Skip to content

Commit

Permalink
WIP/concurrency
Browse files Browse the repository at this point in the history
  • Loading branch information
@melton1968
melton1968 committed Jun 9, 2023
1 parent ae2f7dd commit 2df3b1c
Show file tree
Hide file tree
Showing 2 changed files with 44 additions and 172 deletions.
32 changes: 32 additions & 0 deletions include/core/sort/array_ring.h
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,32 @@
// Copyright (C) 2022, 2023 by Mark Melton
//

#pragma once
#include <algorithm>

namespace core::sort {

template<class T, size_t Size, size_t Align>
class ArrayRing {
public:
auto size() const { return head_ - tail_; }
auto& next() { return data_[head_ % Size]; }
auto& front() { return data_[tail_ % Size]; }
const auto& front() const { return data_[tail_ % Size]; }
auto& back() { return data_[(head_ + Size - 1) % Size]; }
const auto& back() const { return data_[(head_ + Size - 1) % Size]; }
auto& head_index() { return head_; }
const auto& head_index() const { return head_; }
auto& tail_index() { return tail_; }
const auto& tail_index() const { return tail_; }
T pop_front() { return data_[tail_++ % Size]; }
void push_back(const T& value) { data_[head_++ % Size] = value; }
T pop_back() { return data_[--head_ % Size]; }
private:
alignas(Align) std::array<T, Size> data_;
size_t head_{}, tail_{};
};



}; // core::sort
184 changes: 12 additions & 172 deletions src/tools/sort6.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,9 @@
#include <array>
namespace core::sort {

#define REPEAT4(code) code; code; code; code
#define REPEAT8(code) REPEAT4(code); REPEAT4(code)

inline namespace qsort_detail {
inline constexpr size_t InsertionSortThreshold = 24;
inline constexpr size_t PseudoMedianThreshold = 128;
Expand Down Expand Up @@ -93,66 +96,6 @@ auto partition_basic(Iter begin, Iter end, Compare compare) {
return piter;
}

template<class T, size_t Size, size_t Align>
class array_ring {
public:
auto size() const {
return head_ - tail_;
}

auto& next() {
return data_[head_ % Size];
}

auto& front() {
return data_[tail_ % Size];
}

const auto& front() const {
return data_[tail_ % Size];
}

auto& back() {
return data_[(head_ + BlockSize - 1) % Size];
}

const auto& back() const {
return data_[(head_ + BlockSize - 1) % Size];
}

auto& head_index() {
return head_;
}

const auto& head_index() const {
return head_;
}

auto& tail_index() {
return tail_;
}

const auto& tail_index() const {
return tail_;
}

T pop_front() {
return data_[tail_++ % Size];
}

void push_back(const T& value) {
data_[head_++ % Size] = value;
}

T pop_back() {
return data_[--head_ % Size];
}

private:
alignas(Align) std::array<T, Size> data_;
size_t head_{}, tail_{};
};

template<class Iter, class Compare>
void output_partition(std::ostream& os, Iter base, Iter begin, Iter end, auto pivot,
Compare compare) {
Expand All @@ -172,92 +115,6 @@ bool paritition_ordered(Iter begin, Iter end, auto pivot, Compare compare, bool
return true;
}

template<class Iter, class Compare>
auto partition0(Iter begin, Iter end, Compare compare) {
using std::swap;
using T = typename Iter::value_type;

T pivot = std::move(*begin);
Iter liter = begin;
Iter riter = end;

array_ring<unsigned char, BlockSize, CacheLineSize> lring, rring;
Iter lbase, rbase;
while (riter - liter > 1) {
if (lring.size() == 0)
lbase = liter + 1;

if (rring.size() == 0)
rbase = riter - 1;

size_t loff = liter - lbase + 1;
size_t roff = rbase - riter + 1;

size_t nspace = riter - liter - 1;
size_t nl = std::min(256 - loff, BlockSize - lring.size());
size_t nr = std::min(256 - roff, BlockSize - rring.size());

if (nl + nr > nspace) {
auto& a = (nl < nr ? nl : nr);
auto& b = (nl < nr ? nr : nl);
a = std::min(a, nspace / 2);
b = nspace - a;
}

constexpr auto Unroll = 4;

if (nl > 0) {
while (nl >= Unroll) {
lring.next() = loff++; lring.head_index() += not compare(*++liter, pivot);
lring.next() = loff++; lring.head_index() += not compare(*++liter, pivot);
lring.next() = loff++; lring.head_index() += not compare(*++liter, pivot);
lring.next() = loff++; lring.head_index() += not compare(*++liter, pivot);
nl -= Unroll;
}
while (nl-- > 0) {
lring.next() = loff++;
lring.head_index() += not compare(*++liter, pivot);
}
}

if (nr > 0) {
while (nr >= Unroll) {
rring.next() = roff++; rring.head_index() += compare(*--riter, pivot);
rring.next() = roff++; rring.head_index() += compare(*--riter, pivot);
rring.next() = roff++; rring.head_index() += compare(*--riter, pivot);
rring.next() = roff++; rring.head_index() += compare(*--riter, pivot);
nr -= Unroll;
}
while (nr-- > 0) {
rring.next() = roff++;
rring.head_index() += compare(*--riter, pivot);
}
}

auto n = std::min(lring.size(), rring.size());
for (auto i = 0; i < n; ++i)
std::iter_swap(lbase + lring.pop_front(), rbase - rring.pop_front());
}

if (lring.size() > 0) {
while (lring.size())
std::iter_swap(lbase + lring.pop_back(), --riter);
liter = riter - 1;
}

if (rring.size() > 0) {
while (rring.size())
std::iter_swap(rbase - rring.pop_back(), ++liter);
riter = liter + 1;
}

Iter piter = liter;
*begin = std::move(*piter);
*piter = std::move(pivot);

return piter;
}

template<class Iter, class Compare>
auto partition(Iter begin, Iter end, Compare compare) {
using std::swap;
Expand All @@ -274,32 +131,17 @@ auto partition(Iter begin, Iter end, Compare compare) {
size_t ldx{}, rdx{};
Iter lbase = liter + 1, rbase = riter - 1;
while (riter - liter > 2 * BlockSize) {

if (ldx == 0) {
lbase = liter + 1;
for (auto i = 0; i < BlockSize;) {
loffsets[ldx] = i++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = i++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = i++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = i++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = i++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = i++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = i++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = i++; ldx += not compare(*++liter, pivot);
REPEAT8(loffsets[ldx] = i++; ldx += not compare(*++liter, pivot));
}
}

if (rdx == 0) {
rbase = riter - 1;
for (auto i = 0; i < BlockSize;) {
roffsets[rdx] = i++; rdx += compare(*--riter, pivot);
roffsets[rdx] = i++; rdx += compare(*--riter, pivot);
roffsets[rdx] = i++; rdx += compare(*--riter, pivot);
roffsets[rdx] = i++; rdx += compare(*--riter, pivot);
roffsets[rdx] = i++; rdx += compare(*--riter, pivot);
roffsets[rdx] = i++; rdx += compare(*--riter, pivot);
roffsets[rdx] = i++; rdx += compare(*--riter, pivot);
roffsets[rdx] = i++; rdx += compare(*--riter, pivot);
REPEAT8(roffsets[rdx] = i++; rdx += compare(*--riter, pivot));
}
}

Expand All @@ -315,20 +157,17 @@ auto partition(Iter begin, Iter end, Compare compare) {
}

while (riter - liter > 1) {
constexpr auto Unroll = 4;
size_t nspace = riter - liter - 1;
size_t nl = ldx == 0 ? std::min(rdx == 0 ? nspace / 2 : nspace, BlockSize) : 0;
size_t nr = rdx == 0 ? std::min(nspace - nl, BlockSize) : 0;

constexpr auto Unroll = 4;

if (nl > 0) {
lbase = liter + 1;
size_t loff{};
while (nl >= Unroll) {
loffsets[ldx] = loff++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = loff++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = loff++; ldx += not compare(*++liter, pivot);
loffsets[ldx] = loff++; ldx += not compare(*++liter, pivot);
REPEAT4(loffsets[ldx] = loff++; ldx += not compare(*++liter, pivot));
nl -= Unroll;
}
while (nl-- > 0) {
Expand All @@ -341,10 +180,7 @@ auto partition(Iter begin, Iter end, Compare compare) {
rbase = riter - 1;
size_t roff{};
while (nr >= Unroll) {
roffsets[rdx] = roff++; rdx += compare(*--riter, pivot);
roffsets[rdx] = roff++; rdx += compare(*--riter, pivot);
roffsets[rdx] = roff++; rdx += compare(*--riter, pivot);
roffsets[rdx] = roff++; rdx += compare(*--riter, pivot);
REPEAT4(roffsets[rdx] = roff++; rdx += compare(*--riter, pivot));
nr -= Unroll;
}
while (nr-- > 0) {
Expand Down Expand Up @@ -401,6 +237,9 @@ void qsort(Iter begin, Iter end, Compare compare) {
}
}

#undef REPEAT4
#undef REPEAT8

}; // core::sort

using std::cout, std::endl;
Expand All @@ -427,3 +266,4 @@ int main(int argc, const char *argv[]) {

return 0;
}

0 comments on commit 2df3b1c

Please sign in to comment.