Built-in Data Types

Data Type List

Ascend C provides four data types with different bit widths ranging from b8 to b64 (8 bits to 64 bits). The following table lists the data types corresponding to different bit widths.

To simplify the descriptions of these data types, the following data type abbreviations are provided:

Applicability

Boolean

The bool type occupies 8 bits. When all bits are 0, it represents false; any non-zero value represents true.

Integer

An integer consists of a sign bit (S) and a magnitude part (M). Different types of integers allocate bits differently between the sign and magnitude. Unsigned integers have no sign bit, and all bits are used to represent the value.

The following figure shows an example of int8_t, where the sign bit occupies 1 bit and the magnitude part occupies 7 bits. S_v = 1 and M_v = 2⁵ + 2⁶ represent the value 96. The subscript _v indicates the specific values of the sign bit and magnitude part.

The following table lists the value ranges of integer data types.

Floating-Point Number

The following table lists the value ranges of floating-point data types.

A floating-point number consists of three parts: sign bit (S), exponent (E), and mantissa (M). The number of bits occupied by the three parts may vary depending on the type of floating-point number.

• fp4x2_e2m1_t

  The following figure shows an example of fp4x2_e2m1_t, where the sign bit occupies 1 bit, the exponent occupies 2 bits, and the mantissa occupies 1 bit.

  

• fp4x2_e1m2_t

  The following figure shows an example of fp4x2_e1m2_t, where the sign bit occupies 1 bit, the exponent occupies 1 bit, and the mantissa occupies 2 bits.

  

• fp8_e8m0_t (The binary format of fp8_e8m0_t is derived from bfloat16 by discarding the sign bit and mantissa bits, keeping only the exponent.)

  The following figure shows an example of fp8_e8m0_t, where the sign bit occupies 1 bit, the exponent occupies 8 bits, and there is no mantissa.

  

• fp8_e5m2_t

  The following figure shows an example of fp8_e5m2_t, where the sign bit occupies 1 bit, the exponent occupies 5 bits, and the mantissa occupies 2 bits. The represented value in the example is (–1)⁰ × (2 – 0.25) × 2^{(30 – 15)} = 1.75 × 2¹⁵.

  

  The bit representations of special values of fp8_e5m2_t are as follows:

  

• fp8_e4m3fn_t

  The following figure shows an example of fp8_e4m3fn_t. The sign bit occupies 1 bit, the exponent occupies 4 bits, and the mantissa occupies 3 bits. The represented value in the example is (–1)¹ × 2^–3 × 2^–6.

  

  The bit representations of special values of fp8_e4m3fn_t are as follows:

  

• hifloat8_t

  Compared with other types, hifloat8_t has an additional exponent-width control field D. Field D indicates how the exponent bits and mantissa bits are encoded.

  hifloat8_t has different encoding modes depending on field D. The following lists the encoding modes. The sign, exponent, and mantissa are abbreviated as S, E, and M, respectively.

  Figure 1 Bit distribution of S, E, and M under different values of control field D
  

  In the following example, the sign bit occupies 1 bit, the exponent occupies 2 bits, and the mantissa occupies 3 bits. The D field is 2-bit with binary value 01, S_v = 1, E_v = 3, and M_v = 2^–1 + 2^–2, and the represented result is 14. The subscript _v indicates the specific value of each part.

  

  The following table lists the value ranges of hifloat8_t.

  Table 4 Value ranges of hifloat8_t

  Sign Bit Width	Bit Width of Field D	Exponent Bit Width	Mantissa Bit Width	Value of Field D	Sign Value Range (Sv)	Exponent Value Range (Ev)	Mantissa Value Range (Mv)	Value Range Calculation Formula
  1	4	0	3	4'b0000	±1	-	[0, 7]	Sv × 2Mv – 23
  1	4	0	3	4'b0001	±1	0	[0, 7 × 2–3]	Sv × 2Ev × (1 + Mv)
  1	3	1	3	4'b001	±1	±1	[0, 7 × 2–3]
  1	2	2	3	2'b01	±1	±[2, 3]	[0, 7 × 2–3]
  1	2	3	2	2'b10	±1	±[4, 7]	[0, 3 × 2–2]
  1	2	4	1	2'b11	±1	±[8, 15]	[0, 2–1]

  The bit representations of special values of hifloat8_t are as follows:

  

  The calculation formula of hifloat8_t is as follows:

  • Sign bit S_v

    If s_bit_val is 1, the value is negative. If s_bit_val is 0, the value is non-negative.

    

  • Exponent E_v

    It consists of Es and Em.

    Table 5 Es and Em bits according to the value of field D

    Value of Field D	Es Value Range	Em Value Range
    3b001	0–1	-
    2b01	0–1	10–11
    2b10	0–1	100–111
    2b11	0–1	1000–1111

    The es_bit_val value indicates the highest bit of E and is used to calculate the sign value of Ev. For example, if E is 0b1100, the es_bit_val value is the highest bit 1. If E is 0b011, the es_bit_val value is the highest bit 0.

    Formula for calculating Es:

    

    Formula for calculating Em: In the formula, D indicates the exponent-width control field. For details, see Table 5.

    

    Formula for calculating Ev:

    

    In the formula for calculating Mv, M indicates the bit value, and bitwidth of M indicates the bit width of M. For details, see Table 5.

    

  • In Normal and Subnormal modes, the formulas for calculating the value of a floating-point number are different:

    Normal: The formula consists of Sv, Ev, and Mv.

    

    Subnormal: The formula consists of Sv and Mv.

    

    

• half

  The following figure shows an example of half. The sign bit occupies 1 bit, the exponent occupies 5 bits, and the mantissa occupies 10 bits. S_v = 1, E_v = 15, and M_v = 2^–1 + 2^–2, indicating that the result is 1.75. The subscript _v indicates the specific value of each part.

  

  The bit representations of special values of half are as follows:

  

• bfloat16_t

  The following figure shows an example of bfloat16_t, where the sign bit occupies 1 bit, the exponent occupies 8 bits, and the mantissa occupies 7 bits.

  

  The bit representations of special values of bfloat16_t are as follows:

  

• float

  The following figure shows an example of float, where the sign bit occupies 1 bit, the exponent occupies 8 bits, and the mantissa occupies 23 bits.

  

  The bit representations of special values of float are as follows:

  

Complex Number

Ascend C provides complex number data types: complex32 and complex64.

Definition:

using complex32 = AscendC::Complex<half>;
using complex64 = AscendC::Complex<float>;

For details about the definition of Complex, see complex32/complex64.

complex32 is a complex number where both the real and imaginary parts are of the half type.

complex64 is a complex number where both the real and imaginary parts are of the float type.

Example:

complex32 value0(1, 2);
value0 indicates a complex number whose real part is 1 and imaginary part is 2, that is, 1 + 2j.
complex32 value1(3);
value1 indicates a complex number whose real part is 3 and imaginary part is 0, that is, 3 + 0j.
complex64 value2 = 4;
value2 indicates a complex number whose real part is 4 and imaginary part is 0, that is, 4 + 0j.

Currently, only Atlas 350 Accelerator Card supports this function.