Optimize merge algorithm for data sizes equal or greater then 4M items
#1933
Conversation
SergeyKopienko
requested review from
dmitriy-sobolev,
danhoeflinger,
mmichel11 and
MikeDvorskiy
SergeyKopienko
force-pushed
the
dev/skopienko/optimize_merge_to_main
branch
5 times, most recently
from
a6164fd
to
d4721ca
Compare
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h
Outdated
Show resolved
Hide resolved
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h
Outdated
Show resolved
Hide resolved
| auto __scratch_idx = __global_idx / __base_diag_part; | ||
|
|
||
| _split_point_t __start; | ||
| if (__global_idx % __base_diag_part != 0) |
There was a problem hiding this comment.
We discussed offline about the approach to partition based on SLM size and then work within the partitioned blocks in the second kernel.
One advantage of this method (beyond working within SLM for all diagonals in this kernel) would be that there would be no work-item divergence with a branch and mod operation like this. The first partitioning kernel would be lightweight and basically only to establish bounds for the second kernel. Then the second kernel would work within SLM loaded data and search for all diagonals within that block then serial merge, and all work items could be the same (with possible exception for the zeroth work item).
| std::forward<_ExecutionPolicy>(__exec), std::forward<_Range1>(__rng1), std::forward<_Range2>(__rng2), | ||
| std::forward<_Range3>(__rng3), __comp); | ||
| } | ||
| else |
There was a problem hiding this comment.
Can we remove this else branch and its kernel? std::uint32_t has a well-defined maximum, and we know __n < 4 * 1'048'576 in this branch, so it can always be indexed with this type.
There was a problem hiding this comment.
Good point.
But another solution - we may use std::uint16_t type too for smaller data sizes.
There was a problem hiding this comment.
Do you have an estimate of how much performance benefit the 16-bit indexing kernel brings? I think it would be best to weigh the impact of this kernel against the increase in JIT time. If the performance benefit is significant, then I am in favor of keeping it.
There was a problem hiding this comment.
No perf profit, removed.
SergeyKopienko
force-pushed
the
dev/skopienko/optimize_merge_to_main
branch
from
0eb4649
to
129898c
Compare
…introduce new function __find_start_point_in Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
…introduce __parallel_merge_submitter_large for merge of biggest data sizes Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
…using __parallel_merge_submitter_large for merge data equal or greater then 4M items Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
…fix compile error Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
…fix Kernel names Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
…rename template parameter names in __parallel_merge_submitter Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
SergeyKopienko
force-pushed
the
dev/skopienko/optimize_merge_to_main
branch
from
d8366c3
to
58eacbf
Compare
…fix review comment Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
SergeyKopienko
force-pushed
the
dev/skopienko/optimize_merge_to_main
branch
from
8f756f0
to
1b6cd34
Compare
…fix review comment Signed-off-by: Sergey Kopienko <sergey.kopienko@intel.com>
In this PR we optimize
mergealgorithm for data sizes equal or greater then 4M items.The main idea - we doing two submits:
submitwe find split point in some"base"diagonal's subset.submitwe find split points in all other diagonal and runserial mergefor each diagonal (as before).But when we find split point on the current diagonal, we setup some indexes limits for
rng1and 'rng2'.For these limits we load split point's data from previous and next
"base"diagonals, calculated on the step (1).Applying this approach we have good perf profit for biggest data sizes with
floatandintdata types.As additional profit, we have sign performance boost for small and middle data sizes in the
merge_sortalgorithm.