updated trmm kernels for armv7

kernel/arm/KERNEL.ARMV7

@@ -1,8 +1,3 @@
-SGEMVNKERNEL = ../arm/gemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-CGEMVNKERNEL = ../arm/zgemv_n.c
-CGEMVTKERNEL = ../arm/zgemv_t.c
-
 
 #################################################################################
 SAMAXKERNEL  = iamax_vfp.S
@@ -77,14 +72,14 @@ DSCALKERNEL  = scal.c
 CSCALKERNEL  = zscal.c
 ZSCALKERNEL  = zscal.c
 
-# BAD SGEMVNKERNEL = gemv_n_vfp.S
+SGEMVNKERNEL = gemv_n_vfp.S
 DGEMVNKERNEL = gemv_n_vfp.S
-#CGEMVNKERNEL = cgemv_n_vfp.S
+CGEMVNKERNEL = cgemv_n_vfp.S
 ZGEMVNKERNEL = zgemv_n_vfp.S
 
-# BAD SGEMVTKERNEL = gemv_t_vfp.S
+SGEMVTKERNEL = gemv_t_vfp.S
 DGEMVTKERNEL = gemv_t_vfp.S
-#CGEMVTKERNEL = cgemv_t_vfp.S
+CGEMVTKERNEL = cgemv_t_vfp.S
 ZGEMVTKERNEL = zgemv_t_vfp.S
 
 STRMMKERNEL  =  strmm_kernel_4x4_vfpv3.S
@@ -92,7 +87,6 @@ DTRMMKERNEL  =  dtrmm_kernel_4x4_vfpv3.S
 CTRMMKERNEL  =  ctrmm_kernel_2x2_vfpv3.S
 ZTRMMKERNEL  =  ztrmm_kernel_2x2_vfpv3.S
 
-#SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
 SGEMMKERNEL    =  sgemm_kernel_4x4_vfpv3.S
 SGEMMINCOPY    =
 SGEMMITCOPY    =

kernel/arm/ctrmm_kernel_2x2_vfpv3.S

@@ -59,6 +59,10 @@ 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, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -134,7 +138,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -351,7 +355,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -529,7 +533,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -706,7 +710,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s24, s16
 	vmov.f32		s25, s16
@@ -852,6 +856,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// 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/dtrmm_kernel_4x4_vfpv3.S

@@ -59,6 +59,11 @@ 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, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
+
 #define ALPHA	[fp, #-276 ]
 
 #define B	[fp, #4 ]
@@ -89,7 +94,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -386,7 +391,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -468,7 +473,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d20, d16
 	vmov.f64		d24, d16
 	vmov.f64		d28, d16
@@ -527,7 +532,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -601,7 +606,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -656,7 +661,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d20, d16
 
 .endm
@@ -699,7 +704,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -753,7 +758,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 
 .endm
@@ -794,7 +799,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 
 .endm
 
@@ -850,6 +855,10 @@ 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/strmm_kernel_4x4_vfpv3.S

@@ -58,6 +58,10 @@ 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 ]
@@ -88,7 +92,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -322,7 +326,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -405,7 +409,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s20, s16
 	vmov.f32		s24, s16
 	vmov.f32		s28, s16
@@ -464,7 +468,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -538,7 +542,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -593,7 +597,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s20, s16
 
 .endm
@@ -636,7 +640,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -690,7 +694,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 
 .endm
@@ -731,7 +735,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 
 .endm
 
@@ -787,6 +791,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// 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/ztrmm_kernel_2x2_vfpv3.S

@@ -59,6 +59,10 @@ 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, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -134,7 +138,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -388,7 +392,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -566,7 +570,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -743,7 +747,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d24, d16
 	vmov.f64		d25, d16
@@ -889,6 +893,10 @@ 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