github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-11-15 17:02:14 +03:00
Code Issues Packages Projects Releases Wiki Activity

Speed up truffle with 256b TBL instructions

Browse Source 
256b wide SVE vectors allow some simplification of truffle.
Up to 40% speedup on graviton3. Going from 12500 MB/s to 17000 MB/s
onhe microbenchmark.
SVE2 also offer this capability for 128b vector with a speedup around
25% compared to normal SVE

Add unit tests and benchmark for this wide variant

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>
This commit is contained in:
Yoan Picchi
2024-04-23 12:04:40 +00:00
parent 154cb6333b
commit 7054378c93
23 changed files with 1125 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
unit/internal/sheng.cpp
View File

@@ -290,19 +290,26 @@ struct NFA *get_expected_nfa_header(u8 type, unsigned int length, unsigned int n
}
struct NFA *get_expected_nfa16_header() {
return get_expected_nfa_header(SHENG_NFA, 4736, 8);
return get_expected_nfa_header(SHENG_NFA, 4736, 8); /* size recorded in 04/2024 */
}
#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
struct NFA *get_expected_nfa32_header() {
return get_expected_nfa_header(SHENG_NFA_32, 17216, 18);
return get_expected_nfa_header(SHENG_NFA_32, 17216, 18); /* size recorded in 04/2024 */
}
#endif /* defined(HAVE_AVX512VBMI) || defined(HAVE_SVE) */
void test_nfa_equal(const NFA& l, const NFA& r)
{
/**
* The length is meant to be a sanity test: it's not 0 (we compiled something) and that it roughly fit the
* expected size for a given sheng implementation (we don't feed compiled sheng32 into sheng16).
* Changes in other nfa algorithms may affect the sheng length, so we accept small variations.
*/
int relative_difference = std::abs((float)(l.length) - r.length) / ((l.length + r.length) / 2);
EXPECT_LE(relative_difference, 0.1); /* same +-10% */
EXPECT_EQ(l.flags, r.flags);
EXPECT_EQ(l.length, r.length);
EXPECT_EQ(l.type, r.type);
EXPECT_EQ(l.rAccelType, r.rAccelType);
EXPECT_EQ(l.rAccelOffset, r.rAccelOffset);