-
Notifications
You must be signed in to change notification settings - Fork 7
/
partition.h
384 lines (306 loc) · 11.4 KB
/
partition.h
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
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
#ifndef CUTOFFFINDER_H
#define CUTOFFFINDER_H
#include <iostream>
#include <numeric>
#include <stdexcept>
#include <vector>
#include <stdint.h>
#include <numeric>
#include <queue>
#include <thrust/device_vector.h>
#include <cuda_runtime_api.h>
#include <sys/time.h>
#include "util/timer.h"
#include "matrix/matrix_input.h"
#include "matrix/factory.h"
using namespace std;
struct Slice{
uint32_t sum;
uint32_t num_rows;
std::vector<uint32_t> row_start;
Slice() : sum(0), num_rows(0){
row_start.clear();
}
void addRow(uint32_t index, uint32_t weight){
sum += weight;
num_rows++;
row_start.push_back(index);
}
bool operator()(const Slice* lhs, const Slice* rhs){
return rhs->sum < lhs->sum;
}
};
struct Result {
string format;
uint32_t startingRow;
uint32_t numRows;
uint32_t rowsPerSlice;
double avgDist;
double time;
uint64_t nnz;
uint32_t ntimes;
string toString() {
ostringstream ostr;
ostr << format << " - row " << startingRow << ", " << numRows
<< " rows, " << rowsPerSlice << " rows per slice, " << nnz << " nnz, avg d:" << avgDist;
return ostr.str();
}
};
const uint32_t NTIMES = 10;
const string FORMATS[] = { // matrix_factory::BIT_BLOCK,
matrix_factory::SLICED_COO_256,
matrix_factory::SLICED_COO_128,
matrix_factory::SLICED_COO_64,
matrix_factory::SLICED_COO_32,
matrix_factory::SLICED_COO_16,
matrix_factory::SLICED_COO_8,
matrix_factory::SLICED_COO_4,
matrix_factory::SLICED_COO_2,
matrix_factory::SLICED_COO_1,
};
uint32_t FORMATS_LEN = (sizeof(FORMATS) / sizeof(FORMATS[0]));
struct CutOffFinderOutput {
vector<string> formats;
vector<uint32_t> startRows;
vector<uint32_t> numRows;
vector<uint32_t> rowsPerSlice;
};
template <class ValueType>
class CutOffFinder {
private:
csr_matrix &mat;
uint32_t numProcs;
MMFormatInput &mf;
vector<uint32_t> new_perm;
vector<uint32_t> prefsum;
CutOffFinderOutput output;
vector<Matrix<ValueType> *> mfh;
thrust::device_vector<ValueType> g_input, g_output;
bool sortcol;
void test();
public:
CutOffFinder( MMFormatInput &f, bool sortcol ): mf(f), mat(f.data) {
this->sortcol = sortcol;
prefsum.resize( mat.num_rows, 0 );
prefsum[ mf.perm[0] ] = mat.row_offsets[ mf.perm[0] + 1 ] - mat.row_offsets[ mf.perm[0] ];
for (int i=1;i<mat.num_rows;i++) {
prefsum[ mf.perm[i] ] = prefsum[ mf.perm[i-1] ] + getCSRNumRows(mat, mf.perm[i] ); //(mat.row_offsets[ mf.perm[i] + 1] - mat.row_offsets[ mf.perm[i] ]);
}
}
Result getPerformance(const string &fmt, uint32_t currentRow);
void execute(bool verbose);
void balance( const CutOffFinderOutput &cutOffOutput, uint32_t numProcs, vector<uint32_t> &permutation);
void printResult(string filename);
void writeCache(string filename);
double theta;
};
template<class T>
Result CutOffFinder<T>::getPerformance(const string &fmt, uint32_t currentRow) {
Result result;
Matrix<T>* bm = matrix_factory::getMatrixObject<T>(fmt);
int numProcs = this->numProcs;
uint32_t minRows = bm->granularity() * numProcs;
if (mat.num_rows - currentRow < minRows) {
minRows = mat.num_rows - currentRow;
}
uint32_t nnz = 0;
nnz = prefsum[ mf.perm[currentRow + minRows - 1] ];
if (currentRow > 0)
nnz -= prefsum[ mf.perm[currentRow - 1] ];
result.avgDist = (nnz*1.0/minRows) / ( SHARED_MEMORY_SIZE*1.0 / bm->granularity() / sizeof(T) ); // ( nnz / ((double) SHARED_MEMORY_SIZE / sizeof(T)) / numProcs );
result.format = fmt;
result.nnz = nnz;
result.ntimes = NTIMES;
result.startingRow = currentRow;
result.numRows = minRows;
result.rowsPerSlice = bm->granularity();
delete bm;
return result;
}
template <class T>
void CutOffFinder<T>::execute (bool verbose) {
cout << " executing ... ";
uint32_t currentRow = 0;
vector<Result> resultVector;
this->numProcs = getNumMultiprocessors();
double min_per_shared = theta;
double max_per_shared = theta*2;
if (verbose) cout << "MIN_ROWS_PER_SHARED: " << min_per_shared << endl;
if (verbose) cout << "MAX_ROWS_PER_SHARED: " << max_per_shared << endl;
for (uint32_t i = 0; i < FORMATS_LEN - 1;) {
if (verbose) cout << "Getting result for " << FORMATS[i] << endl;
Result result = getPerformance(FORMATS[i], currentRow);
if (verbose) cout << "Evaluate " << result.toString() << endl;
if (result.avgDist < min_per_shared)
for (uint32_t j = i + 1; j < FORMATS_LEN; j++) {
Result comp = getPerformance(FORMATS[j], currentRow);
if (verbose) cout << comp.toString() << endl;
if (result.avgDist < min_per_shared){
result = comp;
i = j;
continue;
}
if (comp.avgDist >= min_per_shared && comp.avgDist <= max_per_shared) {
result = comp;
i=j;
}
else if (comp.avgDist > max_per_shared) {
break;
}
}
if (verbose) cout << "Chosen: " << result.toString() << endl;
resultVector.push_back(result);
currentRow += result.numRows;
if (currentRow >= mat.num_rows) break;
}
if (verbose) cout << "Result:" << endl;
for (uint32_t i = 0; i < resultVector.size(); i++) {
if (verbose) cout << resultVector[i].toString() << endl;
output.formats.push_back(resultVector[i].format);
output.startRows.push_back(resultVector[i].startingRow);
output.numRows.push_back(resultVector[i].numRows);
output.rowsPerSlice.push_back(resultVector[i].rowsPerSlice);
}
if (verbose) cout << endl;
string lastformat = output.formats.back();
uint32_t curNumRows = accumulate(output.numRows.begin(), output.numRows.end(), 0);
Matrix<T>* lastf = matrix_factory::getMatrixObject<T>( lastformat );
for(int i=0; i<2; i++){
uint32_t nr = lastf->granularity() * numProcs;
if(nr + curNumRows <= mat.num_rows){
output.formats.push_back(lastformat);
output.startRows.push_back(curNumRows);
output.numRows.push_back(nr);
output.rowsPerSlice.push_back(lastf->granularity());
if (verbose) cout << "Adding " << lastformat << " from row " << curNumRows << " with " << nr << " rows" << endl;
curNumRows += nr;
} else{
break;
}
}
delete lastf;
balance(output, numProcs ,new_perm);
cout << " OK " << endl;
}
template <class T>
void CutOffFinder<T>::balance(
const CutOffFinderOutput &cutOffOutput,
uint32_t numProcs,
vector<uint32_t> &permutation){
uint32_t baseRow = 0;
permutation.clear();
permutation.reserve(max(mat.num_rows, mat.num_cols));
const vector<uint32_t> &oldPerm = mf.perm;
// for each horizontal split
uint32_t curOffset = 0;
// cout << "Base row: " << baseRow << endl;
for(uint32_t i=0; i<cutOffOutput.formats.size(); i++){
if(cutOffOutput.formats[i].rfind("scoo") != string::npos){
int numProcss = numProcs;
vector<Slice> slices(numProcss);
priority_queue<Slice*, vector<Slice*>, Slice> pq;
for(uint32_t j=0; j<numProcss; ++j){
pq.push(&slices[j]);
}
for(uint32_t j=0; j<cutOffOutput.numRows[i]; ++j){
Slice *sc = pq.top();
pq.pop();
uint32_t index = baseRow+curOffset+j;
sc->addRow(oldPerm[index], mat.row_offsets[oldPerm[index] + 1] - mat.row_offsets[oldPerm[index]] );
if(sc->num_rows < cutOffOutput.rowsPerSlice[i]){
pq.push(sc);
}
}
for(uint32_t j=0; j<numProcss; ++j){
permutation.insert(permutation.end(), slices[j].row_start.begin(), slices[j].row_start.end());
//cout << slices[j].sum << " ";
}
//cout << endl;
}else{
permutation.insert(
permutation.end(),
oldPerm.begin()+baseRow+curOffset,
oldPerm.begin()+baseRow+curOffset+cutOffOutput.numRows[i]);
}
curOffset += cutOffOutput.numRows[i];
}
uint32_t nextBaseRow = baseRow + max(mat.num_rows, mat.num_cols);
permutation.insert(permutation.end(),
oldPerm.begin()+baseRow+curOffset,
oldPerm.begin()+nextBaseRow);
baseRow = nextBaseRow;
}
template <class T>
void CutOffFinder<T>::printResult(string mfile) {
cout << " printResult ... " ;
char outfile[255];
sprintf(outfile, FORMAT_PERM, mfile.c_str());
ofstream permout( outfile );
DataOutputStream out2(permout);
out2.writeVector( new_perm );
permout.close();
sprintf(outfile, FORMAT_OUTPUT, mfile.c_str());
ofstream out;
out.open(outfile);
out.exceptions(ios::failbit | ios::badbit);
for (int i=1; i<output.formats.size(); ) {
if (output.formats[i] == output.formats[i-1]){
output.formats.erase(output.formats.begin() + i);
output.startRows.erase(output.startRows.begin() + i);
}
else i++;
}
out << output.formats.size() << endl;
copy(output.formats.begin(), output.formats.end(), //+firstLscooLoc,
ostream_iterator<string> (out, " "));
out << endl;
// --- output start row --- //
copy(output.startRows.begin(), output.startRows.end(), //+ firstLscooLoc,
ostream_iterator<uint32_t> (out, " "));
out << endl;
out.close();
cout << " OK " << endl;
}
template<class T>
void CutOffFinder<T>::test(){
g_input.resize(max(mat.num_rows, mat.num_cols),1);
g_output.resize(max(mat.num_rows, mat.num_cols),0);
T* v = p(g_input);
T* r = p(g_output);
uint32_t cur_row = 0;
for(uint32_t j=0; j<mfh.size(); j++){
mfh[j]->multiply(v, r+cur_row);
cur_row += mfh[j]->get_num_rows();
}
cudaDeviceSynchronize();
checkCUDAError("kernel finish");
}
template <class T>
void CutOffFinder<T>::writeCache(string output_path){
cout << " writing cache ... ";
uint32_t currentRow = 0;
output.startRows.push_back(mat.num_rows);
mf.perm = new_perm;
for(int i=0; i<output.startRows.size()-1; i++){
Matrix<T> *bm = matrix_factory::getMatrixObject<T>(output.formats[i]);
bm->set_num_cols(mat.num_cols);
bm->set_num_rows(output.startRows[i+1] - output.startRows[i]);
bm->set_sort_col(sortcol);
cout << "Building cache for " << output.formats[i] << endl;
bm->readMatrix(mf, new_perm);
stringstream cacheOutput;
cacheOutput << output_path << "-" << currentRow << bm->getCacheName();
cout << "Writing cache to " << cacheOutput.str() << endl;
ofstream out(cacheOutput.str().c_str());
out.exceptions(ios_base::failbit | ios_base::badbit);
bm->buildCache(out);
out.close();
bm->transferToDevice();
currentRow += bm->get_num_rows();
mfh.push_back(bm);
}
cout << "Testing 1 iteration ... " ;
test();
cout << " All OK." << endl;
}
#endif