Delete optimization

[kstppd]

• This ports over the optimizations done to the insertion kernels during summer 2023 to the delete kernels.
• On top of that hashers.h is now made smaller by separating the nvidia and amd kernels to different files namely kernels_NVIDIA.h and kernels_AMD.h. There is some level of duplication in those files which should be cleaned up in the near future.

[kstppd]

Some stats from VOLTAR:

DEV

void Hashinator::Hashers::delete_kernel<unsigned int, unsigned int, (unsigned int)4294967295, (unsigned int)4294967294, Hashinator::HashFunctions::Fibonacci, (int)32, (int)4>(T1 *, Hashinator::hash_pair<T1, T2> *, unsigned long *, int, unsigned long, unsigned long) (131072, 1, 1)x(1024, 1, 1), Device 0, CC 8.0, Invocations 10
Section: GPU Speed Of Light Throughput
Metric Name Metric Unit Minimum Maximum Average
---------------------------------------------------------------- ------------- -------------- -------------- --------------
dram__cycles_elapsed.avg.per_second cycle/nsecond 1.505338 1.516568 1.511397
gpc__cycles_elapsed.avg.per_second cycle/nsecond 1.060273 1.068183 1.064541
gpc__cycles_elapsed.max cycle 7446972.000000 7500421.000000 7464503.200000
gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed % 17.833765 17.962199 17.920048
gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed % 17.833765 17.962199 17.920048
gpu__time_duration.sum msecond 6.994560 7.028800 7.011821
l1tex__throughput.avg.pct_of_peak_sustained_active % 9.222978 9.263118 9.239684
lts__throughput.avg.pct_of_peak_sustained_elapsed % 26.295844 26.489994 26.420818
sm__cycles_active.avg cycle 7435975.435185 7468405.527778 7455003.432407
sm__throughput.avg.pct_of_peak_sustained_elapsed % 10.428739 10.503618 10.478983

void Hashinator::Hashers::insert_kernel<unsigned int, unsigned int, (unsigned int)4294967295, Hashinator::HashFunctions::Fibonacci, (int)32, (int)4>(T1 *, T2 *, Hashinator::hash_pair<T1, T2> *, int, unsigned long, unsigned long *, unsigned long *, unsigned long, Hashinator::status *) (131072, 1, 1)x(1024, 1, 1), Device 0, CC 8.0, Invocations 10
Section: GPU Speed Of Light Throughput
Metric Name Metric Unit Minimum Maximum Average
---------------------------------------------------------------- ------------- -------------- -------------- --------------
dram__cycles_elapsed.avg.per_second cycle/nsecond 1.510651 1.512987 1.511663
gpc__cycles_elapsed.avg.per_second cycle/nsecond 1.063859 1.065505 1.064571
gpc__cycles_elapsed.max cycle 4213775.000000 4218969.000000 4216988.200000
gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed % 35.815730 35.857784 35.832524
gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed % 32.626122 32.670232 32.642962
gpu__time_duration.sum msecond 3.956960 3.963264 3.960874
l1tex__throughput.avg.pct_of_peak_sustained_active % 35.873580 35.904683 35.888333
lts__throughput.avg.pct_of_peak_sustained_elapsed % 58.533260 58.634845 58.583288
sm__cycles_active.avg cycle 4207948.657407 4212078.842593 4210073.062963
sm__throughput.avg.pct_of_peak_sustained_elapsed % 46.175313 46.232249 46.197265

void Hashinator::Hashers::retrieve_kernel<unsigned int, unsigned int, (unsigned int)4294967295, Hashinator::HashFunctions::Fibonacci, (int)32, (int)4>(T1 *, T2 *, Hashinator::hash_pair<T1, T2> *, int, unsigned long) (131072, 1, 1)x(1024, 1, 1), Device 0, CC 8.0, Invocations 10
Section: GPU Speed Of Light Throughput
Metric Name Metric Unit Minimum Maximum Average
---------------------------------------------------------------- ------------- -------------- -------------- --------------
dram__cycles_elapsed.avg.per_second cycle/nsecond 1.506765 1.511590 1.510017
gpc__cycles_elapsed.avg.per_second cycle/nsecond 1.061036 1.064440 1.063329
gpc__cycles_elapsed.max cycle 2537513.000000 2539980.000000 2539065.500000
gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed % 45.088881 45.133333 45.106795
gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed % 45.088881 45.133333 45.106795
gpu__time_duration.sum msecond 2.385408 2.391648 2.387571
l1tex__throughput.avg.pct_of_peak_sustained_active % 22.094788 22.132451 22.114941
lts__throughput.avg.pct_of_peak_sustained_elapsed % 59.425200 59.494961 59.452679
sm__cycles_active.avg cycle 2520763.879630 2525127.953704 2522717.659259
sm__throughput.avg.pct_of_peak_sustained_elapsed % 26.102052 26.128094 26.112027

DELETE_OPTIMIZATIONS

void Hashinator::Hashers::delete_kernel<unsigned int, unsigned int, (unsigned int)4294967295, (unsigned int)4294967294, Hashinator::HashFunctions::Fibonacci, (int)32, (int)4>(T1 *, Hashinator::hash_pair<T1, T2> *, unsigned long *, int, unsigned long, unsigned long) (131072, 1, 1)x(1024, 1, 1), Device 0, CC 8.0, Invocations 10
Section: GPU Speed Of Light Throughput
Metric Name Metric Unit Minimum Maximum Average
---------------------------------------------------------------- ------------- -------------- -------------- --------------
dram__cycles_elapsed.avg.per_second cycle/nsecond 1.507232 1.517378 1.510987
gpc__cycles_elapsed.avg.per_second cycle/nsecond 1.057371 1.068437 1.063286
gpc__cycles_elapsed.max cycle 3677466.000000 3697696.000000 3686546.700000
gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed % 36.137017 36.322093 36.237116
gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed % 36.137017 36.322093 36.237116
** gpu__time_duration.sum msecond 3.458112 3.473248 3.465114 **
l1tex__throughput.avg.pct_of_peak_sustained_active % 31.170985 31.314450 31.260865
lts__throughput.avg.pct_of_peak_sustained_elapsed % 49.128153 49.428405 49.272549
sm__cycles_active.avg cycle 3669260.083333 3685783.824074 3675393.850000
sm__throughput.avg.pct_of_peak_sustained_elapsed % 30.080758 30.368656 30.183720

void Hashinator::Hashers::insert_kernel<unsigned int, unsigned int, (unsigned int)4294967295, Hashinator::HashFunctions::Fibonacci, (int)32, (int)4>(T1 *, T2 *, Hashinator::hash_pair<T1, T2> *, int, unsigned long, unsigned long *, unsigned long *, unsigned long, Hashinator::status *) (131072, 1, 1)x(1024, 1, 1), Device 0, CC 8.0, Invocations 10
Section: GPU Speed Of Light Throughput
Metric Name Metric Unit Minimum Maximum Average
---------------------------------------------------------------- ------------- -------------- -------------- --------------
dram__cycles_elapsed.avg.per_second cycle/nsecond 1.510984 1.513487 1.512004
gpc__cycles_elapsed.avg.per_second cycle/nsecond 1.064106 1.065860 1.064825
gpc__cycles_elapsed.max cycle 4216802.000000 4221979.000000 4218887.900000
gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed % 35.793318 35.837510 35.817612
gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed % 32.599702 32.640707 32.624702
gpu__time_duration.sum msecond 3.957856 3.964288 3.961763
l1tex__throughput.avg.pct_of_peak_sustained_active % 35.817812 35.892894 35.868637
lts__throughput.avg.pct_of_peak_sustained_elapsed % 58.519449 58.586801 58.561299
sm__cycles_active.avg cycle 4209233.851852 4218398.981481 4212581.847222
sm__throughput.avg.pct_of_peak_sustained_elapsed % 46.145086 46.202654 46.179337

void Hashinator::Hashers::retrieve_kernel<unsigned int, unsigned int, (unsigned int)4294967295, Hashinator::HashFunctions::Fibonacci, (int)32, (int)4>(T1 *, T2 *, Hashinator::hash_pair<T1, T2> *, int, unsigned long) (131072, 1, 1)x(1024, 1, 1), Device 0, CC 8.0, Invocations 10
Section: GPU Speed Of Light Throughput
Metric Name Metric Unit Minimum Maximum Average
---------------------------------------------------------------- ------------- -------------- -------------- --------------
dram__cycles_elapsed.avg.per_second cycle/nsecond 1.506886 1.511546 1.509261
gpc__cycles_elapsed.avg.per_second cycle/nsecond 1.061131 1.064427 1.062812
gpc__cycles_elapsed.max cycle 2538442.000000 2542006.000000 2540214.800000
gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed % 45.049734 45.114206 45.081456
gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed % 45.049734 45.114206 45.081456
gpu__time_duration.sum msecond 2.387328 2.393696 2.389834
l1tex__throughput.avg.pct_of_peak_sustained_active % 22.086863 22.117189 22.101961
lts__throughput.avg.pct_of_peak_sustained_elapsed % 59.407263 59.487880 59.437980
sm__cycles_active.avg cycle 2522794.518519 2526302.138889 2524542.429630
sm__throughput.avg.pct_of_peak_sustained_elapsed % 26.087542 26.124117 26.105620

So essentially we double the throughput!

• dev deletion-> gpu__time_duration.sum msecond 6.994560 7.028800 7.011821
• this PR-> gpu__time_duration.sum msecond 3.458112 3.473248 3.465114

[kstppd]

Will be kept in PR state for a while