bfloat16 Namespace Reference

Constants for BFloat16 number format handling. More...

Variables
constexpr uint16_t	EXPONENT_MASK = 0x7F80
constexpr uint16_t	MANTISSA_MASK = 0x007F
constexpr uint16_t	SIGN_BIT = 0x8000
constexpr uint16_t	ZERO = 0x0000
constexpr uint16_t	NEG_ZERO = 0x8000
constexpr int	SHIFT_BITS = 16

Detailed Description

Constants for BFloat16 number format handling.

BFloat16 is a 16-bit floating point format that uses 1 sign bit, 8 exponent bits, and 7 mantissa bits.

Variable Documentation

EXPONENT_MASK

constexpr uint16_t bfloat16::EXPONENT_MASK = 0x7F80

constexpr

Mask for extracting exponent bits

Definition at line 22 of file model_constants.h.

Referenced by bfloat16_to_float32().

MANTISSA_MASK

constexpr uint16_t bfloat16::MANTISSA_MASK = 0x007F

constexpr

Mask for extracting mantissa bits

Definition at line 23 of file model_constants.h.

Referenced by bfloat16_to_float32().

NEG_ZERO

constexpr uint16_t bfloat16::NEG_ZERO = 0x8000

constexpr

Representation of negative zero

Definition at line 26 of file model_constants.h.

Referenced by bfloat16_to_float32().

SHIFT_BITS

constexpr int bfloat16::SHIFT_BITS = 16

constexpr

Number of bits to shift for conversion

Definition at line 27 of file model_constants.h.

Referenced by bfloat16_to_float32().

SIGN_BIT

constexpr uint16_t bfloat16::SIGN_BIT = 0x8000

constexpr

Mask for extracting sign bit

Definition at line 24 of file model_constants.h.

Referenced by bfloat16_to_float32().

ZERO

constexpr uint16_t bfloat16::ZERO = 0x0000

constexpr

Representation of positive zero

Definition at line 25 of file model_constants.h.

Referenced by bfloat16_to_float32(), and TinyLlamaModel::initialize_gpu_and_rope().