void col2blk(const int M, const int N, const TYPE *A, const int lda, TYPE *V,
const SCALAR alpha)
{
const int nMb = ATL_DivByNB(M), ib = M - ATL_MulByNB(nMb);
const int incA = (lda - M)<<1, incv = ATL_MulByNB(N);
const int incV = (incv<<1) - NB;
int i, ii, j;
TYPE *rp = V+ATL_MulByNB(N), *ip = V, *prp, *pip;
#ifdef ALPHAXI0
#ifdef Conj_
const register TYPE ralpha = *alpha, calpha = -ralpha;
#else
const register TYPE ralpha = *alpha;
#endif
#elif defined(ALPHAX)
const register TYPE ralpha = *alpha, ialpha = alpha[1];
register TYPE ra, ia;
#endif
pip = V + (M-ib)*(N<<1);
prp = pip + ib*N;
for (j=N; j; j--, V += NB, A += incA)
{
ip = V;
rp = V + incv;
for (ii=nMb; ii; ii--, rp += incV, ip += incV)
{
for (i=NB; i; i--, A += 2, rp++, ip++) scalcp(A, rp, ip);
}
for (i=ib; i; i--, A += 2, prp++, pip++) scalcp(A, prp, pip);
}
}
static void row2blkT_NB(const int M, const int N, const TYPE *A, const int lda,
TYPE *vr, TYPE *vi, const SCALAR alpha)
{
const int incA = lda<<2, incv = 2 - NBNB;
const TYPE *pA0 = A, *pA1 = A + (lda<<1);
int i, j;
#ifdef ALPHAXI0
#ifdef Conj_
const register TYPE ralpha = *alpha, calpha = -ralpha;
#else
const register TYPE ralpha = *alpha;
#endif
#elif defined(ALPHAX)
const register TYPE ralpha = *alpha, ialpha = alpha[1];
register TYPE ra, ia;
#endif
#if ((NB/2)*2 != NB) /* ATLAS should ensure NB divisible by 2 */
ATL_assert((NB/2)*2 == NB);
#endif
for (j=(NB>>1); j; j --, pA0 += incA, pA1 += incA, vr += incv, vi += incv)
{
for (i=0; i != NB2; i += 2, vr += NB, vi += NB)
{
scalcp(pA0+i, vr, vi);
scalcp(pA1+i, vr+1, vi+1);
}
}
}
void Mjoin(prow2blkT,_blk)(const int blk, const int M, const int N,
const SCALAR 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
*/
{
const int kb = Mmin(blk,N);
const int ncb = N / kb, nr = N - ncb*kb;
const int incV = kb*M - kb;
const int VN = kb*M, vn = nr*M;
int jb, i, j;
TYPE *v;
#ifdef ALPHAXI0
#ifdef Conj_
const register TYPE ralpha = *alpha, calpha = -ralpha;
#else
const register TYPE ralpha = *alpha;
#endif
#elif defined(ALPHAX)
register const TYPE ralpha=(*alpha), ialpha = alpha[1];
register TYPE ra, ia;
#endif
if (ldainc == -1) lda--;
lda -= M;
lda += lda;
for (jb=ncb; jb; jb--)
{
for (j=kb; j; j--)
{
v = V++;
for (i=0; i != M; i++, v += kb, A += 2) scalcp(A, v+VN, v);
A += lda;
lda += ldainc;
}
V += incV;
}
for (j=nr; j; j--)
{
v = V++;
for (i=0; i != M; i++, v += nr, A += 2) scalcp(A, v+vn, v);
A += lda;
lda += ldainc;
}
}
void Mjoin(pcol2blk,_blk)(const int blk, const int M, const int N,
const SCALAR 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,blk);
const int nrb = M / kb, mr = M - nrb*kb;
const int nv = kb*N, nvv = mr*N;
const int NN = nv+nv - kb;
const int ldainc2 = ldainc+ldainc, M2 = M+M;
int i, ib, j, J;
TYPE *v = V + nrb*(NN+kb);
#ifdef ALPHAXI0
#ifdef Conj_
const register TYPE ralpha = *alpha, calpha = -ralpha;
#else
const register TYPE ralpha = *alpha;
#endif
#elif defined(ALPHAX)
const register TYPE ralpha=(*alpha), ialpha = alpha[1];
register TYPE ra, ia;
#endif
if (ldainc == -1) lda--;
lda += lda;
ATL_assert(N <= blk);
for (j=0; j != N; j++)
{
for (ib=nrb; ib; ib--)
{
for (i=0; i < kb; i++, A += 2, V++) scalcp(A, V+nv, V);
V += NN;
}
if (mr)
{
for (i=0; i < mr; i++, A += 2, v++) scalcp(A, v+nvv, v);
}
V += kb - nrb*(NN+kb);
A += lda - M2;
lda += ldainc2;
}
}
void prow2blk_KB(const int mb, const int nb, const SCALAR alpha, const TYPE *A,
int lda, const int ldainc, TYPE *V)
/*
* This routine used by full copy to copy one mbxnb block of a matrix A to
* block-major nbxmb storage (A is transposed during the copy)
*/
{
TYPE *v;
const int mn = mb * nb, ldainc2 = ldainc+ldainc;
int i, j;
#ifdef ALPHAXI0
#ifdef Conj_
const register TYPE ralpha = *alpha, calpha = -ralpha;
#else
const register TYPE ralpha = *alpha;
#endif
#elif defined(ALPHAX)
register const TYPE ralpha=(*alpha), ialpha = alpha[1];
register TYPE ra, ia;
#endif
if (ldainc == -1) lda--;
lda -= mb;
lda += lda;
for (j=nb; j; j--)
{
v = V++;
for (i=0; i != mb; i++, v += nb, A += 2) scalcp(A, v+mn, v);
A += lda;
lda += ldainc2;
}
}
static void row2blkT_KB(const int M, const int N, const TYPE *A, const int lda,
TYPE *vr, TYPE *vi, const SCALAR alpha)
{
const int M2 = M<<1, lda2 = lda<<1, incv = 1 - M*N;
int i, j;
#ifdef ALPHAXI0
#ifdef Conj_
const register TYPE ralpha = *alpha, calpha = -ralpha;
#else
const register TYPE ralpha = *alpha;
#endif
#elif defined(ALPHAX)
const register TYPE ralpha = *alpha, ialpha = alpha[1];
register TYPE ra, ia;
#endif
for (j=N; j; j--, A += lda2, vr += incv, vi += incv)
{
for (i=0; i != M2; i += 2, vr += N, vi += N) scalcp(A+i, vr, vi);
}
}