LLM AVX

// My Version
void clCompute::IMPL_BASE_UnPackBits_Scatter(i64 num, const u32 *src, u8 *dst)
{
  i64 bitIndex = 0;
  for (i32 b = 31; b >= 0; b--)
  {
    u32 mask = 1 << b;
    i64 localBitIndex = bitIndex;
    for (i64 i = 0; i < num; i++, localBitIndex++)
      dst[localBitIndex >> 3] |= ((src[i] & mask) >> b) << (7 - (localBitIndex & 7));
    bitIndex += num;
  }
}

// AI Version
void clCompute::IMPL_AVX2_UnPackBits_Scatter(i64 num, const u32 * src, u8 * dst)
{
  memset(dst, 0, num * sizeof(u32));

  // Process 8 elements at a time
  const int64_t numChunks = num >> 3;
  const int64_t remainder = num & 3;
  for (int bit = 31; bit >= 0; bit--)
  {
    __m256i bitMask = _mm256_set1_epi32(1 << bit);
    int64_t bitOffset = (31 - bit) * num;
    for (int64_t i = 0; i < numChunks; i++)
    {
      __m256i data = _mm256_loadu_si256((__m256i *) & src[i * 8]);
      __m256i masked = _mm256_and_si256(data, bitMask);
      __m256i shifted = _mm256_srli_epi32(masked, bit);
      uint32_t bitResults =
        ((_mm256_extract_epi32(shifted, 0) & 1) << 7) |
        ((_mm256_extract_epi32(shifted, 1) & 1) << 6) |
        ((_mm256_extract_epi32(shifted, 2) & 1) << 5) |
        ((_mm256_extract_epi32(shifted, 3) & 1) << 4) |
        ((_mm256_extract_epi32(shifted, 4) & 1) << 3) |
        ((_mm256_extract_epi32(shifted, 5) & 1) << 2) |
        ((_mm256_extract_epi32(shifted, 6) & 1) << 1) |
        ((_mm256_extract_epi32(shifted, 7) & 1) << 0);
      dst[(bitOffset + i * 8) >> 3] = (uint8_t)bitResults;
    }

    // Process remaining elements
    int64_t offset = numChunks * 8;
    for (int64_t i = 0; i < remainder; i++)
    {
      uint32_t bitValue = (src[offset + i] & (1 << bit)) >> bit;
      dst[(bitOffset + offset + i) >> 3] |= (bitValue << (7 - ((bitOffset + offset + i) & 7)));
    }
  }
}