public static unsafe void V256Filter(int *src, int srcCount, int target, int *dst, out int dstCount)
{
var alignedCount = srcCount & ~7;
var dstPtr = dst;
int i = 0;
for (; i < alignedCount; i += 8)
{
var val = mm256_loadu_si256(src + i);
var cmp = mm256_cmpeq_epi32(val, mm256_set1_epi32(target));
var packed = SIMDHelpers.LeftPack8PS(cmp, val);
mm256_storeu_si256(dstPtr, packed);
var mask = mm256_movemask_ps(cmp);
dstPtr += popcnt_u32((uint)mask);
}
for (; i < srcCount; i++)
{
if (src[i] == target)
{
*(dstPtr++) = src[i];
}
}
dstCount = (int)(dstPtr - dst);
}
private static unsafe void V256Filter([NoAlias] int *src, int srcCount, int greaterThan, int lessThan, [NoAlias] int *dst, [NoAlias] out int dstCount)
{
var alignedCount = srcCount & ~7;
var dstPtr = dst;
int i = 0;
for (; i < alignedCount; i += 8)
{
var val = mm256_loadu_si256(src + i);
var cmpLt = mm256_cmpgt_epi32(mm256_set1_epi32(lessThan), val);
var cmpGt = mm256_cmpgt_epi32(val, mm256_set1_epi32(greaterThan));
var cmp = mm256_and_ps(cmpLt, cmpGt);
var packed = SIMDHelpers.LeftPack8PS(cmp, val);
mm256_storeu_si256(dstPtr, packed);
var mask = mm256_movemask_ps(cmp);
dstPtr += popcnt_u32((uint)mask);
}
for (; i < srcCount; i++)
{
if (src[i] < lessThan && src[i] > greaterThan)
{
*(dstPtr++) = src[i];
}
}
dstCount = (int)(dstPtr - dst);
}