diff options
| author | Zhang Xianyi <traits.zhang@gmail.com> | 2015-10-13 04:46:08 +0800 |
|---|---|---|
| committer | Zhang Xianyi <traits.zhang@gmail.com> | 2015-10-13 04:46:08 +0800 |
| commit | 94b125255f35615a6ec2431651690ef1e5b174ac (patch) | |
| tree | 0dcd732c6d2c50c61208a2cd62e5cf52b2cfe9e3 /interface | |
| parent | 17ee2237c382e8ea3f9d3a8aa74aef4a1d12ff17 (diff) | |
| parent | 3684706a121f9d9e1ccfc4a2bbb98f698eb04514 (diff) | |
| download | openblas-94b125255f35615a6ec2431651690ef1e5b174ac.tar.gz openblas-94b125255f35615a6ec2431651690ef1e5b174ac.tar.bz2 openblas-94b125255f35615a6ec2431651690ef1e5b174ac.zip | |
Merge branch 'develop' into cmake
Conflicts:
driver/others/memory.c
Diffstat (limited to 'interface')
| -rw-r--r-- | interface/imatcopy.c | 35 | ||||
| -rw-r--r-- | interface/zimatcopy.c | 50 |
2 files changed, 82 insertions, 3 deletions
diff --git a/interface/imatcopy.c b/interface/imatcopy.c index 89f0ec823..f4309a85c 100644 --- a/interface/imatcopy.c +++ b/interface/imatcopy.c @@ -26,7 +26,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. *****************************************************************************/ /*********************************************************** - * 2014/06/10 Saar + * 2014-06-10 Saar + * 2015-09-07 grisuthedragon ***********************************************************/ #include <stdio.h> @@ -50,6 +51,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #undef malloc #undef free +/* Enables the New IMATCOPY code with inplace operation if lda == ldb */ +#define NEW_IMATCOPY + #ifndef CBLAS void NAME( char* ORDER, char* TRANS, blasint *rows, blasint *cols, FLOAT *alpha, FLOAT *a, blasint *lda, blasint *ldb) { @@ -75,7 +79,6 @@ void NAME( char* ORDER, char* TRANS, blasint *rows, blasint *cols, FLOAT *alpha, #else void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows, blasint ccols, FLOAT calpha, FLOAT *a, blasint clda, blasint cldb) { - char Order, Trans; int order=-1,trans=-1; blasint info = -1; FLOAT *b; @@ -117,6 +120,34 @@ 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 ) { + if ( order == BlasColMajor ) + { + if ( trans == BlasNoTrans ) + { + IMATCOPY_K_CN(*rows, *cols, *alpha, a, *lda ); + } + else + { + IMATCOPY_K_CT(*rows, *cols, *alpha, a, *lda ); + } + } + else + { + if ( trans == BlasNoTrans ) + { + IMATCOPY_K_RN(*rows, *cols, *alpha, a, *lda ); + } + else + { + IMATCOPY_K_RT(*rows, *cols, *alpha, a, *lda ); + } + } + return; + } + +#endif if ( *lda > *ldb ) msize = (*lda) * (*ldb) * sizeof(FLOAT); diff --git a/interface/zimatcopy.c b/interface/zimatcopy.c index 3f273cf13..b1e1d15dc 100644 --- a/interface/zimatcopy.c +++ b/interface/zimatcopy.c @@ -26,7 +26,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. *****************************************************************************/ /*********************************************************** - * 2014/06/10 Saar + * 2014-06-10 Saar + * 2015-09-07 grisuthedragon ***********************************************************/ #include <stdio.h> @@ -49,6 +50,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define BlasTransConj 2 #define BlasConj 3 +#define NEW_IMATCOPY #ifndef CBLAS void NAME( char* ORDER, char* TRANS, blasint *rows, blasint *cols, FLOAT *alpha, FLOAT *a, blasint *lda, blasint *ldb) @@ -124,6 +126,52 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows, return; } +#ifdef NEW_IMATCOPY + if (*lda == *ldb) { + if ( order == BlasColMajor ) + { + + if ( trans == BlasNoTrans ) + { + IMATCOPY_K_CN(*rows, *cols, alpha[0], alpha[1], a, *lda ); + } + if ( trans == BlasConj ) + { + IMATCOPY_K_CNC(*rows, *cols, alpha[0], alpha[1], a, *lda ); + } + if ( trans == BlasTrans ) + { + IMATCOPY_K_CT(*rows, *cols, alpha[0], alpha[1], a, *lda ); + } + if ( trans == BlasTransConj ) + { + IMATCOPY_K_CTC(*rows, *cols, alpha[0], alpha[1], a, *lda ); + } + } + else + { + + if ( trans == BlasNoTrans ) + { + IMATCOPY_K_RN(*rows, *cols, alpha[0], alpha[1], a, *lda ); + } + if ( trans == BlasConj ) + { + IMATCOPY_K_RNC(*rows, *cols, alpha[0], alpha[1], a, *lda ); + } + if ( trans == BlasTrans ) + { + IMATCOPY_K_RT(*rows, *cols, alpha[0], alpha[1], a, *lda ); + } + if ( trans == BlasTransConj ) + { + IMATCOPY_K_RTC(*rows, *cols, alpha[0], alpha[1], a, *lda ); + } + } + return; + } +#endif + if ( *lda > *ldb ) msize = (*lda) * (*ldb) * sizeof(FLOAT) * 2; else |