index;
}
auto compress(uint32x4_t mask, float32x4_t a) {
static constexpr std::array<uint32_t, 4> weights{1 + 16, 2 + 16, 4 + 16, 8 + 16};
const size_t s = vaddvq_u32(vandq_u32(mask, vld1q_u32(weights.data())));
const size_t count = s >> 4;
const size_t idx = s & 15;

static constexpr auto index_table = make_index_table();

const auto index = vld1q_u8(index_table[idx].data());
return std::pair{vreinterpretq_f32_u8(vqtbl1q_u8(vreinterpretq_u8_f32(a), index)), count};
}
auto copy_if_neon_unroll(const float* __restrict a,
float* __restrict out,
float threshold,
size_t n) {
auto thd = vdupq_n_f32(threshold);
size_t j = 0;
size_t i = 0;
for (; i + 16 <= n; i += 16) {
#pragma unroll
for (size_t i0 = 0; i0 < 16; i0 += 4) {
auto v = vld1q_f32(a + i + i0);
auto mask = vcgtq_f32(v, thd);
auto [packed, cnt] = compress(mask, v);
vst1q_f32(out + j, packed);
j += cnt;
}
}
for (; i + 4 <= n; i += 4) {
auto v = vld1q_f32(a + i);
auto mask = vcgtq_f32(v, thd);

auto [packed, cnt] = compress(mask, v);
vst1q_f32(out + j, packed);
j += cnt;
}
return j;
}
```
`tbl` and the index table provide `compress`, something that NEON doesn't have out of the box.
This isn't just about `> threshold`. Filter, remove and other data-dependent functions are built the same way.


https://redd.it/1uqscnr
@r_cpp