//
// FIXME: Report correct return value - Needs change in KPRINTF
//
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
DUMMY_ARGS_USAGE_2(subdims, pgran);
CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra;
char tempTemplate[32*1024];
if ( buf == NULL) // return buffer size
{
buflen = (32 * 1024 * sizeof(char));
return (ssize_t)buflen;
}
#ifdef DEBUG_ROTMG
printf("dataType : %c\n", Prefix[extraFlags->dtype]);
#endif
strcpy( tempTemplate, (char*)rotmg_kernel );
kprintf kobj( Prefix[extraFlags->dtype], 1, false, false);
kobj.spit((char*)buf, tempTemplate);
return (32 * 1024 * sizeof(char));
}
BencodeObject* BencodeObject::setValueForKey(const char* key, BencodeObject* val) {
if (_type != BencodeTypeDictionary) {
return NULL;
}
removeValueForKey(key);
BencodeObject kobj(key, BencodeModeCopy);
return const_cast<BencodeObject*>(&_dictValue->insert(BencodeDictStorage::value_type(kobj, *val)).first->second);
}
//
// FIXME: Report correct return value - Needs change in KPRINTF
//
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
DUMMY_ARG_USAGE(subdims);
size_t BLOCKSIZE = pgran->wgSize[0];
char tempTemplate[32*1024];
if ( buf == NULL) // return buffer size
{
buflen = (32 * 1024 * sizeof(char));
return (ssize_t)buflen;
}
CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra;
#ifdef DEBUG_ASUM
printf("ASUM GENERATOR called....\n");
printf("dataType : %c\n", Prefix[extraFlags->dtype]);
#endif
unsigned int vecLenA = extraFlags->vecLenA;
#ifdef DEBUG_ASUM
printf("Vector length used : %d\n\n", vecLenA);
#endif
bool doVLOAD = false;
if( extraFlags->flags & KEXTRA_NO_COPY_VEC_A )
{
doVLOAD = true;
#ifdef DEBUG_ASUM
printf("DOing VLOAD as Aligned Data Pointer not Availabe\n");
#endif
}
else
{
#ifdef DEBUG_ASUM
printf("Using Aligned Data Pointer \n");
#endif
}
strcpy( tempTemplate, (char*)asum_kernel );
kprintf kobj( Prefix[extraFlags->dtype], vecLenA, doVLOAD, doVLOAD, BLOCKSIZE);
kobj.spit((char*)buf, tempTemplate);
return (32 * 1024 * sizeof(char));
}
//
// FIXME: Report correct return value - Needs change in KPRINTF
//
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
size_t BLOCKSIZE = pgran->wgSize[0];
char tempTemplate[32*1024];
SolutionStep *step = container_of(subdims, subdims, SolutionStep);
if ( buf == NULL) // return buffer size
{
buflen = (32 * 1024 * sizeof(char));
return (ssize_t)buflen;
}
CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra;
unsigned int vecLenA = extraFlags->vecLenA;
bool doVLOAD = false;
if( extraFlags->flags & KEXTRA_NO_COPY_VEC_A )
{
doVLOAD = true;
}
const char *kernName;
if(step->args.redctnType == REDUCE_BY_HYPOT) {
kernName = nrm2_hypot_kernel;
} else if (step->args.redctnType == REDUCE_BY_SSQ) {
kernName = nrm2_ssq_kernel;
} else {
printf(" Error in selecting kernel!\n");
return 0;
}
strcpy( tempTemplate, kernName );
kprintf kobj( Prefix[extraFlags->dtype], vecLenA, doVLOAD, doVLOAD, BLOCKSIZE);
kobj.spit((char*)buf, tempTemplate);
return (32 * 1024 * sizeof(char));
}
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
CLBLASKernExtra *kextra = (CLBLASKernExtra*)extra;
KernelExtraFlags kflags = kextra->flags;
DataType dtype = kextra->dtype;
char tempTemplate[32*1024];
char itemx[10], itemy[10], width[10], itemy_by_width[10];
size_t Y, X, BLOCKSIZE, ITEMX, ITEMY;
if (buf == NULL)
{
buflen = 32*1024*sizeof(char);
return (ssize_t)buflen;
}
//
// Row-major is implemented in terms of column major routines
//
if ((kflags & KEXTRA_COLUMN_MAJOR) == 0)
{
return 0;
}
kprintf kobj(Prefix[dtype], kextra->vecLenA, true, true);
BLOCKSIZE = pgran->wgSize[0];
#ifdef DEBUG_SYMM
printf("SYMM- generator(): Blocksize passed = %lu, subdimy = %lu, subdimx = %lu, veclen = %lu \n", BLOCKSIZE, subdims->y, subdims->x, kextra->vecLenA);
#endif
Y = 16;
while (Y*(kextra->vecLenA) > subdims->y)
{
Y /= 2;
}
X = BLOCKSIZE/Y;
ITEMY = (subdims->y) / Y;
ITEMX = (subdims->x) / X;
if (ITEMX == 0)
{
ITEMX = 1;
}
if ((BLOCKSIZE % Y) || ((subdims->y) % Y) || ((subdims->x)%X) || (ITEMY % kextra->vecLenA))
{
printf("WARNING: SYMM- generator: subdim and blocksize in-compatible.\n");
}
sprintf(width, "%" SPREFIX "u", Y);
sprintf(itemy, "%" SPREFIX "u", ITEMY);
sprintf(itemx, "%" SPREFIX "u", ITEMX);
sprintf(itemy_by_width, "%" SPREFIX "u", (size_t) ITEMY/kextra->vecLenA);
kobj.put("%WIDTH", width);
kobj.put("%ITEMX", itemx);
kobj.put("%ITEMY", itemy);
kobj.put("%ITEMY_BY_V", itemy_by_width);
#ifdef DEBUG_SYMM
printf("ColMajor SYMM - WIDTH = %s, ITEMX = %s, ITEMY = %s\n", width, itemx, itemy);
#endif
strcpy(tempTemplate, SYMM_C_KERNEL);
kobj.spit(buf, tempTemplate);
#ifdef DEBUG_SYMM
printf("Kernel = \n%s\n", buf);
#endif
size_t tail = strlen(buf) + 1;
while(tail < 32*1024)
{
buf[tail++] = 0;
}
return 32*1024*sizeof(char);
}
void BencodeObject::removeValueForKey(const char* key) {
BencodeObject kobj(key, BencodeModeCopy);
_dictValue->erase(kobj);
}
//
// FIXME: Report correct return value when "buf" is NULL - Needs change in KPRINTF
// FIXME: Return correct return value when "buf" is NON NULL - Needs change in KPRINTF
// FIXME: "buflen" check needs to be more accurate. Relies on above changes to KPRINTF
//
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra;
unsigned int vecLenA = extraFlags->vecLenA;
char tempTemplate[32*1024];
char TARGETROWS_S[10], NLOOPS_S[10], TARGETWIDTH_S[10];
size_t TARGETROWS, NLOOPS, TARGETWIDTH;
char TARGETHEIGHT_S[10], BLOCKSIZE_S[10], TRIANGLE_HEIGHT_S[10];
size_t TARGETHEIGHT;
bool doVLOAD = false;
int BLOCKSIZE = pgran->wgSize[0] * pgran->wgSize[1]; // [1] will always be 1 since we are a 1D implementation
if (buf == NULL) // PENDING: Return correct buffer size
{
return (32 * 1024 * sizeof(char));
}
if (buflen > 32*1024)
{
#ifdef DEBUG_TRSV_GEMV
printf("TRSV GEMV: generator(): WARNING: Returning 0 as buflen is > 32K\n");
#endif
return 0;
}
if( extraFlags->flags & KEXTRA_NO_COPY_VEC_A )
{
doVLOAD = true;
#ifdef DEBUG_TRSV_GEMV
printf("DOing VLOAD as Aligned Data Pointer not Availabe\n");
#endif
}
else
{
#ifdef DEBUG_TRSV_GEMV
printf("Using Aligned Data Pointer .........................\n");
#endif
}
kprintf kobj( Prefix[extraFlags->dtype], vecLenA, doVLOAD);
#ifdef DEBUG_TRSV_GEMV
printf("TRSV GEMV GENERATOR called....\n");
#endif
clblasUplo uplo = ( extraFlags->flags & KEXTRA_UPPER_TRIANG) ? clblasUpper : clblasLower;
clblasOrder order = ( extraFlags->flags & KEXTRA_COLUMN_MAJOR) ? clblasColumnMajor: clblasRowMajor;
clblasTranspose trans =
(extraFlags->flags & KEXTRA_TRANS_A) ? clblasTrans : (( extraFlags->flags & KEXTRA_CONJUGATE_A) ? clblasConjTrans: clblasNoTrans);
bool unit = (((extraFlags->flags) & KEXTRA_UNIT_DIAGONAL) != 0);
// unity and doConj handled in setKernelArgs
if ( order == clblasRowMajor )
{
order = clblasColumnMajor;
if ( trans == clblasNoTrans)
{
trans = clblasTrans;
}
else if ( trans == clblasTrans )
{
trans = clblasNoTrans;
}
else // clblasConjTrans
{
trans = clblasNoTrans;
}
uplo = ( uplo == clblasUpper)? clblasLower : clblasUpper;
}
//
// Check Feasibility and then generate the code.
//
if ( trans != clblasNoTrans)
{
if (isTransposeFeasible(subdims->y, BLOCKSIZE, vecLenA, TARGETHEIGHT) == false)
{
return 0;
}
sprintf( TARGETHEIGHT_S, "%" SPREFIX "u", TARGETHEIGHT );
sprintf( BLOCKSIZE_S, "%d", BLOCKSIZE );
sprintf( TRIANGLE_HEIGHT_S, "%" SPREFIX "u", subdims->y );
kobj.put("%TARGET_HEIGHT", TARGETHEIGHT_S);
kobj.put("%BLOCKSIZE", BLOCKSIZE_S);
kobj.put("%TRIANGLE_HEIGHT", TRIANGLE_HEIGHT_S);
( uplo == clblasLower )?
(strcpy(tempTemplate, (char*)trsv_CLT_ComputeRectangle_kernel)) :
(strcpy(tempTemplate, (char*)trsv_CUT_ComputeRectangle_kernel));
}
else // No-Transpose cases...
{
if (isNoTransposeFeasible(subdims->y, BLOCKSIZE, vecLenA, TARGETROWS, TARGETWIDTH, NLOOPS) == false)
{
return 0;
}
sprintf( TARGETROWS_S, "%" SPREFIX "u", TARGETROWS );
sprintf( TARGETWIDTH_S, "%" SPREFIX "u", TARGETWIDTH );
sprintf( NLOOPS_S, "%" SPREFIX "u", NLOOPS );
kobj.put("%TARGET_ROWS", TARGETROWS_S);
kobj.put("%TARGET_WIDTH", TARGETWIDTH_S);
kobj.put("%NLOOPS", NLOOPS_S);
if (unit)
{
( uplo == clblasLower )?
(strcpy(tempTemplate, (char*)trsv_CL_ComputeRectangle_kernel)) : (strcpy(tempTemplate, (char*)trsv_CU_ComputeRectangle_kernel));
} else {
( uplo == clblasLower )?
(strcpy(tempTemplate, (char*)trsv_CL_ComputeRectangle_NonUnity_kernel)) : (strcpy(tempTemplate, (char*)trsv_CU_ComputeRectangle_NonUnity_kernel));
}
}
#ifdef DEBUG_TRSV_GEMV
printf("dataType : %c\n", Prefix[extraFlags->dtype]);
#endif
// FIXME: VECTORSIZE HARD CODED
// FIXME : SetKernelArgs.. sends offa, offx, and lda should be received as uint
#ifdef DEBUG_TRSV_GEMV
printf("Vector length used : %d\n\n", vecLenA);
#endif
kobj.spit((char*)buf, tempTemplate);
return (32 * 1024 * sizeof(char));
}
//
// FIXME: Report correct return value - Needs change in KPRINTF
//
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
int BLOCKSIZE = pgran->wgSize[0];
char tempTemplate[64*1024];
char targetRows[10], blockSize[10];
if ( buf == NULL) // return buffer size
{
buflen = (64 * 1024 * sizeof(char));
return (ssize_t)buflen;
}
CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra;
#ifdef DEBUG_HER2
printf("HER2 GENERATOR called....\n");
#endif
clblasUplo uplo = ( extraFlags->flags & KEXTRA_UPPER_TRIANG) ? clblasUpper : clblasLower;
if ((subdims->y % extraFlags->vecLenA) != 0)
{
printf("WARNING: HER2: generator: TARGETROWS must be divisible by Vector Length\n");
return 0;
}
size_t TARGETROWS = 0;
( uplo == clblasLower )?
(strcpy(tempTemplate, (char*)syr2_her2_CL_kernel)) : (strcpy(tempTemplate, (char*)syr2_her2_CU_kernel));
TARGETROWS = subdims->y;
if ((BLOCKSIZE % TARGETROWS) != 0)
{
printf("WARNING: HER2: generator: Invalid Block Size\n");
return 0;
}
#ifdef DEBUG_HER2
printf("dataType : %c\n", Prefix[extraFlags->dtype]);
#endif
// FIXME: VECTORSIZE HARD CODED
// FIXME : SetKernelArgs.. sends offa, offx, and lda should be received as uint
unsigned int vecLenA = extraFlags->vecLenA;
#ifdef DEBUG_HER2
printf("Vector length used : %d\n\n", vecLenA);
#endif
bool doVLOAD = false;
if( extraFlags->flags & KEXTRA_NO_COPY_VEC_A )
{
doVLOAD = true;
#ifdef DEBUG_HER2
printf("DOing VLOAD as Aligned Data Pointer not Availabe\n");
#endif
}
else
{
#ifdef DEBUG_HER2
printf("Using Aligned Data Pointer .........................\n");
#endif
}
kprintf kobj( Prefix[extraFlags->dtype], vecLenA, doVLOAD, doVLOAD);
sprintf( targetRows, "%" SPREFIX "u", TARGETROWS );
sprintf( blockSize, "%d", BLOCKSIZE );
#ifdef DEBUG_HER2
printf("TARGET ROWS = %s\n", targetRows);
printf("BLOCK SIZE = %s\n", blockSize);
#endif
kobj.put("%TARGET_ROWS", (const char *)targetRows);
kobj.put("%BLOCKSIZE", (const char *) blockSize);
kobj.spit((char*)buf, tempTemplate);
return (64 * 1024 * sizeof(char));
// return 0;//(ret < 0) ? -EOVERFLOW : ret;
}
//
// FIXME: Report correct return value - Needs change in KPRINTF
//
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
size_t BLOCKSIZE = pgran->wgSize[0];
char tempTemplate[32*1024];
char targetRows[10], blockSize[10];
if ( buf == NULL) // return buffer size
{
buflen = (64 * 1024 * sizeof(char));
return (ssize_t)buflen;
}
CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra;
#ifdef DEBUG_TRMV
printf("TRMV GENERATOR called....\n");
#endif
if((( extraFlags->flags & KEXTRA_TRANS_A) || ( extraFlags ->flags & KEXTRA_CONJUGATE_A )))
{
#ifdef DEBUG_TRMV
printf("A is trans or CONJ-TRANS\n");
#endif
}
else
{
#ifdef DEBUG_TRMV
printf("A is noTrans...\n");
#endif
}
clblasUplo uplo = ( extraFlags->flags & KEXTRA_UPPER_TRIANG) ? clblasUpper : clblasLower;
clblasOrder order = ( extraFlags->flags & KEXTRA_COLUMN_MAJOR) ? clblasColumnMajor: clblasRowMajor;
clblasTranspose trans = ( extraFlags->flags & KEXTRA_TRANS_A) ? clblasTrans : (( extraFlags->flags & KEXTRA_CONJUGATE_A) ? clblasConjTrans: clblasNoTrans);
// unity and doConj handled in setKernelArgs
if ( order == clblasRowMajor )
{
order = clblasColumnMajor;
if ( trans == clblasNoTrans)
{
trans = clblasTrans;
}
else if ( trans == clblasTrans )
{
trans = clblasNoTrans;
}
else // clblasConjTrans
{
trans = clblasNoTrans;
}
uplo = ( uplo == clblasUpper)? clblasLower : clblasUpper;
}
if ((subdims->y % extraFlags->vecLenA) != 0)
{
printf("WARNING: TRMV: generator: TARGETROWS must be divisible by Vector Length\n");
return 0;
}
size_t TARGETROWS = 0;
if ( trans == clblasNoTrans)
{
#ifdef DEBUG_TRMV
printf("clblasNoTrans....%s\n", ( uplo == clblasLower )?"LOWER":"UPPER");
#endif
( uplo == clblasLower )?
(strcpy(tempTemplate, (char*)trmv_CL_kernel)) : (strcpy(tempTemplate, (char*)trmv_CU_kernel));
TARGETROWS = subdims->y;
if ((BLOCKSIZE % TARGETROWS) != 0)
{
printf("WARNING: TRMV: generator: Invalid Block Size\n");
return 0;
}
}
else // Transpose cases...
{
#ifdef DEBUG_TRMV
printf("clblasTrans....%s\n", ( uplo == clblasLower )?"LOWER":"UPPER");
#endif
( uplo == clblasLower )?
(strcpy(tempTemplate, (char*)trmv_CLT_kernel)) : (strcpy(tempTemplate, (char*)trmv_CUT_kernel));
if ((BLOCKSIZE % (subdims->y / extraFlags->vecLenA)) != 0)
{
printf("WARNING: TRMV: generator: Invalid Block Size\n");
return 0;
}
TARGETROWS = BLOCKSIZE/(subdims->y / extraFlags->vecLenA);
}
#ifdef DEBUG_TRMV
printf("dataType : %c\n", Prefix[extraFlags->dtype]);
#endif
// FIXME: VECTORSIZE HARD CODED
// FIXME : SetKernelArgs.. sends offa, offx, and lda should be received as uint
unsigned int vecLenA = extraFlags->vecLenA;
#ifdef DEBUG_TRMV
printf("Vector length used : %d\n\n", vecLenA);
#endif
bool doVLOAD = false;
if( extraFlags->flags & KEXTRA_NO_COPY_VEC_A )
{
doVLOAD = true;
#ifdef DEBUG_TRMV
printf("DOing VLOAD as Aligned Data Pointer not Availabe\n");
#endif
}
else
{
#ifdef DEBUG_TRMV
printf("Using Aligned Data Pointer .........................\n");
#endif
}
kprintf kobj( Prefix[extraFlags->dtype], vecLenA, doVLOAD);
sprintf( targetRows, "%" SPREFIX "u", TARGETROWS );
sprintf( blockSize, "%" SPREFIX "u", BLOCKSIZE );
#ifdef DEBUG_TRMV
printf("TARGET ROWS = %s\n", targetRows);
printf("BLOCK SIZE = %s\n", blockSize);
#endif
kobj.put("%TARGET_ROWS", (const char *)targetRows);
kobj.put("%BLOCKSIZE", (const char *) blockSize);
kobj.spit((char*)buf, tempTemplate);
return (64 * 1024 * sizeof(char));
// return 0;//(ret < 0) ? -EOVERFLOW : ret;
}
int main(int argc, char** argv)
{
int ppw=10; // Point per wavelength
std::string filename="kitenormcond10.txt";
std::vector<double> freqs;
freqs.push_back(5);
freqs.push_back(10);
freqs.push_back(20);
freqs.push_back(40);
freqs.push_back(80);
freqs.push_back(160);
freqs.push_back(320);
freqs.push_back(640);
std::vector<double> norm_sl(freqs.size());
std::vector<double> norm_dl(freqs.size());
std::vector<double> norm_combined1(freqs.size());
std::vector<double> norm_combined2(freqs.size());
std::vector<double> cond_sl(freqs.size());
std::vector<double> cond_dl(freqs.size());
std::vector<double> cond_combined1(freqs.size());
std::vector<double> cond_combined2(freqs.size());
clock_t start, finish;
double time;
start=clock();
#ifdef BEM2DMPI
MPI_Init(&argc, &argv);
int nprow=4; // Number of rows in process grid
int npcol=2; // Number of columns in process grid
int mb=24; // Row Block size
int nb=24; // Column Block size
bem2d::BlacsSystem* b=bem2d::BlacsSystem::Initialize(nprow,npcol,mb,nb);
// Exit if Context could not be created or process does not belong to context
if (!b) {
std::cout << "Could not create Blacs context" << std::endl;
MPI_Finalize();
exit(1);
}
if ((b->get_myrow()==-1)&&(b->get_mycol()==-1)) {
MPI_Finalize();
exit(0);
}
#endif
for (int j=0; j<freqs.size(); j++) {
bem2d::freqtype k= {(double)freqs[j],0};
double eta1=k.re; // Coupling between conj. double and single layer pot.
double eta2=cbrt(k.re*k.re);
bem2d::pCurve kobj(new bem2d::Kite);
int n=(int)(kobj->Length()*k.re*ppw/2.0/bem2d::PI);
bem2d::AnalyticCurve kite(n,kobj);
bem2d::pGeometry pgeom=kite.GetGeometry();
bem2d::PolBasis::AddBasis(0,pgeom); // Add constant basis functions
// Discretize the single and double layer potential
bem2d::SingleLayer sl(k);
bem2d::ConjDoubleLayer cdl(k);
bem2d::DoubleLayer dl(k);
bem2d::QuadOption quadopts;
quadopts.L=3;
quadopts.N=5;
quadopts.sigma=0.15;
#ifdef BEM2DMPI
if (b->IsRoot()) {
std::cout << "Discretize Kernels with n=" << n << std::endl;
}
#else
std::cout << "Discretize Kernels with n=" << n << std::endl;
#endif
bem2d::Matrix dsl=*(DiscreteKernel(*pgeom,quadopts,sl));
bem2d::Matrix ddl=(*DiscreteKernel(*pgeom,quadopts,dl));
bem2d::Matrix dcdl=*(DiscreteKernel(*pgeom,quadopts,cdl));
bem2d::Matrix Id=*(EvalIdent(*pgeom, quadopts));
bem2d::Matrix combined1=Id+2.0*dcdl-bem2d::complex(0,2.0)*eta1*dsl;
bem2d::Matrix combined2=Id+2.0*dcdl-bem2d::complex(0,2.0)*eta2*dsl;
dsl=2.0*bem2d::ChangeBasis(dsl,Id);
ddl=2.0*bem2d::ChangeBasis(ddl,Id);
dcdl=2.0*bem2d::ChangeBasis(dcdl,Id);
combined1=bem2d::ChangeBasis(combined1,Id);
combined2=bem2d::ChangeBasis(combined2,Id);
#ifdef BEM2DMPI
if (b->IsRoot()) {
std::cout << "Compute norms and condition numbers" << std::endl;
}
#else
std::cout << "Compute norms and condition numbers" << std::endl;
#endif
bem2d::L2NormCond(dsl,norm_sl[j],cond_sl[j]);
bem2d::L2NormCond(ddl,norm_dl[j],cond_dl[j]);
bem2d::L2NormCond(combined1,norm_combined1[j],cond_combined1[j]);
bem2d::L2NormCond(combined2,norm_combined2[j],cond_combined2[j]);
}
finish=clock();
time=(double(finish)-double(start))/CLOCKS_PER_SEC/60;
#ifdef BEM2DMPI
if (b->IsRoot()) {
#endif
std::ofstream out(filename.c_str());
out << "Single Layer" << std::endl;
for (int j=0; j<freqs.size(); j++) {
out << "k=" << freqs[j] << " Norm: " << norm_sl[j] << " Norm of Inverse: " << cond_sl[j]/norm_sl[j] << " Condition Nr.: " << cond_sl[j] << std::endl;
}
out << "Double Layer" << std::endl;
for (int j=0; j<freqs.size(); j++) {
out << "k=" << freqs[j] << " Norm: " << norm_dl[j] << " Norm of Inverse: " << cond_dl[j]/norm_dl[j] << " Condition Nr.: " << cond_dl[j] << std::endl;
}
out << "Combined Layer eta=k" << std::endl;
for (int j=0; j<freqs.size(); j++) {
out << "k=" << freqs[j] << " Norm: " << norm_combined1[j] << " Norm of Inverse: " << cond_combined1[j]/norm_combined1[j] << " Condition Nr.: " << cond_combined1[j] << std::endl;
}
out << "Combined Layer eta=k^(2/3)" << std::endl;
for (int j=0; j<freqs.size(); j++) {
out << "k=" << freqs[j] << " Norm: " << norm_combined2[j] << " Norm of Inverse: " << cond_combined2[j]/norm_combined2[j] << " Condition Nr.: " << cond_combined2[j] << std::endl;
}
out << "Overalll time (minutes): " << time << std::endl;
std::cout << "Overall time (minutes): " << time << std::endl;
out.close();
#ifdef BEM2DMPI
}
#endif
#ifdef BEM2DMPI
bem2d::BlacsSystem::Release();
MPI_Finalize();
#endif
}
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
size_t BLOCKSIZE = pgran->wgSize[0];
size_t H = subdims->x;
char tempTemplate[64*1024];
char def_target_rows[10], def_h[10];
SolutionStep *step = container_of( pgran , pgran, SolutionStep); // NOTE: using container_of() to get pigFuncID
CLBlasKargs* kargs = (CLBlasKargs*) &(step->args);
if ( buf == NULL) // return buffer size
{
buflen = (64 * 1024 * sizeof(char));
return (ssize_t)buflen;
}
CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra;
//clblasUplo uplo = ( extraFlags->flags & KEXTRA_UPPER_TRIANG) ? clblasUpper : clblasLower;
clblasOrder order = ( extraFlags->flags & KEXTRA_COLUMN_MAJOR) ? clblasColumnMajor: clblasRowMajor;
clblasTranspose trans = ( extraFlags->flags & KEXTRA_TRANS_A) ? clblasTrans : (( extraFlags->flags & KEXTRA_CONJUGATE_A) ? clblasConjTrans: clblasNoTrans);
if ( order == clblasColumnMajor )
{
order = clblasRowMajor;
if ( trans == clblasNoTrans)
{
trans = clblasTrans;
}
else if ( trans == clblasTrans )
{
trans = clblasNoTrans;
}
else // clblasConjTrans
{
trans = clblasNoTrans;
}
}
if( (kargs->pigFuncID == CLBLAS_SBMV) || (kargs->pigFuncID == CLBLAS_HBMV) ) // Only NT kernel is used
{
trans = clblasNoTrans;
}
if ((BLOCKSIZE % H) != 0)
{
printf("WARNING: GBMV: generator: Invalid Block Size\n");
return 0;
}
size_t TARGET_ROWS = BLOCKSIZE / H;
if ( trans == clblasNoTrans)
{
strcpy(tempTemplate, (char*)gbmv_RNT_kernel);
}
else // Transpose cases...
{
strcpy(tempTemplate, (char*)gbmv_RT_kernel);;
}
unsigned int vecLenA = extraFlags->vecLenA;
bool doVLOAD = false; // Always scalar load for banded matrices
kprintf kobj( Prefix[extraFlags->dtype], vecLenA, doVLOAD);
sprintf( def_target_rows, "%d", (int)TARGET_ROWS );
sprintf( def_h, "%d", (int)H );
#ifdef DEBUG_GBMV
printf("GBMV GENERATOR called....\n");
if((( extraFlags->flags & KEXTRA_TRANS_A) || ( extraFlags ->flags & KEXTRA_CONJUGATE_A )))
{
printf("A is trans or CONJ-TRANS\n");
}
else
{
printf("A is noTrans...\n");
}
printf("TARGET ROWS = %s\n", def_target_rows);
printf("H = %s\n", def_h);
printf("dataType : %c\n", Prefix[extraFlags->dtype]);
#endif
kobj.put("%DEF_H", (const char *)def_h);
kobj.put("%DEF_TARGET_ROWS", (const char *)def_target_rows);
kobj.spit((char*)buf, tempTemplate);
return (64 * 1024 * sizeof(char));
}