arm: add softfp support in sgemm/strmm vfp kernels

6 files changed, 52 insertions, 22 deletions

diff --git a/kernel/arm/KERNEL.ARMV6 b/kernel/arm/KERNEL.ARMV6
index 022a93183..18d9869de 100644
--- a/kernel/arm/KERNEL.ARMV6
+++ b/kernel/arm/KERNEL.ARMV6
@@ -73,6 +73,7 @@ CGEMVTKERNEL = cgemv_t_vfp.S
 ZGEMVTKERNEL = zgemv_t_vfp.S
 
 SGEMMKERNEL    = ../generic/gemmkernel_4x2.c
+SGEMMKERNEL    = sgemm_kernel_4x2_vfp.S
 ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
 SGEMMINCOPY    = sgemm_ncopy_4_vfp.S
 SGEMMITCOPY    = sgemm_tcopy_4_vfp.S
@@ -97,6 +98,7 @@ DGEMMONCOPYOBJ = dgemm_oncopy.o
 DGEMMOTCOPYOBJ = dgemm_otcopy.o
 
 STRMMKERNEL    = ../generic/trmmkernel_4x2.c
+STRMMKERNEL    = strmm_kernel_4x2_vfp.S
 DTRMMKERNEL    = ../generic/trmmkernel_4x2.c
 
 CGEMMONCOPY    = cgemm_ncopy_2_vfp.S
@@ -111,12 +113,10 @@ ZGEMMOTCOPYOBJ = zgemm_otcopy.o
 
 ifeq ($(ARM_ABI),hard)
 
-STRMMKERNEL    = strmm_kernel_4x2_vfp.S
 DTRMMKERNEL    = dtrmm_kernel_4x2_vfp.S
 CTRMMKERNEL    = ctrmm_kernel_2x2_vfp.S
 ZTRMMKERNEL    = ztrmm_kernel_2x2_vfp.S
 
-SGEMMKERNEL    = sgemm_kernel_4x2_vfp.S
 DGEMMKERNEL    = dgemm_kernel_4x2_vfp.S
 CGEMMKERNEL    = cgemm_kernel_2x2_vfp.S
 ZGEMMKERNEL    = zgemm_kernel_2x2_vfp.S
diff --git a/kernel/arm/KERNEL.ARMV7 b/kernel/arm/KERNEL.ARMV7
index 0872cb8cd..e2044133d 100644
--- a/kernel/arm/KERNEL.ARMV7
+++ b/kernel/arm/KERNEL.ARMV7
@@ -11,7 +11,7 @@ DGEMVNKERNEL = gemv_n_vfpv3.S
 STRMMKERNEL    = ../generic/trmmkernel_4x4.c
 DTRMMKERNEL    = ../generic/trmmkernel_4x4.c
 
-SGEMMKERNEL    = ../generic/gemmkernel_4x4.c
+SGEMMKERNEL    =  sgemm_kernel_4x4_vfpv3.S
 SGEMMONCOPY    =  sgemm_ncopy_4_vfp.S
 SGEMMOTCOPY    =  sgemm_tcopy_4_vfp.S
 SGEMMONCOPYOBJ =  sgemm_oncopy.o
@@ -30,7 +30,6 @@ DTRMMKERNEL  =  dtrmm_kernel_4x4_vfpv3.S
 CTRMMKERNEL  =  ctrmm_kernel_2x2_vfpv3.S
 ZTRMMKERNEL  =  ztrmm_kernel_2x2_vfpv3.S
 
-SGEMMKERNEL    =  sgemm_kernel_4x4_vfpv3.S
 DGEMMKERNEL    =  dgemm_kernel_4x4_vfpv3.S
 
 CGEMMKERNEL    = cgemm_kernel_2x2_vfpv3.S
diff --git a/kernel/arm/sgemm_kernel_4x2_vfp.S b/kernel/arm/sgemm_kernel_4x2_vfp.S
index e8b44b742..1f21e5a1f 100644
--- a/kernel/arm/sgemm_kernel_4x2_vfp.S
+++ b/kernel/arm/sgemm_kernel_4x2_vfp.S
@@ -62,9 +62,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	r3
+#define OLD_A_SOFTFP	[fp, #4 ]
+#define B	[fp, #8 ]
+#define C	[fp, #12 ]
+#define OLD_LDC	[fp, #16 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -416,6 +424,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K
diff --git a/kernel/arm/sgemm_kernel_4x4_vfpv3.S b/kernel/arm/sgemm_kernel_4x4_vfpv3.S
index 86198ac90..6491d3571 100644
--- a/kernel/arm/sgemm_kernel_4x4_vfpv3.S
+++ b/kernel/arm/sgemm_kernel_4x4_vfpv3.S
@@ -58,14 +58,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define	OLD_M	r0
 #define	OLD_N	r1
 #define	OLD_K	r2
-
-#ifdef ARM_SOFTFP_ABI
-#define OLD_ALPHA r3
-//#define OLD_A	
-#else //hard
 #define	OLD_A	r3
 #define OLD_ALPHA s0
-#endif
 
 /******************************************************
 * [fp, #-128] - [fp, #-64] is reserved
@@ -77,10 +71,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define M	[fp, #-256 ]
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
-
-#ifndef ARM_SOFTFP_ABI
 #define A	[fp, #-268 ]
-#endif
 
 #define FP_ZERO [fp, #-240]
 #define FP_ZERO_0 [fp, #-240]
@@ -88,17 +79,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA   [fp, #-280]
 
-#ifdef ARM_SOFTFP_ABI
-#define A	[fp, #4 ]
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	r3
+#define OLD_A_SOFTFP	[fp, #4 ]
 #define B	[fp, #8 ]
 #define C	[fp, #12 ]
 #define OLD_LDC	[fp, #16 ]
-#else //hard
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #endif
-	
+
 #define I	r0
 #define J	r1
 #define L	r2
@@ -867,16 +859,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
+
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K
-	
-#ifdef ARM_SOFTFP_ABI
-	str	OLD_ALPHA, ALPHA
-#else //hard
 	str	OLD_A, A
 	vstr	OLD_ALPHA, ALPHA
-#endif
+
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
diff --git a/kernel/arm/strmm_kernel_4x2_vfp.S b/kernel/arm/strmm_kernel_4x2_vfp.S
index 8f97644ec..635b1dd13 100644
--- a/kernel/arm/strmm_kernel_4x2_vfp.S
+++ b/kernel/arm/strmm_kernel_4x2_vfp.S
@@ -65,10 +65,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-276 ]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	r3
+#define OLD_A_SOFTFP	[fp, #4 ]
+#define B	[fp, #8 ]
+#define OLD_C	[fp, #12 ]
+#define OLD_LDC	[fp, #16 ]
+#define OFFSET  [fp, #20 ]
+#else
 #define B	[fp, #4 ]
 #define OLD_C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -395,6 +404,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K
diff --git a/kernel/arm/strmm_kernel_4x4_vfpv3.S b/kernel/arm/strmm_kernel_4x4_vfpv3.S
index 0dd03ac85..e24d24eba 100644
--- a/kernel/arm/strmm_kernel_4x4_vfpv3.S
+++ b/kernel/arm/strmm_kernel_4x4_vfpv3.S
@@ -64,10 +64,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	r3
+#define OLD_A_SOFTFP	[fp, #4 ]
+#define B	[fp, #8 ]
+#define C	[fp, #12 ]
+#define OLD_LDC	[fp, #16 ]
+#define OFFSET  [fp, #20 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -782,6 +791,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K