github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-11-16 09:21:52 +03:00
Code Issues Packages Projects Releases Wiki Activity

Implement new DoubleVermicelli16 acceleration functions using SVE2

Browse Source 
Change-Id: Id4a8ffca840caab930a6e78cc0dfd0fe7d320b4e
This commit is contained in:
George Wort
2021-06-28 16:29:43 +01:00
committed by Konstantinos Margaritis
parent 91f5f10831
commit e1f0f6baf7
9 changed files with 874 additions and 159 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
src/nfa/accelcompile.cpp
View File

@@ -207,16 +207,45 @@ void buildAccelDouble(const AccelInfo &info, AccelAux *aux) {
u8 m2;
if (buildDvermMask(info.double_stop2, &m1, &m2)) {
u8 c1 = info.double_stop2.begin()->first & m1;
u8 c2 = info.double_stop2.begin()->second & m2;
#ifdef HAVE_SVE2
if (vermicelliDoubleMasked16Build(c1, c2, m1, m2, (u8 *)&aux->mdverm16.mask)) {
aux->accel_type = ACCEL_DVERM16_MASKED;
aux->mdverm16.offset = offset;
aux->mdverm16.c1 = c1;
aux->mdverm16.m1 = m1;
DEBUG_PRINTF("building maskeddouble16-vermicelli for 0x%02hhx%02hhx\n",
c1, c2);
return;
} else if (outs2 <= 8 &&
vermicelliDouble16Build(info.double_stop2, (u8 *)&aux->dverm16.mask,
(u8 *)&aux->dverm16.firsts)) {
aux->accel_type = ACCEL_DVERM16;
aux->dverm16.offset = offset;
DEBUG_PRINTF("building double16-vermicelli\n");
return;
}
#endif // HAVE_SVE2
aux->accel_type = ACCEL_DVERM_MASKED;
aux->dverm.offset = offset;
aux->dverm.c1 = info.double_stop2.begin()->first & m1;
aux->dverm.c2 = info.double_stop2.begin()->second & m2;
aux->dverm.c1 = c1;
aux->dverm.c2 = c2;
aux->dverm.m1 = m1;
aux->dverm.m2 = m2;
DEBUG_PRINTF("building maskeddouble-vermicelli for 0x%02hhx%02hhx\n",
aux->dverm.c1, aux->dverm.c2);
DEBUG_PRINTF("building maskeddouble-vermicelli for 0x%02hhx%02hhx\n", c1, c2);
return;
}
#ifdef HAVE_SVE2
if (outs2 <= 8 &&
vermicelliDouble16Build(info.double_stop2, (u8 *)&aux->dverm16.mask,
(u8 *)&aux->dverm16.firsts)) {
aux->accel_type = ACCEL_DVERM16;
aux->dverm16.offset = offset;
DEBUG_PRINTF("building double16-vermicelli\n");
return;
}
#endif // HAVE_SVE2
}
if (outs1 < outs2 && outs1 <= 2) { // Heuristic from UE-438.