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

Implement AVX512BW 16-bit shift by immediate (srai_epi16) with simd_shr instead of LLVM intrinsics

Browse source
This commit is contained in:
Eduardo Sánchez Muñoz 2023-08-12 12:42:11 +02:00 • committed by Amanieu d'Antras
parent 427bb149f0
commit 7f5b70be72
1 changed files with 10 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

35
library/stdarch/crates/core_arch/src/x86/avx512bw.rs
View file

@ -5996,9 +5996,7 @@ pub unsafe fn _mm_maskz_sra_epi16(k: __mmask8, a: __m128i, count: __m128i) -> __
#[rustc_legacy_const_generics(1)]
pub unsafe fn _mm512_srai_epi16<const IMM8: u32>(a: __m512i) -> __m512i {
static_assert_uimm_bits!(IMM8, 8);
let a = a.as_i16x32();
let r = vpsraiw(a, IMM8);
transmute(r)
transmute(simd_shr(a.as_i16x32(), i16x32::splat(IMM8.min(15) as i16)))
}
/// Shift packed 16-bit integers in a right by imm8 while shifting in sign bits, and store the results in dst using writemask k (elements are copied from src when the corresponding mask bit is not set).
@ -6014,8 +6012,7 @@ pub unsafe fn _mm512_mask_srai_epi16<const IMM8: u32>(
a: __m512i,
) -> __m512i {
static_assert_uimm_bits!(IMM8, 8);
let a = a.as_i16x32();
let shf = vpsraiw(a, IMM8);
let shf = simd_shr(a.as_i16x32(), i16x32::splat(IMM8.min(15) as i16));
transmute(simd_select_bitmask(k, shf, src.as_i16x32()))
}
@ -6028,9 +6025,8 @@ pub unsafe fn _mm512_mask_srai_epi16<const IMM8: u32>(
#[rustc_legacy_const_generics(2)]
pub unsafe fn _mm512_maskz_srai_epi16<const IMM8: u32>(k: __mmask32, a: __m512i) -> __m512i {
static_assert_uimm_bits!(IMM8, 8);
let a = a.as_i16x32();
let shf = vpsraiw(a, IMM8);
let zero = _mm512_setzero_si512().as_i16x32();
let shf = simd_shr(a.as_i16x32(), i16x32::splat(IMM8.min(15) as i16));
let zero = i16x32::splat(0);
transmute(simd_select_bitmask(k, shf, zero))
}
@ -6047,8 +6043,7 @@ pub unsafe fn _mm256_mask_srai_epi16<const IMM8: u32>(
a: __m256i,
) -> __m256i {
static_assert_uimm_bits!(IMM8, 8);
let imm8 = IMM8 as i32;
let r = psraiw256(a.as_i16x16(), imm8);
let r = simd_shr(a.as_i16x16(), i16x16::splat(IMM8.min(15) as i16));
transmute(simd_select_bitmask(k, r, src.as_i16x16()))
}
@ -6061,9 +6056,8 @@ pub unsafe fn _mm256_mask_srai_epi16<const IMM8: u32>(
#[rustc_legacy_const_generics(2)]
pub unsafe fn _mm256_maskz_srai_epi16<const IMM8: u32>(k: __mmask16, a: __m256i) -> __m256i {
static_assert_uimm_bits!(IMM8, 8);
let imm8 = IMM8 as i32;
let r = psraiw256(a.as_i16x16(), imm8);
let zero = _mm256_setzero_si256().as_i16x16();
let r = simd_shr(a.as_i16x16(), i16x16::splat(IMM8.min(15) as i16));
let zero = i16x16::splat(0);
transmute(simd_select_bitmask(k, r, zero))
}
@ -6080,8 +6074,7 @@ pub unsafe fn _mm_mask_srai_epi16<const IMM8: u32>(
a: __m128i,
) -> __m128i {
static_assert_uimm_bits!(IMM8, 8);
let imm8 = IMM8 as i32;
let r = psraiw128(a.as_i16x8(), imm8);
let r = simd_shr(a.as_i16x8(), i16x8::splat(IMM8.min(15) as i16));
transmute(simd_select_bitmask(k, r, src.as_i16x8()))
}
@ -6094,9 +6087,8 @@ pub unsafe fn _mm_mask_srai_epi16<const IMM8: u32>(
#[rustc_legacy_const_generics(2)]
pub unsafe fn _mm_maskz_srai_epi16<const IMM8: u32>(k: __mmask8, a: __m128i) -> __m128i {
static_assert_uimm_bits!(IMM8, 8);
let imm8 = IMM8 as i32;
let r = psraiw128(a.as_i16x8(), imm8);
let zero = _mm_setzero_si128().as_i16x8();
let r = simd_shr(a.as_i16x8(), i16x8::splat(IMM8.min(15) as i16));
let zero = i16x8::splat(0);
transmute(simd_select_bitmask(k, r, zero))
}
@ -10013,13 +10005,6 @@ extern "C" {
#[link_name = "llvm.x86.avx512.psra.w.512"]
fn vpsraw(a: i16x32, count: i16x8) -> i16x32;
#[link_name = "llvm.x86.avx512.psrai.w.512"]
fn vpsraiw(a: i16x32, imm8: u32) -> i16x32;
#[link_name = "llvm.x86.avx2.psrai.w"]
fn psraiw256(a: i16x16, imm8: i32) -> i16x16;
#[link_name = "llvm.x86.sse2.psrai.w"]
fn psraiw128(a: i16x8, imm8: i32) -> i16x8;
#[link_name = "llvm.x86.avx512.psrav.w.512"]
fn vpsravw(a: i16x32, count: i16x32) -> i16x32;