Actual source code: baij.c
petsc-3.14.6 2021-03-30
2: /*
3: Defines the basic matrix operations for the BAIJ (compressed row)
4: matrix storage format.
5: */
6: #include <../src/mat/impls/baij/seq/baij.h>
7: #include <petscblaslapack.h>
8: #include <petsc/private/kernels/blockinvert.h>
9: #include <petsc/private/kernels/blockmatmult.h>
11: #if defined(PETSC_HAVE_HYPRE)
12: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*);
13: #endif
15: #if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
16: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqBAIJMKL(Mat,MatType,MatReuse,Mat*);
17: #endif
18: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
20: PetscErrorCode MatInvertBlockDiagonal_SeqBAIJ(Mat A,const PetscScalar **values)
21: {
22: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*) A->data;
24: PetscInt *diag_offset,i,bs = A->rmap->bs,mbs = a->mbs,ipvt[5],bs2 = bs*bs,*v_pivots;
25: MatScalar *v = a->a,*odiag,*diag,work[25],*v_work;
26: PetscReal shift = 0.0;
27: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
30: allowzeropivot = PetscNot(A->erroriffailure);
32: if (a->idiagvalid) {
33: if (values) *values = a->idiag;
34: return(0);
35: }
36: MatMarkDiagonal_SeqBAIJ(A);
37: diag_offset = a->diag;
38: if (!a->idiag) {
39: PetscMalloc1(bs2*mbs,&a->idiag);
40: PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));
41: }
42: diag = a->idiag;
43: if (values) *values = a->idiag;
44: /* factor and invert each block */
45: switch (bs) {
46: case 1:
47: for (i=0; i<mbs; i++) {
48: odiag = v + 1*diag_offset[i];
49: diag[0] = odiag[0];
51: if (PetscAbsScalar(diag[0] + shift) < PETSC_MACHINE_EPSILON) {
52: if (allowzeropivot) {
53: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
54: A->factorerror_zeropivot_value = PetscAbsScalar(diag[0]);
55: A->factorerror_zeropivot_row = i;
56: PetscInfo1(A,"Zero pivot, row %D\n",i);
57: } else SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D pivot value %g tolerance %g",i,(double)PetscAbsScalar(diag[0]),(double)PETSC_MACHINE_EPSILON);
58: }
60: diag[0] = (PetscScalar)1.0 / (diag[0] + shift);
61: diag += 1;
62: }
63: break;
64: case 2:
65: for (i=0; i<mbs; i++) {
66: odiag = v + 4*diag_offset[i];
67: diag[0] = odiag[0]; diag[1] = odiag[1]; diag[2] = odiag[2]; diag[3] = odiag[3];
68: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
69: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
70: diag += 4;
71: }
72: break;
73: case 3:
74: for (i=0; i<mbs; i++) {
75: odiag = v + 9*diag_offset[i];
76: diag[0] = odiag[0]; diag[1] = odiag[1]; diag[2] = odiag[2]; diag[3] = odiag[3];
77: diag[4] = odiag[4]; diag[5] = odiag[5]; diag[6] = odiag[6]; diag[7] = odiag[7];
78: diag[8] = odiag[8];
79: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
80: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
81: diag += 9;
82: }
83: break;
84: case 4:
85: for (i=0; i<mbs; i++) {
86: odiag = v + 16*diag_offset[i];
87: PetscArraycpy(diag,odiag,16);
88: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
89: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
90: diag += 16;
91: }
92: break;
93: case 5:
94: for (i=0; i<mbs; i++) {
95: odiag = v + 25*diag_offset[i];
96: PetscArraycpy(diag,odiag,25);
97: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
98: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
99: diag += 25;
100: }
101: break;
102: case 6:
103: for (i=0; i<mbs; i++) {
104: odiag = v + 36*diag_offset[i];
105: PetscArraycpy(diag,odiag,36);
106: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
107: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
108: diag += 36;
109: }
110: break;
111: case 7:
112: for (i=0; i<mbs; i++) {
113: odiag = v + 49*diag_offset[i];
114: PetscArraycpy(diag,odiag,49);
115: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
116: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
117: diag += 49;
118: }
119: break;
120: default:
121: PetscMalloc2(bs,&v_work,bs,&v_pivots);
122: for (i=0; i<mbs; i++) {
123: odiag = v + bs2*diag_offset[i];
124: PetscArraycpy(diag,odiag,bs2);
125: PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
126: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
127: diag += bs2;
128: }
129: PetscFree2(v_work,v_pivots);
130: }
131: a->idiagvalid = PETSC_TRUE;
132: return(0);
133: }
135: PetscErrorCode MatSOR_SeqBAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
136: {
137: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
138: PetscScalar *x,*work,*w,*workt,*t;
139: const MatScalar *v,*aa = a->a, *idiag;
140: const PetscScalar *b,*xb;
141: PetscScalar s[7], xw[7]={0}; /* avoid some compilers thinking xw is uninitialized */
142: PetscErrorCode ierr;
143: PetscInt m = a->mbs,i,i2,nz,bs = A->rmap->bs,bs2 = bs*bs,k,j,idx,it;
144: const PetscInt *diag,*ai = a->i,*aj = a->j,*vi;
147: its = its*lits;
148: if (flag & SOR_EISENSTAT) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for Eisenstat");
149: if (its <= 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);
150: if (fshift) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Sorry, no support for diagonal shift");
151: if (omega != 1.0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Sorry, no support for non-trivial relaxation factor");
152: if ((flag & SOR_APPLY_UPPER) || (flag & SOR_APPLY_LOWER)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Sorry, no support for applying upper or lower triangular parts");
154: if (!a->idiagvalid) {MatInvertBlockDiagonal(A,NULL);}
156: if (!m) return(0);
157: diag = a->diag;
158: idiag = a->idiag;
159: k = PetscMax(A->rmap->n,A->cmap->n);
160: if (!a->mult_work) {
161: PetscMalloc1(k+1,&a->mult_work);
162: }
163: if (!a->sor_workt) {
164: PetscMalloc1(k,&a->sor_workt);
165: }
166: if (!a->sor_work) {
167: PetscMalloc1(bs,&a->sor_work);
168: }
169: work = a->mult_work;
170: t = a->sor_workt;
171: w = a->sor_work;
173: VecGetArray(xx,&x);
174: VecGetArrayRead(bb,&b);
176: if (flag & SOR_ZERO_INITIAL_GUESS) {
177: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
178: switch (bs) {
179: case 1:
180: PetscKernel_v_gets_A_times_w_1(x,idiag,b);
181: t[0] = b[0];
182: i2 = 1;
183: idiag += 1;
184: for (i=1; i<m; i++) {
185: v = aa + ai[i];
186: vi = aj + ai[i];
187: nz = diag[i] - ai[i];
188: s[0] = b[i2];
189: for (j=0; j<nz; j++) {
190: xw[0] = x[vi[j]];
191: PetscKernel_v_gets_v_minus_A_times_w_1(s,(v+j),xw);
192: }
193: t[i2] = s[0];
194: PetscKernel_v_gets_A_times_w_1(xw,idiag,s);
195: x[i2] = xw[0];
196: idiag += 1;
197: i2 += 1;
198: }
199: break;
200: case 2:
201: PetscKernel_v_gets_A_times_w_2(x,idiag,b);
202: t[0] = b[0]; t[1] = b[1];
203: i2 = 2;
204: idiag += 4;
205: for (i=1; i<m; i++) {
206: v = aa + 4*ai[i];
207: vi = aj + ai[i];
208: nz = diag[i] - ai[i];
209: s[0] = b[i2]; s[1] = b[i2+1];
210: for (j=0; j<nz; j++) {
211: idx = 2*vi[j];
212: it = 4*j;
213: xw[0] = x[idx]; xw[1] = x[1+idx];
214: PetscKernel_v_gets_v_minus_A_times_w_2(s,(v+it),xw);
215: }
216: t[i2] = s[0]; t[i2+1] = s[1];
217: PetscKernel_v_gets_A_times_w_2(xw,idiag,s);
218: x[i2] = xw[0]; x[i2+1] = xw[1];
219: idiag += 4;
220: i2 += 2;
221: }
222: break;
223: case 3:
224: PetscKernel_v_gets_A_times_w_3(x,idiag,b);
225: t[0] = b[0]; t[1] = b[1]; t[2] = b[2];
226: i2 = 3;
227: idiag += 9;
228: for (i=1; i<m; i++) {
229: v = aa + 9*ai[i];
230: vi = aj + ai[i];
231: nz = diag[i] - ai[i];
232: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2];
233: while (nz--) {
234: idx = 3*(*vi++);
235: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
236: PetscKernel_v_gets_v_minus_A_times_w_3(s,v,xw);
237: v += 9;
238: }
239: t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2];
240: PetscKernel_v_gets_A_times_w_3(xw,idiag,s);
241: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2];
242: idiag += 9;
243: i2 += 3;
244: }
245: break;
246: case 4:
247: PetscKernel_v_gets_A_times_w_4(x,idiag,b);
248: t[0] = b[0]; t[1] = b[1]; t[2] = b[2]; t[3] = b[3];
249: i2 = 4;
250: idiag += 16;
251: for (i=1; i<m; i++) {
252: v = aa + 16*ai[i];
253: vi = aj + ai[i];
254: nz = diag[i] - ai[i];
255: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3];
256: while (nz--) {
257: idx = 4*(*vi++);
258: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx];
259: PetscKernel_v_gets_v_minus_A_times_w_4(s,v,xw);
260: v += 16;
261: }
262: t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2]; t[i2 + 3] = s[3];
263: PetscKernel_v_gets_A_times_w_4(xw,idiag,s);
264: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3];
265: idiag += 16;
266: i2 += 4;
267: }
268: break;
269: case 5:
270: PetscKernel_v_gets_A_times_w_5(x,idiag,b);
271: t[0] = b[0]; t[1] = b[1]; t[2] = b[2]; t[3] = b[3]; t[4] = b[4];
272: i2 = 5;
273: idiag += 25;
274: for (i=1; i<m; i++) {
275: v = aa + 25*ai[i];
276: vi = aj + ai[i];
277: nz = diag[i] - ai[i];
278: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4];
279: while (nz--) {
280: idx = 5*(*vi++);
281: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; xw[4] = x[4+idx];
282: PetscKernel_v_gets_v_minus_A_times_w_5(s,v,xw);
283: v += 25;
284: }
285: t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2]; t[i2+3] = s[3]; t[i2+4] = s[4];
286: PetscKernel_v_gets_A_times_w_5(xw,idiag,s);
287: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4];
288: idiag += 25;
289: i2 += 5;
290: }
291: break;
292: case 6:
293: PetscKernel_v_gets_A_times_w_6(x,idiag,b);
294: t[0] = b[0]; t[1] = b[1]; t[2] = b[2]; t[3] = b[3]; t[4] = b[4]; t[5] = b[5];
295: i2 = 6;
296: idiag += 36;
297: for (i=1; i<m; i++) {
298: v = aa + 36*ai[i];
299: vi = aj + ai[i];
300: nz = diag[i] - ai[i];
301: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5];
302: while (nz--) {
303: idx = 6*(*vi++);
304: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
305: xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx];
306: PetscKernel_v_gets_v_minus_A_times_w_6(s,v,xw);
307: v += 36;
308: }
309: t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2];
310: t[i2+3] = s[3]; t[i2+4] = s[4]; t[i2+5] = s[5];
311: PetscKernel_v_gets_A_times_w_6(xw,idiag,s);
312: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5];
313: idiag += 36;
314: i2 += 6;
315: }
316: break;
317: case 7:
318: PetscKernel_v_gets_A_times_w_7(x,idiag,b);
319: t[0] = b[0]; t[1] = b[1]; t[2] = b[2];
320: t[3] = b[3]; t[4] = b[4]; t[5] = b[5]; t[6] = b[6];
321: i2 = 7;
322: idiag += 49;
323: for (i=1; i<m; i++) {
324: v = aa + 49*ai[i];
325: vi = aj + ai[i];
326: nz = diag[i] - ai[i];
327: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2];
328: s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; s[6] = b[i2+6];
329: while (nz--) {
330: idx = 7*(*vi++);
331: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
332: xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; xw[6] = x[6+idx];
333: PetscKernel_v_gets_v_minus_A_times_w_7(s,v,xw);
334: v += 49;
335: }
336: t[i2] = s[0]; t[i2+1] = s[1]; t[i2+2] = s[2];
337: t[i2+3] = s[3]; t[i2+4] = s[4]; t[i2+5] = s[5]; t[i2+6] = s[6];
338: PetscKernel_v_gets_A_times_w_7(xw,idiag,s);
339: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2];
340: x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; x[i2+6] = xw[6];
341: idiag += 49;
342: i2 += 7;
343: }
344: break;
345: default:
346: PetscKernel_w_gets_Ar_times_v(bs,bs,b,idiag,x);
347: PetscArraycpy(t,b,bs);
348: i2 = bs;
349: idiag += bs2;
350: for (i=1; i<m; i++) {
351: v = aa + bs2*ai[i];
352: vi = aj + ai[i];
353: nz = diag[i] - ai[i];
355: PetscArraycpy(w,b+i2,bs);
356: /* copy all rows of x that are needed into contiguous space */
357: workt = work;
358: for (j=0; j<nz; j++) {
359: PetscArraycpy(workt,x + bs*(*vi++),bs);
360: workt += bs;
361: }
362: PetscKernel_w_gets_w_minus_Ar_times_v(bs,bs*nz,w,v,work);
363: PetscArraycpy(t+i2,w,bs);
364: PetscKernel_w_gets_Ar_times_v(bs,bs,w,idiag,x+i2);
366: idiag += bs2;
367: i2 += bs;
368: }
369: break;
370: }
371: /* for logging purposes assume number of nonzero in lower half is 1/2 of total */
372: PetscLogFlops(1.0*bs2*a->nz);
373: xb = t;
374: }
375: else xb = b;
376: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
377: idiag = a->idiag+bs2*(a->mbs-1);
378: i2 = bs * (m-1);
379: switch (bs) {
380: case 1:
381: s[0] = xb[i2];
382: PetscKernel_v_gets_A_times_w_1(xw,idiag,s);
383: x[i2] = xw[0];
384: i2 -= 1;
385: for (i=m-2; i>=0; i--) {
386: v = aa + (diag[i]+1);
387: vi = aj + diag[i] + 1;
388: nz = ai[i+1] - diag[i] - 1;
389: s[0] = xb[i2];
390: for (j=0; j<nz; j++) {
391: xw[0] = x[vi[j]];
392: PetscKernel_v_gets_v_minus_A_times_w_1(s,(v+j),xw);
393: }
394: PetscKernel_v_gets_A_times_w_1(xw,idiag,s);
395: x[i2] = xw[0];
396: idiag -= 1;
397: i2 -= 1;
398: }
399: break;
400: case 2:
401: s[0] = xb[i2]; s[1] = xb[i2+1];
402: PetscKernel_v_gets_A_times_w_2(xw,idiag,s);
403: x[i2] = xw[0]; x[i2+1] = xw[1];
404: i2 -= 2;
405: idiag -= 4;
406: for (i=m-2; i>=0; i--) {
407: v = aa + 4*(diag[i] + 1);
408: vi = aj + diag[i] + 1;
409: nz = ai[i+1] - diag[i] - 1;
410: s[0] = xb[i2]; s[1] = xb[i2+1];
411: for (j=0; j<nz; j++) {
412: idx = 2*vi[j];
413: it = 4*j;
414: xw[0] = x[idx]; xw[1] = x[1+idx];
415: PetscKernel_v_gets_v_minus_A_times_w_2(s,(v+it),xw);
416: }
417: PetscKernel_v_gets_A_times_w_2(xw,idiag,s);
418: x[i2] = xw[0]; x[i2+1] = xw[1];
419: idiag -= 4;
420: i2 -= 2;
421: }
422: break;
423: case 3:
424: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2];
425: PetscKernel_v_gets_A_times_w_3(xw,idiag,s);
426: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2];
427: i2 -= 3;
428: idiag -= 9;
429: for (i=m-2; i>=0; i--) {
430: v = aa + 9*(diag[i]+1);
431: vi = aj + diag[i] + 1;
432: nz = ai[i+1] - diag[i] - 1;
433: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2];
434: while (nz--) {
435: idx = 3*(*vi++);
436: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
437: PetscKernel_v_gets_v_minus_A_times_w_3(s,v,xw);
438: v += 9;
439: }
440: PetscKernel_v_gets_A_times_w_3(xw,idiag,s);
441: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2];
442: idiag -= 9;
443: i2 -= 3;
444: }
445: break;
446: case 4:
447: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3];
448: PetscKernel_v_gets_A_times_w_4(xw,idiag,s);
449: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3];
450: i2 -= 4;
451: idiag -= 16;
452: for (i=m-2; i>=0; i--) {
453: v = aa + 16*(diag[i]+1);
454: vi = aj + diag[i] + 1;
455: nz = ai[i+1] - diag[i] - 1;
456: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3];
457: while (nz--) {
458: idx = 4*(*vi++);
459: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx];
460: PetscKernel_v_gets_v_minus_A_times_w_4(s,v,xw);
461: v += 16;
462: }
463: PetscKernel_v_gets_A_times_w_4(xw,idiag,s);
464: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3];
465: idiag -= 16;
466: i2 -= 4;
467: }
468: break;
469: case 5:
470: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; s[4] = xb[i2+4];
471: PetscKernel_v_gets_A_times_w_5(xw,idiag,s);
472: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4];
473: i2 -= 5;
474: idiag -= 25;
475: for (i=m-2; i>=0; i--) {
476: v = aa + 25*(diag[i]+1);
477: vi = aj + diag[i] + 1;
478: nz = ai[i+1] - diag[i] - 1;
479: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; s[4] = xb[i2+4];
480: while (nz--) {
481: idx = 5*(*vi++);
482: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; xw[4] = x[4+idx];
483: PetscKernel_v_gets_v_minus_A_times_w_5(s,v,xw);
484: v += 25;
485: }
486: PetscKernel_v_gets_A_times_w_5(xw,idiag,s);
487: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4];
488: idiag -= 25;
489: i2 -= 5;
490: }
491: break;
492: case 6:
493: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; s[4] = xb[i2+4]; s[5] = xb[i2+5];
494: PetscKernel_v_gets_A_times_w_6(xw,idiag,s);
495: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5];
496: i2 -= 6;
497: idiag -= 36;
498: for (i=m-2; i>=0; i--) {
499: v = aa + 36*(diag[i]+1);
500: vi = aj + diag[i] + 1;
501: nz = ai[i+1] - diag[i] - 1;
502: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2]; s[3] = xb[i2+3]; s[4] = xb[i2+4]; s[5] = xb[i2+5];
503: while (nz--) {
504: idx = 6*(*vi++);
505: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
506: xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx];
507: PetscKernel_v_gets_v_minus_A_times_w_6(s,v,xw);
508: v += 36;
509: }
510: PetscKernel_v_gets_A_times_w_6(xw,idiag,s);
511: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2]; x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5];
512: idiag -= 36;
513: i2 -= 6;
514: }
515: break;
516: case 7:
517: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2];
518: s[3] = xb[i2+3]; s[4] = xb[i2+4]; s[5] = xb[i2+5]; s[6] = xb[i2+6];
519: PetscKernel_v_gets_A_times_w_7(x,idiag,b);
520: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2];
521: x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; x[i2+6] = xw[6];
522: i2 -= 7;
523: idiag -= 49;
524: for (i=m-2; i>=0; i--) {
525: v = aa + 49*(diag[i]+1);
526: vi = aj + diag[i] + 1;
527: nz = ai[i+1] - diag[i] - 1;
528: s[0] = xb[i2]; s[1] = xb[i2+1]; s[2] = xb[i2+2];
529: s[3] = xb[i2+3]; s[4] = xb[i2+4]; s[5] = xb[i2+5]; s[6] = xb[i2+6];
530: while (nz--) {
531: idx = 7*(*vi++);
532: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
533: xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; xw[6] = x[6+idx];
534: PetscKernel_v_gets_v_minus_A_times_w_7(s,v,xw);
535: v += 49;
536: }
537: PetscKernel_v_gets_A_times_w_7(xw,idiag,s);
538: x[i2] = xw[0]; x[i2+1] = xw[1]; x[i2+2] = xw[2];
539: x[i2+3] = xw[3]; x[i2+4] = xw[4]; x[i2+5] = xw[5]; x[i2+6] = xw[6];
540: idiag -= 49;
541: i2 -= 7;
542: }
543: break;
544: default:
545: PetscArraycpy(w,xb+i2,bs);
546: PetscKernel_w_gets_Ar_times_v(bs,bs,w,idiag,x+i2);
547: i2 -= bs;
548: idiag -= bs2;
549: for (i=m-2; i>=0; i--) {
550: v = aa + bs2*(diag[i]+1);
551: vi = aj + diag[i] + 1;
552: nz = ai[i+1] - diag[i] - 1;
554: PetscArraycpy(w,xb+i2,bs);
555: /* copy all rows of x that are needed into contiguous space */
556: workt = work;
557: for (j=0; j<nz; j++) {
558: PetscArraycpy(workt,x + bs*(*vi++),bs);
559: workt += bs;
560: }
561: PetscKernel_w_gets_w_minus_Ar_times_v(bs,bs*nz,w,v,work);
562: PetscKernel_w_gets_Ar_times_v(bs,bs,w,idiag,x+i2);
564: idiag -= bs2;
565: i2 -= bs;
566: }
567: break;
568: }
569: PetscLogFlops(1.0*bs2*(a->nz));
570: }
571: its--;
572: }
573: while (its--) {
574: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
575: idiag = a->idiag;
576: i2 = 0;
577: switch (bs) {
578: case 1:
579: for (i=0; i<m; i++) {
580: v = aa + ai[i];
581: vi = aj + ai[i];
582: nz = ai[i+1] - ai[i];
583: s[0] = b[i2];
584: for (j=0; j<nz; j++) {
585: xw[0] = x[vi[j]];
586: PetscKernel_v_gets_v_minus_A_times_w_1(s,(v+j),xw);
587: }
588: PetscKernel_v_gets_A_times_w_1(xw,idiag,s);
589: x[i2] += xw[0];
590: idiag += 1;
591: i2 += 1;
592: }
593: break;
594: case 2:
595: for (i=0; i<m; i++) {
596: v = aa + 4*ai[i];
597: vi = aj + ai[i];
598: nz = ai[i+1] - ai[i];
599: s[0] = b[i2]; s[1] = b[i2+1];
600: for (j=0; j<nz; j++) {
601: idx = 2*vi[j];
602: it = 4*j;
603: xw[0] = x[idx]; xw[1] = x[1+idx];
604: PetscKernel_v_gets_v_minus_A_times_w_2(s,(v+it),xw);
605: }
606: PetscKernel_v_gets_A_times_w_2(xw,idiag,s);
607: x[i2] += xw[0]; x[i2+1] += xw[1];
608: idiag += 4;
609: i2 += 2;
610: }
611: break;
612: case 3:
613: for (i=0; i<m; i++) {
614: v = aa + 9*ai[i];
615: vi = aj + ai[i];
616: nz = ai[i+1] - ai[i];
617: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2];
618: while (nz--) {
619: idx = 3*(*vi++);
620: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
621: PetscKernel_v_gets_v_minus_A_times_w_3(s,v,xw);
622: v += 9;
623: }
624: PetscKernel_v_gets_A_times_w_3(xw,idiag,s);
625: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2];
626: idiag += 9;
627: i2 += 3;
628: }
629: break;
630: case 4:
631: for (i=0; i<m; i++) {
632: v = aa + 16*ai[i];
633: vi = aj + ai[i];
634: nz = ai[i+1] - ai[i];
635: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3];
636: while (nz--) {
637: idx = 4*(*vi++);
638: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx];
639: PetscKernel_v_gets_v_minus_A_times_w_4(s,v,xw);
640: v += 16;
641: }
642: PetscKernel_v_gets_A_times_w_4(xw,idiag,s);
643: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; x[i2+3] += xw[3];
644: idiag += 16;
645: i2 += 4;
646: }
647: break;
648: case 5:
649: for (i=0; i<m; i++) {
650: v = aa + 25*ai[i];
651: vi = aj + ai[i];
652: nz = ai[i+1] - ai[i];
653: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4];
654: while (nz--) {
655: idx = 5*(*vi++);
656: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; xw[4] = x[4+idx];
657: PetscKernel_v_gets_v_minus_A_times_w_5(s,v,xw);
658: v += 25;
659: }
660: PetscKernel_v_gets_A_times_w_5(xw,idiag,s);
661: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; x[i2+3] += xw[3]; x[i2+4] += xw[4];
662: idiag += 25;
663: i2 += 5;
664: }
665: break;
666: case 6:
667: for (i=0; i<m; i++) {
668: v = aa + 36*ai[i];
669: vi = aj + ai[i];
670: nz = ai[i+1] - ai[i];
671: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5];
672: while (nz--) {
673: idx = 6*(*vi++);
674: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
675: xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx];
676: PetscKernel_v_gets_v_minus_A_times_w_6(s,v,xw);
677: v += 36;
678: }
679: PetscKernel_v_gets_A_times_w_6(xw,idiag,s);
680: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2];
681: x[i2+3] += xw[3]; x[i2+4] += xw[4]; x[i2+5] += xw[5];
682: idiag += 36;
683: i2 += 6;
684: }
685: break;
686: case 7:
687: for (i=0; i<m; i++) {
688: v = aa + 49*ai[i];
689: vi = aj + ai[i];
690: nz = ai[i+1] - ai[i];
691: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2];
692: s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; s[6] = b[i2+6];
693: while (nz--) {
694: idx = 7*(*vi++);
695: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
696: xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; xw[6] = x[6+idx];
697: PetscKernel_v_gets_v_minus_A_times_w_7(s,v,xw);
698: v += 49;
699: }
700: PetscKernel_v_gets_A_times_w_7(xw,idiag,s);
701: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2];
702: x[i2+3] += xw[3]; x[i2+4] += xw[4]; x[i2+5] += xw[5]; x[i2+6] += xw[6];
703: idiag += 49;
704: i2 += 7;
705: }
706: break;
707: default:
708: for (i=0; i<m; i++) {
709: v = aa + bs2*ai[i];
710: vi = aj + ai[i];
711: nz = ai[i+1] - ai[i];
713: PetscArraycpy(w,b+i2,bs);
714: /* copy all rows of x that are needed into contiguous space */
715: workt = work;
716: for (j=0; j<nz; j++) {
717: PetscArraycpy(workt,x + bs*(*vi++),bs);
718: workt += bs;
719: }
720: PetscKernel_w_gets_w_minus_Ar_times_v(bs,bs*nz,w,v,work);
721: PetscKernel_w_gets_w_plus_Ar_times_v(bs,bs,w,idiag,x+i2);
723: idiag += bs2;
724: i2 += bs;
725: }
726: break;
727: }
728: PetscLogFlops(2.0*bs2*a->nz);
729: }
730: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
731: idiag = a->idiag+bs2*(a->mbs-1);
732: i2 = bs * (m-1);
733: switch (bs) {
734: case 1:
735: for (i=m-1; i>=0; i--) {
736: v = aa + ai[i];
737: vi = aj + ai[i];
738: nz = ai[i+1] - ai[i];
739: s[0] = b[i2];
740: for (j=0; j<nz; j++) {
741: xw[0] = x[vi[j]];
742: PetscKernel_v_gets_v_minus_A_times_w_1(s,(v+j),xw);
743: }
744: PetscKernel_v_gets_A_times_w_1(xw,idiag,s);
745: x[i2] += xw[0];
746: idiag -= 1;
747: i2 -= 1;
748: }
749: break;
750: case 2:
751: for (i=m-1; i>=0; i--) {
752: v = aa + 4*ai[i];
753: vi = aj + ai[i];
754: nz = ai[i+1] - ai[i];
755: s[0] = b[i2]; s[1] = b[i2+1];
756: for (j=0; j<nz; j++) {
757: idx = 2*vi[j];
758: it = 4*j;
759: xw[0] = x[idx]; xw[1] = x[1+idx];
760: PetscKernel_v_gets_v_minus_A_times_w_2(s,(v+it),xw);
761: }
762: PetscKernel_v_gets_A_times_w_2(xw,idiag,s);
763: x[i2] += xw[0]; x[i2+1] += xw[1];
764: idiag -= 4;
765: i2 -= 2;
766: }
767: break;
768: case 3:
769: for (i=m-1; i>=0; i--) {
770: v = aa + 9*ai[i];
771: vi = aj + ai[i];
772: nz = ai[i+1] - ai[i];
773: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2];
774: while (nz--) {
775: idx = 3*(*vi++);
776: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
777: PetscKernel_v_gets_v_minus_A_times_w_3(s,v,xw);
778: v += 9;
779: }
780: PetscKernel_v_gets_A_times_w_3(xw,idiag,s);
781: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2];
782: idiag -= 9;
783: i2 -= 3;
784: }
785: break;
786: case 4:
787: for (i=m-1; i>=0; i--) {
788: v = aa + 16*ai[i];
789: vi = aj + ai[i];
790: nz = ai[i+1] - ai[i];
791: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3];
792: while (nz--) {
793: idx = 4*(*vi++);
794: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx];
795: PetscKernel_v_gets_v_minus_A_times_w_4(s,v,xw);
796: v += 16;
797: }
798: PetscKernel_v_gets_A_times_w_4(xw,idiag,s);
799: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; x[i2+3] += xw[3];
800: idiag -= 16;
801: i2 -= 4;
802: }
803: break;
804: case 5:
805: for (i=m-1; i>=0; i--) {
806: v = aa + 25*ai[i];
807: vi = aj + ai[i];
808: nz = ai[i+1] - ai[i];
809: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4];
810: while (nz--) {
811: idx = 5*(*vi++);
812: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx]; xw[3] = x[3+idx]; xw[4] = x[4+idx];
813: PetscKernel_v_gets_v_minus_A_times_w_5(s,v,xw);
814: v += 25;
815: }
816: PetscKernel_v_gets_A_times_w_5(xw,idiag,s);
817: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2]; x[i2+3] += xw[3]; x[i2+4] += xw[4];
818: idiag -= 25;
819: i2 -= 5;
820: }
821: break;
822: case 6:
823: for (i=m-1; i>=0; i--) {
824: v = aa + 36*ai[i];
825: vi = aj + ai[i];
826: nz = ai[i+1] - ai[i];
827: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2]; s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5];
828: while (nz--) {
829: idx = 6*(*vi++);
830: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
831: xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx];
832: PetscKernel_v_gets_v_minus_A_times_w_6(s,v,xw);
833: v += 36;
834: }
835: PetscKernel_v_gets_A_times_w_6(xw,idiag,s);
836: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2];
837: x[i2+3] += xw[3]; x[i2+4] += xw[4]; x[i2+5] += xw[5];
838: idiag -= 36;
839: i2 -= 6;
840: }
841: break;
842: case 7:
843: for (i=m-1; i>=0; i--) {
844: v = aa + 49*ai[i];
845: vi = aj + ai[i];
846: nz = ai[i+1] - ai[i];
847: s[0] = b[i2]; s[1] = b[i2+1]; s[2] = b[i2+2];
848: s[3] = b[i2+3]; s[4] = b[i2+4]; s[5] = b[i2+5]; s[6] = b[i2+6];
849: while (nz--) {
850: idx = 7*(*vi++);
851: xw[0] = x[idx]; xw[1] = x[1+idx]; xw[2] = x[2+idx];
852: xw[3] = x[3+idx]; xw[4] = x[4+idx]; xw[5] = x[5+idx]; xw[6] = x[6+idx];
853: PetscKernel_v_gets_v_minus_A_times_w_7(s,v,xw);
854: v += 49;
855: }
856: PetscKernel_v_gets_A_times_w_7(xw,idiag,s);
857: x[i2] += xw[0]; x[i2+1] += xw[1]; x[i2+2] += xw[2];
858: x[i2+3] += xw[3]; x[i2+4] += xw[4]; x[i2+5] += xw[5]; x[i2+6] += xw[6];
859: idiag -= 49;
860: i2 -= 7;
861: }
862: break;
863: default:
864: for (i=m-1; i>=0; i--) {
865: v = aa + bs2*ai[i];
866: vi = aj + ai[i];
867: nz = ai[i+1] - ai[i];
869: PetscArraycpy(w,b+i2,bs);
870: /* copy all rows of x that are needed into contiguous space */
871: workt = work;
872: for (j=0; j<nz; j++) {
873: PetscArraycpy(workt,x + bs*(*vi++),bs);
874: workt += bs;
875: }
876: PetscKernel_w_gets_w_minus_Ar_times_v(bs,bs*nz,w,v,work);
877: PetscKernel_w_gets_w_plus_Ar_times_v(bs,bs,w,idiag,x+i2);
879: idiag -= bs2;
880: i2 -= bs;
881: }
882: break;
883: }
884: PetscLogFlops(2.0*bs2*(a->nz));
885: }
886: }
887: VecRestoreArray(xx,&x);
888: VecRestoreArrayRead(bb,&b);
889: return(0);
890: }
893: /*
894: Special version for direct calls from Fortran (Used in PETSc-fun3d)
895: */
896: #if defined(PETSC_HAVE_FORTRAN_CAPS)
897: #define matsetvaluesblocked4_ MATSETVALUESBLOCKED4
898: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
899: #define matsetvaluesblocked4_ matsetvaluesblocked4
900: #endif
902: PETSC_EXTERN void matsetvaluesblocked4_(Mat *AA,PetscInt *mm,const PetscInt im[],PetscInt *nn,const PetscInt in[],const PetscScalar v[])
903: {
904: Mat A = *AA;
905: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
906: PetscInt *rp,k,low,high,t,ii,jj,row,nrow,i,col,l,N,m = *mm,n = *nn;
907: PetscInt *ai =a->i,*ailen=a->ilen;
908: PetscInt *aj =a->j,stepval,lastcol = -1;
909: const PetscScalar *value = v;
910: MatScalar *ap,*aa = a->a,*bap;
911: PetscErrorCode ierr;
914: if (A->rmap->bs != 4) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Can only be called with a block size of 4");
915: stepval = (n-1)*4;
916: for (k=0; k<m; k++) { /* loop over added rows */
917: row = im[k];
918: rp = aj + ai[row];
919: ap = aa + 16*ai[row];
920: nrow = ailen[row];
921: low = 0;
922: high = nrow;
923: for (l=0; l<n; l++) { /* loop over added columns */
924: col = in[l];
925: if (col <= lastcol) low = 0;
926: else high = nrow;
927: lastcol = col;
928: value = v + k*(stepval+4 + l)*4;
929: while (high-low > 7) {
930: t = (low+high)/2;
931: if (rp[t] > col) high = t;
932: else low = t;
933: }
934: for (i=low; i<high; i++) {
935: if (rp[i] > col) break;
936: if (rp[i] == col) {
937: bap = ap + 16*i;
938: for (ii=0; ii<4; ii++,value+=stepval) {
939: for (jj=ii; jj<16; jj+=4) {
940: bap[jj] += *value++;
941: }
942: }
943: goto noinsert2;
944: }
945: }
946: N = nrow++ - 1;
947: high++; /* added new column index thus must search to one higher than before */
948: /* shift up all the later entries in this row */
949: for (ii=N; ii>=i; ii--) {
950: rp[ii+1] = rp[ii];
951: PetscArraycpy(ap+16*(ii+1),ap+16*(ii),16);CHKERRV(ierr);
952: }
953: if (N >= i) {
954: PetscArrayzero(ap+16*i,16);CHKERRV(ierr);
955: }
956: rp[i] = col;
957: bap = ap + 16*i;
958: for (ii=0; ii<4; ii++,value+=stepval) {
959: for (jj=ii; jj<16; jj+=4) {
960: bap[jj] = *value++;
961: }
962: }
963: noinsert2:;
964: low = i;
965: }
966: ailen[row] = nrow;
967: }
968: PetscFunctionReturnVoid();
969: }
971: #if defined(PETSC_HAVE_FORTRAN_CAPS)
972: #define matsetvalues4_ MATSETVALUES4
973: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
974: #define matsetvalues4_ matsetvalues4
975: #endif
977: PETSC_EXTERN void matsetvalues4_(Mat *AA,PetscInt *mm,PetscInt *im,PetscInt *nn,PetscInt *in,PetscScalar *v)
978: {
979: Mat A = *AA;
980: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
981: PetscInt *rp,k,low,high,t,row,nrow,i,col,l,N,n = *nn,m = *mm;
982: PetscInt *ai=a->i,*ailen=a->ilen;
983: PetscInt *aj=a->j,brow,bcol;
984: PetscInt ridx,cidx,lastcol = -1;
985: MatScalar *ap,value,*aa=a->a,*bap;
989: for (k=0; k<m; k++) { /* loop over added rows */
990: row = im[k]; brow = row/4;
991: rp = aj + ai[brow];
992: ap = aa + 16*ai[brow];
993: nrow = ailen[brow];
994: low = 0;
995: high = nrow;
996: for (l=0; l<n; l++) { /* loop over added columns */
997: col = in[l]; bcol = col/4;
998: ridx = row % 4; cidx = col % 4;
999: value = v[l + k*n];
1000: if (col <= lastcol) low = 0;
1001: else high = nrow;
1002: lastcol = col;
1003: while (high-low > 7) {
1004: t = (low+high)/2;
1005: if (rp[t] > bcol) high = t;
1006: else low = t;
1007: }
1008: for (i=low; i<high; i++) {
1009: if (rp[i] > bcol) break;
1010: if (rp[i] == bcol) {
1011: bap = ap + 16*i + 4*cidx + ridx;
1012: *bap += value;
1013: goto noinsert1;
1014: }
1015: }
1016: N = nrow++ - 1;
1017: high++; /* added new column thus must search to one higher than before */
1018: /* shift up all the later entries in this row */
1019: PetscArraymove(rp+i+1,rp+i,N-i+1);CHKERRV(ierr);
1020: PetscArraymove(ap+16*i+16,ap+16*i,16*(N-i+1));CHKERRV(ierr);
1021: PetscArrayzero(ap+16*i,16);CHKERRV(ierr);
1022: rp[i] = bcol;
1023: ap[16*i + 4*cidx + ridx] = value;
1024: noinsert1:;
1025: low = i;
1026: }
1027: ailen[brow] = nrow;
1028: }
1029: PetscFunctionReturnVoid();
1030: }
1032: /*
1033: Checks for missing diagonals
1034: */
1035: PetscErrorCode MatMissingDiagonal_SeqBAIJ(Mat A,PetscBool *missing,PetscInt *d)
1036: {
1037: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1039: PetscInt *diag,*ii = a->i,i;
1042: MatMarkDiagonal_SeqBAIJ(A);
1043: *missing = PETSC_FALSE;
1044: if (A->rmap->n > 0 && !ii) {
1045: *missing = PETSC_TRUE;
1046: if (d) *d = 0;
1047: PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");
1048: } else {
1049: PetscInt n;
1050: n = PetscMin(a->mbs, a->nbs);
1051: diag = a->diag;
1052: for (i=0; i<n; i++) {
1053: if (diag[i] >= ii[i+1]) {
1054: *missing = PETSC_TRUE;
1055: if (d) *d = i;
1056: PetscInfo1(A,"Matrix is missing block diagonal number %D\n",i);
1057: break;
1058: }
1059: }
1060: }
1061: return(0);
1062: }
1064: PetscErrorCode MatMarkDiagonal_SeqBAIJ(Mat A)
1065: {
1066: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1068: PetscInt i,j,m = a->mbs;
1071: if (!a->diag) {
1072: PetscMalloc1(m,&a->diag);
1073: PetscLogObjectMemory((PetscObject)A,m*sizeof(PetscInt));
1074: a->free_diag = PETSC_TRUE;
1075: }
1076: for (i=0; i<m; i++) {
1077: a->diag[i] = a->i[i+1];
1078: for (j=a->i[i]; j<a->i[i+1]; j++) {
1079: if (a->j[j] == i) {
1080: a->diag[i] = j;
1081: break;
1082: }
1083: }
1084: }
1085: return(0);
1086: }
1089: static PetscErrorCode MatGetRowIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool blockcompressed,PetscInt *nn,const PetscInt *inia[],const PetscInt *inja[],PetscBool *done)
1090: {
1091: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1093: PetscInt i,j,n = a->mbs,nz = a->i[n],*tia,*tja,bs = A->rmap->bs,k,l,cnt;
1094: PetscInt **ia = (PetscInt**)inia,**ja = (PetscInt**)inja;
1097: *nn = n;
1098: if (!ia) return(0);
1099: if (symmetric) {
1100: MatToSymmetricIJ_SeqAIJ(n,a->i,a->j,PETSC_TRUE,0,0,&tia,&tja);
1101: nz = tia[n];
1102: } else {
1103: tia = a->i; tja = a->j;
1104: }
1106: if (!blockcompressed && bs > 1) {
1107: (*nn) *= bs;
1108: /* malloc & create the natural set of indices */
1109: PetscMalloc1((n+1)*bs,ia);
1110: if (n) {
1111: (*ia)[0] = oshift;
1112: for (j=1; j<bs; j++) {
1113: (*ia)[j] = (tia[1]-tia[0])*bs+(*ia)[j-1];
1114: }
1115: }
1117: for (i=1; i<n; i++) {
1118: (*ia)[i*bs] = (tia[i]-tia[i-1])*bs + (*ia)[i*bs-1];
1119: for (j=1; j<bs; j++) {
1120: (*ia)[i*bs+j] = (tia[i+1]-tia[i])*bs + (*ia)[i*bs+j-1];
1121: }
1122: }
1123: if (n) {
1124: (*ia)[n*bs] = (tia[n]-tia[n-1])*bs + (*ia)[n*bs-1];
1125: }
1127: if (inja) {
1128: PetscMalloc1(nz*bs*bs,ja);
1129: cnt = 0;
1130: for (i=0; i<n; i++) {
1131: for (j=0; j<bs; j++) {
1132: for (k=tia[i]; k<tia[i+1]; k++) {
1133: for (l=0; l<bs; l++) {
1134: (*ja)[cnt++] = bs*tja[k] + l;
1135: }
1136: }
1137: }
1138: }
1139: }
1141: if (symmetric) { /* deallocate memory allocated in MatToSymmetricIJ_SeqAIJ() */
1142: PetscFree(tia);
1143: PetscFree(tja);
1144: }
1145: } else if (oshift == 1) {
1146: if (symmetric) {
1147: nz = tia[A->rmap->n/bs];
1148: /* add 1 to i and j indices */
1149: for (i=0; i<A->rmap->n/bs+1; i++) tia[i] = tia[i] + 1;
1150: *ia = tia;
1151: if (ja) {
1152: for (i=0; i<nz; i++) tja[i] = tja[i] + 1;
1153: *ja = tja;
1154: }
1155: } else {
1156: nz = a->i[A->rmap->n/bs];
1157: /* malloc space and add 1 to i and j indices */
1158: PetscMalloc1(A->rmap->n/bs+1,ia);
1159: for (i=0; i<A->rmap->n/bs+1; i++) (*ia)[i] = a->i[i] + 1;
1160: if (ja) {
1161: PetscMalloc1(nz,ja);
1162: for (i=0; i<nz; i++) (*ja)[i] = a->j[i] + 1;
1163: }
1164: }
1165: } else {
1166: *ia = tia;
1167: if (ja) *ja = tja;
1168: }
1169: return(0);
1170: }
1172: static PetscErrorCode MatRestoreRowIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool blockcompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
1173: {
1177: if (!ia) return(0);
1178: if ((!blockcompressed && A->rmap->bs > 1) || (symmetric || oshift == 1)) {
1179: PetscFree(*ia);
1180: if (ja) {PetscFree(*ja);}
1181: }
1182: return(0);
1183: }
1185: PetscErrorCode MatDestroy_SeqBAIJ(Mat A)
1186: {
1187: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1191: #if defined(PETSC_USE_LOG)
1192: PetscLogObjectState((PetscObject)A,"Rows=%D, Cols=%D, NZ=%D",A->rmap->N,A->cmap->n,a->nz);
1193: #endif
1194: MatSeqXAIJFreeAIJ(A,&a->a,&a->j,&a->i);
1195: ISDestroy(&a->row);
1196: ISDestroy(&a->col);
1197: if (a->free_diag) {PetscFree(a->diag);}
1198: PetscFree(a->idiag);
1199: if (a->free_imax_ilen) {PetscFree2(a->imax,a->ilen);}
1200: PetscFree(a->solve_work);
1201: PetscFree(a->mult_work);
1202: PetscFree(a->sor_workt);
1203: PetscFree(a->sor_work);
1204: ISDestroy(&a->icol);
1205: PetscFree(a->saved_values);
1206: PetscFree2(a->compressedrow.i,a->compressedrow.rindex);
1208: MatDestroy(&a->sbaijMat);
1209: MatDestroy(&a->parent);
1210: PetscFree(A->data);
1212: PetscObjectChangeTypeName((PetscObject)A,NULL);
1213: PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJGetArray_C",NULL);
1214: PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJRestoreArray_C",NULL);
1215: PetscObjectComposeFunction((PetscObject)A,"MatInvertBlockDiagonal_C",NULL);
1216: PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);
1217: PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);
1218: PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetColumnIndices_C",NULL);
1219: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqaij_C",NULL);
1220: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqsbaij_C",NULL);
1221: PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetPreallocation_C",NULL);
1222: PetscObjectComposeFunction((PetscObject)A,"MatSeqBAIJSetPreallocationCSR_C",NULL);
1223: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_seqbstrm_C",NULL);
1224: PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);
1225: #if defined(PETSC_HAVE_HYPRE)
1226: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_hypre_C",NULL);
1227: #endif
1228: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqbaij_is_C",NULL);
1229: return(0);
1230: }
1232: PetscErrorCode MatSetOption_SeqBAIJ(Mat A,MatOption op,PetscBool flg)
1233: {
1234: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1238: switch (op) {
1239: case MAT_ROW_ORIENTED:
1240: a->roworiented = flg;
1241: break;
1242: case MAT_KEEP_NONZERO_PATTERN:
1243: a->keepnonzeropattern = flg;
1244: break;
1245: case MAT_NEW_NONZERO_LOCATIONS:
1246: a->nonew = (flg ? 0 : 1);
1247: break;
1248: case MAT_NEW_NONZERO_LOCATION_ERR:
1249: a->nonew = (flg ? -1 : 0);
1250: break;
1251: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1252: a->nonew = (flg ? -2 : 0);
1253: break;
1254: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1255: a->nounused = (flg ? -1 : 0);
1256: break;
1257: case MAT_NEW_DIAGONALS:
1258: case MAT_IGNORE_OFF_PROC_ENTRIES:
1259: case MAT_USE_HASH_TABLE:
1260: case MAT_SORTED_FULL:
1261: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1262: break;
1263: case MAT_SPD:
1264: case MAT_SYMMETRIC:
1265: case MAT_STRUCTURALLY_SYMMETRIC:
1266: case MAT_HERMITIAN:
1267: case MAT_SYMMETRY_ETERNAL:
1268: case MAT_SUBMAT_SINGLEIS:
1269: case MAT_STRUCTURE_ONLY:
1270: /* These options are handled directly by MatSetOption() */
1271: break;
1272: default:
1273: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1274: }
1275: return(0);
1276: }
1278: /* used for both SeqBAIJ and SeqSBAIJ matrices */
1279: PetscErrorCode MatGetRow_SeqBAIJ_private(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v,PetscInt *ai,PetscInt *aj,PetscScalar *aa)
1280: {
1282: PetscInt itmp,i,j,k,M,bn,bp,*idx_i,bs,bs2;
1283: MatScalar *aa_i;
1284: PetscScalar *v_i;
1287: bs = A->rmap->bs;
1288: bs2 = bs*bs;
1289: if (row < 0 || row >= A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range", row);
1291: bn = row/bs; /* Block number */
1292: bp = row % bs; /* Block Position */
1293: M = ai[bn+1] - ai[bn];
1294: *nz = bs*M;
1296: if (v) {
1297: *v = NULL;
1298: if (*nz) {
1299: PetscMalloc1(*nz,v);
1300: for (i=0; i<M; i++) { /* for each block in the block row */
1301: v_i = *v + i*bs;
1302: aa_i = aa + bs2*(ai[bn] + i);
1303: for (j=bp,k=0; j<bs2; j+=bs,k++) v_i[k] = aa_i[j];
1304: }
1305: }
1306: }
1308: if (idx) {
1309: *idx = NULL;
1310: if (*nz) {
1311: PetscMalloc1(*nz,idx);
1312: for (i=0; i<M; i++) { /* for each block in the block row */
1313: idx_i = *idx + i*bs;
1314: itmp = bs*aj[ai[bn] + i];
1315: for (j=0; j<bs; j++) idx_i[j] = itmp++;
1316: }
1317: }
1318: }
1319: return(0);
1320: }
1322: PetscErrorCode MatGetRow_SeqBAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1323: {
1324: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1328: MatGetRow_SeqBAIJ_private(A,row,nz,idx,v,a->i,a->j,a->a);
1329: return(0);
1330: }
1332: PetscErrorCode MatRestoreRow_SeqBAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1333: {
1337: if (idx) {PetscFree(*idx);}
1338: if (v) {PetscFree(*v);}
1339: return(0);
1340: }
1342: PetscErrorCode MatTranspose_SeqBAIJ(Mat A,MatReuse reuse,Mat *B)
1343: {
1344: Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data,*at;
1345: Mat C;
1347: PetscInt i,j,k,*aj=a->j,*ai=a->i,bs=A->rmap->bs,mbs=a->mbs,nbs=a->nbs,*atfill;
1348: PetscInt bs2=a->bs2,*ati,*atj,anzj,kr;
1349: MatScalar *ata,*aa=a->a;
1352: PetscCalloc1(1+nbs,&atfill);
1353: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
1354: for (i=0; i<ai[mbs]; i++) atfill[aj[i]] += 1; /* count num of non-zeros in row aj[i] */
1356: MatCreate(PetscObjectComm((PetscObject)A),&C);
1357: MatSetSizes(C,A->cmap->n,A->rmap->N,A->cmap->n,A->rmap->N);
1358: MatSetType(C,((PetscObject)A)->type_name);
1359: MatSeqBAIJSetPreallocation(C,bs,0,atfill);
1361: at = (Mat_SeqBAIJ*)C->data;
1362: ati = at->i;
1363: for (i=0; i<nbs; i++) at->ilen[i] = at->imax[i] = ati[i+1] - ati[i];
1364: } else {
1365: C = *B;
1366: at = (Mat_SeqBAIJ*)C->data;
1367: ati = at->i;
1368: }
1370: atj = at->j;
1371: ata = at->a;
1373: /* Copy ati into atfill so we have locations of the next free space in atj */
1374: PetscArraycpy(atfill,ati,nbs);
1376: /* Walk through A row-wise and mark nonzero entries of A^T. */
1377: for (i=0; i<mbs; i++) {
1378: anzj = ai[i+1] - ai[i];
1379: for (j=0; j<anzj; j++) {
1380: atj[atfill[*aj]] = i;
1381: for (kr=0; kr<bs; kr++) {
1382: for (k=0; k<bs; k++) {
1383: ata[bs2*atfill[*aj]+k*bs+kr] = *aa++;
1384: }
1385: }
1386: atfill[*aj++] += 1;
1387: }
1388: }
1389: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
1390: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
1392: /* Clean up temporary space and complete requests. */
1393: PetscFree(atfill);
1395: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1396: MatSetBlockSizes(C,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
1397: *B = C;
1398: } else {
1399: MatHeaderMerge(A,&C);
1400: }
1401: return(0);
1402: }
1404: PetscErrorCode MatIsTranspose_SeqBAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
1405: {
1407: Mat Btrans;
1410: *f = PETSC_FALSE;
1411: MatTranspose_SeqBAIJ(A,MAT_INITIAL_MATRIX,&Btrans);
1412: MatEqual_SeqBAIJ(B,Btrans,f);
1413: MatDestroy(&Btrans);
1414: return(0);
1415: }
1417: /* Used for both SeqBAIJ and SeqSBAIJ matrices */
1418: PetscErrorCode MatView_SeqBAIJ_Binary(Mat mat,PetscViewer viewer)
1419: {
1420: Mat_SeqBAIJ *A = (Mat_SeqBAIJ*)mat->data;
1421: PetscInt header[4],M,N,m,bs,nz,cnt,i,j,k,l;
1422: PetscInt *rowlens,*colidxs;
1423: PetscScalar *matvals;
1427: PetscViewerSetUp(viewer);
1429: M = mat->rmap->N;
1430: N = mat->cmap->N;
1431: m = mat->rmap->n;
1432: bs = mat->rmap->bs;
1433: nz = bs*bs*A->nz;
1435: /* write matrix header */
1436: header[0] = MAT_FILE_CLASSID;
1437: header[1] = M; header[2] = N; header[3] = nz;
1438: PetscViewerBinaryWrite(viewer,header,4,PETSC_INT);
1440: /* store row lengths */
1441: PetscMalloc1(m,&rowlens);
1442: for (cnt=0, i=0; i<A->mbs; i++)
1443: for (j=0; j<bs; j++)
1444: rowlens[cnt++] = bs*(A->i[i+1] - A->i[i]);
1445: PetscViewerBinaryWrite(viewer,rowlens,m,PETSC_INT);
1446: PetscFree(rowlens);
1448: /* store column indices */
1449: PetscMalloc1(nz,&colidxs);
1450: for (cnt=0, i=0; i<A->mbs; i++)
1451: for (k=0; k<bs; k++)
1452: for (j=A->i[i]; j<A->i[i+1]; j++)
1453: for (l=0; l<bs; l++)
1454: colidxs[cnt++] = bs*A->j[j] + l;
1455: if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);
1456: PetscViewerBinaryWrite(viewer,colidxs,nz,PETSC_INT);
1457: PetscFree(colidxs);
1459: /* store nonzero values */
1460: PetscMalloc1(nz,&matvals);
1461: for (cnt=0, i=0; i<A->mbs; i++)
1462: for (k=0; k<bs; k++)
1463: for (j=A->i[i]; j<A->i[i+1]; j++)
1464: for (l=0; l<bs; l++)
1465: matvals[cnt++] = A->a[bs*(bs*j + l) + k];
1466: if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);
1467: PetscViewerBinaryWrite(viewer,matvals,nz,PETSC_SCALAR);
1468: PetscFree(matvals);
1470: /* write block size option to the viewer's .info file */
1471: MatView_Binary_BlockSizes(mat,viewer);
1472: return(0);
1473: }
1475: static PetscErrorCode MatView_SeqBAIJ_ASCII_structonly(Mat A,PetscViewer viewer)
1476: {
1478: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1479: PetscInt i,bs = A->rmap->bs,k;
1482: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
1483: for (i=0; i<a->mbs; i++) {
1484: PetscViewerASCIIPrintf(viewer,"row %D-%D:",i*bs,i*bs+bs-1);
1485: for (k=a->i[i]; k<a->i[i+1]; k++) {
1486: PetscViewerASCIIPrintf(viewer," (%D-%D) ",bs*a->j[k],bs*a->j[k]+bs-1);
1487: }
1488: PetscViewerASCIIPrintf(viewer,"\n");
1489: }
1490: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
1491: return(0);
1492: }
1494: static PetscErrorCode MatView_SeqBAIJ_ASCII(Mat A,PetscViewer viewer)
1495: {
1496: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1497: PetscErrorCode ierr;
1498: PetscInt i,j,bs = A->rmap->bs,k,l,bs2=a->bs2;
1499: PetscViewerFormat format;
1502: if (A->structure_only) {
1503: MatView_SeqBAIJ_ASCII_structonly(A,viewer);
1504: return(0);
1505: }
1507: PetscViewerGetFormat(viewer,&format);
1508: if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1509: PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);
1510: } else if (format == PETSC_VIEWER_ASCII_MATLAB) {
1511: const char *matname;
1512: Mat aij;
1513: MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&aij);
1514: PetscObjectGetName((PetscObject)A,&matname);
1515: PetscObjectSetName((PetscObject)aij,matname);
1516: MatView(aij,viewer);
1517: MatDestroy(&aij);
1518: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1519: return(0);
1520: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
1521: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
1522: for (i=0; i<a->mbs; i++) {
1523: for (j=0; j<bs; j++) {
1524: PetscViewerASCIIPrintf(viewer,"row %D:",i*bs+j);
1525: for (k=a->i[i]; k<a->i[i+1]; k++) {
1526: for (l=0; l<bs; l++) {
1527: #if defined(PETSC_USE_COMPLEX)
1528: if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) > 0.0 && PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) {
1529: PetscViewerASCIIPrintf(viewer," (%D, %g + %gi) ",bs*a->j[k]+l,
1530: (double)PetscRealPart(a->a[bs2*k + l*bs + j]),(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));
1531: } else if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) < 0.0 && PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) {
1532: PetscViewerASCIIPrintf(viewer," (%D, %g - %gi) ",bs*a->j[k]+l,
1533: (double)PetscRealPart(a->a[bs2*k + l*bs + j]),-(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));
1534: } else if (PetscRealPart(a->a[bs2*k + l*bs + j]) != 0.0) {
1535: PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)PetscRealPart(a->a[bs2*k + l*bs + j]));
1536: }
1537: #else
1538: if (a->a[bs2*k + l*bs + j] != 0.0) {
1539: PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)a->a[bs2*k + l*bs + j]);
1540: }
1541: #endif
1542: }
1543: }
1544: PetscViewerASCIIPrintf(viewer,"\n");
1545: }
1546: }
1547: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
1548: } else {
1549: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
1550: for (i=0; i<a->mbs; i++) {
1551: for (j=0; j<bs; j++) {
1552: PetscViewerASCIIPrintf(viewer,"row %D:",i*bs+j);
1553: for (k=a->i[i]; k<a->i[i+1]; k++) {
1554: for (l=0; l<bs; l++) {
1555: #if defined(PETSC_USE_COMPLEX)
1556: if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) > 0.0) {
1557: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i) ",bs*a->j[k]+l,
1558: (double)PetscRealPart(a->a[bs2*k + l*bs + j]),(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));
1559: } else if (PetscImaginaryPart(a->a[bs2*k + l*bs + j]) < 0.0) {
1560: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i) ",bs*a->j[k]+l,
1561: (double)PetscRealPart(a->a[bs2*k + l*bs + j]),-(double)PetscImaginaryPart(a->a[bs2*k + l*bs + j]));
1562: } else {
1563: PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)PetscRealPart(a->a[bs2*k + l*bs + j]));
1564: }
1565: #else
1566: PetscViewerASCIIPrintf(viewer," (%D, %g) ",bs*a->j[k]+l,(double)a->a[bs2*k + l*bs + j]);
1567: #endif
1568: }
1569: }
1570: PetscViewerASCIIPrintf(viewer,"\n");
1571: }
1572: }
1573: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
1574: }
1575: PetscViewerFlush(viewer);
1576: return(0);
1577: }
1579: #include <petscdraw.h>
1580: static PetscErrorCode MatView_SeqBAIJ_Draw_Zoom(PetscDraw draw,void *Aa)
1581: {
1582: Mat A = (Mat) Aa;
1583: Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data;
1584: PetscErrorCode ierr;
1585: PetscInt row,i,j,k,l,mbs=a->mbs,color,bs=A->rmap->bs,bs2=a->bs2;
1586: PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r;
1587: MatScalar *aa;
1588: PetscViewer viewer;
1589: PetscViewerFormat format;
1592: PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);
1593: PetscViewerGetFormat(viewer,&format);
1594: PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);
1596: /* loop over matrix elements drawing boxes */
1598: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
1599: PetscDrawCollectiveBegin(draw);
1600: /* Blue for negative, Cyan for zero and Red for positive */
1601: color = PETSC_DRAW_BLUE;
1602: for (i=0,row=0; i<mbs; i++,row+=bs) {
1603: for (j=a->i[i]; j<a->i[i+1]; j++) {
1604: y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0;
1605: x_l = a->j[j]*bs; x_r = x_l + 1.0;
1606: aa = a->a + j*bs2;
1607: for (k=0; k<bs; k++) {
1608: for (l=0; l<bs; l++) {
1609: if (PetscRealPart(*aa++) >= 0.) continue;
1610: PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);
1611: }
1612: }
1613: }
1614: }
1615: color = PETSC_DRAW_CYAN;
1616: for (i=0,row=0; i<mbs; i++,row+=bs) {
1617: for (j=a->i[i]; j<a->i[i+1]; j++) {
1618: y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0;
1619: x_l = a->j[j]*bs; x_r = x_l + 1.0;
1620: aa = a->a + j*bs2;
1621: for (k=0; k<bs; k++) {
1622: for (l=0; l<bs; l++) {
1623: if (PetscRealPart(*aa++) != 0.) continue;
1624: PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);
1625: }
1626: }
1627: }
1628: }
1629: color = PETSC_DRAW_RED;
1630: for (i=0,row=0; i<mbs; i++,row+=bs) {
1631: for (j=a->i[i]; j<a->i[i+1]; j++) {
1632: y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0;
1633: x_l = a->j[j]*bs; x_r = x_l + 1.0;
1634: aa = a->a + j*bs2;
1635: for (k=0; k<bs; k++) {
1636: for (l=0; l<bs; l++) {
1637: if (PetscRealPart(*aa++) <= 0.) continue;
1638: PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);
1639: }
1640: }
1641: }
1642: }
1643: PetscDrawCollectiveEnd(draw);
1644: } else {
1645: /* use contour shading to indicate magnitude of values */
1646: /* first determine max of all nonzero values */
1647: PetscReal minv = 0.0, maxv = 0.0;
1648: PetscDraw popup;
1650: for (i=0; i<a->nz*a->bs2; i++) {
1651: if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]);
1652: }
1653: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1654: PetscDrawGetPopup(draw,&popup);
1655: PetscDrawScalePopup(popup,0.0,maxv);
1657: PetscDrawCollectiveBegin(draw);
1658: for (i=0,row=0; i<mbs; i++,row+=bs) {
1659: for (j=a->i[i]; j<a->i[i+1]; j++) {
1660: y_l = A->rmap->N - row - 1.0; y_r = y_l + 1.0;
1661: x_l = a->j[j]*bs; x_r = x_l + 1.0;
1662: aa = a->a + j*bs2;
1663: for (k=0; k<bs; k++) {
1664: for (l=0; l<bs; l++) {
1665: MatScalar v = *aa++;
1666: color = PetscDrawRealToColor(PetscAbsScalar(v),minv,maxv);
1667: PetscDrawRectangle(draw,x_l+k,y_l-l,x_r+k,y_r-l,color,color,color,color);
1668: }
1669: }
1670: }
1671: }
1672: PetscDrawCollectiveEnd(draw);
1673: }
1674: return(0);
1675: }
1677: static PetscErrorCode MatView_SeqBAIJ_Draw(Mat A,PetscViewer viewer)
1678: {
1680: PetscReal xl,yl,xr,yr,w,h;
1681: PetscDraw draw;
1682: PetscBool isnull;
1685: PetscViewerDrawGetDraw(viewer,0,&draw);
1686: PetscDrawIsNull(draw,&isnull);
1687: if (isnull) return(0);
1689: xr = A->cmap->n; yr = A->rmap->N; h = yr/10.0; w = xr/10.0;
1690: xr += w; yr += h; xl = -w; yl = -h;
1691: PetscDrawSetCoordinates(draw,xl,yl,xr,yr);
1692: PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);
1693: PetscDrawZoom(draw,MatView_SeqBAIJ_Draw_Zoom,A);
1694: PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);
1695: PetscDrawSave(draw);
1696: return(0);
1697: }
1699: PetscErrorCode MatView_SeqBAIJ(Mat A,PetscViewer viewer)
1700: {
1702: PetscBool iascii,isbinary,isdraw;
1705: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1706: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1707: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1708: if (iascii) {
1709: MatView_SeqBAIJ_ASCII(A,viewer);
1710: } else if (isbinary) {
1711: MatView_SeqBAIJ_Binary(A,viewer);
1712: } else if (isdraw) {
1713: MatView_SeqBAIJ_Draw(A,viewer);
1714: } else {
1715: Mat B;
1716: MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);
1717: MatView(B,viewer);
1718: MatDestroy(&B);
1719: }
1720: return(0);
1721: }
1724: PetscErrorCode MatGetValues_SeqBAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[])
1725: {
1726: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1727: PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j;
1728: PetscInt *ai = a->i,*ailen = a->ilen;
1729: PetscInt brow,bcol,ridx,cidx,bs=A->rmap->bs,bs2=a->bs2;
1730: MatScalar *ap,*aa = a->a;
1733: for (k=0; k<m; k++) { /* loop over rows */
1734: row = im[k]; brow = row/bs;
1735: if (row < 0) {v += n; continue;} /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); */
1736: if (row >= A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D too large", row);
1737: rp = aj + ai[brow]; ap = aa + bs2*ai[brow];
1738: nrow = ailen[brow];
1739: for (l=0; l<n; l++) { /* loop over columns */
1740: if (in[l] < 0) {v++; continue;} /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column"); */
1741: if (in[l] >= A->cmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column %D too large", in[l]);
1742: col = in[l];
1743: bcol = col/bs;
1744: cidx = col%bs;
1745: ridx = row%bs;
1746: high = nrow;
1747: low = 0; /* assume unsorted */
1748: while (high-low > 5) {
1749: t = (low+high)/2;
1750: if (rp[t] > bcol) high = t;
1751: else low = t;
1752: }
1753: for (i=low; i<high; i++) {
1754: if (rp[i] > bcol) break;
1755: if (rp[i] == bcol) {
1756: *v++ = ap[bs2*i+bs*cidx+ridx];
1757: goto finished;
1758: }
1759: }
1760: *v++ = 0.0;
1761: finished:;
1762: }
1763: }
1764: return(0);
1765: }
1767: PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
1768: {
1769: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1770: PetscInt *rp,k,low,high,t,ii,jj,row,nrow,i,col,l,rmax,N,lastcol = -1;
1771: PetscInt *imax=a->imax,*ai=a->i,*ailen=a->ilen;
1772: PetscErrorCode ierr;
1773: PetscInt *aj =a->j,nonew=a->nonew,bs2=a->bs2,bs=A->rmap->bs,stepval;
1774: PetscBool roworiented=a->roworiented;
1775: const PetscScalar *value = v;
1776: MatScalar *ap=NULL,*aa = a->a,*bap;
1779: if (roworiented) {
1780: stepval = (n-1)*bs;
1781: } else {
1782: stepval = (m-1)*bs;
1783: }
1784: for (k=0; k<m; k++) { /* loop over added rows */
1785: row = im[k];
1786: if (row < 0) continue;
1787: if (PetscUnlikelyDebug(row >= a->mbs)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Block row index too large %D max %D",row,a->mbs-1);
1788: rp = aj + ai[row];
1789: if (!A->structure_only) ap = aa + bs2*ai[row];
1790: rmax = imax[row];
1791: nrow = ailen[row];
1792: low = 0;
1793: high = nrow;
1794: for (l=0; l<n; l++) { /* loop over added columns */
1795: if (in[l] < 0) continue;
1796: if (PetscUnlikelyDebug(in[l] >= a->nbs)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Block column index too large %D max %D",in[l],a->nbs-1);
1797: col = in[l];
1798: if (!A->structure_only) {
1799: if (roworiented) {
1800: value = v + (k*(stepval+bs) + l)*bs;
1801: } else {
1802: value = v + (l*(stepval+bs) + k)*bs;
1803: }
1804: }
1805: if (col <= lastcol) low = 0;
1806: else high = nrow;
1807: lastcol = col;
1808: while (high-low > 7) {
1809: t = (low+high)/2;
1810: if (rp[t] > col) high = t;
1811: else low = t;
1812: }
1813: for (i=low; i<high; i++) {
1814: if (rp[i] > col) break;
1815: if (rp[i] == col) {
1816: if (A->structure_only) goto noinsert2;
1817: bap = ap + bs2*i;
1818: if (roworiented) {
1819: if (is == ADD_VALUES) {
1820: for (ii=0; ii<bs; ii++,value+=stepval) {
1821: for (jj=ii; jj<bs2; jj+=bs) {
1822: bap[jj] += *value++;
1823: }
1824: }
1825: } else {
1826: for (ii=0; ii<bs; ii++,value+=stepval) {
1827: for (jj=ii; jj<bs2; jj+=bs) {
1828: bap[jj] = *value++;
1829: }
1830: }
1831: }
1832: } else {
1833: if (is == ADD_VALUES) {
1834: for (ii=0; ii<bs; ii++,value+=bs+stepval) {
1835: for (jj=0; jj<bs; jj++) {
1836: bap[jj] += value[jj];
1837: }
1838: bap += bs;
1839: }
1840: } else {
1841: for (ii=0; ii<bs; ii++,value+=bs+stepval) {
1842: for (jj=0; jj<bs; jj++) {
1843: bap[jj] = value[jj];
1844: }
1845: bap += bs;
1846: }
1847: }
1848: }
1849: goto noinsert2;
1850: }
1851: }
1852: if (nonew == 1) goto noinsert2;
1853: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new blocked index new nonzero block (%D, %D) in the matrix", row, col);
1854: if (A->structure_only) {
1855: MatSeqXAIJReallocateAIJ_structure_only(A,a->mbs,bs2,nrow,row,col,rmax,ai,aj,rp,imax,nonew,MatScalar);
1856: } else {
1857: MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
1858: }
1859: N = nrow++ - 1; high++;
1860: /* shift up all the later entries in this row */
1861: PetscArraymove(rp+i+1,rp+i,N-i+1);
1862: rp[i] = col;
1863: if (!A->structure_only) {
1864: PetscArraymove(ap+bs2*(i+1),ap+bs2*i,bs2*(N-i+1));
1865: bap = ap + bs2*i;
1866: if (roworiented) {
1867: for (ii=0; ii<bs; ii++,value+=stepval) {
1868: for (jj=ii; jj<bs2; jj+=bs) {
1869: bap[jj] = *value++;
1870: }
1871: }
1872: } else {
1873: for (ii=0; ii<bs; ii++,value+=stepval) {
1874: for (jj=0; jj<bs; jj++) {
1875: *bap++ = *value++;
1876: }
1877: }
1878: }
1879: }
1880: noinsert2:;
1881: low = i;
1882: }
1883: ailen[row] = nrow;
1884: }
1885: return(0);
1886: }
1888: PetscErrorCode MatAssemblyEnd_SeqBAIJ(Mat A,MatAssemblyType mode)
1889: {
1890: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
1891: PetscInt fshift = 0,i,*ai = a->i,*aj = a->j,*imax = a->imax;
1892: PetscInt m = A->rmap->N,*ip,N,*ailen = a->ilen;
1894: PetscInt mbs = a->mbs,bs2 = a->bs2,rmax = 0;
1895: MatScalar *aa = a->a,*ap;
1896: PetscReal ratio=0.6;
1899: if (mode == MAT_FLUSH_ASSEMBLY) return(0);
1901: if (m) rmax = ailen[0];
1902: for (i=1; i<mbs; i++) {
1903: /* move each row back by the amount of empty slots (fshift) before it*/
1904: fshift += imax[i-1] - ailen[i-1];
1905: rmax = PetscMax(rmax,ailen[i]);
1906: if (fshift) {
1907: ip = aj + ai[i];
1908: ap = aa + bs2*ai[i];
1909: N = ailen[i];
1910: PetscArraymove(ip-fshift,ip,N);
1911: if (!A->structure_only) {
1912: PetscArraymove(ap-bs2*fshift,ap,bs2*N);
1913: }
1914: }
1915: ai[i] = ai[i-1] + ailen[i-1];
1916: }
1917: if (mbs) {
1918: fshift += imax[mbs-1] - ailen[mbs-1];
1919: ai[mbs] = ai[mbs-1] + ailen[mbs-1];
1920: }
1922: /* reset ilen and imax for each row */
1923: a->nonzerorowcnt = 0;
1924: if (A->structure_only) {
1925: PetscFree2(a->imax,a->ilen);
1926: } else { /* !A->structure_only */
1927: for (i=0; i<mbs; i++) {
1928: ailen[i] = imax[i] = ai[i+1] - ai[i];
1929: a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0);
1930: }
1931: }
1932: a->nz = ai[mbs];
1934: /* diagonals may have moved, so kill the diagonal pointers */
1935: a->idiagvalid = PETSC_FALSE;
1936: if (fshift && a->diag) {
1937: PetscFree(a->diag);
1938: PetscLogObjectMemory((PetscObject)A,-(mbs+1)*sizeof(PetscInt));
1939: a->diag = NULL;
1940: }
1941: if (fshift && a->nounused == -1) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB, "Unused space detected in matrix: %D X %D block size %D, %D unneeded", m, A->cmap->n, A->rmap->bs, fshift*bs2);
1942: PetscInfo5(A,"Matrix size: %D X %D, block size %D; storage space: %D unneeded, %D used\n",m,A->cmap->n,A->rmap->bs,fshift*bs2,a->nz*bs2);
1943: PetscInfo1(A,"Number of mallocs during MatSetValues is %D\n",a->reallocs);
1944: PetscInfo1(A,"Most nonzeros blocks in any row is %D\n",rmax);
1946: A->info.mallocs += a->reallocs;
1947: a->reallocs = 0;
1948: A->info.nz_unneeded = (PetscReal)fshift*bs2;
1949: a->rmax = rmax;
1951: if (!A->structure_only) {
1952: MatCheckCompressedRow(A,a->nonzerorowcnt,&a->compressedrow,a->i,mbs,ratio);
1953: }
1954: return(0);
1955: }
1957: /*
1958: This function returns an array of flags which indicate the locations of contiguous
1959: blocks that should be zeroed. for eg: if bs = 3 and is = [0,1,2,3,5,6,7,8,9]
1960: then the resulting sizes = [3,1,1,3,1] correspondig to sets [(0,1,2),(3),(5),(6,7,8),(9)]
1961: Assume: sizes should be long enough to hold all the values.
1962: */
1963: static PetscErrorCode MatZeroRows_SeqBAIJ_Check_Blocks(PetscInt idx[],PetscInt n,PetscInt bs,PetscInt sizes[], PetscInt *bs_max)
1964: {
1965: PetscInt i,j,k,row;
1966: PetscBool flg;
1969: for (i=0,j=0; i<n; j++) {
1970: row = idx[i];
1971: if (row%bs!=0) { /* Not the begining of a block */
1972: sizes[j] = 1;
1973: i++;
1974: } else if (i+bs > n) { /* complete block doesn't exist (at idx end) */
1975: sizes[j] = 1; /* Also makes sure atleast 'bs' values exist for next else */
1976: i++;
1977: } else { /* Begining of the block, so check if the complete block exists */
1978: flg = PETSC_TRUE;
1979: for (k=1; k<bs; k++) {
1980: if (row+k != idx[i+k]) { /* break in the block */
1981: flg = PETSC_FALSE;
1982: break;
1983: }
1984: }
1985: if (flg) { /* No break in the bs */
1986: sizes[j] = bs;
1987: i += bs;
1988: } else {
1989: sizes[j] = 1;
1990: i++;
1991: }
1992: }
1993: }
1994: *bs_max = j;
1995: return(0);
1996: }
1998: PetscErrorCode MatZeroRows_SeqBAIJ(Mat A,PetscInt is_n,const PetscInt is_idx[],PetscScalar diag,Vec x, Vec b)
1999: {
2000: Mat_SeqBAIJ *baij=(Mat_SeqBAIJ*)A->data;
2001: PetscErrorCode ierr;
2002: PetscInt i,j,k,count,*rows;
2003: PetscInt bs=A->rmap->bs,bs2=baij->bs2,*sizes,row,bs_max;
2004: PetscScalar zero = 0.0;
2005: MatScalar *aa;
2006: const PetscScalar *xx;
2007: PetscScalar *bb;
2010: /* fix right hand side if needed */
2011: if (x && b) {
2012: VecGetArrayRead(x,&xx);
2013: VecGetArray(b,&bb);
2014: for (i=0; i<is_n; i++) {
2015: bb[is_idx[i]] = diag*xx[is_idx[i]];
2016: }
2017: VecRestoreArrayRead(x,&xx);
2018: VecRestoreArray(b,&bb);
2019: }
2021: /* Make a copy of the IS and sort it */
2022: /* allocate memory for rows,sizes */
2023: PetscMalloc2(is_n,&rows,2*is_n,&sizes);
2025: /* copy IS values to rows, and sort them */
2026: for (i=0; i<is_n; i++) rows[i] = is_idx[i];
2027: PetscSortInt(is_n,rows);
2029: if (baij->keepnonzeropattern) {
2030: for (i=0; i<is_n; i++) sizes[i] = 1;
2031: bs_max = is_n;
2032: } else {
2033: MatZeroRows_SeqBAIJ_Check_Blocks(rows,is_n,bs,sizes,&bs_max);
2034: A->nonzerostate++;
2035: }
2037: for (i=0,j=0; i<bs_max; j+=sizes[i],i++) {
2038: row = rows[j];
2039: if (row < 0 || row > A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range",row);
2040: count = (baij->i[row/bs +1] - baij->i[row/bs])*bs;
2041: aa = ((MatScalar*)(baij->a)) + baij->i[row/bs]*bs2 + (row%bs);
2042: if (sizes[i] == bs && !baij->keepnonzeropattern) {
2043: if (diag != (PetscScalar)0.0) {
2044: if (baij->ilen[row/bs] > 0) {
2045: baij->ilen[row/bs] = 1;
2046: baij->j[baij->i[row/bs]] = row/bs;
2048: PetscArrayzero(aa,count*bs);
2049: }
2050: /* Now insert all the diagonal values for this bs */
2051: for (k=0; k<bs; k++) {
2052: (*A->ops->setvalues)(A,1,rows+j+k,1,rows+j+k,&diag,INSERT_VALUES);
2053: }
2054: } else { /* (diag == 0.0) */
2055: baij->ilen[row/bs] = 0;
2056: } /* end (diag == 0.0) */
2057: } else { /* (sizes[i] != bs) */
2058: if (PetscUnlikelyDebug(sizes[i] != 1)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal Error. Value should be 1");
2059: for (k=0; k<count; k++) {
2060: aa[0] = zero;
2061: aa += bs;
2062: }
2063: if (diag != (PetscScalar)0.0) {
2064: (*A->ops->setvalues)(A,1,rows+j,1,rows+j,&diag,INSERT_VALUES);
2065: }
2066: }
2067: }
2069: PetscFree2(rows,sizes);
2070: MatAssemblyEnd_SeqBAIJ(A,MAT_FINAL_ASSEMBLY);
2071: return(0);
2072: }
2074: PetscErrorCode MatZeroRowsColumns_SeqBAIJ(Mat A,PetscInt is_n,const PetscInt is_idx[],PetscScalar diag,Vec x, Vec b)
2075: {
2076: Mat_SeqBAIJ *baij=(Mat_SeqBAIJ*)A->data;
2077: PetscErrorCode ierr;
2078: PetscInt i,j,k,count;
2079: PetscInt bs =A->rmap->bs,bs2=baij->bs2,row,col;
2080: PetscScalar zero = 0.0;
2081: MatScalar *aa;
2082: const PetscScalar *xx;
2083: PetscScalar *bb;
2084: PetscBool *zeroed,vecs = PETSC_FALSE;
2087: /* fix right hand side if needed */
2088: if (x && b) {
2089: VecGetArrayRead(x,&xx);
2090: VecGetArray(b,&bb);
2091: vecs = PETSC_TRUE;
2092: }
2094: /* zero the columns */
2095: PetscCalloc1(A->rmap->n,&zeroed);
2096: for (i=0; i<is_n; i++) {
2097: if (is_idx[i] < 0 || is_idx[i] >= A->rmap->N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range",is_idx[i]);
2098: zeroed[is_idx[i]] = PETSC_TRUE;
2099: }
2100: for (i=0; i<A->rmap->N; i++) {
2101: if (!zeroed[i]) {
2102: row = i/bs;
2103: for (j=baij->i[row]; j<baij->i[row+1]; j++) {
2104: for (k=0; k<bs; k++) {
2105: col = bs*baij->j[j] + k;
2106: if (zeroed[col]) {
2107: aa = ((MatScalar*)(baij->a)) + j*bs2 + (i%bs) + bs*k;
2108: if (vecs) bb[i] -= aa[0]*xx[col];
2109: aa[0] = 0.0;
2110: }
2111: }
2112: }
2113: } else if (vecs) bb[i] = diag*xx[i];
2114: }
2115: PetscFree(zeroed);
2116: if (vecs) {
2117: VecRestoreArrayRead(x,&xx);
2118: VecRestoreArray(b,&bb);
2119: }
2121: /* zero the rows */
2122: for (i=0; i<is_n; i++) {
2123: row = is_idx[i];
2124: count = (baij->i[row/bs +1] - baij->i[row/bs])*bs;
2125: aa = ((MatScalar*)(baij->a)) + baij->i[row/bs]*bs2 + (row%bs);
2126: for (k=0; k<count; k++) {
2127: aa[0] = zero;
2128: aa += bs;
2129: }
2130: if (diag != (PetscScalar)0.0) {
2131: (*A->ops->setvalues)(A,1,&row,1,&row,&diag,INSERT_VALUES);
2132: }
2133: }
2134: MatAssemblyEnd_SeqBAIJ(A,MAT_FINAL_ASSEMBLY);
2135: return(0);
2136: }
2138: PetscErrorCode MatSetValues_SeqBAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
2139: {
2140: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
2141: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N,lastcol = -1;
2142: PetscInt *imax=a->imax,*ai=a->i,*ailen=a->ilen;
2143: PetscInt *aj =a->j,nonew=a->nonew,bs=A->rmap->bs,brow,bcol;
2145: PetscInt ridx,cidx,bs2=a->bs2;
2146: PetscBool roworiented=a->roworiented;
2147: MatScalar *ap=NULL,value=0.0,*aa=a->a,*bap;
2150: for (k=0; k<m; k++) { /* loop over added rows */
2151: row = im[k];
2152: brow = row/bs;
2153: if (row < 0) continue;
2154: if (PetscUnlikelyDebug(row >= A->rmap->N)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->N-1);
2155: rp = aj + ai[brow];
2156: if (!A->structure_only) ap = aa + bs2*ai[brow];
2157: rmax = imax[brow];
2158: nrow = ailen[brow];
2159: low = 0;
2160: high = nrow;
2161: for (l=0; l<n; l++) { /* loop over added columns */
2162: if (in[l] < 0) continue;
2163: if (PetscUnlikelyDebug(in[l] >= A->cmap->n)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1);
2164: col = in[l]; bcol = col/bs;
2165: ridx = row % bs; cidx = col % bs;
2166: if (!A->structure_only) {
2167: if (roworiented) {
2168: value = v[l + k*n];
2169: } else {
2170: value = v[k + l*m];
2171: }
2172: }
2173: if (col <= lastcol) low = 0; else high = nrow;
2174: lastcol = col;
2175: while (high-low > 7) {
2176: t = (low+high)/2;
2177: if (rp[t] > bcol) high = t;
2178: else low = t;
2179: }
2180: for (i=low; i<high; i++) {
2181: if (rp[i] > bcol) break;
2182: if (rp[i] == bcol) {
2183: bap = ap + bs2*i + bs*cidx + ridx;
2184: if (!A->structure_only) {
2185: if (is == ADD_VALUES) *bap += value;
2186: else *bap = value;
2187: }
2188: goto noinsert1;
2189: }
2190: }
2191: if (nonew == 1) goto noinsert1;
2192: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) in the matrix", row, col);
2193: if (A->structure_only) {
2194: MatSeqXAIJReallocateAIJ_structure_only(A,a->mbs,bs2,nrow,brow,bcol,rmax,ai,aj,rp,imax,nonew,MatScalar);
2195: } else {
2196: MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
2197: }
2198: N = nrow++ - 1; high++;
2199: /* shift up all the later entries in this row */
2200: PetscArraymove(rp+i+1,rp+i,N-i+1);
2201: rp[i] = bcol;
2202: if (!A->structure_only) {
2203: PetscArraymove(ap+bs2*(i+1),ap+bs2*i,bs2*(N-i+1));
2204: PetscArrayzero(ap+bs2*i,bs2);
2205: ap[bs2*i + bs*cidx + ridx] = value;
2206: }
2207: a->nz++;
2208: A->nonzerostate++;
2209: noinsert1:;
2210: low = i;
2211: }
2212: ailen[brow] = nrow;
2213: }
2214: return(0);
2215: }
2217: PetscErrorCode MatILUFactor_SeqBAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info)
2218: {
2219: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)inA->data;
2220: Mat outA;
2222: PetscBool row_identity,col_identity;
2225: if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for in-place ILU");
2226: ISIdentity(row,&row_identity);
2227: ISIdentity(col,&col_identity);
2228: if (!row_identity || !col_identity) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for in-place ILU");
2230: outA = inA;
2231: inA->factortype = MAT_FACTOR_LU;
2232: PetscFree(inA->solvertype);
2233: PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);
2235: MatMarkDiagonal_SeqBAIJ(inA);
2237: PetscObjectReference((PetscObject)row);
2238: ISDestroy(&a->row);
2239: a->row = row;
2240: PetscObjectReference((PetscObject)col);
2241: ISDestroy(&a->col);
2242: a->col = col;
2244: /* Create the invert permutation so that it can be used in MatLUFactorNumeric() */
2245: ISDestroy(&a->icol);
2246: ISInvertPermutation(col,PETSC_DECIDE,&a->icol);
2247: PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);
2249: MatSeqBAIJSetNumericFactorization_inplace(inA,(PetscBool)(row_identity && col_identity));
2250: if (!a->solve_work) {
2251: PetscMalloc1(inA->rmap->N+inA->rmap->bs,&a->solve_work);
2252: PetscLogObjectMemory((PetscObject)inA,(inA->rmap->N+inA->rmap->bs)*sizeof(PetscScalar));
2253: }
2254: MatLUFactorNumeric(outA,inA,info);
2255: return(0);
2256: }
2258: PetscErrorCode MatSeqBAIJSetColumnIndices_SeqBAIJ(Mat mat,PetscInt *indices)
2259: {
2260: Mat_SeqBAIJ *baij = (Mat_SeqBAIJ*)mat->data;
2261: PetscInt i,nz,mbs;
2264: nz = baij->maxnz;
2265: mbs = baij->mbs;
2266: for (i=0; i<nz; i++) {
2267: baij->j[i] = indices[i];
2268: }
2269: baij->nz = nz;
2270: for (i=0; i<mbs; i++) {
2271: baij->ilen[i] = baij->imax[i];
2272: }
2273: return(0);
2274: }
2276: /*@
2277: MatSeqBAIJSetColumnIndices - Set the column indices for all the rows
2278: in the matrix.
2280: Input Parameters:
2281: + mat - the SeqBAIJ matrix
2282: - indices - the column indices
2284: Level: advanced
2286: Notes:
2287: This can be called if you have precomputed the nonzero structure of the
2288: matrix and want to provide it to the matrix object to improve the performance
2289: of the MatSetValues() operation.
2291: You MUST have set the correct numbers of nonzeros per row in the call to
2292: MatCreateSeqBAIJ(), and the columns indices MUST be sorted.
2294: MUST be called before any calls to MatSetValues();
2296: @*/
2297: PetscErrorCode MatSeqBAIJSetColumnIndices(Mat mat,PetscInt *indices)
2298: {
2304: PetscUseMethod(mat,"MatSeqBAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));
2305: return(0);
2306: }
2308: PetscErrorCode MatGetRowMaxAbs_SeqBAIJ(Mat A,Vec v,PetscInt idx[])
2309: {
2310: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
2312: PetscInt i,j,n,row,bs,*ai,*aj,mbs;
2313: PetscReal atmp;
2314: PetscScalar *x,zero = 0.0;
2315: MatScalar *aa;
2316: PetscInt ncols,brow,krow,kcol;
2319: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2320: bs = A->rmap->bs;
2321: aa = a->a;
2322: ai = a->i;
2323: aj = a->j;
2324: mbs = a->mbs;
2326: VecSet(v,zero);
2327: VecGetArray(v,&x);
2328: VecGetLocalSize(v,&n);
2329: if (n != A->rmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2330: for (i=0; i<mbs; i++) {
2331: ncols = ai[1] - ai[0]; ai++;
2332: brow = bs*i;
2333: for (j=0; j<ncols; j++) {
2334: for (kcol=0; kcol<bs; kcol++) {
2335: for (krow=0; krow<bs; krow++) {
2336: atmp = PetscAbsScalar(*aa);aa++;
2337: row = brow + krow; /* row index */
2338: if (PetscAbsScalar(x[row]) < atmp) {x[row] = atmp; if (idx) idx[row] = bs*(*aj) + kcol;}
2339: }
2340: }
2341: aj++;
2342: }
2343: }
2344: VecRestoreArray(v,&x);
2345: return(0);
2346: }
2348: PetscErrorCode MatCopy_SeqBAIJ(Mat A,Mat B,MatStructure str)
2349: {
2353: /* If the two matrices have the same copy implementation, use fast copy. */
2354: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2355: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
2356: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)B->data;
2357: PetscInt ambs=a->mbs,bmbs=b->mbs,abs=A->rmap->bs,bbs=B->rmap->bs,bs2=abs*abs;
2359: if (a->i[ambs] != b->i[bmbs]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Number of nonzero blocks in matrices A %D and B %D are different",a->i[ambs],b->i[bmbs]);
2360: if (abs != bbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Block size A %D and B %D are different",abs,bbs);
2361: PetscArraycpy(b->a,a->a,bs2*a->i[ambs]);
2362: PetscObjectStateIncrease((PetscObject)B);
2363: } else {
2364: MatCopy_Basic(A,B,str);
2365: }
2366: return(0);
2367: }
2369: PetscErrorCode MatSetUp_SeqBAIJ(Mat A)
2370: {
2374: MatSeqBAIJSetPreallocation(A,A->rmap->bs,PETSC_DEFAULT,NULL);
2375: return(0);
2376: }
2378: static PetscErrorCode MatSeqBAIJGetArray_SeqBAIJ(Mat A,PetscScalar *array[])
2379: {
2380: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
2383: *array = a->a;
2384: return(0);
2385: }
2387: static PetscErrorCode MatSeqBAIJRestoreArray_SeqBAIJ(Mat A,PetscScalar *array[])
2388: {
2390: *array = NULL;
2391: return(0);
2392: }
2394: PetscErrorCode MatAXPYGetPreallocation_SeqBAIJ(Mat Y,Mat X,PetscInt *nnz)
2395: {
2396: PetscInt bs = Y->rmap->bs,mbs = Y->rmap->N/bs;
2397: Mat_SeqBAIJ *x = (Mat_SeqBAIJ*)X->data;
2398: Mat_SeqBAIJ *y = (Mat_SeqBAIJ*)Y->data;
2402: /* Set the number of nonzeros in the new matrix */
2403: MatAXPYGetPreallocation_SeqX_private(mbs,x->i,x->j,y->i,y->j,nnz);
2404: return(0);
2405: }
2407: PetscErrorCode MatAXPY_SeqBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2408: {
2409: Mat_SeqBAIJ *x = (Mat_SeqBAIJ*)X->data,*y = (Mat_SeqBAIJ*)Y->data;
2411: PetscInt bs=Y->rmap->bs,bs2=bs*bs;
2412: PetscBLASInt one=1;
2415: if (str == SAME_NONZERO_PATTERN) {
2416: PetscScalar alpha = a;
2417: PetscBLASInt bnz;
2418: PetscBLASIntCast(x->nz*bs2,&bnz);
2419: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2420: PetscObjectStateIncrease((PetscObject)Y);
2421: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2422: MatAXPY_Basic(Y,a,X,str);
2423: } else {
2424: Mat B;
2425: PetscInt *nnz;
2426: if (bs != X->rmap->bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrices must have same block size");
2427: PetscMalloc1(Y->rmap->N,&nnz);
2428: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2429: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2430: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2431: MatSetBlockSizesFromMats(B,Y,Y);
2432: MatSetType(B,(MatType) ((PetscObject)Y)->type_name);
2433: MatAXPYGetPreallocation_SeqBAIJ(Y,X,nnz);
2434: MatSeqBAIJSetPreallocation(B,bs,0,nnz);
2435: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2436: MatHeaderReplace(Y,&B);
2437: PetscFree(nnz);
2438: }
2439: return(0);
2440: }
2442: PetscErrorCode MatRealPart_SeqBAIJ(Mat A)
2443: {
2444: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
2445: PetscInt i,nz = a->bs2*a->i[a->mbs];
2446: MatScalar *aa = a->a;
2449: for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]);
2450: return(0);
2451: }
2453: PetscErrorCode MatImaginaryPart_SeqBAIJ(Mat A)
2454: {
2455: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
2456: PetscInt i,nz = a->bs2*a->i[a->mbs];
2457: MatScalar *aa = a->a;
2460: for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
2461: return(0);
2462: }
2464: /*
2465: Code almost identical to MatGetColumnIJ_SeqAIJ() should share common code
2466: */
2467: PetscErrorCode MatGetColumnIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
2468: {
2469: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
2471: PetscInt bs = A->rmap->bs,i,*collengths,*cia,*cja,n = A->cmap->n/bs,m = A->rmap->n/bs;
2472: PetscInt nz = a->i[m],row,*jj,mr,col;
2475: *nn = n;
2476: if (!ia) return(0);
2477: if (symmetric) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for BAIJ matrices");
2478: else {
2479: PetscCalloc1(n,&collengths);
2480: PetscMalloc1(n+1,&cia);
2481: PetscMalloc1(nz,&cja);
2482: jj = a->j;
2483: for (i=0; i<nz; i++) {
2484: collengths[jj[i]]++;
2485: }
2486: cia[0] = oshift;
2487: for (i=0; i<n; i++) {
2488: cia[i+1] = cia[i] + collengths[i];
2489: }
2490: PetscArrayzero(collengths,n);
2491: jj = a->j;
2492: for (row=0; row<m; row++) {
2493: mr = a->i[row+1] - a->i[row];
2494: for (i=0; i<mr; i++) {
2495: col = *jj++;
2497: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
2498: }
2499: }
2500: PetscFree(collengths);
2501: *ia = cia; *ja = cja;
2502: }
2503: return(0);
2504: }
2506: PetscErrorCode MatRestoreColumnIJ_SeqBAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
2507: {
2511: if (!ia) return(0);
2512: PetscFree(*ia);
2513: PetscFree(*ja);
2514: return(0);
2515: }
2517: /*
2518: MatGetColumnIJ_SeqBAIJ_Color() and MatRestoreColumnIJ_SeqBAIJ_Color() are customized from
2519: MatGetColumnIJ_SeqBAIJ() and MatRestoreColumnIJ_SeqBAIJ() by adding an output
2520: spidx[], index of a->a, to be used in MatTransposeColoringCreate() and MatFDColoringCreate()
2521: */
2522: PetscErrorCode MatGetColumnIJ_SeqBAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
2523: {
2524: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data;
2526: PetscInt i,*collengths,*cia,*cja,n=a->nbs,m=a->mbs;
2527: PetscInt nz = a->i[m],row,*jj,mr,col;
2528: PetscInt *cspidx;
2531: *nn = n;
2532: if (!ia) return(0);
2534: PetscCalloc1(n,&collengths);
2535: PetscMalloc1(n+1,&cia);
2536: PetscMalloc1(nz,&cja);
2537: PetscMalloc1(nz,&cspidx);
2538: jj = a->j;
2539: for (i=0; i<nz; i++) {
2540: collengths[jj[i]]++;
2541: }
2542: cia[0] = oshift;
2543: for (i=0; i<n; i++) {
2544: cia[i+1] = cia[i] + collengths[i];
2545: }
2546: PetscArrayzero(collengths,n);
2547: jj = a->j;
2548: for (row=0; row<m; row++) {
2549: mr = a->i[row+1] - a->i[row];
2550: for (i=0; i<mr; i++) {
2551: col = *jj++;
2552: cspidx[cia[col] + collengths[col] - oshift] = a->i[row] + i; /* index of a->j */
2553: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
2554: }
2555: }
2556: PetscFree(collengths);
2557: *ia = cia;
2558: *ja = cja;
2559: *spidx = cspidx;
2560: return(0);
2561: }
2563: PetscErrorCode MatRestoreColumnIJ_SeqBAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
2564: {
2568: MatRestoreColumnIJ_SeqBAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);
2569: PetscFree(*spidx);
2570: return(0);
2571: }
2573: PetscErrorCode MatShift_SeqBAIJ(Mat Y,PetscScalar a)
2574: {
2576: Mat_SeqBAIJ *aij = (Mat_SeqBAIJ*)Y->data;
2579: if (!Y->preallocated || !aij->nz) {
2580: MatSeqBAIJSetPreallocation(Y,Y->rmap->bs,1,NULL);
2581: }
2582: MatShift_Basic(Y,a);
2583: return(0);
2584: }
2586: /* -------------------------------------------------------------------*/
2587: static struct _MatOps MatOps_Values = {MatSetValues_SeqBAIJ,
2588: MatGetRow_SeqBAIJ,
2589: MatRestoreRow_SeqBAIJ,
2590: MatMult_SeqBAIJ_N,
2591: /* 4*/ MatMultAdd_SeqBAIJ_N,
2592: MatMultTranspose_SeqBAIJ,
2593: MatMultTransposeAdd_SeqBAIJ,
2594: NULL,
2595: NULL,
2596: NULL,
2597: /* 10*/ NULL,
2598: MatLUFactor_SeqBAIJ,
2599: NULL,
2600: NULL,
2601: MatTranspose_SeqBAIJ,
2602: /* 15*/ MatGetInfo_SeqBAIJ,
2603: MatEqual_SeqBAIJ,
2604: MatGetDiagonal_SeqBAIJ,
2605: MatDiagonalScale_SeqBAIJ,
2606: MatNorm_SeqBAIJ,
2607: /* 20*/ NULL,
2608: MatAssemblyEnd_SeqBAIJ,
2609: MatSetOption_SeqBAIJ,
2610: MatZeroEntries_SeqBAIJ,
2611: /* 24*/ MatZeroRows_SeqBAIJ,
2612: NULL,
2613: NULL,
2614: NULL,
2615: NULL,
2616: /* 29*/ MatSetUp_SeqBAIJ,
2617: NULL,
2618: NULL,
2619: NULL,
2620: NULL,
2621: /* 34*/ MatDuplicate_SeqBAIJ,
2622: NULL,
2623: NULL,
2624: MatILUFactor_SeqBAIJ,
2625: NULL,
2626: /* 39*/ MatAXPY_SeqBAIJ,
2627: MatCreateSubMatrices_SeqBAIJ,
2628: MatIncreaseOverlap_SeqBAIJ,
2629: MatGetValues_SeqBAIJ,
2630: MatCopy_SeqBAIJ,
2631: /* 44*/ NULL,
2632: MatScale_SeqBAIJ,
2633: MatShift_SeqBAIJ,
2634: NULL,
2635: MatZeroRowsColumns_SeqBAIJ,
2636: /* 49*/ NULL,
2637: MatGetRowIJ_SeqBAIJ,
2638: MatRestoreRowIJ_SeqBAIJ,
2639: MatGetColumnIJ_SeqBAIJ,
2640: MatRestoreColumnIJ_SeqBAIJ,
2641: /* 54*/ MatFDColoringCreate_SeqXAIJ,
2642: NULL,
2643: NULL,
2644: NULL,
2645: MatSetValuesBlocked_SeqBAIJ,
2646: /* 59*/ MatCreateSubMatrix_SeqBAIJ,
2647: MatDestroy_SeqBAIJ,
2648: MatView_SeqBAIJ,
2649: NULL,
2650: NULL,
2651: /* 64*/ NULL,
2652: NULL,
2653: NULL,
2654: NULL,
2655: NULL,
2656: /* 69*/ MatGetRowMaxAbs_SeqBAIJ,
2657: NULL,
2658: MatConvert_Basic,
2659: NULL,
2660: NULL,
2661: /* 74*/ NULL,
2662: MatFDColoringApply_BAIJ,
2663: NULL,
2664: NULL,
2665: NULL,
2666: /* 79*/ NULL,
2667: NULL,
2668: NULL,
2669: NULL,
2670: MatLoad_SeqBAIJ,
2671: /* 84*/ NULL,
2672: NULL,
2673: NULL,
2674: NULL,
2675: NULL,
2676: /* 89*/ NULL,
2677: NULL,
2678: NULL,
2679: NULL,
2680: NULL,
2681: /* 94*/ NULL,
2682: NULL,
2683: NULL,
2684: NULL,
2685: NULL,
2686: /* 99*/ NULL,
2687: NULL,
2688: NULL,
2689: NULL,
2690: NULL,
2691: /*104*/ NULL,
2692: MatRealPart_SeqBAIJ,
2693: MatImaginaryPart_SeqBAIJ,
2694: NULL,
2695: NULL,
2696: /*109*/ NULL,
2697: NULL,
2698: NULL,
2699: NULL,
2700: MatMissingDiagonal_SeqBAIJ,
2701: /*114*/ NULL,
2702: NULL,
2703: NULL,
2704: NULL,
2705: NULL,
2706: /*119*/ NULL,
2707: NULL,
2708: MatMultHermitianTranspose_SeqBAIJ,
2709: MatMultHermitianTransposeAdd_SeqBAIJ,
2710: NULL,
2711: /*124*/ NULL,
2712: NULL,
2713: MatInvertBlockDiagonal_SeqBAIJ,
2714: NULL,
2715: NULL,
2716: /*129*/ NULL,
2717: NULL,
2718: NULL,
2719: NULL,
2720: NULL,
2721: /*134*/ NULL,
2722: NULL,
2723: NULL,
2724: NULL,
2725: NULL,
2726: /*139*/ MatSetBlockSizes_Default,
2727: NULL,
2728: NULL,
2729: MatFDColoringSetUp_SeqXAIJ,
2730: NULL,
2731: /*144*/MatCreateMPIMatConcatenateSeqMat_SeqBAIJ,
2732: MatDestroySubMatrices_SeqBAIJ
2733: };
2735: PetscErrorCode MatStoreValues_SeqBAIJ(Mat mat)
2736: {
2737: Mat_SeqBAIJ *aij = (Mat_SeqBAIJ*)mat->data;
2738: PetscInt nz = aij->i[aij->mbs]*aij->bs2;
2742: if (aij->nonew != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
2744: /* allocate space for values if not already there */
2745: if (!aij->saved_values) {
2746: PetscMalloc1(nz+1,&aij->saved_values);
2747: PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));
2748: }
2750: /* copy values over */
2751: PetscArraycpy(aij->saved_values,aij->a,nz);
2752: return(0);
2753: }
2755: PetscErrorCode MatRetrieveValues_SeqBAIJ(Mat mat)
2756: {
2757: Mat_SeqBAIJ *aij = (Mat_SeqBAIJ*)mat->data;
2759: PetscInt nz = aij->i[aij->mbs]*aij->bs2;
2762: if (aij->nonew != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
2763: if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first");
2765: /* copy values over */
2766: PetscArraycpy(aij->a,aij->saved_values,nz);
2767: return(0);
2768: }
2770: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat, MatType,MatReuse,Mat*);
2771: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqSBAIJ(Mat, MatType,MatReuse,Mat*);
2773: PetscErrorCode MatSeqBAIJSetPreallocation_SeqBAIJ(Mat B,PetscInt bs,PetscInt nz,PetscInt *nnz)
2774: {
2775: Mat_SeqBAIJ *b;
2777: PetscInt i,mbs,nbs,bs2;
2778: PetscBool flg = PETSC_FALSE,skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE;
2781: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
2782: if (nz == MAT_SKIP_ALLOCATION) {
2783: skipallocation = PETSC_TRUE;
2784: nz = 0;
2785: }
2787: MatSetBlockSize(B,PetscAbs(bs));
2788: PetscLayoutSetUp(B->rmap);
2789: PetscLayoutSetUp(B->cmap);
2790: PetscLayoutGetBlockSize(B->rmap,&bs);
2792: B->preallocated = PETSC_TRUE;
2794: mbs = B->rmap->n/bs;
2795: nbs = B->cmap->n/bs;
2796: bs2 = bs*bs;
2798: if (mbs*bs!=B->rmap->n || nbs*bs!=B->cmap->n) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number rows %D, cols %D must be divisible by blocksize %D",B->rmap->N,B->cmap->n,bs);
2800: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
2801: if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz);
2802: if (nnz) {
2803: for (i=0; i<mbs; i++) {
2804: if (nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be less than 0: local row %D value %D",i,nnz[i]);
2805: if (nnz[i] > nbs) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be greater than block row length: local row %D value %D rowlength %D",i,nnz[i],nbs);
2806: }
2807: }
2809: b = (Mat_SeqBAIJ*)B->data;
2810: PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Optimize options for SEQBAIJ matrix 2 ","Mat");
2811: PetscOptionsBool("-mat_no_unroll","Do not optimize for block size (slow)",NULL,flg,&flg,NULL);
2812: PetscOptionsEnd();
2814: if (!flg) {
2815: switch (bs) {
2816: case 1:
2817: B->ops->mult = MatMult_SeqBAIJ_1;
2818: B->ops->multadd = MatMultAdd_SeqBAIJ_1;
2819: break;
2820: case 2:
2821: B->ops->mult = MatMult_SeqBAIJ_2;
2822: B->ops->multadd = MatMultAdd_SeqBAIJ_2;
2823: break;
2824: case 3:
2825: B->ops->mult = MatMult_SeqBAIJ_3;
2826: B->ops->multadd = MatMultAdd_SeqBAIJ_3;
2827: break;
2828: case 4:
2829: B->ops->mult = MatMult_SeqBAIJ_4;
2830: B->ops->multadd = MatMultAdd_SeqBAIJ_4;
2831: break;
2832: case 5:
2833: B->ops->mult = MatMult_SeqBAIJ_5;
2834: B->ops->multadd = MatMultAdd_SeqBAIJ_5;
2835: break;
2836: case 6:
2837: B->ops->mult = MatMult_SeqBAIJ_6;
2838: B->ops->multadd = MatMultAdd_SeqBAIJ_6;
2839: break;
2840: case 7:
2841: B->ops->mult = MatMult_SeqBAIJ_7;
2842: B->ops->multadd = MatMultAdd_SeqBAIJ_7;
2843: break;
2844: case 9:
2845: #if defined(PETSC_HAVE_IMMINTRIN_H) && defined(__AVX2__) && defined(__FMA__) && defined(PETSC_USE_REAL_DOUBLE) && !defined(PETSC_USE_COMPLEX) && !defined(PETSC_USE_64BIT_INDICES)
2846: B->ops->mult = MatMult_SeqBAIJ_9_AVX2;
2847: B->ops->multadd = MatMultAdd_SeqBAIJ_9_AVX2;
2848: #else
2849: B->ops->mult = MatMult_SeqBAIJ_N;
2850: B->ops->multadd = MatMultAdd_SeqBAIJ_N;
2851: #endif
2852: break;
2853: case 11:
2854: B->ops->mult = MatMult_SeqBAIJ_11;
2855: B->ops->multadd = MatMultAdd_SeqBAIJ_11;
2856: break;
2857: case 15:
2858: B->ops->mult = MatMult_SeqBAIJ_15_ver1;
2859: B->ops->multadd = MatMultAdd_SeqBAIJ_N;
2860: break;
2861: default:
2862: B->ops->mult = MatMult_SeqBAIJ_N;
2863: B->ops->multadd = MatMultAdd_SeqBAIJ_N;
2864: break;
2865: }
2866: }
2867: B->ops->sor = MatSOR_SeqBAIJ;
2868: b->mbs = mbs;
2869: b->nbs = nbs;
2870: if (!skipallocation) {
2871: if (!b->imax) {
2872: PetscMalloc2(mbs,&b->imax,mbs,&b->ilen);
2873: PetscLogObjectMemory((PetscObject)B,2*mbs*sizeof(PetscInt));
2875: b->free_imax_ilen = PETSC_TRUE;
2876: }
2877: /* b->ilen will count nonzeros in each block row so far. */
2878: for (i=0; i<mbs; i++) b->ilen[i] = 0;
2879: if (!nnz) {
2880: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
2881: else if (nz < 0) nz = 1;
2882: nz = PetscMin(nz,nbs);
2883: for (i=0; i<mbs; i++) b->imax[i] = nz;
2884: PetscIntMultError(nz,mbs,&nz);
2885: } else {
2886: PetscInt64 nz64 = 0;
2887: for (i=0; i<mbs; i++) {b->imax[i] = nnz[i]; nz64 += nnz[i];}
2888: PetscIntCast(nz64,&nz);
2889: }
2891: /* allocate the matrix space */
2892: MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);
2893: if (B->structure_only) {
2894: PetscMalloc1(nz,&b->j);
2895: PetscMalloc1(B->rmap->N+1,&b->i);
2896: PetscLogObjectMemory((PetscObject)B,(B->rmap->N+1)*sizeof(PetscInt)+nz*sizeof(PetscInt));
2897: } else {
2898: PetscMalloc3(bs2*nz,&b->a,nz,&b->j,B->rmap->N+1,&b->i);
2899: PetscLogObjectMemory((PetscObject)B,(B->rmap->N+1)*sizeof(PetscInt)+nz*(bs2*sizeof(PetscScalar)+sizeof(PetscInt)));
2900: PetscArrayzero(b->a,nz*bs2);
2901: }
2902: PetscArrayzero(b->j,nz);
2904: if (B->structure_only) {
2905: b->singlemalloc = PETSC_FALSE;
2906: b->free_a = PETSC_FALSE;
2907: } else {
2908: b->singlemalloc = PETSC_TRUE;
2909: b->free_a = PETSC_TRUE;
2910: }
2911: b->free_ij = PETSC_TRUE;
2913: b->i[0] = 0;
2914: for (i=1; i<mbs+1; i++) {
2915: b->i[i] = b->i[i-1] + b->imax[i-1];
2916: }
2918: } else {
2919: b->free_a = PETSC_FALSE;
2920: b->free_ij = PETSC_FALSE;
2921: }
2923: b->bs2 = bs2;
2924: b->mbs = mbs;
2925: b->nz = 0;
2926: b->maxnz = nz;
2927: B->info.nz_unneeded = (PetscReal)b->maxnz*bs2;
2928: B->was_assembled = PETSC_FALSE;
2929: B->assembled = PETSC_FALSE;
2930: if (realalloc) {MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);}
2931: return(0);
2932: }
2934: PetscErrorCode MatSeqBAIJSetPreallocationCSR_SeqBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[])
2935: {
2936: PetscInt i,m,nz,nz_max=0,*nnz;
2937: PetscScalar *values=NULL;
2938: PetscBool roworiented = ((Mat_SeqBAIJ*)B->data)->roworiented;
2942: if (bs < 1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive but it is %D",bs);
2943: PetscLayoutSetBlockSize(B->rmap,bs);
2944: PetscLayoutSetBlockSize(B->cmap,bs);
2945: PetscLayoutSetUp(B->rmap);
2946: PetscLayoutSetUp(B->cmap);
2947: PetscLayoutGetBlockSize(B->rmap,&bs);
2948: m = B->rmap->n/bs;
2950: if (ii[0] != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "ii[0] must be 0 but it is %D",ii[0]);
2951: PetscMalloc1(m+1, &nnz);
2952: for (i=0; i<m; i++) {
2953: nz = ii[i+1]- ii[i];
2954: if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D",i,nz);
2955: nz_max = PetscMax(nz_max, nz);
2956: nnz[i] = nz;
2957: }
2958: MatSeqBAIJSetPreallocation(B,bs,0,nnz);
2959: PetscFree(nnz);
2961: values = (PetscScalar*)V;
2962: if (!values) {
2963: PetscCalloc1(bs*bs*(nz_max+1),&values);
2964: }
2965: for (i=0; i<m; i++) {
2966: PetscInt ncols = ii[i+1] - ii[i];
2967: const PetscInt *icols = jj + ii[i];
2968: if (bs == 1 || !roworiented) {
2969: const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0);
2970: MatSetValuesBlocked_SeqBAIJ(B,1,&i,ncols,icols,svals,INSERT_VALUES);
2971: } else {
2972: PetscInt j;
2973: for (j=0; j<ncols; j++) {
2974: const PetscScalar *svals = values + (V ? (bs*bs*(ii[i]+j)) : 0);
2975: MatSetValuesBlocked_SeqBAIJ(B,1,&i,1,&icols[j],svals,INSERT_VALUES);
2976: }
2977: }
2978: }
2979: if (!V) { PetscFree(values); }
2980: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2981: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2982: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
2983: return(0);
2984: }
2986: /*@C
2987: MatSeqBAIJGetArray - gives access to the array where the data for a MATSEQBAIJ matrix is stored
2989: Not Collective
2991: Input Parameter:
2992: . mat - a MATSEQBAIJ matrix
2994: Output Parameter:
2995: . array - pointer to the data
2997: Level: intermediate
2999: .seealso: MatSeqBAIJRestoreArray(), MatSeqAIJGetArray(), MatSeqAIJRestoreArray()
3000: @*/
3001: PetscErrorCode MatSeqBAIJGetArray(Mat A,PetscScalar **array)
3002: {
3006: PetscUseMethod(A,"MatSeqBAIJGetArray_C",(Mat,PetscScalar**),(A,array));
3007: return(0);
3008: }
3010: /*@C
3011: MatSeqBAIJRestoreArray - returns access to the array where the data for a MATSEQBAIJ matrix is stored obtained by MatSeqBAIJGetArray()
3013: Not Collective
3015: Input Parameters:
3016: + mat - a MATSEQBAIJ matrix
3017: - array - pointer to the data
3019: Level: intermediate
3021: .seealso: MatSeqBAIJGetArray(), MatSeqAIJGetArray(), MatSeqAIJRestoreArray()
3022: @*/
3023: PetscErrorCode MatSeqBAIJRestoreArray(Mat A,PetscScalar **array)
3024: {
3028: PetscUseMethod(A,"MatSeqBAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));
3029: return(0);
3030: }
3032: /*MC
3033: MATSEQBAIJ - MATSEQBAIJ = "seqbaij" - A matrix type to be used for sequential block sparse matrices, based on
3034: block sparse compressed row format.
3036: Options Database Keys:
3037: . -mat_type seqbaij - sets the matrix type to "seqbaij" during a call to MatSetFromOptions()
3039: Level: beginner
3041: Notes:
3042: MatSetOptions(,MAT_STRUCTURE_ONLY,PETSC_TRUE) may be called for this matrix type. In this no
3043: space is allocated for the nonzero entries and any entries passed with MatSetValues() are ignored
3045: .seealso: MatCreateSeqBAIJ()
3046: M*/
3048: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqBSTRM(Mat, MatType,MatReuse,Mat*);
3050: PETSC_EXTERN PetscErrorCode MatCreate_SeqBAIJ(Mat B)
3051: {
3053: PetscMPIInt size;
3054: Mat_SeqBAIJ *b;
3057: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
3058: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Comm must be of size 1");
3060: PetscNewLog(B,&b);
3061: B->data = (void*)b;
3062: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
3064: b->row = NULL;
3065: b->col = NULL;
3066: b->icol = NULL;
3067: b->reallocs = 0;
3068: b->saved_values = NULL;
3070: b->roworiented = PETSC_TRUE;
3071: b->nonew = 0;
3072: b->diag = NULL;
3073: B->spptr = NULL;
3074: B->info.nz_unneeded = (PetscReal)b->maxnz*b->bs2;
3075: b->keepnonzeropattern = PETSC_FALSE;
3077: PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJGetArray_C",MatSeqBAIJGetArray_SeqBAIJ);
3078: PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJRestoreArray_C",MatSeqBAIJRestoreArray_SeqBAIJ);
3079: PetscObjectComposeFunction((PetscObject)B,"MatInvertBlockDiagonal_C",MatInvertBlockDiagonal_SeqBAIJ);
3080: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqBAIJ);
3081: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqBAIJ);
3082: PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetColumnIndices_C",MatSeqBAIJSetColumnIndices_SeqBAIJ);
3083: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_seqaij_C",MatConvert_SeqBAIJ_SeqAIJ);
3084: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_seqsbaij_C",MatConvert_SeqBAIJ_SeqSBAIJ);
3085: PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetPreallocation_C",MatSeqBAIJSetPreallocation_SeqBAIJ);
3086: PetscObjectComposeFunction((PetscObject)B,"MatSeqBAIJSetPreallocationCSR_C",MatSeqBAIJSetPreallocationCSR_SeqBAIJ);
3087: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqBAIJ);
3088: #if defined(PETSC_HAVE_HYPRE)
3089: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_hypre_C",MatConvert_AIJ_HYPRE);
3090: #endif
3091: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqbaij_is_C",MatConvert_XAIJ_IS);
3092: PetscObjectChangeTypeName((PetscObject)B,MATSEQBAIJ);
3093: return(0);
3094: }
3096: PetscErrorCode MatDuplicateNoCreate_SeqBAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace)
3097: {
3098: Mat_SeqBAIJ *c = (Mat_SeqBAIJ*)C->data,*a = (Mat_SeqBAIJ*)A->data;
3100: PetscInt i,mbs = a->mbs,nz = a->nz,bs2 = a->bs2;
3103: if (a->i[mbs] != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupt matrix");
3105: if (cpvalues == MAT_SHARE_NONZERO_PATTERN) {
3106: c->imax = a->imax;
3107: c->ilen = a->ilen;
3108: c->free_imax_ilen = PETSC_FALSE;
3109: } else {
3110: PetscMalloc2(mbs,&c->imax,mbs,&c->ilen);
3111: PetscLogObjectMemory((PetscObject)C,2*mbs*sizeof(PetscInt));
3112: for (i=0; i<mbs; i++) {
3113: c->imax[i] = a->imax[i];
3114: c->ilen[i] = a->ilen[i];
3115: }
3116: c->free_imax_ilen = PETSC_TRUE;
3117: }
3119: /* allocate the matrix space */
3120: if (mallocmatspace) {
3121: if (cpvalues == MAT_SHARE_NONZERO_PATTERN) {
3122: PetscCalloc1(bs2*nz,&c->a);
3123: PetscLogObjectMemory((PetscObject)C,a->i[mbs]*bs2*sizeof(PetscScalar));
3125: c->i = a->i;
3126: c->j = a->j;
3127: c->singlemalloc = PETSC_FALSE;
3128: c->free_a = PETSC_TRUE;
3129: c->free_ij = PETSC_FALSE;
3130: c->parent = A;
3131: C->preallocated = PETSC_TRUE;
3132: C->assembled = PETSC_TRUE;
3134: PetscObjectReference((PetscObject)A);
3135: MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3136: MatSetOption(C,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3137: } else {
3138: PetscMalloc3(bs2*nz,&c->a,nz,&c->j,mbs+1,&c->i);
3139: PetscLogObjectMemory((PetscObject)C,a->i[mbs]*(bs2*sizeof(PetscScalar)+sizeof(PetscInt))+(mbs+1)*sizeof(PetscInt));
3141: c->singlemalloc = PETSC_TRUE;
3142: c->free_a = PETSC_TRUE;
3143: c->free_ij = PETSC_TRUE;
3145: PetscArraycpy(c->i,a->i,mbs+1);
3146: if (mbs > 0) {
3147: PetscArraycpy(c->j,a->j,nz);
3148: if (cpvalues == MAT_COPY_VALUES) {
3149: PetscArraycpy(c->a,a->a,bs2*nz);
3150: } else {
3151: PetscArrayzero(c->a,bs2*nz);
3152: }
3153: }
3154: C->preallocated = PETSC_TRUE;
3155: C->assembled = PETSC_TRUE;
3156: }
3157: }
3159: c->roworiented = a->roworiented;
3160: c->nonew = a->nonew;
3162: PetscLayoutReference(A->rmap,&C->rmap);
3163: PetscLayoutReference(A->cmap,&C->cmap);
3165: c->bs2 = a->bs2;
3166: c->mbs = a->mbs;
3167: c->nbs = a->nbs;
3169: if (a->diag) {
3170: if (cpvalues == MAT_SHARE_NONZERO_PATTERN) {
3171: c->diag = a->diag;
3172: c->free_diag = PETSC_FALSE;
3173: } else {
3174: PetscMalloc1(mbs+1,&c->diag);
3175: PetscLogObjectMemory((PetscObject)C,(mbs+1)*sizeof(PetscInt));
3176: for (i=0; i<mbs; i++) c->diag[i] = a->diag[i];
3177: c->free_diag = PETSC_TRUE;
3178: }
3179: } else c->diag = NULL;
3181: c->nz = a->nz;
3182: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
3183: c->solve_work = NULL;
3184: c->mult_work = NULL;
3185: c->sor_workt = NULL;
3186: c->sor_work = NULL;
3188: c->compressedrow.use = a->compressedrow.use;
3189: c->compressedrow.nrows = a->compressedrow.nrows;
3190: if (a->compressedrow.use) {
3191: i = a->compressedrow.nrows;
3192: PetscMalloc2(i+1,&c->compressedrow.i,i+1,&c->compressedrow.rindex);
3193: PetscLogObjectMemory((PetscObject)C,(2*i+1)*sizeof(PetscInt));
3194: PetscArraycpy(c->compressedrow.i,a->compressedrow.i,i+1);
3195: PetscArraycpy(c->compressedrow.rindex,a->compressedrow.rindex,i);
3196: } else {
3197: c->compressedrow.use = PETSC_FALSE;
3198: c->compressedrow.i = NULL;
3199: c->compressedrow.rindex = NULL;
3200: }
3201: C->nonzerostate = A->nonzerostate;
3203: PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);
3204: return(0);
3205: }
3207: PetscErrorCode MatDuplicate_SeqBAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B)
3208: {
3212: MatCreate(PetscObjectComm((PetscObject)A),B);
3213: MatSetSizes(*B,A->rmap->N,A->cmap->n,A->rmap->N,A->cmap->n);
3214: MatSetType(*B,MATSEQBAIJ);
3215: MatDuplicateNoCreate_SeqBAIJ(*B,A,cpvalues,PETSC_TRUE);
3216: return(0);
3217: }
3219: /* Used for both SeqBAIJ and SeqSBAIJ matrices */
3220: PetscErrorCode MatLoad_SeqBAIJ_Binary(Mat mat,PetscViewer viewer)
3221: {
3222: PetscInt header[4],M,N,nz,bs,m,n,mbs,nbs,rows,cols,sum,i,j,k;
3223: PetscInt *rowidxs,*colidxs;
3224: PetscScalar *matvals;
3228: PetscViewerSetUp(viewer);
3230: /* read matrix header */
3231: PetscViewerBinaryRead(viewer,header,4,NULL,PETSC_INT);
3232: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Not a matrix object in file");
3233: M = header[1]; N = header[2]; nz = header[3];
3234: if (M < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix row size (%D) in file is negative",M);
3235: if (N < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix column size (%D) in file is negative",N);
3236: if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk, cannot load as SeqBAIJ");
3238: /* set block sizes from the viewer's .info file */
3239: MatLoad_Binary_BlockSizes(mat,viewer);
3240: /* set local and global sizes if not set already */
3241: if (mat->rmap->n < 0) mat->rmap->n = M;
3242: if (mat->cmap->n < 0) mat->cmap->n = N;
3243: if (mat->rmap->N < 0) mat->rmap->N = M;
3244: if (mat->cmap->N < 0) mat->cmap->N = N;
3245: PetscLayoutSetUp(mat->rmap);
3246: PetscLayoutSetUp(mat->cmap);
3248: /* check if the matrix sizes are correct */
3249: MatGetSize(mat,&rows,&cols);
3250: if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%D, %D) than the input matrix (%D, %D)",M,N,rows,cols);
3251: MatGetBlockSize(mat,&bs);
3252: MatGetLocalSize(mat,&m,&n);
3253: mbs = m/bs; nbs = n/bs;
3255: /* read in row lengths, column indices and nonzero values */
3256: PetscMalloc1(m+1,&rowidxs);
3257: PetscViewerBinaryRead(viewer,rowidxs+1,m,NULL,PETSC_INT);
3258: rowidxs[0] = 0; for (i=0; i<m; i++) rowidxs[i+1] += rowidxs[i];
3259: sum = rowidxs[m];
3260: if (sum != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent matrix data in file: nonzeros = %D, sum-row-lengths = %D\n",nz,sum);
3262: /* read in column indices and nonzero values */
3263: PetscMalloc2(rowidxs[m],&colidxs,nz,&matvals);
3264: PetscViewerBinaryRead(viewer,colidxs,rowidxs[m],NULL,PETSC_INT);
3265: PetscViewerBinaryRead(viewer,matvals,rowidxs[m],NULL,PETSC_SCALAR);
3267: { /* preallocate matrix storage */
3268: PetscBT bt; /* helper bit set to count nonzeros */
3269: PetscInt *nnz;
3270: PetscBool sbaij;
3272: PetscBTCreate(nbs,&bt);
3273: PetscCalloc1(mbs,&nnz);
3274: PetscObjectTypeCompare((PetscObject)mat,MATSEQSBAIJ,&sbaij);
3275: for (i=0; i<mbs; i++) {
3276: PetscBTMemzero(nbs,bt);
3277: for (k=0; k<bs; k++) {
3278: PetscInt row = bs*i + k;
3279: for (j=rowidxs[row]; j<rowidxs[row+1]; j++) {
3280: PetscInt col = colidxs[j];
3281: if (!sbaij || col >= row)
3282: if (!PetscBTLookupSet(bt,col/bs)) nnz[i]++;
3283: }
3284: }
3285: }
3286: PetscBTDestroy(&bt);
3287: MatSeqBAIJSetPreallocation(mat,bs,0,nnz);
3288: MatSeqSBAIJSetPreallocation(mat,bs,0,nnz);
3289: PetscFree(nnz);
3290: }
3292: /* store matrix values */
3293: for (i=0; i<m; i++) {
3294: PetscInt row = i, s = rowidxs[i], e = rowidxs[i+1];
3295: (*mat->ops->setvalues)(mat,1,&row,e-s,colidxs+s,matvals+s,INSERT_VALUES);
3296: }
3298: PetscFree(rowidxs);
3299: PetscFree2(colidxs,matvals);
3300: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
3301: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
3302: return(0);
3303: }
3305: PetscErrorCode MatLoad_SeqBAIJ(Mat mat,PetscViewer viewer)
3306: {
3308: PetscBool isbinary;
3311: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
3312: if (!isbinary) SETERRQ2(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)mat)->type_name);
3313: MatLoad_SeqBAIJ_Binary(mat,viewer);
3314: return(0);
3315: }
3317: /*@C
3318: MatCreateSeqBAIJ - Creates a sparse matrix in block AIJ (block
3319: compressed row) format. For good matrix assembly performance the
3320: user should preallocate the matrix storage by setting the parameter nz
3321: (or the array nnz). By setting these parameters accurately, performance
3322: during matrix assembly can be increased by more than a factor of 50.
3324: Collective
3326: Input Parameters:
3327: + comm - MPI communicator, set to PETSC_COMM_SELF
3328: . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row
3329: blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs()
3330: . m - number of rows
3331: . n - number of columns
3332: . nz - number of nonzero blocks per block row (same for all rows)
3333: - nnz - array containing the number of nonzero blocks in the various block rows
3334: (possibly different for each block row) or NULL
3336: Output Parameter:
3337: . A - the matrix
3339: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3340: MatXXXXSetPreallocation() paradigm instead of this routine directly.
3341: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
3343: Options Database Keys:
3344: + -mat_no_unroll - uses code that does not unroll the loops in the
3345: block calculations (much slower)
3346: - -mat_block_size - size of the blocks to use
3348: Level: intermediate
3350: Notes:
3351: The number of rows and columns must be divisible by blocksize.
3353: If the nnz parameter is given then the nz parameter is ignored
3355: A nonzero block is any block that as 1 or more nonzeros in it
3357: The block AIJ format is fully compatible with standard Fortran 77
3358: storage. That is, the stored row and column indices can begin at
3359: either one (as in Fortran) or zero. See the users' manual for details.
3361: Specify the preallocated storage with either nz or nnz (not both).
3362: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3363: allocation. See Users-Manual: ch_mat for details.
3364: matrices.
3366: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateBAIJ()
3367: @*/
3368: PetscErrorCode MatCreateSeqBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A)
3369: {
3373: MatCreate(comm,A);
3374: MatSetSizes(*A,m,n,m,n);
3375: MatSetType(*A,MATSEQBAIJ);
3376: MatSeqBAIJSetPreallocation(*A,bs,nz,(PetscInt*)nnz);
3377: return(0);
3378: }
3380: /*@C
3381: MatSeqBAIJSetPreallocation - Sets the block size and expected nonzeros
3382: per row in the matrix. For good matrix assembly performance the
3383: user should preallocate the matrix storage by setting the parameter nz
3384: (or the array nnz). By setting these parameters accurately, performance
3385: during matrix assembly can be increased by more than a factor of 50.
3387: Collective
3389: Input Parameters:
3390: + B - the matrix
3391: . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row
3392: blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs()
3393: . nz - number of block nonzeros per block row (same for all rows)
3394: - nnz - array containing the number of block nonzeros in the various block rows
3395: (possibly different for each block row) or NULL
3397: Options Database Keys:
3398: + -mat_no_unroll - uses code that does not unroll the loops in the
3399: block calculations (much slower)
3400: - -mat_block_size - size of the blocks to use
3402: Level: intermediate
3404: Notes:
3405: If the nnz parameter is given then the nz parameter is ignored
3407: You can call MatGetInfo() to get information on how effective the preallocation was;
3408: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3409: You can also run with the option -info and look for messages with the string
3410: malloc in them to see if additional memory allocation was needed.
3412: The block AIJ format is fully compatible with standard Fortran 77
3413: storage. That is, the stored row and column indices can begin at
3414: either one (as in Fortran) or zero. See the users' manual for details.
3416: Specify the preallocated storage with either nz or nnz (not both).
3417: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3418: allocation. See Users-Manual: ch_mat for details.
3420: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateBAIJ(), MatGetInfo()
3421: @*/
3422: PetscErrorCode MatSeqBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt nz,const PetscInt nnz[])
3423: {
3430: PetscTryMethod(B,"MatSeqBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[]),(B,bs,nz,nnz));
3431: return(0);
3432: }
3434: /*@C
3435: MatSeqBAIJSetPreallocationCSR - Creates a sparse parallel matrix in BAIJ format using the given nonzero structure and (optional) numerical values
3437: Collective
3439: Input Parameters:
3440: + B - the matrix
3441: . i - the indices into j for the start of each local row (starts with zero)
3442: . j - the column indices for each local row (starts with zero) these must be sorted for each row
3443: - v - optional values in the matrix
3445: Level: advanced
3447: Notes:
3448: The order of the entries in values is specified by the MatOption MAT_ROW_ORIENTED. For example, C programs
3449: may want to use the default MAT_ROW_ORIENTED=PETSC_TRUE and use an array v[nnz][bs][bs] where the second index is
3450: over rows within a block and the last index is over columns within a block row. Fortran programs will likely set
3451: MAT_ROW_ORIENTED=PETSC_FALSE and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a
3452: block column and the second index is over columns within a block.
3454: Though this routine has Preallocation() in the name it also sets the exact nonzero locations of the matrix entries and usually the numerical values as well
3456: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatSeqBAIJSetPreallocation(), MATSEQBAIJ
3457: @*/
3458: PetscErrorCode MatSeqBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
3459: {
3466: PetscTryMethod(B,"MatSeqBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));
3467: return(0);
3468: }
3471: /*@
3472: MatCreateSeqBAIJWithArrays - Creates an sequential BAIJ matrix using matrix elements provided by the user.
3474: Collective
3476: Input Parameters:
3477: + comm - must be an MPI communicator of size 1
3478: . bs - size of block
3479: . m - number of rows
3480: . n - number of columns
3481: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row block row of the matrix
3482: . j - column indices
3483: - a - matrix values
3485: Output Parameter:
3486: . mat - the matrix
3488: Level: advanced
3490: Notes:
3491: The i, j, and a arrays are not copied by this routine, the user must free these arrays
3492: once the matrix is destroyed
3494: You cannot set new nonzero locations into this matrix, that will generate an error.
3496: The i and j indices are 0 based
3498: When block size is greater than 1 the matrix values must be stored using the BAIJ storage format (see the BAIJ code to determine this).
3500: The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is
3501: the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first
3502: block, followed by the second column of the first block etc etc. That is, the blocks are contiguous in memory
3503: with column-major ordering within blocks.
3505: .seealso: MatCreate(), MatCreateBAIJ(), MatCreateSeqBAIJ()
3507: @*/
3508: PetscErrorCode MatCreateSeqBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat)
3509: {
3511: PetscInt ii;
3512: Mat_SeqBAIJ *baij;
3515: if (bs != 1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"block size %D > 1 is not supported yet",bs);
3516: if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
3518: MatCreate(comm,mat);
3519: MatSetSizes(*mat,m,n,m,n);
3520: MatSetType(*mat,MATSEQBAIJ);
3521: MatSeqBAIJSetPreallocation(*mat,bs,MAT_SKIP_ALLOCATION,NULL);
3522: baij = (Mat_SeqBAIJ*)(*mat)->data;
3523: PetscMalloc2(m,&baij->imax,m,&baij->ilen);
3524: PetscLogObjectMemory((PetscObject)*mat,2*m*sizeof(PetscInt));
3526: baij->i = i;
3527: baij->j = j;
3528: baij->a = a;
3530: baij->singlemalloc = PETSC_FALSE;
3531: baij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
3532: baij->free_a = PETSC_FALSE;
3533: baij->free_ij = PETSC_FALSE;
3535: for (ii=0; ii<m; ii++) {
3536: baij->ilen[ii] = baij->imax[ii] = i[ii+1] - i[ii];
3537: if (PetscUnlikelyDebug(i[ii+1] - i[ii] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row length in i (row indices) row = %d length = %d",ii,i[ii+1] - i[ii]);
3538: }
3539: if (PetscDefined(USE_DEBUG)) {
3540: for (ii=0; ii<baij->i[m]; ii++) {
3541: if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %d index = %d",ii,j[ii]);
3542: if (j[ii] > n - 1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column index to large at location = %d index = %d",ii,j[ii]);
3543: }
3544: }
3546: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
3547: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
3548: return(0);
3549: }
3551: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqBAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
3552: {
3554: PetscMPIInt size;
3557: MPI_Comm_size(comm,&size);
3558: if (size == 1 && scall == MAT_REUSE_MATRIX) {
3559: MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);
3560: } else {
3561: MatCreateMPIMatConcatenateSeqMat_MPIBAIJ(comm,inmat,n,scall,outmat);
3562: }
3563: return(0);
3564: }