0

block_class: add aarch64 SIMD

Speedup of 2.3x for this particular function. Can probably be
improved somewhat still.
This commit is contained in:
Nicolas F
2021-05-11 11:57:52 +02:00
parent 86963c5de9
commit 772d73ce04

View File

@@ -21,6 +21,9 @@
#if defined(__i386__) || defined(__x86_64__)
#include <immintrin.h>
#endif
#if defined(__aarch64__)
#include <arm_neon.h>
#endif
bool block_class_is_subset(
mc_block_t block,
@@ -39,6 +42,18 @@ bool block_class_is_subset(
}
}
#endif
#if defined(__aarch64__)
for (; i / 8 < block_class_len / 8; i += 8) {
const uint16x8_t block_class_vec = vld1q_u16(
(uint16_t*)&block_class[i]);
const uint16x8_t block_vec = vmovq_n_u16(block);
const uint16x8_t block_cmp = vceqq_u16(block_vec, (uint16x8_t) block_class_vec);
if(vgetq_lane_s64((int64x2_t) block_cmp, 0) |
vgetq_lane_s64((int64x2_t) block_cmp, 1)) {
return true;
}
}
#endif
#ifdef __MMX__
for (; i / 4 < block_class_len / 4; i += 4) {
const __m64 block_class_vec = _mm_cvtsi64_m64(