int Mjoin(PATL,mmIJK)(const enum ATLAS_TRANS TA, const enum ATLAS_TRANS TB,
const int M, const int N0, const int K,
const SCALAR alpha, const TYPE *A, const int lda0,
const TYPE *B, const int ldb0, const SCALAR beta,
TYPE *C, const int ldc0)
{
size_t incA, incB, incC;
const size_t lda=lda0, ldb=ldb0, ldc=ldc0;
const size_t incK = ATL_MulByNB((size_t)K);
int N = N0;
int nMb, nNb, nKb, ib, jb, kb, jb2, h, i, j, k, n;
void *vA=NULL, *vC=NULL;
TYPE *pA, *pB, *pC;
MAT2BLK A2blk, B2blk;
PUTBLK putblk;
NBMM0 NBmm0;
nMb = ATL_DivByNB(M);
nNb = ATL_DivByNB(N);
nKb = ATL_DivByNB(K);
ib = M - ATL_MulByNB(nMb);
jb = N - ATL_MulByNB(nNb);
kb = K - ATL_MulByNB(nKb);
/*
* If K sufficiently large, write to temporary C as safety measure; otherwise
* write directly to C
*/
if (nKb < 12)
{
putblk = NULL;
pC = C;
if ( SCALAR_IS_ONE(beta) ) NBmm0 = NBmm_b1;
else if ( SCALAR_IS_ZERO(beta) ) NBmm0 = NBmm_b0;
else NBmm0 = NBmm_bX;
}
else
{
NBmm0 = NBmm_b0;
vC = malloc(ATL_Cachelen + ATL_MulBySize(NBNB));
if (!vC) return(-1);
pC = ATL_AlignPtr(vC);
if ( SCALAR_IS_ONE(beta) ) putblk = Mjoin(PATL,putblk_b1);
else if ( SCALAR_IS_ZERO(beta) ) putblk = Mjoin(PATL,putblk_b0);
else if ( SCALAR_IS_NONE(beta) ) putblk = Mjoin(PATL,putblk_bn1);
else putblk = Mjoin(PATL,putblk_bX);
}
/*
* Special case if we don't need to copy one or more input matrix
*/
if (K == NB && TB == AtlasNoTrans && ldb == NB && ATL_DataIsMinAligned(B))
{
if (lda == NB && TA == AtlasTrans && SCALAR_IS_ONE(alpha) &&
ATL_DataIsMinAligned(A))
{
i = NBNB;
pA = (TYPE *) A;
A = NULL;
A2blk = NULL;
incA = 0;
}
else
{
vA = malloc(ATL_Cachelen + ATL_MulBySize(incK));
if (!vA)
{
free(vC);
return(-1);
}
pA = ATL_AlignPtr(vA);
if (TA == AtlasNoTrans)
{
incA = NB;
if ( SCALAR_IS_ONE(alpha) ) A2blk = Mjoin(PATL,row2blkT_a1);
else A2blk = Mjoin(PATL,row2blkT_aX);
}
else
{
incA = ATL_MulByNB(lda);
if ( SCALAR_IS_ONE(alpha) ) A2blk = Mjoin(PATL,col2blk_a1);
else A2blk = Mjoin(PATL,col2blk_aX);
}
}
Mjoin(PATL,mmIJK2)(K, nMb, nNb, nKb, ib, jb, kb, alpha, A, lda, pA,
incA, A2blk, B, beta, C, ldc, pC, putblk, NBmm0);
if (vA) free(vA);
if (vC) free(vC);
return(0);
}
i = ATL_Cachelen + ATL_MulBySize(N*K + incK);
if (i <= ATL_MaxMalloc) vA = malloc(i);
if (!vA)
{
if (TA == AtlasNoTrans && TB == AtlasNoTrans)
{
if (vC) free(vC);
return(1);
}
if (jb) n = nNb + 1;
else n = nNb;
for (j=2; !vA; j++)
{
k = n / j;
if (k < 1) break;
if (k*j < n) k++;
h = ATL_Cachelen + ATL_MulBySize((k+1)*incK);
if (h <= ATL_MaxMalloc) vA = malloc(h);
}
if (!vA)
{
if (vC) free(vC);
return(-1);
}
n = ATL_MulByNB(k);
jb2 = 0;
}
else
{
jb2 = jb;
k = nNb;
n = N;
}
pA = ATL_AlignPtr(vA);
if (TB == AtlasNoTrans)
{
incB = ldb*n;
if ( SCALAR_IS_ONE(alpha) ) B2blk = Mjoin(PATL,col2blk2_a1);
else B2blk = Mjoin(PATL,col2blk2_aX);
}
else
{
incB = n;
if ( SCALAR_IS_ONE(alpha) ) B2blk = Mjoin(PATL,row2blkT2_a1);
else B2blk = Mjoin(PATL,row2blkT2_aX);
}
if (TA == AtlasNoTrans)
{
incA = NB;
A2blk = Mjoin(PATL,row2blkT_a1);
}
else
{
incA = ATL_MulByNB(lda);
A2blk = Mjoin(PATL,col2blk_a1);
}
incC = ldc*n;
pB = pA + incK;
do
{
if (TB == AtlasNoTrans) B2blk(K, n, B, ldb, pB, alpha);
else B2blk(n, K, B, ldb, pB, alpha);
Mjoin(PATL,mmIJK2)(K, nMb, k, nKb, ib, jb2, kb, alpha, A, lda, pA,
incA, A2blk, pB, beta, C, ldc, pC, putblk, NBmm0);
N -= n;
nNb -= k;
if (N < n)
{
jb2 = jb;
n = N;
k = nNb;
}
C += incC;
B += incB;
if (!putblk) pC = C;
}
while (N);
if (vC) free(vC);
free(vA);
return(0);
}
int Mjoin(PATL,mmIJK)(const enum ATLAS_TRANS TA, const enum ATLAS_TRANS TB,
const int M, const int N0, const int K,
const SCALAR alpha, const TYPE *A, const int lda,
const TYPE *B, const int ldb, const SCALAR beta,
TYPE *C, const int ldc)
{
int N = N0;
int nMb, nNb, nKb, ib, jb, kb, jb2, h, i, j, k, n, incA, incB, incC;
const int incK = ATL_MulByNB(K);
void *vA=NULL;
TYPE *pA, *pB;
MAT2BLK A2blk, B2blk;
MATSCAL gescal;
NBMM0 NBmm0;
nMb = ATL_DivByNB(M);
nNb = ATL_DivByNB(N);
nKb = ATL_DivByNB(K);
ib = M - ATL_MulByNB(nMb);
jb = N - ATL_MulByNB(nNb);
kb = K - ATL_MulByNB(nKb);
if (beta[1] == ATL_rzero)
{
gescal = NULL;
if (*beta == ATL_rone) NBmm0 = Mjoin(PATL,CNBmm_b1);
else if (*beta == ATL_rzero) NBmm0 = Mjoin(PATL,CNBmm_b0);
else NBmm0 = Mjoin(PATL,CNBmm_bX);
}
else
{
gescal = Mjoin(PATL,gescal_bX);
NBmm0 = Mjoin(PATL,CNBmm_b1);
}
i = ATL_Cachelen + ATL_MulBySize(N*K + incK);
if (i <= ATL_MaxMalloc) vA = malloc(i);
if (!vA)
{
if (TA == AtlasNoTrans && TB == AtlasNoTrans) return(1);
if (jb) n = nNb + 1;
else n = nNb;
for (j=2; !vA; j++)
{
k = n / j;
if (k < 1) break;
if (k*j < n) k++;
h = ATL_Cachelen + ATL_MulBySize((k+1)*incK);
if (h <= ATL_MaxMalloc) vA = malloc(h);
}
if (!vA) return(-1);
n = ATL_MulByNB(k);
jb2 = 0;
}
else
{
jb2 = jb;
k = nNb;
n = N;
}
pA = ATL_AlignPtr(vA);
if (TB == AtlasNoTrans)
{
incB = ldb*n<<1;
if (alpha[1] == ATL_rzero)
{
if (*alpha == ATL_rone) B2blk = Mjoin(PATL,col2blk2_a1);
else B2blk = Mjoin(PATL,col2blk2_aXi0);
}
else B2blk = Mjoin(PATL,col2blk2_aX);
}
else if (TB == AtlasConjTrans)
{
incB = n<<1;
if (alpha[1] == ATL_rzero)
{
if (*alpha == ATL_rone) B2blk = Mjoin(PATL,row2blkC2_a1);
else B2blk = Mjoin(PATL,row2blkC2_aXi0);
}
else B2blk = Mjoin(PATL,row2blkC2_aX);
}
else
{
incB = n<<1;
if (alpha[1] == ATL_rzero)
{
if (*alpha == ATL_rone) B2blk = Mjoin(PATL,row2blkT2_a1);
else B2blk = Mjoin(PATL,row2blkT2_aXi0);
}
else B2blk = Mjoin(PATL,row2blkT2_aX);
}
if (TA == AtlasNoTrans)
{
incA = NB<<1;
A2blk = Mjoin(PATL,row2blkT_a1);
}
else if (TA == AtlasConjTrans)
{
incA = ATL_MulByNB(lda)<<1;
A2blk = Mjoin(PATL,col2blkConj_a1);
}
else
{
incA = ATL_MulByNB(lda)<<1;
A2blk = Mjoin(PATL,col2blk_a1);
}
incC = ldc*n<<1;
pB = pA + (incK<<1);
do
{
if (TB == AtlasNoTrans) B2blk(K, n, B, ldb, pB, alpha);
else B2blk(n, K, B, ldb, pB, alpha);
Mjoin(PATL,mmIJK2)(K, nMb, k, nKb, ib, jb2, kb, alpha, A, lda, pA,
incA, A2blk, pB, beta, C, ldc, gescal, NBmm0);
N -= n;
nNb -= k;
if (N < n)
{
jb2 = jb;
n = N;
k = nNb;
}
C += incC;
B += incB;
}
while (N);
free(vA);
return(0);
}
