float-conv

Go application to convert a number to a floating-point format and print details like the hex/bit representation of the converted number, conversion error etc.

• IEEE-754 Float32
• BFloat16
• Float16

Usage

float-conv --num=<number> [--format=<format>] [--round-mode=<rounding mode>] [--overflow-mode=<overflow mode>] [--underflow-mode=<underflow-mode>]

• The option --num is used to provide the input number. The number can be input as either decimal, scientific or in hexfloat formats.
• The --format option is used to specify which floating point format to convert to. Valid values are float32 [Default], bfloat16, float16.
• The --round-mode option is used to specify which rounding mode to use, if the number cannot be exactly represented in the desired format. Supported options are
  • rtz: Round Towards Zero
  • rtposinf: Round Towards Positive Infinity
  • rtneginf: Round Towards Negative Infinity
  • rthalfzero: Round to the closest number, break ties by rounding towards zero
  • rthalfposinf: Round to the closest number, break ties by rounding towards positive infinity
  • rthalfneginf: Round to the closest number, break ties by rounding towards negative infinity
  • rne: Round towards the nearest even number (LSB is 0) [Default]
  • rno: Round towards the nearest odd number (LSB is 1)
• The --overflow-mode option is used to specify the response if the number (in magnitude) is larger than the maximum representable (in magnitude) in the target format. Supported options are
  • satinf: Saturate the number to infinity with the same sign as the input
  • satmax: Saturate the number to the maximum possible in the format, with the same sign as the input [Default]
• The --underflow-mode option is used to specify the response if the number (in magnitude) is smaller than the minimum representable (in magnitude) in the target format. Supported options are
  • flushzero: Flush the number to 0. If the target format supports signed zeros, then the sign is same as that of the input
  • satmin: Saturates the number to the minimum representable, with the same sign as the input [Default]
• The --precision flag is used to augment the precision to use when parsing the input. The default is 53.

Example

$ float-conv --num=0.125 --format=float32
Float32
|Sign|Exponent|               Mantissa|
|   0|00111110|00000000000000000000000|
Decimal: 1.25e-01
Hexfloat: 0x1p-03
Conversion Error: 0e+00 (Exact)
Binary: 0b00111110000000000000000000000000
Hexadecimal: 0x3e000000

$ float-conv --num=1e-256 --format=bfloat16 --underflow-mode=flushzero
BFloat16
|Sign|Exponent|Mantissa|
|   0|00000000| 0000000|
Decimal: 0e+00
Hexfloat: 0x0p+00
Conversion Error: -1e-256 (Below)
Binary: 0b0000000000000000
Hexadecimal: 0x0000
UNDERFLOW

Roadmap

• Add support for OCP Microscaling 8-bit Formats