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

Implement AVX512BW 16-bit shift by immediate (srli_epi16) with simd_shr instead of LLVM intrinsics

Browse source
This commit is contained in:
Eduardo Sánchez Muñoz 2023-08-12 12:36:31 +02:00 • committed by Amanieu d'Antras
parent 0c0f72ee7f
commit 427bb149f0
1 changed files with 18 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

30
library/stdarch/crates/core_arch/src/x86/avx512bw.rs
View file

@ -5675,9 +5675,11 @@ pub unsafe fn _mm_maskz_srl_epi16(k: __mmask8, a: __m128i, count: __m128i) -> __
#[rustc_legacy_const_generics(1)]
pub unsafe fn _mm512_srli_epi16<const IMM8: u32>(a: __m512i) -> __m512i {
static_assert_uimm_bits!(IMM8, 8);
let a = a.as_i16x32();
let r = vpsrliw(a, IMM8);
transmute(r)
if IMM8 >= 16 {
_mm512_setzero_si512()
} else {
transmute(simd_shr(a.as_u16x32(), u16x32::splat(IMM8 as u16)))
}
}
/// Shift packed 16-bit integers in a right by imm8 while shifting in zeros, and store the results in dst using writemask k (elements are copied from src when the corresponding mask bit is not set).
@ -5693,9 +5695,12 @@ pub unsafe fn _mm512_mask_srli_epi16<const IMM8: u32>(
a: __m512i,
) -> __m512i {
static_assert_uimm_bits!(IMM8, 8);
let a = a.as_i16x32();
let shf = vpsrliw(a, IMM8);
transmute(simd_select_bitmask(k, shf, src.as_i16x32()))
let shf = if IMM8 >= 16 {
u16x32::splat(0)
} else {
simd_shr(a.as_u16x32(), u16x32::splat(IMM8 as u16))
};
transmute(simd_select_bitmask(k, shf, src.as_u16x32()))
}
/// Shift packed 16-bit integers in a right by imm8 while shifting in zeros, and store the results in dst using zeromask k (elements are zeroed out when the corresponding mask bit is not set).
@ -5708,10 +5713,13 @@ pub unsafe fn _mm512_mask_srli_epi16<const IMM8: u32>(
pub unsafe fn _mm512_maskz_srli_epi16<const IMM8: i32>(k: __mmask32, a: __m512i) -> __m512i {
static_assert_uimm_bits!(IMM8, 8);
//imm8 should be u32, it seems the document to verify is incorrect
let a = a.as_i16x32();
let shf = vpsrliw(a, IMM8 as u32);
let zero = _mm512_setzero_si512().as_i16x32();
transmute(simd_select_bitmask(k, shf, zero))
if IMM8 >= 16 {
_mm512_setzero_si512()
} else {
let shf = simd_shr(a.as_u16x32(), u16x32::splat(IMM8 as u16));
let zero = u16x32::splat(0);
transmute(simd_select_bitmask(k, shf, zero))
}
}
/// Shift packed 16-bit integers in a right by imm8 while shifting in zeros, and store the results in dst using writemask k (elements are copied from src when the corresponding mask bit is not set).
@ -9995,8 +10003,6 @@ extern "C" {
#[link_name = "llvm.x86.avx512.psrl.w.512"]
fn vpsrlw(a: i16x32, count: i16x8) -> i16x32;
#[link_name = "llvm.x86.avx512.psrli.w.512"]
fn vpsrliw(a: i16x32, imm8: u32) -> i16x32;
#[link_name = "llvm.x86.avx512.psrlv.w.512"]
fn vpsrlvw(a: i16x32, b: i16x32) -> i16x32;