Fix axpby_rvv kernels for cases where inc_y = 0

The following openblas_utest tests fail when the RISCV64_ZVL128B is
enabled.

TEST 89/103 axpby:zaxpby_inc_0 [FAIL]
TEST 92/103 axpby:caxpby_inc_0 [FAIL]
TEST 95/103 axpby:daxpby_inc_0 [FAIL]
TEST 98/103 axpby:saxpby_inc_0 [FAIL]

The issue is that the vectorized kernels do not work when inc_y == 0.
This patch updates the kernels to fall back to the scalar algorithms
when inc_y == 0, fixing the failing tests.

Signed-off-by: Mark Ryan <markdryan@rivosinc.com>
This commit is contained in:
Mark Ryan 2024-07-12 11:16:48 +00:00
parent 6013b36b16
commit 67bf4b6998
2 changed files with 70 additions and 35 deletions

View File

@ -114,6 +114,11 @@ int CNAME(BLASLONG n, FLOAT alpha, FLOAT *x, BLASLONG inc_x, FLOAT beta, FLOAT *
vy = VFMULVF_FLOAT(vy, beta, vl);
VSEV_FLOAT (y, vy, vl);
}
} else if (inc_y == 0) {
FLOAT vf = y[0];
for (; n > 0; n--)
vf *= beta;
y[0] = vf;
} else {
BLASLONG stride_y = inc_y * sizeof(FLOAT);
for (size_t vl; n > 0; n -= vl, y += vl*inc_y) {
@ -134,6 +139,13 @@ int CNAME(BLASLONG n, FLOAT alpha, FLOAT *x, BLASLONG inc_x, FLOAT beta, FLOAT *
vy = VFMACCVF_FLOAT(vy, alpha, vx, vl);
VSEV_FLOAT (y, vy, vl);
}
} else if (inc_y == 0) {
FLOAT vf = y[0];
for (; n > 0; n--) {
vf = (vf * beta) + (x[0] * alpha);
x += inc_x;
}
y[0] = vf;
} else if (1 == inc_x) {
BLASLONG stride_y = inc_y * sizeof(FLOAT);
for (size_t vl; n > 0; n -= vl, x += vl, y += vl*inc_y) {

View File

@ -79,8 +79,10 @@ int CNAME(BLASLONG n, FLOAT alpha_r, FLOAT alpha_i, FLOAT *x, BLASLONG inc_x, FL
BLASLONG stride_x = inc_x2 * sizeof(FLOAT);
BLASLONG stride_y = inc_y2 * sizeof(FLOAT);
BLASLONG ix;
FLOAT_V_T vx0, vx1, vy0, vy1;
FLOAT_VX2_T vxx2, vyx2;
FLOAT temp;
if ( beta_r == 0.0 && beta_i == 0.0)
{
@ -125,6 +127,14 @@ int CNAME(BLASLONG n, FLOAT alpha_r, FLOAT alpha_i, FLOAT *x, BLASLONG inc_x, FL
if ( alpha_r == 0.0 && alpha_i == 0.0 )
{
if ( inc_y == 0 ) {
for (; n > 0; n--)
{
temp = (beta_r * y[0] - beta_i * y[1]);
y[1] = (beta_r * y[1] + beta_i * y[0]);
y[0] = temp;
}
} else {
for (size_t vl; n > 0; n -= vl, y += vl*inc_y2)
{
vl = VSETVL(n);
@ -144,8 +154,20 @@ int CNAME(BLASLONG n, FLOAT alpha_r, FLOAT alpha_i, FLOAT *x, BLASLONG inc_x, FL
VSSSEG_FLOAT(y, stride_y, v_x2, vl);
}
}
}
else
{
if ( inc_y == 0 ) {
ix = 0;
for (; n > 0; n--) {
temp = (alpha_r * x[ix] - alpha_i * x[ix+1] ) +
(beta_r * y[0] - beta_i * y[1]);
y[1] = (alpha_r * x[ix+1] + alpha_i * x[ix]) +
(beta_r * y[1] + beta_i * y[0]);
y[0] = temp;
ix += inc_x2;
}
} else {
for (size_t vl; n > 0; n -= vl, x += vl*inc_x2, y += vl*inc_y2)
{
vl = VSETVL(n);
@ -175,6 +197,7 @@ int CNAME(BLASLONG n, FLOAT alpha_r, FLOAT alpha_i, FLOAT *x, BLASLONG inc_x, FL
}
}
}
}
return(0);
}