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

Optimize the performance of dot by using universal intrinsics in X86/ARM

Browse Source
This commit is contained in:
Qiyu8
2020-09-28 20:36:53 +08:00
parent 0d98ce202c
commit 1b1a757f5f
7 changed files with 177 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
kernel/generic/dot.c
View File

@@ -47,27 +47,59 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
if ( (inc_x == 1) && (inc_y == 1) )
{
int n1 = n & -4;
while(i < n1)
#if V_SIMD && !defined(DSDOT)
const int vstep = v_nlanes_f32;
const int unrollx4 = n & (-vstep * 4);
const int unrollx = n & -vstep;
v_f32 vsum0 = v_zero_f32();
v_f32 vsum1 = v_zero_f32();
v_f32 vsum2 = v_zero_f32();
v_f32 vsum3 = v_zero_f32();
while(i < unrollx4)
{
vsum0 = v_muladd_f32(
v_loadu_f32(x + i), v_loadu_f32(y + i), vsum0
);
vsum1 = v_muladd_f32(
v_loadu_f32(x + i + vstep), v_loadu_f32(y + i + vstep), vsum1
);
vsum2 = v_muladd_f32(
v_loadu_f32(x + i + vstep*2), v_loadu_f32(y + i + vstep*2), vsum2
);
vsum3 = v_muladd_f32(
v_loadu_f32(x + i + vstep*3), v_loadu_f32(y + i + vstep*3), vsum3
);
i += vstep*4;
}
vsum0 = v_add_f32(
v_add_f32(vsum0, vsum1), v_add_f32(vsum2 , vsum3)
);
while(i < unrollx)
{
vsum0 = v_muladd_f32(
v_loadu_f32(x + i), v_loadu_f32(y + i), vsum0
);
i += vstep;
}
dot = v_sum_f32(vsum0);
#elif defined(DSDOT)
for (; i < n1; i += 4)
{
#if defined(DSDOT)
dot += (double) y[i] * (double) x[i]
+ (double) y[i+1] * (double) x[i+1]
+ (double) y[i+2] * (double) x[i+2]
+ (double) y[i+3] * (double) x[i+3] ;
}
#else
for (; i < n1; i += 4)
{
dot += y[i] * x[i]
+ y[i+1] * x[i+1]
+ y[i+2] * x[i+2]
+ y[i+3] * x[i+3] ;
#endif
i+=4 ;
}
#endif
while(i < n)
{