github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-11-19 18:44:24 +03:00
Code Issues Packages Projects Releases Wiki Activity

Implement new Vermicelli16 acceleration functions using SVE2.

Browse Source 
The scheme utilises the MATCH and NMATCH instructions to
scan for 16 characters at the same rate as vermicelli
scans for one.

Change-Id: Ie2cef904c56651e6108593c668e9b65bc001a886
This commit is contained in:
George Wort
2021-06-28 16:29:43 +01:00
committed by Konstantinos Margaritis
parent b6a7ee7e84
commit b54710d208
25 changed files with 1153 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
src/nfa/vermicelli_sve.h
View File

@@ -232,10 +232,9 @@ const u8 *rdvermSearchLoopBody(svuint16_t chars, const u8 *buf) {
}
static really_inline
const u8 *vermSearch(char c, bool nocase, const u8 *buf, const u8 *buf_end,
const u8 *vermSearch(svuint8_t chars, const u8 *buf, const u8 *buf_end,
bool negate) {
assert(buf < buf_end);
svuint8_t chars = getCharMaskSingle(c, nocase);
size_t len = buf_end - buf;
if (len <= svcntb()) {
return vermSearchOnce(chars, buf, buf_end, negate);
@@ -267,10 +266,9 @@ const u8 *vermSearch(char c, bool nocase, const u8 *buf, const u8 *buf_end,
}
static really_inline
const u8 *rvermSearch(char c, bool nocase, const u8 *buf, const u8 *buf_end,
const u8 *rvermSearch(svuint8_t chars, const u8 *buf, const u8 *buf_end,
bool negate) {
assert(buf < buf_end);
svuint8_t chars = getCharMaskSingle(c, nocase);
size_t len = buf_end - buf;
if (len <= svcntb()) {
return rvermSearchOnce(chars, buf, buf_end, negate);
@@ -353,7 +351,8 @@ const u8 *vermicelliExec(char c, bool nocase, const u8 *buf,
const u8 *buf_end) {
DEBUG_PRINTF("verm scan %s\\x%02hhx over %td bytes\n",
nocase ? "nocase " : "", c, buf_end - buf);
const u8 *ptr = vermSearch(c, nocase, buf, buf_end, false);
svuint8_t chars = getCharMaskSingle(c, nocase);
const u8 *ptr = vermSearch(chars, buf, buf_end, false);
return ptr ? ptr : buf_end;
}
@@ -364,7 +363,8 @@ const u8 *nvermicelliExec(char c, bool nocase, const u8 *buf,
const u8 *buf_end) {
DEBUG_PRINTF("nverm scan %s\\x%02hhx over %td bytes\n",
nocase ? "nocase " : "", c, buf_end - buf);
const u8 *ptr = vermSearch(c, nocase, buf, buf_end, true);
svuint8_t chars = getCharMaskSingle(c, nocase);
const u8 *ptr = vermSearch(chars, buf, buf_end, true);
return ptr ? ptr : buf_end;
}
@@ -375,7 +375,8 @@ const u8 *rvermicelliExec(char c, bool nocase, const u8 *buf,
const u8 *buf_end) {
DEBUG_PRINTF("rev verm scan %s\\x%02hhx over %td bytes\n",
nocase ? "nocase " : "", c, buf_end - buf);
const u8 *ptr = rvermSearch(c, nocase, buf, buf_end, false);
svuint8_t chars = getCharMaskSingle(c, nocase);
const u8 *ptr = rvermSearch(chars, buf, buf_end, false);
return ptr ? ptr : buf - 1;
}
@@ -386,7 +387,8 @@ const u8 *rnvermicelliExec(char c, bool nocase, const u8 *buf,
const u8 *buf_end) {
DEBUG_PRINTF("rev verm scan %s\\x%02hhx over %td bytes\n",
nocase ? "nocase " : "", c, buf_end - buf);
const u8 *ptr = rvermSearch(c, nocase, buf, buf_end, true);
svuint8_t chars = getCharMaskSingle(c, nocase);
const u8 *ptr = rvermSearch(chars, buf, buf_end, true);
return ptr ? ptr : buf - 1;
}
@@ -427,4 +429,45 @@ const u8 *rvermicelliDoubleExec(char c1, char c2, bool nocase, const u8 *buf,
}
}
return buf - 1;
}
static really_inline
svuint8_t getDupSVEMaskFrom128(m128 _mask) {
return svld1rq_u8(svptrue_b8(), (const uint8_t *)&_mask);
}
static really_inline
const u8 *vermicelli16Exec(const m128 _chars, const u8 *buf,
const u8 *buf_end) {
DEBUG_PRINTF("verm16 scan over %td bytes\n", buf_end - buf);
svuint8_t chars = getDupSVEMaskFrom128(_chars);
const u8 *ptr = vermSearch(chars, buf, buf_end, false);
return ptr ? ptr : buf_end;
}
static really_inline
const u8 *nvermicelli16Exec(const m128 _chars, const u8 *buf,
const u8 *buf_end) {
DEBUG_PRINTF("nverm16 scan over %td bytes\n", buf_end - buf);
svuint8_t chars = getDupSVEMaskFrom128(_chars);
const u8 *ptr = vermSearch(chars, buf, buf_end, true);
return ptr ? ptr : buf_end;
}
static really_inline
const u8 *rvermicelli16Exec(const m128 _chars, const u8 *buf,
const u8 *buf_end) {
DEBUG_PRINTF("rverm16 scan over %td bytes\n", buf_end - buf);
svuint8_t chars = getDupSVEMaskFrom128(_chars);
const u8 *ptr = rvermSearch(chars, buf, buf_end, false);
return ptr ? ptr : buf - 1;
}
static really_inline
const u8 *rnvermicelli16Exec(const m128 _chars, const u8 *buf,
const u8 *buf_end) {
DEBUG_PRINTF("rnverm16 scan over %td bytes\n", buf_end - buf);
svuint8_t chars = getDupSVEMaskFrom128(_chars);
const u8 *ptr = rvermSearch(chars, buf, buf_end, true);
return ptr ? ptr : buf - 1;
}