github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-09-30 11:44:28 +03:00
Code Issues Packages Projects Releases Wiki Activity

avx512: noodle

Browse Source
This commit is contained in:
Matthew Barr
2016-08-08 10:55:52 +10:00
parent ec7869711f
commit 1089fa5018
2 changed files with 236 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
src/hwlm/noodle_engine.c
View File

@@ -36,12 +36,14 @@
#include "util/arch.h"
#include "util/bitutils.h"
#include "util/compare.h"
#include "util/join.h"
#include "util/masked_move.h"
#include "util/simd_utils.h"
#include <ctype.h>
#include <stdbool.h>
#include <string.h>
#include <x86intrin.h>
/** \brief Noodle runtime context. */
struct cb_info {
@@ -51,6 +53,24 @@ struct cb_info {
size_t offsetAdj; //!< used in streaming mode
};
#if defined(HAVE_AVX512)
#define CHUNKSIZE 64
#define MASK_TYPE m512
#define Z_BITS 64
#define Z_TYPE u64a
#elif defined(HAVE_AVX2)
#define CHUNKSIZE 32
#define MASK_TYPE m256
#define Z_BITS 32
#define Z_TYPE u32
#else
#define CHUNKSIZE 16
#define MASK_TYPE m128
#define Z_BITS 32
#define Z_TYPE u32
#endif
#define RETURN_IF_TERMINATED(x) \
{ \
if ((x) == HWLM_TERMINATED) { \
@@ -61,8 +81,9 @@ struct cb_info {
#define SINGLE_ZSCAN() \
do { \
while (unlikely(z)) { \
u32 pos = findAndClearLSB_32(&z); \
Z_TYPE pos = JOIN(findAndClearLSB_, Z_BITS)(&z); \
size_t matchPos = d - buf + pos; \
DEBUG_PRINTF("match pos %zu\n", matchPos); \
hwlmcb_rv_t rv = final(buf, len, key, 1, 0, 0, noCase, cbi, \
matchPos); \
RETURN_IF_TERMINATED(rv); \
@@ -72,8 +93,9 @@ struct cb_info {
#define DOUBLE_ZSCAN() \
do { \
while (unlikely(z)) { \
u32 pos = findAndClearLSB_32(&z); \
Z_TYPE pos = JOIN(findAndClearLSB_, Z_BITS)(&z); \
size_t matchPos = d - buf + pos - 1; \
DEBUG_PRINTF("match pos %zu\n", matchPos); \
hwlmcb_rv_t rv = final(buf, len, key, keyLen, keyOffset, 1, \
noCase, cbi, matchPos); \
RETURN_IF_TERMINATED(rv); \
@@ -110,7 +132,11 @@ hwlm_error_t final(const u8 *buf, size_t len, const u8 *key, size_t keyLen,
return HWLM_SUCCESS;
}
#if defined(HAVE_AVX2)
#if defined(HAVE_AVX512)
#define CHUNKSIZE 64
#define MASK_TYPE m512
#include "noodle_engine_avx512.c"
#elif defined(HAVE_AVX2)
#define CHUNKSIZE 32
#define MASK_TYPE m256
#include "noodle_engine_avx2.c"
@@ -123,12 +149,14 @@ hwlm_error_t final(const u8 *buf, size_t len, const u8 *key, size_t keyLen,
static really_inline
hwlm_error_t scanSingleMain(const u8 *buf, size_t len, const u8 *key,
bool noCase, const struct cb_info *cbi) {
hwlm_error_t rv;
size_t end = len;
const MASK_TYPE mask1 = getMask(key[0], noCase);
const MASK_TYPE caseMask = getCaseMask();
#if !defined(HAVE_AVX512)
hwlm_error_t rv;
size_t end = len;
if (len < CHUNKSIZE) {
rv = scanSingleShort(buf, len, key, noCase, caseMask, mask1, cbi, 0, len);
return rv;
@@ -173,13 +201,15 @@ hwlm_error_t scanSingleMain(const u8 *buf, size_t len, const u8 *key,
cbi, s2End, end);
return rv;
#else // HAVE_AVX512
return scanSingle512(buf, len, key, noCase, caseMask, mask1, cbi);
#endif
}
static really_inline
hwlm_error_t scanDoubleMain(const u8 *buf, size_t len, const u8 *key,
size_t keyLen, size_t keyOffset, bool noCase,
const struct cb_info *cbi) {
hwlm_error_t rv;
// we stop scanning for the key-fragment when the rest of the key can't
// possibly fit in the remaining buffer
size_t end = len - keyLen + keyOffset + 2;
@@ -188,6 +218,9 @@ hwlm_error_t scanDoubleMain(const u8 *buf, size_t len, const u8 *key,
const MASK_TYPE mask1 = getMask(key[keyOffset + 0], noCase);
const MASK_TYPE mask2 = getMask(key[keyOffset + 1], noCase);
#if !defined(HAVE_AVX512)
hwlm_error_t rv;
if (end - keyOffset < CHUNKSIZE) {
rv = scanDoubleShort(buf, len, key, keyLen, keyOffset, noCase, caseMask,
mask1, mask2, cbi, keyOffset, end);
@@ -244,6 +277,10 @@ hwlm_error_t scanDoubleMain(const u8 *buf, size_t len, const u8 *key,
caseMask, mask1, mask2, cbi, off, end);
return rv;
#else // AVX512
return scanDouble512(buf, len, key, keyLen, keyOffset, noCase, caseMask,
mask1, mask2, cbi, keyOffset, end);
#endif // AVX512
}