static void hegen
(enum CBLAS_ORDER Order, enum CBLAS_UPLO Uplo, int N, TYPE *A, int lda)
{
#ifdef POSDEFGEN
PosDefGen(Order, Uplo, N, A, lda);
#else
MakeHEDiagDom(Order, Uplo, N, A, lda);
CrapUpTri(Order, Uplo, N, A, lda);
#endif
}
double GetTimeWithReps_LLT
(int mflopF, int lda, int M, int N, int nb, int Uplo, int Side, int flsizeKB)
{
double mflop, t0, t1, drep;
char *wrksets; /* working sets for kernel calls */
#ifdef TCPLX
const int lda2 = lda+lda;
#else
const int lda2 = lda;
#endif
size_t setsz, setszT; /* work set size in memory, and amnt of it touched */
size_t nrep; /* # of reps required to force mflopF flops */
size_t nset; /* # of working sets allocated */
int i;
setsz=lda*N*ATL_sizeof; /* matrix is entire working set of LLt */
setszT=N*N*ATL_sizeof; /* only touch N*N portion */
mflop = GetFlopCount(LApotrf, Uplo, M, N, 0, 0, CAN_NB);
/*
* Cannot reuse matrices (bogus to factor an already factored matrix), so we
* must take as our total memspace MAX(nrep,nset)*setsz
*/
ATL_assert(mflop > 0.0);
drep = (mflopF*1.0e6) / mflop;
nrep = (int)(drep+0.999999);
/*
* If cacheline flush doesn't work, then we must use this method
*/
#if ATL_LINEFLUSH
if (nrep < 2)
return(-1.0); /* do wt normal timer */
#else
nrep = (nrep >= 1) ? nrep : 1;
#endif
nset = (flsizeKB*1024+setszT-1)/setszT;
if (nset < nrep)
nset = nrep;
wrksets = malloc(nset * setsz);
ATL_assert(wrksets);
for (i=0; i < nset; i++)
PosDefGen(CblasColMajor, Uplo_LA2ATL(Uplo), N,
(TYPE*)(wrksets+i*setsz), lda);
t0 = time00();
for (i=0; i < nrep; i++)
{
test_potrf(Uplo, N, (TYPE*)(wrksets+i*setsz), lda);
}
t1 = time00();
free(wrksets);
return((t1-t0)/((double)nrep));
}
double GetTime(int rout, int mflopF, int lda, int M, int N, int nb, int Uplo,
int Side, int flsizeKB)
{
#if ATL_LINEFLUSH
FLSTRUCT *flp;
#endif
TYPE *A, *wrk=NULL, dtmp, dtmp1, *tau=NULL;
int *ipiv=NULL, itmp, wlen;
double t0, t1;
/*
* Call routs that force particular flop count if requested; they return -1.0
* if one invocation will suffice to force mflopF, in which case do the timing
* in this routine, which is simpler & doesn't require LRU & as much workspace
* If we don't have the ability to do cacheline flushing, must use LRU rout!
*/
#if ATL_LINEFLUSH
if (mflopF > 0)
{
#endif
if (rout == LApotrf)
t1 = GetTimeWithReps_LLT(mflopF, lda, M, N, nb, Uplo, Side, flsizeKB);
else if (rout == LAgeqrf)
{
if (Side == LARight)
{
if (Uplo == LAUpper)
t1 = GetTimeWithReps_QR(mflopF, lda, M, N, nb, Uplo, Side,
flsizeKB);
else
t1 = GetTimeWithReps_QL(mflopF, lda, M, N, nb, Uplo, Side,
flsizeKB);
}
else if (Uplo == LAUpper)
t1 = GetTimeWithReps_RQ(mflopF, lda, M, N, nb, Uplo, Side,
flsizeKB);
else
t1 = GetTimeWithReps_LQ(mflopF, lda, M, N, nb, Uplo, Side,
flsizeKB);
}
else
t1 = GetTimeWithReps_LU(mflopF, lda, M, N, nb, Uplo, Side, flsizeKB);
#if ATL_LINEFLUSH == 0
return(t1);
#else
if (t1 >= 0.0)
return(t1);
}
#endif
#if ATL_LINEFLUSH != 0
/*
* Generate operands
*/
A = GetGE(M, N, lda);
ATL_assert(A);
flp = ATL_GetFlushStruct(A, N*((size_t)lda)*ATL_sizeof, NULL);
if (rout == LApotrf)
PosDefGen(CblasColMajor, Uplo_LA2ATL(Uplo), N, A, lda);
else if (rout & LAgeqrf)
{ /* QR must allocate workspace */
if (Side == LARight)
{
if (Uplo == LAUpper)
{
test_geqrf(CblasColMajor, M, N, A, lda, &dtmp1, &dtmp, -1);
}
else
{
test_geqlf(CblasColMajor, M, N, A, lda, &dtmp1, &dtmp, -1);
}
}
else if (Uplo == LAUpper)
{
test_gerqf(CblasColMajor, M, N, A, lda, &dtmp1, &dtmp, -1);
}
else
{
test_gelqf(CblasColMajor, M, N, A, lda, &dtmp1, &dtmp, -1);
}
wlen = dtmp;
wrk = calloc(wlen, ATL_sizeof);
ATL_assert(wrk);
flp = ATL_GetFlushStruct(wrk, wlen*ATL_sizeof, flp);
itmp = (M >= N) ? M : N;
tau = calloc(itmp, ATL_sizeof);
flp = ATL_GetFlushStruct(tau, itmp*ATL_sizeof, flp);
}
else
{
ipiv = calloc(M, sizeof(int));
ATL_assert(ipiv);
flp = ATL_GetFlushStruct(ipiv, M*sizeof(int), flp);
}
/*
* Flush cache, and do timing
*/
ATL_FlushAreasByCL(flp);
if (rout == LApotrf)
{
t0 = time00();
test_potrf(Uplo, N, A, lda);
t1 = time00();
}
else if (rout == LAgeqrf)
{
if (Side == LARight)
{
if (Uplo == LAUpper)
{
t0 = time00();
test_geqrf(CblasColMajor, M, N, A, lda, tau, wrk, wlen);
t1 = time00();
}
else
{
t0 = time00();
test_geqlf(CblasColMajor, M, N, A, lda, tau, wrk, wlen);
t1 = time00();
}
}
else if (Uplo == LAUpper)
{
t0 = time00();
test_gerqf(CblasColMajor, M, N, A, lda, tau, wrk, wlen);
t1 = time00();
}
else
{
t0 = time00();
test_gelqf(CblasColMajor, M, N, A, lda, tau, wrk, wlen);
t1 = time00();
}
}
else
{
t0 = time00();
test_getrf(CblasColMajor, M, N, A, lda, ipiv);
t1 = time00();
}
if (tau)
free(tau);
if (wrk)
free(wrk);
if (ipiv)
free(ipiv);
free(A);
ATL_KillAllFlushStructs(flp);
return(t1 - t0);
#endif
}