convert_external - convert external binary data formats
cc [ flag ... ] file ... -lsunmath -lm [ library ... ] #include <sunmath.h>
enum convert_external_arch_t = {
convert_external_sparc, convert_external_pc,
convert_external_vax, convert_external_vaxg,
convert_external_s370, convert_external_cray };
enum convert_external_type_t = {
convert_external_signed,
convert_external_unsigned,
convert_external_float };
enum convert_external_rounding_t = {
convert_external_common = -1,
convert_external_biased = -2,
convert_external_away = -3 };
typedef struct {
enum convert_external_arch_t arch ; /* format architecture */
enum convert_external_type_t type ; /* format type */
int size ; /* format size in 8-bit bytes */
} convert_external_t ;
fp_exception_field_type convert_external(const char *src_p,
convert_external_t src_format, char *dst_p,
convert_external_t dst_format, int rounding, int n);
fp_exception_field_type convert_external_(const char *src_p,
const convert_external_t *src_format_p, char *dst_p,
const convert_external_t *dst_format_p, const int *rounding_p,
const int *n_p);
convert_external() is used in C programs to convert between binary data on non-SPARC systems and SPARC binary data. The data may be signed integers, unsigned integers, or floating-point data. convert_external_() is used in Fortran programs (CALL CONVERT_EXTERNAL(...)) for the same purpose, with all parameters call-by-reference.
The convert_external_t type describes the supported formats; for each architecture, signed and unsigned integers of sizes 1, 2, 4, 8, or 16 bytes, and floating-point types of sizes 4, 8, 10, 12, or 16 bytes in size are supported. If an improper size is specified for the source or destination, no conversion occurs and (1<<fp_division) is returned, an exception that can't arise otherwise.
Supported architectures and formats include
external_t endian- float*4 float*8 float*10 float*12 float*16
architec- ness format format format format format
ture
sparc big IEEE IEEE none MC68881 IEEE
single double extended quad
pc little IEEE IEEE i80x87 i80960KB IEEE
single double extended extended quad
vax little VAX F VAX D none none VAX H
vaxg little VAX F VAX G none none VAX H
s370 big S/370 S/370 none none S/370
single double extended
cray big none Cray-1 none none Cray-1
single double
Note that MC68000 and SPARC formats are identical big-endian formats; the Intel PC formats for 8086/8087, 80x86/7, and 80486 are identical little-endian. Table entries ``none'' represent improper sizes; no conversion occurs. ``IEEE quad'' refers to the natural big- or little-endian extension of IEEE single and double precision to a format with 15 exponent bits, one implicit significand bit, and 112 explicit significand bits.
excep = convert_external(src_p, src_format, dst_p, dst_format,
rounding, n);
The function value is an accumulation of all the exceptions encountered during the conversion. If detecting individual exceptions is important, n should be 1 so that only one data item is converted at a time.
A pointer to the data to be converted stored as a contiguous array. The pointer is declared char * to emphasize that there are no alignment requirements, although data being converted to SPARC format for subsequent processing by the same program should be appropriately aligned.
A pointer to the converted data stored as a contiguous array.
Description of the format of the source data.
Description of the format of the destination data.
The rounding mode to apply to the destination data.
Number of data items to be converted.
Rounding choices; the first is intended to be satisfactory for almost all applications:
integer and unsigned destination formats round toward zero (2.9 rounds to 2, -2.9 rounds to -2), IEEE floating-point destinations round to nearest unbiased, VAX floating-point destinations round to nearest biased, IBM 370 and Cray floating-point destinations round toward zero.
to round to nearest biased (for integer destinations, 1.5 rounds to 2, 2.5 rounds to 3).
to round away from zero (for integer destinations, 2.1 rounds to 3, -2.1 rounds to -3).
to obtain rounding in one of the modes specified in IEEE 754.
After the conversion, if (excep & (1 << fp_invalid)) != 0, then one or more invalid exceptions occurred; similarly for the other exceptions, which are defined this way:
An input argument was an IEEE signaling NaN or a VAX reserved operand, or an input argument was an infinity and the destination was an integer or unsigned type, or a floating-point type with no infinity representation.
An input argument was finite and large but could not be represented in the destination format with only normal rounding error; if the destination format is integer or unsigned then fp_overflow represents integer overflow.
An input argument was finite and small and could not be represented in the destination format with only normal rounding error. This can only arise with floating-point destinations.
An input argument was finite but could not be represented exactly in the destination format.
The source or destination formats aren't supported. No conversions are done.
Note that converting external data to SPARC format, processing on a SPARC, then converting back to external format will almost never yield the same results as processing the data entirely on the external system, even if the conversions and processing yield no exceptions on the SPARC.
The following code demonstrates typical Fortran generation of binary floating-point data on an external system:
REAL A(100) WRITE(9) A
Suppose the previous binary data file was created on a tape on a Cray and read on a SPARC system. To convert to IEEE double-precision floating-point format, but using biased rounding,
#include <sunmath.h>
char datasrc[800];
double datadest[100];
fp_exception_field_type excep;
int rounding;
int i;
convert_external_t src, dest;
/* read the Cray data into the array datasrc somehow, then ... */
src.arch = convert_external_cray;
src.type = convert_external_float;
src.size = 8;
dest.arch = convert_external_sparc;
dest.type = convert_external_float;
dest.size = 8;
rounding = convert_external_biased;
excep = convert_external((char *) datasrc, src, (char *) datadest,
dest, rounding, 100);
/*
* usually you are not interested in inexact exceptions
* and you'd do this
*/
excep &= ~(1 << fp_inexact);
/*
* other exceptions are possible in this example - after the fact,
* you can find out where they happened this way
*/
if (excep != 0) for (i = 0 ; i < 100 ; i++) {
excep = convert_external((char *)&datasrc[8 * i], src,
(char *)&datadest[i], dest, rounding, 1);
if (excep != 0) {
/* do something specific about datadest[i] */
}
}
Suppose data created in a SPARC program is to be read into an IBM PC. To convert,
#include <floatingpoint.h>
REAL*4 datasrc(100)
REAL*4 datadest(100)
INTEGER excep, convert_external, round
INTEGER src(4), dest(4)
c create datasrc array somehow, then ...
src(1) = convert_external_sparc
src(2) = convert_external_float
src(3) = 4
dest(1) = convert_external_pc
dest(2) = convert_external_float
dest(3) = 4
round = convert_external_common
excep = convert_external(datasrc, src, datadest, dest, round,
100);
c The only exception that can arise is fp_invalid,
c by converting a signaling NaN.
c Now write the data out in a file that an IBM PC can read
See attributes(5) for descriptions of the following attributes:
|
dd (1M) for reading tapes from foreign systems; xdr (3NSL) for formatting arbitrary data structures in a machine-independent fashion.
Conversions are performed one at a time by converting a source datum into an internal format large enough to hold any input exactly, then by converting the internal data to the destination form. Thus any pair of source and destination formats is permitted, but efficiency is not as high as would be the case for programs written to convert from one specific format to another.