Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add b200 tunings for radix_sort.pairs #3626

Merged
merged 2 commits into from
Feb 4, 2025
Merged

Add b200 tunings for radix_sort.pairs #3626

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e3a0aea
Add b200 tunings for radix_sort.pairs
gonidelis Jan 10, 2025
66dcbc9
Default back i8,i64,* and i16,i64,* tunings due regressions
gonidelis Feb 4, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
134 changes: 134 additions & 0 deletions cub/cub/device/dispatch/tuning/tuning_radix_sort.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -143,6 +143,140 @@ template <> struct sm100_small_key_tuning<float, 4, 0, 8> { static constexpr in

// ipt_21.tpb_256 1.068590 0.986635 1.059704 1.144921
template <> struct sm100_small_key_tuning<double, 8, 0, 8> { static constexpr int threads = 256; static constexpr int items = 21; };

// pairs 1-byte key

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 1, 4> : sm90_small_key_tuning<1, 1, 4> {};

// ipt_18.tpb_512 1.011463 0.978807 1.010106 1.024056
// todo(@gonidelis): insignificant performance gain, need more runs.
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 2, 4> { static constexpr int threads = 512; static constexpr int items = 18; };

// ipt_18.tpb_512 1.008207 0.980377 1.007132 1.022155
// todo(@gonidelis): insignificant performance gain, need more runs.
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 4, 4> { static constexpr int threads = 512; static constexpr int items = 18; };

// todo(@gonidelis): regresses for large problem sizes.
// template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 8, 4> { static constexpr int threads = 288; static constexpr int items = 16; };

// ipt_21.tpb_576 1.044274 0.979145 1.038723 1.072068
// todo(@gonidelis): insignificant performance gain, need more runs.
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 16, 4> { static constexpr int threads = 576; static constexpr int items = 21; };

// ipt_20.tpb_384 1.008881 0.968750 1.006846 1.026910
// todo(@gonidelis): insignificant performance gain, need more runs.
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 1, 8> { static constexpr int threads = 384; static constexpr int items = 20; };

// ipt_22.tpb_256 1.015597 0.966038 1.011167 1.045921
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 2, 8> { static constexpr int threads = 256; static constexpr int items = 22; };

// ipt_15.tpb_384 1.029730 0.972699 1.029066 1.067894
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 4, 8> { static constexpr int threads = 384; static constexpr int items = 15; };

// todo(@gonidelis): regresses for large problem sizes.
// template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 8, 8> { static constexpr int threads = 256; static constexpr int items = 17; };

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 1, 16, 8> : sm90_small_key_tuning<1, 16, 8> {};


// pairs 2-byte key

// ipt_20.tpb_448 1.031929 0.936849 1.023411 1.075172
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 1, 4> { static constexpr int threads = 448; static constexpr int items = 20; };

// ipt_23.tpb_384 1.104683 0.939335 1.087342 1.234988
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 2, 4> { static constexpr int threads = 384; static constexpr int items = 23; };

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 4, 4> : sm90_small_key_tuning<2, 4, 4> {};

// todo(@gonidelis): regresses for large problem sizes.
// template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 8, 4> { static constexpr int threads = 256; static constexpr int items = 17; };

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 16, 4> : sm90_small_key_tuning<2, 16, 4> {};

// ipt_15.tpb_384 1.093598 1.000000 1.088111 1.183369
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 1, 8> { static constexpr int threads = 384; static constexpr int items = 15; };

// ipt_15.tpb_576 1.040476 1.000333 1.037060 1.084850
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 2, 8> { static constexpr int threads = 576; static constexpr int items = 15; };

// ipt_18.tpb_512 1.096819 0.953488 1.082026 1.209533
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 4, 8> { static constexpr int threads = 512; static constexpr int items = 18; };

// todo(@gonidelis): regresses for large problem sizes.
// template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 8, 8> { static constexpr int threads = 288; static constexpr int items = 16; };

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 2, 16, 8> : sm90_small_key_tuning<2, 16, 8> {};


// pairs 4-byte key

// ipt_21.tpb_416 1.237956 1.001909 1.210882 1.469981
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 1, 4> { static constexpr int threads = 416; static constexpr int items = 21; };

// ipt_17.tpb_512 1.022121 1.012346 1.022439 1.038524
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 2, 4> { static constexpr int threads = 512; static constexpr int items = 17; };

// ipt_20.tpb_448 1.012688 0.999531 1.011865 1.028513
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 4, 4> { static constexpr int threads = 448; static constexpr int items = 20; };

// ipt_15.tpb_384 1.006872 0.998651 1.008374 1.026118
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 8, 4> { static constexpr int threads = 384; static constexpr int items = 15; };

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 16, 4> : sm90_small_key_tuning<4, 16, 4> {};

// ipt_17.tpb_512 1.080000 0.927362 1.066211 1.172959
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 1, 8> { static constexpr int threads = 512; static constexpr int items = 17; };

// ipt_15.tpb_384 1.068529 1.000000 1.062277 1.135281
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 2, 8> { static constexpr int threads = 384; static constexpr int items = 15; };

// ipt_21.tpb_448 1.080642 0.927713 1.064758 1.191177
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 4, 8> { static constexpr int threads = 448; static constexpr int items = 21; };

// ipt_13.tpb_448 1.019046 0.991228 1.016971 1.039712
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 8, 8> { static constexpr int threads = 448; static constexpr int items = 13; };

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 4, 16, 8> : sm90_small_key_tuning<4, 16, 8> {};

// pairs 8-byte key

// ipt_17.tpb_256 1.276445 1.025562 1.248511 1.496947
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 1, 4> { static constexpr int threads = 256; static constexpr int items = 17; };

// ipt_12.tpb_352 1.128086 1.040000 1.117960 1.207254
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 2, 4> { static constexpr int threads = 352; static constexpr int items = 12; };

// ipt_12.tpb_352 1.132699 1.040000 1.122676 1.207716
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 4, 4> { static constexpr int threads = 352; static constexpr int items = 12; };

// ipt_18.tpb_256 1.266745 0.995432 1.237754 1.460538
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 8, 4> { static constexpr int threads = 256; static constexpr int items = 18; };

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 16, 4> : sm90_small_key_tuning<8, 16, 4> {};

// ipt_15.tpb_384 1.007343 0.997656 1.006929 1.047208
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 1, 8> { static constexpr int threads = 384; static constexpr int items = 15; };

// ipt_14.tpb_256 1.186477 1.012683 1.167150 1.332313
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 2, 8> { static constexpr int threads = 256; static constexpr int items = 14; };

// ipt_21.tpb_256 1.220607 1.000239 1.196400 1.390471
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 4, 8> { static constexpr int threads = 256; static constexpr int items = 21; };

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 8, 8> : sm90_small_key_tuning<8, 8, 8> {};

// same as previous tuning
template <typename ValueT> struct sm100_small_key_tuning<ValueT, 8, 16, 8> : sm90_small_key_tuning<8, 16, 8> {};
// clang-format on

/**
Expand Down
Loading