user0/rust
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Replace addsub variations

Browse source
This commit is contained in:
Tobias Decking 2024-06-16 17:47:52 +02:00 • committed by Amanieu d'Antras
parent 3e95b6133a
commit f0f309bc52
1 changed files with 25 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

54
library/stdarch/crates/core_arch/src/x86/fma.rs
View file

@ -19,7 +19,7 @@
//! [wiki_fma]: https://en.wikipedia.org/wiki/Fused_multiply-accumulate
use crate::core_arch::x86::*;
use crate::intrinsics::simd::{simd_fma, simd_insert, simd_neg};
use crate::intrinsics::simd::{simd_fma, simd_insert, simd_neg, simd_shuffle};
use crate::intrinsics::{fmaf32, fmaf64};
#[cfg(test)]
@ -119,7 +119,9 @@ pub unsafe fn _mm_fmadd_ss(a: __m128, b: __m128, c: __m128) -> __m128 {
#[cfg_attr(test, assert_instr(vfmaddsub))]
#[stable(feature = "simd_x86", since = "1.27.0")]
pub unsafe fn _mm_fmaddsub_pd(a: __m128d, b: __m128d, c: __m128d) -> __m128d {
vfmaddsubpd(a, b, c)
let add = simd_fma(a, b, c);
let sub = simd_fma(a, b, simd_neg(c));
simd_shuffle!(add, sub, [2, 1])
}
/// Multiplies packed double-precision (64-bit) floating-point elements in `a`
@ -132,7 +134,9 @@ pub unsafe fn _mm_fmaddsub_pd(a: __m128d, b: __m128d, c: __m128d) -> __m128d {
#[cfg_attr(test, assert_instr(vfmaddsub))]
#[stable(feature = "simd_x86", since = "1.27.0")]
pub unsafe fn _mm256_fmaddsub_pd(a: __m256d, b: __m256d, c: __m256d) -> __m256d {
vfmaddsubpd256(a, b, c)
let add = simd_fma(a, b, c);
let sub = simd_fma(a, b, simd_neg(c));
simd_shuffle!(add, sub, [4, 1, 6, 3])
}
/// Multiplies packed single-precision (32-bit) floating-point elements in `a`
@ -145,7 +149,9 @@ pub unsafe fn _mm256_fmaddsub_pd(a: __m256d, b: __m256d, c: __m256d) -> __m256d
#[cfg_attr(test, assert_instr(vfmaddsub))]
#[stable(feature = "simd_x86", since = "1.27.0")]
pub unsafe fn _mm_fmaddsub_ps(a: __m128, b: __m128, c: __m128) -> __m128 {
vfmaddsubps(a, b, c)
let add = simd_fma(a, b, c);
let sub = simd_fma(a, b, simd_neg(c));
simd_shuffle!(add, sub, [4, 1, 6, 3])
}
/// Multiplies packed single-precision (32-bit) floating-point elements in `a`
@ -158,7 +164,9 @@ pub unsafe fn _mm_fmaddsub_ps(a: __m128, b: __m128, c: __m128) -> __m128 {
#[cfg_attr(test, assert_instr(vfmaddsub))]
#[stable(feature = "simd_x86", since = "1.27.0")]
pub unsafe fn _mm256_fmaddsub_ps(a: __m256, b: __m256, c: __m256) -> __m256 {
vfmaddsubps256(a, b, c)
let add = simd_fma(a, b, c);
let sub = simd_fma(a, b, simd_neg(c));
simd_shuffle!(add, sub, [8, 1, 10, 3, 12, 5, 14, 7])
}
/// Multiplies packed double-precision (64-bit) floating-point elements in `a`
@ -255,7 +263,9 @@ pub unsafe fn _mm_fmsub_ss(a: __m128, b: __m128, c: __m128) -> __m128 {
#[cfg_attr(test, assert_instr(vfmsubadd))]
#[stable(feature = "simd_x86", since = "1.27.0")]
pub unsafe fn _mm_fmsubadd_pd(a: __m128d, b: __m128d, c: __m128d) -> __m128d {
vfmsubaddpd(a, b, c)
let add = simd_fma(a, b, c);
let sub = simd_fma(a, b, simd_neg(c));
simd_shuffle!(add, sub, [0, 3])
}
/// Multiplies packed double-precision (64-bit) floating-point elements in `a`
@ -268,7 +278,9 @@ pub unsafe fn _mm_fmsubadd_pd(a: __m128d, b: __m128d, c: __m128d) -> __m128d {
#[cfg_attr(test, assert_instr(vfmsubadd))]
#[stable(feature = "simd_x86", since = "1.27.0")]
pub unsafe fn _mm256_fmsubadd_pd(a: __m256d, b: __m256d, c: __m256d) -> __m256d {
vfmsubaddpd256(a, b, c)
let add = simd_fma(a, b, c);
let sub = simd_fma(a, b, simd_neg(c));
simd_shuffle!(add, sub, [0, 5, 2, 7])
}
/// Multiplies packed single-precision (32-bit) floating-point elements in `a`
@ -281,7 +293,9 @@ pub unsafe fn _mm256_fmsubadd_pd(a: __m256d, b: __m256d, c: __m256d) -> __m256d
#[cfg_attr(test, assert_instr(vfmsubadd))]
#[stable(feature = "simd_x86", since = "1.27.0")]
pub unsafe fn _mm_fmsubadd_ps(a: __m128, b: __m128, c: __m128) -> __m128 {
vfmsubaddps(a, b, c)
let add = simd_fma(a, b, c);
let sub = simd_fma(a, b, simd_neg(c));
simd_shuffle!(add, sub, [0, 5, 2, 7])
}
/// Multiplies packed single-precision (32-bit) floating-point elements in `a`
@ -294,7 +308,9 @@ pub unsafe fn _mm_fmsubadd_ps(a: __m128, b: __m128, c: __m128) -> __m128 {
#[cfg_attr(test, assert_instr(vfmsubadd))]
#[stable(feature = "simd_x86", since = "1.27.0")]
pub unsafe fn _mm256_fmsubadd_ps(a: __m256, b: __m256, c: __m256) -> __m256 {
vfmsubaddps256(a, b, c)
let add = simd_fma(a, b, c);
let sub = simd_fma(a, b, simd_neg(c));
simd_shuffle!(add, sub, [0, 9, 2, 11, 4, 13, 6, 15])
}
/// Multiplies packed double-precision (64-bit) floating-point elements in `a`
@ -471,26 +487,6 @@ pub unsafe fn _mm_fnmsub_ss(a: __m128, b: __m128, c: __m128) -> __m128 {
)
}
#[allow(improper_ctypes)]
extern "C" {
#[link_name = "llvm.x86.fma.vfmaddsub.pd"]
fn vfmaddsubpd(a: __m128d, b: __m128d, c: __m128d) -> __m128d;
#[link_name = "llvm.x86.fma.vfmaddsub.pd.256"]
fn vfmaddsubpd256(a: __m256d, b: __m256d, c: __m256d) -> __m256d;
#[link_name = "llvm.x86.fma.vfmaddsub.ps"]
fn vfmaddsubps(a: __m128, b: __m128, c: __m128) -> __m128;
#[link_name = "llvm.x86.fma.vfmaddsub.ps.256"]
fn vfmaddsubps256(a: __m256, b: __m256, c: __m256) -> __m256;
#[link_name = "llvm.x86.fma.vfmsubadd.pd"]
fn vfmsubaddpd(a: __m128d, b: __m128d, c: __m128d) -> __m128d;
#[link_name = "llvm.x86.fma.vfmsubadd.pd.256"]
fn vfmsubaddpd256(a: __m256d, b: __m256d, c: __m256d) -> __m256d;
#[link_name = "llvm.x86.fma.vfmsubadd.ps"]
fn vfmsubaddps(a: __m128, b: __m128, c: __m128) -> __m128;
#[link_name = "llvm.x86.fma.vfmsubadd.ps.256"]
fn vfmsubaddps256(a: __m256, b: __m256, c: __m256) -> __m256;
}
#[cfg(test)]
mod tests {