170baa948SKarl Rupp 270baa948SKarl Rupp /* 370baa948SKarl Rupp Include files needed for the ViennaCL row-scaling preconditioner: 470baa948SKarl Rupp pcimpl.h - private include file intended for use by all preconditioners 570baa948SKarl Rupp */ 670baa948SKarl Rupp #define PETSC_SKIP_SPINLOCK 799acd6aaSStefano Zampini #define PETSC_SKIP_IMMINTRIN_H_CUDAWORKAROUND 1 899acd6aaSStefano Zampini 970baa948SKarl Rupp #include <petsc/private/pcimpl.h> /*I "petscpc.h" I*/ 1070baa948SKarl Rupp #include <../src/mat/impls/aij/seq/aij.h> 1170baa948SKarl Rupp #include <../src/vec/vec/impls/dvecimpl.h> 1270baa948SKarl Rupp #include <../src/mat/impls/aij/seq/seqviennacl/viennaclmatimpl.h> 1370baa948SKarl Rupp #include <../src/vec/vec/impls/seq/seqviennacl/viennaclvecimpl.h> 1470baa948SKarl Rupp #include <viennacl/linalg/row_scaling.hpp> 1570baa948SKarl Rupp 1670baa948SKarl Rupp /* 1770baa948SKarl Rupp Private context (data structure) for the ROWSCALINGVIENNACL preconditioner. 1870baa948SKarl Rupp */ 1970baa948SKarl Rupp typedef struct { 2070baa948SKarl Rupp viennacl::linalg::row_scaling<viennacl::compressed_matrix<PetscScalar>> *ROWSCALINGVIENNACL; 2170baa948SKarl Rupp } PC_ROWSCALINGVIENNACL; 2270baa948SKarl Rupp 2370baa948SKarl Rupp /* 2470baa948SKarl Rupp PCSetUp_ROWSCALINGVIENNACL - Prepares for the use of the ROWSCALINGVIENNACL preconditioner 2570baa948SKarl Rupp by setting data structures and options. 2670baa948SKarl Rupp 2770baa948SKarl Rupp Input Parameter: 2870baa948SKarl Rupp . pc - the preconditioner context 2970baa948SKarl Rupp 3070baa948SKarl Rupp Application Interface Routine: PCSetUp() 3170baa948SKarl Rupp 32f1580f4eSBarry Smith Note: 3370baa948SKarl Rupp The interface routine PCSetUp() is not usually called directly by 3470baa948SKarl Rupp the user, but instead is called by PCApply() if necessary. 3570baa948SKarl Rupp */ 369371c9d4SSatish Balay static PetscErrorCode PCSetUp_ROWSCALINGVIENNACL(PC pc) { 3770baa948SKarl Rupp PC_ROWSCALINGVIENNACL *rowscaling = (PC_ROWSCALINGVIENNACL *)pc->data; 3870baa948SKarl Rupp PetscBool flg = PETSC_FALSE; 3970baa948SKarl Rupp Mat_SeqAIJViennaCL *gpustruct; 4070baa948SKarl Rupp 4170baa948SKarl Rupp PetscFunctionBegin; 429566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)pc->pmat, MATSEQAIJVIENNACL, &flg)); 4328b400f6SJacob Faibussowitsch PetscCheck(flg, PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Currently only handles ViennaCL matrices"); 4470baa948SKarl Rupp if (pc->setupcalled != 0) { 4570baa948SKarl Rupp try { 4670baa948SKarl Rupp delete rowscaling->ROWSCALINGVIENNACL; 479371c9d4SSatish Balay } catch (char *ex) { SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "ViennaCL error: %s", ex); } 4870baa948SKarl Rupp } 4970baa948SKarl Rupp try { 5070baa948SKarl Rupp #if defined(PETSC_USE_COMPLEX) 5170baa948SKarl Rupp gpustruct = NULL; 5270baa948SKarl Rupp SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "No support for complex arithmetic in ROWSCALINGVIENNACL preconditioner"); 5370baa948SKarl Rupp #else 549566063dSJacob Faibussowitsch PetscCall(MatViennaCLCopyToGPU(pc->pmat)); 5570baa948SKarl Rupp gpustruct = (Mat_SeqAIJViennaCL *)(pc->pmat->spptr); 5670baa948SKarl Rupp 5770baa948SKarl Rupp viennacl::linalg::row_scaling_tag pc_tag(1); 5870baa948SKarl Rupp ViennaCLAIJMatrix *mat = (ViennaCLAIJMatrix *)gpustruct->mat; 5970baa948SKarl Rupp rowscaling->ROWSCALINGVIENNACL = new viennacl::linalg::row_scaling<viennacl::compressed_matrix<PetscScalar>>(*mat, pc_tag); 6070baa948SKarl Rupp #endif 619371c9d4SSatish Balay } catch (char *ex) { SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "ViennaCL error: %s", ex); } 6270baa948SKarl Rupp PetscFunctionReturn(0); 6370baa948SKarl Rupp } 6470baa948SKarl Rupp 6570baa948SKarl Rupp /* 6670baa948SKarl Rupp PCApply_ROWSCALINGVIENNACL - Applies the ROWSCALINGVIENNACL preconditioner to a vector. 6770baa948SKarl Rupp 6870baa948SKarl Rupp Input Parameters: 6970baa948SKarl Rupp . pc - the preconditioner context 7070baa948SKarl Rupp . x - input vector 7170baa948SKarl Rupp 7270baa948SKarl Rupp Output Parameter: 7370baa948SKarl Rupp . y - output vector 7470baa948SKarl Rupp 7570baa948SKarl Rupp Application Interface Routine: PCApply() 7670baa948SKarl Rupp */ 779371c9d4SSatish Balay static PetscErrorCode PCApply_ROWSCALINGVIENNACL(PC pc, Vec x, Vec y) { 7870baa948SKarl Rupp PC_ROWSCALINGVIENNACL *ilu = (PC_ROWSCALINGVIENNACL *)pc->data; 7970baa948SKarl Rupp PetscBool flg1, flg2; 8070baa948SKarl Rupp viennacl::vector<PetscScalar> const *xarray = NULL; 8170baa948SKarl Rupp viennacl::vector<PetscScalar> *yarray = NULL; 8270baa948SKarl Rupp 8370baa948SKarl Rupp PetscFunctionBegin; 8470baa948SKarl Rupp /*how to apply a certain fixed number of iterations?*/ 859566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)x, VECSEQVIENNACL, &flg1)); 869566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)y, VECSEQVIENNACL, &flg2)); 8708401ef6SPierre Jolivet PetscCheck((flg1 && flg2), PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Currently only handles ViennaCL vectors"); 8848a46eb9SPierre Jolivet if (!ilu->ROWSCALINGVIENNACL) PetscCall(PCSetUp_ROWSCALINGVIENNACL(pc)); 899566063dSJacob Faibussowitsch PetscCall(VecSet(y, 0.0)); 909566063dSJacob Faibussowitsch PetscCall(VecViennaCLGetArrayRead(x, &xarray)); 919566063dSJacob Faibussowitsch PetscCall(VecViennaCLGetArrayWrite(y, &yarray)); 9270baa948SKarl Rupp try { 9370baa948SKarl Rupp #if defined(PETSC_USE_COMPLEX) 9470baa948SKarl Rupp 9570baa948SKarl Rupp #else 9670baa948SKarl Rupp *yarray = *xarray; 9770baa948SKarl Rupp ilu->ROWSCALINGVIENNACL->apply(*yarray); 9870baa948SKarl Rupp #endif 999371c9d4SSatish Balay } catch (char *ex) { SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "ViennaCL error: %s", ex); } 1009566063dSJacob Faibussowitsch PetscCall(VecViennaCLRestoreArrayRead(x, &xarray)); 1019566063dSJacob Faibussowitsch PetscCall(VecViennaCLRestoreArrayWrite(y, &yarray)); 1029566063dSJacob Faibussowitsch PetscCall(PetscObjectStateIncrease((PetscObject)y)); 10370baa948SKarl Rupp PetscFunctionReturn(0); 10470baa948SKarl Rupp } 105f1580f4eSBarry Smith 10670baa948SKarl Rupp /* 10770baa948SKarl Rupp PCDestroy_ROWSCALINGVIENNACL - Destroys the private context for the ROWSCALINGVIENNACL preconditioner 10870baa948SKarl Rupp that was created with PCCreate_ROWSCALINGVIENNACL(). 10970baa948SKarl Rupp 11070baa948SKarl Rupp Input Parameter: 11170baa948SKarl Rupp . pc - the preconditioner context 11270baa948SKarl Rupp 11370baa948SKarl Rupp Application Interface Routine: PCDestroy() 11470baa948SKarl Rupp */ 1159371c9d4SSatish Balay static PetscErrorCode PCDestroy_ROWSCALINGVIENNACL(PC pc) { 11670baa948SKarl Rupp PC_ROWSCALINGVIENNACL *rowscaling = (PC_ROWSCALINGVIENNACL *)pc->data; 11770baa948SKarl Rupp 11870baa948SKarl Rupp PetscFunctionBegin; 11970baa948SKarl Rupp if (rowscaling->ROWSCALINGVIENNACL) { 12070baa948SKarl Rupp try { 12170baa948SKarl Rupp delete rowscaling->ROWSCALINGVIENNACL; 1229371c9d4SSatish Balay } catch (char *ex) { SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "ViennaCL error: %s", ex); } 12370baa948SKarl Rupp } 12470baa948SKarl Rupp 12570baa948SKarl Rupp /* 12670baa948SKarl Rupp Free the private data structure that was hanging off the PC 12770baa948SKarl Rupp */ 1289566063dSJacob Faibussowitsch PetscCall(PetscFree(pc->data)); 12970baa948SKarl Rupp PetscFunctionReturn(0); 13070baa948SKarl Rupp } 13170baa948SKarl Rupp 1329371c9d4SSatish Balay static PetscErrorCode PCSetFromOptions_ROWSCALINGVIENNACL(PC pc, PetscOptionItems *PetscOptionsObject) { 13370baa948SKarl Rupp PetscFunctionBegin; 134d0609cedSBarry Smith PetscOptionsHeadBegin(PetscOptionsObject, "ROWSCALINGVIENNACL options"); 135d0609cedSBarry Smith PetscOptionsHeadEnd(); 13670baa948SKarl Rupp PetscFunctionReturn(0); 13770baa948SKarl Rupp } 13870baa948SKarl Rupp 13970baa948SKarl Rupp /*MC 14070baa948SKarl Rupp PCRowScalingViennaCL - A diagonal preconditioner (scaling rows of matrices by their norm) that can be used via the CUDA, OpenCL, and OpenMP backends of ViennaCL 14170baa948SKarl Rupp 14270baa948SKarl Rupp Level: advanced 14370baa948SKarl Rupp 144f1580f4eSBarry Smith Developer Note: 145f1580f4eSBarry Smith This `PCType` does not appear to be registered 14670baa948SKarl Rupp 147f1580f4eSBarry Smith .seealso: `PCCreate()`, `PCSetType()`, `PCType`, `PC` 14870baa948SKarl Rupp M*/ 14970baa948SKarl Rupp 1509371c9d4SSatish Balay PETSC_EXTERN PetscErrorCode PCCreate_ROWSCALINGVIENNACL(PC pc) { 15170baa948SKarl Rupp PC_ROWSCALINGVIENNACL *rowscaling; 15270baa948SKarl Rupp 15370baa948SKarl Rupp PetscFunctionBegin; 15470baa948SKarl Rupp /* 15570baa948SKarl Rupp Creates the private data structure for this preconditioner and 15670baa948SKarl Rupp attach it to the PC object. 15770baa948SKarl Rupp */ 158*4dfa11a4SJacob Faibussowitsch PetscCall(PetscNew(&rowscaling)); 15970baa948SKarl Rupp pc->data = (void *)rowscaling; 16070baa948SKarl Rupp 16170baa948SKarl Rupp /* 16270baa948SKarl Rupp Initialize the pointer to zero 16370baa948SKarl Rupp Initialize number of v-cycles to default (1) 16470baa948SKarl Rupp */ 16570baa948SKarl Rupp rowscaling->ROWSCALINGVIENNACL = 0; 16670baa948SKarl Rupp 16770baa948SKarl Rupp /* 16870baa948SKarl Rupp Set the pointers for the functions that are provided above. 16970baa948SKarl Rupp Now when the user-level routines (such as PCApply(), PCDestroy(), etc.) 17070baa948SKarl Rupp are called, they will automatically call these functions. Note we 17170baa948SKarl Rupp choose not to provide a couple of these functions since they are 17270baa948SKarl Rupp not needed. 17370baa948SKarl Rupp */ 17470baa948SKarl Rupp pc->ops->apply = PCApply_ROWSCALINGVIENNACL; 17570baa948SKarl Rupp pc->ops->applytranspose = 0; 17670baa948SKarl Rupp pc->ops->setup = PCSetUp_ROWSCALINGVIENNACL; 17770baa948SKarl Rupp pc->ops->destroy = PCDestroy_ROWSCALINGVIENNACL; 17870baa948SKarl Rupp pc->ops->setfromoptions = PCSetFromOptions_ROWSCALINGVIENNACL; 17970baa948SKarl Rupp pc->ops->view = 0; 18070baa948SKarl Rupp pc->ops->applyrichardson = 0; 18170baa948SKarl Rupp pc->ops->applysymmetricleft = 0; 18270baa948SKarl Rupp pc->ops->applysymmetricright = 0; 18370baa948SKarl Rupp PetscFunctionReturn(0); 18470baa948SKarl Rupp } 185