Class: Flt::Support::Formatter

Inherits:
Object
  • Object
show all
Defined in:
lib/flt/support.rb

Overview

Burger and Dybvig free formatting algorithm, from their paper: “Printing Floating-Point Numbers Quickly and Accurately” (Robert G. Burger, R. Kent Dybvig)

This algorithm formats arbitrary base floating point numbers as decimal text literals. The floating-point (with fixed precision) is interpreted as an approximated value, representing any value in its ‘rounding-range’ (the interval where all values round to the floating-point value, with the given precision and rounding mode). An alternative approach which is not taken here would be to represent the exact floating-point value with some given precision and rounding mode requirements; that can be achieved with Clinger algorithm (which may fail for exact precision).

The variables used by the algorithm are stored in instance variables: Quotients of integers will be used to hold the magnitudes: All numbers in the randound-range are rounded to @v (with the given precision p) significant digit right after the radix point. @b**@k is the first power of @b >= u

The rounding range of @v is the interval of values that round to @v under the runding-mode. If the rounding mode is one of the round-to-nearest variants (even, up, down), then it is ((v+v-)/2 = (@v-@m_m)/@s, (v+v+)/2 = (@v+@m_)/2) whith the boundaries open or closed as explained below. In this case:

@m_m/@s = (@v - (v + v-)/2) where v- = @v.next_minus is the lower adjacent to v floating point value
@m_p/@s = ((v + v+)/2 - @v) where v+ = @v.next_plus is the upper adjacent to v floating point value

If the rounding is directed, then the rounding interval is either (v-, @v] or [@v, v+] if @roundh, then @k is the minimum @k with (@r+@m_p)/@s <= @output_b**@k

@k = ceil(logB((@r+@m_p)/2)) with lobB the @output_b base logarithm

if @roundh, then @k is the minimum @k with (@r+@m_p)/@s < @output_b**@k

@k = 1+floor(logB((@r+@m_p)/2))

p is the input floating point precision

Constant Summary collapse

ESTIMATE_FLOAT_LOG_B = 
{2=>1/Math.log(2), 10=>1/Math.log(10), 16=>1/Math.log(16)}

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(input_b, input_min_e, output_b) ⇒ Formatter

This Object-oriented implementation is slower than the functional one for two reasons:

  • The overhead of object creation

  • The use of instance variables instead of local variables

But if scale is optimized or local variables are used in the inner loops, then this implementation is on par with the functional one for Float and it is more efficient for Flt types, where the variables passed as parameters hold larger objects.



832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
 
# File 'lib/flt/support.rb', line 832

def initialize(input_b, input_min_e, output_b)
  @b = input_b
  @min_e = input_min_e
  @output_b = output_b
  # result of last operation
  @adjusted_digits = @digits = nil
  # for "all-digits" mode results (which are truncated, rather than rounded),
  # round_up contains information to round the result:
  # * it is nil if the rest of digits are zero (the result is exact)
  # * it is :lo if there exist non-zero digits beyond the significant ones (those returned), but
  #   the value is below the tie (the value must be rounded up only for :up rounding mode)
  # * it is :tie if there exists exactly one nonzero digit after the significant and it is radix/2,
  #   for round-to-nearest it is atie.
  # * it is :hi otherwise (the value should be rounded-up except for the :down mode)
  @round_up = nil
end

Instance Attribute Details

#round_up ⇒ Object (readonly)

Returns the value of attribute round_up.



979
980
981
 
# File 'lib/flt/support.rb', line 979

def round_up
  @round_up
end

Instance Method Details

#adjusted_digits(round_mode) ⇒ Object

Access rounded result of format operation: scaling (position of radix point) and digits



983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
 
# File 'lib/flt/support.rb', line 983

def adjusted_digits(round_mode)
  round_mode = Support.simplified_round_mode(round_mode, @minus)
  if @adjusted_digits.nil? && !@digits.nil?
    increment = (@round_up && (round_mode != :down)) &&
                ((round_mode == :up) ||
                (@round_up == :hi) ||
                ((@round_up == :tie) &&
                 ((round_mode==:half_up) || ((round_mode==:half_even) && ((@digits.last % 2)==1)))))
    # increment = (@round_up == :tie) || (@round_up == :hi) # old behaviour (:half_up)
    if increment
      base = @output_b
      dec_pos = @k
      digits = @digits.dup
      # carry = increment ? 1 : 0
      # digits = digits.reverse.map{|d| d += carry; d>=base ? 0 : (carry=0;d)}.reverse
      # if carry != 0
      #   digits.unshift carry
      #   dec_pos += 1
      # end
      i = digits.size - 1
      while i>=0
        digits[i] += 1
        if digits[i] == base
          digits[i] = 0
        else
          break
        end
        i -= 1
      end
      if i<0
        dec_pos += 1
        digits.unshift 1
      end
      @adjusted_k = dec_pos
      @adjusted_digits = digits
    else
      @adjusted_k = @k
      @adjusted_digits = @digits
    end
  end
  return @adjusted_k, @adjusted_digits
end

#b_power(n) ⇒ Object 



1073
1074
1075
 
# File 'lib/flt/support.rb', line 1073

def b_power(n)
  @b**n
end

#digits ⇒ Object

Access result of format operation: scaling (position of radix point) and digits



975
976
977
 
# File 'lib/flt/support.rb', line 975

def digits
  return @k, @digits
end

#format(v, f, e, round_mode, p = nil, all = false) ⇒ Object

This method converts v = f*b**e into a sequence of output_b-base digits, so that if the digits are converted back to a floating-point value of precision p (correctly rounded), the result is v. If round_mode is not nil, just enough digits to produce v using that rounding is used; otherwise enough digits to produce v with any rounding are delivered.

If the all parameter is true, all significant digits are generated without rounding, i.e. all digits that, if used on input, cannot arbitrarily change while preserving the parsed value of the floating point number. Since the digits are not rounded more digits may be needed to assure round-trip value preservation. This is useful to reflect the precision of the floating point value in the output; in particular trailing significant zeros are shown. But note that, for directed rounding and base conversion this may need to produce an infinite number of digits, in which case an exception will be raised. This is specially frequent for the :up rounding mode, in which any number with a finite number of nonzero digits equal to or less than the precision will haver and infinite sequence of zero significant digits.

With :down rounding (truncation) this could be used to show the exact value of the floating point but beware: when used with directed rounding, if the value has not an exact representation in the output base this will lead to an infinite loop. formatting ‘0.1’ (as a decimal floating-point number) in base 2 with :down rounding

When the all parameters is used the result is not rounded (is truncated), and the round_up flag is set to indicate that nonzero digits exists beyond the returned digits; the possible values of the round_up flag are:

  • nil : the rest of digits are zero (the result is exact)

  • :lo : there exist non-zero digits beyond the significant ones (those returned), but the value is below the tie (the value must be rounded up only for :up rounding mode)

  • :tie : there exists exactly one nonzero digit after the significant and it is radix/2, for round-to-nearest it is atie.

  • :hi : the value is closer to the rounded-up value (incrementing the last significative digit.)

Note that the round_mode here is not the rounding mode applied to the output; it is the rounding mode that applied to input preserves the original floating-point value (with the same precision as input). should be rounded-up.



886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
 
# File 'lib/flt/support.rb', line 886

def format(v, f, e, round_mode, p=nil, all=false)
  context = v.class.context
  # TODO: consider removing parameters f,e and using v.split instead
  @minus = (context.sign(v)==-1)
  @v = context.copy_sign(v, +1) # don't use context.abs(v) because it rounds (and may overflow also)
  @f = f.abs
  @e = e
  @round_mode = round_mode
  @all_digits = all
  p ||= context.precision

  # adjust the rounding mode to work only with positive numbers
  @round_mode = Support.simplified_round_mode(@round_mode, @minus)

  # determine the high,low inclusion flags of the rounding limits
  case @round_mode
    when :half_even
      # rounding rage is (v-m-,v+m+) if v is odd and [v+m-,v+m+] if even
      @round_l = @round_h = ((@f%2)==0)
    when :up
      # rounding rage is (v-,v]
      # ceiling is treated here assuming f>0
      @round_l, @round_h = false, true
    when :down
      # rounding rage is [v,v+)
      # floor is treated here assuming f>0
      @round_l, @round_h = true, false
    when :half_up
      # rounding rage is [v+m-,v+m+)
      @round_l, @round_h = true, false
    when :half_down
      # rounding rage is (v+m-,v+m+]
      @round_l, @round_h = false, true
    else # :nearest
      # Here assume only that round-to-nearest will be used, but not which variant of it
      # The result is valid for any rounding (to nearest) but may produce more digits
      # than stricly necessary for specific rounding modes.
      # That is, enough digits are generated so that when the result is
      # converted to floating point with the specified precision and
      # correct rounding (to nearest), the result is the original number.
      # rounding range is (v+m-,v+m+)
      @round_l = @round_h = false
  end

  # TODO: use context.next_minus, next_plus instead of direct computing, don't require min_e & ps
  # Now compute the working quotients @r/@s, @m_p/@s = (v+ - @v), @m_m/@s = (@v - v-) and scale them.
  if @e >= 0
    if @f != b_power(p-1)
      be = b_power(@e)
      @r, @s, @m_p, @m_m = @f*be*2, 2, be, be
    else
      be = b_power(@e)
      be1 = be*@b
      @r, @s, @m_p, @m_m = @f*be1*2, @b*2, be1, be
    end
  else
    if @e==@min_e or @f != b_power(p-1)
      @r, @s, @m_p, @m_m = @f*2, b_power(-@e)*2, 1, 1
    else
      @r, @s, @m_p, @m_m = @f*@b*2, b_power(1-@e)*2, @b, 1
    end
  end
  @k = 0
  @context = context
  scale_optimized!


  # The value to be formatted is @v=@r/@s; m- = @m_m/@s = (@v - v-)/@s; m+ = @m_p/@s = (v+ - @v)/@s
  # Now adjust @m_m, @m_p so that they define the rounding range
  case @round_mode
  when :up
    # ceiling is treated here assuming @f>0
    # rounding range is -v,@v
    @m_m, @m_p = @m_m*2, 0
  when :down
    # floor is treated here assuming #f>0
    # rounding range is @v,v+
    @m_m, @m_p = 0, @m_p*2
  else
    # rounding range is v-,v+
    # @m_m, @m_p = @m_m, @m_p
  end

  # Now m_m, m_p define the rounding range
  all ? generate_max : generate

end

#generate ⇒ Object 



1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
 
# File 'lib/flt/support.rb', line 1121

def generate
  list = []
  r, s, m_p, m_m, = @r, @s, @m_p, @m_m
  loop do
    d,r = (r*@output_b).divmod(s)
    m_p *= @output_b
    m_m *= @output_b
    tc1 = @round_l ? (r<=m_m) : (r<m_m)
    tc2 = @round_h ? (r+m_p >= s) : (r+m_p > s)

    if not tc1
      if not tc2
        list << d
      else
        list << d+1
        break
      end
    else
      if not tc2
        list << d
        break
      else
        if r*2 < s
          list << d
          break
        else
          list << d+1
          break
        end
      end
    end

  end
  @digits = list
end

#generate_max ⇒ Object 



1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
 
# File 'lib/flt/support.rb', line 1081

def generate_max
  @round_up = false
  list = []
  r, s, m_p, m_m, = @r, @s, @m_p, @m_m
  n_iters, rs = 0, []
  loop do
    if (n_iters > 10000)
      raise "Infinite digit sequence." if rs.include?(r)
      rs << r
    else
      n_iters += 1
    end

    d,r = (r*@output_b).divmod(s)

    m_p *= @output_b
    m_m *= @output_b

    list << d

    tc1 = @round_l ? (r<=m_m) : (r<m_m)
    tc2 = @round_h ? (r+m_p >= s) : (r+m_p > s)

    if tc1 && tc2
      if r != 0
        r *= 2
        if r > s
          @round_up = :hi
        elsif r == s
          @round_up = :tie
        else
          @rund_up = :lo
        end
      end
      break
    end
  end
  @digits = list
end

#output_b_power(n) ⇒ Object 



1077
1078
1079
 
# File 'lib/flt/support.rb', line 1077

def output_b_power(n)
  @output_b**n
end

#scale! ⇒ Object

using local vars instead of instance vars: it makes a difference in performance



1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
 
# File 'lib/flt/support.rb', line 1051

def scale!
  r, s, m_p, m_m, k,output_b = @r, @s, @m_p, @m_m, @k,@output_b
  loop do
    if (@round_h ? (r+m_p >= s) : (r+m_p > s)) # k is too low
      s *= output_b
      k += 1
    elsif (@round_h ? ((r+m_p)*output_b<s) : ((r+m_p)*output_b<=s)) # k is too high
      r *= output_b
      m_p *= output_b
      m_m *= output_b
      k -= 1
    else
      @s = s
      @r = r
      @m_p = m_p
      @m_m = m_m
      @k = k
      break
    end
  end
end

#scale_fixup! ⇒ Object

fix up scaling (final step): specialized version of scale! This performs a single up scaling step, i.e. behaves like scale2, but the input must be at most one step down from the final result



1231
1232
1233
1234
1235
1236
 
# File 'lib/flt/support.rb', line 1231

def scale_fixup!
  if (@round_h ? (@r+@m_p >= @s) : (@r+@m_p > @s)) # too low?
    @s *= @output_b
    @k += 1
  end
end

#scale_optimized! ⇒ Object

scale_o1! is an optimized version of scale!; it requires an additional parameters with the floating-point number v=r/s

It uses a Float estimate of ceil(logB(v)) that may need to adjusted one unit up TODO: find easy to use estimate; determine max distance to correct value and use it for fixing,

or use the general scale! for fixing (but remembar to multiply by exptt(...))
(determine when Math.log is aplicable, etc.)



1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
 
# File 'lib/flt/support.rb', line 1165

def scale_optimized!
  context = @context # @v.class.context
  return scale! if context.zero?(@v)

  # 1. compute estimated_scale

  # 1.1. try to use Float logarithms (Math.log)
  v = @v
  v_abs = context.copy_sign(v, +1) # don't use v.abs because it rounds (and may overflow also)
  v_flt = v_abs.to_f
  b = @output_b
  log_b = ESTIMATE_FLOAT_LOG_B[b]
  log_b = ESTIMATE_FLOAT_LOG_B[b] = 1.0/Math.log(b) if log_b.nil?
  estimated_scale = nil
  fixup = false
  begin
    l = ((b==10) ? Math.log10(v_flt) : Math.log(v_flt)*log_b)
    estimated_scale =(l - 1E-10).ceil
    fixup = true
  rescue
    # rescuing errors is more efficient than checking (v_abs < Float::MAX.to_i) && (v_flt > Float::MIN) when v is a Flt
  else
    # estimated_scale = nil
  end

  # 1.2. Use Flt::DecNum logarithm
  if estimated_scale.nil?
    v.to_decimal_exact(:precision=>12) if v.is_a?(BinNum)
    if v.is_a?(DecNum)
      l = nil
      DecNum.context(:precision=>12) do
        case b
        when 10
          l = v_abs.log10
        else
          l = v_abs.ln/Flt.DecNum(b).ln
        end
      end
      l -= Flt.DecNum(+1,1,-10)
      estimated_scale = l.ceil
      fixup = true
    end
  end

  # 1.3 more rough Float aproximation
    # TODO: optimize denominator, correct numerator for more precision with first digit or part
    # of the coefficient (like _log_10_lb)
  estimated_scale ||= (v.adjusted_exponent.to_f * Math.log(v.class.context.radix) * log_b).ceil

  if estimated_scale >= 0
    @k = estimated_scale
    @s *= output_b_power(estimated_scale)
  else
    sc = output_b_power(-estimated_scale)
    @k = estimated_scale
    @r *= sc
    @m_p *= sc
    @m_m *= sc
  end
  fixup ? scale_fixup! : scale!

end

#scale_original!(really = false) ⇒ Object

Given r/s = v (number to convert to text), m_m/s = (v - v-)/s, m_p/s = (v+ - v)/s Scale the fractions so that the first significant digit is right after the radix point, i.e. find k = ceil(logB((r+m_p)/s)), the smallest integer such that (r+m_p)/s <= B^k if k>=0 return:

r=r, s=s*B^k, m_p=m_p, m_m=m_m

if k<0 return:

r=r*B^k, s=s, m_p=m_p*B^k, m_m=m_m*B^k

scale! is a general iterative method using only (multiprecision) integer arithmetic.



1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
 
# File 'lib/flt/support.rb', line 1035

def scale_original!(really=false)
  loop do
    if (@round_h ? (@r+@m_p >= @s) : (@r+@m_p > @s)) # k is too low
      @s *= @output_b
      @k += 1
    elsif (@round_h ? ((@r+@m_p)*@output_b<@s) : ((@r+@m_p)*@output_b<=@s)) # k is too high
      @r *= @output_b
      @m_p *= @output_b
      @m_m *= @output_b
      @k -= 1
    else
      break
    end
  end
end