void Mjoin(PATL,gpmm)
(const enum PACK_UPLO UA, const enum PACK_TRANS TA,
const enum PACK_UPLO UB, const enum PACK_TRANS TB, const enum PACK_UPLO UC,
const int M, const int N, const int K, const SCALAR alpha,
const TYPE *A, const int IA, const int JA, const int lda,
const TYPE *B, const int IB, const int JB, const int ldb,
const SCALAR beta, TYPE *C, const int IC, const int JC, const int ldc)
{
int j;
#ifdef CacheEdge
static const int CE_K = ((ATL_DivBySize(CacheEdge)-(NBNB SHIFT)) /
(NB*(NB+NB)))*NB;
#else
#define CE_K K
#endif
if (!M || !N) return;
if (!K || SCALAR_IS_ZERO(alpha))
{
for (j=0; j != N; j++)
Mjoin(PATL,scal)(M, beta, C+MindexP(UC,IC,JC+j,ldc), 1);
return;
}
/*
* Packed gpmm not yet implemented for complex,
* so die if not really a dense gemm
*/
#ifdef TCPLX
ATL_assert (UA == PackGen && UB == PackGen && UC == PackGen);
Mjoin(PATL,gemm)(TA, TB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);
#else
Mjoin(PATL,prankK)(UA, TA, UB, TB, M, N, K, CE_K, alpha,
A+MindexP(UA,IA,JA,lda), Mpld(UA,JA,lda), B+MindexP(UB,IB,JB,ldb),
Mpld(UB,JB,ldb), beta, UC, C+MindexP(UC,IB,JB,ldc), Mpld(UC,JC,ldc));
#endif
}
static void ATL_rk_recUT
(const enum PACK_UPLO UA, const enum PACK_TRANS TA,
const enum ATLAS_UPLO UC, const int CP,
const int N, const int K, const SCALAR alpha,
const TYPE *A, int lda, const SCALAR beta, TYPE *C, const int ldc)
/*
* For upper trans matrix, use recursion to reduce N until enough memory
* can be allocated
*/
{
int Nright, Nleft;
const enum PACK_UPLO UC2 = (CP ? UC : PackGen);
if (Mjoin(PATL,prk_kmm)(UC, UA, TA, N, K, alpha, A, lda, beta, CP, C, ldc))
{
Nleft = N >> 1;
#ifdef NB
if (Nleft > NB) Nleft = ATL_MulByNB(ATL_DivByNB(Nleft));
#endif
Nright = N - Nleft;
ATL_rk_recUT(UA, TA, UC, CP, Nleft, K, alpha, A, lda, beta, C, ldc);
Mjoin(PATL,gpmm)(PackGen, TA, PackGen, AtlasNoTrans, CP?PackUpper:PackGen,
Nleft, Nright, K, alpha, A, 0, 0, lda,
A+Nleft*(lda SHIFT), 0, 0, lda, beta,
C+MindexP(UC2,0,Nleft,ldc), 0, 0, Mpld(UC2,Nleft,ldc));
ATL_rk_recUT(UA, TA, UC, CP, Nright, K, alpha, A+(Nleft SHIFT)*lda, lda,
beta, C+MindexP(UC2,Nleft,Nleft,ldc), Mpld(UC2,Nleft,ldc));
}
}
void Mjoin(PATL,pcol2blkF)
(const int M, const int N, const SCALAR alpha, const TYPE *A, int lda,
const int ldainc, TYPE *V)
/*
* Copies entire MxN matrix to block major format
*/
{
int j, jb;
const int incV = ATL_MulByNB(M);
const enum PACK_UPLO UA = (ldainc == 1) ? PackUpper :
( (lda == -1) ? PackLower : PackGen );
void (*col2blk)(const int M, const int N, const SCALAR alpha, const TYPE *A,
int lda, const int ldainc, TYPE *V);
if (ldainc)
{
if (alpha == ATL_rone) col2blk = Mjoin(PATL,pcol2blk_a1);
else col2blk = Mjoin(PATL,pcol2blk_aX);
for (j=0; j < N; j += NB)
{
jb = N-j;
jb = Mmin(jb, NB);
col2blk(M, jb, alpha, A+MindexP(UA,0,j,lda), Mpld(UA,j,lda), ldainc,V);
V += incV;
}
}
else if (alpha == ATL_rone)
Mjoin(PATL,col2blk2_a1)(M, N, A, lda, V, alpha);
else
Mjoin(PATL,col2blk2_aX)(M, N, A, lda, V, alpha);
}
void Mjoin(PATL,prow2blkTF)(const int M, const int N, const SCALAR alpha,
const TYPE *A, int lda, const int ldainc, TYPE *V)
{
const int mb = Mmin(NB,M), nMb = ATL_DivByNB(M);
const int m = ATL_MulByNB(nMb), n = ATL_MulByNB(ATL_DivByNB(N));
const int nr = N - n, mr = M - m;
const int incVm = ATL_MulByNB(N), incVV = ATL_MulByNB(mr);
int i, j, ib, jb;
const enum PACK_UPLO UA = (ldainc == 1) ? PackUpper :
( (ldainc == -1) ? PackLower : PackGen );
TYPE *v, *vv = V+nMb*incVm;
void (*row2blk)(const int M, const int N, const TYPE alpha, const TYPE *A,
int lda, const int ldainc, TYPE *V);
if (ldainc)
{
if (alpha == ATL_rone) row2blk = ATL_prow2blk_KB_a1;
else row2blk = ATL_prow2blk_KB_aX;
for (j=0; j < n; j += NB)
{
for (v=V, i=0; i < m; i += NB, v += incVm)
row2blk(NB, NB, alpha, A+MindexP(UA,i,j,lda), Mpld(UA,j,lda),
ldainc, v);
if (mr)
{
row2blk(mr, NB, alpha, A+MindexP(UA,m,j,lda), Mpld(UA,j,lda),
ldainc, vv);
vv += incVV;
}
V += NBNB;
}
if (nr)
{
for (v=V, i=0; i < m; i += NB, v += incVm)
row2blk(NB, nr, alpha, A+MindexP(UA,i,n,lda), Mpld(UA,n,lda),
ldainc, v);
if (mr)
row2blk(mr, nr, alpha, A+MindexP(UA,m,n,lda), Mpld(UA,n,lda),
ldainc, vv);
}
}
else if (SCALAR_IS_ONE(alpha))
Mjoin(PATL,row2blkT2_a1)(M, N, A, lda, V, alpha);
else
Mjoin(PATL,row2blkT2_aX)(M, N, A, lda, V, alpha);
}
void Mjoin(pcol2blkF,_blk)
(const int blk, const int M, const int N, const SCALAR alpha, const TYPE *A,
int lda, const int ldainc, TYPE *V)
/*
* Copies entire MxN matrix to block major format
*/
{
int j, jb;
const int incV = blk*(M+M);
const enum PACK_UPLO UA = (ldainc == 1) ? PackUpper :
( (lda == -1) ? PackLower : PackGen );
void (*col2blk)(const int blk, const int M, const int N, const SCALAR alpha,
const TYPE *A, int lda, const int ldainc, TYPE *V);
#ifdef Conj_
if (alpha[1] == ATL_rzero)
{
if (*alpha == ATL_rone) col2blk = Mjoin(Mjoin(PATL,pcol2blkConj_a1),_blk);
else col2blk = Mjoin(Mjoin(PATL,pcol2blkConj_aXi0),_blk);
}
else col2blk = Mjoin(Mjoin(PATL,pcol2blkConj_aX),_blk);
#else
if (alpha[1] == ATL_rzero)
{
if (*alpha == ATL_rone) col2blk = Mjoin(Mjoin(PATL,pcol2blk_a1),_blk);
else col2blk = Mjoin(Mjoin(PATL,pcol2blk_aXi0),_blk);
}
else col2blk = Mjoin(Mjoin(PATL,pcol2blk_aX),_blk);
#endif
for (j=0; j < N; j += blk)
{
jb = N-j;
jb = Mmin(jb, blk);
col2blk(blk, M, jb, alpha, A+MindexP(UA,0,j,lda), Mpld(UA,j,lda),
ldainc, V);
V += incV;
}
}
int ATL_pmmJIK(const enum PACK_UPLO UA, const enum ATLAS_TRANS TA,
const enum PACK_UPLO UB, const enum ATLAS_TRANS TB,
const int M, const int N, const int K, const SCALAR alpha,
const TYPE *A, const int lda, const TYPE *B, const int ldb,
const SCALAR beta, const enum PACK_UPLO UC,
TYPE *C, const int ldc)
/*
* Special packed matmul, calls dense gemm kernel using at most
* K*NB + 2*NB*NB space. $B$ is copied only once, but $A$ is copied
* ceil(N/NB) times. However, $A$ should start in-cache for kernel call.
*/
{
const int nKb = ATL_DivByNB(K), kb = K - ATL_MulByNB(nKb);
const int incK = ATL_MulByNB(K);
const int ldainc = (UA == AtlasUpper) ? 1 : ((UA == AtlasLower) ? -1 : 0);
const int ldbinc = (UB == AtlasUpper) ? 1 : ((UB == AtlasLower) ? -1 : 0);
const int ldcinc = (UC == AtlasUpper) ? 1 : ((UC == AtlasLower) ? -1 : 0);
int ib, jb, i, j, k;
void *vC;
TYPE *pC, *pA, *pB;
NBMM0 pNBmm, pNBmm0;
vC = malloc(ATL_Cachelen + ATL_MulBySize(NBNB+NBNB+incK));
if (!vC) return(-1);
pC = ATL_AlignPtr(vC);
pA = pC + NBNB;
pB = pA + NBNB;
/*
* Loop over column panels of $B$
*/
for (j=0; j < N; j += NB)
{
jb = N - j;
jb = Mmin(jb, NB);
/*
* Copy column-panel of B to block-major storage
*/
if (alpha == 1.0)
{
if (TB == AtlasNoTrans)
ATL_pcol2blk(K, jb, alpha, B+MindexP(UB,0,j,ldb), Mpld(UB,j,ldb),
ldbinc, pB);
else /* TB == AtlasTrans */
ATL_prow2blkT(jb, K, alpha, B+MindexP(UB,j,0,ldb), ldb, ldbinc, pB);
}
else if (TB == AtlasNoTrans)
ATL_pcol2blk_aX(K, jb, alpha, B+MindexP(UB,0,j,ldb), Mpld(UB,j,ldb),
ldbinc, pB);
else /* TB == AtlasTrans */
ATL_prow2blkT_aX(jb, K, alpha, B+MindexP(UB,j,0,ldb), ldb, ldbinc, pB);
/*
* Loop over row-panels of A
*/
for (i=0; i < M; i += MB)
{
ib = M - i;
ib = Mmin(ib, MB);
if (jb != NB || ib != MB)
{
pNBmm0 = pNBmm = ATL_gNBmm;
if (ib != NB && jb != NB) Mjoin(PATL,gezero)(MB, NB, pC, MB);
}
else
{
pNBmm = NBmm;
pNBmm0 = NBmm_b0;
}
/*
* Handle full blocks of K
*/
if (nKb)
{
if (TA == AtlasNoTrans)
ATL_prow2blkT(ib, NB, 1.0, A+MindexP(UA,i,0,lda),
lda, ldainc, pA);
else
ATL_pcol2blk(NB, ib, 1.0, A+MindexP(UA,0,i,lda),
Mpld(UA,i,lda), ldainc, pA);
pNBmm0(ib, jb, NB, ATL_rone, pA, NB, pB, NB, ATL_rzero, pC, ib);
for (k=1; k != nKb; k++)
{
if (TA == AtlasNoTrans)
ATL_prow2blkT(ib, NB, 1.0, A+MindexP(UA,i,ATL_MulByNB(k),lda),
Mpld(UA,ATL_MulByNB(k),lda), ldainc, pA);
else
ATL_pcol2blk(NB, ib, 1.0, A+MindexP(UA,ATL_MulByNB(k),i,lda),
Mpld(UA,i,lda), ldainc, pA);
pNBmm(ib, jb, NB, ATL_rone, pA, NB, pB+jb*NB*k, NB,
ATL_rone, pC, ib);
}
if (kb)
{
if (TA == AtlasNoTrans)
ATL_prow2blkT(ib, kb, 1.0,
A+MindexP(UA,i,ATL_MulByNB(nKb),lda),
Mpld(UA,ATL_MulByNB(nKb),lda), ldainc, pA);
else
ATL_pcol2blk(kb, ib, 1.0,
A+MindexP(UA,ATL_MulByNB(nKb),i,lda),
Mpld(UA,i,lda), ldainc, pA);
ATL_gNBmm(ib, jb, kb, ATL_rone, pA, kb, pB+jb*NB*nKb, kb,
ATL_rone, pC, ib);
}
}
else if (kb)
{
Mjoin(PATL,gezero)(ib, jb, pC, ib);
if (TA == AtlasNoTrans)
ATL_prow2blkT(ib, kb, 1.0, A+MindexP(UA,i,0,lda),
lda, ldainc, pA);
else
ATL_pcol2blk(kb, ib, 1.0, A+MindexP(UA,0,i,lda),
Mpld(UA,i,lda), ldainc, pA);
ATL_gNBmm(ib, jb, kb, ATL_rone, pA, kb, pB, kb, ATL_rzero, pC, ib);
}
ATL_pputblk(ib, jb, pC, C+MindexP(UC,i,j,ldc), Mpld(UC,j,ldc),
ldcinc, beta);
}
}
free(vC);
return(0);
}
int ATL_pmmJIKF(const enum PACK_UPLO UA, const enum ATLAS_TRANS TA,
const enum PACK_UPLO UB, const enum ATLAS_TRANS TB,
const int M, const int N, const int K, const SCALAR alpha,
const TYPE *A, const int lda, const TYPE *B, const int ldb,
const SCALAR beta, const enum PACK_UPLO UC,
TYPE *C, const int ldc)
/*
* Special packed matmul, calls dense gemm kernel using at most
* M*K + K*NB + NB*NB space. If this exceeds ATL_pkMaxMalloc or fails,
* operates using at most 2*K*NB + NB*NB. If this fails, returns non-zero.
* If full space is malloced, both matrices are copied exactly once. If
* the smaller space is used, $A$ will be copied ceil(N/NB) times.
*/
{
const int nKb = ATL_DivByNB(K), kb = K - ATL_MulByNB(nKb);
const int incK = ATL_MulByNB(K);
const int ldainc = (UA == AtlasUpper) ? 1 : ((UA == AtlasLower) ? -1 : 0);
const int ldbinc = (UB == AtlasUpper) ? 1 : ((UB == AtlasLower) ? -1 : 0);
const int ldcinc = (UC == AtlasUpper) ? 1 : ((UC == AtlasLower) ? -1 : 0);
int ib, jb, i, j, k;
void *vC=NULL;
TYPE *pC, *pA, *pB, *pA0;
NBMM0 pNBmm, pNBmm0;
void (*A2blk)(const int M, const int N, const TYPE alpha, const TYPE *A,
int lda, const int ldainc, TYPE *V);
i = ATL_Cachelen + ATL_MulBySize(NBNB+ATL_MulByNB(K)+M*K);
if (i <= ATL_pkMaxMalloc) vC = malloc(i);
if (!vC)
{
vC = malloc(ATL_Cachelen + ATL_MulBySize(NBNB+ATL_MulByNB(K+K)));
if (TA == AtlasNoTrans) A2blk = ATL_prow2blkT;
else A2blk = ATL_pcol2blk;
}
else A2blk = NULL;
if (!vC) return(-1);
pC = ATL_AlignPtr(vC);
pB = pC + NBNB;
pA = pB + ATL_MulByNB(K);
/*
* If we've got the space, copy all of A up front
*/
if (!A2blk)
{
if (TA == AtlasNoTrans)
ATL_prow2blkTF(M, K, ATL_rone, A, lda, ldainc, pA);
else ATL_pcol2blkF(K, M, ATL_rone, A, lda, ldainc, pA);
pA -= ATL_MulByNB(K);
}
pA0 = pA;
/*
* Loop over column panels of $B$
*/
for (j=0; j < N; j += NB)
{
jb = N - j;
jb = Mmin(jb, NB);
/*
* Copy column-panel of B to block-major storage
*/
if (alpha == 1.0)
{
if (TB == AtlasNoTrans)
ATL_pcol2blk(K, jb, alpha, B+MindexP(UB,0,j,ldb), Mpld(UB,j,ldb),
ldbinc, pB);
else /* TB == AtlasTrans */
ATL_prow2blkT(jb, K, alpha, B+MindexP(UB,j,0,ldb), ldb, ldbinc, pB);
}
else if (TB == AtlasNoTrans)
ATL_pcol2blk_aX(K, jb, alpha, B+MindexP(UB,0,j,ldb), Mpld(UB,j,ldb),
ldbinc, pB);
else /* TB == AtlasTrans */
ATL_prow2blkT_aX(jb, K, alpha, B+MindexP(UB,j,0,ldb), ldb, ldbinc, pB);
/*
* Loop over row-panels of A
*/
for (i=0; i < M; i += MB)
{
ib = M - i;
ib = Mmin(ib, MB);
if (A2blk)
{
if (TA == AtlasNoTrans)
ATL_prow2blkT(ib, K, ATL_rone, A+MindexP(UA,i,0,lda), lda,
ldainc, pA);
else /* TA == AtlasTrans */
ATL_pcol2blk(K, ib, ATL_rone, A+MindexP(UA,0,i,lda),
Mpld(UA,i,lda), ldainc, pA);
}
else pA += ATL_MulByNB(K);
if (jb != NB || ib != MB)
{
pNBmm0 = pNBmm = ATL_gNBmm;
if (ib != NB && jb != NB) Mjoin(PATL,gezero)(MB, NB, pC, MB);
}
else
{
pNBmm = NBmm;
pNBmm0 = NBmm_b0;
}
/*
* Handle full blocks of K
*/
if (nKb)
{
pNBmm0(ib, jb, NB, ATL_rone, pA, NB, pB, NB, ATL_rzero, pC, ib);
for (k=1; k != nKb; k++)
{
pNBmm(ib, jb, NB, ATL_rone, pA+ib*NB*k, NB, pB+jb*NB*k, NB,
ATL_rone, pC, ib);
}
if (kb)
ATL_gNBmm(ib, jb, kb, ATL_rone, pA+ib*NB*nKb, kb,
pB+jb*NB*nKb, kb, ATL_rone, pC, ib);
}
else if (kb)
{
Mjoin(PATL,gezero)(ib, jb, pC, ib);
ATL_gNBmm(ib, jb, kb, ATL_rone, pA, kb, pB, kb, ATL_rzero, pC, ib);
}
ATL_pputblk(ib, jb, pC, C+MindexP(UC,i,j,ldc), Mpld(UC,j,ldc),
ldcinc, beta);
}
pA = pA0;
}
free(vC);
return(0);
}
void Mjoin(PATL,sprk_rK)
(const enum PACK_UPLO UA, const enum PACK_TRANS TA,
const enum ATLAS_UPLO UC, const int CP,
const int N, const int K, int R, const SCALAR alpha,
const TYPE *A, int lda, const SCALAR beta0, TYPE *C, const int ldc)
/*
* This routine does the packed symmetric rank-K update by doing ceil(K/R)
* rank-R updates of C. This primarily done for CacheEdge, but is also
* useful to auto-reduce R until enough workspace may be allocated.
*/
{
const enum PACK_UPLO UC2 = ((CP) ? UC : PackGen);
int k=0, kb, ierr;
const int ldainc = (UA == AtlasUpper) ? 1 : ((UA == AtlasLower) ? -1 : 0);
const int ldcinc = (UC2 == AtlasUpper) ? 1 : ((UC2 == AtlasLower) ? -1 : 0);
#ifdef TREAL
TYPE beta = beta0;
#else
TYPE beta[2];
*beta = *beta0;
beta[1] = beta0[1];
#endif
if (R < NB) R = NB<<4;
if ((K - R) < 2*NB) R = K;
do
{
kb = K - k;
if (kb - R < 2*NB) R = kb;
kb = Mmin(R, kb);
/*
* If we can't do the rank-R update, reduce R until we can, or R = 1
*/
ierr = Mjoin(PATL,prk_kmm)(UC, UA, TA, N, kb, alpha,
A, lda, beta, CP, C, ldc);
if (ierr && R <= NB*8)
{
if (UC == AtlasUpper)
{
if (TA == AtlasNoTrans)
ATL_rk_recUN(UA, TA, UC, CP, N, kb, alpha, A, lda, beta, C, ldc);
else
ATL_rk_recUT(UA, TA, UC, CP, N, kb, alpha, A, lda, beta, C, ldc);
}
else
{
if (TA == AtlasNoTrans)
ATL_rk_recLN(UA, TA, UC, CP, N, kb, alpha, A, lda, beta, C, ldc);
else
ATL_rk_recLT(UA, TA, UC, CP, N, kb, alpha, A, lda, beta, C, ldc);
}
ierr = 0;
}
if (ierr)
{
R = Mmin(NB*8, R>>1);
ATL_assert(R);
}
/*
* Subsequent updates use beta = 1
*/
else
{
#ifdef TREAL
beta = ATL_rone;
#else
*beta = ATL_rone;
beta[1] = ATL_rzero;
#endif
if (TA == AtlasNoTrans)
{
A += MindexP(UA, 0, R, lda);
lda = Mpld(UA, R, lda);
}
else A += MindexP(UA, R, 0, lda);
k += R;
}
}
