prenex/simap
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

simd-map multimap search fix in the AVX2 code

Browse Source
This commit is contained in:
Richard Thier 2024-10-23 18:06:01 +02:00
parent 41988a0dee
commit 968f95734d
2 changed files with 5 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
makefile
View File

@ -5,6 +5,6 @@ release:
debug-avx2: debug-avx2:
g++ main.cpp -g -mavx2 -Wall -o main g++ main.cpp -g -mavx2 -Wall -o main
release-avx2: release-avx2:
g++ main.cpp -mavx2 -O3 -Wall -o main g++ main.cpp -mavx2 -O2 -Wall -o main
release-avx2-asm: release-avx2-asm:
g++ main.cpp -S -fopt-info-vec-missed -masm=intel -mavx2 -O3 -Wall -o main g++ main.cpp -S -fopt-info-vec-missed -masm=intel -mavx2 -O2 -Wall -o main

7
simd_map.h
View File

@ -103,16 +103,16 @@ static inline SM_ALWAYS_INLINE uint32_t *simd_map_lane_find(
/* The 's' means "single" (float precision), and mask will have [0..7] bits set! */ /* The 's' means "single" (float precision), and mask will have [0..7] bits set! */
uint32_t mask = (uint32_t) _mm256_movemask_ps((__m256) m); uint32_t mask = (uint32_t) _mm256_movemask_ps((__m256) m);
if(SM_UNLIKELY(mask != 0)) { if(SM_UNLIKELY(mask != 0)) {
int ipc = __builtin_popcount(mask);
/* 00000000 00000000 00000000 01000100 -> 6 */ /* 00000000 00000000 00000000 01000100 -> 6 */
int i = (31 - __builtin_clz(mask)); int i = (31 - __builtin_clz(mask));
uint32_t *ptr = &values[i]; uint32_t *ptr = &values[i];
if(SM_LIKELY(ipc == 1) || i >= lane_begin) { if(SM_LIKELY(lane_begin == 0)) {
/* Only one match in the lane OR first matching in find/find_all */ /* Fast-path: Only one match per lane OR first matching in lane for this find/find_all */
*lane_next_begin = (i + 1) % SM_LANE_SPAN; *lane_next_begin = (i + 1) % SM_LANE_SPAN;
return ptr; return ptr;
} else { } else {
/* We did a find_all(..) AND there is more than one match in the lane /* We did a find_all(..) AND there is more than one match in the lane
* and its not first find_all(..) on the lane in question...
* *
* This might be suboptimal, but not so bad: * This might be suboptimal, but not so bad:
* *
@ -123,7 +123,6 @@ static inline SM_ALWAYS_INLINE uint32_t *simd_map_lane_find(
* *
* I guess its fine as it should happen statistically rarely anyways * I guess its fine as it should happen statistically rarely anyways
*/ */
/* TODO: Can this be solved more optimal by specialized function? */
goto non_simd_modulo; goto non_simd_modulo;
} }
} }