logos-storage/plonky2
1
0
Fork 0
mirror of https://github.com/logos-storage/plonky2.git synced 2026-01-07 08:13:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

Daniel PR comments + delete throughput figures

Browse Source
This commit is contained in:
Jakub Nabaglo 2021-09-08 11:41:13 -07:00
parent 87f5201e6f
commit 5d69d85d42
1 changed files with 1 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
src/field/packed_crandall_avx2.rs
View File

@ -303,10 +303,6 @@ unsafe fn add_no_canonicalize_64_64s_s(x: __m256i, y_s: __m256i) -> __m256i {
res_s
}
// Theoretical throughput (Skylake)
// Scalar version (compiled): 1.75 cycles/(op * word)
// Scalar version (optimized asm): 1 cycle/(op * word)
// Below (256-bit vectors): .75 cycles/(op * word)
#[inline]
unsafe fn add(x: __m256i, y: __m256i) -> __m256i {
let y_s = shift(y);
@ -314,10 +310,6 @@ unsafe fn add(x: __m256i, y: __m256i) -> __m256i {
shift(res_s)
}
// Theoretical throughput (Skylake)
// Scalar version (compiled): 1.75 cycles/(op * word)
// Scalar version (optimized asm): 1 cycle/(op * word)
// Below (256-bit vectors): .75 cycles/(op * word)
#[inline]
unsafe fn sub(x: __m256i, y: __m256i) -> __m256i {
let mut y_s = shift(y);
@ -330,20 +322,10 @@ unsafe fn sub(x: __m256i, y: __m256i) -> __m256i {
res
}
// Theoretical throughput (Skylake)
// Scalar version (compiled): 1 cycle/(op * word)
// Scalar version (optimized asm): .5 cycles/(op * word)
// Below (256-bit vectors): .42 cycles/(op * word)
#[inline]
unsafe fn neg(y: __m256i) -> __m256i {
let y_s = shift(y);
let field_order_s = shift(field_order());
// mask is -1 if sub will underflow (y > field_order) else 0.
let mask = _mm256_cmpgt_epi64(y_s, field_order_s);
let wrapback_amt = _mm256_and_si256(mask, epsilon()); // -FIELD_ORDER if underflow else 0.
let res_wrapped = _mm256_sub_epi64(field_order_s, y_s);
let res = _mm256_sub_epi64(res_wrapped, wrapback_amt);
res
_mm256_sub_epi64(shift(field_order()), canonicalize_s(y_s))
}
/// Full 64-bit by 64-bit multiplication. This emulated multiplication is 1.5x slower than the