floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Use vec_vsx_ld/st to fix misaligned accesses flagged by asan

This commit is contained in:
Martin Kroeker 2020-07-16 23:32:54 +02:00 committed by GitHub
parent d8e2edfc20
commit 0033f8be0d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 93 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

96
kernel/power/saxpy.c
View File

@ -28,6 +28,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"
#define offset_0 0
#define offset_1 16
#define offset_2 32
#define offset_3 48
#define offset_4 64
#define offset_5 80
#define offset_6 96
#define offset_7 112
#define offset_8 128
#define offset_9 144
#define offset_10 160
#define offset_11 176
#define offset_12 192
#define offset_13 208
#define offset_14 224
#define offset_15 240
@ -37,12 +53,85 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
static void saxpy_kernel_64(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT alpha)
{
BLASLONG i = 0;
__vector float v_a = {alpha,alpha,alpha,alpha};
__vector float * v_y=(__vector float *)y;
__vector float * v_x=(__vector float *)x;
__vector float v_a __attribute((aligned(16))) = {alpha,alpha,alpha,alpha};
__vector float * vptr_y =(__vector float *)y;
__vector float * vptr_x =(__vector float *)x;
for(; i<n/4; i+=16){
register __vector float vy_0 = vec_vsx_ld( offset_0 ,vptr_y ) ;
register __vector float vy_1 = vec_vsx_ld( offset_1 ,vptr_y ) ;
register __vector float vy_2 = vec_vsx_ld( offset_2 ,vptr_y ) ;
register __vector float vy_3 = vec_vsx_ld( offset_3 ,vptr_y ) ;
register __vector float vy_4 = vec_vsx_ld( offset_4 ,vptr_y ) ;
register __vector float vy_5 = vec_vsx_ld( offset_5 ,vptr_y ) ;
register __vector float vy_6 = vec_vsx_ld( offset_6 ,vptr_y ) ;
register __vector float vy_7 = vec_vsx_ld( offset_7 ,vptr_y ) ;
register __vector float vy_8 = vec_vsx_ld( offset_8 ,vptr_y ) ;
register __vector float vy_9 = vec_vsx_ld( offset_9 ,vptr_y ) ;
register __vector float vy_10 = vec_vsx_ld( offset_10 ,vptr_y ) ;
register __vector float vy_11 = vec_vsx_ld( offset_11 ,vptr_y ) ;
register __vector float vy_12 = vec_vsx_ld( offset_12 ,vptr_y ) ;
register __vector float vy_13 = vec_vsx_ld( offset_13 ,vptr_y ) ;
register __vector float vy_14 = vec_vsx_ld( offset_14 ,vptr_y ) ;
register __vector float vy_15 = vec_vsx_ld( offset_15 ,vptr_y ) ;
register __vector float vx_0 = vec_vsx_ld( offset_0 ,vptr_x ) ;
register __vector float vx_1 = vec_vsx_ld( offset_1 ,vptr_x ) ;
register __vector float vx_2 = vec_vsx_ld( offset_2 ,vptr_x ) ;
register __vector float vx_3 = vec_vsx_ld( offset_3 ,vptr_x ) ;
register __vector float vx_4 = vec_vsx_ld( offset_4 ,vptr_x ) ;
register __vector float vx_5 = vec_vsx_ld( offset_5 ,vptr_x ) ;
register __vector float vx_6 = vec_vsx_ld( offset_6 ,vptr_x ) ;
register __vector float vx_7 = vec_vsx_ld( offset_7 ,vptr_x ) ;
register __vector float vx_8 = vec_vsx_ld( offset_8 ,vptr_x ) ;
register __vector float vx_9 = vec_vsx_ld( offset_9 ,vptr_x ) ;
register __vector float vx_10 = vec_vsx_ld( offset_10 ,vptr_x ) ;
register __vector float vx_11 = vec_vsx_ld( offset_11 ,vptr_x ) ;
register __vector float vx_12 = vec_vsx_ld( offset_12 ,vptr_x ) ;
register __vector float vx_13 = vec_vsx_ld( offset_13 ,vptr_x ) ;
register __vector float vx_14 = vec_vsx_ld( offset_14 ,vptr_x ) ;
register __vector float vx_15 = vec_vsx_ld( offset_15 ,vptr_x ) ;
vy_0 += vx_0*v_a;
vy_1 += vx_1*v_a;
vy_2 += vx_2*v_a;
vy_3 += vx_3*v_a;
vy_4 += vx_4*v_a;
vy_5 += vx_5*v_a;
vy_6 += vx_6*v_a;
vy_7 += vx_7*v_a;
vy_8 += vx_8*v_a;
vy_9 += vx_9*v_a;
vy_10 += vx_10*v_a;
vy_11 += vx_11*v_a;
vy_12 += vx_12*v_a;
vy_13 += vx_13*v_a;
vy_14 += vx_14*v_a;
vy_15 += vx_15*v_a;
vec_vsx_st( vy_0, offset_0 ,vptr_y ) ;
vec_vsx_st( vy_1, offset_1 ,vptr_y ) ;
vec_vsx_st( vy_2, offset_2 ,vptr_y ) ;
vec_vsx_st( vy_3, offset_3 ,vptr_y ) ;
vec_vsx_st( vy_4, offset_4 ,vptr_y ) ;
vec_vsx_st( vy_5, offset_5 ,vptr_y ) ;
vec_vsx_st( vy_6, offset_6 ,vptr_y ) ;
vec_vsx_st( vy_7, offset_7 ,vptr_y ) ;
vec_vsx_st( vy_8, offset_8 ,vptr_y ) ;
vec_vsx_st( vy_9, offset_9 ,vptr_y ) ;
vec_vsx_st( vy_10, offset_10 ,vptr_y ) ;
vec_vsx_st( vy_11, offset_11 ,vptr_y ) ;
vec_vsx_st( vy_12, offset_12 ,vptr_y ) ;
vec_vsx_st( vy_13, offset_13 ,vptr_y ) ;
vec_vsx_st( vy_14, offset_14 ,vptr_y ) ;
vec_vsx_st( vy_15, offset_15 ,vptr_y ) ;
vptr_x+=16;
vptr_y+=16;
/*
v_y[i] += v_a * v_x[i];
v_y[i+1] += v_a * v_x[i+1];
v_y[i+2] += v_a * v_x[i+2];
@ -59,6 +148,7 @@ static void saxpy_kernel_64(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT alpha)
v_y[i+13] += v_a * v_x[i+13];
v_y[i+14] += v_a * v_x[i+14];
v_y[i+15] += v_a * v_x[i+15];
*/
}
}
#endif