mklspmm.c 25.7 KB
Newer Older
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
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
/*includes {{{*/
#include <stdio.h>
#include <assert.h>
#include <math.h>
#include "mkl.h"
#include "mkl_types.h"
#include "mkl_spblas.h"

#include "commonmm.h"
#include "mmio.h"
/*}}}*/
/* global varibles and defines {{{*/
int option_host;
#define OPTION_HOST_CPU -1
#define OPTION_HOST_MIC 0
int micdev;
__declspec(target(mic)) int max_threads = 0;
__declspec(target(mic)) int nworker  = 0;
__declspec(target(mic)) int num_repeat;
__declspec(target(mic)) int option_print_matrices = 0;
#define OPTION_NOPRINT_MATRICES 0
#define OPTION_PRINT_MATRICES 1
__declspec(target(mic)) int option_printfile_matrices = 0;
#define OPTION_NOPRINTFILE_MATRICES 0
#define OPTION_PRINTFILE_MATRICES 1
__declspec(target(mic)) char transa = 'n';
__declspec(target(mic)) double time_mic_numeric_mm;
__declspec(target(mic)) double time_mic_symbolic_mm;
/*}}}*/
/* method declarations {{{*/
	int 
main(int argc, char* argv[]);
	__declspec(target(mic)) void 
printmm_one(int m, double* Aval, int* AJ, int* AI);
	__declspec(target(mic)) void 
printfilemm_one(char* file, int m, int n, double* Aval, int* AJ, int* AI);
	void 
printmm_zero(int m, double* Aval, int* AJ, int* AI);
/*}}}*/

/*csr util{{{*/
//#include "../../csr/sspmxm.h"
#define CS_MAX(a,b) (((a) > (b)) ? (a) : (b))

csr *csr_spfree(csr *A);
/* free workspace and return a sparse matrix result */
csr *csr_done(csr *C, void *w, void *x, csi ok)
{
//#ifdef KADEBUG
//      printf("%d:%s:%s():%d, %s(%lu,%lu,%lu)\n", me, __FILE__, __FUNCTION__,
//              __LINE__, C->name, C->m, C->n, C->nzmax);
//#endif
//      cs_free(w); /* free workspace */
//      cs_free(x);
        return(ok ? C : csr_spfree(C)); /* return result if OK, else free it */
}


/* wrapper for free */
void *cs_free(void *p)
{
        if (p)
                free(p); /* free p if it is not already NULL */
        return(NULL); /* return NULL to simplify the use of cs_free */
}
/* wrapper for realloc */
void *csr_realloc(void *p, csi n, size_t size, csi *ok)
{
        void *pnew = NULL;
        pnew = realloc(p, CS_MAX(n, 1) * size); /* realloc the block */
        *ok = (pnew != NULL); /* realloc fails if pnew is NULL */
        if (pnew == NULL) {
                printf("%d:reallocation failed, pnew is NULL\n", __LINE__);
        }
//printf("%s:%d: n=%d ok=%d\n", __FUNCTION__, __LINE__, n, *ok);
        return((*ok) ? pnew : p); /* return original p if failure */
}

/* wrapper for realloc */
void *cs_realloc(void *p, csi n, size_t size, csi *ok)
{
        void *pnew = NULL;
        pnew = realloc(p, CS_MAX(n, 1) * size); /* realloc the block */
        *ok = (pnew != NULL); /* realloc fails if pnew is NULL */
        if (pnew == NULL) {
                printf("reallocation failed\n");
        }
        return((*ok) ? pnew : p); /* return original p if failure */
}


/* change the max # of entries sparse matrix */
csi csr_sprealloc(csr *A, csi nzmax)
{
//printf("%d: nzmax=%d\n", __LINE__, nzmax);
        csi ok, oki = 0, okj = 1, okx = 1;
        if (!A)
                return(0);
        if (nzmax <= 0)
                nzmax = A->p[A->m];
        A->j = (int*)csr_realloc(A->j, nzmax, sizeof(csi), &oki);
        if (A->x)
                A->x = (csv*)csr_realloc(A->x, nzmax, sizeof(csv), &okx);
//printf("%d: ok=%d oki=%d okj=%d okx=%d\n", __LINE__, ok, oki, okj, okx);
        ok = (oki && okj && okx);
//printf("%d: ok=%d oki=%d okj=%d okx=%d\n", __LINE__, ok, oki, okj, okx);

//printf("%s:%d: A->nzmax=%d nzmax=%d\n", __FUNCTION__, __LINE__, A->nzmax, nzmax);
        if (ok)
                A->nzmax = nzmax;
//printf("%s:%d: A->nzmax=%d nzmax=%d\n", __FUNCTION__, __LINE__, A->nzmax, nzmax);
        return(ok);
}

/* free a sparse matrix */
csr *csr_spfree(csr *A)
{
        if (!A)
                return(NULL); /* do nothing if A already NULL */
        cs_free(A->p);
        A->p = NULL;
        cs_free(A->j);
        A->j = NULL;
        cs_free(A->x);
        A->x = NULL;
        cs_free(A->r);
        A->r = NULL;
//      cs_free(A->name);
        cs_free(A); /* free the cs struct and return NULL */
	return NULL;
}
/* allocate a sparse matrix (triplet form or compressed-ROW form) */
csr *csr_spalloc(csi m, csi n, csi nzmax, int values, int triplet, csv f) // floatType f is to suppress compile error
{
//printf("m=%d n=%d nzmax=%d\n", m, n, nzmax);
        csr* A = (csr*)calloc(1, sizeof(csr)); /* allocate the cs struct */
//      A->name = calloc(MTX_NAME_LEN, sizeof(char));
//      sprintf(A->name, "N/A");
        if (!A) {
                perror("sparse allocation failed");
                return(NULL); /* out of memory */
        }
        A->m = m; /* define dimensions and nzmax */
        A->n = n;
        A->nzmax = nzmax = CS_MAX(nzmax, 0);
//printf("A m=%d n=%d nzmax=%d\n", A->m, A->n, A->nzmax);
        A->nr = 0;  // number of nonzero rows
        A->p = (csi*)calloc(m + 2, sizeof(csi));
        A->j = (csi*)calloc(CS_MAX(nzmax,1), sizeof(csi));
        A->x = (csv*)calloc(CS_MAX(nzmax,1), sizeof(csv));
        return((!A->p || !A->j || !A->x) ? csr_spfree(A) : A);
}/*}}}*/
/*csr_multiply{{{*/
/* C = A*B 
If nonzero counts of matrices are zero or number of rows/cols is zero, it is advised not to call this function.
*/
// run 8 CUT 0.1 NNZ MTX ~/matrices/nlpkkt200.mtx ~/matrices/nlpkkt200.mtx tmp stdout MNAC SERIAL
csr *csr_multiply(csi Am, csi An, csi Anzmax, const csi* Ap, const csi* Aj, const csv* Ax, csi Bm, csi Bn, csi Bnzmax, const csi* Bp, const csi* Bj, const csv* Bx, long* nummult, csi* xb, csv* x)
{
	csv tf = 0;
	csi p, jp, j, kp, k, i, nz = 0, anz, *Cp, *Cj, m, n,
		bnz, values = 1;
	csv *Cx;
	csr *C;
	//printf("%s:%d\n", __FILE__, __LINE__);
	if (An != Bm)
		return(NULL);
	//printf("%s:%d\n", __FILE__, __LINE__);
	if (Anzmax == 0 || Bnzmax == 0) {
		C = csr_spalloc(Am, Bn, 0, values, 0, tf);
/*#ifdef KADEBUG
	printf("%d:%s:%s():%d, A(%s,%lu,%lu,%lu), B(%s,%lu,%lu,%lu),  \
		 One of matrice is Zero! C(%lu,%lu,%lu)\n", me, __FILE__, __FUNCTION__,
		__LINE__, A->name, A->m, A->n, A->nzmax, B->name, B->m, B->n,
		B->nzmax, C->m, C->n, C->nzmax);
#endif*/
		return C;
	}
//	printf("%s:%d\n", __FILE__, __LINE__);
	m = Am;
	anz = Ap[Am];
	n = Bn;
	bnz = Bp[Bm];
	for(i = 0; i < n; i++) xb[i] = 0;
	//csi* xb  = ka_calloc(n, sizeof(csi)); /* get workspace */
	for(i = 0; i < n; i++)
		xb[i] = 0;
	values = (Ax != NULL) && (Bx != NULL);
	//csv* x = values ? ka_calloc(n, sizeof(csv)) : NULL; /* get workspace */
	csi tnz = (anz + bnz) * 2;
//printf("A->m=%d B->n=%d tnz=%d\n", A->m, B->n, tnz);
	C = csr_spalloc(m, n, tnz, values, 0, tf); /* allocate result */
//	sprintf(C->name, "C=(%s)*(%s)", A->name, B->name);
/*#ifdef KADEBUG
	printf("%d:%s:%s():%d, A(%s,%lu,%lu,%lu), B(%s,%lu,%lu,%lu),  \
		allocated C(%lu,%lu,%lu)\n", me, __FILE__, __FUNCTION__,
		__LINE__, A->name, A->m, A->n, A->nzmax, B->name, B->m, B->n,
		B->nzmax, C->m, C->n, C->nzmax);
#endif*/
	if (!C || !xb || (values && !x))
		return (csr_done(C, xb, x, 0));
//	csi zero = 0;
//	for(i = 0; i < n; i++)
//		xb[i] = zero;
	Cp = C->p;
	//csi* oldj = C->j;
//printf("C->m=%d C->n=%d C->nzmax=%d C->j[0]=%d\n", C->m, C->n, C->nzmax, C->j[0]);
	for (i = 0; i < m; i++) {
//C->j != oldj ? printf("Changed old=%u new=%u\n", oldj, C->j):0;
		if ( ( (nz + n) > C->nzmax ) ) {
			if(!csr_sprealloc(C, (2 * (C->nzmax) + n) ) ) {
/*#ifdef KADEBUG
	printf("%d:%s:%s():%d, A(%s,%lu,%lu,%lu), B(%s,%lu,%lu,%lu), CANNOT BE\
		enlargened C(%lu,%lu,%lu)\n", me, __FILE__, __FUNCTION__, 
		__LINE__, A->name, A->m, A->n, A->nzmax, B->name, B->m, B->n, 
		B->nzmax, C->m, C->n, C->nzmax);
#endif	*/
				return (csr_done(C, xb, x, 0)); // out of memory
			} else {
/*#ifdef KADEBUG
	printf("%d:%s:%s():%d, A(%s,%lu,%lu,%lu), B(%s,%lu,%lu,%lu), enlargened \
		C(%lu,%lu,%lu)\n", me, __FILE__, __FUNCTION__, __LINE__, 
		A->name, A->m, A->n, A->nzmax, B->name, B->m, B->n, 
		B->nzmax, C->m, C->n, C->nzmax);
#endif		*/
			}
//printf("After expansion, C->nzmax=%d\n", C->nzmax);
		}
		Cj = C->j;
		Cx = C->x; /* C->j and C->x may be reallocated */
		//printf("i=%d C->j[0]=%d\n", i, C->j[0]);
		Cp[i] = nz; /* row i of C starts here */
		for (jp = Ap[i]; jp < Ap[i + 1]; jp++) {
			j = Aj[jp];
			for (kp = Bp[j]; kp < Bp[j + 1]; kp++) {
				k = Bj[kp]; /* B(i,j) is nonzero */
//if(k > n || k < 0)printf("k=%d kp=%d n=%d: %d\n", k, kp, n,__LINE__);
				if (xb[k] != i + 1) {
					xb[k] = i + 1; /* i is new entry in column j */
			//		if(nz > C->nzmax) {
			//			printf("%d: Error nz>nzmax %d>%d\n", __LINE__, nz, C->nzmax);
			//		}
//if(nz > C->nzmax || nz < 0)printf("nz: %d\n", __LINE__);
					Cj[nz++] = k; /* add i to pattern of C(:,j) */

					if (x) {
				//		C->j != oldj ? printf("B Changed old=%u new=%u\n", oldj, C->j):0;
						x[k] = Ax[jp] * Bx[kp]; /* x(i) = beta*A(i,j) */
				//		C->j != oldj ? printf("A Changed old=%u new=%u\n", oldj, C->j):0;
						(*nummult)++;
					}
				} else if (x) {
					x[k] += (Ax[jp] * Bx[kp]); /* i exists in C(:,j) already */
					(*nummult)++;
				}
			}
		}
		if (values)
			for (p = Cp[i]; p < nz; p++)
				Cx[p] = x[Cj[p]];
	}

//	printf("%s:%d nz=%d\n", __FILE__, __LINE__, nz);
	Cp[m] = nz; /* finalize the last row of C */
	csr_sprealloc(C, 0); /* remove extra space from C */
/*#ifdef KADEBUG
	printf("%d:%s:%s():%d, A(%s,%lu,%lu,%lu), B(%s,%lu,%lu,%lu), trimmed \
		C(%lu,%lu,%lu)\n", me, __FILE__, __FUNCTION__, __LINE__, 
		A->name, A->m, A->n, A->nzmax, B->name, B->m, B->n, 
		B->nzmax,  C->m, C->n, C->nzmax);
#endif*/
//	cs_free(xb);
	xb = NULL;
//	cs_free(x);
	x = NULL;
	return C;//(csr_done(C, xb, x, 1)); /* success; free workspace, return C */
}/*}}}*/


	void 
mkl_cpu_spmm(MKL_INT Am, MKL_INT An, MKL_INT Annz, double* Aval, MKL_INT* AJ, MKL_INT* AI, /*{{{*/
	MKL_INT Bn, MKL_INT Bnnz, double* Bval, MKL_INT* BJ, MKL_INT* BI, double** pCval, MKL_INT** pCJ, MKL_INT** pCI) { 

MKL_INT* CJ = NULL;
double* Cval = NULL;
MKL_INT sort = 3;	// sort everything
MKL_INT* CI = (MKL_INT*)mkl_malloc( (Am+2) * sizeof( MKL_INT ), 64 );
MKL_INT nnzmax = 0;	// nnzmax is zero in case of symbolic&numeric usage of mkl_?csrmultcsr
MKL_INT ierr;
MKL_INT request = 1;	// symbolic multiplication
mkl_dcsrmultcsr(&transa, &request, &sort, &Am, &An, &Bn, Aval, AJ, AI, Bval, BJ, BI, Cval, CJ, CI, &nnzmax, &ierr);

request = 2;		//numeric multiplication
int Cnnz = CI[Am]-1;
int Cval_size = Cnnz + 1;
CJ = (MKL_INT*)mkl_malloc( Cval_size	*	sizeof( MKL_INT ), 64 );
Cval = (double*)mkl_malloc( Cval_size	*	sizeof( double ), 64 );
mkl_dcsrmultcsr(&transa, &request, &sort, &Am, &An, &Bn, Aval, AJ, AI, Bval, BJ, BI, Cval, CJ, CI, &nnzmax, &ierr);
//printmm_one(Am, Cval, CJ, CI);//printf("Cnnz=%d\n", Cnnz);
*pCval = Cval; *pCJ = CJ; *pCI = CI; //printf("Cnnz_host:%d\n", Cnnz);
} /* ENDOF mkl_cpu_spmm }}}*/
	void
mkl_mic_spmm(MKL_INT Am, MKL_INT An, MKL_INT Annz, double* Aval, MKL_INT* AJ, MKL_INT* AI,/*{{{*/ 
	MKL_INT Bn, MKL_INT Bnnz, double* Bval, MKL_INT* BJ, MKL_INT* BI, double** pCval, MKL_INT** pCJ, MKL_INT** pCI) {

	int i;
	#pragma offload target(mic:micdev) inout(max_threads)
	{
		max_threads = mkl_get_max_threads();
		mkl_set_num_threads(nworker);
	}  // printf("MIC started \n"); fflush(stdout);
	#pragma offload target(mic:micdev) \
		in(transa) \
		in(Am) \
		in(An) \
		in(Bn) \
		in(Aval:length(Annz)	free_if(0)) \
		in(AI:length(Am+1)	free_if(0)) \
		in(AJ:length(Annz)	free_if(0)) \
		in(Bval:length(Bnnz)	free_if(0)) \
		in(BI:length(An+1)	free_if(0)) \
		in(BJ:length(Bnnz)	free_if(0)) 
	{}
	

	//void mkl_dcsrmultcsr (char *transa, MKL_INT *job, MKL_INT *sort, MKL_INT *m, MKL_INT *n, MKL_INT *k, double *a, MKL_INT *ja, MKL_INT *ia, double *b, MKL_INT *jb, MKL_INT *ib, double *c, MKL_INT *jc, MKL_INT *ic, MKL_INT *nnzmax, MKL_INT *ierr);


//	double s_initial = dsecnd();
//	double s_elapsed = dsecnd() - s_initial; // seconds

	MKL_INT	Cnnz_host = -1;
	MKL_INT* CI = NULL;
	MKL_INT* CJ = NULL;
	double* Cval = NULL;
	MKL_INT nnzmax = 0;	// nnzmax is zero in case of symbolic&numeric usage of mkl_?csrmultcsr
	MKL_INT ierr;
	MKL_INT request = 2;	//numeric multiplication
	MKL_INT sort = 3;	// sort everything
	time_mic_symbolic_mm = 0.0;
	#pragma offload target(mic:micdev) in(i) in(ierr) in(nnzmax) in(request) in(sort) in(transa) in(Am)  in(An) in(Bn) \
		out(time_mic_symbolic_mm) out(Cnnz_host)\
		in(Aval:length(Annz)	alloc_if(0)	free_if(0)) \
		in(AI:length(Am+1)	alloc_if(0)	free_if(0)) \
		in(AJ:length(Annz)	alloc_if(0)	free_if(0)) \
		in(Bval:length(Bnnz)	alloc_if(0)	free_if(0)) \
		in(BI:length(An+1)	alloc_if(0)	free_if(0)) \
		in(BJ:length(Bnnz)	alloc_if(0)	free_if(0)) \
		nocopy(CI:	alloc_if(0)	free_if(0)) \
		nocopy(CJ:	alloc_if(0)	free_if(0)) \
		nocopy(Cval:	alloc_if(0)	free_if(0))
	{
		CI = (MKL_INT*)mkl_malloc( (Am+2) * sizeof( MKL_INT ), 64 );
		MKL_INT nnzmax = 0;	// nnzmax is zero in case of symbolic&numeric usage of mkl_?csrmultcsr
		MKL_INT ierr;
		MKL_INT request = 1;	// symbolic multiplication
		//MKL_INT sort = 7;	// do not sort anything

		for(i = 0; i < 10; i++) {
			mkl_dcsrmultcsr(&transa, &request, &sort, &Am, &An, &Bn, Aval, AJ, AI, Bval, BJ, BI, Cval, CJ, CI, &nnzmax, &ierr);
		}	
		double s_initial = dsecnd();
		for(i = 0; i < num_repeat; i++) {
			mkl_dcsrmultcsr(&transa, &request, &sort, &Am, &An, &Bn, Aval, AJ, AI, Bval, BJ, BI, Cval, CJ, CI, &nnzmax, &ierr);
		}	
		time_mic_symbolic_mm = (dsecnd() - s_initial) / num_repeat; // seconds

		request = 2;		//numeric multiplication
		int Cnnz = CI[Am]-1;
		int Cval_size = Cnnz + 1; Cnnz_host = Cnnz;
		CJ		= (MKL_INT*)mkl_malloc( Cval_size	*	sizeof( MKL_INT ), 64 );
		Cval		= (double*)mkl_malloc( Cval_size	*	sizeof( double ), 64 );
		mkl_dcsrmultcsr(&transa, &request, &sort, &Am, &An, &An, Aval, AJ, AI, Bval, BJ, BI, Cval, CJ, CI, &nnzmax, &ierr);
		printmm_one(Am, Cval, CJ, CI);//printf("Cnnz=%d\n", Cnnz);
		sort = 7;		// do not sort anything
		request = 2;		// numeric multiplication
	}//printf("Cnnz_mic: %d\n", Cnnz_host);exit(-1);
	time_mic_numeric_mm = 0.0;
	#pragma offload target(mic:micdev) nocopy(request) nocopy(sort) nocopy(i) nocopy(ierr) nocopy(nnzmax)  nocopy(transa) nocopy(Am) nocopy(An) nocopy(Bn) \
		out(time_mic_numeric_mm) \
		nocopy(Aval:length(Annz)	alloc_if(0)	free_if(0)) \
		nocopy(AI:length(Am+1)	alloc_if(0)	free_if(0)) \
		nocopy(AJ:length(Annz)	alloc_if(0)	free_if(0)) \
		nocopy(Bval:length(Bnnz)	alloc_if(0)	free_if(0)) \
		nocopy(BI:length(An+1)	alloc_if(0)	free_if(0)) \
		nocopy(BJ:length(Bnnz)	alloc_if(0)	free_if(0)) \
		nocopy(CI:	alloc_if(0)	free_if(0)) \
		nocopy(CJ:	alloc_if(0)	free_if(0)) \
		nocopy(Cval:	alloc_if(0)	free_if(0))
	{
		//sort = 7
		//request = 1; 2'ye gore sure iki katina cikiyor
		//request = 0; 2'ye gore sure uc katina cikiyor

		//request = 2
		//sort = 3; 7'ye gore %30 daha yavas
		//printf("sort:%d request:%d\n", sort, request); // prints sort:7 request:2
		for(i = 0; i < 10; i++) {
			mkl_dcsrmultcsr(&transa, &request, &sort, &Am, &An, &An, Aval, AJ, AI, Bval, BJ, BI, Cval, CJ, CI, &nnzmax, &ierr);
		}
		double s_initial = dsecnd();
		for(i = 0; i < num_repeat; i++) {
			mkl_dcsrmultcsr(&transa, &request, &sort, &Am, &An, &An, Aval, AJ, AI, Bval, BJ, BI, Cval, CJ, CI, &nnzmax, &ierr);
		}
		time_mic_numeric_mm = (dsecnd() - s_initial) / num_repeat; // seconds
	}
	if (option_printfile_matrices == OPTION_PRINTFILE_MATRICES) {
	int nelm_CI = Am + 2; 
	int nelm_CJ = Cnnz_host + 1; 
	int nelm_Cval = Cnnz_host + 1; 
	__declspec(target(mic)) MKL_INT* CI_host = (MKL_INT*)mkl_malloc( (nelm_CI) * sizeof( MKL_INT ), 64 );
	__declspec(target(mic)) MKL_INT* CJ_host = (MKL_INT*)mkl_malloc( (nelm_CJ) * sizeof( MKL_INT ), 64 );
	__declspec(target(mic)) double* Cval_host = (double*)mkl_malloc( (nelm_Cval) * sizeof( double ), 64 );
	#pragma offload target(mic:micdev) in(nelm_CI)\ 
		inout(CI_host:length(nelm_CI)	alloc_if(1)	free_if(0)) \
		inout(CJ_host:length(nelm_CJ)	alloc_if(1)	free_if(0)) \
		inout(Cval_host:length(nelm_Cval)	alloc_if(1)	free_if(0)) \
		nocopy(CI:	alloc_if(0)	free_if(0)) \
		nocopy(CJ:	alloc_if(0)	free_if(0)) \
		nocopy(Cval:	alloc_if(0)	free_if(0))
	{
		int i;
		for(i = 0; i < nelm_CI; i++) CI_host[i] = CI[i];
		for(i = 0; i < nelm_CJ; i++) {CJ_host[i] = CJ[i]; Cval_host[i] = Cval[i];}
	}
	*pCval = Cval_host; *pCJ = CJ_host; *pCI = CI_host; //printf("Cnnz_host:%d\n", Cnnz_host);
	}
} /* ENDOF mkl_mic_spmm }}}*/
	void
read_mm(char* strpath, int* pM, int* pN, int* prealnnz, int** pI, int** pJ, double** pval){ /*{{{*/

int i, M, N, nz, *I, *J;
double* val;
int ret_code;
MM_typecode matcode;
FILE* f;
if ((f = fopen(strpath, "r")) == NULL) {fprintf(stderr, "Input matrix file %s cannot be opened to read.", strpath);exit(1);}
/* READ MATRIX */
if (mm_read_banner(f, &matcode) != 0) {
	printf("Could not process Matrix Market banner.\n");	exit(1);
}
/*  This is how one can screen matrix types if their application */
/*  only supports a subset of the Matrix Market data types.      */
if (mm_is_complex(matcode) && mm_is_matrix(matcode) && mm_is_sparse(matcode) ) {
	printf("Sorry, this application does not support ");
	printf("Market Market type: [%s]\n", mm_typecode_to_str(matcode));exit(1);
}
/* find out size of sparse matrix .... */
if ((ret_code = mm_read_mtx_crd_size(f, &M, &N, &nz)) !=0) exit(1);
/* reseve memory for matrices */
I = (int *) malloc((mm_is_symmetric(matcode) ? 2*nz : nz) * sizeof(int));
J = (int *) malloc((mm_is_symmetric(matcode) ? 2*nz : nz) * sizeof(int));
val = (double *) malloc((mm_is_symmetric(matcode) ? 2*nz : nz) * sizeof(double));
*pI = I;
*pJ = J;
*pval = val;
/* NOTE: when reading in doubles, ANSI C requires the use of the "l"  */
/*   specifier as in "%lg", "%lf", "%le", otherwise errors will occur */
/*  (ANSI C X3.159-1989, Sec. 4.9.6.2, p. 136 lines 13-15)            */
int realnnz = 0;
for (i=0; i<nz; i++) {
	if(mm_is_pattern(matcode)) {
		fscanf(f, "%d %d\n", &I[realnnz], &J[realnnz]);
		val[realnnz] = 1.0;
	}
	else
		fscanf(f, "%d %d %lg\n", &I[realnnz], &J[realnnz], &val[realnnz]);
	I[realnnz]--;  /* adjust from 1-based to 0-based */
        J[realnnz]--;
	if(mm_is_symmetric(matcode) && I[realnnz] != J[realnnz]) {
		I[realnnz+1] = J[realnnz];
		J[realnnz+1] = I[realnnz];
		val[realnnz+1] = val[realnnz];
		realnnz++;
	}
	realnnz++;
}
if (f !=stdin) fclose(f);
*pM = M;
*pN = N;
*prealnnz = realnnz;
} /* ENDOF read_mm }}}*/
	void
coo_to_csr(int m, int nnz, int* I, int* J, double* val, MKL_INT* AI, MKL_INT* AJ, double* Aval) { /*{{{*/

MKL_INT info = 0;
MKL_INT job[8];   
//job[1]=0; // zero based indexing in csr
job[1]=1; // one based indexing in csr

job[2]=0; // zero based indexing in coo
job[3]=2; // I don't know
job[4]=nnz; // nnz

job[0]=1;  // coo to csr
job[5]=0;  // Acsr and AJR allocated by user
//void mkl_dcsrcoo (MKL_INT * job, MKL_INT * n, double *Acsr, MKL_INT * AJR, MKL_INT *AIR, MKL_INT * nnz, double *Acoo, MKL_INT * ir, MKL_INT * jc, MKL_INT * info);
mkl_dcsrcoo (job,&m, Aval, AJ, AI, &nnz, val, I, J, &info);
} /* ENDOF coo_to_csr }}}*/
	int* 
read_perm(int n, char* fnrcperm){/*{{{*/

int i;
if(!strcasecmp(fnrcperm, "null")) {
} else {
	int* rperm = (int*)calloc(n, sizeof(int));
	FILE* fpPerm = fopen(fnrcperm, "r");
	if(fpPerm == NULL) {
		fprintf(stderr, "%s cannot be opened\n", fnrcperm);
       		exit(-3);
	}
	for(i = 0; i < n; i++){
		fscanf(fpPerm, "%d\n", &rperm[i]);
	}
	fclose(fpPerm);
	return rperm;
}
return NULL;
} /* ENDOF read_perm }}}*/ 
	void
permute_array(int n, int* array, int* perm){/*{{{1 */

if(perm == NULL)
	return;	
int i;
for (i=0; i<n; i++){
	array[i] = perm[array[i]];
}
} /* ENDOF permute_array */ 
/*1}}}*/
	int 
main(int argc, char* argv[]) { /*{{{*/

FILE *f;
int i;
int nrequired_args = 16;
if (argc != nrequired_args){
	fprintf(stderr, "Usage: %s [dataset name] [method] [cache size] [martix-market-filename A] [matrix-market-filename B] [martix-market-filename C] [nworker] [num repeat]  [-1: CPU|0,1,2,3: MIC] [compile flag] [null|Arow perm] [null|Acol/Brow perm] [null|Bcol perm] [PRINTMATRICES|NOPRINTMATRICES] [PRINTFILEMATRICES|NOPRINTFILEMATRICES]\n", argv[0]);
	exit(1);
}
int iarg = 1;
char* dataset_name = argv[iarg];	iarg++;
char* method = argv[iarg];		iarg++;
int cache_size = atoi(argv[iarg]);		iarg++;
char* strpathA = argv[iarg];	iarg++;
char* strpathB = argv[iarg];	iarg++;
char* strpathC = argv[iarg];	iarg++;
nworker = atoi(argv[iarg]);	iarg++;
num_repeat = atoi(argv[iarg]);	iarg++;
int option_host = atoi(argv[iarg]);	micdev = option_host;	iarg++;
char* compile_flag = argv[iarg];	iarg++;
char* fnArperm = argv[iarg];	iarg++;
char* fnAcBrperm = argv[iarg];	iarg++;
char* fnBcperm = argv[iarg];	iarg++;
option_print_matrices = strcmp(argv[iarg], "PRINTMATRICES")==0?OPTION_PRINT_MATRICES:OPTION_NOPRINT_MATRICES;	iarg++;
option_printfile_matrices = strcmp(argv[iarg], "PRINTFILEMATRICES")==0?OPTION_PRINTFILE_MATRICES:OPTION_NOPRINTFILE_MATRICES;	iarg++;
assert(nrequired_args == iarg);
MKL_INT Am, An, Annz;
int *Ax, *Ay;  double *Anz;
read_mm(strpathA, &Am, &An, &Annz, &Ax, &Ay, &Anz);

permute_array(Annz, Ax, read_perm(Am, fnArperm));	

int* permAcBr = read_perm(An, fnAcBrperm);
permute_array(Annz, Ay, permAcBr);	

double* Aval = (double*) mkl_malloc( Annz * sizeof( double ),  64 );
MKL_INT* AJ = (MKL_INT*) mkl_malloc( Annz * sizeof( MKL_INT ), 64 );
MKL_INT* AI = (MKL_INT*) mkl_malloc( (Am+1) * sizeof( MKL_INT ), 64 );
coo_to_csr(Am, Annz, Ax, Ay, Anz, AI, AJ, Aval);	


MKL_INT Bm, Bn, Bnnz;   
int *Bx, *By;  double *Bnz;
read_mm(strpathB, &Bm, &Bn, &Bnnz, &Bx, &By, &Bnz);

permute_array(Bnnz, Bx, permAcBr);	

permute_array(Bnnz, By, read_perm(Bn, fnBcperm));	

double* Bval = (double*) mkl_malloc( Bnnz * sizeof( double ),  64 );
MKL_INT* BJ = (MKL_INT*) mkl_malloc( Bnnz * sizeof( MKL_INT ), 64 );
MKL_INT* BI = (MKL_INT*) mkl_malloc( (Bm+1) * sizeof( MKL_INT ), 64 );
coo_to_csr(Bm, Bnnz, Bx, By, Bnz, BI, BJ, Bval);	

//printmm_one(Am, Aval, AJ, AI);
//printmm_one(Bm, Bval, BJ, BI);

double* Cval; MKL_INT* CJ; MKL_INT* CI;
if ( micdev == OPTION_HOST_CPU) {	
	mkl_cpu_spmm(Am, An, Annz, Aval, AJ, AI, Bn, Bnnz, Bval, BJ, BI, &Cval, &CJ, &CI);
} else {	
	mkl_mic_spmm(Am, An, Annz, Aval, AJ, AI, Bn, Bnnz, Bval, BJ, BI, &Cval, &CJ, &CI);
}
for(i=0;i<=Am;i++)AI[i]--;	for(i=0;i<Annz;i++)AJ[i]--;
for(i=0;i<=Bm;i++)BI[i]--;	for(i=0;i<Bnnz;i++)BJ[i]--;
printmm_one(Am, Cval, CJ, CI);
printfilemm_one(strpathC, Am, Bn, Cval, CJ, CI);
long nummult = 0;
csi* xb = (csi*)calloc(Bn, sizeof(csi));
csv* x = (csv*)calloc(Bn, sizeof(csv));
csr* C = csr_multiply(Am, An, Annz, AI, AJ, Aval, Bm, Bn, Bnnz, BI, BJ, Bval, &nummult, xb, x);
//        sprintf(benchName, "%s\t%s\t%d\t%s\t%s\t%d\t%d\t%s\t%s\t%s\t%g\t%g\t%g\t%g\t%g\t%g\t%ld\t%s\t%s\n", dataset_name, method, cache_size,target_str[target], dpTest ? "DP":"SP",numInMats, n_mic_omp_threads, str_datastructure, str_algorithm, compile_flag, gflop/avgTime, gflop/(avgTime+iTransferTime), gflop/(avgTime+iTransferTime+oTransferTime), avgTime, (avgTime+iTransferTime), (avgTime+iTransferTime+oTransferTime),  nummult, inFileName.c_str(), mm_C_filename.empty()?"N/A":mm_C_filename.c_str());
//results{{{
double time_cpu_numeric_mm = 0.0, time_cpu_symbolic_mm = 0.0;//TODO add time for mkl_cpu_spmm
double time_numeric_mm = 0.0, time_symbolic_mm = 0.0;
if ( micdev == OPTION_HOST_CPU) {	
	time_symbolic_mm = time_cpu_symbolic_mm;
	time_numeric_mm = time_cpu_numeric_mm;
} else {
	time_symbolic_mm = time_mic_symbolic_mm;
	time_numeric_mm = time_mic_numeric_mm;
}	
        double gflop = 2 * (double) nummult / 1e9;
	printf("MKL\t");
	printf("%s\t", dataset_name);
	printf("%s\t", method);
	printf("N/A\t");
	printf("%d\t", cache_size);
	printf("%s\t", micdev==OPTION_HOST_CPU?"CPU":"MIC");
	printf("DP\t");
	//printf("%s ", sizeof(csv)==sizeof(double)? "DP": sizeof(csv)==sizeof(long double)? "LDP":"SP");
	printf("1\t"); // numInMats
	printf("%d\t", nworker);
	printf("CSR\t"); // str_datastructure
	printf("INNER\t"); // str_algorithm
	printf("%s\t", compile_flag);
	printf("%g\t", (gflop/time_numeric_mm));
	printf("N/A\t");
	printf("N/A\t");
	printf("%g\t", time_numeric_mm);
	printf("N/A\t");
	printf("N/A\t");
	printf("%ld\t",nummult);
	printf("%s\t", strpathA);
	printf("%s\t", strpathC);
	printf("N/A\t");
	printf("N/A\t");
	printf("N/A\t");
	printf("N/A\t");
	printf("MKL\t"); ///PARTMODEL
	printf("N/A\t");
	printf("N/A\t");
	printf("N/A\t");
	printf("N/A\t");
	printf("N/A\t");
	printf("%d\t", max_threads);
	printf("%g\t", time_symbolic_mm);
	printf("%d\t", num_repeat);
	printf("%s\t", fnArperm);
	printf("%s\t", fnAcBrperm);
	printf("%s\t", fnBcperm);
	printf("\n");
//}}}
//	mkl_free(A);

	return 0;
} /* ENDOF main }}}*/
	__declspec(target(mic)) void 
printmm_one(int m, double* Aval, int* AJ, int* AI){ //{{{

	if (option_print_matrices == OPTION_NOPRINT_MATRICES)
		return;
	int i;
	for(i = 0; i < m; i++) {
		printf("%d: ", i+1);
		int j;
		for(j = AI[i]-1; j < AI[i+1]-1; j++) {
			printf("%d:%g  ", AJ[j], Aval[j]);
		}
		printf("\n");
	}
	printf("\n");
}//}}}
	__declspec(target(mic)) void 
printfilemm_one(char* file, int m, int n, double* Aval, int* AJ, int* AI){//{{{

	if (option_printfile_matrices == OPTION_NOPRINTFILE_MATRICES)
		return;
	FILE* f = fopen(file, "w");
	if(f == NULL){
		printf("%s %s %d: %s cannot be opened to write matrix\n", __FILE__, __PRETTY_FUNCTION__, __LINE__, file);
		exit(1); 
	}
	int i;
	fprintf(f, "%%%%MatrixMarket matrix coordinate real general\n");
	fprintf(f, "%d %d %d\n", m, n, AI[m]-1);
	for(i = 0; i < m; i++) {
		//printf("%d: ", i+1);
		int j;
		for(j = AI[i]-1; j < AI[i+1]-1; j++) {
			fprintf(f, "%d %d %g\n", i+1, AJ[j], Aval[j]);
		}
	//	printf("\n");
	}
	//printf("\n");
	fclose(f);
}//}}}
	void 
printmm_zero(int m, double* Aval, int* AJ, int* AI){//{{{

	int i;
	for(i = 0; i < m; i++) {
		printf("%d: ", i+1);
		int j;
		for(j = AI[i]; j < AI[i+1]; j++) {
			printf("%d:%g  ", AJ[j]+1, Aval[j]);
		}
		printf("\n");
	}
}//}}}
//commenteds{{{
/*	// for dense result
	double* Cdense		= (double*)mkl_malloc( (m+1)*(m+1)	*	sizeof( double ), 64 );
	assert(m==n);
	mkl_dcsrmultd(&transa,  &m, &n, &m, Aval, AJ, AI, Aval, AJ, AI, Cdense, &m);
	for(i = 0; i < m;i++){
		int j;
		for(j = 0; j < n; j++){
			printf("%g\t", Cdense[(j)*m+(i)]);
		}
		printf("\n");
	}
	printf("\n");
*/
//	double maxdiff = 0.00001;
/*	for(i = 0; i < m; i++){
		double diff = fabs(vecout[i] - vecoutcpu[i]);
		if(diff>maxdiff){
			printf("%g\t%g\t%g\n", diff, vecout[i], vecoutcpu[i]);
		}
	}
	printf("\n");
*///}}}