void Mjoin(Mjoin(PATL,trcopyU2U_N),NM)
#endif
(const int N, const SCALAR alpha0, const TYPE *A, const int lda, TYPE *C)
/*
* Copies an Upper matrix to a dense matrix with zeros below the diagonal
*/
{
int i, j;
const register TYPE alpha=alpha0;
const TYPE *Ac = A;
if (N > 1)
{
for (j=0; j != N; j++)
{
for (i=0; i != j; i++) C[i] = ATL_MulByALPHA(Ac[i]);
#ifdef UnitDiag_
C[j] = alpha;
#else
C[j] = ATL_MulByALPHA(Ac[j]);
#endif
for (i=j+1; i < N; i++) C[i] = 0.0;
C += N;
Ac += lda;
}
}
else if (N == 1)
{
#ifdef UnitDiag_
*C = alpha;
#else
*C = ATL_MulByALPHA(*A);
#endif
}
}
void Mjoin(PATL,Mjoin(gemove,NM))
(const int M0, const int N, const SCALAR alpha0,
const TYPE *A, const int lda, TYPE *C, const int ldc)
/*
* C <- alpha * A
*/
{
#ifdef ALPHA0
Mjoin(PATL,gezero)(M0, N, C, ldc);
#elif defined(ALPHA1)
Mjoin(PATL,gecopy)(M0, N, A, lda, C, ldc);
#elif defined(TREAL) || defined(ALPHAXI0)
int i, j;
const int n = N>>1;
#ifdef TREAL
#define M M0
const register TYPE alpha = alpha0;
const int incA = lda<<1, incC = ldc<<1;
#else
const register TYPE ralpha = *alpha0;
const int M = M0<<1, incA = lda<<2, incC = ldc<<2;
#endif
const TYPE *A0 = A, *A1 = A+(lda SHIFT);
TYPE *C0 = C, *C1 = C+(ldc SHIFT);
for (j=n; j; j--, A0 += incA, A1 += incA, C0 += incC, C1 += incC)
{
for (i=0; i != M; i++)
{
C0[i] = ATL_MulByALPHA(A0[i]);
C1[i] = ATL_MulByALPHA(A1[i]);
}
}
if (N != (n<<1)) for (i=0; i != M; i++) C0[i] = ATL_MulByALPHA(A0[i]);
#elif defined(TCPLX)
const int incA = (lda-M0)<<1, incC = (ldc-M0)<<1;
int i, j;
const register TYPE ralpha = *alpha0, ialpha = alpha0[1];
register TYPE rtmp, itmp;
for (j=N; j; j--, A += incA, C += incC)
{
for(i=M0; i; i--, A += 2, C += 2)
{
rtmp = *A; itmp = A[1];
*C = rtmp * ralpha - itmp * ialpha;
C[1] = rtmp * ialpha + itmp * ralpha;
}
}
#endif
}
void Mjoin(Mjoin(PATL,pcol2blk),NM)
(const int M, const int N, const TYPE alpha, const TYPE *A, int lda,
const int ldainc, TYPE *V)
/*
* Given a packed matrix A, copies N columns starting at A into
* block-major column panel
* ldainc = 0 : General
* ldainc = 1 : Upper
* ldainc = -1 : Lower
* NOTE: specialize to alpha cases after it works!
*/
{
const int kb = Mmin(M,NB);
const int nrb = M / kb, mr = M - nrb*kb;
int i, ib, j, J;
const int NN = N*kb;
TYPE *v = V + nrb*NN;
if (ldainc)
{
if (ldainc == -1) lda--;
ATL_assert(N <= NB);
for (j=0; j != N; j++)
{
for (ib=nrb; ib; ib--)
{
for (i=0; i < kb; i++) V[i] = ATL_MulByALPHA(A[i]);
V += NN;
A += kb;
}
if (mr)
{
for (i=0; i < mr; i++) v[i] = ATL_MulByALPHA(A[i]);
v += mr;
}
V += kb - nrb*NN;
A += lda - nrb*kb;
lda += ldainc;
}
}
else Mjoin(Mjoin(PATL,col2blk),NM)(M, N, A, lda, V, alpha);
}
void Mjoin(Mjoin(PATL,prow2blkT),NM)
(const int M, const int N, const TYPE alpha, const TYPE *A, int lda,
const int ldainc, TYPE *V)
/*
* Given a packed Upper matrix A, copies & transposes M rows starting at A into
* block-major row panel
* ldainc = 0 : General rectangular
* ldainc = 1 : Upper
* ldainc = -1 : Lower
* NOTE: specialize to alpha cases after it works!
*/
{
const int kb = Mmin(NB,N);
const int ncb = N / kb, nr = N - ncb*kb;
const int incV = kb*M - kb;
int jb, i, j;
TYPE *v;
if (ldainc)
{
if (ldainc == -1) lda--;
for (jb=ncb; jb; jb--)
{
for (j=kb; j; j--)
{
v = V++;
for (i=0; i != M; i++, v += kb) *v = ATL_MulByALPHA(A[i]);
A += lda;
lda += ldainc;
}
V += incV;
}
for (j=nr; j; j--)
{
v = V++;
for (i=0; i != M; i++, v += nr) *v = ATL_MulByALPHA(A[i]);
A += lda;
lda += ldainc;
}
}
else Mjoin(Mjoin(PATL,row2blkT),NM)(N, M, A, lda, V, alpha);
}
void Mjoin(Mjoin(PATL,trcopyU2L_N),NM)
#endif
(const int N, const SCALAR alpha0, const TYPE *A, const int lda, TYPE *C)
#endif
/*
* Takes an Upper matrix, transposes it, and copies it to a dense matrix with
* zeroes above the diagonal (i.e., lower form)
*/
{
int i, j;
const int ldap1 = lda+1;
const register TYPE alpha=alpha0;
const TYPE *Ac = A, *Ar;
TYPE *rC, *cC=C;
if (N > 1)
{
for (j=0; j != N; j++)
{
for (i=0; i != j; i++) C[i] = 0.0;
#ifdef UnitDiag_
C[j] = alpha;
#else
C[j] = ATL_MulByALPHA(*Ac);
#endif
Ar = Ac + lda;
for (i=j+1; i < N; i++, Ar += lda) C[i] = ATL_MulByALPHA(*Ar);
C += N;
Ac += ldap1;
}
}
else if (N == 1)
{
#ifdef UnitDiag_
*C = alpha;
#else
*C = ATL_MulByALPHA(*A);
#endif
}
}
void Mjoin(Mjoin(PATL,sycopyU),NM)
(const int N, const SCALAR alpha0, const TYPE *A, const int lda, TYPE *C)
/*
* Copies a symmetric matrix stored in Upper format to a dense matrix
*/
{
int i, j;
const register TYPE alpha=alpha0;
const TYPE *Ar, *Ac = A;
TYPE *rC, *cC=C;
if (N > 1)
{
for (j=0; j != N; j++)
{
for (i=0; i <= j; i++) C[i] = ATL_MulByALPHA(Ac[i]);
Ar = Ac + j + lda;
for (i=j+1; i < N; i++, Ar += lda) C[i] = ATL_MulByALPHA(*Ar);
C += N;
Ac += lda;
}
}
else if (N == 1) *C = ATL_MulByALPHA(*A);
}
