static TYPE uumtest(enum ATLAS_ORDER Order, enum ATLAS_UPLO Uplo,
int CacheSize, int N, int lda, double *tim)
{
TYPE *A, *Ag, *LmLt;
double t0, t1;
TYPE normA, eps, resid;
enum ATLAS_UPLO MyUplo = Uplo;
if (Order == CblasRowMajor)
{
if (Uplo == CblasUpper) MyUplo = CblasLower;
else MyUplo = CblasUpper;
}
eps = Mjoin(PATL,epsilon)();
A = malloc(ATL_MulBySize(lda)*N + ATL_MulBySize(N)*N);
if (A == NULL) return(-1);
Ag = A + lda*(N SHIFT);
t0 = ATL_flushcache(CacheSize);
lltgen(MyUplo, N, A, lda, N*1029+lda);
lltgen(MyUplo, N, Ag, N, N*1029+lda);
normA = lltnrm1(MyUplo, N, A, lda);
#ifdef DEBUG
Mjoin(PATL,geprint)("A", N, N, A, lda);
Mjoin(PATL,geprint)("Ag", N, N, Ag, N);
#endif
t0 = ATL_flushcache(-1);
t0 = time00();
test_lauum(Order, Uplo, N, A, lda);
t1 = time00() - t0;
*tim = t1;
t0 = ATL_flushcache(0);
ATL_checkpad(MyUplo, N, A, lda);
if (Uplo == CblasUpper) LmLt = ATL_UmulUt(Order, N, Ag, N);
else LmLt = ATL_LtmulL(Order, N, Ag, N);
#ifdef DEBUG
Mjoin(PATL,geprint)("A", N, N, A, lda);
Mjoin(PATL,geprint)("Ag", N, N, LmLt, N);
#endif
lltdiff(MyUplo, N, A, lda, LmLt, N);
#ifdef DEBUG
Mjoin(PATL,geprint)("A-L*Lt", N, N, LmLt, N);
#endif
resid = lltnrm1(MyUplo, N, LmLt, N) / (normA * eps * N);
if (resid > 10.0 || resid != resid)
fprintf(stderr, "normA=%e, eps=%e, num=%e\n", normA, eps, resid);
free(LmLt);
free(A);
return(resid);
}
static TYPE llttest(enum ATLAS_UPLO Uplo, int CacheSize, int N, int lda,
double *tim)
{
TYPE *A, *LmLt;
int i;
double t0, t1;
TYPE normA, eps, resid;
eps = Mjoin(PATL,epsilon)();
A = malloc(ATL_MulBySize(lda)*N);
if (A == NULL) return(-1);
t0 = ATL_flushcache(CacheSize);
lltgen(Uplo, N, A, lda, N*1029+lda);
normA = lltnrm1(Uplo, N, A, lda);
#ifdef DEBUG
Mjoin(PATL,geprint)("A0", N, N, A, lda);
#endif
t0 = ATL_flushcache(-1);
t0 = time00();
test_potrf(Uplo, N, A, lda);
t1 = time00() - t0;
*tim = t1;
t0 = ATL_flushcache(0);
#ifdef DEBUG
Mjoin(PATL,geprint)("L", N, N, A, lda);
#endif
ATL_checkpad(Uplo, N, A, lda);
if (Uplo == AtlasUpper) LmLt = ATL_UtmulU(N, A, lda);
else LmLt = ATL_LmulLt(N, A, lda);
#ifdef DEBUG
Mjoin(PATL,geprint)("L*Lt", N, N, LmLt, N);
#endif
lltgen(Uplo, N, A, lda, N*1029+lda); /* regen A over LLt */
lltdiff(Uplo, N, A, lda, LmLt, N);
#ifdef DEBUG
Mjoin(PATL,geprint)("A-L*Lt", N, N, LmLt, N);
#endif
resid = lltnrm1(Uplo, N, LmLt, N);
#ifdef DEBUG
if (resid/(normA*eps*N) > 10.0)
fprintf(stderr, "normA=%e, eps=%e, num=%e\n", normA, eps, resid);
#endif
resid /= (normA * eps * N);
free(LmLt);
free(A);
return(resid);
}
static TYPE trtritest(enum ATLAS_ORDER Order, enum ATLAS_UPLO Uplo,
enum ATLAS_DIAG Diag, int CacheSize, int N, int lda,
double *tim)
{
TYPE *A, *Acompare;
int i;
double t0, t1;
TYPE normA, eps, resid;
/*int ierr;*/
#ifdef TCPLX
const TYPE one[2]={ATL_rone, ATL_rzero};
#else
const TYPE one = ATL_rone;
#endif
eps = Mjoin(PATL,epsilon)();
A = malloc(ATL_MulBySize(lda)*N);
Acompare = malloc(ATL_MulBySize(lda)*N);
if (A == NULL) return(-1);
if (Acompare == NULL) return(-1);
t0 = ATL_flushcache(CacheSize);
/* create random, diagonally dominant matrix with magic value at
unused places. Last number is just the random seed. */
trigen(Order, Uplo, Diag, N, A, lda, PADVAL, N*1029+lda);
/* Create backup to calculate residual. This one has to be used
as a full matrix, so it has zero fills and correct diagonal. */
trigen(Order, Uplo, Diag, N, Acompare, lda, ATL_rzero, N*1029+lda);
if (Diag==AtlasUnit)
for (i=0; i < N; i++)
Acompare[(i*(lda+1)) SHIFT] = ATL_rone;
normA = trinrm1(Order,Uplo, Diag, N, A, lda);
#ifdef DEBUG
Mjoin(PATL,geprint)("A0", N, N, A, lda);
#endif
t0 = ATL_flushcache(-1);
/* Calculate and time a solution */
t0 = time00();
test_trtri(Order, Uplo, Diag, N, A, lda);
t1 = time00() - t0;
*tim = t1;
/* if (ierr != 0)
{
fprintf(stderr, "Return values != 0 : %d \n",ierr);
return(9999.9999);
}*/
t0 = ATL_flushcache(0);
/* Instroduce a padding error. */
/* A[(5+5*lda)SHIFT]=114.0; */
#ifdef DEBUG
Mjoin(PATL,geprint)("L", N, N, A, lda);
#endif
ATL_checkpad(Order, Uplo, Diag, N, A, lda);
/* Calculate A^{-1}*A */
cblas_trmm(Order,CblasLeft,Uplo,AtlasNoTrans,Diag,
N,N,one,A,lda,Acompare,lda);
#ifdef DEBUG
Mjoin(PATL,geprint)("A^{-1}*A", N, N, Acompare, N);
#endif
/* Subtract diagonal */
for (i=0; i < N; i++)
Acompare[i*((lda+1) SHIFT)] -= ATL_rone;
/*
resid = trinrm1(Order, Uplo,AtlasNonUnit,N,Acompare,lda);
fprintf(stderr, "normA=%e, eps=%e, num=%e\n", normA, eps, resid);
*/
resid = Mjoin(PATL,genrm1)(N, N, Acompare, lda);
#ifdef DEBUG
if (resid/(normA*eps*N) > 10.0)
fprintf(stderr, "normA=%e, eps=%e, num=%e\n", normA, eps, resid);
#endif
resid /= (normA * eps * N);
free(Acompare);
free(A);
return(resid);
}