1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
#pragma once
#include "common/Timer.hpp"
#include "common/thread.hpp"
#include <atomic>
#include <cstdint>
#include <span>
#include <vector>
class ParallelRadixSort {
public:
void sort(std::vector<uint32_t>& data, std::vector<uint32_t>& buffer, bool verbose = false)
{
int const num_bits = 32; // Assuming 32-bit integers
int const radix_bits = 8; // on pi 5: 30/10 is fastest.
int const radix = 1 << radix_bits; // Base (e.g., 8 bits at a time)
int const radix_mask = radix - 1;
int const num_passes = num_bits / radix_bits;
int const num_threads = std::thread::hardware_concurrency();
Timer timer;
if (verbose) {
std::cout << "ParallelRadixSort: Sorting " << data.size() << " elements with "
<< num_threads << " threads" << std::endl;
timer.start();
}
std::vector<std::vector<int>> counts_by_thread(num_threads, std::vector<int>(radix, 0));
// get each threads start and end index
size_t const num_elements_per_thread = data.size() / num_threads;
for (int pass = 0; pass < num_passes; ++pass) {
if (verbose)
std::cout << "----- Pass " << pass << " -----" << std::endl;
int shift = pass * radix_bits;
// Count phase
Timer countPhaseTimer;
if (verbose)
countPhaseTimer.start("Count phase");
{
std::vector<thread> threads;
for (int t = 0; t < num_threads; ++t) {
threads.emplace_back([&, t]() {
// fill counts to zero
for (int r = 0; r < radix; ++r) {
counts_by_thread[t][r] = 0;
}
size_t start = num_elements_per_thread * t;
size_t end = (t == num_threads - 1) ? data.size()
: num_elements_per_thread * (t + 1);
for (size_t i = start; i < end; ++i) {
uint32_t key = (data[i] >> shift) & radix_mask;
counts_by_thread[t][key]++;
}
});
}
}
if (verbose) {
countPhaseTimer.stop();
countPhaseTimer.start("Prefix sum phase");
}
// now merge all counts into one global count
std::vector<uint32_t> counts(radix, 0);
for (int t = 0; t < num_threads; ++t) {
for (int r = 0; r < radix; ++r) {
counts[r] += counts_by_thread[t][r];
}
}
// Prefix sum phase
std::vector<std::vector<int>> offsets_for_thread(
num_threads, std::vector<int>(radix, 0));
// first get global offsets
std::vector<uint32_t> offsets(radix, 0);
for (int i = 1; i < radix; ++i) {
offsets[i] = offsets[i - 1] + counts[i - 1];
}
// then update offsets for each thread, with thread 0 using the global offsets and
// others threads building on that. t 0 offsets first
for (int r = 0; r < radix; ++r) {
offsets_for_thread[0][r] = offsets[r];
}
// t > 0 offsets build on previous ones
for (int i = 1; i < num_threads; i++) {
for (int r = 0; r < radix; ++r) {
offsets_for_thread[i][r]
= offsets_for_thread[i - 1][r] + counts_by_thread[i - 1][r];
}
}
if (verbose) {
countPhaseTimer.stop();
}
// now we know each threads own bucket counts, so when the thread scans the same data,
// it can place the data in the correct bucket with the offset assigned to it from all
// the threads.
// Redistribution phase
Timer redistributionTimer;
if (verbose)
redistributionTimer.start("Redistribution phase");
{
std::vector<thread> threads;
for (int t = 0; t < num_threads; ++t) {
threads.emplace_back([&, t]() {
size_t start = num_elements_per_thread * t;
size_t end = (t == num_threads - 1) ? data.size()
: num_elements_per_thread * (t + 1);
for (size_t i = start; i < end; ++i) {
uint32_t key = (data[i] >> shift) & radix_mask;
int outpos = offsets_for_thread[t][key]++;
buffer[outpos] = data[i];
}
});
}
}
redistributionTimer.stop();
std::swap(data, buffer);
}
}
void sortByKey(std::vector<uint32_t>& keys,
std::vector<uint32_t>& values,
std::vector<uint32_t>& keyBuffer,
std::vector<uint32_t>& valueBuffer,
int num_bits,
int radix_bits = -1,
bool verbose = false)
{
// num_bits = 28; // Assuming 32-bit integers
// radix_bits = 14; // on pi 5: 30/10 is fastest.
if (radix_bits == -1) {
if (num_bits == 30) {
// on pi 5: 30/10 is fastest.
// radix_bits = 15; // 1139ms ryzen 5600
radix_bits = 10; // 515ms ryzen 5600
// radix_bits = 8; // 478ms ryzen 5600
// radix_bits = 6; // 734ms ryzen 5600
}
else if (num_bits == 28) {
// on pi5 28/14 is fastest?
radix_bits = 10; // 128ms ryzen 5600
// radix_bits = 14; // 143ms ryzen 5600 <-- use for pi5?
// radix_bits = 8; // 197ms ryzen 5600
// radix_bits = 7; // 195ms ryzen 5600
}
else {
radix_bits = 8;
}
}
int radix = 1 << radix_bits; // Base (e.g., 8 bits at a time)
int radix_mask = radix - 1;
int const num_passes = (num_bits + radix_bits - 1) / radix_bits;
int const num_threads = std::thread::hardware_concurrency();
if (verbose) {
std::cout << "ParallelRadixSort: Sorting " << keys.size() << " key-value pairs with "
<< num_threads << " threads" << std::endl;
}
std::vector<std::vector<int>> counts_by_thread(num_threads, std::vector<int>(radix, 0));
size_t const num_elements_per_thread = keys.size() / num_threads;
for (int pass = 0; pass < num_passes; ++pass) {
if (verbose) {
std::cout << "----- Pass " << pass << " -----" << std::endl;
}
int shift = pass * radix_bits;
if (pass == num_passes - 1) {
// For the last pass, we might not need all radix_bits.
radix_bits = (num_bits - shift);
radix = 1 << radix_bits;
radix_mask = (1 << radix_bits) - 1;
}
// Count phase
Timer countPhaseTimer;
if (verbose)
countPhaseTimer.start("Count phase");
// Count phase
{
std::vector<thread> threads;
for (int t = 0; t < num_threads; ++t) {
threads.emplace_back([&, t]() {
for (int r = 0; r < radix; ++r) {
counts_by_thread[t][r] = 0;
}
size_t start = num_elements_per_thread * t;
size_t end = (t == num_threads - 1) ? keys.size()
: num_elements_per_thread * (t + 1);
for (size_t i = start; i < end; ++i) {
uint32_t key = (keys[i] >> shift) & radix_mask;
counts_by_thread[t][key]++;
}
});
}
}
if (verbose) {
countPhaseTimer.stop();
countPhaseTimer.start("Prefix sum phase");
}
// Merge counts
std::vector<uint32_t> counts(radix, 0);
for (int t = 0; t < num_threads; ++t) {
for (int r = 0; r < radix; ++r) {
counts[r] += counts_by_thread[t][r];
}
}
// Prefix sum phase
std::vector<std::vector<int>> offsets_for_thread(
num_threads, std::vector<int>(radix, 0));
std::vector<uint32_t> offsets(radix, 0);
for (int i = 1; i < radix; ++i) {
offsets[i] = offsets[i - 1] + counts[i - 1];
}
for (int r = 0; r < radix; ++r) {
offsets_for_thread[0][r] = offsets[r];
}
for (int i = 1; i < num_threads; i++) {
for (int r = 0; r < radix; ++r) {
offsets_for_thread[i][r]
= offsets_for_thread[i - 1][r] + counts_by_thread[i - 1][r];
}
}
if (verbose) {
countPhaseTimer.stop();
countPhaseTimer.start("Redistribution phase");
}
// Redistribution phase
{
std::vector<thread> threads;
for (int t = 0; t < num_threads; ++t) {
threads.emplace_back([&, t]() {
size_t start = num_elements_per_thread * t;
size_t end = (t == num_threads - 1) ? keys.size()
: num_elements_per_thread * (t + 1);
for (size_t i = start; i < end; ++i) {
uint32_t key = (keys[i] >> shift) & radix_mask;
int outpos = offsets_for_thread[t][key]++;
keyBuffer[outpos] = keys[i];
valueBuffer[outpos] = values[i];
}
});
}
}
if (verbose) {
countPhaseTimer.stop();
}
std::swap(keys, keyBuffer);
std::swap(values, valueBuffer);
}
}
};