c10/util/Float8_e5m2fnuz.cpp

#include <c10/util/Float8_e5m2fnuz.h>
#include <array>
#include <iostream>

namespace c10 {

namespace detail {

C10_HOST_DEVICE float fp8e5m2fnuz_to_fp32_value(uint8_t input) {
  constexpr std::array<float, 256> e5m2fnuz_lut = {
      0.0f,
      7.62939453125e-06f,
      1.52587890625e-05f,
      2.288818359375e-05f,
      3.0517578125e-05f,
      3.814697265625e-05f,
      4.57763671875e-05f,
      5.340576171875e-05f,
      6.103515625e-05f,
      7.62939453125e-05f,
      9.1552734375e-05f,
      0.0001068115234375f,
      0.0001220703125f,
      0.000152587890625f,
      0.00018310546875f,
      0.000213623046875f,
      0.000244140625f,
      0.00030517578125f,
      0.0003662109375f,
      0.00042724609375f,
      0.00048828125f,
      0.0006103515625f,
      0.000732421875f,
      0.0008544921875f,
      0.0009765625f,
      0.001220703125f,
      0.00146484375f,
      0.001708984375f,
      0.001953125f,
      0.00244140625f,
      0.0029296875f,
      0.00341796875f,
      0.00390625f,
      0.0048828125f,
      0.005859375f,
      0.0068359375f,
      0.0078125f,
      0.009765625f,
      0.01171875f,
      0.013671875f,
      0.015625f,
      0.01953125f,
      0.0234375f,
      0.02734375f,
      0.03125f,
      0.0390625f,
      0.046875f,
      0.0546875f,
      0.0625f,
      0.078125f,
      0.09375f,
      0.109375f,
      0.125f,
      0.15625f,
      0.1875f,
      0.21875f,
      0.25f,
      0.3125f,
      0.375f,
      0.4375f,
      0.5f,
      0.625f,
      0.75f,
      0.875f,
      1.0f,
      1.25f,
      1.5f,
      1.75f,
      2.0f,
      2.5f,
      3.0f,
      3.5f,
      4.0f,
      5.0f,
      6.0f,
      7.0f,
      8.0f,
      10.0f,
      12.0f,
      14.0f,
      16.0f,
      20.0f,
      24.0f,
      28.0f,
      32.0f,
      40.0f,
      48.0f,
      56.0f,
      64.0f,
      80.0f,
      96.0f,
      112.0f,
      128.0f,
      160.0f,
      192.0f,
      224.0f,
      256.0f,
      320.0f,
      384.0f,
      448.0f,
      512.0f,
      640.0f,
      768.0f,
      896.0f,
      1024.0f,
      1280.0f,
      1536.0f,
      1792.0f,
      2048.0f,
      2560.0f,
      3072.0f,
      3584.0f,
      4096.0f,
      5120.0f,
      6144.0f,
      7168.0f,
      8192.0f,
      10240.0f,
      12288.0f,
      14336.0f,
      16384.0f,
      20480.0f,
      24576.0f,
      28672.0f,
      32768.0f,
      40960.0f,
      49152.0f,
      57344.0f,
      std::numeric_limits<float>::signaling_NaN(),
      -7.62939453125e-06f,
      -1.52587890625e-05f,
      -2.288818359375e-05f,
      -3.0517578125e-05f,
      -3.814697265625e-05f,
      -4.57763671875e-05f,
      -5.340576171875e-05f,
      -6.103515625e-05f,
      -7.62939453125e-05f,
      -9.1552734375e-05f,
      -0.0001068115234375f,
      -0.0001220703125f,
      -0.000152587890625f,
      -0.00018310546875f,
      -0.000213623046875f,
      -0.000244140625f,
      -0.00030517578125f,
      -0.0003662109375f,
      -0.00042724609375f,
      -0.00048828125f,
      -0.0006103515625f,
      -0.000732421875f,
      -0.0008544921875f,
      -0.0009765625f,
      -0.001220703125f,
      -0.00146484375f,
      -0.001708984375f,
      -0.001953125f,
      -0.00244140625f,
      -0.0029296875f,
      -0.00341796875f,
      -0.00390625f,
      -0.0048828125f,
      -0.005859375f,
      -0.0068359375f,
      -0.0078125f,
      -0.009765625f,
      -0.01171875f,
      -0.013671875f,
      -0.015625f,
      -0.01953125f,
      -0.0234375f,
      -0.02734375f,
      -0.03125f,
      -0.0390625f,
      -0.046875f,
      -0.0546875f,
      -0.0625f,
      -0.078125f,
      -0.09375f,
      -0.109375f,
      -0.125f,
      -0.15625f,
      -0.1875f,
      -0.21875f,
      -0.25f,
      -0.3125f,
      -0.375f,
      -0.4375f,
      -0.5f,
      -0.625f,
      -0.75f,
      -0.875f,
      -1.0f,
      -1.25f,
      -1.5f,
      -1.75f,
      -2.0f,
      -2.5f,
      -3.0f,
      -3.5f,
      -4.0f,
      -5.0f,
      -6.0f,
      -7.0f,
      -8.0f,
      -10.0f,
      -12.0f,
      -14.0f,
      -16.0f,
      -20.0f,
      -24.0f,
      -28.0f,
      -32.0f,
      -40.0f,
      -48.0f,
      -56.0f,
      -64.0f,
      -80.0f,
      -96.0f,
      -112.0f,
      -128.0f,
      -160.0f,
      -192.0f,
      -224.0f,
      -256.0f,
      -320.0f,
      -384.0f,
      -448.0f,
      -512.0f,
      -640.0f,
      -768.0f,
      -896.0f,
      -1024.0f,
      -1280.0f,
      -1536.0f,
      -1792.0f,
      -2048.0f,
      -2560.0f,
      -3072.0f,
      -3584.0f,
      -4096.0f,
      -5120.0f,
      -6144.0f,
      -7168.0f,
      -8192.0f,
      -10240.0f,
      -12288.0f,
      -14336.0f,
      -16384.0f,
      -20480.0f,
      -24576.0f,
      -28672.0f,
      -32768.0f,
      -40960.0f,
      -49152.0f,
      -57344.0f,
  };

  return e5m2fnuz_lut[input];
}

} // namespace detail

static_assert(
    std::is_standard_layout_v<Float8_e5m2fnuz>,
    "c10::Float8_e5m2 must be standard layout.");

std::ostream& operator<<(std::ostream& out, const Float8_e5m2fnuz& value) {
  out << (float)value;
  return out;
}

} // namespace c10