Skip to content

Commit

Permalink
Drop 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 66cbb57 commit 30f5b21
Showing 1 changed file with 3 additions and 97 deletions.
100 changes: 3 additions & 97 deletions cub/cub/device/dispatch/tuning/tuning_histogram.cuh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -151,17 +151,7 @@ struct sm100_tuning<true, SampleT, 1, 1, counter_size::_4, primitive_sample::yes
static constexpr int vec_size = 1 << 2;
};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<true, SampleT, 1, 1, counter_size::_4, primitive_sample::yes, sample_size::_2> {};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<true, SampleT, 1, 1, counter_size::_4, primitive_sample::yes, sample_size::_4> {};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<true, SampleT, 1, 1, counter_size::_4, primitive_sample::yes, sample_size::_8> {};
// sample_size 2/4/8 showed no benefit over SM90 during verification benchmarks

// range
template <class SampleT>
Expand All @@ -178,93 +168,9 @@ struct sm100_tuning<false, SampleT, 1, 1, counter_size::_4, primitive_sample::ye
static constexpr int vec_size = 1 << 2;
};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<false, SampleT, 1, 1, counter_size::_4, primitive_sample::yes, sample_size::_2> {};

// TODO(gonidelis): we found the below tuning but the verification benchmark showed regressions, so it's disabled
// template <class SampleT>
// struct sm100_tuning<false, SampleT, 1, 1, counter_size::_4, primitive_sample::yes, sample_size::_4>
// {
// // ipt_9.tpb_1024.rle_1.ws_0.mem_1.ld_0.laid_1.vec_0 1.358537 1.001009 1.373329 2.614104
// static constexpr int items = 9;
// static constexpr int threads = 1024;
// static constexpr bool rle_compress = true;
// static constexpr bool work_stealing = false;
// static constexpr BlockHistogramMemoryPreference mem_preference = SMEM;
// static constexpr CacheLoadModifier load_modifier = LOAD_DEFAULT;
// static constexpr BlockLoadAlgorithm load_algorithm = BLOCK_LOAD_WARP_TRANSPOSE;
// static constexpr int vec_size = 1 << 0;
// };

// TODO(gonidelis): we found the below tuning but the verification benchmark showed regressions, so it's disabled
// template <class SampleT>
// struct sm100_tuning<false, SampleT, 1, 1, counter_size::_4, primitive_sample::yes, sample_size::_8>
// {
// // ipt_7.tpb_544.rle_1.ws_0.mem_1.ld_1.laid_0.vec_0 1.105331 0.934888 1.108557 1.391657
// static constexpr int items = 7;
// static constexpr int threads = 544;
// static constexpr bool rle_compress = true;
// static constexpr bool work_stealing = false;
// static constexpr BlockHistogramMemoryPreference mem_preference = SMEM;
// static constexpr CacheLoadModifier load_modifier = LOAD_LDG;
// static constexpr BlockLoadAlgorithm load_algorithm = BLOCK_LOAD_DIRECT;
// static constexpr int vec_size = 1 << 0;
// };

// multi.even
template <class SampleT>
struct sm100_tuning<true, SampleT, 4, 3, counter_size::_4, primitive_sample::yes, sample_size::_1>
{
// ipt_9.tpb_1024.rle_0.ws_0.mem_1.ld_1.laid_1.vec_0 1.629591 0.997416 1.570900 2.772504
static constexpr int items = 9;
static constexpr int threads = 1024;
static constexpr bool rle_compress = false;
static constexpr bool work_stealing = false;
static constexpr BlockHistogramMemoryPreference mem_preference = SMEM;
static constexpr CacheLoadModifier load_modifier = LOAD_LDG;
static constexpr BlockLoadAlgorithm load_algorithm = BLOCK_LOAD_WARP_TRANSPOSE;
static constexpr int vec_size = 1 << 0;
};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<true, SampleT, 4, 3, counter_size::_4, primitive_sample::yes, sample_size::_2> {};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<true, SampleT, 4, 3, counter_size::_4, primitive_sample::yes, sample_size::_4> {};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<true, SampleT, 4, 3, counter_size::_4, primitive_sample::yes, sample_size::_8> {};

// multi.range
template <class SampleT>
struct sm100_tuning<false, SampleT, 4, 3, counter_size::_4, primitive_sample::yes, sample_size::_1>
{
// ipt_7.tpb_160.rle_0.ws_0.mem_1.ld_1.laid_1.vec_1 1.210837 0.99556 1.189049 1.939584
static constexpr int items = 7;
static constexpr int threads = 160;
static constexpr bool rle_compress = false;
static constexpr bool work_stealing = false;
static constexpr BlockHistogramMemoryPreference mem_preference = SMEM;
static constexpr CacheLoadModifier load_modifier = LOAD_LDG;
static constexpr BlockLoadAlgorithm load_algorithm = BLOCK_LOAD_WARP_TRANSPOSE;
static constexpr int vec_size = 1 << 1;
};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<false, SampleT, 4, 3, counter_size::_4, primitive_sample::yes, sample_size::_2> {};

// same as SM90
// template <class SampleT>
// struct sm100_tuning<false, SampleT, 4, 3, counter_size::_4, primitive_sample::yes, sample_size::_4> {};
// sample_size 2/4/8 showed no benefit over SM90 during verification benchmarks

// same as SM90
// template <class SampleT>
// struct sm100_tuning<false, SampleT, 4, 3, counter_size::_4, primitive_sample::yes, sample_size::_8> {};
// multi.even and multi.range: none of the found tunings surpassed the SM90 tuning during verification benchmarks

template <class SampleT, class CounterT, int NumChannels, int NumActiveChannels, bool IsEven>
struct policy_hub
Expand Down

0 comments on commit 30f5b21

Please sign in to comment.