module Flt module Support # Floating-point reading and printing (from/to text literals). # # Here are methods for floating-point reading, using algorithms by William D. Clinger, and # printing, using algorithms by Robert G. Burger and R. Kent Dybvig. # # Reading and printing can also viewed as floating-point conversion between a fixed-precision # floating-point format (the floating-point numbers) and and a free floating-point format (text), # which may use different numerical bases. # # The Reader class, in the default :free mode, converts a free-form numeric value # (as a text literal, i.e. a free floating-point format, usually in base 10) which is taken # as an exact value, to a correctly-rounded floating-point of specified precision and with a # specified rounding mode. It also has a :fixed mode that uses the Formatter class indirectly. # # The Formatter class implements the Burger-Dybvig printing algorithm which converts a # fixed-precision floating point value and produces a text literal in some base, usually 10, # (equivalently, it produces a floating-point free-format value) so that it rounds back to # the original value (with some specified rounding-mode or any round-to-nearest mode) and with # the same original precision (e.g. using the Clinger algorithm) # Clinger algorithms to read floating point numbers from text literals with correct rounding. # from his paper: "How to Read Floating Point Numbers Accurately" # (William D. Clinger) class Reader # There are three different reading approaches, selected by the :mode parameter: # * :fixed (the destination context defines the resulting precision) input is rounded as specified # by the context; if the context precision is 'exact', the exact input value will be represented # in the destination base, which can lead to a Inexact exception (or a NaN result and an Inexact flag) # * :free The input precision is preserved, and the destination context precision is ignored; # in this case the result can be converted back to the original number (with the same precision) # a rounding mode for the back conversion may be passed; otherwise any round-to-nearest is assumed. # (to increase the precision of the result the input precision must be increased --adding trailing zeros) # * :short is like :free, but the minumum number of digits that preserve the original value # are generated (with :free, all significant digits are generated) # # For the fixed mode there are three conversion algorithms available that can be selected with the # :algorithm parameter: # * :A Arithmetic algorithm, using correctly rounded Flt::Num arithmetic. # * :M The Clinger Algorithm M is the slowest method, but it was the first implemented and testes and # is kept as a reference for testing. # * :R The Clinger Algorithm R, which requires an initial approximation is currently only implemented # for Float and is the fastest by far. def initialize(options={}) @exact = nil @algorithm = options[:algorithm] @mode = options[:mode] || :fixed end def exact? @exact end # Given exact integers f and e, with f nonnegative, returns the floating-point number # closest to f * eb**e # (eb is the input radix) # # If the context precision is exact an Inexact exception may occur (an NaN be returned) # if an exact conversion is not possible. # # round_mode: in :fixed mode it specifies how to round the result (to the context precision); it # is passed separate from context for flexibility. # in :free mode it specifies what rounding would be used to convert back the output to the # input base eb (using the same precision that f has). def read(context, round_mode, sign, f, e, eb=10) @exact = true case @mode when :free, :short all_digits = (@mode == :free) # for free mode, (any) :nearest rounding is used by default Num.convert(Num[eb].Num(sign, f, e), context.num_class, :rounding=>round_mode||:nearest, :all_digits=>all_digits) when :fixed if context.exact? a,b = [eb, context.radix].sort m = (Math.log(b)/Math.log(a)).round if b == a**m # conmensurable bases if eb > context.radix n = AuxiliarFunctions._ndigits(f, eb)*m else n = (AuxiliarFunctions._ndigits(f, eb)+m-1)/m end else # inconmesurable bases; exact result may not be possible x = Num[eb].Num(sign, f, e) x = Num.convert_exact(x, context.num_class, context) @exact = !x.nan? return x end else n = context.precision end if round_mode == :nearest # :nearest is not meaningful here in :fixed mode; replace it if [:half_even, :half_up, :half_down].include?(context.rounding) round_mode = context.rounding else round_mode = :half_even end end # for fixed mode, use the context rounding by default round_mode ||= context.rounding alg = @algorithm if (context.radix == 2 && alg.nil?) || alg==:R z0 = _alg_r_approx(context, round_mode, sign, f, e, eb, n) alg = z0 && :R end alg ||= :A case alg when :M, :R round_mode = Support.simplified_round_mode(round_mode, sign == -1) case alg when :M _alg_m(context, round_mode, sign, f, e, eb, n) when :R _alg_r(z0, context, round_mode, sign, f, e, eb, n) end else # :A # direct arithmetic conversion if round_mode == context.rounding x = Num.convert_exact(Num[eb].Num(sign, f, e), context.num_class, context) x = context.normalize(x) unless !context.respond_to?(:normalize) || context.exact? x else if context.num_class == Float float = true context = BinNum::FloatContext end x = context.num_class.context(context) do |local_context| local_context.rounding = round_mode Num.convert_exact(Num[eb].Num(sign, f, e), local_context.num_class, local_context) end if float x = x.to_f else x = context.normalize(x) unless context.exact? end x end end end end def _alg_r_approx(context, round_mode, sign, f, e, eb, n) return nil if context.radix != Float::RADIX || context.exact? || context.precision > Float::MANT_DIG # Compute initial approximation; if Float uses IEEE-754 binary arithmetic, the approximation # is good enough to be adjusted in just one step. @good_approx = true ndigits = Support::AuxiliarFunctions._ndigits(f, eb) adj_exp = e + ndigits - 1 min_exp, max_exp = Reader.float_min_max_adj_exp(eb) if adj_exp >= min_exp && adj_exp <= max_exp if eb==2 z0 = Math.ldexp(f,e) elsif eb==10 unless Flt.float_correctly_rounded? min_exp_norm, max_exp_norm = Reader.float_min_max_adj_exp(eb, true) @good_approx = false return nil if e <= min_exp_norm end z0 = Float("#{f}E#{e}") else ff = f ee = e min_exp_norm, max_exp_norm = Reader.float_min_max_adj_exp(eb, true) if e <= min_exp_norm # avoid loss of precision due to gradual underflow return nil if e <= min_exp @good_approx = false ff = Float(f)*Float(eb)**(e-min_exp_norm-1) ee = min_exp_norm + 1 end # if ee < 0 # z0 = Float(ff)/Float(eb**(-ee)) # else # z0 = Float(ff)*Float(eb**ee) # end z0 = Float(ff)*Float(eb)**ee end if z0 && context.num_class != Float @good_approx = false z0 = context.Num(z0).plus(context) # context.plus(z0) ? else z0 = context.Num(z0) end end end def _alg_r(z0, context, round_mode, sign, f, e, eb, n) # Fast for Float #raise InvalidArgument, "Reader Algorithm R only supports base 2" if context.radix != 2 @z = z0 @r = context.radix @rp_n_1 = context.int_radix_power(n-1) @round_mode = round_mode ret = nil loop do m, k = context.to_int_scale(@z) # TODO: replace call to compare by setting the parameters in local variables, # then insert the body of compare here; # then eliminate innecesary instance variables if e >= 0 && k >= 0 ret = compare m, f*eb**e, m*@r**k, context elsif e >= 0 && k < 0 ret = compare m, f*eb**e*@r**(-k), m, context elsif e < 0 && k >= 0 ret = compare m, f, m*@r**k*eb**(-e), context else # e < 0 && k < 0 ret = compare m, f*@r**(-k), m*eb**(-e), context end break if ret end ret && context.copy_sign(ret, sign) # TODO: normalize? end @float_min_max_exp_values = { 10 => [Float::MIN_10_EXP, Float::MAX_10_EXP], Float::RADIX => [Float::MIN_EXP, Float::MAX_EXP], -Float::RADIX => [Float::MIN_EXP-Float::MANT_DIG, Float::MAX_EXP-Float::MANT_DIG] } class < d2 < y # d < 0 <=> x < y <=> v < z directed_rounding = [:up, :down].include?(@round_mode) if directed_rounding if @round_mode==:up ? (d <= 0) : (d < 0) # v <(=) z chk = (m == @rp_n_1) ? d2*@r : d2 if (@round_mode == :up) && (chk < 2*y) # eps < 1 ret = @z else @z = context.next_minus(@z) end else # @round_mode==:up ? (d > 0) : (d >= 0) # v >(=) z if (@round_mode == :down) && (d2 < 2*y) # eps < 1 ret = @z else @z = context.next_plus(@z) end end else if d2 < y # eps < 1/2 if (m == @rp_n_1) && (d < 0) && (y < @r*d2) # z has the minimum normalized significand, i.e. is a power of @r # and v < z # and @r*eps > 1/2 # On the left of z the ulp is 1/@r than the ulp on the right; if v < z we # must require an error @r times smaller. @z = context.next_minus(@z) else # unambiguous nearest ret = @z end elsif d2 == y # eps == 1/2 # round-to-nearest tie if @round_mode == :half_even if (m%2) == 0 # m is even if (m == @rp_n_1) && (d < 0) # z is power of @r and v < z; this wasn't really a tie because # there are closer values on the left @z = context.next_minus(@z) else # m is even => round tie to z ret = @z end elsif d < 0 # m is odd, v < z => round tie to prev ret = context.next_minus(@z) elsif d > 0 # m is odd, v > z => round tie to next ret = context.next_plus(@z) end elsif @round_mode == :half_up if d < 0 # v < z if (m == @rp_n_1) # this was not really a tie @z = context.next_minus(@z) else ret = @z end else # d > 0 # v >= z ret = context.next_plus(@z) end else # @round_mode == :half_down if d < 0 # v < z if (m == @rp_n_1) # this was not really a tie @z = context.next_minus(@z) else ret = context.next_minus(@z) end else # d < 0 # v > z ret = @z end end elsif d < 0 # eps > 1/2 and v < z @z = context.next_minus(@z) elsif d > 0 # eps > 1/2 and v > z @z = context.next_plus(@z) end end # Assume the initial approx is good enough (uses IEEE-754 arithmetic with round-to-nearest), # so we can avoid further iteration, except for directed rounding ret ||= @z unless directed_rounding || !@good_approx return ret end # Algorithm M to read floating point numbers from text literals with correct rounding # from his paper: "How to Read Floating Point Numbers Accurately" (William D. Clinger) def _alg_m(context, round_mode, sign, f, e, eb, n) if e<0 u,v,k = f,eb**(-e),0 else u,v,k = f*(eb**e),1,0 end min_e = context.etiny max_e = context.etop rp_n = context.int_radix_power(n) rp_n_1 = context.int_radix_power(n-1) r = context.radix loop do x = u.div(v) # bottleneck if (x>=rp_n_1 && x context.maximum_coefficient context.exception(Num::Overflow,"Input literal out of range") end end context.exception Num::Inexact if !exact end return z.copy_sign(sign) elsif x=rp_n v *= r k += 1 end end end # Given exact positive integers u and v with beta**(n-1) <= u/v < beta**n # and exact integer k, returns the floating point number closest to u/v * beta**n # (beta is the floating-point radix) def self.ratio_float(context, u, v, k, round_mode) # since this handles only positive numbers and ceiling and floor # are not symmetrical, they should have been swapped before calling this. q = u.div v r = u-q*v v_r = v-r z = context.Num(+1,q,k) exact = (r==0) if round_mode == :down # z = z elsif (round_mode == :up) && r>0 z = context.next_plus(z) elsif rv_r z = context.next_plus(z) else # tie if (round_mode == :half_down) || (round_mode == :half_even && ((q%2)==0)) || (round_mode == :down) # z = z else z = context.next_plus(z) end end return z, exact end end # Reader end end