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

Merge pull request #2981 from Qiyu8/fix-sum 

Fix sum optimize issues
This commit is contained in:
Martin Kroeker 2020-11-16 08:40:46 +01:00 committed by GitHub
parent be075d53cf c4c591ac5a
commit 7e9cb39a25
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 20 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

39
kernel/arm/sum.c
View File

@ -42,24 +42,27 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
n *= inc_x;
if (inc_x == 1)
{
#if V_SIMD
#if V_SIMD && (!defined(DOUBLE) || (defined(DOUBLE) && V_SIMD_F64 && V_SIMD > 128))
#ifdef DOUBLE
const int vstep = v_nlanes_f64;
const int unrollx2 = n & (-vstep * 2);
const int unrollx4 = n & (-vstep * 4);
const int unrollx = n & -vstep;
v_f64 vsum0 = v_zero_f64();
v_f64 vsum1 = v_zero_f64();
while (i < unrollx2)
{
vsum0 = v_add_f64(vsum0, v_loadu_f64(x));
vsum1 = v_add_f64(vsum1, v_loadu_f64(x + vstep));
i += vstep * 2;
}
vsum0 = v_add_f64(vsum0, vsum1);
while (i < unrollx)
v_f64 vsum2 = v_zero_f64();
v_f64 vsum3 = v_zero_f64();
for (; i < unrollx4; i += vstep * 4)
{
vsum0 = v_add_f64(vsum0, v_loadu_f64(x + i));
vsum1 = v_add_f64(vsum1, v_loadu_f64(x + i + vstep));
vsum2 = v_add_f64(vsum2, v_loadu_f64(x + i + vstep * 2));
vsum3 = v_add_f64(vsum3, v_loadu_f64(x + i + vstep * 3));
}
vsum0 = v_add_f64(
v_add_f64(vsum0, vsum1), v_add_f64(vsum2, vsum3));
for (; i < unrollx; i += vstep)
{
vsum0 = v_add_f64(vsum0, v_loadu_f64(x + i));
i += vstep;
}
sumf = v_sum_f64(vsum0);
#else
@ -70,20 +73,18 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v_f32 vsum1 = v_zero_f32();
v_f32 vsum2 = v_zero_f32();
v_f32 vsum3 = v_zero_f32();
while (i < unrollx4)
for (; i < unrollx4; i += vstep * 4)
{
vsum0 = v_add_f32(vsum0, v_loadu_f32(x));
vsum1 = v_add_f32(vsum1, v_loadu_f32(x + vstep));
vsum2 = v_add_f32(vsum2, v_loadu_f32(x + vstep * 2));
vsum3 = v_add_f32(vsum3, v_loadu_f32(x + vstep * 3));
i += vstep * 4;
vsum0 = v_add_f32(vsum0, v_loadu_f32(x + i));
vsum1 = v_add_f32(vsum1, v_loadu_f32(x + i + vstep));
vsum2 = v_add_f32(vsum2, v_loadu_f32(x + i + vstep * 2));
vsum3 = v_add_f32(vsum3, v_loadu_f32(x + i + vstep * 3));
}
vsum0 = v_add_f32(
v_add_f32(vsum0, vsum1), v_add_f32(vsum2, vsum3));
while (i < unrollx)
for (; i < unrollx; i += vstep)
{
vsum0 = v_add_f32(vsum0, v_loadu_f32(x + i));
i += vstep;
}
sumf = v_sum_f32(vsum0);
#endif