i32,
])
}
}
}

/// Converts the f32 elements within this struct to i32 elements.
///
/// The decimal portions of the values are rounded to the nearest integer.
///
/// This is a faster implementation than `to_i32x4_round`,
/// but it doesn't handle out of range values or NaNs. For those
/// values you get implementation defined behavior.
#[inline]
#[must_use]
pub fn to_i32x4_round_fast(self) -> i32x4 {
pick! {
if #[cfg(target_feature="sse2")] {
cast(convert_to_i32_m128i_from_m128(self.sse))
} else {
self.to_i32x4_round()
}
}
}

/// Converts the f32 elements within this struct to i32 elements.
///
/// The decimal portions of the values are rounded to the nearest integer.
///
/// Rounds each lane into an integer. This saturates out of range values and
/// turns NaNs into 0. Use `to_i32x4_round_fast` for a faster implementation that
/// doesn't handle out of range values or NaNs.
#[inline]
#[must_use]
pub fn to_i32x4_round(self) -> i32x4 {
pick! {
if #[cfg(target_feature="sse2")] {
// Based on: https://github.com/v8/v8/blob/210987a552a2bf2a854b0baa9588a5959ff3979d/src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h#L489-L504
let non_nan_mask = self.cmp_eq(self);
let non_nan = self & non_nan_mask;
let flip_to_max: i32x4 = cast(self.cmp_ge(Self::splat(2147483648.0)));
let cast: i32x4 = cast(convert_to_i32_m128i_from_m128(non_nan.sse));
flip_to_max ^ cast
} else if #[cfg(target_feature="simd128")] {
cast(Self { simd: i32x4_trunc_sat_f32x4(f32x4_nearest(self.simd)) })
} else {
let rounded: [f32; 4] = cast(self.round());
cast([
rounded[0] as i32,
rounded[1] as i32,
rounded[2] as i32,
rounded[3] as i32,
])
}
}
}
}
Loading