updated gemm_kernels for armv6

kernel/arm/cgemm_kernel_2x2_vfp.S

@@ -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

kernel/arm/dgemm_kernel_4x2_vfp.S

@@ -56,8 +56,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
+
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
@@ -85,7 +90,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -173,7 +178,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -233,7 +238,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d12, d8
 
 .endm
@@ -283,7 +288,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -338,7 +343,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -380,7 +385,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 
 .endm
 
@@ -433,6 +438,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// 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 * 8
 	str	r3, LDC

kernel/arm/sgemm_kernel_4x2_vfp.S

@@ -56,6 +56,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
 #define B	[fp, #4 ]
@@ -85,7 +89,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -161,7 +165,7 @@ 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		s12, s8
 	vmov.f32		s13, s8
@@ -221,7 +225,7 @@ 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		s12, s8
 
 .endm
@@ -271,7 +275,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -326,7 +330,7 @@ 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
 
 .endm
@@ -368,7 +372,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 
 .endm
 
@@ -421,6 +425,10 @@ 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, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/zgemm_kernel_2x2_vfp.S

@@ -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]
 
@@ -131,7 +135,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -383,7 +387,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -557,7 +561,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -724,7 +728,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -869,6 +873,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC