| 1 |
#ifndef lint
|
| 2 |
static const char RCSid[] = "$Id: pvsum.c,v 2.9 2025/10/29 02:48:50 greg Exp $";
|
| 3 |
#endif
|
| 4 |
/*
|
| 5 |
* pvsum.c - add together spectral and/or float pictures
|
| 6 |
* based on vector or matrix, similar to dctimestep
|
| 7 |
*/
|
| 8 |
|
| 9 |
#include <math.h>
|
| 10 |
#include "rtio.h"
|
| 11 |
#include "resolu.h"
|
| 12 |
#include "platform.h"
|
| 13 |
#include "random.h"
|
| 14 |
#include "rmatrix.h"
|
| 15 |
#if !defined(_WIN32) && !defined(_WIN64)
|
| 16 |
#include <sys/mman.h>
|
| 17 |
#include <sys/wait.h>
|
| 18 |
#endif
|
| 19 |
|
| 20 |
#define VIEWSTR "VIEW=" /* borrowed from common/view.h */
|
| 21 |
#define VIEWSTRL 5
|
| 22 |
|
| 23 |
int nprocs = 1; /* # of calculation processes (Unix) */
|
| 24 |
int in_type = DTfromHeader; /* input data type */
|
| 25 |
int out_type = DTfromHeader; /* output data type */
|
| 26 |
char *in_spec = NULL; /* input specification */
|
| 27 |
char *out_spec = NULL; /* output file specification */
|
| 28 |
|
| 29 |
int iswapped = 0; /* input data is byte-swapped? */
|
| 30 |
int ncomp = 3; /* # input components */
|
| 31 |
int xres=0, yres=0; /* input image dimensions */
|
| 32 |
char viewspec[128] = ""; /* VIEW= line from first header */
|
| 33 |
char pixasp[48] = ""; /* PIXASPECT= line from header */
|
| 34 |
|
| 35 |
int gargc; /* global argc */
|
| 36 |
char **gargv; /* global argv */
|
| 37 |
|
| 38 |
RMATRIX *cmtx; /* coefficient matrix */
|
| 39 |
int row0, rowN; /* rows for current pass */
|
| 40 |
|
| 41 |
/* does the given spec contain integer format? */
|
| 42 |
int
|
| 43 |
hasFormat(const char *s)
|
| 44 |
{
|
| 45 |
restart:
|
| 46 |
if (s) s = strchr(s, '%');
|
| 47 |
if (!s) return(0);
|
| 48 |
if (s[1] == '%') { /* "%%" ? */
|
| 49 |
s += 2;
|
| 50 |
goto restart;
|
| 51 |
}
|
| 52 |
while (*++s) {
|
| 53 |
if (strchr("diouxX", *s))
|
| 54 |
return(1); /* integer format */
|
| 55 |
if (strchr("%fFeEgGaAcsb", *s))
|
| 56 |
break; /* non-integer format */
|
| 57 |
}
|
| 58 |
return(0);
|
| 59 |
}
|
| 60 |
|
| 61 |
/* get first input header data we'll need */
|
| 62 |
int
|
| 63 |
iheadline(char *s, void *p)
|
| 64 |
{
|
| 65 |
char fmt[MAXFMTLEN];
|
| 66 |
int i;
|
| 67 |
|
| 68 |
if (!strncmp(s, "NCOLS=", 6)) {
|
| 69 |
xres = atoi(s+6);
|
| 70 |
return(1);
|
| 71 |
}
|
| 72 |
if (!strncmp(s, "NROWS=", 6)) {
|
| 73 |
yres = atoi(s+6);
|
| 74 |
return(1);
|
| 75 |
}
|
| 76 |
if (isncomp(s)) {
|
| 77 |
ncomp = ncompval(s);
|
| 78 |
return(1);
|
| 79 |
}
|
| 80 |
if (formatval(fmt, s)) {
|
| 81 |
for (in_type = DTend; --in_type > DTfromHeader; )
|
| 82 |
if (!strcmp(fmt, cm_fmt_id[in_type]))
|
| 83 |
return(1);
|
| 84 |
return(-1);
|
| 85 |
}
|
| 86 |
i = isbigendian(s);
|
| 87 |
if (i >= 0) {
|
| 88 |
iswapped = (i != nativebigendian());
|
| 89 |
return(1);
|
| 90 |
}
|
| 91 |
if (!strncmp(s, VIEWSTR, VIEWSTRL)) {
|
| 92 |
strcpy(viewspec, s);
|
| 93 |
return(1);
|
| 94 |
}
|
| 95 |
if (isaspect(s)) {
|
| 96 |
strcpy(pixasp, s);
|
| 97 |
return(1);
|
| 98 |
}
|
| 99 |
return(0);
|
| 100 |
}
|
| 101 |
|
| 102 |
/* open initial file and get relevant dimensions and type */
|
| 103 |
int
|
| 104 |
get_iotypes(void)
|
| 105 |
{
|
| 106 |
char fbuf[256];
|
| 107 |
FILE *fp;
|
| 108 |
|
| 109 |
sprintf(fbuf, in_spec, 0);
|
| 110 |
fp = fopen(fbuf, "rb");
|
| 111 |
if (!fp) {
|
| 112 |
fprintf(stderr, "%s: cannot open for reading\n", fbuf);
|
| 113 |
return(0);
|
| 114 |
}
|
| 115 |
if (getheader(fp, iheadline, NULL) < 0) {
|
| 116 |
fprintf(stderr, "%s: bad header - wrong format?\n", fbuf);
|
| 117 |
fclose(fp);
|
| 118 |
return(0);
|
| 119 |
}
|
| 120 |
if ((xres <= 0) | (yres <= 0) && !fscnresolu(&xres, &yres, fp)) {
|
| 121 |
fprintf(stderr, "%s: missing input resolution\n", fbuf);
|
| 122 |
fclose(fp);
|
| 123 |
return(0);
|
| 124 |
}
|
| 125 |
if (nprocs > 1 && (in_type==DTrgbe) | (in_type==DTxyze)) {
|
| 126 |
long data_start = ftell(fp); /* make sure input flat */
|
| 127 |
off_t dend = lseek(fileno(fp), 0, SEEK_END);
|
| 128 |
if (dend < data_start + 4L*xres*yres) {
|
| 129 |
fprintf(stderr, "%s: warning - multi-processing requires flat input files\n",
|
| 130 |
gargv[0]);
|
| 131 |
nprocs = 1;
|
| 132 |
}
|
| 133 |
}
|
| 134 |
fclose(fp);
|
| 135 |
if ((cmtx->ncomp == 1) & (ncomp != 1)) {
|
| 136 |
double xfm[MAXCSAMP];
|
| 137 |
RMATRIX *nmtx;
|
| 138 |
int i;
|
| 139 |
for (i = ncomp; i--; )
|
| 140 |
xfm[i] = 1.;
|
| 141 |
nmtx = rmx_transform(cmtx, ncomp, xfm);
|
| 142 |
if (!nmtx)
|
| 143 |
return(0);
|
| 144 |
rmx_free(cmtx);
|
| 145 |
cmtx = nmtx;
|
| 146 |
} else if (cmtx->ncomp != ncomp) {
|
| 147 |
fprintf(stderr, "%s: operation %s needs %d components, has %d\n",
|
| 148 |
gargv[0], cmtx->ncols == 1 ? "vector" : "matrix",
|
| 149 |
ncomp, cmtx->ncomp);
|
| 150 |
return(0);
|
| 151 |
}
|
| 152 |
if ((in_type != DTrgbe) & (in_type != DTxyze) & (in_type != DTspec) &
|
| 153 |
(in_type != DTfloat)) {
|
| 154 |
fprintf(stderr, "%s: unsupported input data type '%s'\n",
|
| 155 |
fbuf, cm_fmt_id[in_type]);
|
| 156 |
return(0);
|
| 157 |
}
|
| 158 |
if ((out_type == DTrgbe) & (ncomp > 3))
|
| 159 |
out_type = DTspec;
|
| 160 |
else if (out_type == DTfromHeader ||
|
| 161 |
(out_type == DTrgbe) & (in_type != DTfloat))
|
| 162 |
out_type = in_type;
|
| 163 |
return(1);
|
| 164 |
}
|
| 165 |
|
| 166 |
struct hdata {
|
| 167 |
int xr, yr; /* resolution */
|
| 168 |
int fno; /* frame # */
|
| 169 |
char fmt[MAXFMTLEN]; /* format */
|
| 170 |
};
|
| 171 |
|
| 172 |
/* check subsequent headers match initial file */
|
| 173 |
int
|
| 174 |
checkline(char *s, void *p)
|
| 175 |
{
|
| 176 |
static int exposWarned = 0;
|
| 177 |
struct hdata *hp = (struct hdata *)p;
|
| 178 |
|
| 179 |
if (!strncmp(s, "NCOLS=", 6)) {
|
| 180 |
hp->xr = atoi(s+6);
|
| 181 |
if (hp->xr <= 0)
|
| 182 |
return(-1);
|
| 183 |
return(1);
|
| 184 |
}
|
| 185 |
if (!strncmp(s, "NROWS=", 6)) {
|
| 186 |
hp->yr = atoi(s+6);
|
| 187 |
if (hp->yr <= 0)
|
| 188 |
return(-1);
|
| 189 |
return(1);
|
| 190 |
}
|
| 191 |
if (!strncmp(s, "FRAME=", 6)) {
|
| 192 |
hp->fno = atoi(s+6);
|
| 193 |
return(1);
|
| 194 |
}
|
| 195 |
if (isncomp(s)) {
|
| 196 |
if (ncompval(s) != ncomp)
|
| 197 |
return(-1);
|
| 198 |
return(1);
|
| 199 |
}
|
| 200 |
if (isexpos(s)) {
|
| 201 |
if (!exposWarned && fabs(1. - exposval(s)) > 0.04) {
|
| 202 |
fprintf(stderr, "%s: warning - ignoring EXPOSURE setting(s)\n",
|
| 203 |
gargv[0]);
|
| 204 |
exposWarned++;
|
| 205 |
}
|
| 206 |
return(1);
|
| 207 |
}
|
| 208 |
if (formatval(hp->fmt, s))
|
| 209 |
return(1);
|
| 210 |
|
| 211 |
return(0);
|
| 212 |
}
|
| 213 |
|
| 214 |
/* open and check input/output file, read/write mode if fno >= 0 */
|
| 215 |
FILE *
|
| 216 |
open_iofile(char *fname, int fno)
|
| 217 |
{
|
| 218 |
struct hdata hd;
|
| 219 |
FILE *fp = fopen(fname, fno>=0 ? "r+b" : "rb");
|
| 220 |
|
| 221 |
if (!fp) {
|
| 222 |
fprintf(stderr, "%s: cannot open for reading%s\n",
|
| 223 |
fname, fno>=0 ? "/writing" : "");
|
| 224 |
return(NULL);
|
| 225 |
}
|
| 226 |
hd.xr = hd.yr = 0;
|
| 227 |
hd.fno = -1;
|
| 228 |
hd.fmt[0] = '\0';
|
| 229 |
if (getheader(fp, checkline, &hd) < 0) {
|
| 230 |
fprintf(stderr, "%s: bad/inconsistent header\n", fname);
|
| 231 |
fclose(fp);
|
| 232 |
return(NULL);
|
| 233 |
}
|
| 234 |
if ((hd.fno >= 0) & (fno >= 0) & (hd.fno != fno)) {
|
| 235 |
fprintf(stderr, "%s: unexpected frame number (%d != %d)\n",
|
| 236 |
fname, hd.fno, fno);
|
| 237 |
fclose(fp);
|
| 238 |
return(NULL);
|
| 239 |
}
|
| 240 |
if (strcmp(hd.fmt, cm_fmt_id[fno>=0 ? out_type : in_type])) {
|
| 241 |
fprintf(stderr, "%s: wrong format\n", fname);
|
| 242 |
fclose(fp);
|
| 243 |
return(NULL);
|
| 244 |
}
|
| 245 |
if ((hd.xr <= 0) | (hd.yr <= 0) &&
|
| 246 |
!fscnresolu(&hd.xr, &hd.yr, fp)) {
|
| 247 |
fprintf(stderr, "%s: missing resolution\n", fname);
|
| 248 |
fclose(fp);
|
| 249 |
return(NULL);
|
| 250 |
}
|
| 251 |
if ((hd.xr != xres) | (hd.yr != yres)) {
|
| 252 |
fprintf(stderr, "%s: mismatched resolution\n", fname);
|
| 253 |
fclose(fp);
|
| 254 |
return(NULL);
|
| 255 |
}
|
| 256 |
return(fp);
|
| 257 |
}
|
| 258 |
|
| 259 |
/* read in previous pixel data from output and rewind to data start */
|
| 260 |
int
|
| 261 |
reload_data(float *osum, FILE *fp)
|
| 262 |
{
|
| 263 |
long dstart;
|
| 264 |
|
| 265 |
if (!osum | !fp)
|
| 266 |
return(0);
|
| 267 |
if ((dstart = ftell(fp)) < 0) {
|
| 268 |
fprintf(stderr, "%s: ftell() error in reload_data()\n",
|
| 269 |
gargv[0]);
|
| 270 |
return(0);
|
| 271 |
}
|
| 272 |
if (in_type == DTfloat) {
|
| 273 |
if (fread(osum, sizeof(float)*ncomp, (size_t)xres*yres, fp) !=
|
| 274 |
(size_t)xres*yres) {
|
| 275 |
fprintf(stderr, "%s: fread() error\n", gargv[0]);
|
| 276 |
return(0);
|
| 277 |
}
|
| 278 |
} else {
|
| 279 |
int y;
|
| 280 |
for (y = 0; y < yres; y++, osum += ncomp*xres)
|
| 281 |
if (freadsscan(osum, ncomp, xres, fp) < 0) {
|
| 282 |
fprintf(stderr, "%s: freadsscan() error\n", gargv[0]);
|
| 283 |
return(0);
|
| 284 |
}
|
| 285 |
}
|
| 286 |
if (fseek(fp, dstart, SEEK_SET) < 0) {
|
| 287 |
fprintf(stderr, "%s: fseek() error in reload_data()\n",
|
| 288 |
gargv[0]);
|
| 289 |
return(0);
|
| 290 |
}
|
| 291 |
return(1);
|
| 292 |
}
|
| 293 |
|
| 294 |
/* open output file or command (if !NULL) and write info header */
|
| 295 |
FILE *
|
| 296 |
open_output(char *ospec, int fno)
|
| 297 |
{
|
| 298 |
FILE *fp;
|
| 299 |
|
| 300 |
if (!ospec) {
|
| 301 |
ospec = "<stdout>";
|
| 302 |
fp = stdout;
|
| 303 |
} else if (ospec[0] == '!') {
|
| 304 |
if (!(fp = popen(ospec+1, "w"))) {
|
| 305 |
fprintf(stderr, "%s: cannot start: %s\n", gargv[0], ospec);
|
| 306 |
return(NULL);
|
| 307 |
}
|
| 308 |
} else if (!(fp = fopen(ospec, "w"))) {
|
| 309 |
fprintf(stderr, "%s: cannot open for writing\n", ospec);
|
| 310 |
return(NULL);
|
| 311 |
}
|
| 312 |
SET_FILE_BINARY(fp);
|
| 313 |
newheader("RADIANCE", fp);
|
| 314 |
if (cmtx->info) /* prepend matrix metadata */
|
| 315 |
fputs(cmtx->info, fp);
|
| 316 |
else
|
| 317 |
fputnow(fp);
|
| 318 |
printargs(gargc, gargv, fp); /* this command */
|
| 319 |
if (fno >= 0)
|
| 320 |
fprintf(fp, "FRAME=%d\n", fno);
|
| 321 |
if (viewspec[0])
|
| 322 |
fputs(viewspec, fp);
|
| 323 |
if (pixasp[0])
|
| 324 |
fputs(pixasp, fp);
|
| 325 |
switch (out_type) {
|
| 326 |
case DTfloat:
|
| 327 |
case DTdouble:
|
| 328 |
fprintf(fp, "NCOLS=%d\nNROWS=%d\n", xres, yres);
|
| 329 |
fputncomp(ncomp, fp);
|
| 330 |
fputendian(fp);
|
| 331 |
fputformat(cm_fmt_id[out_type], fp);
|
| 332 |
fputc('\n', fp);
|
| 333 |
break;
|
| 334 |
case DTrgbe:
|
| 335 |
fputformat(COLRFMT, fp);
|
| 336 |
fputc('\n', fp);
|
| 337 |
fprtresolu(xres, yres, fp);
|
| 338 |
break;
|
| 339 |
case DTxyze:
|
| 340 |
fputformat(CIEFMT, fp);
|
| 341 |
fputc('\n', fp);
|
| 342 |
fprtresolu(xres, yres, fp);
|
| 343 |
break;
|
| 344 |
case DTspec:
|
| 345 |
fputncomp(ncomp, fp);
|
| 346 |
fputwlsplit(cmtx->wlpart, fp);
|
| 347 |
fputformat(SPECFMT, fp);
|
| 348 |
fputc('\n', fp);
|
| 349 |
fprtresolu(xres, yres, fp);
|
| 350 |
break;
|
| 351 |
default:
|
| 352 |
fprintf(stderr, "%s: unsupported output type!\n", gargv[0]);
|
| 353 |
return(NULL);
|
| 354 |
}
|
| 355 |
if (fflush(fp) < 0) {
|
| 356 |
fprintf(stderr, "%s: write error\n", ospec);
|
| 357 |
fclose(fp);
|
| 358 |
return(NULL);
|
| 359 |
}
|
| 360 |
return(fp);
|
| 361 |
}
|
| 362 |
|
| 363 |
/* run calculation from a single process */
|
| 364 |
int
|
| 365 |
solo_process(void)
|
| 366 |
{
|
| 367 |
float *osum = (float *)calloc((size_t)xres*yres, sizeof(float)*ncomp);
|
| 368 |
COLORV *iscan = (COLORV *)malloc(sizeof(COLORV)*ncomp*xres);
|
| 369 |
char fbuf[512];
|
| 370 |
int c;
|
| 371 |
|
| 372 |
if (!osum | !iscan) {
|
| 373 |
fprintf(stderr, "%s: annot allocate %dx%d %d-component accumulator\n",
|
| 374 |
gargv[0], xres, yres, ncomp);
|
| 375 |
return(0);
|
| 376 |
}
|
| 377 |
if (sizeof(float) != sizeof(COLORV)) {
|
| 378 |
fprintf(stderr, "%s: Code Error 1 in solo_process()\n", gargv[0]);
|
| 379 |
return(0);
|
| 380 |
}
|
| 381 |
/* run through each column/output */
|
| 382 |
for (c = 0; c < cmtx->ncols; c++) {
|
| 383 |
int rc = rowN - row0;
|
| 384 |
FILE *fout;
|
| 385 |
int y;
|
| 386 |
/* open output (load if multipass) */
|
| 387 |
if (out_spec) { /* file or command */
|
| 388 |
if (cmtx->ncols > 1 && !hasFormat(out_spec)) {
|
| 389 |
fprintf(stderr, "%s: sequential result must go to stdout\n",
|
| 390 |
gargv[0]);
|
| 391 |
return(0);
|
| 392 |
}
|
| 393 |
sprintf(fbuf, out_spec, c);
|
| 394 |
if (row0) { /* another pass -- get prev. data */
|
| 395 |
fout = open_iofile(fbuf, c);
|
| 396 |
if (!reload_data(osum, fout))
|
| 397 |
return(0);
|
| 398 |
} else /* else new output (clobber prev. files) */
|
| 399 |
fout = open_output(fbuf, c-(cmtx->ncols==1));
|
| 400 |
} else { /* else stdout */
|
| 401 |
if ((out_type == DTfloat) & (cmtx->ncols > 1)) {
|
| 402 |
fprintf(stderr, "%s: float outputs must have separate destinations\n",
|
| 403 |
gargv[0]);
|
| 404 |
return(0);
|
| 405 |
}
|
| 406 |
strcpy(fbuf, "<stdout>");
|
| 407 |
fout = open_output(NULL, c-(cmtx->ncols==1));
|
| 408 |
}
|
| 409 |
if (!fout)
|
| 410 |
return(0); /* assume error was reported */
|
| 411 |
if (!row0 & (c > 0)) /* clear accumulator? */
|
| 412 |
memset(osum, 0, sizeof(float)*ncomp*xres*yres);
|
| 413 |
while (rc-- > 0) { /* run through each input file */
|
| 414 |
const int r = c&1 ? row0 + rc : rowN-1 - rc;
|
| 415 |
const rmx_dtype *cval = rmx_val(cmtx, r, c);
|
| 416 |
FILE *finp;
|
| 417 |
int i, x;
|
| 418 |
for (i = ncomp; i--; )
|
| 419 |
if (cval[i] != 0) break;
|
| 420 |
if (i < 0) /* this coefficient is zero, skip */
|
| 421 |
continue;
|
| 422 |
sprintf(fbuf, in_spec, r);
|
| 423 |
finp = open_iofile(fbuf, -1);
|
| 424 |
if (!finp)
|
| 425 |
return(0);
|
| 426 |
for (y = 0; y < yres; y++) {
|
| 427 |
float *dst = osum + y*xres*ncomp;
|
| 428 |
if (in_type == DTfloat ? getbinary(iscan, sizeof(float)*ncomp,
|
| 429 |
xres, finp) != xres
|
| 430 |
: freadsscan(iscan, ncomp, xres, finp) < 0)
|
| 431 |
goto readerr;
|
| 432 |
if ((in_type == DTfloat) & iswapped)
|
| 433 |
swap32((char *)iscan, ncomp*xres);
|
| 434 |
for (x = 0; x < xres; x++, dst += ncomp)
|
| 435 |
for (i = ncomp; i--; )
|
| 436 |
dst[i] += cval[i]*iscan[x*ncomp + i];
|
| 437 |
}
|
| 438 |
fclose(finp);
|
| 439 |
} /* write accumulated picture */
|
| 440 |
if (out_type != DTfloat) {
|
| 441 |
for (y = 0; y < yres; y++)
|
| 442 |
if (fwritesscan(osum + (size_t)y*xres*ncomp,
|
| 443 |
ncomp, xres, fout) < 0)
|
| 444 |
goto writerr;
|
| 445 |
} else if (fwrite(osum, sizeof(float)*ncomp, (size_t)xres*yres, fout) !=
|
| 446 |
(size_t)xres*yres)
|
| 447 |
goto writerr;
|
| 448 |
|
| 449 |
if (fbuf[0] == '!') {
|
| 450 |
if (pclose(fout) != 0) {
|
| 451 |
fprintf(stderr, "%s: bad status from: %s\n", gargv[0], fbuf);
|
| 452 |
return(0);
|
| 453 |
}
|
| 454 |
} else if (fout != stdout && fclose(fout) == EOF)
|
| 455 |
goto writerr;
|
| 456 |
}
|
| 457 |
free(osum); /* clean up on success */
|
| 458 |
free(iscan);
|
| 459 |
return(1);
|
| 460 |
readerr:
|
| 461 |
fprintf(stderr, "%s: read error\n", fbuf);
|
| 462 |
return(0);
|
| 463 |
writerr:
|
| 464 |
fprintf(stderr, "%s: write error\n", fbuf);
|
| 465 |
return(0);
|
| 466 |
}
|
| 467 |
|
| 468 |
#if defined(_WIN32) || defined(_WIN64)
|
| 469 |
#define multi_process solo_process
|
| 470 |
#else
|
| 471 |
|
| 472 |
/* allocate a scrambled index array of the specified length */
|
| 473 |
int *
|
| 474 |
scramble(int n)
|
| 475 |
{
|
| 476 |
int *scarr = (int *)malloc(sizeof(int)*n);
|
| 477 |
int i;
|
| 478 |
|
| 479 |
if (!scarr) {
|
| 480 |
fprintf(stderr, "%s: out of memory in scramble(%d)\n", gargv[0], n);
|
| 481 |
exit(1);
|
| 482 |
}
|
| 483 |
for (i = n; i--; )
|
| 484 |
scarr[i] = i;
|
| 485 |
/* perform Fisher-Yates shuffle */
|
| 486 |
for (i = 0; i < n-1; i++) {
|
| 487 |
int ix = irandom(n-i) + i;
|
| 488 |
int ndx = scarr[i];
|
| 489 |
scarr[i] = scarr[ix];
|
| 490 |
scarr[ix] = ndx;
|
| 491 |
}
|
| 492 |
return(scarr);
|
| 493 |
}
|
| 494 |
|
| 495 |
/* run calculation on multiple processes, using memory maps and fork() */
|
| 496 |
int
|
| 497 |
multi_process(void)
|
| 498 |
{
|
| 499 |
int coff = nprocs;
|
| 500 |
int odd = 0;
|
| 501 |
float *osum = NULL;
|
| 502 |
int *syarr = NULL;
|
| 503 |
char fbuf[512];
|
| 504 |
int c;
|
| 505 |
/* sanity check */
|
| 506 |
if (sizeof(float) != sizeof(COLORV)) {
|
| 507 |
fprintf(stderr, "%s: code Error 1 in multi_process()\n", gargv[0]);
|
| 508 |
return(0);
|
| 509 |
}
|
| 510 |
fflush(NULL); /* parent births helper subprocs */
|
| 511 |
while (--coff > 0) {
|
| 512 |
int pid = fork();
|
| 513 |
if (pid < 0) {
|
| 514 |
fprintf(stderr, "%s: fork() call failed!\n", gargv[0]);
|
| 515 |
return(0);
|
| 516 |
}
|
| 517 |
if (!pid) break; /* new child gets to work */
|
| 518 |
}
|
| 519 |
if (!row0 | (out_type != DTfloat)) {
|
| 520 |
osum = (float *)calloc((size_t)xres*yres, sizeof(float)*ncomp);
|
| 521 |
if (!osum) {
|
| 522 |
fprintf(stderr, "%s: cannot allocate %dx%d %d-component accumulator\n",
|
| 523 |
gargv[0], xres, yres, ncomp);
|
| 524 |
return(0);
|
| 525 |
}
|
| 526 |
}
|
| 527 |
srandom(113*coff + 5669); /* randomize row access for this process */
|
| 528 |
syarr = scramble(yres);
|
| 529 |
/* run through our unique set of columns */
|
| 530 |
for (c = coff; c < cmtx->ncols; c += nprocs) {
|
| 531 |
int rc = rowN - row0;
|
| 532 |
void *omap = NULL;
|
| 533 |
size_t omaplen = 0;
|
| 534 |
long dstart;
|
| 535 |
FILE *fout;
|
| 536 |
int y;
|
| 537 |
/* create/load output */
|
| 538 |
sprintf(fbuf, out_spec, c);
|
| 539 |
if (row0) { /* making another pass? */
|
| 540 |
fout = open_iofile(fbuf, c);
|
| 541 |
if (!fout) return(0);
|
| 542 |
if (out_type == DTfloat) {
|
| 543 |
dstart = ftell(fout);
|
| 544 |
if ((dstart < 0) | (dstart % sizeof(float))) {
|
| 545 |
fprintf(stderr, "%s: bad seek/alignment\n", fbuf);
|
| 546 |
return(0);
|
| 547 |
}
|
| 548 |
omaplen = dstart + sizeof(float)*ncomp*xres*yres;
|
| 549 |
omap = mmap(NULL, omaplen, PROT_READ|PROT_WRITE,
|
| 550 |
MAP_FILE|MAP_SHARED, fileno(fout), 0);
|
| 551 |
if (omap == MAP_FAILED) {
|
| 552 |
fprintf(stderr, "%s: cannot map file '%s'\n",
|
| 553 |
gargv[0], fbuf);
|
| 554 |
return(0);
|
| 555 |
}
|
| 556 |
osum = (float *)((char *)omap + dstart);
|
| 557 |
} else if (!reload_data(osum, fout))
|
| 558 |
return(0);
|
| 559 |
} else { /* else new output (clobber prev. files) */
|
| 560 |
fout = open_output(fbuf, c);
|
| 561 |
if (!fout) return(0);
|
| 562 |
if (c > coff) /* clear accumulator? */
|
| 563 |
memset(osum, 0, sizeof(float)*ncomp*xres*yres);
|
| 564 |
}
|
| 565 |
while (rc-- > 0) { /* map & sum each input file */
|
| 566 |
const int r = odd ? row0 + rc : rowN-1 - rc;
|
| 567 |
const rmx_dtype *cval = rmx_val(cmtx, r, c);
|
| 568 |
size_t imaplen;
|
| 569 |
void *imap;
|
| 570 |
FILE *finp;
|
| 571 |
float *dst;
|
| 572 |
int i, x;
|
| 573 |
for (i = ncomp; i--; )
|
| 574 |
if (cval[i] != 0) break;
|
| 575 |
if (i < 0) /* this coefficient is zero, skip */
|
| 576 |
continue;
|
| 577 |
sprintf(fbuf, in_spec, r);
|
| 578 |
finp = open_iofile(fbuf, -1);
|
| 579 |
if (!finp)
|
| 580 |
return(0);
|
| 581 |
dstart = ftell(finp);
|
| 582 |
if (dstart < 0) {
|
| 583 |
fprintf(stderr, "%s: ftell() failed!\n", fbuf);
|
| 584 |
return(0);
|
| 585 |
}
|
| 586 |
if (in_type == DTfloat && dstart%sizeof(float)) {
|
| 587 |
fprintf(stderr, "%s: float header misalignment\n", fbuf);
|
| 588 |
return(0);
|
| 589 |
}
|
| 590 |
i = in_type==DTfloat ? ncomp*(int)sizeof(float) : ncomp+1;
|
| 591 |
imaplen = dstart + (size_t)yres*xres*i;
|
| 592 |
imap = mmap(NULL, imaplen, PROT_READ,
|
| 593 |
MAP_FILE|MAP_SHARED, fileno(finp), 0);
|
| 594 |
fclose(finp); /* will read from map (randomly) */
|
| 595 |
if (imap == MAP_FAILED) {
|
| 596 |
fprintf(stderr, "%s: unable to map input file\n", fbuf);
|
| 597 |
return(0);
|
| 598 |
}
|
| 599 |
if (in_type == DTfloat)
|
| 600 |
for (y = yres; y-- > 0; ) {
|
| 601 |
const float *fvp = (float *)((char *)imap + dstart) +
|
| 602 |
(size_t)ncomp*xres*syarr[y];
|
| 603 |
dst = osum + (size_t)ncomp*xres*syarr[y];
|
| 604 |
for (x = xres; x-- > 0; dst += ncomp, fvp += ncomp)
|
| 605 |
for (i = ncomp; i--; )
|
| 606 |
dst[i] += cval[i]*fvp[i];
|
| 607 |
}
|
| 608 |
else
|
| 609 |
for (y = yres; y-- > 0; ) {
|
| 610 |
const COLRV *cvp = (COLRV *)((char *)imap + dstart) +
|
| 611 |
(ncomp+1L)*xres*syarr[y];
|
| 612 |
dst = osum + (size_t)ncomp*xres*syarr[y];
|
| 613 |
for (x = xres; x-- > 0; dst += ncomp, cvp += ncomp+1) {
|
| 614 |
const rmx_dtype fe = cxponent[cvp[ncomp]];
|
| 615 |
for (i = ncomp; i--; )
|
| 616 |
dst[i] += cval[i]*(cvp[i]+(rmx_dtype).5)*fe;
|
| 617 |
}
|
| 618 |
}
|
| 619 |
munmap(imap, imaplen);
|
| 620 |
} /* write accumulated column picture */
|
| 621 |
if (out_type != DTfloat) {
|
| 622 |
for (y = 0; y < yres; y++)
|
| 623 |
if (fwritesscan(osum + (size_t)y*xres*ncomp,
|
| 624 |
ncomp, xres, fout) < 0)
|
| 625 |
goto writerr;
|
| 626 |
} else if (omap) {
|
| 627 |
if (munmap(omap, omaplen) < 0)
|
| 628 |
goto writerr;
|
| 629 |
osum = NULL;
|
| 630 |
} else if (fwrite(osum, sizeof(float)*ncomp, (size_t)xres*yres, fout) !=
|
| 631 |
(size_t)xres*yres)
|
| 632 |
goto writerr;
|
| 633 |
|
| 634 |
if (fbuf[0] == '!') {
|
| 635 |
if (pclose(fout) != 0) {
|
| 636 |
fprintf(stderr, "%s: bad status from: %s\n", gargv[0], fbuf);
|
| 637 |
return(0);
|
| 638 |
}
|
| 639 |
} else if (fclose(fout) == EOF)
|
| 640 |
goto writerr;
|
| 641 |
odd = !odd; /* go back & forth to milk page cache */
|
| 642 |
}
|
| 643 |
if (coff) _exit(0); /* child exits here */
|
| 644 |
/* but parent waits for children */
|
| 645 |
if (osum) free(osum);
|
| 646 |
free(syarr);
|
| 647 |
c = 0;
|
| 648 |
while (++coff < nprocs) {
|
| 649 |
int st;
|
| 650 |
if (wait(&st) < 0) {
|
| 651 |
fprintf(stderr, "%s: warning - wait() call failed unexpectedly\n", gargv[0]);
|
| 652 |
break;
|
| 653 |
}
|
| 654 |
if (st) c = st;
|
| 655 |
}
|
| 656 |
return(c == 0);
|
| 657 |
writerr:
|
| 658 |
fprintf(stderr, "%s: write error\n", fbuf);
|
| 659 |
return(0);
|
| 660 |
}
|
| 661 |
|
| 662 |
#endif /* ! Windows */
|
| 663 |
|
| 664 |
int
|
| 665 |
main(int argc, char *argv[])
|
| 666 |
{
|
| 667 |
double cacheGB = 0;
|
| 668 |
int rintvl;
|
| 669 |
int a;
|
| 670 |
|
| 671 |
gargc = argc; /* for header output */
|
| 672 |
gargv = argv;
|
| 673 |
|
| 674 |
for (a = 1; a < argc-1 && argv[a][0] == '-'; a++)
|
| 675 |
switch (argv[a][1]) {
|
| 676 |
case 'o': /* output spec/format */
|
| 677 |
switch (argv[a][2]) {
|
| 678 |
case '\0':
|
| 679 |
out_spec = argv[++a];
|
| 680 |
break;
|
| 681 |
case 'f':
|
| 682 |
out_type = DTfloat;
|
| 683 |
break;
|
| 684 |
case 'c':
|
| 685 |
out_type = DTrgbe;
|
| 686 |
break;
|
| 687 |
default:
|
| 688 |
goto badopt;
|
| 689 |
}
|
| 690 |
break;
|
| 691 |
case 'm': /* cache size in GigaBytes */
|
| 692 |
cacheGB = atof(argv[++a]);
|
| 693 |
break;
|
| 694 |
case 'N': /* number of desired processes */
|
| 695 |
case 'n': /* quietly supported alternate */
|
| 696 |
nprocs = atoi(argv[++a]);
|
| 697 |
if (nprocs <= 0)
|
| 698 |
goto userr;
|
| 699 |
break;
|
| 700 |
default:;
|
| 701 |
badopt: fprintf(stderr, "%s: bad option: %s\n", argv[0], argv[a]);
|
| 702 |
goto userr;
|
| 703 |
return(1);
|
| 704 |
}
|
| 705 |
if ((argc-a < 1) | (argc-a > 2) || argv[a][0] == '-')
|
| 706 |
goto userr;
|
| 707 |
in_spec = argv[a];
|
| 708 |
cmtx = rmx_load(argv[a+1]); /* loads from stdin if a+1==argc */
|
| 709 |
if (cmtx == NULL)
|
| 710 |
return(1); /* error reported */
|
| 711 |
cacheGB *= (cmtx->ncols > 1);
|
| 712 |
if (cacheGB > 0 && (!out_spec || *out_spec == '!')) {
|
| 713 |
fprintf(stderr, "%s: -m option incompatible with output to %s\n",
|
| 714 |
argv[0], out_spec ? "command" : "stdout");
|
| 715 |
return(1);
|
| 716 |
}
|
| 717 |
if (nprocs > cmtx->ncols)
|
| 718 |
nprocs = cmtx->ncols;
|
| 719 |
#if defined(_WIN32) || defined(_WIN64)
|
| 720 |
if (nprocs > 1) {
|
| 721 |
fprintf(stderr, "%s: warning - Windows only allows -N 1\n", argv[0]);
|
| 722 |
nprocs = 1;
|
| 723 |
}
|
| 724 |
#else
|
| 725 |
if ((nprocs > 1) & !out_spec) {
|
| 726 |
fprintf(stderr, "%s: multi-processing result cannot go to stdout\n",
|
| 727 |
argv[0]);
|
| 728 |
nprocs = 1;
|
| 729 |
}
|
| 730 |
if ((nprocs > 1) & iswapped && (in_type==DTfloat) | (in_type==DTdouble)) {
|
| 731 |
fprintf(stderr, "%s: multi-processing unsupported on swapped input\n",
|
| 732 |
argv[0]);
|
| 733 |
nprocs = 1;
|
| 734 |
}
|
| 735 |
#endif
|
| 736 |
if (cmtx->nrows > 1 && !hasFormat(in_spec)) {
|
| 737 |
fprintf(stderr, "%s: input specification '%s' needs %%d format\n",
|
| 738 |
argv[0], in_spec);
|
| 739 |
goto userr;
|
| 740 |
}
|
| 741 |
if (!get_iotypes())
|
| 742 |
return(1);
|
| 743 |
if (cacheGB > 1e-4) { /* figure out # of passes => rintvl */
|
| 744 |
size_t inp_bytes = (in_type==DTfloat ? sizeof(float)*ncomp
|
| 745 |
: (size_t)(ncomp+1)) * xres*yres;
|
| 746 |
size_t mem_bytes = sizeof(float)*ncomp*xres*yres;
|
| 747 |
int npasses = (double)inp_bytes*cmtx->nrows /
|
| 748 |
(cacheGB*(1L<<30) - (double)mem_bytes*nprocs) + 1;
|
| 749 |
if ((npasses <= 0) | (npasses*8 >= cmtx->nrows))
|
| 750 |
npasses = 1; /* let's not go there... */
|
| 751 |
rintvl = cmtx->nrows / npasses;
|
| 752 |
rintvl += (rintvl*npasses < cmtx->nrows);
|
| 753 |
} else
|
| 754 |
rintvl = cmtx->nrows;
|
| 755 |
/* make our passes */
|
| 756 |
for (row0 = 0; row0 < cmtx->nrows; row0 += rintvl) {
|
| 757 |
if ((rowN = row0 + rintvl) > cmtx->nrows)
|
| 758 |
rowN = cmtx->nrows;
|
| 759 |
if (nprocs==1 ? !solo_process() : !multi_process())
|
| 760 |
return(1);
|
| 761 |
}
|
| 762 |
return(0);
|
| 763 |
userr:
|
| 764 |
fprintf(stderr, "Usage: %s [-oc | -of][-o ospec][-N nproc][-m cacheGB] inpspec [mtx]\n",
|
| 765 |
argv[0]);
|
| 766 |
return(1);
|
| 767 |
}
|
