Skip to content

Commit

Permalink
Enhance adversary data generation
Browse files Browse the repository at this point in the history
  • Loading branch information
@GabTux
GabTux committed Mar 21, 2024
1 parent cf1601c commit 041f56e
Show file tree
Hide file tree
Showing 2 changed files with 66 additions and 33 deletions.
32 changes: 2 additions & 30 deletions benchmark/source/fixtures.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -189,36 +189,8 @@ concept SizeFit = std::numeric_limits<T>::max() > (Size - 1);

template <typename T = int, std::size_t Size = ELEMENTS_VECTOR_DEFAULT>
class AdversaryVectorFixture : public VectorFixture<T, Size> {
public:
void generate_data() requires SizeFit<T, Size> {
// inspired from this paper: https://www.cs.dartmouth.edu/~doug/mdmspe.pdf
int candidate = 0;
int nsolid = 0;
const int gas = Size - 1;

// initially, all values are gas
std::ranges::fill(this->data_, gas);

// fill with values from 0 to Size-1
// will be used as indices to this->data
std::vector<int> asc_vals(Size);
std::iota(asc_vals.begin(), asc_vals.end(), 0);

auto cmp = [&](int x, int y) {
if (this->data_[x] == gas && this->data_[y] == gas)
{
if (x == candidate)
this->data_[x] = nsolid++;
else
this->data_[y] = nsolid++;
}
if (this->data_[x] == gas)
candidate = x;
else if (this->data_[y] == gas)
candidate = y;
return this->data_[x] < this->data_[y];
};
std::ranges::sort(asc_vals, cmp);
void generate_data() override {
// nothing in general, specific to each sort
}
};

Expand Down
67 changes: 64 additions & 3 deletions benchmark/source/ppqsort.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -94,8 +94,7 @@ register_benchmark_default(sort_signature, sort_name, Rotated, int, default, pre
register_benchmark_default(sort_signature, sort_name, Rotated, double, default, prepare_code); \
register_benchmark_default(sort_signature, sort_name, Heap, short, default, prepare_code); \
register_benchmark_default(sort_signature, sort_name, Heap, int, default, prepare_code); \
register_benchmark_default(sort_signature, sort_name, Heap, double, default, prepare_code); \
register_benchmark_size(sort_signature, sort_name, Adversary, int, size##_##1e8, int(1e8), prepare_code);
register_benchmark_default(sort_signature, sort_name, Heap, double, default, prepare_code);

#define small_range_benchmark(sort_signature, sort_name, prepare_code) \
register_benchmark_range(sort_signature, sort_name, Random, int, small_range##_##1, 0, 0, prepare_code); \
Expand Down Expand Up @@ -131,6 +130,65 @@ register_matrix_benchmark(sort_signature, sort_name, SparseMatrixSoA, af_shell10
register_matrix_benchmark(sort_signature, sort_name, SparseMatrixSoA, cage15, false, prepare_code) \
register_matrix_benchmark(sort_signature, sort_name, SparseMatrixSoA, fem_hifreq_circuit, true, prepare_code)

#define prepare_adversary(Size) \
int candidate = 0; \
int nsolid = 0; \
const int gas = Size - 1; \
std::ranges::fill(data_, gas); \
std::vector<int> asc_vals(Size); \
std::iota(asc_vals.begin(), asc_vals.end(), 0); \
auto cmp = [&](int x, int y) { \
if (data_[x] == gas && data_[y] == gas) { \
if (x == candidate) data_[x] = nsolid++; \
else data_[y] = nsolid++; \
} \
if (this->data_[x] == gas) candidate = x; \
else if (data_[y] == gas) candidate = y; \
return data_[x] < data_[y]; \
};

#define adversary_benchmark(sort_signature, sort_name, prepare_code) \
register_benchmark_size(sort_signature, sort_name, Adversary, int, size##_##1e8, int(1e8), prepare_code)

#define adversary_benchmarks_all \
adversary_benchmark(ppqsort::sort(ppqsort::execution::par, data_.begin(), data_.end()), ppqsort_par, \
prepare_adversary(int(1e8)); \
ppqsort::sort(ppqsort::execution::par_force_branchless, asc_vals.begin(), asc_vals.end(), cmp);) \
adversary_benchmark(__gnu_parallel::sort(data_.begin(), data_.end(), __gnu_parallel::balanced_quicksort_tag()), bqs, \
prepare_adversary(int(1e8)); \
__gnu_parallel::sort(asc_vals.begin(), asc_vals.end(), cmp, __gnu_parallel::balanced_quicksort_tag());) \
adversary_benchmark(boost::sort::block_indirect_sort(data_.begin(), data_.end()), block_indirect_sort, \
prepare_adversary(int(1e8)); \
boost::sort::block_indirect_sort(asc_vals.begin(), asc_vals.end(), cmp);) \
adversary_benchmark(std::sort(std::execution::par, data_.begin(), data_.end()), std_sort_par, \
prepare_adversary(int(1e8)); \
std::sort(std::execution::par, asc_vals.begin(), asc_vals.end(), cmp);) \
adversary_benchmark(std::sort(poolstl::par, data_.begin(), data_.end()), poolstl_sort_par, \
prepare_adversary(int(1e8)); \
std::sort(poolstl::par, asc_vals.begin(), asc_vals.end(), cmp);) \
adversary_benchmark(tbb::parallel_sort(data_.begin(), data_.end()), tbb_parallel_sort, \
prepare_adversary(int(1e8)); \
tbb::parallel_sort(asc_vals.begin(), asc_vals.end(), cmp);) \
adversary_benchmark(thrust::sort(thrust::host, data_.begin(), data_.end()), thrust_device_sort, \
prepare_adversary(int(1e8)); \
thrust::sort(thrust::host, asc_vals.begin(), asc_vals.end(), cmp);) \
adversary_benchmark(mpqsort::sort(mpqsort::execution::par, data_.begin(), data_.end()), mpqsort_par, \
prepare_adversary(int(1e8)); \
mpqsort::sort(mpqsort::execution::par, asc_vals.begin(), asc_vals.end(), cmp);) \
adversary_benchmark(ips4o::parallel::sort(data_.begin(), data_.end()), ips4o, \
prepare_adversary(int(1e8)); \
ips4o::parallel::sort(asc_vals.begin(), asc_vals.end(), cmp);) \
adversary_benchmark(aqsort::sort(data_.size(), &cmp_normal, &sw), aqsort, \
auto sw = [&](std::size_t i, std::size_t j) { std::swap(data_[i], data_[j]); }; \
auto cmp_normal = [&](std::size_t i, std::size_t j) { return data_[i] < data_[j]; }; \
prepare_adversary(int(1e8)); \
aqsort::sort(asc_vals.size(), &cmp, &sw);)
/* CPP11Sort tbd:
adversary_benchmark(cpp11sort::sort(data_.begin(), data_.end()), cpp11sort_par, \
prepare_adversary(int(1e8)); \
cpp11sort::sort(asc_vals.begin(), asc_vals.end(), cmp);) \
*/

#define complete_benchmark_set(sort_signature, sort_name, prepare_code) \
default_benchmark(sort_signature, sort_name, prepare_code) \
small_range_benchmark(sort_signature, sort_name, prepare_code) \
Expand Down Expand Up @@ -183,4 +241,7 @@ complete_benchmark_set_aq(aqsort::sort(data_.size(), &cmp, &sw), aqsort,
);

complete_benchmark_set(mpqsort::sort(mpqsort::execution::par, data_.begin(), data_.end()), mpqsort_par, "");
complete_benchmark_set(ips4o::parallel::sort(data_.begin(), data_.end()), ips4o, "");
complete_benchmark_set(ips4o::parallel::sort(data_.begin(), data_.end()), ips4o, "");

// Adversary benchmarks, can take a lot of time
adversary_benchmarks_all

0 comments on commit 041f56e

Please sign in to comment.