github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-09-29 19:24:25 +03:00
Code Issues Packages Projects Releases Wiki Activity

prefetch works best when addresses are 64-byte aligned

Browse Source
This commit is contained in:
Konstantinos Margaritis
2021-03-12 10:10:53 +02:00
parent 521f233cfd
commit d3ff893871
6 changed files with 26 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/hwlm/noodle_engine_avx2.c
View File

@@ -95,7 +95,7 @@ hwlm_error_t scanSingleFast(const struct noodTable *n, const u8 *buf,
u32 z = movemask256(eq256(mask1, v));
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d + 128);
__builtin_prefetch(ROUNDDOWN_PTR(d + 128, 64));
hwlm_error_t result = single_zscan(n, d, buf, z, len, cbi);
if (unlikely(result != HWLM_SUCCESS))
@@ -126,7 +126,7 @@ hwlm_error_t scanDoubleFast(const struct noodTable *n, const u8 *buf,
lastz0 = z0 >> 31;
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d + 128);
__builtin_prefetch(ROUNDDOWN_PTR(d + 128, 64));
hwlm_error_t result = double_zscan(n, d, buf, z, len, cbi);
if (unlikely(result != HWLM_SUCCESS))

4
src/hwlm/noodle_engine_sse.c
View File

@@ -91,7 +91,7 @@ hwlm_error_t scanSingleFast(const struct noodTable *n, const u8 *buf,
u32 z = movemask128(eq128(mask1, v));
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d + 128);
__builtin_prefetch(ROUNDDOWN_PTR(d + 128, 64));
DEBUG_PRINTF("z 0x%08x\n", z);
hwlm_error_t result = single_zscan(n, d, buf, z, len, cbi);
@@ -118,7 +118,7 @@ hwlm_error_t scanDoubleFast(const struct noodTable *n, const u8 *buf,
lastz1 = z1;
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(d + 128);
__builtin_prefetch(ROUNDDOWN_PTR(d + 128, 64));
DEBUG_PRINTF("z 0x%08x\n", z);
hwlm_error_t result = double_zscan(n, d, buf, z, len, cbi);