Function quantize_row_q8_K_ref

src/quant.cpp:616–653

}

void quantize_row_q8_K_ref(const float * __restrict__ x, block_q8_K * __restrict__ y, int64_t k) {
  assert(k % QK_K == 0);
  const int64_t nb = k / QK_K;

  for (int i = 0; i < nb; i++) {

    float max = 0;
    float amax = 0;
    for (int j = 0; j < QK_K; ++j) {
      float ax = fabsf(x[j]);
      if (ax > amax) {
        amax = ax; max = x[j];
      }
    }
    if (!amax) {
      y[i].d = 0;
      memset(y[i].qs, 0, QK_K);
      x += QK_K;
      continue;
    }
    //const float iscale = -128.f/max;
    // We need this change for IQ2_XXS, else the AVX implementation becomes very awkward
    const float iscale = -127.f/max;
    for (int j = 0; j < QK_K; ++j) {
      int v = nearest_int(iscale*x[j]);
      y[i].qs[j] = std::min(127, v);
    }
    for (int j = 0; j < QK_K/16; ++j) {
      int sum = 0;
      for (int ii = 0; ii < 16; ++ii) {
        sum += y[i].qs[j*16 + ii];
      }
      y[i].bsums[j] = sum;
    }
    y[i].d = 1/iscale;
    x += QK_K;
  }
}

void dequantize_row_q8_K(const block_q8_K * __restrict__ x, float * __restrict__ y, int64_t k) {
  assert(k % QK_K == 0);