/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/utility/numeric_utils.hpp Source File

• develop
• include
• ck
• utility

numeric_utils.hpp

Go to the documentation of this file.

1 // SPDX-License-Identifier: MIT

2 // // Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.

3 #pragma once

4 #include "ck/utility/data_type.hpp"

5 

6 namespace ck {

7 

8 template <typename T>

9 struct NumericUtils

10 {

11 };

12 

13 template <>

14 struct NumericUtils<e8m0_bexp_t>

15 {

16  static constexpr int exp = 8;

17  static constexpr int mant = 0;

18  static constexpr int bias = 127;

19 

20  static constexpr int unbiased_exp_min = -127;

21  static constexpr int unbiased_exp_max = 127;

22  static constexpr int biased_exp_min = 0;

23  static constexpr int biased_exp_max = 254;

24 

25  using bitwise_type = uint8_t;

26 };

27 

28 template <>

29 struct NumericUtils<float>

30 {

31  static constexpr int exp = 8;

32  static constexpr int mant = 23;

33  static constexpr int bias = 127;

34  static constexpr uint32_t nan_mask = 0x7F800000;

35  static constexpr uint32_t head_mask = 0xFF800000;

36  static constexpr uint32_t mant_mask = 0x7FFFFF;

37  static constexpr uint32_t exp_mask = 0xFF;

38  static constexpr uint32_t Inf = 0x7F800000;

39  static constexpr uint32_t NegInf = 0xFF800000;

40  static constexpr uint32_t NaN = 0x7F800001;

41  static constexpr uint32_t Neg0 = 0x80000000;

42  static constexpr bool has_inf = true;

43  using bitwise_type = uint32_t;

44 };

45 

46 template <>

47 struct NumericUtils<half_t>

48 {

49  static constexpr int exp = 5;

50  static constexpr int mant = 10;

51  static constexpr int bias = 15;

52  static constexpr uint16_t nan_mask = 0x7C00;

53  static constexpr uint16_t head_mask = 0xFC00;

54  static constexpr uint16_t mant_mask = 0x3FF;

55  static constexpr uint16_t exp_mask = 0x1F;

56  static constexpr uint32_t Inf = 0x7C00;

57  static constexpr uint32_t NegInf = 0xFC00;

58  static constexpr uint32_t NaN = 0x7C01;

59  static constexpr uint32_t Neg0 = 0x8000;

60  static constexpr bool has_inf = true;

61  using bitwise_type = uint16_t;

62 };

63 

64 template <>

65 struct NumericUtils<bhalf_t>

66 {

67  static constexpr int exp = 8;

68  static constexpr int mant = 7;

69  static constexpr int bias = 128; // negative zero nan mode

70  // static constexpr int bias = 127; // ieee mode

71 };

72 

73 template <>

74 struct NumericUtils<f8_fnuz_t>

75 {

76  static constexpr int exp = 4;

77  static constexpr int mant = 3;

78  static constexpr int bias = 8; // negative zero nan mode

79  // static constexpr int bias = 7; // ieee mode

80  static constexpr bool has_inf = false;

81 };

82 

83 template <>

84 struct NumericUtils<bf8_fnuz_t>

85 {

86  static constexpr int exp = 5;

87  static constexpr int mant = 2;

88  static constexpr int bias = 16; // negative zero nan mode

89  // static constexpr int bias = 15; // ieee mode

90  static constexpr bool has_inf = false;

91 };

92 template <>

93 struct NumericUtils<f8_ocp_t>

94 {

95  static constexpr int exp = 4;

96  static constexpr int mant = 3;

97  static constexpr int bias = 7;

98 };

99 

100 template <>

101 struct NumericUtils<bf8_ocp_t>

102 {

103  static constexpr int exp = 5;

104  static constexpr int mant = 2;

105  static constexpr int bias = 15;

106 };

107 

108 template <>

109 struct NumericUtils<f4_t>

110 {

111  static constexpr int exp = 2;

112  static constexpr int mant = 1;

113  static constexpr int bias = 1;

114  static constexpr uint32_t sr_shift = 10;

115 

116  static constexpr int unbiased_exp_min = 0;

117  static constexpr int unbiased_exp_max = 2;

118  static constexpr int biased_exp_min = 1;

119  static constexpr int biased_exp_max = 3;

120 

121  static constexpr uint8_t positive_zero_mask = 0b0000;

122  static constexpr uint8_t negative_zero_mask = 0b1000;

123 

124  static constexpr uint8_t one_mask = 0b0010;

125  static constexpr uint8_t set_sign_mask = 0b0111;

126 

127  static constexpr uint8_t data_max_positive_normal_mask = 0b0111;

128  static constexpr uint8_t data_max_negative_normal_mask = 0b1111;

129 

130  static constexpr uint8_t data_max_positive_subnormal_mask = 0b0001;

131  static constexpr uint8_t data_max_negative_subnormal_mask = 0b1001;

132 

133  static constexpr bool has_inf = false;

134 

135  using bitwise_type = uint8_t;

136 };

137 

138 template <>

139 struct NumericUtils<f6_t>

140 {

141  static constexpr int exp = 2;

142  static constexpr int mant = 3;

143  static constexpr int bias = 1;

144  static constexpr uint32_t sr_shift = 12;

145 

146  static constexpr int unbiased_exp_min = 0;

147  static constexpr int unbiased_exp_max = 2;

148  static constexpr int biased_exp_min = 1;

149  static constexpr int biased_exp_max = 3;

150 

151  static constexpr uint8_t positive_zero_mask = 0b000000;

152  static constexpr uint8_t negative_zero_mask = 0b100000;

153 

154  static constexpr uint8_t set_sign_mask = 0b011111;

155 

156  static constexpr uint8_t data_max_positive_normal_mask = 0b011111;

157  static constexpr uint8_t data_max_negative_normal_mask = 0b111111;

158 

159  static constexpr uint8_t data_max_positive_subnormal_mask = 0b000111;

160  static constexpr uint8_t data_max_negative_subnormal_mask = 0b100111;

161 

162  static constexpr bool has_inf = false;

163  static constexpr bool has_nan = false;

164  static constexpr bool has_zero = true;

165 

166  using bitwise_type = uint8_t;

167 };

168 

169 template <>

170 struct NumericUtils<bf6_t>

171 {

172  static constexpr int exp = 3;

173  static constexpr int mant = 2;

174  static constexpr int bias = 3;

175  static constexpr uint32_t sr_shift = 11;

176 

177  static constexpr int unbiased_exp_min = -2;

178  static constexpr int unbiased_exp_max = 4;

179  static constexpr int biased_exp_min = 1;

180  static constexpr int biased_exp_max = 7;

181 

182  static constexpr uint8_t positive_zero_mask = 0b000000;

183  static constexpr uint8_t negative_zero_mask = 0b100000;

184 

185  static constexpr uint8_t set_sign_mask = 0b011111;

186 

187  static constexpr uint8_t data_max_positive_normal_mask = 0b011111;

188  static constexpr uint8_t data_max_negative_normal_mask = 0b111111;

189 

190  static constexpr uint8_t data_max_positive_subnormal_mask = 0b000011;

191  static constexpr uint8_t data_max_negative_subnormal_mask = 0b100011;

192 

193  static constexpr bool has_inf = false;

194  static constexpr bool has_nan = false;

195  static constexpr bool has_zero = true;

196 

197  using bitwise_type = uint8_t;

198 };

199 } // namespace ck

data_type.hpp

ck::math::exp

__host__ T exp(T x)

Definition: math_v2.hpp:391

ck

Definition: ck.hpp:267

ck::f4_t

unsigned _BitInt(4) f4_t

Definition: data_type.hpp:32

ck::f6_t

_BitInt(6) f6_t

Definition: data_type.hpp:33

ck::half_t

_Float16 half_t

Definition: data_type.hpp:30

ck::bhalf_t

ushort bhalf_t

Definition: data_type.hpp:29

ck::bf6_t

unsigned _BitInt(6) bf6_t

Definition: data_type.hpp:34

ck::bf8_fnuz_t

unsigned _BitInt(8) bf8_fnuz_t

Definition: amd_ck_fp8.hpp:37

ck::f8_fnuz_t

_BitInt(8) f8_fnuz_t

Definition: amd_ck_fp8.hpp:36

uint16_t

unsigned short uint16_t

Definition: stdint.h:125

uint32_t

unsigned int uint32_t

Definition: stdint.h:126

uint8_t

unsigned char uint8_t

Definition: stdint.h:124

ck::NumericUtils< bf6_t >::bitwise_type

uint8_t bitwise_type

Definition: numeric_utils.hpp:197

ck::NumericUtils< e8m0_bexp_t >::bitwise_type

uint8_t bitwise_type

Definition: numeric_utils.hpp:25

ck::NumericUtils< f4_t >::bitwise_type

uint8_t bitwise_type

Definition: numeric_utils.hpp:135

ck::NumericUtils< f6_t >::bitwise_type

uint8_t bitwise_type

Definition: numeric_utils.hpp:166

ck::NumericUtils< float >::bitwise_type

uint32_t bitwise_type

Definition: numeric_utils.hpp:43

ck::NumericUtils< half_t >::bitwise_type

uint16_t bitwise_type

Definition: numeric_utils.hpp:61

ck::NumericUtils

Definition: numeric_utils.hpp:10

ck::bf8_ocp_t

Definition: amd_ck_fp8.hpp:344

ck::e8m0_bexp_t

Unsigned representation of a conventional biased Float32 exponent.

Definition: e8m0.hpp:25

ck::f8_ocp_t

Definition: amd_ck_fp8.hpp:298