|
@ -120,11 +120,6 @@ void scan_a(size_t gid, size_t tid) { |
|
|
BARRIERS_[gid]->arrive_and_drop(); |
|
|
BARRIERS_[gid]->arrive_and_drop(); |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
template <size_t size> |
|
|
|
|
|
__m512i AggrFn(uint64_t* chunk_ptr, uint16_t* mask_ptr_a, const uint32_t tid, const uint32_t gid, __m512i aggregator, uint64_t* load_time) { |
|
|
|
|
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
void aggr_j(size_t gid, size_t tid) { |
|
|
void aggr_j(size_t gid, size_t tid) { |
|
|
constexpr size_t SUBCHUNK_SIZE_B = CHUNK_SIZE_B / TC_AGGRJ; |
|
|
constexpr size_t SUBCHUNK_SIZE_B = CHUNK_SIZE_B / TC_AGGRJ; |
|
|
constexpr size_t LAST_CHUNK_SIZE_B = SUBCHUNK_SIZE_B + (CHUNK_SIZE_B % (TC_AGGRJ * GROUP_COUNT)); |
|
|
constexpr size_t LAST_CHUNK_SIZE_B = SUBCHUNK_SIZE_B + (CHUNK_SIZE_B % (TC_AGGRJ * GROUP_COUNT)); |
|
|