include/ck/utility/mxf4_utils.hpp Source File

include/ck/utility/mxf4_utils.hpp Source File#

Composable Kernel: include/ck/utility/mxf4_utils.hpp Source File
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck/utility/data_type.hpp"
 #include "ck/utility/mxfp_utils.hpp"
  
 namespace ck::utils {
  
 template <>
 __host__ __device__ inline bool is_nan<f4_t>(e8m0_bexp_t const scale,
                                              f4_t const dataBytes [[maybe_unused]])
 {
     // no need to check for data as it does not have NaN representation
     return scale == NumericLimits<e8m0_bexp_t>::QuietNaN();
 }
  
 // no infinity representation in ocp_e2m1_mxfp4 will always return false
 template <>
 __host__ __device__ inline bool is_inf<f4_t>(e8m0_bexp_t const scale [[maybe_unused]],
                                              f4_t const data [[maybe_unused]])
 {
     // no inf representation for ocp_e2m1_mxfp4
     return false;
 }
  
 template <>
 __host__ __device__ inline bool is_zero<f4_t>(e8m0_bexp_t const scale, f4_t const data)
 {
     if(is_nan<f4_t>(scale, data))
         return false;
  
     // no need to check for scale as it does not have a 0 representation
     f4_t result = (data & 0b00001111) & NumericUtils<f4_t>::set_sign_mask;
  
     return result == 0b0;
 }
  
 template <>
 __host__ __device__ inline float to_float<f4_t>(e8m0_bexp_t const scale, f4_t const data)
 {
     if(is_nan<f4_t>(scale, data))
         return std::numeric_limits<float>::quiet_NaN();
  
     if(is_zero<f4_t>(scale, data))
         return 0.0f;
  
     f4_t prepared_data = data & 0b00001111;
  
     int scale_exp = get_exponent_value<e8m0_bexp_t>(scale);
  
     return convert_to_float<f4_t>(prepared_data, scale_exp);
 }
  
 template <>
 __host__ __device__ inline f4_t sat_convert_to_type<f4_t>(float value)
 {
     cvt t;
     t.value_float = value;
     uint32_t sign = t.value_bitwise >> 31;
  
     if(std::isnan(value))
     {
  
         return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
                     : NumericUtils<f4_t>::data_max_positive_normal_mask;
     }
  
     if(std::abs(value) > NumericLimits<f4_t>::Max()) // covers inf case as well
         return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
                     : NumericUtils<f4_t>::data_max_positive_normal_mask;
  
     f4_t res = convert_to_type<f4_t>(value);
  
     if(std::abs(to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
        NumericLimits<f4_t>::DataMinSubnorm())
         return value < 0 ? NumericUtils<f4_t>::negative_zero_mask
                          : NumericUtils<f4_t>::positive_zero_mask;
  
     return res;
 }
  
 template <>
 __host__ __device__ inline f4_t sat_convert_to_type_sr<f4_t>(float value, uint32_t seed)
 {
     cvt t;
     t.value_float = value;
     uint32_t sign = t.value_bitwise >> 31;
  
     if(std::isnan(value))
         return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
                     : NumericUtils<f4_t>::data_max_positive_normal_mask;
  
     if(std::abs(value) > NumericLimits<f4_t>::Max()) // covers inf case as well
         return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
                     : NumericUtils<f4_t>::data_max_positive_normal_mask;
  
     f4_t res = convert_to_type_sr<f4_t>(value, seed);
  
     if(std::abs(to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
        NumericLimits<f4_t>::DataMinSubnorm())
         return value < 0 ? NumericUtils<f4_t>::negative_zero_mask
                          : NumericUtils<f4_t>::positive_zero_mask;
  
     return res;
 }
  
 } // namespace ck::utils