Skip to main content

ClikaRT::dtype::FloatInfo

struct

Header: ClikaRT/compute/finfo.h

Numeric properties of a floating-point DataType.

Data members​

bits​

int bits

total stored bits (Float6_E2M3 -> 6, Float4_E2M1 -> 4).

Declared in ClikaRT/compute/finfo.h, line 20

exponent_bits​

int exponent_bits

exponent field width, bits

Declared in ClikaRT/compute/finfo.h, line 21

mantissa_bits​

int mantissa_bits

mantissa (fraction) field width, bits

Declared in ClikaRT/compute/finfo.h, line 22

bias​

int bias

exponent bias.

Declared in ClikaRT/compute/finfo.h, line 23

is_signed​

bool is_signed

has a sign bit (false for E8M0).

Declared in ClikaRT/compute/finfo.h, line 24

has_infinity​

bool has_infinity

representation has infinities

Declared in ClikaRT/compute/finfo.h, line 25

has_nan​

bool has_nan

representation has NaN

Declared in ClikaRT/compute/finfo.h, line 26

has_denorm​

bool has_denorm

representation has subnormals

Declared in ClikaRT/compute/finfo.h, line 27

is_lookup​

bool is_lookup

true for an NF4-style lookup-quantized format.

Declared in ClikaRT/compute/finfo.h, line 28

eps​

float eps

1 ULP at 1.0.

Declared in ClikaRT/compute/finfo.h, line 29

smallest_normal​

float smallest_normal

smallest positive NORMAL value

Declared in ClikaRT/compute/finfo.h, line 30

denorm_min​

float denorm_min

smallest positive subnormal (== smallest_normal when has_denorm is false)

Declared in ClikaRT/compute/finfo.h, line 31

min_value​

float min_value

most-negative finite value.

Declared in ClikaRT/compute/finfo.h, line 32

max_value​

float max_value

most-positive finite value.

Declared in ClikaRT/compute/finfo.h, line 33

dtype​

DataType dtype

the dtype these facts describe

Declared in ClikaRT/compute/finfo.h, line 34