Skip to content

Latest commit

 

History

History
80 lines (64 loc) · 4.29 KB

README.md

File metadata and controls

80 lines (64 loc) · 4.29 KB

SoftPositE

C-based Function Extension of SoftPosit Library

About SoftPositE

Extension of SoftPosit, aiming to address the challenges of SoftPosit Library.

Currently, we have implemented the following extended functions. We are still working on..., according to our own requirements.

Note: the following extended functions have not been exhaustively tested, that is, there may be unknown bugs. :)

// posit-based fma unit, supporting arbitrary posit format & mixed-precision strategy
uint32_t posit_muladd_mixed(
	uint_fast32_t uiA, uint_fast32_t uiB, uint_fast32_t uiC, uint_fast32_t op, 
	int n_i, int es_i, int n_o, int es_o);

// posit-based fma unit, supporting arbitrary posit format
uint32_t posit_muladd(
	uint_fast32_t uiA, uint_fast32_t uiB, uint_fast32_t uiC, uint_fast32_t op, int n, int es);

// convert arbitrary posit format to double format
double convertPositToDouble(posit32_t pA, int n, int es);

// convert double format to arbitrary posit format
posit32_t convertDoubleToPosit(double f32, int n, int es);

Appendix: About SoftPosit

SoftPosit is a comprehensive C library for posits based on Berkeley SoftFloat, developed by NGA reseach team.

The library offers an array of posit-based operations, such as basic addition and multiplication, fused multiply-add, and format consersion. Additionally, it provides support for the quire format.

Features

The main functions provided are as follows (take posit16_t as an example):

Main Posit Functionalities
  • Add: posit16_t p16_add(posit16_t, posit16_t)
  • Subtract: posit16_t p16_sub(posit16_t, posit16_t)
  • Divide: posit16_t p16_div(posit16_t, posit16_t)
  • Multiply: posit16_t p16_mul(posit16_t, posit16_t)
  • Fused Multiply Add: posit16_t p16_mulAdd(posit16_t, posit16_t, posit16_t)
Main Quire Functionalities
  • Fused Dot Product-Add: quire16_t q16_fdp_add(quire16_t, posit16_t, posit16_t)
  • Fused Dot Product-Subtract: quire16_t q16_fdp_sub(quire16_t, posit16_t, posit16_t)
  • Set Quire Variable to Zero: quire16_t q16_clr(quire16_t)
  • Convert Quire to Posit: posit16_t q16_to_p16(quire16_t)
Functionalities in Posit Standard
  • Square Root: posit16_t p16_sqrt(posit16_t)
  • Round to Nearest Integer: posit16_t p16_roundToInt(posit16_t)
  • Check Equal: bool p16_eq(posit16_t, posit16_t)
  • Check Less than Equal: bool p16_le(posit16_t, posit16_t)
  • Check Less than: bool p16_lt(posit16_t, posit16_t)
  • Convert Posit to Integer (32-bit): int_fast32_t p16_to_i32(posit16_t)
  • Convert Posit to Long Long Integer (64-bit): int_fast64_t p16_to_i64(posit16_t)
  • Convert Unsigned Integer (32-bit) to Posit: posit16_t ui32_to_p16(uint32_t a)
  • Convert Unsigned Long Long Int (64-bit) to Posit: posit16_t ui64_to_p16(uint64_t a)
  • Convert Integer (32-bit) to Posit: posit16_t i32_to_p16(int32_t a)
  • Convert Long Integer (64 bits) to Posit: posit16_t i64_to_p16(int64_t a)
  • Convert Posit to Unsigned Integer (32-bit): uint_fast32_t p16_to_ui32(posit16_t)
  • Convert Posit to Unsigned Long Long Integer (64-bit): uint_fast64_t p16_to_ui64(posit16_t)
  • Convert Posit to Posit of Another Size: posit8_t p16_to_p8(posit16_t)
Helper Functionalities (Not in Posit Standard)
  • Convert Posit to Double (64-bit): double convertP16ToDouble(posit16_t)
  • Convert Double (64-bit) to Posit: posit16_t convertDoubleToP16(double)
  • Convert Binary Expressed in Unsigned Integer to Posit: posit16_t castP16(uint16_t)
  • Cast Posit into Binary expressed in Unsigned Integer: uint16_t castUI(posit16_t)

Challenges

  • Limited Format Supports: The functions listed above may only be available in a specific posit format, which can limit flexibility. The main supported formats are as follows:
    • posit32_t: 32-bit with two exponent bits (n=32, es=2)
    • posit16_t: 16-bit with one exponent bit (n=16, es=1)
    • posit8_t: 8-bit with zero exponent bit (n=8, es=0)
    • posit_2_t: 2-bit to 32-bit with two exponent bits (es=2)
    • posit_1_t: 2-bit to 32-bit with one exponent bit (es=1)
  • Limited Function Capabilities: The function can be further extended, e.g., adding support for mixed-precision arithmetic.
  • Potential Bugs: As mentioned in README, some functions have not been exhaustively tested, that is, there may be unknown bugs.