Merge pull request #4039 from klho/develop

Bug fix and improvements for [z]imatcopy interface.
This commit is contained in:
Martin Kroeker 2023-05-14 10:51:24 +02:00 committed by GitHub
commit ccad94162a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 64 additions and 73 deletions

View File

@ -120,17 +120,20 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows,
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#ifdef NEW_IMATCOPY
if ( *lda == *ldb && *rows == *cols) {
if ( *lda == *ldb ) {
if ( order == BlasColMajor )
{
if ( trans == BlasNoTrans )
{
IMATCOPY_K_CN(*rows, *cols, *alpha, a, *lda );
return;
}
else
else if ( *rows == *cols )
{
IMATCOPY_K_CT(*rows, *cols, *alpha, a, *lda );
return;
}
}
else
@ -138,26 +141,23 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows,
if ( trans == BlasNoTrans )
{
IMATCOPY_K_RN(*rows, *cols, *alpha, a, *lda );
return;
}
else
else if ( *rows == *cols )
{
IMATCOPY_K_RT(*rows, *cols, *alpha, a, *lda );
return;
}
}
return;
}
#endif
if ( *lda > *ldb )
msize = (size_t)(*lda) * (*ldb) * sizeof(FLOAT);
else
msize = (size_t)(*ldb) * (*ldb) * sizeof(FLOAT);
msize = (size_t)(*rows) * (*cols) * sizeof(FLOAT);
b = malloc(msize);
if ( b == NULL )
{
printf("Memory alloc failed\n");
printf("Memory alloc failed in imatcopy\n");
exit(1);
}
@ -165,26 +165,26 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows,
{
if ( trans == BlasNoTrans )
{
OMATCOPY_K_CN(*rows, *cols, *alpha, a, *lda, b, *ldb );
OMATCOPY_K_CN(*rows, *cols, (FLOAT) 1.0 , b, *ldb, a, *ldb );
OMATCOPY_K_CN(*rows, *cols, *alpha, a, *lda, b, *rows );
OMATCOPY_K_CN(*rows, *cols, (FLOAT) 1.0 , b, *rows, a, *ldb );
}
else
{
OMATCOPY_K_CT(*rows, *cols, *alpha, a, *lda, b, *ldb );
OMATCOPY_K_CN(*cols, *rows, (FLOAT) 1.0, b, *ldb, a, *ldb );
OMATCOPY_K_CT(*rows, *cols, *alpha, a, *lda, b, *cols );
OMATCOPY_K_CN(*cols, *rows, (FLOAT) 1.0, b, *cols, a, *ldb );
}
}
else
{
if ( trans == BlasNoTrans )
{
OMATCOPY_K_RN(*rows, *cols, *alpha, a, *lda, b, *ldb );
OMATCOPY_K_RN(*rows, *cols, (FLOAT) 1.0, b, *ldb, a, *ldb );
OMATCOPY_K_RN(*rows, *cols, *alpha, a, *lda, b, *cols );
OMATCOPY_K_RN(*rows, *cols, (FLOAT) 1.0, b, *cols, a, *ldb );
}
else
{
OMATCOPY_K_RT(*rows, *cols, *alpha, a, *lda, b, *ldb );
OMATCOPY_K_RN(*cols, *rows, (FLOAT) 1.0, b, *ldb, a, *ldb );
OMATCOPY_K_RT(*rows, *cols, *alpha, a, *lda, b, *rows );
OMATCOPY_K_RN(*cols, *rows, (FLOAT) 1.0, b, *rows, a, *ldb );
}
}

View File

@ -125,27 +125,33 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows,
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#ifdef NEW_IMATCOPY
if (*lda == *ldb && *cols == *rows) {
if (*lda == *ldb ) {
if ( order == BlasColMajor )
{
if ( trans == BlasNoTrans )
{
IMATCOPY_K_CN(*rows, *cols, alpha[0], alpha[1], a, *lda );
return;
}
if ( trans == BlasConj )
{
IMATCOPY_K_CNC(*rows, *cols, alpha[0], alpha[1], a, *lda );
return;
}
if ( trans == BlasTrans )
if ( trans == BlasTrans && *rows == *cols )
{
IMATCOPY_K_CT(*rows, *cols, alpha[0], alpha[1], a, *lda );
return;
}
if ( trans == BlasTransConj )
if ( trans == BlasTransConj && *rows == *cols )
{
IMATCOPY_K_CTC(*rows, *cols, alpha[0], alpha[1], a, *lda );
return;
}
}
else
{
@ -153,67 +159,59 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows,
if ( trans == BlasNoTrans )
{
IMATCOPY_K_RN(*rows, *cols, alpha[0], alpha[1], a, *lda );
return;
}
if ( trans == BlasConj )
{
IMATCOPY_K_RNC(*rows, *cols, alpha[0], alpha[1], a, *lda );
return;
}
if ( trans == BlasTrans )
if ( trans == BlasTrans && *rows == *cols )
{
IMATCOPY_K_RT(*rows, *cols, alpha[0], alpha[1], a, *lda );
return;
}
if ( trans == BlasTransConj )
if ( trans == BlasTransConj && *rows == *cols )
{
IMATCOPY_K_RTC(*rows, *cols, alpha[0], alpha[1], a, *lda );
return;
}
}
return;
}
#endif
if ( *lda > *ldb )
msize = (size_t)(*lda) * (*ldb) * sizeof(FLOAT) * 2;
else
msize = (size_t)(*ldb) * (*ldb) * sizeof(FLOAT) * 2;
b = malloc(msize);
if ( b == NULL )
{
printf("Memory alloc failed in zimatcopy\n");
exit(1);
}
msize = (size_t)(*rows) * (*cols) * sizeof(FLOAT) * 2;
b = malloc(msize);
if ( b == NULL )
{
printf("Memory alloc failed in zimatcopy\n");
exit(1);
}
if ( order == BlasColMajor )
{
if ( trans == BlasNoTrans )
{
OMATCOPY_K_CN(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *ldb );
OMATCOPY_K_CN(*rows, *cols, (FLOAT) 1.0, (FLOAT) 0.0 , b, *ldb, a, *ldb );
free(b);
return;
OMATCOPY_K_CN(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *rows );
OMATCOPY_K_CN(*rows, *cols, (FLOAT) 1.0, (FLOAT) 0.0 , b, *rows, a, *ldb );
}
if ( trans == BlasConj )
else if ( trans == BlasConj )
{
OMATCOPY_K_CNC(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *ldb );
OMATCOPY_K_CN(*rows, *cols, (FLOAT) 1.0, (FLOAT) 0.0 , b, *ldb, a, *ldb );
free(b);
return;
OMATCOPY_K_CNC(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *rows );
OMATCOPY_K_CN(*rows, *cols, (FLOAT) 1.0, (FLOAT) 0.0 , b, *rows, a, *ldb );
}
if ( trans == BlasTrans )
else if ( trans == BlasTrans )
{
OMATCOPY_K_CT(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *ldb );
OMATCOPY_K_CN(*cols, *rows, (FLOAT) 1.0, (FLOAT) 0.0 , b, *ldb, a, *ldb );
free(b);
return;
OMATCOPY_K_CT(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *cols );
OMATCOPY_K_CN(*cols, *rows, (FLOAT) 1.0, (FLOAT) 0.0 , b, *cols, a, *ldb );
}
if ( trans == BlasTransConj )
else if ( trans == BlasTransConj )
{
OMATCOPY_K_CTC(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *ldb );
OMATCOPY_K_CN(*cols, *rows, (FLOAT) 1.0, (FLOAT) 0.0 , b, *ldb, a, *ldb );
free(b);
return;
OMATCOPY_K_CTC(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *cols );
OMATCOPY_K_CN(*cols, *rows, (FLOAT) 1.0, (FLOAT) 0.0 , b, *cols, a, *ldb );
}
}
@ -222,34 +220,27 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows,
if ( trans == BlasNoTrans )
{
OMATCOPY_K_RN(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *ldb );
OMATCOPY_K_RN(*rows, *cols, (FLOAT) 1.0, (FLOAT) 0.0 , b, *ldb, a, *ldb );
free(b);
return;
OMATCOPY_K_RN(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *cols );
OMATCOPY_K_RN(*rows, *cols, (FLOAT) 1.0, (FLOAT) 0.0 , b, *cols, a, *ldb );
}
if ( trans == BlasConj )
else if ( trans == BlasConj )
{
OMATCOPY_K_RNC(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *ldb );
OMATCOPY_K_RN(*rows, *cols, (FLOAT) 1.0, (FLOAT) 0.0 , b, *ldb, a, *ldb );
free(b);
return;
OMATCOPY_K_RNC(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *cols );
OMATCOPY_K_RN(*rows, *cols, (FLOAT) 1.0, (FLOAT) 0.0 , b, *cols, a, *ldb );
}
if ( trans == BlasTrans )
else if ( trans == BlasTrans )
{
OMATCOPY_K_RT(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *ldb );
OMATCOPY_K_RN(*cols, *rows, (FLOAT) 1.0, (FLOAT) 0.0 , b, *ldb, a, *ldb );
free(b);
return;
OMATCOPY_K_RT(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *rows );
OMATCOPY_K_RN(*cols, *rows, (FLOAT) 1.0, (FLOAT) 0.0 , b, *rows, a, *ldb );
}
if ( trans == BlasTransConj )
else if ( trans == BlasTransConj )
{
OMATCOPY_K_RTC(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *ldb );
OMATCOPY_K_RN(*cols, *rows, (FLOAT) 1.0, (FLOAT) 0.0 , b, *ldb, a, *ldb );
free(b);
return;
OMATCOPY_K_RTC(*rows, *cols, alpha[0], alpha[1], a, *lda, b, *rows );
OMATCOPY_K_RN(*cols, *rows, (FLOAT) 1.0, (FLOAT) 0.0 , b, *rows, a, *ldb );
}
}
free(b);
return;