Actual source code: vecimpl.h
1: #pragma once
3: /*
4: This private file should not be included in users' code. Defines the fields shared by all
5: vector implementations.
6: */
8: #include <petscvec.h>
9: #include <petsc/private/petscimpl.h>
11: PETSC_INTERN PetscBool VecRegisterAllCalled;
12: PETSC_EXTERN MPI_Op MPIU_MAXLOC;
13: PETSC_EXTERN MPI_Op MPIU_MINLOC;
14: PETSC_INTERN MPI_Op VecStrideMax_Op;
15: PETSC_INTERN MPI_Op VecStrideMin_Op;
16: PETSC_INTERN void MPIAPI VecStrideMax_Local(void *, void *, PetscMPIInt *, MPI_Datatype *);
17: PETSC_INTERN void MPIAPI VecStrideMin_Local(void *, void *, PetscMPIInt *, MPI_Datatype *);
19: /* ----------------------------------------------------------------------------*/
21: typedef struct _VecOps *VecOps;
22: struct _VecOps {
23: PetscErrorCode (*duplicate)(Vec, Vec *); /* get single vector */
24: PetscErrorCode (*duplicatevecs)(Vec, PetscInt, Vec **); /* get array of vectors */
25: PetscErrorCode (*destroyvecs)(PetscInt, Vec[]); /* free array of vectors */
26: PetscErrorCode (*dot)(Vec, Vec, PetscScalar *); /* z = x^H * y */
27: PetscErrorCode (*mdot)(Vec, PetscInt, const Vec[], PetscScalar *); /* z[j] = x dot y[j] */
28: PetscErrorCode (*norm)(Vec, NormType, PetscReal *); /* z = sqrt(x^H * x) */
29: PetscErrorCode (*tdot)(Vec, Vec, PetscScalar *); /* x'*y */
30: PetscErrorCode (*mtdot)(Vec, PetscInt, const Vec[], PetscScalar *); /* z[j] = x dot y[j] */
31: PetscErrorCode (*scale)(Vec, PetscScalar); /* x = alpha * x */
32: PetscErrorCode (*copy)(Vec, Vec); /* y = x */
33: PetscErrorCode (*set)(Vec, PetscScalar); /* y = alpha */
34: PetscErrorCode (*swap)(Vec, Vec); /* exchange x and y */
35: PetscErrorCode (*axpy)(Vec, PetscScalar, Vec); /* y = y + alpha * x */
36: PetscErrorCode (*axpby)(Vec, PetscScalar, PetscScalar, Vec); /* y = alpha * x + beta * y*/
37: PetscErrorCode (*maxpy)(Vec, PetscInt, const PetscScalar *, Vec *); /* y = y + alpha[j] x[j] */
38: PetscErrorCode (*aypx)(Vec, PetscScalar, Vec); /* y = x + alpha * y */
39: PetscErrorCode (*waxpy)(Vec, PetscScalar, Vec, Vec); /* w = y + alpha * x */
40: PetscErrorCode (*axpbypcz)(Vec, PetscScalar, PetscScalar, PetscScalar, Vec, Vec); /* z = alpha * x + beta *y + gamma *z*/
41: PetscErrorCode (*pointwisemult)(Vec, Vec, Vec); /* w = x .* y */
42: PetscErrorCode (*pointwisedivide)(Vec, Vec, Vec); /* w = x ./ y */
43: PetscErrorCode (*setvalues)(Vec, PetscInt, const PetscInt[], const PetscScalar[], InsertMode);
44: PetscErrorCode (*assemblybegin)(Vec); /* start global assembly */
45: PetscErrorCode (*assemblyend)(Vec); /* end global assembly */
46: PetscErrorCode (*getarray)(Vec, PetscScalar **); /* get data array */
47: PetscErrorCode (*getsize)(Vec, PetscInt *);
48: PetscErrorCode (*getlocalsize)(Vec, PetscInt *);
49: PetscErrorCode (*restorearray)(Vec, PetscScalar **); /* restore data array */
50: PetscErrorCode (*max)(Vec, PetscInt *, PetscReal *); /* z = max(x); idx=index of max(x) */
51: PetscErrorCode (*min)(Vec, PetscInt *, PetscReal *); /* z = min(x); idx=index of min(x) */
52: PetscErrorCode (*setrandom)(Vec, PetscRandom); /* set y[j] = random numbers */
53: PetscErrorCode (*setoption)(Vec, VecOption, PetscBool);
54: PetscErrorCode (*setvaluesblocked)(Vec, PetscInt, const PetscInt[], const PetscScalar[], InsertMode);
55: PetscErrorCode (*destroy)(Vec);
56: PetscErrorCode (*view)(Vec, PetscViewer);
57: PetscErrorCode (*placearray)(Vec, const PetscScalar *); /* place data array */
58: PetscErrorCode (*replacearray)(Vec, const PetscScalar *); /* replace data array */
59: PetscErrorCode (*dot_local)(Vec, Vec, PetscScalar *);
60: PetscErrorCode (*tdot_local)(Vec, Vec, PetscScalar *);
61: PetscErrorCode (*norm_local)(Vec, NormType, PetscReal *);
62: PetscErrorCode (*mdot_local)(Vec, PetscInt, const Vec[], PetscScalar *);
63: PetscErrorCode (*mtdot_local)(Vec, PetscInt, const Vec[], PetscScalar *);
64: PetscErrorCode (*load)(Vec, PetscViewer);
65: PetscErrorCode (*reciprocal)(Vec);
66: PetscErrorCode (*conjugate)(Vec);
67: PetscErrorCode (*setlocaltoglobalmapping)(Vec, ISLocalToGlobalMapping);
68: PetscErrorCode (*getlocaltoglobalmapping)(Vec, ISLocalToGlobalMapping *);
69: PetscErrorCode (*resetarray)(Vec); /* vector points to its original array, i.e. undoes any VecPlaceArray() */
70: PetscErrorCode (*setfromoptions)(Vec, PetscOptionItems);
71: PetscErrorCode (*maxpointwisedivide)(Vec, Vec, PetscReal *); /* m = max abs(x ./ y) */
72: PetscErrorCode (*pointwisemax)(Vec, Vec, Vec);
73: PetscErrorCode (*pointwisemaxabs)(Vec, Vec, Vec);
74: PetscErrorCode (*pointwisemin)(Vec, Vec, Vec);
75: PetscErrorCode (*getvalues)(Vec, PetscInt, const PetscInt[], PetscScalar[]);
76: PetscErrorCode (*sqrt)(Vec);
77: PetscErrorCode (*abs)(Vec);
78: PetscErrorCode (*exp)(Vec);
79: PetscErrorCode (*log)(Vec);
80: PetscErrorCode (*shift)(Vec, PetscScalar);
81: PetscErrorCode (*create)(Vec);
82: PetscErrorCode (*stridegather)(Vec, PetscInt, Vec, InsertMode);
83: PetscErrorCode (*stridescatter)(Vec, PetscInt, Vec, InsertMode);
84: PetscErrorCode (*dotnorm2)(Vec, Vec, PetscScalar *, PetscScalar *);
85: PetscErrorCode (*getsubvector)(Vec, IS, Vec *);
86: PetscErrorCode (*restoresubvector)(Vec, IS, Vec *);
87: PetscErrorCode (*getarrayread)(Vec, const PetscScalar **);
88: PetscErrorCode (*restorearrayread)(Vec, const PetscScalar **);
89: PetscErrorCode (*stridesubsetgather)(Vec, PetscInt, const PetscInt[], const PetscInt[], Vec, InsertMode);
90: PetscErrorCode (*stridesubsetscatter)(Vec, PetscInt, const PetscInt[], const PetscInt[], Vec, InsertMode);
91: PetscErrorCode (*viewnative)(Vec, PetscViewer);
92: PetscErrorCode (*loadnative)(Vec, PetscViewer);
93: PetscErrorCode (*createlocalvector)(Vec, Vec *);
94: PetscErrorCode (*getlocalvector)(Vec, Vec);
95: PetscErrorCode (*restorelocalvector)(Vec, Vec);
96: PetscErrorCode (*getlocalvectorread)(Vec, Vec);
97: PetscErrorCode (*restorelocalvectorread)(Vec, Vec);
98: PetscErrorCode (*bindtocpu)(Vec, PetscBool);
99: PetscErrorCode (*getarraywrite)(Vec, PetscScalar **);
100: PetscErrorCode (*restorearraywrite)(Vec, PetscScalar **);
101: PetscErrorCode (*getarrayandmemtype)(Vec, PetscScalar **, PetscMemType *);
102: PetscErrorCode (*restorearrayandmemtype)(Vec, PetscScalar **);
103: PetscErrorCode (*getarrayreadandmemtype)(Vec, const PetscScalar **, PetscMemType *);
104: PetscErrorCode (*restorearrayreadandmemtype)(Vec, const PetscScalar **);
105: PetscErrorCode (*getarraywriteandmemtype)(Vec, PetscScalar **, PetscMemType *);
106: PetscErrorCode (*restorearraywriteandmemtype)(Vec, PetscScalar **, PetscMemType *);
107: PetscErrorCode (*concatenate)(PetscInt, const Vec[], Vec *, IS *[]);
108: PetscErrorCode (*sum)(Vec, PetscScalar *);
109: PetscErrorCode (*setpreallocationcoo)(Vec, PetscCount, const PetscInt[]);
110: PetscErrorCode (*setvaluescoo)(Vec, const PetscScalar[], InsertMode);
111: PetscErrorCode (*errorwnorm)(Vec, Vec, Vec, NormType, PetscReal, Vec, PetscReal, Vec, PetscReal, PetscReal *, PetscInt *, PetscReal *, PetscInt *, PetscReal *, PetscInt *);
112: PetscErrorCode (*maxpby)(Vec, PetscInt, const PetscScalar *, PetscScalar, Vec *); /* y = beta y + alpha[j] x[j] */
113: PetscErrorCode (*setstdbasis)(Vec, PetscInt);
114: };
116: #if defined(offsetof) && (defined(__cplusplus) || (PETSC_C_VERSION >= 23))
117: static_assert(offsetof(struct _VecOps, duplicate) == sizeof(PetscErrorCodeFn *) * VECOP_DUPLICATE, "");
118: static_assert(offsetof(struct _VecOps, set) == sizeof(PetscErrorCodeFn *) * VECOP_SET, "");
119: static_assert(offsetof(struct _VecOps, view) == sizeof(PetscErrorCodeFn *) * VECOP_VIEW, "");
120: static_assert(offsetof(struct _VecOps, load) == sizeof(PetscErrorCodeFn *) * VECOP_LOAD, "");
121: static_assert(offsetof(struct _VecOps, viewnative) == sizeof(PetscErrorCodeFn *) * VECOP_VIEWNATIVE, "");
122: static_assert(offsetof(struct _VecOps, loadnative) == sizeof(PetscErrorCodeFn *) * VECOP_LOADNATIVE, "");
123: #endif
125: /*
126: The stash is used to temporarily store inserted vec values that
127: belong to another processor. During the assembly phase the stashed
128: values are moved to the correct processor and
129: */
131: typedef struct {
132: PetscInt nmax; /* maximum stash size */
133: PetscInt umax; /* max stash size user wants */
134: PetscInt oldnmax; /* the nmax value used previously */
135: PetscInt n; /* stash size */
136: PetscInt bs; /* block size of the stash */
137: PetscInt reallocs; /* preserve the no of mallocs invoked */
138: PetscInt *idx; /* global row numbers in stash */
139: PetscScalar *array; /* array to hold stashed values */
140: /* The following variables are used for communication */
141: MPI_Comm comm;
142: PetscMPIInt size, rank;
143: PetscMPIInt tag1, tag2;
144: MPI_Request *send_waits; /* array of send requests */
145: MPI_Request *recv_waits; /* array of receive requests */
146: MPI_Status *send_status; /* array of send status */
147: PetscMPIInt nsends, nrecvs; /* numbers of sends and receives */
148: PetscScalar *svalues, *rvalues; /* sending and receiving data */
149: PetscInt *sindices, *rindices;
150: PetscInt rmax; /* maximum message length */
151: PetscInt *nprocs; /* tmp data used both during scatterbegin and end */
152: PetscMPIInt nprocessed; /* number of messages already processed */
153: PetscBool donotstash;
154: PetscBool ignorenegidx; /* ignore negative indices passed into VecSetValues/VetGetValues */
155: InsertMode insertmode;
156: PetscInt *bowners;
157: } VecStash;
159: struct _p_Vec {
160: PETSCHEADER(struct _VecOps);
161: PetscLayout map;
162: void *data; /* implementation-specific data */
163: PetscBool array_gotten;
164: VecStash stash, bstash; /* used for storing off-proc values during assembly */
165: PetscBool petscnative; /* means the ->data starts with VECHEADER and can use VecGetArrayFast()*/
166: PetscInt lock; /* lock state. vector can be free (=0), locked for read (>0) or locked for write(<0) */
167: #if PetscDefined(USE_DEBUG)
168: PetscStack lockstack; /* the file,func,line of where locks are added */
169: #endif
170: PetscOffloadMask offloadmask; /* a mask which indicates where the valid vector data is (GPU, CPU or both) */
171: #if PetscDefined(HAVE_DEVICE)
172: void *spptr; /* this is the special pointer to the array on the GPU */
173: PetscBool boundtocpu;
174: PetscBool bindingpropagates;
175: size_t minimum_bytes_pinned_memory; /* minimum data size in bytes for which pinned memory will be allocated */
176: PetscBool pinned_memory; /* PETSC_TRUE if the current host allocation has been made from pinned memory. */
177: #endif
178: char *defaultrandtype;
179: };
181: PETSC_EXTERN PetscLogEvent VEC_SetRandom;
182: PETSC_EXTERN PetscLogEvent VEC_View;
183: PETSC_EXTERN PetscLogEvent VEC_Max;
184: PETSC_EXTERN PetscLogEvent VEC_Min;
185: PETSC_EXTERN PetscLogEvent VEC_Dot;
186: PETSC_EXTERN PetscLogEvent VEC_MDot;
187: PETSC_EXTERN PetscLogEvent VEC_TDot;
188: PETSC_EXTERN PetscLogEvent VEC_MTDot;
189: PETSC_EXTERN PetscLogEvent VEC_Norm;
190: PETSC_EXTERN PetscLogEvent VEC_Normalize;
191: PETSC_EXTERN PetscLogEvent VEC_Scale;
192: PETSC_EXTERN PetscLogEvent VEC_Shift;
193: PETSC_EXTERN PetscLogEvent VEC_Copy;
194: PETSC_EXTERN PetscLogEvent VEC_Set;
195: PETSC_EXTERN PetscLogEvent VEC_AXPY;
196: PETSC_EXTERN PetscLogEvent VEC_AYPX;
197: PETSC_EXTERN PetscLogEvent VEC_WAXPY;
198: PETSC_EXTERN PetscLogEvent VEC_MAXPY;
199: PETSC_EXTERN PetscLogEvent VEC_AssemblyEnd;
200: PETSC_EXTERN PetscLogEvent VEC_PointwiseMult;
201: PETSC_EXTERN PetscLogEvent VEC_PointwiseDivide;
202: PETSC_EXTERN PetscLogEvent VEC_Reciprocal;
203: PETSC_EXTERN PetscLogEvent VEC_SetValues;
204: PETSC_EXTERN PetscLogEvent VEC_SetPreallocateCOO;
205: PETSC_EXTERN PetscLogEvent VEC_SetValuesCOO;
206: PETSC_EXTERN PetscLogEvent VEC_Load;
207: PETSC_EXTERN PetscLogEvent VEC_ScatterBegin;
208: PETSC_EXTERN PetscLogEvent VEC_ScatterEnd;
209: PETSC_EXTERN PetscLogEvent VEC_ReduceArithmetic;
210: PETSC_EXTERN PetscLogEvent VEC_ReduceCommunication;
211: PETSC_EXTERN PetscLogEvent VEC_ReduceBegin;
212: PETSC_EXTERN PetscLogEvent VEC_ReduceEnd;
213: PETSC_EXTERN PetscLogEvent VEC_Swap;
214: PETSC_EXTERN PetscLogEvent VEC_AssemblyBegin;
215: PETSC_EXTERN PetscLogEvent VEC_DotNorm2;
216: PETSC_EXTERN PetscLogEvent VEC_AXPBYPCZ;
217: PETSC_EXTERN PetscLogEvent VEC_Ops;
218: PETSC_EXTERN PetscLogEvent VEC_ViennaCLCopyToGPU;
219: PETSC_EXTERN PetscLogEvent VEC_ViennaCLCopyFromGPU;
220: PETSC_EXTERN PetscLogEvent VEC_CUDACopyToGPU;
221: PETSC_EXTERN PetscLogEvent VEC_CUDACopyFromGPU;
222: PETSC_EXTERN PetscLogEvent VEC_HIPCopyToGPU;
223: PETSC_EXTERN PetscLogEvent VEC_HIPCopyFromGPU;
225: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode VecView_Seq(Vec, PetscViewer);
226: #if PetscDefined(HAVE_VIENNACL)
227: PETSC_EXTERN PetscErrorCode VecViennaCLAllocateCheckHost(Vec v);
228: PETSC_EXTERN PetscErrorCode VecViennaCLCopyFromGPU(Vec v);
229: #endif
231: /*
232: Common header shared by array based vectors,
233: currently Vec_Seq and Vec_MPI
234: */
235: #define VECHEADER \
236: PetscScalar *array; \
237: PetscScalar *array_allocated; /* if the array was allocated by PETSc this is its pointer */ \
238: PetscScalar *unplacedarray; /* if one called VecPlaceArray(), this is where it stashed the original */
240: /* Get Root type of vector. e.g. VECSEQ -> VECSTANDARD, VECMPICUDA -> VECCUDA */
241: PETSC_EXTERN PetscErrorCode VecGetRootType_Private(Vec, VecType *);
243: /* Default obtain and release vectors; can be used by any implementation */
244: PETSC_INTERN PetscErrorCode VecDuplicateVecs_Default(Vec, PetscInt, Vec *[]);
245: PETSC_INTERN PetscErrorCode VecDestroyVecs_Default(PetscInt, Vec[]);
246: PETSC_INTERN PetscErrorCode VecView_Binary(Vec, PetscViewer);
248: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode VecLoad_Default(Vec, PetscViewer);
250: PETSC_INTERN PetscInt NormIds[4]; /* map from NormType to IDs used to cache/retrieve values of norms, 1_AND_2 is excluded */
252: PETSC_INTERN PetscErrorCode VecStashCreate_Private(MPI_Comm, PetscInt, VecStash *);
253: PETSC_INTERN PetscErrorCode VecStashDestroy_Private(VecStash *);
254: PETSC_INTERN PetscErrorCode VecStashScatterEnd_Private(VecStash *);
255: PETSC_INTERN PetscErrorCode VecStashSetInitialSize_Private(VecStash *, PetscInt);
256: PETSC_INTERN PetscErrorCode VecStashGetInfo_Private(VecStash *, PetscInt *, PetscInt *);
257: PETSC_INTERN PetscErrorCode VecStashScatterBegin_Private(VecStash *, const PetscInt *);
258: PETSC_INTERN PetscErrorCode VecStashScatterGetMesg_Private(VecStash *, PetscMPIInt *, PetscInt **, PetscScalar **, PetscInt *);
259: PETSC_INTERN PetscErrorCode VecStashSortCompress_Private(VecStash *);
260: PETSC_INTERN PetscErrorCode VecStashGetOwnerList_Private(VecStash *, PetscLayout, PetscMPIInt *, PetscMPIInt **);
262: PETSC_INTERN PetscErrorCode VecStashExpand_Private(VecStash *, PetscInt);
264: /*
265: VecStashValue_Private - inserts a single value into the stash.
267: Input Parameters:
268: stash - the stash
269: idx - the global of the inserted value
270: values - the value inserted
271: */
272: static inline PetscErrorCode VecStashValue_Private(VecStash *stash, PetscInt row, PetscScalar value)
273: {
274: /* Check and see if we have sufficient memory */
275: PetscFunctionBegin;
276: if ((stash->n + 1) > stash->nmax) PetscCall(VecStashExpand_Private(stash, 1));
277: stash->idx[stash->n] = row;
278: stash->array[stash->n] = value;
279: stash->n++;
280: PetscFunctionReturn(PETSC_SUCCESS);
281: }
283: /*
284: VecStashValuesBlocked_Private - inserts 1 block of values into the stash.
286: Input Parameters:
287: stash - the stash
288: idx - the global block index
289: values - the values inserted
290: */
291: static inline PetscErrorCode VecStashValuesBlocked_Private(VecStash *stash, PetscInt row, PetscScalar *values)
292: {
293: PetscInt stash_bs = stash->bs;
294: PetscScalar *array;
296: PetscFunctionBegin;
297: if ((stash->n + 1) > stash->nmax) PetscCall(VecStashExpand_Private(stash, 1));
298: array = stash->array + stash_bs * stash->n;
299: stash->idx[stash->n] = row;
300: if (values) PetscCall(PetscArraycpy(array, values, stash_bs));
301: else PetscCall(PetscArrayzero(array, stash_bs));
302: stash->n++;
303: PetscFunctionReturn(PETSC_SUCCESS);
304: }
306: PETSC_INTERN PetscErrorCode VecStrideGather_Default(Vec, PetscInt, Vec, InsertMode);
307: PETSC_INTERN PetscErrorCode VecStrideScatter_Default(Vec, PetscInt, Vec, InsertMode);
308: PETSC_INTERN PetscErrorCode VecStrideSubSetGather_Default(Vec, PetscInt, const PetscInt[], const PetscInt[], Vec, InsertMode);
309: PETSC_INTERN PetscErrorCode VecStrideSubSetScatter_Default(Vec, PetscInt, const PetscInt[], const PetscInt[], Vec, InsertMode);
311: PETSC_INTERN PetscErrorCode VecReciprocal_Default(Vec);
312: #if PetscDefined(HAVE_MATLAB)
313: PETSC_EXTERN PetscErrorCode VecMatlabEnginePut_Default(PetscObject, void *);
314: PETSC_EXTERN PetscErrorCode VecMatlabEngineGet_Default(PetscObject, void *);
315: #endif
317: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode PetscSectionGetField_Internal(PetscSection, PetscSection, Vec, PetscInt, PetscInt, PetscInt, IS *, Vec *);
318: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode PetscSectionRestoreField_Internal(PetscSection, PetscSection, Vec, PetscInt, PetscInt, PetscInt, IS *, Vec *);
320: #define VecCheckSameLocalSize(x, ar1, y, ar2) \
321: do { \
322: PetscCheck((x)->map->n == (y)->map->n, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Incompatible vector local lengths parameter # %d local size %" PetscInt_FMT " != parameter # %d local size %" PetscInt_FMT, ar1, (x)->map->n, ar2, (y)->map->n); \
323: } while (0)
325: #define VecCheckSameSize(x, ar1, y, ar2) \
326: do { \
327: PetscCheck((x)->map->N == (y)->map->N, PetscObjectComm((PetscObject)(x)), PETSC_ERR_ARG_INCOMP, "Incompatible vector global lengths parameter # %d global size %" PetscInt_FMT " != parameter # %d global size %" PetscInt_FMT, ar1, (x)->map->N, ar2, \
328: (y)->map->N); \
329: VecCheckSameLocalSize(x, ar1, y, ar2); \
330: } while (0)
332: #define VecCheckLocalSize(x, ar1, n) \
333: do { \
334: PetscCheck((x)->map->n == (n), PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Incorrect vector local size: parameter # %d local size %" PetscInt_FMT " != %" PetscInt_FMT, ar1, (x)->map->n, n); \
335: } while (0)
337: #define VecCheckSize(x, ar1, n, N) \
338: do { \
339: PetscCheck((x)->map->N == (N), PetscObjectComm((PetscObject)(x)), PETSC_ERR_ARG_INCOMP, "Incorrect vector global size: parameter # %d global size %" PetscInt_FMT " != %" PetscInt_FMT, ar1, (x)->map->N, N); \
340: VecCheckLocalSize(x, ar1, n); \
341: } while (0)
343: typedef struct _VecTaggerOps *VecTaggerOps;
344: struct _VecTaggerOps {
345: PetscErrorCode (*create)(VecTagger);
346: PetscErrorCode (*destroy)(VecTagger);
347: PetscErrorCode (*setfromoptions)(VecTagger, PetscOptionItems);
348: PetscErrorCode (*setup)(VecTagger);
349: PetscErrorCode (*view)(VecTagger, PetscViewer);
350: PetscErrorCode (*computeboxes)(VecTagger, Vec, PetscInt *, VecTaggerBox **, PetscBool *);
351: PetscErrorCode (*computeis)(VecTagger, Vec, IS *, PetscBool *);
352: };
353: struct _p_VecTagger {
354: PETSCHEADER(struct _VecTaggerOps);
355: void *data;
356: PetscInt blocksize;
357: PetscBool invert;
358: PetscBool setupcalled;
359: };
361: PETSC_INTERN PetscBool VecTaggerRegisterAllCalled;
362: PETSC_INTERN PetscErrorCode VecTaggerComputeIS_FromBoxes(VecTagger, Vec, IS *, PetscBool *);
363: PETSC_INTERN PetscMPIInt Petsc_Reduction_keyval;
365: PETSC_INTERN PetscInt VecGetSubVectorSavedStateId;
366: PETSC_INTERN PetscErrorCode VecGetSubVectorContiguityAndBS_Private(Vec, IS, PetscBool *, PetscInt *, PetscInt *);
367: PETSC_INTERN PetscErrorCode VecGetSubVectorThroughVecScatter_Private(Vec, IS, PetscInt, Vec *);
369: #if PetscDefined(HAVE_CUDA)
370: PETSC_INTERN PetscErrorCode VecCreate_CUDA(Vec);
371: PETSC_INTERN PetscErrorCode VecCreate_SeqCUDA(Vec);
372: PETSC_INTERN PetscErrorCode VecCreate_MPICUDA(Vec);
373: PETSC_INTERN PetscErrorCode VecCUDAGetArrays_Private(Vec, const PetscScalar **, const PetscScalar **, PetscOffloadMask *);
374: PETSC_INTERN PetscErrorCode VecConvert_Seq_SeqCUDA_inplace(Vec);
375: PETSC_INTERN PetscErrorCode VecConvert_MPI_MPICUDA_inplace(Vec);
376: #endif
378: #if PetscDefined(HAVE_HIP)
379: PETSC_INTERN PetscErrorCode VecCreate_HIP(Vec);
380: PETSC_INTERN PetscErrorCode VecCreate_SeqHIP(Vec);
381: PETSC_INTERN PetscErrorCode VecCreate_MPIHIP(Vec);
382: PETSC_INTERN PetscErrorCode VecHIPGetArrays_Private(Vec, const PetscScalar **, const PetscScalar **, PetscOffloadMask *);
383: PETSC_INTERN PetscErrorCode VecConvert_Seq_SeqHIP_inplace(Vec);
384: PETSC_INTERN PetscErrorCode VecConvert_MPI_MPIHIP_inplace(Vec);
385: #endif
387: #if PetscDefined(HAVE_KOKKOS)
388: PETSC_INTERN PetscErrorCode VecCreateMPIKokkosWithArrays_Private(MPI_Comm, PetscInt, PetscInt, PetscInt, const PetscScalar *, const PetscScalar *, Vec *);
389: PETSC_INTERN PetscErrorCode VecConvert_Seq_SeqKokkos_inplace(Vec, PetscScalar *);
390: PETSC_INTERN PetscErrorCode VecConvert_MPI_MPIKokkos_inplace(Vec);
391: #endif
393: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode VecCreateWithLayout_Private(PetscLayout, Vec *);
394: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode VecCreateSeqWithLayoutAndArray_Private(PetscLayout, const PetscScalar *, Vec *);
395: PETSC_SINGLE_LIBRARY_INTERN PetscErrorCode VecCreateMPIWithLayoutAndArray_Private(PetscLayout, const PetscScalar *, Vec *);
397: /* std::upper_bound(): Given a sorted array, return index of the first element in range [first,last) whose value
398: is greater than value, or last if there is no such element.
399: */
400: static inline PetscErrorCode PetscSortedIntUpperBound(const PetscInt *array, PetscCount first, PetscCount last, PetscInt value, PetscCount *upper)
401: {
402: PetscCount count = last - first;
404: PetscFunctionBegin;
405: while (count > 0) {
406: const PetscCount step = count / 2;
407: PetscCount it = first + step;
409: if (value < array[it]) {
410: count = step;
411: } else {
412: first = it + 1;
413: count -= step + 1;
414: }
415: }
416: *upper = first;
417: PetscFunctionReturn(PETSC_SUCCESS);
418: }
420: #define VEC_ASYNC_FN_NAME(Base) "Vec" Base "Async_Private_C"
421: #define VecAsyncFnName(Base) VEC_##Base##_ASYNC_FN_NAME
423: #define VEC_Abs_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Abs")
424: #define VEC_AXPBY_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("AXPBY")
425: #define VEC_AXPBYPCZ_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("AXPBYPCZ")
426: #define VEC_AXPY_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("AXPY")
427: #define VEC_AYPX_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("AYPX")
428: #define VEC_Conjugate_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Conjugate")
429: #define VEC_Copy_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Copy")
430: #define VEC_Exp_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Exp")
431: #define VEC_Log_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Log")
432: #define VEC_MAXPY_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("MAXPY")
433: #define VEC_PointwiseDivide_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("PointwiseDivide")
434: #define VEC_PointwiseMax_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("PointwiseMax")
435: #define VEC_PointwiseMaxAbs_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("PointwiseMaxAbs")
436: #define VEC_PointwiseMin_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("PointwiseMin")
437: #define VEC_PointwiseMult_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("PointwiseMult")
438: #define VEC_PointwiseSign_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("PointwiseSign")
439: #define VEC_Reciprocal_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Reciprocal")
440: #define VEC_Scale_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Scale")
441: #define VEC_Set_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Set")
442: #define VEC_Shift_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Shift")
443: #define VEC_SqrtAbs_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("SqrtAbs")
444: #define VEC_Swap_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("Swap")
445: #define VEC_WAXPY_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("WAXPY")
446: #define VEC_SetStdBasis_ASYNC_FN_NAME VEC_ASYNC_FN_NAME("SetStdBasis")
448: PETSC_INTERN PetscErrorCode VecAbsAsync_Private(Vec, PetscDeviceContext);
449: PETSC_INTERN PetscErrorCode VecAXPBYAsync_Private(Vec, PetscScalar, PetscScalar, Vec, PetscDeviceContext);
450: PETSC_INTERN PetscErrorCode VecAXPBYPCZAsync_Private(Vec, PetscScalar, PetscScalar, PetscScalar, Vec, Vec, PetscDeviceContext);
451: PETSC_INTERN PetscErrorCode VecAXPYAsync_Private(Vec, PetscScalar, Vec, PetscDeviceContext);
452: PETSC_INTERN PetscErrorCode VecAYPXAsync_Private(Vec, PetscScalar, Vec, PetscDeviceContext);
453: PETSC_INTERN PetscErrorCode VecConjugateAsync_Private(Vec, PetscDeviceContext);
454: PETSC_INTERN PetscErrorCode VecCopyAsync_Private(Vec, Vec, PetscDeviceContext);
455: PETSC_INTERN PetscErrorCode VecExpAsync_Private(Vec, PetscDeviceContext);
456: PETSC_INTERN PetscErrorCode VecLogAsync_Private(Vec, PetscDeviceContext);
457: PETSC_INTERN PetscErrorCode VecMAXPYAsync_Private(Vec, PetscInt, const PetscScalar *, Vec[], PetscDeviceContext);
458: PETSC_INTERN PetscErrorCode VecPointwiseDivideAsync_Private(Vec, Vec, Vec, PetscDeviceContext);
459: PETSC_INTERN PetscErrorCode VecPointwiseMaxAsync_Private(Vec, Vec, Vec, PetscDeviceContext);
460: PETSC_INTERN PetscErrorCode VecPointwiseMaxAbsAsync_Private(Vec, Vec, Vec, PetscDeviceContext);
461: PETSC_INTERN PetscErrorCode VecPointwiseMinAsync_Private(Vec, Vec, Vec, PetscDeviceContext);
462: PETSC_INTERN PetscErrorCode VecPointwiseMultAsync_Private(Vec, Vec, Vec, PetscDeviceContext);
463: PETSC_INTERN PetscErrorCode VecPointwiseSignAsync_Private(Vec, Vec, VecSignMode, PetscDeviceContext);
464: PETSC_INTERN PetscErrorCode VecReciprocalAsync_Private(Vec, PetscDeviceContext);
465: PETSC_INTERN PetscErrorCode VecScaleAsync_Private(Vec, PetscScalar, PetscDeviceContext);
466: PETSC_INTERN PetscErrorCode VecSetAsync_Private(Vec, PetscScalar, PetscDeviceContext);
467: PETSC_INTERN PetscErrorCode VecShiftAsync_Private(Vec, PetscScalar, PetscDeviceContext);
468: PETSC_INTERN PetscErrorCode VecSqrtAbsAsync_Private(Vec, PetscDeviceContext);
469: PETSC_INTERN PetscErrorCode VecSwapAsync_Private(Vec, Vec, PetscDeviceContext);
470: PETSC_INTERN PetscErrorCode VecWAXPYAsync_Private(Vec, PetscScalar, Vec, Vec, PetscDeviceContext);
472: #define VecMethodDispatch(v, dctx, async_name, name, async_arg_types, ...) \
473: do { \
474: PetscErrorCode(*_8_f) async_arg_types = NULL; \
475: if (dctx) PetscCall(PetscObjectQueryFunction((PetscObject)(v), async_name, &_8_f)); \
476: if (_8_f) { \
477: PetscCall((*_8_f)(v, __VA_ARGS__, dctx)); \
478: } else { \
479: PetscUseTypeMethod(v, name, __VA_ARGS__); \
480: } \
481: } while (0)
483: // return in lda the vector's local size aligned to bytes, where lda is an integer pointer
484: #define VecGetLocalSizeAligned(v, alignment, lda) \
485: do { \
486: PetscInt n = (v)->map->n; \
487: const size_t s = (alignment) / sizeof(PetscScalar); \
488: *(lda) = ((n + s - 1) / s) * s; \
489: } while (0)
491: #define VecSignZeroToZero_Private(y) (((y) > 0) - ((y) < 0))
493: #define VecSignZeroToSignedUnit_Private(y) PetscCopysignReal(1.0, (y))
495: #define VecSignZeroToSignedZero_Private(y) (PetscCopysignReal(1.0, (y)) * ((y) != 0.0))