diff options
| author | Hans Kratz <hans@appfour.com> | 2024-10-23 00:42:29 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-10-22 15:42:29 -0700 |
| commit | 7e162d19dd2e245dbba0e37fd12fe2cdaafdfed8 (patch) | |
| tree | 6deba2eccab0d851f6f5f7d4a61516e1315e3ee0 | |
| parent | 158e2409fe479cd509d1549ce629a40a1588f1bf (diff) | |
| download | rust-7e162d19dd2e245dbba0e37fd12fe2cdaafdfed8.tar.gz rust-7e162d19dd2e245dbba0e37fd12fe2cdaafdfed8.zip | |
rust-lang/portable-simd#443: Add armv7 neon mplementation for `Simd<u8, 16>::swizzle_dyn`
Use arm neon intrinsics to swizzle two u8x8 blocks with a u8x8x2 lookup table.
| -rw-r--r-- | crates/core_simd/src/swizzle_dyn.rs | 29 |
1 files changed, 29 insertions, 0 deletions
diff --git a/crates/core_simd/src/swizzle_dyn.rs b/crates/core_simd/src/swizzle_dyn.rs index eaf297ba3e3..0619404e5f7 100644 --- a/crates/core_simd/src/swizzle_dyn.rs +++ b/crates/core_simd/src/swizzle_dyn.rs @@ -57,6 +57,13 @@ where target_endian = "little" ))] 16 => transize(vqtbl1q_u8, self, idxs), + #[cfg(all( + target_arch = "arm", + target_feature = "v7", + target_feature = "neon", + target_endian = "little" + ))] + 16 => transize(armv7_neon_swizzle_u8x16, self, idxs), #[cfg(all(target_feature = "avx2", not(target_feature = "avx512vbmi")))] 32 => transize(avx2_pshufb, self, idxs), #[cfg(all(target_feature = "avx512vl", target_feature = "avx512vbmi"))] @@ -98,6 +105,28 @@ where } } +/// armv7 neon supports swizzling `u8x16` by swizzling two u8x8 blocks +/// with a u8x8x2 lookup table. +/// +/// # Safety +/// This requires armv7 neon to work +#[cfg(all( + target_arch = "arm", + target_feature = "v7", + target_feature = "neon", + target_endian = "little" +))] +unsafe fn armv7_neon_swizzle_u8x16(bytes: Simd<u8, 16>, idxs: Simd<u8, 16>) -> Simd<u8, 16> { + use core::arch::arm::{uint8x8x2_t, vcombine_u8, vget_high_u8, vget_low_u8, vtbl2_u8}; + // SAFETY: Caller promised arm neon support + unsafe { + let bytes = uint8x8x2_t(vget_low_u8(bytes.into()), vget_high_u8(bytes.into())); + let lo = vtbl2_u8(bytes, vget_low_u8(idxs.into())); + let hi = vtbl2_u8(bytes, vget_high_u8(idxs.into())); + vcombine_u8(lo, hi).into() + } +} + /// "vpshufb like it was meant to be" on AVX2 /// /// # Safety |