CxxUtils::FloatCompressor Class Reference

Class implementing a lossy float compression. More...

#include <FloatCompressor.h>

Collaboration diagram for CxxUtils::FloatCompressor:

Classes
union	floatint_t
	Type used in the compression. More...

Public Member Functions
	FloatCompressor (unsigned int mantissaBits=7)
	Constructor with the number of mantissa bits to retain.
float	reduceFloatPrecision (float value) const
	Function returning a reduced precision float value.

Private Attributes
unsigned int	m_mantissaBits
	Number of mantissa bits to keep.
uint32_t	m_mantissaBitmask
	Bitmask for zeroing out the non-interesting bits.
Magic numbers
uint32_t	m_rounding
uint32_t	m_vmax
	Largest possible positive 32bit float minus the rounding.

Detailed Description

Class implementing a lossy float compression.

This is just a copy of the code implemented first in CaloUtils. Used to reduce the precision of select floating point quantities during the POOL conversion.

Author

Scott Snyder snyde.nosp@m.r@bn.nosp@m.l.gov

Attila Krasznahorkay Attil.nosp@m.a.Kr.nosp@m.aszna.nosp@m.hork.nosp@m.ay@ce.nosp@m.rn.c.nosp@m.h

Revision

789425

Date

2016-12-13 11:50:12 +0100 (Tue, 13 Dec 2016)

Definition at line 30 of file FloatCompressor.h.

Constructor & Destructor Documentation

FloatCompressor()

CxxUtils::FloatCompressor::FloatCompressor

(

unsigned int

mantissaBits = 7

)

Constructor with the number of mantissa bits to retain.

Definition at line 18 of file FloatCompressor.cxx.

      : m_mantissaBits( mantissaBits ), m_mantissaBitmask( 0 ) {
 
      // IEEE754 single-precision float
      // SEEE EEEE EMMM MMMM MMMM MMMM MMMM MMMM
      //   F   F     8   0    0    0    7    F  
 
      // Definition:
      //
      // Assume that we'd like to keep only 7 bits in the mantissa
      // In this case the memory layout of the bits will be:
      //
      // Sign |  Exp (8 bits) | Frac (23 bits)
      // S       EEEEEEEE       FFFFFFLRTTTTTTTTTTTTTTT
      //
      // where
      //
      // S : Sign bit
      // E : Exponent bits
      // F : Fraction bits
      // L : Least significant bit (lsb) for 7 bits mantissa precision
      // R : Rounding bit
      // T : Sticky bits (i.e any bit after lsb + 1)
      //
      // In the current implementation there are essentially 4 cases:
      //
      // Case 1: L = 0 and R = 0
      // In this case there'll be rounding down
      //
      // Case 2: L = 1 and R = 0
      // In this case there'll be rounding down
      //
      // Case 3: L = 0 and R = 1
      // In this case there'll be rounding up
      //
      // Note: This scenario can be different than bfloat16 implementation
      //       of TensorFlow, where they round down if all the Ts are zero.
      //       Otherwise, they also round up.
      //
      // Case 4: L = 1 and R = 1
      // In this case there'll be rounding up
      //
      // In all cases, we do an extra check to avoid overflow.
      //
      // From a technical point of view, the rounding is computed
      // to be the half of the lsb(=1) and added to the original value
      // as long as the new value doesn't overflow. Then the
      // undesired bits are masked. We never go below 5 bits in the
      // mantissa.
 
      // Adjust the received bit number to some reasonable value:
      if( m_mantissaBits < 5 ) {
         m_mantissaBits = 5;
      }
      if( m_mantissaBits > NMANTISSA ) {
         m_mantissaBits = NMANTISSA;
      }
 
      // Fill up the lower N bits:
      for( unsigned int i = 0; i < ( NMANTISSA - m_mantissaBits ); ++i ) {
         m_mantissaBitmask |= ( 0x1 << i );
      }
      // And now negate it to get the correct mask:
      m_mantissaBitmask = ~m_mantissaBitmask;
 
      // Set the Magic numbers
      if (m_mantissaBits == NMANTISSA) {
        m_rounding = 0;
      }
      else {
        m_rounding = 0x1 << ( 32 - (1 + 8 + m_mantissaBits) - 1 );
      }
      // The part below is taken from AthenaPoolCnvSvc/Compressor
      // and would work the same as long as the user doesn't
      // compress lower than 3 mantissa bits, which is not allowed
      // in any case.
      m_vmax = 0x7f7 << 20;
      m_vmax |= 0x000fffff xor (m_rounding);
   }

Member Function Documentation

reduceFloatPrecision()

float CxxUtils::FloatCompressor::reduceFloatPrecision

(

float

value

)

const

Function returning a reduced precision float value.

Definition at line 98 of file FloatCompressor.cxx.

                                                                  {
 
      // Check if any compression is to be made:
      if( m_mantissaBits == NMANTISSA ) {
         return value;
      }
 
      // Check for NaN, etc:
      if( ! std::isfinite( value ) ) {
         return value;
      }
 
      // Create the helper object:
      floatint_t fi;
      fi.fvalue = value;
 
      //safety-check if value (omitting the sign-bit) is lower than vmax
      //(avoid overflow)
      if( ( fi.ivalue & 0x7fffffff ) < m_vmax ) {
         fi.ivalue += m_rounding;
      }
 
      // Do the compression:
      fi.ivalue &= m_mantissaBitmask;
      return fi.fvalue;
   }

Member Data Documentation

m_mantissaBitmask

uint32_t CxxUtils::FloatCompressor::m_mantissaBitmask

private

Bitmask for zeroing out the non-interesting bits.

Definition at line 49 of file FloatCompressor.h.

m_mantissaBits

unsigned int CxxUtils::FloatCompressor::m_mantissaBits

private

Number of mantissa bits to keep.

Definition at line 47 of file FloatCompressor.h.

m_rounding

uint32_t CxxUtils::FloatCompressor::m_rounding

private

Definition at line 55 of file FloatCompressor.h.

m_vmax

uint32_t CxxUtils::FloatCompressor::m_vmax

private

Largest possible positive 32bit float minus the rounding.

Definition at line 57 of file FloatCompressor.h.

---

The documentation for this class was generated from the following files:

• FloatCompressor.h
• FloatCompressor.cxx