updated gemm_kernels for armv6
This commit is contained in:
@@ -57,6 +57,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
#define N [fp, #-260 ]
|
||||
#define K [fp, #-264 ]
|
||||
|
||||
#define FP_ZERO [fp, #-240]
|
||||
#define FP_ZERO_0 [fp, # -240]
|
||||
#define FP_ZERO_1 [fp, # -236]
|
||||
|
||||
#define ALPHA_I [fp, #-272]
|
||||
#define ALPHA_R [fp, #-280]
|
||||
|
||||
@@ -138,7 +142,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT2x2
|
||||
|
||||
vsub.f32 s8 , s8 , s8
|
||||
flds s8 , FP_ZERO
|
||||
vmov.f32 s9 , s8
|
||||
vmov.f32 s10, s8
|
||||
vmov.f32 s11, s8
|
||||
@@ -340,7 +344,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT1x2
|
||||
|
||||
vsub.f32 s8 , s8 , s8
|
||||
flds s8 , FP_ZERO
|
||||
vmov.f32 s9 , s8
|
||||
vmov.f32 s12, s8
|
||||
vmov.f32 s13, s8
|
||||
@@ -514,7 +518,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT2x1
|
||||
|
||||
vsub.f32 s8 , s8 , s8
|
||||
flds s8 , FP_ZERO
|
||||
vmov.f32 s9 , s8
|
||||
vmov.f32 s10, s8
|
||||
vmov.f32 s11, s8
|
||||
@@ -681,7 +685,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
.macro INIT1x1
|
||||
|
||||
vsub.f32 s8 , s8 , s8
|
||||
flds s8 , FP_ZERO
|
||||
vmov.f32 s9 , s8
|
||||
|
||||
.endm
|
||||
@@ -822,6 +826,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
sub r3, fp, #128
|
||||
vstm r3, { s8 - s15} // store floating point registers
|
||||
|
||||
movs r4, #0
|
||||
str r4, FP_ZERO
|
||||
str r4, FP_ZERO_1
|
||||
|
||||
ldr r3, OLD_LDC
|
||||
lsl r3, r3, #3 // ldc = ldc * 4 * 2
|
||||
str r3, LDC
|
||||
|
||||
Reference in New Issue
Block a user