floraachy/OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

changes for compatibility with Pathscale compiler

Browse Source
This commit is contained in:
wernsaar
2013-11-13 17:59:11 +01:00
parent 6216ab8a7e
commit f1db386211
5 changed files with 178 additions and 238 deletions
Download Patch File Download Diff File Expand all files Collapse all files
kernel/x86_64/zgemm_kernel_4x2_haswell.S
+47 -79
View File
@@ -26,7 +26,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
**********************************************************************************/
/********************************************************************************
* 2013/10/28 Saar
* 2013/11/13 Saar
* BLASTEST : OK
* CTEST : OK
* TEST : OK
@@ -137,43 +137,27 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
.macro VFMADDPD_R y0,y1,y2
vfmaddpd \y0,\y1,\y2,\y0
.endm
#define VFMADDPD_R( y0,y1,y2 ) vfmaddpd y0,y1,y2,y0
.macro VFMADDPD_I y0,y1,y2
vfmaddpd \y0,\y1,\y2,\y0
.endm
#define VFMADDPD_I( y0,y1,y2 ) vfmaddpd y0,y1,y2,y0
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
.macro VFMADDPD_R y0,y1,y2
vfnmaddpd \y0,\y1,\y2,\y0
.endm
#define VFMADDPD_R( y0,y1,y2 ) vfnmaddpd y0,y1,y2,y0
.macro VFMADDPD_I y0,y1,y2
vfmaddpd \y0,\y1,\y2,\y0
.endm
#define VFMADDPD_I( y0,y1,y2 ) vfmaddpd y0,y1,y2,y0
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
.macro VFMADDPD_R y0,y1,y2
vfmaddpd \y0,\y1,\y2,\y0
.endm
#define VFMADDPD_R( y0,y1,y2 ) vfmaddpd y0,y1,y2,y0
.macro VFMADDPD_I y0,y1,y2
vfnmaddpd \y0,\y1,\y2,\y0
.endm
#define VFMADDPD_I( y0,y1,y2 ) vfnmaddpd y0,y1,y2,y0
#else
.macro VFMADDPD_R y0,y1,y2
vfnmaddpd \y0,\y1,\y2,\y0
.endm
#define VFMADDPD_R( y0,y1,y2 ) vfnmaddpd y0,y1,y2,y0
.macro VFMADDPD_I y0,y1,y2
vfnmaddpd \y0,\y1,\y2,\y0
.endm
#define VFMADDPD_I( y0,y1,y2 ) vfnmaddpd y0,y1,y2,y0
#endif
@@ -181,43 +165,27 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
.macro VFMADDPD_R y0,y1,y2
vfmadd231pd \y1,\y2,\y0
.endm
#define VFMADDPD_R( y0,y1,y2 ) vfmadd231pd y1,y2,y0
.macro VFMADDPD_I y0,y1,y2
vfmadd231pd \y1,\y2,\y0
.endm
#define VFMADDPD_I( y0,y1,y2 ) vfmadd231pd y1,y2,y0
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
.macro VFMADDPD_R y0,y1,y2
vfnmadd231pd \y1,\y2,\y0
.endm
#define VFMADDPD_R( y0,y1,y2 ) vfnmadd231pd y1,y2,y0
.macro VFMADDPD_I y0,y1,y2
vfmadd231pd \y1,\y2,\y0
.endm
#define VFMADDPD_I( y0,y1,y2 ) vfmadd231pd y1,y2,y0
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
.macro VFMADDPD_R y0,y1,y2
vfmadd231pd \y1,\y2,\y0
.endm
#define VFMADDPD_R( y0,y1,y2 ) vfmadd231pd y1,y2,y0
.macro VFMADDPD_I y0,y1,y2
vfnmadd231pd \y1,\y2,\y0
.endm
#define VFMADDPD_I( y0,y1,y2 ) vfnmadd231pd y1,y2,y0
#else
.macro VFMADDPD_R y0,y1,y2
vfnmadd231pd \y1,\y2,\y0
.endm
#define VFMADDPD_R( y0,y1,y2 ) vfnmadd231pd y1,y2,y0
.macro VFMADDPD_I y0,y1,y2
vfnmadd231pd \y1,\y2,\y0
.endm
#define VFMADDPD_I( y0,y1,y2 ) vfnmadd231pd y1,y2,y0
#endif
@@ -233,16 +201,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
vbroadcastsd -8 * SIZE(BO, BI, SIZE), %ymm4
vbroadcastsd -7 * SIZE(BO, BI, SIZE), %ymm5
VFMADDPD_R %ymm8 ,%ymm4,%ymm0
VFMADDPD_R %ymm12,%ymm4,%ymm1
VFMADDPD_R( %ymm8 ,%ymm4,%ymm0 )
VFMADDPD_R( %ymm12,%ymm4,%ymm1 )
vbroadcastsd -6 * SIZE(BO, BI, SIZE), %ymm6
VFMADDPD_I %ymm9 ,%ymm5,%ymm0
VFMADDPD_I %ymm13,%ymm5,%ymm1
VFMADDPD_I( %ymm9 ,%ymm5,%ymm0 )
VFMADDPD_I( %ymm13,%ymm5,%ymm1 )
vbroadcastsd -5 * SIZE(BO, BI, SIZE), %ymm7
VFMADDPD_R %ymm10,%ymm6,%ymm0
VFMADDPD_R %ymm14,%ymm6,%ymm1
VFMADDPD_I %ymm11,%ymm7,%ymm0
VFMADDPD_I %ymm15,%ymm7,%ymm1
VFMADDPD_R( %ymm10,%ymm6,%ymm0 )
VFMADDPD_R( %ymm14,%ymm6,%ymm1 )
VFMADDPD_I( %ymm11,%ymm7,%ymm0 )
VFMADDPD_I( %ymm15,%ymm7,%ymm1 )
addq $4, BI
addq $8, %rax
@@ -337,17 +305,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
vmovups -8 * SIZE(AO, %rax, SIZE), %xmm0
vmovddup -8 * SIZE(BO, BI, SIZE), %xmm4
vmovups -6 * SIZE(AO, %rax, SIZE), %xmm1
VFMADDPD_R %xmm8,%xmm4,%xmm0
VFMADDPD_R %xmm12,%xmm4,%xmm1
VFMADDPD_R( %xmm8,%xmm4,%xmm0 )
VFMADDPD_R( %xmm12,%xmm4,%xmm1 )
vmovddup -7 * SIZE(BO, BI, SIZE), %xmm5
VFMADDPD_I %xmm9,%xmm5,%xmm0
VFMADDPD_I %xmm13,%xmm5,%xmm1
VFMADDPD_I( %xmm9,%xmm5,%xmm0 )
VFMADDPD_I( %xmm13,%xmm5,%xmm1 )
vmovddup -6 * SIZE(BO, BI, SIZE), %xmm6
VFMADDPD_R %xmm10,%xmm6,%xmm0
VFMADDPD_R %xmm14,%xmm6,%xmm1
VFMADDPD_R( %xmm10,%xmm6,%xmm0 )
VFMADDPD_R( %xmm14,%xmm6,%xmm1 )
vmovddup -5 * SIZE(BO, BI, SIZE), %xmm7
VFMADDPD_I %xmm11,%xmm7,%xmm0
VFMADDPD_I %xmm15,%xmm7,%xmm1
VFMADDPD_I( %xmm11,%xmm7,%xmm0 )
VFMADDPD_I( %xmm15,%xmm7,%xmm1 )
addq $4, BI
addq $4, %rax
.endm
@@ -441,12 +409,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
vmovups -8 * SIZE(AO, %rax, SIZE), %xmm0
vmovddup -8 * SIZE(BO, BI, SIZE), %xmm4
vmovddup -7 * SIZE(BO, BI, SIZE), %xmm5
VFMADDPD_R %xmm8,%xmm4,%xmm0
VFMADDPD_I %xmm9,%xmm5,%xmm0
VFMADDPD_R( %xmm8,%xmm4,%xmm0 )
VFMADDPD_I( %xmm9,%xmm5,%xmm0 )
vmovddup -6 * SIZE(BO, BI, SIZE), %xmm6
vmovddup -5 * SIZE(BO, BI, SIZE), %xmm7
VFMADDPD_R %xmm10,%xmm6,%xmm0
VFMADDPD_I %xmm11,%xmm7,%xmm0
VFMADDPD_R( %xmm10,%xmm6,%xmm0 )
VFMADDPD_I( %xmm11,%xmm7,%xmm0 )
addq $4, BI
addq $2, %rax
.endm
@@ -513,10 +481,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
vmovups -4 * SIZE(AO, %rax, SIZE), %ymm1
vbroadcastsd -4 * SIZE(BO, BI, SIZE) , %ymm4
vbroadcastsd -3 * SIZE(BO, BI, SIZE) , %ymm5
VFMADDPD_R %ymm8 ,%ymm4,%ymm0
VFMADDPD_R %ymm12,%ymm4,%ymm1
VFMADDPD_I %ymm9 ,%ymm5,%ymm0
VFMADDPD_I %ymm13,%ymm5,%ymm1
VFMADDPD_R( %ymm8 ,%ymm4,%ymm0 )
VFMADDPD_R( %ymm12,%ymm4,%ymm1 )
VFMADDPD_I( %ymm9 ,%ymm5,%ymm0 )
VFMADDPD_I( %ymm13,%ymm5,%ymm1 )
addq $2, BI
addq $8, %rax
@@ -585,12 +553,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
.macro KERNEL2x1_SUB
vmovups -8 * SIZE(AO, %rax, SIZE), %xmm0
vmovddup -4 * SIZE(BO, BI, SIZE), %xmm4
VFMADDPD_R %xmm8,%xmm4,%xmm0
VFMADDPD_R( %xmm8,%xmm4,%xmm0 )
vmovups -6 * SIZE(AO, %rax, SIZE), %xmm1
VFMADDPD_R %xmm12,%xmm4,%xmm1
VFMADDPD_R( %xmm12,%xmm4,%xmm1 )
vmovddup -3 * SIZE(BO, BI, SIZE), %xmm5
VFMADDPD_I %xmm9,%xmm5,%xmm0
VFMADDPD_I %xmm13,%xmm5,%xmm1
VFMADDPD_I( %xmm9,%xmm5,%xmm0 )
VFMADDPD_I( %xmm13,%xmm5,%xmm1 )
addq $2, BI
addq $4, %rax
.endm
@@ -655,9 +623,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
.macro KERNEL1x1_SUB
vmovups -8 * SIZE(AO, %rax, SIZE), %xmm0
vmovddup -4 * SIZE(BO, BI, SIZE), %xmm4
VFMADDPD_R %xmm8,%xmm4,%xmm0
VFMADDPD_R( %xmm8,%xmm4,%xmm0 )
vmovddup -3 * SIZE(BO, BI, SIZE), %xmm5
VFMADDPD_I %xmm9,%xmm5,%xmm0
VFMADDPD_I( %xmm9,%xmm5,%xmm0 )
addq $2, BI
addq $2, %rax
.endm