filter-chain: handle unaligned sum_simd

Fall back to slow path if the input is not aligned, which can happen
after resampling.

Fixes #1659
This commit is contained in:
Wim Taymans 2021-09-29 16:24:58 +02:00
parent d2f0573369
commit 275dfed92a

View file

@ -2128,15 +2128,19 @@ static void zconvolve_simd(PFFFT_Setup * s, const float *a, const float *b,
static void sum_simd(const float *a, const float *b, float *ab, int len)
{
const v4sf *RESTRICT va = (const v4sf *)a;
const v4sf *RESTRICT vb = (const v4sf *)b;
v4sf *RESTRICT vab = (v4sf *) ab;
int i;
const int end4 = len / SIMD_SZ;
int i = 0;
for (i = 0; i < end4; i += 1)
vab[i] = VADD(va[i],vb[i]);
for (i = i * 4; i < len; ++i)
if (VALIGNED(a) && VALIGNED(b) && VALIGNED(ab)) {
const v4sf *RESTRICT va = (const v4sf *)a;
const v4sf *RESTRICT vb = (const v4sf *)b;
v4sf *RESTRICT vab = (v4sf *) ab;
const int end4 = len / SIMD_SZ;
for (; i < end4; i += 1)
vab[i] = VADD(va[i],vb[i]);
i *= 4;
}
for (; i < len; ++i)
ab[i] = a[i] + b[i];
}