sad/raptorq
1
0
Fork 0
mirror of https://github.com/cberner/raptorq.git synced 2024-06-27 09:19:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

Use vshrq_n_u8 in neon optimizations

Browse Source 
Now that https://github.com/rust-lang/rust/issues/82072 is fixed this
intrinsic works and improves mulassign & FMA performance by ~30% on
Raspberry Pi 3 B+. End to end speedup is ~5%
This commit is contained in:
Christopher Berner 2021-10-16 17:37:01 -07:00
parent 1a4a62e64a
commit 88959e05e6
1 changed files with 2 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
src/octets.rs
View File

@ -290,11 +290,7 @@ unsafe fn mulassign_scalar_neon(octets: &mut [u8], scalar: &Octet) {
let self_vec = vld1q_u8(self_neon_ptr.add(i * mem::size_of::<uint8x16_t>()));
let low = vandq_u8(self_vec, mask);
let low_result = vqtbl1q_u8(low_table, low);
// TODO: replace with vshrq. Right now it doesn't optimize correctly (it generates 16 ubfx instructions)
// let hi = vshrq_n_u8(self_vec, 4);
let mut hi: u128 = mem::transmute(self_vec);
hi >>= 4;
let hi: uint8x16_t = mem::transmute(hi);
let hi = vshrq_n_u8(self_vec, 4);
let hi = vandq_u8(hi, mask);
let hi_result = vqtbl1q_u8(hi_table, hi);
let result = veorq_u8(hi_result, low_result);
@ -463,11 +459,7 @@ unsafe fn fused_addassign_mul_scalar_neon(octets: &mut [u8], other: &[u8], scala
let other_vec = vld1q_u8(other_neon_ptr.add(i * mem::size_of::<uint8x16_t>()));
let low = vandq_u8(other_vec, mask);
let low_result = vqtbl1q_u8(low_table, low);
// TODO: replace with vshrq. Right now it doesn't optimize correctly (it generates 16 ubfx instructions)
// let hi = vshrq_n_u8(other_vec, 4);
let mut hi: u128 = mem::transmute(other_vec);
hi >>= 4;
let hi: uint8x16_t = mem::transmute(hi);
let hi = vshrq_n_u8(other_vec, 4);
let hi = vandq_u8(hi, mask);
let hi_result = vqtbl1q_u8(hi_table, hi);
let other_vec = veorq_u8(hi_result, low_result);