Skip to content

Commit

Permalink
Remove reduce tunings with no benefit
Browse files Browse the repository at this point in the history
  • Loading branch information
@bernhardmgruber
bernhardmgruber committed Feb 6, 2025
1 parent 044cabf commit b2b016c
Showing 1 changed file with 2 additions and 174 deletions.
176 changes: 2 additions & 174 deletions cub/cub/device/dispatch/tuning/tuning_reduce.cuh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -156,33 +156,8 @@ template <class AccumT,
struct sm100_tuning;

// sum
template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_4, accum_size::_1>
{
// todo(gonidelis): Very low performance, we need more runs.
// ipt_16.tpb_256.ipv_2 1.001174 1.0 1.001044 1.004175
static constexpr int items = 16;
static constexpr int threads = 256;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_4, accum_size::_2>
{
// ipt_18.tpb_288.ipv_2 1.032068 0.997167 1.028244 1.115809
static constexpr int items = 18;
static constexpr int threads = 288;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_4, accum_size::_4>
{
// ipt_15.tpb_960.ipv_1 1.040241 0.988042 1.038795 1.167139
static constexpr int items = 15;
static constexpr int threads = 960;
static constexpr int items_per_vec_load = 1;
};
// Tunings for offset size 4/8 and accum size 1/2/4 all showed no significant improvement during verification

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_4, accum_size::_8>
Expand All @@ -193,33 +168,6 @@ struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_4, accum_size::_8>
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_8, accum_size::_1>
{
// ipt_14.tpb_288.ipv_2 1.036897 1.000000 1.032813 1.13125
static constexpr int items = 14;
static constexpr int threads = 288;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_8, accum_size::_2>
{
// ipt_12.tpb_224.ipv_2 1.032496 1.000000 1.028899 1.115596
static constexpr int items = 12;
static constexpr int threads = 224;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_8, accum_size::_4>
{
// ipt_14.tpb_288.ipv_1 1.050725 1.000000 1.048286 1.181818
static constexpr int items = 14;
static constexpr int threads = 288;
static constexpr int items_per_vec_load = 1;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::plus, offset_size::_8, accum_size::_8>
{
Expand Down Expand Up @@ -247,127 +195,7 @@ struct sm100_tuning<double, OffsetT, op_type::plus, offset_size::_4, accum_size:
static constexpr int items_per_vec_load = 1;
};

// min or max
template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_4, accum_size::_1>
{
// ipt_16.tpb_128.ipv_2 1.021369 0.998557 1.019009 1.077479
static constexpr int items = 16;
static constexpr int threads = 128;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_4, accum_size::_2>
{
// ipt_16.tpb_256.ipv_2 1.038750 1.0 1.034382 1.117647
static constexpr int items = 16;
static constexpr int threads = 256;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_4, accum_size::_4>
{
// ipt_12.tpb_448.ipv_1 1.037834 1.000000 1.036212 1.144847
static constexpr int items = 12;
static constexpr int threads = 448;
static constexpr int items_per_vec_load = 1;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_4, accum_size::_8>
{
// ipt_15.tpb_512.ipv_2 1.020165 1.0 1.018162 1.058036
static constexpr int items = 15;
static constexpr int threads = 512;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_4, accum_size::_16>
{
// ipt_16.tpb_320.ipv_2 1.009217 1.0 1.008197 1.032787
static constexpr int items = 16;
static constexpr int threads = 320;
static constexpr int items_per_vec_load = 2;
};

template <class OffsetT>
struct sm100_tuning<float, OffsetT, op_type::min_or_max, offset_size::_4, accum_size::_4>
{
// ipt_18.tpb_448.ipv_2 1.032745 0.966480 1.032123 1.162011
static constexpr int items = 18;
static constexpr int threads = 448;
static constexpr int items_per_vec_load = 2;
};

template <class OffsetT>
struct sm100_tuning<double, OffsetT, op_type::min_or_max, offset_size::_4, accum_size::_4>
{
// ipt_15.tpb_512.ipv_2 1.019901 1.0 1.017648 1.058036
static constexpr int items = 15;
static constexpr int threads = 512;
static constexpr int items_per_vec_load = 2;
};

// same as base, so fall back to Policy600
// template <class T, class OffsetT>
// struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_8, accum_size::_1> {};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_8, accum_size::_2>
{
// ipt_16.tpb_224.ipv_2 1.031922 0.997989 1.028396 1.115596
static constexpr int items = 16;
static constexpr int threads = 224;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_8, accum_size::_4>
{
// ipt_14.tpb_416.ipv_1 1.047490 1.000000 1.045455 1.181818
static constexpr int items = 14;
static constexpr int threads = 416;
static constexpr int items_per_vec_load = 1;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_8, accum_size::_8>
{
// ipt_21.tpb_384.ipv_2 1.021487 1.0 1.019033 1.057143
static constexpr int items = 21;
static constexpr int threads = 384;
static constexpr int items_per_vec_load = 2;
};

template <class T, class OffsetT>
struct sm100_tuning<T, OffsetT, op_type::min_or_max, offset_size::_8, accum_size::_16>
{
// ipt_17.tpb_512.ipv_2 1.003412 0.980713 1.003111 1.031730
static constexpr int items = 17;
static constexpr int threads = 512;
static constexpr int items_per_vec_load = 2;
};

template <class OffsetT>
struct sm100_tuning<float, OffsetT, op_type::min_or_max, offset_size::_8, accum_size::_4>
{
// ipt_18.tpb_448.ipv_1 1.023427 1.000000 1.022287 1.083333
static constexpr int items = 18;
static constexpr int threads = 448;
static constexpr int items_per_vec_load = 1;
};

template <class OffsetT>
struct sm100_tuning<double, OffsetT, op_type::min_or_max, offset_size::_8, accum_size::_8>
{
// ipt_16.tpb_320.ipv_2 1.018602 1.0 1.016518 1.059821
static constexpr int items = 16;
static constexpr int threads = 320;
static constexpr int items_per_vec_load = 1;
};
// For min or max, verification showed the benefits were too small (within noise)

template <typename AccumT, typename OffsetT, typename ReductionOpT>
struct policy_hub
Expand Down

0 comments on commit b2b016c

Please sign in to comment.