From: Taylor R Campbell Date: Sun, 27 Apr 2014 03:17:43 +0000 (+0000) Subject: Add portable IEEE 754 format utilities. X-Git-Tag: release-9.2.0~14 X-Git-Url: https://birchwood-abbey.net/git?a=commitdiff_plain;h=762db656c2ef16ca20f88aef7db118fb1bb708cd;p=mit-scheme.git Add portable IEEE 754 format utilities. Not hooked up to anything, but these have been floating around my disk for months and were likely to get lost. Feel free to hook these up and start using them. If you do, you should start by writing some automatic tests. --- diff --git a/src/runtime/ieee754.scm b/src/runtime/ieee754.scm new file mode 100644 index 000000000..50d83f2a1 --- /dev/null +++ b/src/runtime/ieee754.scm @@ -0,0 +1,222 @@ +#| -*-Scheme-*- + +Copyright (C) 2013 Taylor R Campbell + +This file is part of MIT/GNU Scheme. + +MIT/GNU Scheme is free software; you can redistribute it and/or modify +it under the terms of the GNU General Public License as published by +the Free Software Foundation; either version 2 of the License, or (at +your option) any later version. + +MIT/GNU Scheme is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of +MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU +General Public License for more details. + +You should have received a copy of the GNU General Public License +along with MIT/GNU Scheme; if not, write to the Free Software +Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, +USA. + +|# + +;;;; IEEE 754 Format + +(define (decompose-ieee754-double x) + (decompose-ieee754-binary x 11 53)) + +(define (decompose-ieee754-binary x exponent-bits precision) + (receive (base emin emax bias exp-subnormal exp-inf/nan) + (ieee754-binary-parameters exponent-bits precision) + (decompose-ieee754 x base emax precision + (lambda (sign) ;if-zero + (values sign 0 0)) + (lambda (sign scaled-significand) ;if-subnormal + (assert (= 0 (shift-right scaled-significand precision))) + (values sign exp-subnormal scaled-significand)) + (lambda (sign exponent scaled-significand) ;if-normal + (assert (<= emin exponent emax)) + ;; The integer part is always 1. Strip it for the binary + ;; interchange format. + (assert (= 1 (shift-right scaled-significand (- precision 1)))) + (values sign + (+ exponent bias) + (extract-bit-field (- precision 1) 0 scaled-significand))) + (lambda (sign) ;if-infinite + (values sign exp-inf/nan 0)) + (lambda (sign quiet payload) ;if-nan + (assert (not (and (zero? quiet) (zero? payload)))) + (assert (zero? (extract-bit-field (- precision 1) 1 payload))) + (values sign + exp-inf/nan + (replace-bit-field (- precision 1) 1 payload quiet)))))) + +(define (ieee754-sign x) + (cond ((< 0 x) 0) + ((< x 0) 1) + ;; Zero -- can't use < directly to detect sign. Elicit a + ;; computational difference. + ((negative? (atan x -1)) 1) + (else 0))) + +(define (decompose-ieee754 x base emax precision + if-zero if-subnormal if-normal if-infinite if-nan) + (cond ((not (= x x)) + ;; There are, of course, b^p different NaNs. There is no + ;; obvious way to computationally detect the sign of a NaN, + ;; and no portable way to get at the quiet bit or the payload + ;; bits, so we'll just assume every NaN is a trivial positive + ;; signalling NaN and hope the caller has a good story... + (if-nan 0 0 1)) + ((and (< 1. (abs x)) (= x (/ x 2))) + (if-infinite (if (< 0. x) 0 1))) + (else + (let ((sign (ieee754-sign x)) (x (abs x)) (emin (- 1 emax))) + (define (significand x) + (truncate->exact (* x (expt base (- precision 1))))) + (cond ((<= 1 x) ;Nonnegative exponent (normal) + (let loop ((exponent 0) (x x)) + (cond ((< emax exponent) (if-infinite sign)) + ((< base x) (loop (+ exponent 1) (/ x base))) + (else (if-normal sign exponent (significand x)))))) + ((< (expt base emin) x) ;Negative exponent, normal + (let loop ((exponent 0) (x x)) + (assert (<= emin exponent)) + (if (<= 1 x) + (if-normal sign exponent (significand x)) + (loop (- exponent 1) (* x base))))) + ((< 0 x) ;Negative exponent, subnormal + (if (<= x (- (expt base emin) (expt base (- 0 precision)))) + (if-zero sign) + (if-subnormal + sign + (significand (/ x (expt base emin)))))) + (else + (if-zero sign))))))) + +(define (compose-ieee754-double sign biased-exponent trailing-significand) + (compose-ieee754-binary sign biased-exponent trailing-significand 11 53)) + +(define (compose-ieee754-binary sign biased-exponent trailing-significand + exponent-bits precision) + (receive (base emin emax bias exp-subnormal exp-inf/nan) + (ieee754-binary-parameters exponent-bits precision) + (let ((exponent (- biased-exponent bias))) + (cond ((= exponent exp-subnormal) + (if (zero? trailing-significand) + (compose-ieee754-zero sign base emax precision) + (compose-ieee754-subnormal sign trailing-significand + base emax precision))) + ((= exponent exp-inf/nan) + (if (zero? trailing-significand) + (compose-ieee754-infinity sign base emax precision) + (let ((p-1 (- precision 1)) + (T trailing-significand)) + (let ((quiet (extract-bit-field 1 p-1 T)) + (payload (extract-bit-field p-1 0 T))) + (compose-ieee754-nan sign quiet payload + base emax precision))))) + (else + (assert (<= emin exponent emax)) + (let ((scaled-significand + ;; Add the implied integer part of 1. + (replace-bit-field 1 (- precision 1) trailing-significand + 1))) + (compose-ieee754-normal sign exponent scaled-significand + base emax precision))))))) + +(define (compose-ieee754-zero sign base emax precision) + base emax precision ;ignore + (* (expt -1 sign) 0)) + +(define (compose-ieee754-subnormal sign significand base emax precision) + (* (expt -1 sign) + (* significand (expt base (- precision emax))))) + +(define (compose-ieee754-normal sign exponent significand base emax precision) + (assert (<= (- 1 emax) exponent emax)) + (* (expt -1 sign) + (expt base exponent) + (/ significand (expt base (- precision 1))))) + +(define (compose-ieee754-infinity sign) + (error "Can't compose an IEEE754 infinity!" sign)) + +(define (compose-ieee754-nan sign quiet payload) + (error "Can't compose an IEEE754 NaN!" sign quiet payload)) + +(define (ieee754-binary-parameters exponent-bits precision) + (assert (zero? (modulo (+ exponent-bits precision) 32))) + (let* ((base 2) + (emax (- (expt base (- exponent-bits 1)) 1))) + (let ((bias emax) + (emin (- 1 emax))) + (let ((exp-subnormal (- emin 1)) + (exp-inf/nan (+ emax 1))) + (values base emin emax bias exp-subnormal exp-inf/nan))))) + +(define (ieee754-double-recomposable? x) + (= x + (receive (sign biased-exponent trailing-significand) + (decompose-ieee754-double x) + (compose-ieee754-double sign biased-exponent trailing-significand)))) + +(define (ieee754-binary-hex-string x exponent-bits precision) + (receive (base emin emax bias exp-subnormal exp-inf/nan) + (ieee754-binary-parameters exponent-bits precision) + (define (symbolic sign name extra) + (assert (or (= sign 0) (= sign 1))) + (assert (<= 0 extra)) + (let ((extra (number->string extra #x10))) + (string-append (if (zero? sign) "+" "-") name "." extra))) + (define (numeric sign integer fractional exponent) + (assert (or (= sign 0) (= sign 1))) + (assert (or (= integer 0) (= integer 1))) + (assert (<= 0 fractional)) + (let ((sign (if (zero? sign) "" "-")) + (integer (if (zero? integer) "0" "1")) + (fractional + (if (zero? fractional) "" (number->string fractional #x10))) + (exponent (number->string exponent #d10))) + (string-append sign "0x" integer "." fractional "p" exponent))) + (decompose-ieee754 x base emax precision + (lambda (sign) ;if-zero + (numeric sign 0 0 0)) + (lambda (sign scaled-significand) ;if-subnormal + (assert (< 0 scaled-significand)) + (let ((start (first-set-bit scaled-significand)) + (end (integer-length scaled-significand))) + (let ((bits (- (- end 1) start))) + (let ((exponent (- emin (- precision end))) + ;; Strip the integer part (1) and the trailing zeros. + (fractional + (extract-bit-field bits start scaled-significand))) + ;; Align to a multiple of four bits (hex digits). + (let* ((alignment (modulo (- 4 (modulo bits 4)) 4)) + (aligned-fractional + (replace-bit-field bits alignment 0 fractional))) + (numeric sign 1 aligned-fractional exponent)))))) + (lambda (sign exponent scaled-significand) + (assert (< 0 scaled-significand)) + (assert (= 1 (shift-right scaled-significand (- precision 1)))) + (let ((useless-zeros (round-down (first-set-bit scaled-significand) 4)) + (fractional + (extract-bit-field (- precision 1) 0 scaled-significand))) + (numeric sign 1 (shift-right fractional useless-zeros) exponent))) + (lambda (sign) + (symbolic sign "inf" 0)) + (lambda (sign quiet payload) + (symbolic sign (if (zero? quiet) "sNaN" "qNaN") payload))))) + +(define (round-up x n) + (* n (quotient (+ x (- n 1)) n))) + +(define (round-down x n) + (* n (quotient x n))) + +(define (round-up2 x n) + (+ 1 (bitwise-ior (- x 1) (- n 1)))) + +(define (round-down2 x n) + (bitwise-andc2 x (- n 1)))