NEW SVE BLAS FILES ADDED

SVE implementation of gemv , scal , swap and rot BLAS routines files has been added

kernel/arm64/dgemv_kernel_sve.c

@@ -0,0 +1,41 @@
+#include "common.h"
+
+#include <arm_sve.h>
+
+#ifdef DOUBLE
+#define SVE_TYPE svfloat64_t
+#define SVE_ZERO svdup_f64(0.0)
+#define SVE_WHILELT svwhilelt_b64
+#define SVE_ALL svptrue_b64()
+#define SVE_WIDTH svcntd()
+#else
+#define SVE_TYPE svfloat32_t
+#define SVE_ZERO svdup_f32(0.0)
+#define SVE_WHILELT svwhilelt_b32
+#define SVE_ALL svptrue_b32()
+#define SVE_WIDTH svcntw()
+#endif
+
+static  FLOAT dgemv_kernel_sve(BLASLONG i, FLOAT  *x, BLASLONG lda, FLOAT  *y, BLASLONG incx, BLASLONG  n){
+        SVE_TYPE acc_a = SVE_ZERO;
+        SVE_TYPE acc_b = SVE_ZERO;
+
+        BLASLONG  sve_width = SVE_WIDTH;
+
+        for (BLASLONG j = 0; j < n; j += sve_width * 2) {
+                svbool_t pg_a = SVE_WHILELT(j, n);
+                svbool_t pg_b = SVE_WHILELT(j + sve_width, n);
+
+                SVE_TYPE x_vec_a = svld1(pg_a, &x[i*lda+j]);
+                SVE_TYPE y_vec_a = svld1(pg_a, &y[j*incx]);
+                SVE_TYPE x_vec_b = svld1(pg_b, &x[i*lda+j + sve_width]);
+                SVE_TYPE y_vec_b = svld1(pg_b, &y[j*incx + sve_width]);
+
+                acc_a = svmla_m(pg_a, acc_a, x_vec_a, y_vec_a);
+                acc_b = svmla_m(pg_b, acc_b, x_vec_b, y_vec_b);
+        }
+
+        return svaddv(SVE_ALL, acc_a) + svaddv(SVE_ALL, acc_b);
+
+}
+

kernel/arm64/gemv_n.c

@@ -0,0 +1,32 @@
+
+#include "common.h"
+
+// Some compilers will report feature support for SVE without the appropriate
+// header available
+#ifdef HAVE_SVE
+#if defined __has_include
+#if __has_include(<arm_sve.h>) && __ARM_FEATURE_SVE
+#define USE_SVE
+#endif
+#endif
+#endif
+
+#include "dgemv_kernel_sve.c"
+#include "dgemv_kernel_c.c"
+
+
+int CNAME(BLASLONG m,  BLASLONG n ,  BLASLONG dummy,  FLOAT  alpha,  FLOAT* a, BLASLONG lda ,  FLOAT *x, BLASLONG incx,  FLOAT *y, BLASLONG incy,  FLOAT *buffer){
+  
+  if (  incx == 1 && incy == 1){
+ // if(alpha!=1) for(BLASLONG i=0; i<n; ++i)X[i]=alpha*X[i]; 
+  for(BLASLONG i=0; i<n; ++i){
+  // Y[i*incy]+= dgemv_kernel_sve(i,A,lda,X,incx,n);
+  y[i]+= dgemv_kernel_sve(i,lda,a,m,x,alpha,n);
+ }
+  }
+//  BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+  else  dgemv_kernel_c( m,  n, dummy,  alpha, a, lda,  x,  incx, y,  incy, buffer );
+	
+ return 0;
+}
+

kernel/arm64/rot.c

@@ -0,0 +1,19 @@
+#include "common.h"
+
+#include "rot_kernel_sve.c"
+#include "rot_kernel_c.c"
+
+int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT c, FLOAT s)
+{
+  
+       
+        if ( n <= 0 )  return(0);
+        if ( inc_x == 1 && inc_y==1)
+	   rot_kernel_sve( n, x, y, c, s);
+        else 
+	   rot_kernel_c ( n, x, inc_x, y, inc_y, c, s);	
+       
+       	return(0);
+
+}
+

kernel/arm64/rot_kernel_c.c

@@ -0,0 +1,25 @@
+#include "common.h"
+
+static int rot_kernel_c(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT c, FLOAT s)
+{
+        BLASLONG i=0;
+        BLASLONG ix=0,iy=0;
+        FLOAT temp;
+
+        if ( n <= 0     )  return(0);
+
+        while(i < n)
+        {
+                temp   = c*x[ix] + s*y[iy] ;
+                y[iy]  = c*y[iy] - s*x[ix] ;
+                x[ix]  = temp ;
+
+                ix += inc_x ;
+                iy += inc_y ;
+                i++ ;
+
+        }
+        return(0);
+
+}
+

kernel/arm64/rot_kernel_sve.c

@@ -0,0 +1,38 @@
+#include "common.h"
+
+#include <arm_sve.h>
+
+#ifdef DOUBLE
+#define SVE_TYPE svfloat64_t
+#define SVE_ZERO svdup_f64(0.0)
+#define SVE_WHILELT svwhilelt_b64
+#define SVE_ALL svptrue_b64()
+#define SVE_WIDTH svcntd()
+#else
+#define SVE_TYPE svfloat32_t
+#define SVE_ZERO svdup_f32(0.0)
+#define SVE_WHILELT svwhilelt_b32
+#define SVE_ALL svptrue_b32()
+#define SVE_WIDTH svcntw()
+#endif
+
+static void  rot_kernel_sve(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT c, FLOAT s){
+       
+       for(int i=0; i<n; i+=SVE_WIDTH){     
+       	svbool_t pg =SVE_WHILELT((uint32_t)i,(uint32_t) n);
+   
+        SVE_TYPE  x_vec = svld1(pg, &x[i]);
+        SVE_TYPE  y_vec = svld1(pg, &y[i]);
+        SVE_TYPE  cx_vec=svmul_z(pg,x_vec,c);
+        SVE_TYPE  sy_vec=svmul_z(pg,y_vec,s);
+
+        SVE_TYPE  sx_vec=svmul_z(pg,x_vec,s);
+        SVE_TYPE  cy_vec=svmul_z(pg,y_vec,c);
+
+       svst1(pg,&x[i],svadd_z(pg,cx_vec,sy_vec));
+       svst1(pg,&y[i],svsub_z(pg,cy_vec,sx_vec));
+	
+        }
+}
+    
+

kernel/arm64/scal.c

@@ -0,0 +1,20 @@
+#include "common.h"
+
+#include "scal_kernel_sve.c"
+#include "scal_kernel_c.c"
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+        
+        if ( (n <= 0) || (inc_x <= 0))
+                return(0);
+         if (inc_x == 1)
+		 scal_kernel_sve( n, x, da);
+	 else
+		 scal_kernel_c(n,dummy0,dummy1,da,x,inc_x,y,inc_y,dummy,dummy2);
+
+        
+        return 0;
+
+}
+

kernel/arm64/scal_kernel_c.c

@@ -0,0 +1,23 @@
+
+#include "common.h"
+
+static int scal_kernel_c(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+         BLASLONG i=0,j=0;
+     
+        while(j < n)
+        {
+
+                if ( da == 0.0 )
+                        x[i]=0.0;
+                else
+                        x[i] = da * x[i] ;
+
+                i += inc_x ;
+                j++;
+
+        }
+        return 0;
+
+}
+

kernel/arm64/scal_kernel_sve.c

@@ -0,0 +1,28 @@
+#include "common.h"
+
+#include <arm_sve.h>
+
+#ifdef DOUBLE
+#define SVE_TYPE svfloat64_t
+#define SVE_ZERO svdup_f64(0.0)
+#define SVE_WHILELT svwhilelt_b64
+#define SVE_ALL svptrue_b64()
+#define SVE_WIDTH svcntd()
+#else
+#define SVE_TYPE svfloat32_t
+#define SVE_ZERO svdup_f32(0.0)
+#define SVE_WHILELT svwhilelt_b32
+#define SVE_ALL svptrue_b32()
+#define SVE_WIDTH svcntw()
+#endif
+static int scal_kernel_sve(int n, FLOAT *x, FLOAT da)
+{
+      for (int i = 0; i < n; i += SVE_WIDTH){
+        svbool_t pg = SVE_WHILELT(i, n);
+        SVE_TYPE  x_vec = svld1(pg, &x[i]);
+        SVE_TYPE  result= svmul_z(pg,x_vec,da);
+        svst1(pg,&x[i],result);
+    }
+   return (0);
+}
+

kernel/arm64/swap.c

@@ -0,0 +1,22 @@
+#include "common.h"
+
+
+#ifdef HAVE_SVE
+#if defined __has_include
+#if __has_include(<arm_sve.h>) && __ARM_FEATURE_SVE
+#define USE_SVE
+#endif
+#endif
+#endif
+
+#include "swap_kernel_sve.c"
+
+//(BLASLONG,  BLASLONG,  BLASLONG,  float,  float *, BLASLONG,  float *, BLASLONG,  float *, BLASLONG)
+//int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT dummy3, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT  dummy3, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT* dummy, BLASLONG dummy2)
+{
+      swap_kernel_sve(n, x,inc_x, y, inc_y);
+	return 0;
+
+}
+

kernel/arm64/swap_kernel_sve.c

@@ -0,0 +1,37 @@
+#include "common.h"
+
+#include <arm_sve.h>
+
+#ifdef DOUBLE
+#define SVE_TYPE svfloat64_t
+#define SVE_ZERO svdup_f64(0.0)
+#define SVE_WHILELT svwhilelt_b64
+#define SVE_ALL svptrue_b64()
+#define SVE_WIDTH svcntd()
+#else
+#define SVE_TYPE svfloat32_t
+#define SVE_ZERO svdup_f32(0.0)
+#define SVE_WHILELT svwhilelt_b32
+#define SVE_ALL svptrue_b32()
+#define SVE_WIDTH svcntw()
+#endif
+
+static int  swap_kernel_sve(BLASLONG n, FLOAT *x,BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) {
+          BLASLONG sve_width = SVE_WIDTH;
+
+        for (BLASLONG i = 0; i < n; i += sve_width * 2) {
+                svbool_t pg_a = SVE_WHILELT(i, n);
+                svbool_t pg_b = SVE_WHILELT((i + sve_width), n);
+                SVE_TYPE x_vec_a = svld1(pg_a, &x[i]);
+                SVE_TYPE y_vec_a = svld1(pg_a, &y[i]);
+                SVE_TYPE x_vec_b = svld1(pg_b, &x[i + sve_width]);
+                SVE_TYPE y_vec_b = svld1(pg_b, &y[i + sve_width]);
+
+        svst1(pg_a, &x[i], y_vec_a);
+        svst1(pg_a, &y[i], x_vec_a);
+	svst1(pg_b, &x[i+sve_width], y_vec_b);
+        svst1(pg_b, &y[i+sve_width], x_vec_b);
+             }
+	return 0;
+}
+