updated gemv kernel for armv6
This commit is contained in:
+23
-12
@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
#define I r12
|
||||
|
||||
#define FP_ZERO [fp, #-228]
|
||||
#define FP_ZERO_0 [fp, #-228]
|
||||
#define FP_ZERO_1 [fp, #-224]
|
||||
|
||||
#define N [fp, #-252 ]
|
||||
#define A [fp, #-256 ]
|
||||
|
||||
@@ -75,8 +79,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT_F2
|
||||
|
||||
vsub.f64 d2 , d2 , d2
|
||||
vsub.f64 d3 , d3 , d3
|
||||
fldd d2, FP_ZERO
|
||||
vmov.f64 d3 , d2
|
||||
|
||||
.endm
|
||||
|
||||
@@ -123,7 +127,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT_F1
|
||||
|
||||
vsub.f64 d2 , d2 , d2
|
||||
fldd d2, FP_ZERO
|
||||
vmov.f64 d3 , d2
|
||||
|
||||
.endm
|
||||
|
||||
@@ -160,8 +165,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT_S2
|
||||
|
||||
vsub.f64 d2 , d2 , d2
|
||||
vsub.f64 d3 , d3 , d3
|
||||
fldd d2, FP_ZERO
|
||||
vmov.f64 d3 , d2
|
||||
|
||||
.endm
|
||||
|
||||
@@ -224,7 +229,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT_S1
|
||||
|
||||
vsub.f64 d2 , d2 , d2
|
||||
fldd d2, FP_ZERO
|
||||
vmov.f64 d3 , d2
|
||||
|
||||
.endm
|
||||
|
||||
@@ -276,8 +282,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT_F2
|
||||
|
||||
vsub.f32 s2 , s2 , s2
|
||||
vsub.f32 s3 , s3 , s3
|
||||
flds s2 , FP_ZERO
|
||||
vmov.f32 s3 , s2
|
||||
|
||||
|
||||
.endm
|
||||
|
||||
@@ -321,7 +328,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT_F1
|
||||
|
||||
vsub.f32 s2 , s2 , s2
|
||||
flds s2 , FP_ZERO
|
||||
|
||||
.endm
|
||||
|
||||
@@ -356,8 +363,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT_S2
|
||||
|
||||
vsub.f32 s2 , s2 , s2
|
||||
vsub.f32 s3 , s3 , s3
|
||||
flds s2 , FP_ZERO
|
||||
vmov.f32 s3 , s2
|
||||
|
||||
.endm
|
||||
|
||||
@@ -418,7 +425,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT_S1
|
||||
|
||||
vsub.f32 s2 , s2 , s2
|
||||
flds s2 , FP_ZERO
|
||||
|
||||
.endm
|
||||
|
||||
@@ -488,6 +495,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
vstm r12, { s8 - s15 } // store floating point registers
|
||||
#endif
|
||||
|
||||
movs r12, #0
|
||||
str r12, FP_ZERO
|
||||
str r12, FP_ZERO_1
|
||||
|
||||
cmp M, #0
|
||||
ble gemvt_kernel_L999
|
||||
|
||||
|
||||
Reference in New Issue
Block a user