From 22f703cab05b7cd368f4de9e03991b7664dc5022 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?J=C3=B6rg=20Frings-F=C3=BCrst?= Date: Mon, 1 Sep 2014 13:56:46 +0200 Subject: Initial import of argyll version 1.5.1-8 --- rspl/rev.c | 6608 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 6608 insertions(+) create mode 100644 rspl/rev.c (limited to 'rspl/rev.c') diff --git a/rspl/rev.c b/rspl/rev.c new file mode 100644 index 0000000..237dbe3 --- /dev/null +++ b/rspl/rev.c @@ -0,0 +1,6608 @@ + +/* + * Argyll Color Correction System + * Multi-dimensional regularized spline data structure + * + * Reverse interpolation support code. + * + * Author: Graeme W. Gill + * Date: 30/1/00 + * + * Copyright 1999 - 2008 Graeme W. Gill + * All rights reserved. + * + * This material is licenced under the GNU AFFERO GENERAL PUBLIC LICENSE Version 3 :- + * see the License.txt file for licencing details. + * + * Latest simplex/linear equation version. + */ + +/* TTBD: + + Should fix the clipping case so that a direction weighting + funtion can be applied. This should be used just like + the perceptual case to increase L* constance for dark + colors. Will this stuff up the geometric consistency though ? + [ See fill_nncell(), fill_nncell() and users of calc_fwd_nn_cell_list(), + ie. nnearest_clip_solve(), clipn_setsort() etc. ] + The SVD least squares computation case makes this hard to change ? + Would have to feed in a weighting function, or can it be general ? + + Allow function callback to set auxiliary values for + flag RSPL_AUXLOCUS. + How to pass enough info back to aux_compute() ? + + Should auxil return multiple solutions if it finds them ??? + + */ + +/* TTBD: + Get rid of error() calls - return status instead + + Need to add a hefty overview and explanation of + how all this works, before I forget it ! + + ie: + + Basic function requirements: exact, auxil, locus, clip + + Fwd cell - reverse cell list lookup + + Basic layout di -> fdi + auxils + ink limit + + Basic search strategy + + Sub Simplex decomposition & properties + + How each type of function finds solutions + Sub-simplex dimensionality & dof + target dim & dof + Linear algebra choices. + + How final solutions are chosen + + */ + +/* PROBLEMS: + + Sometimes the aux locus doesn't correspond exactly to + the inversion :- ie. one locus segment is returned, + yet the inversion can't return a solution with + a particular aux target that lies within that segment. + (1150 near black, k ~= 0.4). + + + */ + +#include +#include +#include +#include +#include +#include + +#ifdef NT +# ifdef WINVER +# undef WINVER +# endif +# define WINVER 0x0500 /* We need 2k features */ +# include +#else +# include +# ifdef __APPLE__ +# include +# include +# include +# endif +#endif + +#define INKSCALE 5000.0 /* For ink limit weighting to fudge SVD least squares solution */ + +#include "rspl_imp.h" +#include "numlib.h" +#include "sort.h" /* Heap sort */ +#include "counters.h" /* Counter macros */ + +//#define DMALLOC_GLOBALS +//#include "dmalloc.h" +//#undef DMALLOC_GLOBALS + +#undef DEBUG1 /* Higher level code */ +#undef DEBUG2 /* Lower level code */ + +/* Debug memory usage accounting */ +#ifdef NEVER +#ifdef NEVER +int thissz, lastsz = -1; +#define INCSZ(s, bbb) { \ + (s)->rev.sz += (bbb); \ + (s)->rev.thissz = (s)->rev.sz/1000000; \ + if ((s)->rev.thissz != (s)->rev.lastsz) fprintf(stderr,"~1 0x%x: %s, %d: rev size = %d Mbytes, delta %d, limit %d\n",((int)(s) >> 8) & 0xf, __FILE__, __LINE__,(s)->rev.thissz,(bbb),(s)->rev.max_sz/1000000); \ + (s)->rev.lastsz = (s)->rev.thissz; \ + } +#define DECSZ(s, bbb) { \ + (s)->rev.sz -= (bbb); \ + (s)->rev.thissz = (s)->rev.sz/1000000; \ + if ((s)->rev.thissz != (s)->rev.lastsz) fprintf(stderr,"~1 0x%x: %s, %d: rev size = %d Mbytes, delta %d, limit %d\n",((int)(s) >> 8) & 0xf, __FILE__, __LINE__,(s)->rev.thissz,-(bbb),(s)->rev.max_sz/1000000); \ + (s)->rev.lastsz = (s)->rev.thissz; \ + } +#else +#define INCSZ(s, bbb) (s)->rev.sz += (bbb); \ + fprintf(stderr,"%s, %d: rev.sz += %d\n",__FILE__, __LINE__, bbb) +#define DECSZ(s, bbb) (s)->rev.sz -= (bbb); \ + fprintf(stderr,"%s, %d: rev.sz -= %d\n",__FILE__, __LINE__, bbb) +#endif +#else +#define INCSZ(s, bbb) (s)->rev.sz += (bbb) +#define DECSZ(s, bbb) (s)->rev.sz -= (bbb) +#endif + +/* Set STATS in rev.h */ + +#define DOSORT /* Cell sort */ + +/* Print a vectors value */ +#define DBGVI(text, dim, out, vec, end) \ +{ int pveci; \ + printf("%s",text); \ + for (pveci = 0 ; pveci < (dim); pveci++) \ + printf(out,(vec)[pveci]); \ + printf(end); \ +} + +/* Print a matrix value */ +#define DBGMI(text, rows, cols, out, mat, end) \ +{ int pveci, pvecr; \ + printf("%s",text); \ + for (pvecr = 0 ; pvecr < (rows); pvecr++) { \ + for (pveci = 0 ; pveci < (cols); pveci++) \ + printf(out,(mat)[pvecr][pveci]); \ + if ((pvecr+1) < (rows)) \ + printf("\n"); \ + } \ + printf(end); \ +} + +/* Do an arbitrary printf */ +#define DBGI(text) printf text ; + +#undef DEBUG +#undef DBG +#undef DBGV +#undef DBGM + +#undef NEVER +#define ALWAYS + +#ifdef DEBUG1 +#undef DBGS +#undef DBG +#undef DBGV +#undef DBGM +#define DEBUG +#define DBGS(xxx) xxx +#define DBG(xxx) DBGI(xxx) +#define DBGV(xxx) DBGVI xxx +#define DBGM(xxx) DBGMI xxx +#else +#undef DEBUG +#undef DBGS +#undef DBG +#undef DBGV +#undef DBGM +#define DBGS(xxx) +#define DBG(xxx) +#define DBGV(xxx) +#define DBGM(xxx) +#endif + +/* Debug string routines */ +static char *pcellorange(cell *c); + +/* Convention is to use: + i to index grid points u.a + n to index data points d.a + e to index position dimension di + f to index output function dimension fdi + j misc and cube corners + k misc + */ + +#define EPS (2e-6) /* 2e-6 Allowance for numeric error */ + +static void make_rev(rspl *s); +static void init_revaccell(rspl *s); + +static cell *get_rcell(schbase *b, int ix, int force); +static void uncache_rcell(revcache *r, cell *cp); +#define unget_rcell(r, cp) uncache_rcell(r, cp) /* These are the same */ +static void invalidate_revaccell(rspl *s); +static int decrease_revcache(revcache *rc); + +/* ====================================================== */ + +static schbase *init_search(rspl *s, int flags, double *av, int *auxm, + double *v, double *cdir, co *cpp, int mxsoln, enum ops op); +static void adjust_search(rspl *s, int flags, double *av, enum ops op); +static schbase *set_search_limit(rspl *s, double (*limit)(void *vcntx, double *in), + void *lcntx, double limitv); +static void set_lsearch(rspl *s, int e); +static void free_search(schbase *b); + +static int *calc_fwd_cell_list(rspl *s, double *v); + +static int *calc_fwd_nn_cell_list(rspl *s, double *v); + +static void init_line_eq(schbase *b, double st[MXRO], double de[MXRO]); +static int *init_line(rspl *s, line *l, double st[MXRO], double de[MXRO]); +static int *next_line_cell(line *l); + +static void search_list(schbase *b, int *rip, unsigned int tcount); + +static void clear_limitv(rspl *s); + +static double get_limitv(schbase *b, int ix, float *fcb, double *p); + +#ifdef STATS +static char *opnames[6] = { "exact", "clipv", "clipn", "auxil", "locus" }; +#endif /* STATS */ + +#define INF_DIST 1e38 /* Stands for infinite "current best" distance */ + +/* ====================================================== */ +/* Globals that track overall usage of reverse cache to aportion memory */ +/* This is incremented for rspl with di > 1 when rev.rev_valid != 0 */ +size_t g_avail_ram = 0; /* Total maximum memory to be used */ +size_t g_test_ram = 0; /* Amount of memory that has been tested to be allocatable */ +int g_no_rev_cache_instances = 0; +rev_struct *g_rev_instances = NULL; + +/* ------------------------------------------------------ */ +/* Retry allocation routines - if the malloc fails, */ +/* try reducing the cache size and trying again */ +/* (This won't catch the problem if it occurs in a malloc outside rev) */ + +/* When a malloc fails, reduce the maximum cache to */ +/* it's current allocation minus the given size. */ +static void rev_reduce_cache(size_t size) { + rev_struct *rsi; + size_t ram; + + /* Compute how much ram is currently allocated */ + for (ram = 0, rsi = g_rev_instances; rsi != NULL; rsi = rsi->next) + ram += rsi->sz; + + if (size > ram) + error("rev_reduce_cache: run out of rev virtual memory!"); + +//printf("~1 size = %d, g_test_ram = %d\n",size,g_test_ram); +//printf("~1 rev: Reducing cache because alloc of %d bytes failed. Reduced from %d to %d MB\n", +//size, g_avail_ram/1000000, (ram - size)/1000000); + ram = g_avail_ram = ram - size; + + /* Aportion the memory, and reduce the cache allocation to match */ + ram /= g_no_rev_cache_instances; + for (rsi = g_rev_instances; rsi != NULL; rsi = rsi->next) { + revcache *rc = rsi->cache; + + rsi->max_sz = ram; + while (rc->nunlocked > 0 && rsi->sz > rsi->max_sz) { + if (decrease_revcache(rc) == 0) + break; + } +//printf("~1 rev instance ram = %d MB\n",rsi->sz/1000000); + } +//fprintf(stdout, "%c~~1 There %s %d rev cache instance%s with %d Mbytes limit\n", +// cr_char, +// g_no_rev_cache_instances > 1 ? "are" : "is", +// g_no_rev_cache_instances, +// g_no_rev_cache_instances > 1 ? "s" : "", +// ram/1000000); +} + +/* Check that the requested allocation plus 20 M Bytes */ +/* can be allocated, and if not, reduce the rev-cache limit. */ +/* This is so as to detect running out of VM before */ +/* we actually run out and (on OS X) avoid emitting a warning. */ +static void rev_test_vram(size_t size) { + char *a1; +#ifdef __APPLE__ + int old_stderr, new_stderr; + + /* OS X malloc() blabs about a malloc failure. This */ + /* will confuse users, so we temporarily redirect stdout */ + fflush(stderr); + old_stderr = dup(fileno(stderr)); + new_stderr = open("/dev/null", O_WRONLY | O_APPEND); + dup2(new_stderr, fileno(stderr)); +#endif + size += 20 * 1024 * 1024; /* This depends on the VM region allocation size */ + if ((a1 = malloc(size)) == NULL) { + rev_reduce_cache(size); + } else { + free(a1); + } + g_test_ram = size/2; /* Allow for twice as much VM to be used for each allocation */ +#ifdef __APPLE__ + fflush(stderr); + dup2(old_stderr, fileno(stderr)); /* Restore stderr */ + close(new_stderr); + close(old_stderr); +#endif +} + +static void *rev_malloc(rspl *s, size_t size) { + void *rv; + + if ((size + 1 * 1024 * 1204) > g_test_ram) + rev_test_vram(size); + if ((rv = malloc(size)) == NULL) { + rev_reduce_cache(size); + rv = malloc(size); + } + if (rv != NULL) + g_test_ram -= size; + + return rv; +} + +static void *rev_calloc(rspl *s, size_t num, size_t size) { + void *rv; + + if (((num * size) + 1 * 1024 * 1204) > g_test_ram) + rev_test_vram(size); + if ((rv = calloc(num, size)) == NULL) { + rev_reduce_cache(num * size); + rv = calloc(num, size); + } + if (rv != NULL) + g_test_ram -= size; + + return rv; +} + +static void *rev_realloc(rspl *s, void *ptr, size_t size) { + void *rv; + + if ((size + 1 * 1024 * 1204) > g_test_ram) + rev_test_vram(size); + if ((rv = realloc(ptr, size)) == NULL) { + rev_reduce_cache(size); /* approximation */ + rv = realloc(ptr, size); + } + if (rv != NULL) + g_test_ram -= size; + + return rv; +} + + +/* ====================================================== */ +/* Set the ink limit information for any reverse interpolation. */ +/* Calling this will clear the reverse interpolaton cache and acceleration structures. */ +static void +rev_set_limit_rspl( + rspl *s, /* this */ + double (*limit)(void *lcntx, double *in), /* Optional input space limit function. Function */ + /* should evaluate in[0..di-1], and return number that is not to exceed */ + /* limitv. NULL if not used */ + void *lcntx, /* Context passed to limit() */ + double limitv /* Value that limit() is not to exceed */ +) { + schbase *b; + + DBG(("rev: setting ink limit function 0x%x and limit %f\n",limit,limitv)); + /* This is a restricted size function */ + if (s->di > MXRI) + error("rspl: rev_set_limit can't handle di = %d",s->di); + if (s->fdi > MXRO) + error("rspl: rev_set_limit can't handle fdi = %d",s->fdi); + + b = set_search_limit(s, limit, lcntx, limitv); /* Init and set limit info */ + + if (s->rev.inited) { /* If cache and acceleration has been allocated */ + invalidate_revaccell(s); /* Invalidate the reverse cache */ + } + + /* Invalidate any ink limit values cached with the fwd grid data */ + clear_limitv(s); +} + +/* Get the ink limit information for any reverse interpolation. */ +static void +rev_get_limit_rspl( + rspl *s, /* this */ + double (**limitf)(void *lcntx, double *in), /* Return pointer to function of NULL if not set */ + void **lcntx, /* return context pointer */ + double *limitv /* Return limit value */ +) { + schbase *b = s->rev.sb; + + /* This is a restricted size function */ + if (s->di > MXRI) + error("rspl: rev_get_limit can't handle di = %d",s->di); + if (s->fdi > MXRO) + error("rspl: rev_get_limit can't handle fdi = %d",s->fdi); + + if (b == NULL) { + *limitf = NULL; + *lcntx = NULL; + *limitv = 0.0; + } else { + *limitf = s->limitf; + *lcntx = s->lcntx; + *limitv = s->limitv/INKSCALE; + } +} + +#define RSPL_CERTAIN 0x80000000 /* WILLCLIP hint is certain */ +#define RSPL_WILLCLIP2 (RSPL_CERTAIN | RSPL_WILLCLIP) /* Clipping will certainly be needed */ + +/* Do reverse interpolation given target output values and (optional) auxiliary target */ +/* input values. Return number of results and clipping flag. If return value == mxsoln, */ +/* then there might be more results. The target values returned will correspond to the */ +/* actual (posssibly clipped) point. The return value is the number of solutions + */ +/* a clipped flag. Properly set hint flags improve performance, but a correct result should */ +/* be returned if the RSPL_NEARCLIP is set, even if they are not set correctly. */ +static int +rev_interp_rspl( + rspl *s, /* this */ + int flags, /* Hint flag */ + int mxsoln, /* Maximum number of solutions allowed for */ + int *auxm, /* Array of di mask flags, !=0 for valid auxliaries (NULL if no auxiliaries) */ + double cdir[MXRO], /* Clip vector direction wrt to cpp[0].v and length - NULL if not used */ + co *cpp /* Given target output space value in cpp[0].v[] + */ + /* target input space auxiliaries in cpp[0].p[], return */ + /* input space solutions in cpp[0..retval-1].p[], and */ +) { + int e, di = s->di; + int fdi = s->fdi; + int i, *rip = NULL; + schbase *b = NULL; /* Base search information */ + double auxv[MXRI]; /* Locus proportional auxiliary values */ + int didclip = 0; /* flag - set if we clipped the target */ + + DBGV(("\nrev interp called with out targets", fdi, " %f", cpp[0].v, "\n")); + + /* This is a restricted size function */ + if (di > MXRI) + error("rspl: rev_interp can't handle di = %d",di); + if (fdi > MXRO) + error("rspl: rev_interp can't handle fdi = %d",fdi); + + if (auxm != NULL) { + double ax[MXRI]; + for (i = 0; i < di; i++) { + if (auxm[i] != 0) + ax[i] = cpp[0].p[i]; + else + ax[i] = 0.0; + } + DBGV((" auxiliaries mask", di, " %d", auxm, "\n")); + DBGV((" auxiliaries values", di, " %f", ax, "\n")); + } + DBG(("di = %d, fdi = %d\n",di, fdi)); + DBG(("flags = 0x%x\n",flags)); + + mxsoln &= RSPL_NOSOLNS; /* Prevent silliness */ + + /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ + /* Auxiliary is proportion of locus, so we need to find locus extent */ + if (flags & RSPL_AUXLOCUS) { + DBG(("rev interp has aux targets as proportion of locus\n")); + + flags &= ~RSPL_WILLCLIP; /* Reset hint flag, as we will figure it out */ + + /* For each valid auxiliary */ + for (e = 0; e < di; e++) { + if (auxm[e] == 0) + continue; /* Skip unsused auxiliaries */ + + /* Do search for min and max */ + DBG(("rev locus searching for aux %d min/max\n", e)); + if (b == NULL) { + b = init_search(s, flags, cpp[0].p, auxm, cpp[0].v, cdir, cpp, mxsoln, locus); +#ifdef STATS + s->rev.st[b->op].searchcalls++; +#endif /* STATS */ + } else + set_lsearch(s, e); /* Reset locus search for next auxiliary */ + + if (rip == NULL) { /* Not done this yet */ + rip = calc_fwd_cell_list(s, cpp[0].v); /* Reverse grid index for out target */ + if (rip == NULL) { + DBG(("Got NULL list (point outside range) for auxiliary locus search\n")); + flags |= RSPL_WILLCLIP2; + break; + } + } + + search_list(b, rip, s->get_next_touch(s)); /* Setup, sort and search the list */ + + if (b->min > b->max) { /* Failed to find locus */ + DBG(("rev interp failed to find locus for aux %d, so expect clip\n",e)); + flags |= RSPL_WILLCLIP2; + break; + } + auxv[e] = (cpp[0].p[e] * (b->max - b->min)) + b->min; + } + + DBG(("rev interp got all locuses, so expect exact result\n",e)); + if (!(flags & RSPL_WILLCLIP)) { + flags |= RSPL_EXACTAUX; /* Got locuses, so expect exact result */ + } + } + + /* Init the search information */ + if (b == NULL) + b = init_search(s, flags, cpp[0].p, auxm, cpp[0].v, cdir, cpp, mxsoln, exact); + else + adjust_search(s, flags, auxv, exact); /* Using proportion of locus aux */ + + /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ + /* If hinted that we will not need to clip, look for exact solution. */ + if (!(flags & RSPL_WILLCLIP)) { + DBG(("Hint we won't clip, so trying exact search\n")); + + /* First do an exact search (init will select auxil if requested) */ + adjust_search(s, flags, NULL, exact); + + /* Figure out the reverse grid index appropriate for this request */ + if (rip == NULL) /* Not done this yet */ + rip = calc_fwd_cell_list(s, cpp[0].v); + +#ifdef STATS + s->rev.st[b->op].searchcalls++; +#endif /* STATS */ + if (rip != NULL) { + /* Setup, sort and search the list */ + search_list(b, rip, s->get_next_touch(s)); + } else { + DBG(("Got NULL list (point outside range) for first exact reverse cell\n")); + } + + /* If we selected exact aux, but failed to find a solution, relax expectation */ + if (b->nsoln == 0 && b->naux > 0 && (flags & RSPL_EXACTAUX)) { +//printf("~1 relaxing notclip expactation when nsoln == %d, naux = %d, falgs & RSPL_EXACTAUX = 0x%x\n", b->nsoln,b->naux,flags & RSPL_EXACTAUX); + DBG(("Searching for exact match to auxiliary target failed, so try again\n")); + adjust_search(s, flags & ~RSPL_EXACTAUX, NULL, exact); + +#ifdef STATS + s->rev.st[b->op].searchcalls++; +#endif /* STATS */ + /* Candidate cell list should be the same */ + if (rip != NULL) { + /* Setup, sort and search the list */ + search_list(b, rip, s->get_next_touch(s)); + } else { + DBG(("Got NULL list (point outside range) for nearest search reverse cell\n")); + } + } + } + + /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ + /* If the exact search failed, and we should look for a nearest solution */ + if (b->nsoln == 0 && (flags & RSPL_NEARCLIP)) { + DBG(("Trying nearest search\n")); + +#ifdef STATS + s->rev.st[b->op].searchcalls++; +#endif /* STATS */ + + /* We get returned a list of cube base indexes of all cubes that have */ + /* the closest valid vertex value to the target value. */ + /* (This may not result in the true closest point if the geometry of */ + /* the vertex values is localy non-smooth or self intersecting, */ + /* but seems to return a good result in most realistic situations ?) */ + + adjust_search(s, flags, NULL, clipn); + + /* Get list of cells enclosing nearest vertex */ + if ((rip = calc_fwd_nn_cell_list(s, cpp[0].v)) != NULL) { + search_list(b, rip, s->get_next_touch(s)); /* Setup, sort and search the list */ + } else { + DBG(("Got NULL list! (point inside gamut \?\?) for nearest search\n")); + } + + if (b->nsoln > 0) + didclip = RSPL_DIDCLIP; + } + + /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ + /* If we still don't have a solution, do a vector direction clip */ + if (b->nsoln == 0 && b->canvecclip) { + /* Find clipping solution in vector direction */ + line ln; /* Structure to hold line context */ + unsigned int tcount; /* grid touch count for this opperation */ + + DBG(("Starting a clipping vector search now!!\n")); + + adjust_search(s, flags, NULL, clipv); + + tcount = s->get_next_touch(s); /* Get next grid touched generation count */ + +#ifdef STATS + s->rev.st[b->op].searchcalls++; +#endif /* STATS */ + init_line_eq(b, b->v, cdir); /* Init the implicit line equation */ + rip = init_line(s, &ln, cpp[0].v, cdir); /* Init the line cell dda */ +//~~1 HACK!!! should be <= 1.0 !!! + for (; ln.t <= 2.0; rip = next_line_cell(&ln)) { + if (rip == NULL) { + DBG(("Got NULL list for this reverse cell\n")); + continue; + } + + /* Setup, sort and search the list */ + search_list(b, rip, tcount); + + /* If we have found a solution, then abort the search - */ + /* this line will be taking us away from the best solution. */ + if (b->nsoln > 0) + break; + } + if (b->nsoln > 0) + didclip = RSPL_DIDCLIP; + } + + /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ + /* If the clipped solution seems to have been jumping to conclusions, */ + /* search for an exact solution. */ + if (didclip && (flags & RSPL_WILLCLIP && !(flags & RSPL_CERTAIN)) + && (b->cdist/s->get_out_scale(s)) < 0.002) { + co c_cpp = b->cpp[0]; /* Save clip solution in case we want it */ + double c_idist = b->idist; + int c_iabove = b->iabove; + int c_nsoln = b->nsoln; + int c_pauxcell = b->pauxcell; + double c_cdist = b->cdist; + int c_iclip = b->iclip; + + DBG(("Trying exact search again\n")); + + /* Do an exact search (init will select auxil if requested) */ + adjust_search(s, flags & ~RSPL_WILLCLIP, NULL, exact); + + /* Figure out the reverse grid index appropriate for this request */ + rip = calc_fwd_cell_list(s, cpp[0].v); + +#ifdef STATS + s->rev.st[b->op].searchcalls++; +#endif /* STATS */ + if (rip != NULL) { + /* Setup, sort and search the list */ + search_list(b, rip, s->get_next_touch(s)); + } else { + DBG(("Got NULL list (point outside range) for first exact reverse cell\n")); + } + + /* If we selected exact aux, but failed to find a solution, relax expectation */ + if (b->nsoln == 0 && b->naux > 0 && (flags & RSPL_EXACTAUX)) { + DBG(("Searching for exact match to auxiliary target failed, so try again\n")); +//printf("~1 relaxing didclip expactation when nsoln == %d, naux = %d, flags & RSPL_EXACTAUX = 0x%x\n", b->nsoln,b->naux,flags & RSPL_EXACTAUX); + adjust_search(s, flags & ~RSPL_EXACTAUX, NULL, exact); + +#ifdef STATS + s->rev.st[b->op].searchcalls++; +#endif /* STATS */ + /* Candidate cell list should be the same */ + if (rip != NULL) { + /* Setup, sort and search the list */ + search_list(b, rip, s->get_next_touch(s)); + } else { + DBG(("Got NULL list (point outside range) for nearest search reverse cell\n")); + } + } + + /* If we did get an exact solution */ + if (b->nsoln > 0) { + DBG(("Deciding to return exact solution after finding clipped\n")); + didclip = 0; /* Reset did-clip and return exact solution */ + + } else { + DBG(("keeping clipped solution\n")); + /* Restore the clipped solution */ + b->cpp[0] = c_cpp; + b->idist = c_idist; + b->iabove = c_iabove; + b->nsoln = c_nsoln; + b->pauxcell = c_pauxcell; + b->cdist = c_cdist; + b->iclip = c_iclip; + } + } + + if (b->nsoln > 0) { + DBGV(("rev interp returning 1st soln: ",di," %f", cpp[0].p, "\n")); + } + DBG(("rev interp returning %d solutions%s\n",b->nsoln, didclip ? " [clip]" : "")); + + return b->nsoln | didclip; +} + +/* ------------------------------------------------------------------------------------ */ +/* Do reverse search for the auxiliary min/max ranges of the solution locus for the */ +/* given target output values. */ +/* Return number of locus segments found, up to mxsoln. 0 will be returned if no solutions */ +/* are found. */ + +static int +rev_locus_segs_rspl ( + rspl *s, /* this */ + int *auxm, /* Array of di mask flags, !=0 for valid auxliaries (NULL if no auxiliaries) */ + co *cpp, /* Input value in cpp[0].v[] */ + int mxsoln, /* Maximum number of solutions allowed for */ + double min[][MXRI], /* Array of min[MXRI] to hold return segment minimum values. */ + double max[][MXRI] /* Array of max[MXRI] to hold return segment maximum values. */ +) { + int e, di = s->di; + int f, fdi = s->fdi; + int six; /* solution index */ + int *rip = NULL; + int rv = 1; /* Return value */ + schbase *b = NULL; /* Base search information */ + + DBGV(("rev locus called with out targets", fdi, " %f", cpp[0].v, "\n")); + + /* This is a restricted size function */ + if (di > MXRI) + error("rspl: rev_locus_segs can't handle di = %d",di); + if (fdi > MXRO) + error("rspl: rev_locus_segs can't handle fdi = %d",fdi); + + if (mxsoln < 1) { + return 0; /* Guard against silliness */ + } + + if (auxm != NULL) { + int i; + double ax[MXRI]; + for (i = 0; i < di; i++) { + if (auxm[i] != 0) + ax[i] = cpp[0].p[i]; + else + ax[i] = 0.0; + } + DBGV((" auxiliaries mask", di, " %d", auxm, "\n")); + DBGV((" auxiliaries values", di, " %f", ax, "\n")); + } + + /* Init default return values */ + for (six = 0; six < mxsoln; six++) { + for (e = 0; e < di; e++) { + if (auxm[e] == 0) { + min[six][e] = max[six][e] = 0; /* Return 0 for unused auxiliaries */ + } else { + min[six][e] = 1.0; /* max < min indicates invalid range */ + max[six][e] = 0.0; + } + } + } + + /* For each valid auxiliary */ + for (e = 0; e < di; e++) { + if (auxm[e] == 0) + continue; /* Skip unsused auxiliaries */ + + /* Do search for min and max */ + DBG(("rev locus searching for aux %d min/max\n", e)); + if (b == NULL) + b = init_search(s, 0, cpp[0].p, auxm, cpp[0].v, NULL, cpp, mxsoln, locus); + else + set_lsearch(s, e); /* Reset locus search for next auxiliary */ + + if (rip == NULL) { /* Not done this yet */ + rip = calc_fwd_cell_list(s, cpp[0].v); /* Reverse grid index for this request */ + if (rip == NULL) { + DBG(("Got NULL list (point outside range) for auxiliary locus search\n")); + rv = 0; + break; + } + } + + search_list(b, rip, s->get_next_touch(s)); /* Setup, sort and search the list */ + + if (b->min > b->max) { + rv = 0; /* Failed to find a result */ + break; + } + + if (b->asegs == 0) { /* Overall min max only */ + + min[0][e] = b->min; /* Save single result */ + max[0][e] = b->max; + + } else { /* Tracking auxiliary segments */ + int si; /* Start i */ + int i, j, ff; + + /* Sort the segment list */ +#define HEAP_COMPARE(A,B) (A.xval < B.xval) + HEAPSORT(axisec, b->axisl, b->axisln) +#undef HEAP_COMPARE + +#ifdef NEVER +for (i = 0; i < b->axisln; i++) { +printf("~2 xval = %f, verts = ",b->axisl[i].xval); +for (f = 0; f < b->axisl[i].nv; f++) +printf(" %d", b->axisl[i].vix[f]); +printf("\n"); +} +#endif + /* Find the segments by finding common verticies */ + six = si = i = 0; + + min[six][e] = b->axisl[i].xval; + + for (i++; i < (b->axisln-1); i++) { + /* Check if any i and i-1 to j are connected */ + for (j = i-1; j >= si; j--) { + for (f = 0; f < b->axisl[j].nv; f++) { + for (ff = 0; ff < b->axisl[i].nv; ff++) { + if (b->axisl[j].vix[f] == b->axisl[i].vix[ff]) + break; /* Found a link */ + } + if (ff < b->axisl[i].nv) + break; + } + if (f < b->axisl[j].nv) + break; + } + if (j < si) { /* Wasn't linked */ + int ii, jj; + /* Think we found a break. Check that all the rest of */ + /* the entries don't have any links to the previous group */ + + /* This could be rather a slow way of checking ! (On^2) */ + for (ii = i+1; ii < (b->axisln); ii++) { + for (jj = i-1; jj >= si; jj--) { + for (f = 0; f < b->axisl[jj].nv; f++) { + for (ff = 0; ff < b->axisl[ii].nv; ff++) { + if (b->axisl[jj].vix[f] == b->axisl[ii].vix[ff]) + break; /* Found a link */ + } + if (ff < b->axisl[ii].nv) + break; + } + if (f < b->axisl[jj].nv) + break; + } + if (jj >= si) + break; + } + if (ii >= b->axisln) { /* Wasn't forward linked */ + /* Nothing ahead links to last group */ + max[six][e] = b->axisl[i-1].xval; + + /* If we run out of solution space */ + /* merge the last segments */ + if ((six+1) < mxsoln) { + six++; + min[six][e] = b->axisl[i].xval; + } + } + } + } + max[six++][e] = b->axisl[i].xval; + + if (six > rv) + rv = six; + } + } + +#ifdef STATS + s->rev.st[b->op].searchcalls++; +#endif /* STATS */ + if (rv) { + for (six = 0; six < rv; six++) { + DBG(("rev locus returning:\n")); + DBGV((" min", di, " %f", min[six], "\n")); + DBGV((" max", di, " %f", max[six], "\n")); + } + } + + DBG(("rev locus returning status %d\n",rv)); + return rv; +} + +/* ------------------------------------------------------------------------------------ */ +typedef double mxdi_ary[MXRI]; + +/* Do reverse search for the locus of the auxiliary input values given a target output. */ +/* Return 1 on finding a valid solution, and 0 if no solutions are found. */ +static int +rev_locus_rspl( + rspl *s, /* this */ + int *auxm, /* Array of di mask flags, !=0 for valid auxliaries (NULL if no auxiliaries) */ + co *cpp, /* Input value in cpp[0].v[] */ + double min[MXRI],/* Return minimum auxiliary values */ + double max[MXRI] /* Return maximum auxiliary values */ +) { + + /* Use segment routine to compute oveall locus */ + return rev_locus_segs_rspl (s, auxm, cpp, 1, (mxdi_ary *)min, (mxdi_ary *)max); +} + +/* ------------------------------------------------------------------------------------ */ + +#ifdef DEBUG2 +#define DEBUG +#undef DBG +#undef DBGV +#undef DBGM +#define DBG(xxx) DBGI(xxx) +#define DBGV(xxx) DBGVI xxx +#define DBGM(xxx) DBGMI xxx +#else +#undef DEBUG +#undef DBG +#undef DBGV +#undef DBGM +#define DBG(xxx) +#define DBGV(xxx) +#define DBGM(xxx) +#endif + +/* ------------------------------------------------ */ +/* subroutines of top level reverse lookup routine */ + +static int exact_setsort(schbase *b, cell *c); +static int exact_compute(schbase *b, simplex *x); + +static int auxil_setsort(schbase *b, cell *c); +static int auxil_check(schbase *b, cell *c); +static int auxil_compute(schbase *b, simplex *x); + +static int locus_setsort(schbase *b, cell *c); +static int locus_check(schbase *b, cell *c); +static int locus_compute(schbase *b, simplex *x); + +static int clipv_setsort(schbase *b, cell *c); +static int clipv_check(schbase *b, cell *c); +static int clipv_compute(schbase *b, simplex *x); + +static int clipn_setsort(schbase *b, cell *c); +static int clipn_check(schbase *b, cell *c); +static int clipn_compute(schbase *b, simplex *x); + +/* Allocate the search base structure */ +static schbase * +alloc_sb(rspl *s) { + schbase *b; + if ((b = s->rev.sb = (schbase *)rev_calloc(s, 1, sizeof(schbase))) == NULL) + error("rspl malloc failed - rev.sb structure"); + INCSZ(s, sizeof(schbase)); + + b->s = s; /* rsp */ + b->pauxcell = /* Previous solution cell indexes */ + b->plmaxcell = + b->plmincell = -1; + + return b; +} + +/* Free the search base structure */ +static void +free_sb(schbase *b) { + DECSZ(b->s, sizeof(schbase)); + free(b); +} + +/* Do the basic search type independent initialization */ +static schbase * /* Return pointer to base search information */ +init_search( + rspl *s, /* rsp; */ + int flags, /* Hint flag */ + + double *av, /* Auxiliary input values - may be NULL */ + int *auxm, /* Array of di mask flags, !=0 for valid auxliaries (NULL if no auxiliaries) */ + /* Locus search will search for max/min of first valid auxlilary */ + double *v, /* Output value target, NULL if none */ + double *cdir, /* Clip vector direction, NULL if none */ + co *cpp, /* Array that hold solutions, NULL if none. */ + int mxsoln, /* Maximum number of solutions allowed for */ + enum ops op /* Type of reverse search operation requested */ +) { + schbase *b = NULL; /* Pointer to search base information structure */ + int e, di = s->di; + int f, fdi = s->fdi; + + DBG(("Initializing search\n")); + + if (s->rev.inited == 0) /* Compute reverse info if it doesn't exist */ + make_rev(s); + + /* If first time initialisation (Fourth section init) */ + if ((b = s->rev.sb) == NULL) { + b = alloc_sb(s); + } + + /* Init some basic search info */ + b->op = op; /* operation */ + b->flags = flags; /* hint flags */ + b->canvecclip = 0; /* Assume invalid clip direction */ + + b->ixc = (1<naux = 0; + b->auxbm = 0; + if (auxm != NULL) { + unsigned bm; + + if (mxsoln > 1) + b->asegs = 1; /* Find all segments */ + else + b->asegs = 0; /* Find only overall aux locus range */ + + for (e = di-1, bm = 1 << e; e >= 0; e--, bm >>= 1) { /* Record auxiliary mask bits */ + if (av != NULL) + b->av[e] = av[e]; /* Auxiliary target values */ + b->auxm[e] = auxm[e]; /* Auxiliary mask */ + if (auxm[e] != 0) { + b->auxbm |= bm; /* Auxiliary bit mask */ + b->auxi[b->naux++] = e; /* Index of next auxiliary input to be used */ + /* Auxiliary locus extent */ + b->lxi = e; /* Assume first one */ + b->max = -INF_DIST; /* In case searching for max */ + b->min = INF_DIST; /* In case searching for minimum */ + b->axisln = 0; /* No intersects in list */ + } + } + } + + /* Figure out if the clip direction is meaningfull */ + /* Check that the clip vector makes sense */ + if (cdir != NULL) { /* Clip vector is specified */ + double ss; + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt = cdir[f]; + b->cdir[f] = tt; + ss += tt * tt; + } + + if (ss > 1e-6) { + b->canvecclip = 1; /* It has a non-zero length */ + ss = sqrt(ss); + /* Compute normalised clip vector direction */ + for (f = 0; f < fdi; f++) { + b->ncdir[f] = b->cdir[f]/ss; + } + } + } + + if (di <= fdi) /* Only allow auxiliaries if di > fdi */ + b->naux = 0; + + /* Switch to appropriate operation */ + if (b->op == exact && (b->naux > 0 || di != fdi)) { + b->op = auxil; + } else if (b->op == auxil && b->naux == 0 && di == fdi) { + b->op = exact; + } + + /* Set appropriate functions for type of operation */ + switch (b->op) { + case exact: + b->setsort = exact_setsort; + b->check = NULL; + b->compute = exact_compute; + b->snsdi = b->ensdi = di; /* Search full dimension simplex, expect point soln. */ + break; + case auxil: + b->setsort = auxil_setsort; + b->check = auxil_check; + b->compute = auxil_compute; + b->snsdi = di; /* Start here DOF = di-fdi locus solutions */ + b->ensdi = fdi; /* End with DOF = 0 for point solutions */ + break; + case locus: + b->setsort = locus_setsort; + b->check = locus_check; + b->compute = locus_compute; + b->snsdi = b->ensdi = fdi; /* Search for point solutions */ + break; + case clipv: + b->setsort = clipv_setsort; + b->check = clipv_check; + b->compute = clipv_compute; + /* Clip vector 1 dimension in output space, */ + b->snsdi = b->ensdi = fdi-1; /* search planes for combined point solution */ + break; + case clipn: + b->setsort = clipn_setsort; + b->check = clipn_check; + b->compute = clipn_compute; + b->snsdi = 0; /* Start with DOF = 0 for point solutions */ + b->ensdi = fdi-1; /* End on DOF = di-fdi-1 on surfaces of simplexes */ + break; + default: + error("init_search: Unknown operation %d\n",b->op); + } + + if (v != NULL) { + for (f = 0; f < fdi; f++) /* Record target output values */ + b->v[f] = v[f]; + b->v[fdi] = s->limitv; /* Limitvalue is output target for limit clip subsimplexes */ + } + + b->mxsoln = mxsoln; /* Allow solutions to be returned */ + b->cpp = cpp; /* Put solutions here */ + b->nsoln = 0; /* No solutions at present */ + b->iclip = 0; /* Default solution isn't above ink limit */ + + if (flags & RSPL_EXACTAUX) /* Expect to be able to match auxiliary target exactly */ + b->idist = 2.0 * EPS; /* Best input distance to beat - helps sort/triage */ + else + b->idist = INF_DIST; /* Best input distance to beat. */ + b->iabove = 0; /* Best isn't known to be above (yet) */ + + b->cdist = INF_DIST; /* Best clip distance to beat. */ + + DBG(("Search initialized\n")); + + return b; +} + +/* Adjust the search */ +static void +adjust_search( + rspl *s, /* rsp; */ + int flags, /* Hint flag */ + double *av, /* Auxiliary input values - may be NULL */ + enum ops op /* Type of reverse search operation requested */ +) { + schbase *b = s->rev.sb; /* Pointer to search base information structure */ + int e, di = s->di; + int fdi = s->fdi; + + DBG(("Adjusting search\n")); + + b->op = op; /* operation */ + b->flags = flags; /* hint flags */ + + /* Switch from exact to aux if we need to */ + if (b->op == exact && (b->naux > 0 || di != fdi)) { + b->op = auxil; + } else if (b->op == auxil && b->naux == 0 && di == fdi) { + b->op = exact; + } + + /* Update auxiliary target values */ + if (av != NULL) { + for (e = 0; e < b->naux; e++) { + int ee = b->auxi[e]; + b->av[ee] = av[ee]; + } + } + + /* Set appropriate functions for type of operation */ + switch (b->op) { + case exact: + b->setsort = exact_setsort; + b->check = NULL; + b->compute = exact_compute; + b->snsdi = b->ensdi = di; /* Expect point solution */ + break; + case auxil: + b->setsort = auxil_setsort; + b->check = auxil_check; + b->compute = auxil_compute; + b->snsdi = di; /* Start here DOF = di-fdi locus solutions */ + b->ensdi = fdi; /* End with DOF = 0 for point solutions, */ + break; /* will early exit DOF if good soln found. */ + case locus: + b->setsort = locus_setsort; + b->check = locus_check; + b->compute = locus_compute; + b->snsdi = b->ensdi = fdi; /* Search for point solutions */ + break; + case clipv: + b->setsort = clipv_setsort; + b->check = clipv_check; + b->compute = clipv_compute; + /* Clip vector 1 dimension in output space, */ + b->snsdi = b->ensdi = fdi-1; /* so the intersection with the simplex is a point. */ + break; + case clipn: + b->setsort = clipn_setsort; + b->check = clipn_check; + b->compute = clipn_compute; + b->snsdi = 0; /* Start with DOF = 0 for point solutions */ + b->ensdi = fdi-1; /* End on DOF = di-fdi-1 on surfaces of simplexes */ + break; /* Will go through all DOF */ + default: + error("init_search: Unknown operation %d\n",b->op); + } + + b->nsoln = 0; /* No solutions at present */ + + if (flags & RSPL_EXACTAUX) /* Expect to be able to match auxiliary target exactly */ + b->idist = 2.0 * EPS; /* Best input distance to beat - helps sort/triage */ + else + b->idist = INF_DIST; /* Best input distance to beat. */ + b->iabove = 0; /* Best isn't known to be above (yet) */ + + b->cdist = INF_DIST; /* Best clip distance to beat. */ + + DBG(("Search adjusted\n")); +} + +/* Adjust existing locus search for a different auxiliary */ +static void +set_lsearch( +rspl *s, +int e /* Next auxiliary */ +) { + schbase *b = s->rev.sb; /* Pointer to search base information structure */ + + b->lxi = e; /* Assume first one */ + b->max = -INF_DIST; /* In case searching for max */ + b->min = INF_DIST; /* In case searching for minimum */ + b->axisln = 0; /* No intersects in list */ +} + +/* Set the limit search information */ +/* Note this doesn't create or init the main rev information. */ +static schbase * /* Return pointer to base search information */ +set_search_limit( + rspl *s, /* rsp; */ + double (*limitf)(void *vcntx, double *in), /* Optional input space limit function. Function */ + /* should evaluate in[0..di-1], and return number that is not to exceed */ + /* limitv. NULL if not used */ + void *lcntx, /* Context passed to limit() */ + double limitv /* Value that limit() is not to exceed */ +) { + schbase *b = NULL; /* Pointer to search base information structure */ + + /* If sb info needs initialising (Fourth section init) */ + if ((b = s->rev.sb) == NULL) { + b = alloc_sb(s); + } + + s->limitf = limitf; /* Input limit function */ + s->lcntx = lcntx; /* Context passed to limit() */ + s->limitv= INKSCALE * limitv; /* Context passed to values not to be exceedded by limit() */ + if (limitf != NULL) { + s->limiten = 1; /* enable limiting by default */ + } else + s->limiten = 0; /* No limit function, so limiting not enabled. */ + + return b; +} + +/* Free any search specific data, plus the search base. */ +static void +free_search( +schbase *b /* Base search information */ +) { + DBG(("Freeing search\n")); + + /* Clip line implicit equation (incuding space for ink target) */ + if (b->cla != NULL) { + int fdi = b->s->fdi; + free_dmatrix(b->cla, 0, fdi-1, 0, fdi); + b->cla = NULL; + } + + /* Auxiliary segment list */ + if (b->axislz > 0) { + free(b->axisl); + DECSZ(b->s, b->axislz * sizeof(axisec)); + b->axisl = NULL; + b->axislz = 0; + b->axisln = 0; + } + + /* Sorted cell list */ + if (b->lclistz > 0) { + free(b->lclist); + DECSZ(b->s, b->lclistz * sizeof(cell *)); + b->lclist = NULL; + b->lclistz = 0; + } + + /* Simplex filter list */ + if (b->lsxfilt > 0) { + free(b->sxfilt); + DECSZ(b->s, b->lsxfilt * sizeof(char)); + b->sxfilt = NULL; + b->lsxfilt = 0; + } + + free_sb(b); +} + +/* Return the pointer to the list of fwd cells given */ +/* the target output values. The pointer will be to the first */ +/* index in the list (ie. list address + 3) */ +/* Return NULL if none in list (out of gamut). */ +static int * +calc_fwd_cell_list( + rspl *s, /* this */ + double *v /* Output values */ +) { + int f, fdi = s->fdi; + int **rpp; + int rgres_1 = s->rev.res - 1; + + if (s->rev.rev_valid == 0) + init_revaccell(s); + + for (rpp = s->rev.rev, f = 0; f < fdi; f++) { + int mi; + double t = (v[f] - s->rev.gl[f])/s->rev.gw[f]; + mi = (int)floor(t); /* Grid coordinate */ + if (mi < 0 || mi > rgres_1) { /* If outside valid reverse range */ + return NULL; + } + rpp += mi * s->rev.coi[f]; /* Accumulate reverse grid pointer */ + } + if (*rpp == NULL) + return NULL; + return (*rpp) + 3; +} + +void alloc_simplexes(cell *c, int nsdi); + +/* Given a pointer to a list of fwd cells, cull cells that */ +/* cannot contain or improve the solution, sort the list, */ +/* and then compute the final best solution. */ +static void +search_list( +schbase *b, /* Base search information */ +int *rip, /* Pointer to first index in cell list */ +unsigned int tcount /* grid touch count for this operation */ +) { + rspl *s = b->s; + int nsdi; + int i; + int nilist; /* Number in cell list */ + unsigned int stouch; /* Simplex touch count */ + + DBG(("search_list called\n")); + + /* (rip[-3] contains allocation for fwd cells in the list) */ + /* (rip[-2] contains the index of the next free entry in the list) */ + /* (rip[-1] contains the reference count for the list) */ + if (b->lclistz < rip[-3]) { /* Allocate more space if needed */ + + if (b->lclistz > 0) { /* Free old space before allocating new */ + free(b->lclist); + DECSZ(b->s, b->lclistz * sizeof(cell *)); + } + b->lclistz = 0; + /* Allocate enough space for all the candidate cells */ + if ((b->lclist = (cell **)rev_malloc(s, rip[-3] * sizeof(cell *))) == NULL) + error("rev: malloc failed - candidate cell list, count %d",rip[-3]); + b->lclistz = rip[-3]; /* Current allocated space */ + INCSZ(b->s, b->lclistz * sizeof(cell *)); + } + + /* Get the next simplex touch count, so that we don't search shared */ + /* face simplexes more than once in this pass through the cells. */ + if ((stouch = ++s->rev.stouch) == 0) { /* If touch count rolls over */ + cell *cp; + stouch = s->rev.stouch = 1; + + DBG(("touch has rolled over, resetting it\n")); + /* For all of the cells */ + for (cp = s->rev.cache->mrubot; cp != NULL; cp = cp->mruup) { + int nsdi; + + if (cp->s == NULL) /* Cell has never been used */ + continue; + + /* For all the simplexes in the cell */ + for (nsdi = 0; nsdi <= s->di; nsdi++) { + if (cp->sx[nsdi] != NULL) { + int si; + + for (si = 0; si < cp->sxno[nsdi]; si++) { + cp->sx[nsdi][si]->touch = 0; + } + } + } + } + } + + /* For each chunk of the list that we can fit in the rcache: */ + for(; *rip != -1;) { + + /* Go through all the candidate fwd cells, and build up the list of search cells */ + for(nilist = 0; *rip != -1; rip++) { + int ix = *rip; /* Fwd cell index */ + float *fcb = s->g.a + ix * s->g.pss; /* Pointer to base float of fwd cell */ + cell *c; + + if (TOUCHF(fcb) >= tcount) { /* If we have visited this cell before */ + DBG((" Already touched cell index %d\n",ix)); + continue; + } + /* Get pointers to cells from cache, and lock it in the cache */ + if ((c = get_rcell(b, ix, nilist == 0 ? 1 : 0)) == NULL) { + static int warned = 0; + if (!warned) { + warning("%cWarning - Reverse Cell Cache exausted, processing in chunks",cr_char); + warned = 1; + } + DBG(("revcache is exausted, do search in chunks\n")); + if (nilist == 0) { + /* This should never happen, because nz force should prevent it */ + revcache *rc = s->rev.cache; + cell *cp; + int nunlk = 0; + /* Double check that there are no unlocked cells */ + for (cp = rc->mrubot; cp != NULL && cp->refcount > 0; cp = cp->mruup) { + if (cp->refcount == 0) + nunlk++; + } + fprintf(stdout,"Diagnostic: rev.sz = %lu, rev.max_sz = %lu, numlocked = %d, nunlk = %d\n", + rc->s->rev.sz, rc->s->rev.max_sz, rc->nunlocked,nunlk); + error("Not enough memory to process in chunks"); + } + break; /* cache has run out of room, so abandon, and do it next time */ + } + + DBG(("checking out cell %d range %s\n",ix,pcellorange(c))); + TOUCHF(fcb) = tcount; /* Touch it */ + + /* Check mandatory conditions, and compute search key */ + if (!b->setsort(b, c)) { + DBG(("cell %d rejected from list\n",ix)); + unget_rcell(s->rev.cache, c); + continue; + } + DBG(("cell %d accepted into list\n",ix)); + + b->lclist[nilist++] = c; /* Cell is accepted as recursion candidate */ + } + + if (nilist == 0) { + DBG(("List was empty\n")); + } + +#ifdef DOSORT + /* If appropriate, sort child cells into best order */ + /* == sort key smallest to largest */ + switch (b->op) { + case locus: + { /* Special case, adjust sort values */ + double min = INF_DIST, max = -INF_DIST; + for (i = 0; i < nilist; i++) { + cell *c = b->lclist[i]; + if (c->sort < min) + min = c->sort; + if (c->sort > max) + max = c->sort; + } + max = min + max; /* Total of min/max */ + min = 0.5 * max; /* Average sort value */ + for (i = 0; i < nilist; i++) { + cell *c = b->lclist[i]; + if (c->ix == b->plmincell || c->ix == b->plmaxcell) { + c->sort = -1.0; /* Put previous solution cells at head of list */ + } else if (c->sort > min) { + c->sort = max - c->sort; /* Reflect about average */ + } + } + } + /* Fall through to sort */ + case auxil: + case clipv: + case clipn: +#define HEAP_COMPARE(A,B) (A->sort < B->sort) + HEAPSORT(cell *,b->lclist, nilist) +#undef HEAP_COMPARE + break; + default: + break; + } +#endif /* DOSORT */ + + DBG(("List sorted, about to search\n")); +#ifdef NEVER + printf("\n~1 Op = %s, Cell sort\n",opnames[b->op]); + for (i = 0; i < nilist; i++) { + printf("~1 List %d, cell %d, sort = %f\n",i,b->lclist[i]->ix,b->lclist[i]->sort); + } +#endif /* NEVER */ + + /* + Tried reversing the "for each cell" and "for each level" loops, + but it made a negligible difference to the performance. + We choose to have cell on the outer so that we can unlock + them as we go, so that they may be freed, even though + this is a couple of percent slower (?). + */ + + /* For each cell in the list */ + for (i = 0; i < nilist; i++) { + cell *c = b->lclist[i]; + +#ifdef STATS + s->rev.st[b->op].csearched++; +#endif /* STATS */ + + /* For each dimensionality of sub-simplexes, in given order */ + DBG(("Searching from level %d to level %d\n",b->snsdi, b->ensdi)); + for (nsdi = b->snsdi;;) { + int j, nospx; /* Number of simplexes in cell */ + + DBG(("\n******************\n")); + DBG(("Searching level %d\n",nsdi)); + + /* For those searches that have an optimisation goal, */ + /* re-check the cell to see if the goal can still improve on. */ + if (b->check != NULL && !b->check(b, c)) + break; + + if (c->sx[nsdi] == NULL) { + alloc_simplexes(c, nsdi); /* Do level 1 initialisation for nsdi */ + } + + /* For each simplex in a cell */ + nospx = c->sxno[nsdi]; /* Number of nsdi simplexes */ + for (j = 0; j < nospx; j++) { + simplex *x = c->sx[nsdi][j]; + + if (x->touch >= stouch) { + continue; /* We've already seen this one */ + } + + if (s->limiten == 0) { + if (x->flags & SPLX_CLIPSX) /* If limiting is disabled, we're */ + continue; /* not interested in clip plane simplexes */ + } +#ifdef STATS + s->rev.st[b->op].ssearched++; +#endif /* STATS */ + if (b->compute(b, x)) { + DBG(("search aborted by compute\n")); + break; /* Found enough solutions */ + } + x->touch = stouch; /* Don't look at it again */ + + } /* Next Simplex */ + + if (nsdi == b->ensdi) + break; /* We're done with levels */ + + /* Next Simplex dimensionality */ + if (b->ensdi < b->snsdi) { + if (nsdi == b->snsdi && b->nsoln > 0 + && (b->op != auxil || b->idist <= 2.0 * EPS)) + break; /* Don't continue though decreasing */ + /* sub-simplex dimensions if we found a solution at */ + /* the highest dimension level. */ + nsdi--; + } else if (b->ensdi > b->snsdi) { + nsdi++; /* Continue through increasing sub-simplex dimenionality */ + } /* until we get to the top. */ + } + /* Unlock the cache cell now that we're done with it */ + unget_rcell(s->rev.cache, b->lclist[i]); + } /* Next cell */ + + } /* Next chunk */ + + DBG(("search_list complete\n")); + return; +} + +/* ------------------------------------- */ +/* Vector search in output space support */ + +/* Setup the line, and fetch the first cell */ +/* Return the pointer to the list of fwd cells, NULL if none in list. */ +static int * +init_line( + rspl *s, /* this */ + line *l, /* line structure */ + double st[MXRO], /* start of line */ + double de[MXRO] /* line direction and length */ +) { + int f, fdi = s->fdi; + int **rpp; + int rgres_1 = s->rev.res - 1; + int nvalid = 0; /* Flag set if outside reverse grid range */ + + DBGV(("Line from ", fdi, " %f", st, "\n")); + DBGV(("In dir ", fdi, " %f", de, "\n")); + DBGV(("gl ", fdi, " %f", s->rev.gl, "\n")); + DBGV(("gh ", fdi, " %f", s->rev.gh, "\n")); + DBGV(("gw ", fdi, " %f", s->rev.gw, "\n")); + + /* Init */ + l->s = s; + for (f = 0; f < fdi; f++) { + l->st[f] = st[f] - s->rev.gl[f]; + l->de[f] = de[f]; + if (de[f] > 0.0) + l->di[f] = 1; /* Axis increments */ + else if (de[f] < 0.0) + l->di[f] = -1; + else + l->di[f] = 0; + } + l->t = 0.0; + DBGV(("increments =", fdi, " %d", l->di, "\n")); + + /* Figure out the starting cell */ + for (rpp = s->rev.rev, f = 0; f < fdi; f++) { + double t = l->st[f]/s->rev.gw[f]; + l->ci[f] = (int)floor(t); /* Grid coordinate */ + if (l->ci[f] < 0 || l->ci[f] > rgres_1) /* If outside valid reverse range */ + nvalid = 1; + rpp += l->ci[f] * s->rev.coi[f]; /* Accumulate reverse grid pointer */ + } + DBGV(("current line cell = ", fdi, " %d", l->ci, "")); DBG((", t = %f, nvalid = %d\n",l->t,nvalid)); +#ifdef DEBUG +{ +int ii; +double tt; +printf("Current cell = "); +for (ii = 0; ii < fdi; ii++) { + tt = l->ci[ii] * s->rev.gw[ii] + s->rev.gl[ii]; + printf(" %f - %f",tt,tt+s->rev.gw[ii]); +} +printf("\n"); +} +#endif /* DEBUG */ + if (nvalid) + return NULL; + if (*rpp == NULL) + return NULL; + return *rpp + 3; +} + +/* Get the next cell on the line. */ +/* Return the pointer to the list of fwd cells, NULL if none in list. */ +static int * +next_line_cell( + line *l /* line structure */ +) { + rspl *s = l->s; + int bf = 0, f, fdi = s->fdi; + int **rpp; + int rgres_1 = s->rev.res - 1; + double bt = 100.0; /* Best (smalest +ve) parameter value to move */ + + /* See which axis cell crossing we will hit next */ + for (f = 0; f < fdi; f++) { + double t; + if (l->de[f] != 0) { + t = ((l->ci[f] + l->di[f]) * s->rev.gw[f] - l->st[f])/l->de[f]; + DBG(("t for dim %d = %f\n",f,t)); + if (t < bt) { + bt = t; + bf = f; /* Best direction to move */ + } + } + } + + /* Move to the next reverse grid coordinate */ + l->ci[bf] += l->di[bf]; + l->t = bt; + + DBGV(("current line cell =", fdi, " %d", l->ci, "")); DBG((", t = %f\n",l->t)); + +#ifdef DEBUG +{ +int ii; +double tt; +printf("Current cell = "); +for (ii = 0; ii < fdi; ii++) { + tt = l->ci[ii] * s->rev.gw[ii] + s->rev.gl[ii]; + printf(" %f - %f",tt,tt+s->rev.gw[ii]); +} +printf("\n"); +} +#endif /* DEBUG */ + + /* Compute reverse cell index */ + for (rpp = s->rev.rev, f = 0; f < fdi; f++) { + if (l->ci[f] < 0 || l->ci[f] > rgres_1) { /* If outside valid reverse range */ + DBG(("Outside list on dim %d, 0 <= %d <= %d\n", f, l->ci[f],rgres_1)); + return NULL; + } + rpp += l->ci[f] * s->rev.coi[f]; /* Accumulate reverse grid pointer */ + } + if (*rpp == NULL) + return NULL; + return *rpp + 3; +} + +/* ------------------------------------- */ +/* Clip nearest support. */ + +/* Track candidate cells nearest and furthest */ +struct _nncell_nf{ + double n, f; +}; typedef struct _nncell_nf nncell_nf; + +/* Given and empty nnrev index, create a list of */ +/* the forward cells that may contain the nearest value by */ +/* using and exaustive search. This is used for faststart. */ +static void fill_nncell( + rspl *s, + int *co, /* Integer coords of cell to be filled */ + int ix /* Index of cell to be filled */ +) { + int i; + int e, di = s->di; + int f, fdi = s->fdi; + double cc[MXDO]; /* Cell center */ + double rr = 0.0; /* Cell radius */ + int **rpp, *rp; + int gno = s->g.no; + float *gp; /* Pointer to fwd grid points */ + nncell_nf *nf; /* cloase and far distances corresponding to list */ + double clfu = 1e38; /* closest furthest distance in list */ + + rpp = s->rev.nnrev + ix; + rp = *rpp; + + /* Compute the center location and radius of the target cell */ + for (f = 0; f < fdi; f++) { + cc[f] = s->rev.gw[f] * (co[f] + 0.5) + s->rev.gl[f]; + rr += 0.25 * s->rev.gw[f] * s->rev.gw[f]; + } + rr = sqrt(rr); +//printf("~1 fill_nncell() cell ix %d, coord %d %d %d, cent %f %f %f, rad %f\n", +//ix, co[0], co[1], co[2], cc[0], cc[1], cc[2], rr); +//printf("~1 total of %d fwd cells\n",gno); + + /* For all the forward cells: */ + for (gp = s->g.a, i = 0; i < gno; gp += s->g.pss, i++) { + int ee; + int uil; /* One is under the ink limit */ + double dn, df; /* Nearest and farthest distance of fwd cell values */ + + /* Skip cubes that are on the outside edge of the grid */ + for (e = 0; e < di; e++) { + if(G_FL(gp, e) == 0) /* At the top edge */ + break; + } + if (e < di) { /* Top edge - skip this cube */ + continue; + } + + /* Compute the closest and furthest distances of nodes of current cell */ + dn = 1e38, df = 0.0; + for (uil = ee = 0; ee < (1 << di); ee++) { /* For all grid points in the cube */ + double r; + float *gt = gp + s->g.fhi[ee]; /* Pointer to cube vertex */ + + if (!s->limiten || gt[-1] <= s->limitv) + uil = 1; + + /* Update bounding box for this grid point */ + for (r = 0.0, f = 0; f < fdi; f++) { + double tt = cc[f] - (double)gt[f]; + r += tt * tt; + } +//printf("~1 grid location %f %f %f rad %f\n",gt[0],gt[1],gt[2],sqrt(r)); + if (r < dn) + dn = r; + if (r > df) + df = r; + } + /* Skip any fwd cells that are over the ink limit */ + if (!uil) + continue; + + dn = sqrt(dn) - rr; + df = sqrt(df) + rr; + +//printf("~1 checking cell %d, near %f, far %f\n",i,dn,df); + + /* Skip any that have a closest distance larger that the lists */ + /* closest furthest distance. */ + if (dn > clfu) { +//printf("~1 skipping cell %d, near %f, far %f clfu %f\n",i,dn,df,clfu); + continue; + } + +//printf("~1 adding cell %d\n",i); + if (rp == NULL) { + if ((nf = (nncell_nf *) rev_malloc(s, 6 * sizeof(nncell_nf))) == NULL) + error("rspl malloc failed - nncell_nf list"); + INCSZ(s, 6 * sizeof(nncell_nf)); + if ((rp = (int *) rev_malloc(s, 6 * sizeof(int))) == NULL) + error("rspl malloc failed - rev.grid entry"); + INCSZ(s, 6 * sizeof(int)); + *rpp = rp; + rp[0] = 6; /* Allocation */ + rp[1] = 4; /* Next empty cell */ + rp[2] = 1; /* Reference count */ + rp[3] = i; + nf[3].n = dn; + nf[3].f = df; + rp[4] = -1; + } else { + int z = rp[1], ll = rp[0]; + if (z >= (ll-1)) { /* Not enough space */ + INCSZ(s, ll * sizeof(nncell_nf)); + INCSZ(s, ll * sizeof(int)); + ll *= 2; + if ((nf = (nncell_nf *) rev_realloc(s, nf, sizeof(nncell_nf) * ll)) == NULL) + error("rspl realloc failed - nncell_nf list"); + if ((rp = (int *) rev_realloc(s, rp, sizeof(int) * ll)) == NULL) + error("rspl realloc failed - rev.grid entry"); + *rpp = rp; + rp[0] = ll; + } + rp[z] = i; + nf[z].n = dn; + nf[z++].f = df; + rp[z] = -1; + rp[1] = z; + } + + if (df < clfu) + clfu = df; + } +//printf("~1 Current list is:\n"); +//for (e = 3; rp[e] != -1; e++) +//printf(" %d: Cell %d near %f far %f\n",e,rp[e],nf[e].n,nf[e].f); + + /* Now filter out any cells that have a closest point that is further than */ + /* closest furthest point */ + { + int z, w, ll = rp[0]; + + /* For all the cells in the current list: */ + for (w = z = 3; rp[z] != -1; z++) { + + /* If the new cell nearest is greater than the existing cell closest, */ + /* then don't omit existing cell from the list. */ + if (clfu >= nf[z].n) { + rp[w] = rp[z]; + nf[w].n = nf[z].n; + nf[w].f = nf[z].f; + w++; + } +//else printf("~1 deleting cell %d because %f >= %f\n",rp[z],clfu, nf[z].f); + } + rp[w] = rp[z]; + } +//printf("~1 Current list is:\n"); +//for (e = 3; rp[e] != -1; e++) +//printf(" %d: Cell %d near %f far %f\n",e,rp[e],nf[e].n,nf[e].f); + free(nf); +//printf("~1 Done\n"); +} + +/* Return the pointer to the list of nearest fwd cells given */ +/* the target output values. The pointer will be to the first */ +/* index in the list (ie. list address + 3) */ +/* Return NULL if none in list (out of gamut). */ +static int * +calc_fwd_nn_cell_list( + rspl *s, /* this */ + double *v /* Output values */ +) { + int f, fdi = s->fdi, ix; + int **rpp; + int rgres_1 = s->rev.res - 1; + int mi[MXDO]; + + if (s->rev.rev_valid == 0) + init_revaccell(s); + + for (ix = 0, f = 0; f < fdi; f++) { + double t = (v[f] - s->rev.gl[f])/s->rev.gw[f]; + mi[f] = (int)floor(t); /* Grid coordinate */ + if (mi[f] < 0) /* Clip to reverse range, so we always return a result */ + mi[f] = 0; + else if (mi[f] > rgres_1) + mi[f] = rgres_1; + ix += mi[f] * s->rev.coi[f]; /* Accumulate reverse grid index */ + } + rpp = s->rev.nnrev + ix; + if (*rpp == NULL) { + if (s->rev.fastsetup) + fill_nncell(s, mi, ix); + if (*rpp == NULL) + rpp = s->rev.rev + ix; /* fall back to in-gamut lookup */ + } + if (*rpp == NULL) + return NULL; + return (*rpp) + 3; +} + +/* =================================================== */ +/* The cell and simplex solver top level routines */ + +static int add_lu_svd(simplex *x); +static int add_locus(schbase *b, simplex *x); +static int add_auxil_lu_svd(schbase *b, simplex *x); +static int within_simplex(simplex *x, double *p); +static void simplex_to_abs(simplex *x, double *in, double *out); + +static int auxil_solve(schbase *b, simplex *x, double *xp); + +/* ---------------------- */ +/* Exact search functions */ +/* Return non-zero if cell is acceptable */ +static int exact_setsort(schbase *b, cell *c) { + rspl *s = b->s; + int f, fdi = s->fdi; + double ss; + + DBG(("Reverse exact search, evaluate and set sort key on cell\n")); + + /* Check that the target lies within the cell bounding sphere */ + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt = c->bcent[f] - b->v[f]; + ss += tt * tt; + } + if (ss > c->bradsq) { + DBG(("Cell rejected - %s outside sphere c %s rad %f\n",icmPdv(fdi,b->v),icmPdv(fdi,c->bcent),sqrt(c->bradsq))); + return 0; + } + + if (s->limiten != 0 && c->limmin > s->limitv) { + DBG(("Cell is rejected - ink limit, min = %f, limit = %f\n",c->limmin,s->limitv)); + return 0; + } + + /* Sort can't be used, because we return all solutions */ + c->sort = 0.0; + + DBG(("Cell is accepted\n")); + + return 1; +} + +/* Compute a solution for a given sub-simplex (if there is one) */ +/* Return 1 if search should be aborted */ +static int exact_compute(schbase *b, simplex *x) { + rspl *s = b->s; + int e, di = s->di, sdi = x->sdi; + int f, fdi = s->fdi; + int i; + datai xp; /* solution in simplex relative coord order */ + datai p; /* absolute solution */ + int wsrv; /* Within simplex return value */ + + DBG(("\nExact: computing possible solution\n")); + +#ifdef DEBUG + /* Sanity check */ + if (sdi != fdi || sdi != di || x->efdi != fdi) { + printf("di = %d, fdi = %d\n",di,fdi); + printf("sdi = %d, efdi = %d\n",sdi,x->efdi); + error("rspl exact reverse interp called with sdi != fdi, sdi != di, efdi != fdi"); + /* !!! could switch to SVD solution if di != fdi ?? !!! */ + } +#endif + + /* This may not be worth it here since it may not filter out */ + /* many more simplexes than the cube check did. */ + /* This is due to full dimension simplexes all sharing the main */ + /* diagonal axis. */ + + /* Check that the target lies within the simplex bounding cube */ + for (f = 0; f < fdi; f++) { + if (b->v[f] < x->min[f] || b->v[f] > x->max[f]) { + DBG(("Simplex is rejected - bounding cube\n")); + return 0; + } + } + + /* Create the LU decomp needed to exactly solve */ + if (add_lu_svd(x)) { + DBG(("LU decomp was singular, skip simplex\n")); + return 0; + } + + /* Init the RHS B[] vector (note di == fdi) */ + for (f = 0; f < fdi; f++) { + xp[f] = b->v[f] - x->v[di][f]; + } + + /* Compute the solution (in simplex space) */ + lu_backsub(x->d_u, sdi, (int *)x->d_w, xp); + + /* Check that the solution is within the simplex */ + if ((wsrv = within_simplex(x, xp)) == 0) { + DBG(("Solution rejected because not in simplex\n")); + return 0; + } + + /* Convert solution from simplex relative to absolute space */ + simplex_to_abs(x, p, xp); + + /* Check if a very similiar input solution has been found before */ + for (i = 0; i < b->nsoln; i++) { + double tt; + for (e = 0; e < di; e++) { + tt = b->cpp[i].p[e] - p[e]; + if (fabs(tt) > (2 * EPS)) + break; /* Mismatch */ + } + if (e >= di) /* Found good match */ + break; + } + + /* Probably alias caused by solution lying close to a simplex boundary */ + if (i < b->nsoln) { + DBG(("Another solution has been found before - index %d\n",i)); + return 0; /* Skip this, since betters been found before */ + } + + /* Check we haven't overflowed space */ + if (i >= b->mxsoln) { + DBG(("Run out of space for new solution\n")); + return 1; /* Abort */ + } + + DBG(("######## Accepting new solution\n")); + + /* Put solution in place */ + for (e = 0; e < di; e++) + b->cpp[i].p[e] = p[e]; + for (f = 0; f < fdi; f++) + b->cpp[i].v[f] = b->v[f]; /* Assumed to be an exact solution */ + if (i == b->nsoln) + b->nsoln++; + if (wsrv == 2) /* Is above (disabled) ink limit */ + b->iclip = 1; + return 0; +} + +/* -------------------------- */ +/* Auxiliary search functions */ +static int auxil_setsort(schbase *b, cell *c) { + rspl *s = b->s; + int f, fdi = b->s->fdi; + int ee, ixc = b->ixc; + double ss, sort, nabove; + + DBG(("Reverse auxiliary search, evaluate and set sort key on cell\n")); + + if (b->s->di <= fdi) { /* Assert */ + error("rspl auxiliary reverse interp called with di <= fdi (%d %d)", b->s->di, fdi); + } + + /* Check that the target lies within the cell bounding sphere */ + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt = c->bcent[f] - b->v[f]; + ss += tt * tt; + } + if (ss > c->bradsq) { + DBG(("Cell rejected - %s outside sphere c %s rad %f\n",icmPdv(fdi,b->v),icmPdv(fdi,c->bcent),sqrt(c->bradsq))); + return 0; + } + + if (s->limiten != 0 && c->limmin > s->limitv) { + DBG(("Cell is rejected - ink limit, min = %f, limit = %f\n",c->limmin,s->limitv)); + return 0; + } + + /* Check if this cell could possible improve b->idist */ + /* and compute sort key as the distance to auxilliary target */ + /* (We may have a non INF_DIST idist before commencing the */ + /* search if we already know that the auxiliary target is */ + /* within gamut - the usual usage case!) */ + for (sort = 0.0, nabove = ee = 0; ee < b->naux; ee++) { + int ei = b->auxi[ee]; + double tt = (c->p[0][ei] + c->p[ixc][ei]) - b->av[ei]; + sort += tt * tt; + if (c->p[ixc][ei] >= (b->av[ei] - EPS)) /* Could be above */ + nabove++; + } + + if (b->flags & RSPL_MAXAUX && nabove < b->iabove) { + DBG(("Doesn't contain solution that has as many aux above auxiliary goal\n")); + return 0; + } + if (!(b->flags & RSPL_MAXAUX) || nabove == b->iabove) { + for (ee = 0; ee < b->naux; ee++) { + int ei = b->auxi[ee]; + if (c->p[0][ei] >= (b->av[ei] + b->idist) + || c->p[ixc][ei] <= (b->av[ei] - b->idist)) { + DBG(("Doesn't contain solution that will be closer to auxiliary goal\n")); + return 0; + } + } + } + c->sort = sort + 0.01 * ss; + + if (c->ix == b->pauxcell) + c->sort = -1.0; /* Put previous calls solution cell at top of sort list */ + + DBG(("Cell is accepted\n")); + return 1; +} + +/* Re-check whether it's worth searching cell */ +static int auxil_check(schbase *b, cell *c) { + int ee, ixc = b->ixc, nabove; + + DBG(("Reverse auxiliary search, re-check cell\n")); + + /* Check if this cell could possible improve b->idist */ + /* and compute sort key as the distance to auxilliary target */ + + for (nabove = ee = 0; ee < b->naux; ee++) { + int ei = b->auxi[ee]; + if (c->p[ixc][ei] >= (b->av[ei] - EPS)) /* Could be above */ + nabove++; + } + + if (b->flags & RSPL_MAXAUX && nabove < b->iabove) { + DBG(("Doesn't contain solution that has as many aux above auxiliary goal\n")); + return 0; + } + if (!(b->flags & RSPL_MAXAUX) || nabove == b->iabove) { + for (ee = 0; ee < b->naux; ee++) { + int ei = b->auxi[ee]; + if (c->p[0][ei] >= (b->av[ei] + b->idist) + || c->p[ixc][ei] <= (b->av[ei] - b->idist)) { + DBG(("Doesn't contain solution that will be closer to auxiliary goal\n")); + return 0; + } + } + } + DBG(("Cell is still ok\n")); + return 1; +} + +/* Compute a solution for a given simplex (if there is one) */ +/* Return 1 if search should be aborted */ +static int auxil_compute(schbase *b, simplex *x) { + rspl *s = b->s; + int e, di = s->di; + int f, fdi = s->fdi; + datai xp; /* solution in simplex relative coord order */ + datai p; /* absolute solution */ + double idist; /* Auxiliary input distance */ + int wsrv; /* Within simplex return value */ + int nabove; /* Number above aux target */ + + DBG(("\nAuxil: computing possible solution\n")); + +#ifdef DEBUG + { + unsigned int sum = 0; + for (f = 0; f <= x->sdi; f++) + sum += x->vix[f]; + printf("Simplex of cell ix %d, sum 0x%x, sdi = %d, efdi = %d\n",x->ix, sum, x->sdi, x->efdi); + printf("Target val %s\n",icmPdv(fdi, b->v)); + for (f = 0; f <= x->sdi; f++) { + int ix = x->vix[f], i; + float *fcb = s->g.a + ix * s->g.pss; /* Pointer to base float of fwd cell */ + printf("Simplex vtx %d [cell ix %d] val %s\n",f,ix,icmPfv(fdi, fcb)); + } + } +#endif + + /* Check that the target lies within the simplex bounding cube */ + for (f = 0; f < fdi; f++) { + if (b->v[f] < x->min[f] || b->v[f] > x->max[f]) { + DBG(("Simplex is rejected - bounding cube\n")); + return 0; + } + } + + /* Check if this cell could possible improve b->idist */ + for (nabove = e = 0; e < b->naux; e++) { + int ei = b->auxi[e]; /* pmin/max[] is indexed in input space */ + if (x->pmax[ei] >= (b->av[ei] - EPS)) /* Could be above */ + nabove++; + } + if ((b->flags & RSPL_MAXAUX) && nabove < b->iabove) { + DBG(("Simplex doesn't contain solution that has as many aux above auxiliary goal\n")); + return 0; + } + if (!(b->flags & RSPL_MAXAUX) || nabove == b->iabove) { + for (nabove = e = 0; e < b->naux; e++) { + int ei = b->auxi[e]; /* pmin/max[] is indexed in input space */ + if (x->pmin[ei] >= (b->av[ei] + b->idist) + || x->pmax[ei] <= (b->av[ei] - b->idist)) { + DBG(("Simplex doesn't contain solution that will be closer to auxiliary goal\n")); + return 0; + } + } + } + +//printf("~~ About to create svd decomp\n"); + /* Create the SVD or LU decomp needed to compute solution or locus */ + if (add_lu_svd(x)) { + DBG(("SVD decomp failed, skip simplex\n")); + return 0; + } + +//printf("~~ About to solve locus for aux target\n"); + /* Now solve for locus parameter that minimises */ + /* distance to auxliary target. */ + if ((wsrv = auxil_solve(b, x, xp)) == 0) { + DBG(("Target auxiliary along locus is outside simplex,\n")); + DBG(("or computation failed, skip simplex\n")); + return 0; + } + +//printf("~~ About to convert solution to absolute space\n"); + /* Convert solution from simplex relative to absolute space */ + simplex_to_abs(x, p, xp); + + DBG(("Got solution at %s\n", icmPdv(di,p))); + +//printf("~~ soln = %f %f %f %f\n",p[0],p[1],p[2],p[3]); +//printf("~~ About to compute auxil distance\n"); + /* Compute distance to auxiliary target */ + for (idist = 0.0, nabove = e = 0; e < b->naux; e++) { + int ei = b->auxi[e]; + double tt = b->av[ei] - p[ei]; + idist += tt * tt; + if (p[ei] >= (b->av[ei] - EPS)) + nabove++; + } + idist = sqrt(idist); +//printf("~1 idist %f, nabove %d\n",idist, nabove); +//printf("~1 best idist %f, best iabove %d\n",b->idist, b->iabove); + + /* We want the smallest error from auxiliary target */ + if (b->flags & RSPL_MAXAUX) { + if (nabove < b->iabove || (nabove == b->iabove && idist >= b->idist)) { + DBG(("nsoln %d, nabove %d, iabove %d, idist = %f, better solution has been found before\n",b->nsoln, nabove, b->iabove, idist)); + return 0; + } + } else { + if (idist >= b->idist) { /* Equal or worse auxiliary solution */ + DBG(("nsoln %d, idist = %f, better solution has been found before\n",b->nsoln,idist)); + return 0; + } + } + + /* Solution is accepted */ + DBG(("######## Accepting new solution with nabove %d <= iabove %d and idist %f <= %f\n",nabove,b->iabove,idist,b->idist)); + for (e = 0; e < di; e++) + b->cpp[0].p[e] = p[e]; + for (f = 0; f < fdi; f++) + b->cpp[0].v[f] = b->v[f]; /* Assumed to be an exact solution */ + b->idist = idist; + b->iabove = nabove; + b->nsoln = 1; + b->pauxcell = x->ix; + if (wsrv == 2) /* Is above (disabled) ink limit */ + b->iclip = 1; + + return 0; +} + +/* ------------------------------------ */ +/* Locus range search functions */ + +static int locus_setsort(schbase *b, cell *c) { + rspl *s = b->s; + int f, fdi = s->fdi; + int lxi = b->lxi; /* Auxiliary we are finding min/max of */ + int ixc = b->ixc; + double sort, ss; + + DBG(("Reverse locus evaluate and set sort key on cell\n")); + +#ifdef DEBUG + if (b->s->di <= fdi) { /* Assert ~1 */ + error("rspl auxiliary locus interp called with di <= fdi"); + } +#endif /* DEBUG */ + + /* Check that the target lies within the cell bounding sphere */ + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt = c->bcent[f] - b->v[f]; + ss += tt * tt; + } + if (ss > c->bradsq) { + DBG(("Cell rejected - %s outside sphere c %s rad %f\n",icmPdv(fdi,b->v),icmPdv(fdi,c->bcent),sqrt(c->bradsq))); + return 0; + } + + if (s->limiten != 0 && c->limmin > s->limitv) { + DBG(("Cell is rejected - ink limit, min = %f, limit = %f\n",c->limmin,s->limitv)); + return 0; + } + + /* Check if this cell could possible improve the locus min/max */ + if (b->asegs == 0) { /* If we aren't find all segments of the locus */ + if (c->p[0][lxi] >= b->min && c->p[ixc][lxi] <= b->max ) { + DBG(("Doesn't contain solution that will expand the locus\n")); + return 0; + } + } + + /* Compute sort index from average of auxiliary values */ + sort = (c->p[0][b->lxi] + c->p[ixc][b->lxi]); + + c->sort = sort + 0.01 * ss; + + DBG(("Cell is accepted\n")); + return 1; +} + +/* Re-check whether it's worth searching simplexes */ +static int locus_check(schbase *b, cell *c) { + int lxi = b->lxi; /* Auxiliary we are finding min/max of */ + int ixc = b->ixc; + + DBG(("Reverse locus re-check\n")); + + /* Check if this cell could possible improve the locus min/max */ + if (b->asegs == 0) { /* If we aren't find all segments of the locus */ + if (c->p[0][lxi] >= b->min && c->p[ixc][lxi] <= b->max ) { + DBG(("Doesn't contain solution that will expand the locus\n")); + return 0; + } + } + + DBG(("Cell is still ok\n")); + return 1; +} + +static int auxil_locus(schbase *b, simplex *x); + +/* We expect to be given a sub-simplex with no DOF, to give an exact solution */ +static int locus_compute(schbase *b, simplex *x) { + rspl *s = b->s; + int f, fdi = s->fdi; + int lxi = b->lxi; /* Auxiliary we are finding min/max of */ + + DBG(("\nLocus: computing possible solution\n")); + +#ifdef DEBUG + { + unsigned int sum = 0; + for (f = 0; f <= x->sdi; f++) + sum += x->vix[f]; + printf("Simplex of cell ix %d, sum 0x%x, sdi = %d, efdi = %d\n",x->ix, sum, x->sdi, x->efdi); + printf("Target val %s\n",icmPdv(fdi, b->v)); + for (f = 0; f <= x->sdi; f++) { + int ix = x->vix[f], i; + float *fcb = s->g.a + ix * s->g.pss; /* Pointer to base float of fwd cell */ + double v[MXDO]; + printf("Simplex vtx %d [cell ix %d] val %s\n",f,ix,icmPfv(fdi, fcb)); + } + } +#endif + + /* Check that the target lies within the simplex bounding cube */ + for (f = 0; f < fdi; f++) { + if (b->v[f] < x->min[f] || b->v[f] > x->max[f]) { + DBG(("Simplex is rejected - bounding cube\n")); + return 0; + } + } + + /* Check if simplex could possible improve the locus min/max */ + if (b->asegs == 0) { /* If we aren't find all segments of the locus */ + if (x->pmin[lxi] >= b->min && x->pmax[lxi] <= b->max ) { + DBG(("Simplex doesn't contain solution that will expand the locus\n")); + return 0; + } + } + +//printf("~~ About to create svd decomp\n"); + /* Create the SVD decomp needed to compute solution extreme points */ + if (add_lu_svd(x)) { + DBG(("SVD decomp failed, skip simplex\n")); + return 0; + } + +//printf("~~ About to solve locus for aux extremes\n"); + /* Now solve for locus parameter that are at the extremes */ + /* of the axiliary we are interested in. */ + if (!auxil_locus(b, x)) { + DBG(("Target auxiliary is outside simplex,\n")); + DBG(("or computation failed, skip simplex\n")); + return 0; + } + + return 0; +} + +/* ------------------- */ +/* Vector clipping search functions */ +static int clipv_setsort(schbase *b, cell *c) { + rspl *s = b->s; + int f, fdi = s->fdi; + double ss, dp; + + DBG(("Reverse clipping search evaluate cell\n")); + +//printf("~~sphere center = %f %f %f, radius %f\n",c->bcent[0],c->bcent[1],c->bcent[2],sqrt(c->bradsq)); + /* Check if the clipping line intersects the bounding sphere */ + /* First compute dot product cdir . (bcent - v) */ + /* == distance to center of sphere in direction of clip vector */ + for (dp = 0.0, f = 0; f < fdi; f++) { + dp += b->ncdir[f] * (c->bcent[f] - b->v[f]); + } + + if (s->limiten != 0 && c->limmin > s->limitv) { + DBG(("Cell is rejected - ink limit, min = %f, limit = %f\n",c->limmin,s->limitv)); + return 0; + } + +//printf("~~ dot product = %f\n",dp); + /* Now compute closest distance to sphere center */ + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt = b->v[f] + dp * b->ncdir[f] - c->bcent[f]; + ss += tt * tt; + } + +//printf("~~ distance to sphere center = %f\n",sqrt(ss)); + if (ss > c->bradsq) { + DBG(("Cell is rejected - wrong direction or bounding sphere\n")); + return 0; + } + c->sort = dp; /* May be -ve if beyond clip target point ? */ + + DBG(("Cell is accepted\n")); + return 1; +} + +/* Clipping check functions */ +/* Note that we don't bother with this check in setsort(), */ +/* because we assume that nothing will set a small cdist */ +/* before the search commences (unlike auxil). */ +/* Note that line search loop exits on finding any solution. */ +static int clipv_check(schbase *b, cell *c) { + + DBG(("Reverse clipping re-check\n")); + + if (b->cdist < INF_DIST) { /* If some clip solution has been found */ + int f, fdi = b->s->fdi; + double dist; + /* Compute a conservative "best possible solution clip distance" */ + for (dist = 0.0, f = 0; f < fdi ; f++) { + double tt = (c->bcent[f] - b->v[f]); + dist += tt * tt; + } + dist = sqrt(dist); /* Target distance to bounding */ + + if (dist >= (c->brad + b->cdist)) { /* Equal or worse clip solution */ + DBG(("Cell best possible solution worse than current\n")); + return 0; + } + } + + DBG(("Cell is still ok\n")); + return 1; +} + +static int vnearest_clip_solve(schbase *b, simplex *x, double *xp, double *xv, double *err); + +/* Compute a clip solution */ +static int clipv_compute(schbase *b, simplex *x) { + rspl *s = b->s; + int f, fdi = s->fdi; + datai p; /* Input space solution */ + datao v; /* Output space solution */ + double err; /* output error of solution */ + int wsrv; /* Within simplex return value */ + + DBG(("Clips: computing possible solution\n")); + + /* Compute a solution value */ + if ((wsrv = vnearest_clip_solve(b, x, p, v, &err)) == 0) { + DBG(("Doesn't contain a solution\n")); + return 0; + } + + /* We want the smallest clip error */ + /* (Should we reject points in -ve vector direction ??) */ + if (err >= b->cdist) { /* Equal or worse clip solution */ + DBG(("better solution has been found before\n")); + return 0; + } + + simplex_to_abs(x, b->cpp[0].p, p); /* Convert to abs. space & copy */ + + DBG(("######## Accepting new clipv solution with error %f\n",err)); +#ifdef DEBUG + if (s->limiten != 0) { + DBG(("######## Ink value = %f, limit %f\n",get_limitv(b, x->ix, NULL, b->cpp[0].p), s->limitv)); + } +#endif + + /* Put solution in place */ + for (f = 0; f < fdi; f++) + b->cpp[0].v[f] = v[f]; + b->cdist = err; + b->nsoln = 1; + if (wsrv == 2) /* Is above (disabled) ink limit */ + b->iclip = 1; + + return 0; +} + +/* ------------------- */ +/* Nearest clipping search functions */ +static int clipn_setsort(schbase *b, cell *c) { + rspl *s = b->s; + int f, fdi = s->fdi; + double ss; + + DBG(("Reverse nearest clipping search evaluate cell\n")); + + /* Compute a conservative "best possible solution clip distance" */ + for (ss = 0.0, f = 0; f < fdi ; f++) { + double tt = (c->bcent[f] - b->v[f]); + ss += tt * tt; + } + ss = sqrt(ss); /* Target distance to bounding sphere */ + ss -= c->brad; + if (ss < 0.0) + ss = 0.0; + + /* Check that the cell could possibly improve the solution */ + if (b->cdist < INF_DIST) { /* If some clip solution has been found */ + if (ss >= b->cdist) { /* Equal or worse clip solution */ + DBG(("Cell best possible solution worse than current\n")); + return 0; + } + } + + if (s->limiten != 0 && c->limmin > s->limitv) { + DBG(("Cell is rejected - ink limit, min = %f, limit = %f\n",c->limmin,s->limitv)); + return 0; + } + + c->sort = ss; /* May be -ve if beyond clip target point ? */ + + DBG(("Cell is accepted\n")); + return 1; +} + +/* Clipping check functions */ +static int clipn_check(schbase *b, cell *c) { + + DBG(("Reverse nearest clipping re-check\n")); + + if (b->cdist < INF_DIST) { /* If some clip solution has been found */ + /* re-use sort value, best possible distance to solution */ + if (c->sort >= b->cdist) { /* Equal or worse clip solution */ + DBG(("Cell best possible solution worse than current\n")); + return 0; + } + } + + DBG(("Cell is still ok\n")); + return 1; +} + +static int nnearest_clip_solve(schbase *b, simplex *x, double *xp, double *xv, double *err); + +/* Compute a clip solution */ +static int clipn_compute(schbase *b, simplex *x) { + rspl *s = b->s; + int f, fdi = s->fdi; + datai p; /* Input space solution */ + datao v; /* Output space solution */ + double err; /* output error of solution */ + int wsrv; /* Within simplex return value */ + + DBG(("Clipn: computing possible solution simplex %d, sdi = %d, efdi = %d\n",x->si,x->sdi,x->efdi)); + + /* Compute a solution value */ + if ((wsrv = nnearest_clip_solve(b, x, p, v, &err)) == 0) { + DBG(("Doesn't contain a solution\n")); + return 0; + } + + /* We want the smallest clip error */ + if (err >= b->cdist) { /* Equal or worse clip solution */ + DBG(("better solution has been found before\n")); + return 0; + } + + DBG(("######## Accepting new clipn solution with error %f\n",err)); + + simplex_to_abs(x, b->cpp[0].p, p); /* Convert to abs. space & copy */ + + /* Put solution in place */ + for (f = 0; f < fdi; f++) + b->cpp[0].v[f] = v[f]; + b->cdist = err; + b->nsoln = 1; + if (wsrv == 2) /* Is above (disabled) ink limit */ + b->iclip = 1; + + return 0; +} + +/* -------------------------------------------------------- */ +/* Cell/simplex solver middle level code */ + +/* Find the point on this sub-simplexes solution locus that is */ +/* closest to the target auxiliary values, and return it in xp[] */ +/* Return zero if this point canot be calculated, */ +/* or it lies outside the simplex. */ +/* Return 1 normally, and 2 if the solution would be over the ink limit */ +static int +auxil_solve( +schbase *b, +simplex *x, +double *xp /* Return solution xp[sdi] */ +) { + rspl *s = b->s; + int ee, e, di = s->di, sdi = x->sdi; + int f, efdi = x->efdi; + int dof = sdi-efdi; /* Degree of freedom of simplex locus */ + int *icomb = x->psxi->icomb; /* abs -> simplex coordinate translation */ + double auxt[MXRI]; /* Simplex relative auxiliary targets */ + double bb[MXRI]; + int wsrv; /* Within simplex return value */ + + DBG(("axuil_solve called\n")); + + if (dof < 0) + error("Error - auxil_solve got sdi < efdi (%d < %d) - don't know how to handle this",sdi, efdi); + + /* If there is no locus, compute an exact solution */ + if (dof == 0) { + DBG(("axuil_solve dof = zero\n")); + + /* Init the RHS B[] vector (note sdi == efdi) */ + for (f = 0; f < efdi; f++) { + xp[f] = b->v[f] - x->v[sdi][f]; + } + + /* Compute the solution (in simplex space) */ + lu_backsub(x->d_u, sdi, (int *)x->d_w, xp); + + if ((wsrv = within_simplex(x, xp)) != 0) { + DBG(("Got solution at %s\n", icmPdv(sdi,xp))); + return wsrv; /* OK, got solution */ + } + + DBG(("No solution (not within simplex)\n")); + return 0; + } + + /* There is a locus, so find solution nearest auxiliaries */ + + /* Compute locus for target function values (if sdi > efdi) */ + if (add_locus(b, x)) { + DBG(("Locus computation failed, skip simplex\n")); + return 0; + } + + /* Convert aux targets from absolute space to simplex relative */ + for (e = 0; e < di; e++) { /* For abs coords */ + int ei = icomb[e]; /* Simplex coord */ + + if (ei >= 0 && b->auxm[e] != 0) { + auxt[ei] = (b->av[e] - x->p0[e])/s->g.w[e]; /* Only sets those needed */ + } + } + + if (dof == 1 && b->naux == 1) { /* Special case, because it's common and easy! */ + int ei = icomb[b->auxi[0]]; /* Simplex relative auxiliary index */ + double tt; + + DBG(("axuil_solve dof = naux = 1\n")); + if (ei < 0) + return 0; /* Not going to find solution */ + if ((tt = x->lo_l[ei][0]) == 0.0) + return 0; + tt = (auxt[ei] - x->lo_bd[ei])/tt; /* Parameter solution for target auxiliary */ + + /* Back substitute parameter */ + for (e = 0; e < sdi; e++) { + xp[e] = x->lo_bd[e] + tt * x->lo_l[e][0]; + } + if ((wsrv = within_simplex(x, xp)) != 0) { + DBG(("Got solution %s\n",icmPdv(di,xp))); + return wsrv; /* OK, got solution */ + } + DBG(("No solution (not within simplex)\n")); + return 0; + } + + /* Compute the locus decompositions needed (info #5) */ + if (add_auxil_lu_svd(b, x)) { /* Will set x->naux */ + DBG(("LU/SVD decomp failed\n")); + return 0; + } + + /* Setup B[], equation RHS */ + for (e = ee = 0; ee < b->naux; ee++) { + int ei = icomb[b->auxi[ee]]; /* Simplex relative auxiliary index */ + if (ei >= 0) /* Usable auxiliary on this sub simplex */ + bb[e++] = auxt[ei] - x->lo_bd[ei]; + } + if (e != x->naux) /* Assert */ + error("Internal error - auxil_solve got mismatching number of auxiliaries"); + + if (x->naux == dof) { /* Use LU decomp to solve */ + DBG(("axuil_solve using LU\n")); + lu_backsub(x->ax_u, dof, (int *)x->ax_w, bb); + + } else if (x->naux > 0) { /* Use SVD to solve least squares */ + DBG(("axuil_solve using SVD\n")); + svdbacksub(x->ax_u, x->ax_w, x->ax_v, bb, bb, x->naux, dof); + + } else { /* x->naux == 0 */ + DBG(("axuil_solve naux = 0\n")); + for (f = 0; f < dof; f++) + bb[f] = 0.0; /* Use base solution ?? */ + } + + /* Now back substitute the locus parameters */ + /* to calculate the solution point (in simplex space) */ + for (e = 0; e < sdi; e++) { + double tt; + for (tt = 0.0, f = 0; f < dof; f++) { + tt += bb[f] * x->lo_l[e][f]; + } + xp[e] = x->lo_bd[e] + tt; + } + + if ((wsrv = within_simplex(x, xp)) != 0) { + DBG(("Got solution %s\n",icmPdv(di,xp))); + return wsrv; /* OK, got solution */ + } + DBG(("No solution (not within simplex)\n")); + return 0; +} + +/* - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ +/* Compute the min/max values for the current auxiliary of interest. */ +/* Return zero if this point canot be calculated, */ +/* or it lies outside the simplex. */ +/* Return 1 normally, 2 if it would be outside the simplex if limting was enabled */ +/* We expect to get a sub-simplex that will give an exact solution. */ +static int +auxil_locus( +schbase *b, +simplex *x +) { + rspl *s = b->s; + int sdi = x->sdi; + int f, efdi = x->efdi; + double pp[MXRI]; + int wsrv; /* Within simplex return value */ + + DBG(("axuil_locus called\n")); + + if (sdi != efdi) + warning("Internal error - auxil_locus got sdi != efdi (%d < %d)",sdi, efdi); + + /* Init the RHS B[] vector (note sdi == efdi) */ + for (f = 0; f < efdi; f++) { + pp[f] = b->v[f] - x->v[sdi][f]; + } + + /* Compute the solution (in simplex space) */ + lu_backsub(x->d_u, sdi, (int *)x->d_w, pp); + + /* Check that the solution is within the simplex */ + if ((wsrv = within_simplex(x, pp)) != 0) { + double xval; + int lxi = b->lxi; /* Auxiliary we are finding min/max of (Abs space) */ + int xlxi = x->psxi->icomb[lxi]; /* Auxiliary we are finding min/max of (simplex space) */ + + DBG(("Got locus solution within simplex\n")); + + /* Compute auxiliary value for this solution (absolute space) */ + xval = x->p0[lxi]; + if (xlxi >= 0) /* Simplex param value */ + xval += s->g.w[lxi] * pp[xlxi]; + else if (xlxi == -2) /* 1 value */ + xval += s->g.w[lxi]; + /* Else 0 value */ + + if (b->asegs != 0) { /* Tracking auxiliary segments */ + if (b->axisln >= b->axislz) { /* Need some more space in list */ + if (b->axislz == 0) { + b->axislz = 10; + if ((b->axisl = (axisec *)rev_malloc(s, b->axislz * sizeof(axisec))) == NULL) + error("rev: malloc failed - Auxiliary intersect list size %d",b->axislz); + INCSZ(b->s, b->axislz * sizeof(axisec)); + } else { + INCSZ(b->s, b->axislz * sizeof(axisec)); + b->axislz *= 2; + if ((b->axisl = (axisec *)rev_realloc(s, b->axisl, b->axislz * sizeof(axisec))) + == NULL) + error("rev: realloc failed - Auxiliary intersect list size %d",b->axislz); + } + } + b->axisl[b->axisln].xval = xval; + b->axisl[b->axisln].nv = x->sdi + 1; + for (f = 0; f <= x->sdi; f++) { + b->axisl[b->axisln].vix[f] = x->vix[f]; + } + b->axisln++; + } + +#ifdef DEBUG + if (xval >= b->min && xval <= b->max) + DBG(("auxil_locus: solution %f doesn't improve on min %f, max %f\n",xval,b->min,b->max)); +#endif + /* If this solution is expands the min or max, save it */ + if (xval < b->min) { + DBG(("######## Improving minimum to %f\n",xval)); + b->min = xval; + b->plmincell = x->ix; + } + if (xval > b->max) { + DBG(("######## Improving maximum to %f\n",xval)); + b->max = xval; + b->plmaxcell = x->ix; + } + } else { + DBG(("Solution wasn't within the simplex\n")); + return 0; + } + + return wsrv; +} + +/* - - - - - - - - - - - - - - - - - - - - - - - */ +/* Find the point on the clip line locus and simplexes */ +/* valid surface, that is closest to the target output value. */ +/* We expect to be given a sub simplex with sdi = fdi-1, and efdi = fdi */ +/* or a limit sub-simplex with sdi = fdi, and efdi = fdi+1 */ +/* Return zero if solution canot be calculated, */ +/* return 1 normally, 2 if solution would be above the (disabled) ink limit */ +static int +vnearest_clip_solve( +schbase *b, +simplex *x, +double *xp, /* Return solution (simplex parameter space) */ +double *xv, /* Return solution (output space) */ +double *err /* Output error distance at solution point */ +) { + rspl *s = b->s; + int e, sdi = x->sdi; + int f, fdi = s->fdi, efdi = x->efdi; + int g; + int wsrv; /* Within simplex return value */ + + double *ta[MXRO], TA[MXRO][MXRO]; + double tb[MXRO]; + + DBG(("Vector nearest clip solution called, cell %d, splx %d\n", x->ix, x->si)); + + /* Setup temporary matricies */ + for (f = 0; f < sdi; f++) { + ta[f] = TA[f]; + } + + /* Substitute simplex equation for output values V */ + /* in terms of sub-simplex parameters P, */ + /* into clip line implicit equation in V, to give */ + /* clip line simplex implicit equation in terms of P (simplex input space) */ + /* If this is a limit sub-simlex, the ink limit part of the clip vector */ + /* equations will be used. */ + + /* LHS: ta[sdi][sdi] = cla[sdi][efdi] * vv[efdi][sdi] */ + /* RHS: tb[sdi] = clb[sdi] - cla[sdi][efdi] * vv_di[efdi] */ + for (f = 0; f < sdi; f++) { + double tt; + for (e = 0; e < sdi; e++) { + for (tt = 0.0, g = 0; g < efdi; g++) + tt += b->cla[f][g] * (x->v[e][g] - x->v[e+1][g]); + ta[f][e] = tt; + } + for (tt = 0.0, g = 0; g < efdi; g++) + tt += b->cla[f][g] * x->v[sdi][g]; + tb[f] = b->clb[f] - tt; + } + + /* Compute the solution */ + if (gen_solve_se(ta, tb, sdi, sdi)) { + DBG(("Equation solution failed!\n")); + return 0; /* No solution */ + } + + /* Check that the solution is within the simplex */ + if ((wsrv = within_simplex(x, tb)) != 0) { + double dist; /* distance to clip target */ + + DBG(("Got solution within simplex %s\n", icmPdv(sdi,tb))); + + /* Compute the output space solution point */ + for (f = 0; f < fdi; f++) { + double tt = 0.0; + for (e = 0; e < sdi; e++) { + tt += (x->v[e][f] - x->v[e+1][f]) * tb[e]; + } + xv[f] = tt + x->v[sdi][f]; + } + + /* Copy to return array */ + for (e = 0; e < sdi; e++) + xp[e] = tb[e]; + + /* Compute distance to clip target */ + for (dist = 0.0, f = 0; f < fdi ; f++) { + double tt = (b->v[f] - xv[f]); + dist += tt * tt; + } + DBGV(("Vector clip output soln: ",fdi," %f", xv, "\n")); + + /* Return the solution in xp[]m xv[] and *err */ + *err = sqrt(dist); + + DBG(("Vector clip returning a solution with error %f\n",*err)); + return wsrv; + } + + DBG(("Vector clip solution not in simplex\n")); + return 0; /* No solution */ +} + +/* - - - - - - - - - - - - - - - - - - - - - - - */ +/* Find the point on the simplexes valid surface, that is closest */ +/* to the target output value. */ +/* We expect to be given a sub simplex with sdi = fdi-1, and efdi = fdi */ +/* or a limit sub-simplex with sdi = fdi, and efdi = fdi+1 */ +/* Return zero if solution canot be calculated, */ +/* return 1 normally, 2 if solution would be above the (disabled) ink limit */ +static int +nnearest_clip_solve( +schbase *b, +simplex *x, +double *xp, /* Return solution (simplex parameter space) */ +double *xv, /* Return solution (output space) */ +double *err /* Output error distance at solution point */ +) { + rspl *s = b->s; + int e, sdi = x->sdi; + int f, fdi = s->fdi, efdi = x->efdi; + double tb[MXRO]; /* RHS & Parameter solution */ + double dist; /* distance to clip target */ + int wsrv = 0; /* Within simplex return value */ + + DBG(("Nearest clip solution called, cell %d, splx %d\n", x->ix, x->si)); + + if (sdi == 0) { /* Solution is vertex */ + wsrv = 1; + for (f = 0; f < efdi; f++) + xv[f] = x->v[sdi][f]; /* Copy vertex value */ + if (x->v[sdi][fdi] > s->limitv) { + if (s->limiten) /* Needed when limiten == 0 */ + return 0; /* Over ink limit - no good */ + wsrv = 2; /* Would be over */ + } + DBG(("Got assumed vertex solution\n")); + } else { +#ifdef NEVER /* Don't specialise ink limit version - use INKSCALE fudge instead */ + if (!(x->flags & SPLX_CLIPSX)) { /* Not an ink limited plane simplex */ + +#endif + /* Create the SVD decomp needed for least squares solution */ + if (add_lu_svd(x)) { + DBG(("SVD decomp failed, skip simplex\n")); + return 0; + } + + /* Setup RHS to solve */ + for (f = 0; f < efdi; f++) + tb[f] = b->v[f] - x->v[sdi][f]; + + /* Find least squares solution */ + svdbacksub(x->d_u, x->d_w, x->d_v, tb, tb, efdi, sdi); + + /* Check that the solution is within the simplex */ + if ((wsrv = within_simplex(x, tb)) == 0) { + DBG(("Nearest clip solution not in simplex\n")); + return 0; /* No solution */ + } + + DBG(("Got solution within simplex %s\n",icmPdv(sdi,tb))); + + /* Compute the output space solution point */ + for (f = 0; f < fdi; f++) { + double tt = 0.0; + for (e = 0; e < sdi; e++) { + tt += (x->v[e][f] - x->v[e+1][f]) * tb[e]; + } + xv[f] = tt + x->v[sdi][f]; + } +#ifdef NEVER /* ~~1 Haven't figured out equations to make this a special case. */ + /* Content to use INKSCALE fudge and rely on SVD least squares. */ + } else { + /* We can't use the given equations, because we want the solution */ + /* to lie exactly on the ink limit plane, and be least squares to the */ + /* other target parameters. */ + /* Extract the ink limit parameters, and transform them into */ + /* a parameterised surface for this simplex. */ + /* Substitute the ink plane equation into the remaining target */ + /* parameter equations, and solve for least squares. */ + + } +#endif + } + + /* Copy to return array */ + for (e = 0; e < sdi; e++) + xp[e] = tb[e]; + + /* Compute distance to clip target */ + for (dist = 0.0, f = 0; f < fdi ; f++) { + double tt = (b->v[f] - xv[f]); + dist += tt * tt; + } + DBGV(("Nearest clip output soln: ",fdi," %f", xv, "\n")); + + /* Return the solution in xp[]m xv[] and *err */ + *err = sqrt(dist); + + DBG(("Nearest clip returning a solution with error %f\n",*err)); + return wsrv; +} + + +#ifdef NEVER +/* Utility to convert an implicit ink limit plane equation */ +/* (held at the end of the simplex output value equations), */ +/* into a parameterized surface equation. */ +static void +compute_param_limit_surface( +schbase *b, +simplex *x +) { + rspl *s = b->s; + int ff, f, fdi = s->fdi; + int i, p; + double lgst; + +double st[MXRO], /* Start point */ +double de[MXRO] /* Delta */ + DBG(("Computing clipping line implicit equation, dim = %d\n", fdi)); + + /* Pick a pivot element - the smallest */ + for (lgst = -1.0, p = -1, f = 0; f < fdi; f++) { + double tt = de[f]; + b->cdir[f] = tt; /* Stash this away */ + tt = fabs(tt); + if (tt > lgst) { + lgst = tt; + p = f; + } + } + if (p < 0) /* Shouldn't happen */ + error("rspl rev, internal, trying to cope with zero length clip line\n"); + + if (b->cla == NULL) + b->cla = dmatrix(0, fdi-1, 0, fdi); /* Allow for ink limit supliment */ + + for (i = ff = 0; ff < fdi; ff++) { /* For the input rows */ + if (ff == p) { + continue; /* Skip pivot row */ + } + for (f = 0; f < fdi; f++) { /* For input & output columns */ + if (f == p) { + b->cla[i][f] = -de[ff]; /* Last column is -ve delta value */ + } else if (f == ff) { + b->cla[i][f] = de[p]; /* Diagonal is pivot value */ + } else { + b->cla[i][f] = 0.0; /* Else zero */ + } + } + b->clb[i] = de[p] * st[ff] - de[ff] * st[p]; + i++; + } + + /* Add ink limit target equation - */ + /* interpolated ink value == target */ + if (s->limitf != NULL) { + for (i = 0; i < (fdi-1); i++) + b->cla[i][fdi] = 0.0; + + for (f = 0; f < fdi; f++) + b->cla[fdi-1][f] = 0.0; + + b->cla[fdi-1][fdi] = 1.0; + b->clb[fdi-1] = s->limitv; + } + +#ifdef NEVER +/* Verify that the implicit equation is correct */ +{ + double pnt[MXRO], v[MXRO]; + double pa; /* Parameter */ + for (pa = 0.0; pa <= 1.0; pa += 0.125) { + for (f = 0; f < fdi; f++) { + pnt[f] = st[f] + pa * de[f]; + } + + /* Verify the implicit equation */ + for (ff = 0; ff < (fdi-1); ff++) { + v[ff] = 0.0; + for (f = 0; f < fdi; f++) { + v[ff] += b->cla[ff][f] * pnt[f]; + } + v[ff] -= b->clb[ff]; + if (v[ff] < 0.0) + v[ff] = -v[ff]; + if (v[ff] > 0.000001) { + printf("Point on clip line = %f %f %f\n",pnt[0],pnt[1],pnt[2]); + printf("Implicit %d error of = %f\n",ff, v[ff]); + } + } + } +} +#endif /* NEVER */ + +} + +#endif + + + + +/* -------------------------------------------------------- */ +/* Cell/simplex object lower level code */ + +/* Utility to get or calculate a vertexes ink limit value */ +static double get_limitv( +schbase *b, /* Base search information */ +int ix, /* fwd index of cell */ +float *fcb, /* Pointer to base of vertex value array (ix is used if NULL) */ +double *p /* Array of input values (can be NULL to compute) */ +) { + rspl *s = b->s; + float *base = fcb; + double lv; + if (base == NULL) + base = s->g.a + ix * s->g.pss; + lv = base[-1]; /* Fetch existing ink limit function value */ + if ((float)lv == L_UNINIT) { /* Not been computed yet */ + if (p != NULL) { + lv = INKSCALE * s->limitf(s->lcntx, p); /* Do it */ + base[-1] = (float)lv; + } else { + int e, di = s->di; + double pp[MXRI]; /* Copy from float to double */ + int tix; /* Temp fwd cell index */ + + for (tix = ix, e = 0; e < di; e++) { + int dix; + dix = tix % s->g.res[e]; + tix /= s->g.res[e]; + pp[e] = s->g.l[e] + (double)dix * s->g.w[e]; /* Base point */ + } + lv = INKSCALE * s->limitf(s->lcntx, pp); /* Do it */ + base[-1] = (float)lv; + } + s->g.limitv_cached = 1; /* At least one limit value is cached */ + } + return lv; +} + +/* Utility to invalidate all the ink limit values */ +/* cached in the main rspl array */ +static void clear_limitv( +rspl *s +) { + int i; + float *gp; /* Grid point pointer */ + + if (s->g.limitv_cached != 0) { /* If any have been set */ + /* Unset them all */ + for (i = 0, gp = s->g.a; i < s->g.no; i++, gp += s->g.pss) { + gp[-1] = L_UNINIT; + } + s->g.limitv_cached = 0; + } +} + +/* Cell code */ + +static void free_cell_contents(cell *c); +static cell *cache_rcell(revcache *r, int ix, int force); +static void uncache_rcell(revcache *r, cell *cp); + +/* Return a pointer to an appropriate reverse cell */ +/* cache structure. None of the sub simplex lists will */ +/* be initialised. */ +/* NOTE: must unget_cell() (== uncache_rcell()) when cell */ +/* is no longer needed */ +/* Return NULL if we ran out of room in the cache. */ +static cell *get_rcell( +schbase *b, /* Base search information */ +int ix, /* fwd index of cell */ +int force /* if nz, force memory allocation, so that we have at least one cell */ +) { + rspl *s = b->s; + int ee, e, di = s->di; + int p2di = (1<fdi; + cell *c; + + c = cache_rcell(s->rev.cache, ix, force); /* Fetch it from the cache and lock it */ + if (c == NULL) + return NULL; + + if (!(c->flags & CELL_FLAG_1)) { /* Have to (re)initialize cell & simplexes */ + int tix; /* Temp fwd cell index */ + float *fcb = s->g.a + ix * s->g.pss; /* Pointer to base float of fwd cell */ + + /* Compute basic Cell info and vertex output values */ + for (ee = 0; ee < p2di; ee++) { + float *vp = fcb + s->g.fhi[ee]; + for (f = 0; f < fdi; f++) /* Transfer cell verticy values from grid */ + c->v[ee][f] = vp[f]; + + /* ~~ reset any other cell info that will be stale */ + } + + /* Convert from cell index, to absolute fwd coord base values */ + c->limmin = INF_DIST; /* and min/max values */ + c->limmax = -INF_DIST; + for (tix = ix, e = 0; e < di; e++) { + int dix; + dix = tix % s->g.res[e]; + tix /= s->g.res[e]; + c->p[0][e] = s->g.l[e] + (double)dix * s->g.w[e]; /* Base point */ + } + if (s->limitf != NULL) { /* Compute ink limit values at base verticy */ + double lv = get_limitv(b, ix, fcb, c->p[0]); /* Fetch or generate limit value */ + c->v[0][fdi] = lv; + if (lv < c->limmin) /* And min/max for this cell */ + c->limmin = lv; + if (lv > c->limmax) + c->limmax = lv; + } + + /* Setup cube verticy input position values, and ink limit values */ + for (ee = 1; ee < p2di; ee++) { + for (e = 0; e < di; e++) { + c->p[ee][e] = c->p[0][e]; + if (ee & (1 << e)) + c->p[ee][e] += s->g.w[e]; /* In input space offset */ + } + if (s->limitf != NULL) { /* Compute ink limit values at cell verticies */ + double lv = get_limitv(b, ix, fcb + s->g.fhi[ee], c->p[ee]); + c->v[ee][fdi] = lv; + if (lv < c->limmin) /* And min/max for this cell */ + c->limmin = lv; + if (lv > c->limmax) + c->limmax = lv; + } + } + + /* Compute the output bounding sphere for fast rejection testing */ + { + double *min[MXRO], *max[MXRO]; /* Pointers to points with min/max values */ + double radsq = -1.0; /* Span/radius squared */ + double rad; + int spf = 0; + + /* Find verticies of cell that have min and max values in output space */ + for (f = 0; f < fdi; f++) + min[f] = max[f] = NULL; + + for (ee = 0; ee < p2di; ee++) { + double *vp = c->v[ee]; + for (f = 0; f < fdi; f++) { + if (min[f] == NULL || min[f][f] > vp[f]) + min[f] = vp; + if (max[f] == NULL || max[f][f] < vp[f]) + max[f] = vp; + } + } + + /* Find the pair of points with the largest span (diameter) in output space */ + for (ff = 0; ff < fdi; ff++) { + double ss; + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt; + tt = max[ff][f] - min[ff][f]; + ss += tt * tt; + } + if (ss > radsq) { + radsq = ss; + spf = ff; /* Output dimension max was in */ + } + } + + /* Set initial bounding sphere */ + for (f = 0; f < fdi; f++) { + c->bcent[f] = (max[spf][f] + min[spf][f])/2.0; + } + radsq /= 4.0; /* diam^2 -> rad^2 */ + c->bradsq = radsq; + rad = c->brad = sqrt(radsq); + + /* Go though all the points again, expanding sphere if necessary */ + for (ee = 0; ee < p2di; ee++) { + double ss; + double *vp = c->v[ee]; + + /* Compute distance squared of point to bounding shere */ + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt = vp[f] - c->bcent[f]; + ss += tt * tt; + } + if (ss > radsq) { + double tt; + /* DBG(("Expanding bounding sphere by %f\n",sqrt(ss) - rad)); */ + + ss = sqrt(ss) + EPS; /* Radius to point */ + rad = (rad + ss)/2.0; + c->bradsq = radsq = rad * rad; + tt = ss - rad; + for (f = 0; f < fdi; f++) { + c->bcent[f] = (rad * c->bcent[f] + tt * vp[f])/ss; + } + + } else { + /* DBG(("Bounding sphere encloses by %f\n",rad - sqrt(ss))); */ + } + } + c->bradsq += EPS; + } + c->flags = CELL_FLAG_1; + } + + return c; +} + +void free_simplex_info(cell *c, int dof); + +/* Free up any allocated simplexes in a cell, */ +/* and set the pointers to NULL. */ +/* Nothing else is changed (ie. it's NOT removed from */ +/* the cache index or unthrheaded from the mru list). */ +static void +free_cell_contents( +cell *c +) { + int nsdi; + + /* Free up all the simplexes */ + if (c->s != NULL) { + for (nsdi = 0; nsdi <= c->s->di; nsdi++) { + if (c->sx[nsdi] != NULL) { + free_simplex_info(c, nsdi); + c->sx[nsdi] = NULL; + } + } + } + /* ~~ free any other cell information */ +} + +/* - - - - - - */ +/* Simplex code */ + +/* Simplex and Cell hash index size increments */ +int primes[] = { + 367, + 853, + 1489, + 3373, + 3373, + 6863, + 12919, + 23333, + 43721, + 97849, + 146221, + 254941, + -1 +}; + +/* Compute a simplex hash index */ +unsigned int simplex_hash(revcache *rc, int sdi, int efdi, int *vix) { + unsigned int hash = 0; + int i; + + for (i = 0; i <= sdi; i++) + hash = hash * 17 + vix[i]; + hash = hash * 17 + sdi; + hash = hash * 17 + efdi; + + hash %= rc->spx_hash_size; + return hash; +} + +/* Allocate and do the basic initialisation for a DOF list of simplexes */ +void alloc_simplexes( +cell *c, +int nsdi /* Non limited sub simplex dimensionality */ +) { + rspl *s = c->s; + schbase *b = s->rev.sb; + revcache *rc = s->rev.cache; + int ee, e, di = s->di; + int f, fdi = s->fdi; + int lsdi; /* Ink limited Sub-simplex sdi */ + int tsxno; /* Total number of DOF simplexes */ + int nsxno; /* Number of non-ink limited DOF simplexes */ + int si, so; /* simplex index in and out */ + + DBG(("Allocating level %d sub simplexes in cell %d\n",nsdi,c->ix)); + if (c->sx[nsdi] != NULL) + error("rspl rev, internal, trying allocate already allocated simplexes\n"); + + /* Figure out how many simplexes will be at this nsdi */ + lsdi = nsdi + 1; /* Ink limit simplexes sdi */ + + tsxno = nsxno = s->rev.sspxi[nsdi].nospx; + + if (s->limitf != NULL && lsdi <= di) + tsxno += s->rev.sspxi[lsdi].nospx; /* Second set with extra input dimension */ + + /* Make sure there is enough space in temp simplex filter list */ + if (b->lsxfilt < tsxno) { /* Allocate more space if needed */ + + if (b->lsxfilt > 0) { /* Free old space before allocating new */ + free(b->sxfilt); + DECSZ(b->s, b->lsxfilt * sizeof(char)); + } + b->lsxfilt = 0; + /* Allocate enough space for all the candidate cells */ + if ((b->sxfilt = (char *)rev_malloc(s, tsxno * sizeof(char))) == NULL) + error("rev: malloc failed - temp simplex filter list, count %d",tsxno); + b->lsxfilt = tsxno; /* Current allocated space */ + INCSZ(b->s, b->lsxfilt * sizeof(char)); + } + + /* Figure out the number of simplexes that will actually be needed */ + for (si = so = 0; si < tsxno; si++) { + psxinfo *psxi = NULL; + int *icomb, *offs; + int sdi = nsdi; + int efdi = fdi; + int ssi = si; + int isclip = 0; + if (si >= nsxno) { /* If limit boundary simplex */ + sdi++; /* One more dimension */ + efdi++; /* One more constraint */ + ssi -= nsxno; /* In second half of list */ + isclip++; /* Limit clipped simplex */ + } + psxi = &s->rev.sspxi[sdi].spxi[ssi]; + icomb = psxi->icomb; + offs = psxi->offs; + + b->sxfilt[si] = 0; /* Assume simplex won't be used */ + + /* Check if simplex should be discared due to the ink limit */ + if (s->limitf != NULL) { + double max = -INF_DIST; + double min = INF_DIST; + + /* Find the range of ink limit values covered by simplex */ + for (e = 0; e <= sdi; e++) { /* For all the simplex verticies */ + int i = offs[e]; + double vv = c->v[i][fdi]; /* Ink limit value */ + if (vv < min) + min = vv; + if (vv > max) + max = vv; + } + +//if ((max - min) > EPS) printf("~1 Found simplex sdi %d, efdi %d, min = %f, max = %f, limitv = %f\n", sdi, efdi, min,max,s->limitv); + if (isclip) { /* Limit clipped simplex */ + /* (Make sure it straddles the limit boundary) */ + if (max < s->limitv || min > s->limitv) + continue; /* Discard this simplex - it can't straddle the ink limit */ +//printf("~1 using sub simplex sdi %d, efdi %d, min = %f, max = %f, limitv = %f\n", sdi, efdi, min,max,s->limitv); + } else { + if (min > s->limitv) + continue; /* Discard this simplex - it is above the ink limit */ + } + } + + b->sxfilt[si] |= 1; /* This cell will be OK */ + so++; + } + + DBG(("There are %d level %d sub simplexes\n",so, nsdi)); + /* Allocate space for all the DOF simplexes that will be used */ + if (so > 0) { + if ((c->sx[nsdi] = (simplex **) rev_calloc(s, so, sizeof(simplex *))) == NULL) + error("rspl malloc failed - reverse cell simplexes - list of pointers"); + INCSZ(s, so * sizeof(simplex *)); + } + + /* Setup SPLX_FLAG_1 level information in the simplex */ + for (si = so = 0; si < tsxno; si++) { + simplex *x; + psxinfo *psxi = NULL; + int *icomb; + int sdi, efdi; + int ssi; + int vix[MXRI+1]; /* fwd cell vertex indexes of this simplex [sdi+1] */ + + if (b->sxfilt[si] == 0) /* Decided not to use this one */ + continue; + +#ifdef STATS + s->rev.st[b->op].sinited++; +#endif /* STATS */ + + sdi = nsdi; + efdi = fdi; + ssi = si; + if (si >= nsxno) { /* If limit boundary simplex */ + sdi++; /* One more dimension */ + efdi++; /* One more constraint */ + ssi -= nsxno; /* In second half of list */ + } + + psxi = &s->rev.sspxi[sdi].spxi[ssi]; + icomb = psxi->icomb; + + /* Compute simplex vertexes so we can match it in the cache */ + for (e = 0; e <= sdi; e++) + vix[e] = c->ix + s->g.hi[psxi->offs[e]]; + + x = c->sx[nsdi][so]; + + /* If this is a shared simplex, see if we already have it in another cell */ + if (x == NULL && psxi->face) { + unsigned int hash; +//printf("~1 looking for existing simplex nsdi = %d\n",nsdi); + hash = simplex_hash(rc, sdi, efdi, vix); + for (x = rc->spxhashtop[hash]; x != NULL; x = x->hlink) { + if (x->sdi != sdi + || x->efdi != efdi) + continue; /* miss */ + for (e = 0; e <= sdi; e++) { + if (x->vix[e] != vix[e]) + break; /* miss */ + } + if (e > sdi) + break; /* hit */ + } + if (x != NULL) { + x->refcount++; +//printf("~1 found hit in simplex face list hash %d, refcount = %d\n",hash,x->refcount); + } + } + /* Doesn't already exist */ + if (x == NULL) { + if ((x = (simplex *) rev_calloc(s, 1, sizeof(simplex))) == NULL) + error("rspl malloc failed - reverse cell simplexes - base simplex %d bytes",sizeof(simplex)); + INCSZ(s, sizeof(simplex)); + x->refcount = 1; + x->touch = s->rev.stouch-1; + x->flags = 0; + + if (si >= nsxno) { /* If limit boundary simplex */ + x->flags |= SPLX_CLIPSX; /* Limit clipped simplex */ + } + + /* Fill in the other simplex details */ + x->s = s; /* Parent rspl */ + x->ix = c->ix; /* Construction cube base index */ + for (e = 0; e <= sdi; e++) /* Indexs of fwd verticies that make up this simplex */ + x->vix[e] = vix[e]; + x->psxi = psxi; /* Pointer to constant per simplex info */ +//printf("~1 set simplex 0x%x psxi = 0x%x\n",x,x->psxi); + x->si = so; /* Diagnostic, simplex offset in list */ + x->sdi = sdi; /* Copy of simplex dimensionaity */ + x->efdi = efdi; /* Copy of effective output dimensionality */ + + /* Copy cell simplex vertex output and limit values */ + for (e = 0; e <= sdi; e++) { /* For all the simplex verticies */ + int i = x->psxi->offs[e]; + + for (f = 0; f <= fdi; f++) /* Copy vertex value + ink sum */ + x->v[e][f] = c->v[i][f]; + + /* Setup output bounding box values (the hard way) */ + if (e == 0) { /* Init to first vertex of simplex */ + for (f = 0; f <= fdi; f++) /* Output space */ + x->min[f] = x->max[f] = c->v[i][f]; + } else { + for (f = 0; f <= fdi; f++) { /* Output space + ink sum */ + double vv; +// if (f == fdi && s->limit == NULL) +// continue; /* Skip ink */ + vv = c->v[i][f]; + if (vv < x->min[f]) + x->min[f] = vv; + else if (vv > x->max[f]) + x->max[f] = vv; + } + } + } + /* Add a margin */ + for (f = 0; f <= fdi; f++) { /* Output space + ink sum */ + x->min[f] -= EPS; + x->max[f] += EPS; + } + + /* Setup input bounding box value pointers (the easy way) */ + for (ee = 0; ee < di; ee++) { + x->p0[ee] = c->p[0][ee]; /* Construction base cube origin */ + x->pmin[ee] = c->p[x->psxi->pmino[ee]][ee] - EPS; + x->pmax[ee] = c->p[x->psxi->pmaxo[ee]][ee] + EPS; + } + + x->flags |= SPLX_FLAG_1; /* vv & iv done, nothing else */ + + x->aloc2 = x->aloc5 = NULL; /* Matrix allocations not done yet */ + + /* Add it to the face shared simplex hash index */ + if (x->psxi->face) { + unsigned int hash; + int i; + /* See if we should re-size the simplex hash index */ + if (++rc->nspx > (HASH_FILL_RATIO * rc->spx_hash_size)) { + for (i = 0; primes[i] > 0 && primes[i] <= rc->spx_hash_size; i++) + ; + if (primes[i] > 0) { + int spx_hash_size = rc->spx_hash_size; /* Old */ + simplex **spxhashtop = rc->spxhashtop; + + rc->spx_hash_size = primes[i]; + + DBG(("Increasing face simplex hash index to %d\n",spx_hash_size)); +//printf("~1 increasing simplex hash index size to %d\n",spx_hash_size); + /* Allocate a new index */ + if ((rc->spxhashtop = (simplex **) rev_calloc(s, rc->spx_hash_size, + sizeof(simplex *))) == NULL) + error("rspl malloc failed - reverse simplex cache index"); + INCSZ(s, rc->spx_hash_size * sizeof(simplex *)); + + /* Transfer all the simplexes to the new index */ + for (i = 0; i < spx_hash_size; i++) { + simplex *x, *nx; + for (x = spxhashtop[i]; x != NULL; x = nx) { + nx = x->hlink; + hash = simplex_hash(rc, x->sdi, x->efdi, x->vix); /* New hash */ + x->hlink = rc->spxhashtop[hash]; /* Add to new hash index */ + rc->spxhashtop[hash] = x; + } + } + free(spxhashtop); /* Done with old index */ + DECSZ(s, spx_hash_size * sizeof(simplex *)); + } + } + hash = simplex_hash(rc, sdi, efdi, vix); + + /* Add this to hash index */ + x->hlink = rc->spxhashtop[hash]; + rc->spxhashtop[hash] = x; +//printf("~1 Added simplex to hash %d, rc->nspx = %d\n",hash,rc->nspx); + } + +//if (rc->nunlocked == 0 && rc->s->rev.sz > rc->s->rev.max_sz) +//printf("~1 unable to decrease_revcache 1\n"); + + /* keep memory in check */ + while (rc->nunlocked > 0 && rc->s->rev.sz > rc->s->rev.max_sz) { + if (decrease_revcache(rc) == 0) + break; + } + } + c->sx[nsdi][so] = x; + so++; + } + c->sxno[nsdi] = so; /* Record actual number in list */ + c->flags |= CELL_FLAG_2; /* Note that cell now has simplexes */ +} + +/* Free up any allocated for a list of sub-simplexes */ +void +free_simplex_info( +cell *c, +int nsdi /* non limit sub simplex dimensionaity */ +) { + int si, sxno = c->sxno[nsdi]; /* Number of simplexes */ + + for (si = 0; si < sxno; si++) { /* For all the simplexes */ + simplex *x = c->sx[nsdi][si]; + int dof = x->sdi - x->efdi; + +//printf("~1 freeing simplex, refcount = %d\n",x->refcount); + if (--x->refcount <= 0) { /* Last reference to this simplex */ + +//printf("~1 freeing simplex 0x%x psxi = 0x%x\n",x,x->psxi); + if (x->psxi->face) { + unsigned int hash; + revcache *rc = c->s->rev.cache; + + hash = simplex_hash(rc, x->sdi, x->efdi, x->vix); + + /* Free it from the hash list */ + if (rc->spxhashtop[hash] == x) { + rc->spxhashtop[hash] = x->hlink; + rc->nspx--; +//printf("~1 removed simplex from hash %d, nspx now = %d\n",hash,rc->nspx); + } else { + simplex *xx; + for (xx = rc->spxhashtop[hash]; xx != NULL && xx->hlink != x; xx = xx->hlink) + ; + if (xx != NULL) { /* Found it */ + xx->hlink = x->hlink; + rc->nspx--; +//printf("~1 removed simplex from hash %d, nspx now = %d\n",hash,rc->nspx); + } +//else +//printf("~1 warning, failed to find face simplex hash %d, sdi = %d in cache index (nspx = %d)!!\n",hash,x->sdi,rc->nspx); + } + } + if (x->aloc2 != NULL) { + int adof = dof >= 0 ? dof : 0; /* Allocation dof */ + int asize; + if (dof == 0) + asize = sizeof(double) * (x->efdi * x->sdi) + + sizeof(double *) * x->efdi + + sizeof(int) * x->sdi; + else + asize = sizeof(double) * (x->sdi * (x->efdi + x->sdi + adof + 2) + x->efdi) + + sizeof(double *) * (x->efdi + 2 * x->sdi); + free(x->aloc2); + DECSZ(x->s, asize); + } + + if (x->aloc5 != NULL) { + int asize; + if (x->naux == dof) + asize = sizeof(double *) * x->naux + + sizeof(double) * (x->naux * dof) + + sizeof(int) * dof; + else + asize = sizeof(double *) * (x->naux + dof) + + sizeof(double) * (dof * (x->naux + dof + 1)); + free(x->aloc5); + DECSZ(x->s, asize); + } + + /* ~~ free any other simplex information */ + + free(x); + DECSZ(c->s, sizeof(simplex)); + c->sx[nsdi][si] = NULL; + } + } + free(c->sx[nsdi]); + DECSZ(c->s, c->sxno[nsdi] * sizeof(simplex *)); + c->sx[nsdi] = NULL; + c->sxno[nsdi] = 0; + + /* ~~ free any other cell information */ +} + +/* - - - - - - - - - - - - */ +/* Check that an input space vector is within a given simplex, */ +/* and that it meets any ink limit. */ +/* Return zero if outside the simplex, */ +/* 1 normally if within the simplex, */ +/* and 2 if it would be over the ink limit if limit was enabled. */ +static int +within_simplex( +simplex *x, /* Simplex */ +double *p /* Input coords in simplex space */ +) { + rspl *s = x->s; + schbase *b = s->rev.sb; + int fdi = s->fdi; + int e, sdi = x->sdi; /* simplex dimensionality */ + double cp, lp; + int rv = 1; + /* EPS is allowance for numeric error */ + /* (Don't want solutions falling down */ + /* the numerical cracks between the simplexes) */ + + /* Check we are within baricentric limits */ + for (lp = 0.0, e = 0; e < sdi; e++) { + cp = p[e]; + if ((cp+EPS) < lp) /* Outside baricentric or not in correct */ + return 0; /* order for this simplex */ + lp = cp; + } + if ((1.0+EPS) < lp) /* outside baricentric range */ + return 0; + + /* Compute limit using interp. - assume simplex would have been trivially rejected */ + if (s->limitf != NULL) { + double sum = 0.0; /* Might be over the limit */ + for (e = 0; e < sdi; e++) + sum += p[e] * (x->v[e][fdi] - x->v[e+1][fdi]); + sum += x->v[sdi][fdi]; + if (sum > s->limitv) { + if (s->limiten != 0) + return 0; /* Exceeds ink limit */ + else + rv = 2; /* would have exceeded limit */ + } + } + +#ifdef NEVER + /* Constrain to legal values */ + /* (Is this needed ?????) */ + for (e = 0; e < sdi; e++) { + cp = p[e]; + if (cp < 0.0) + p[e] = 0.0; + else if (cp > 1.0) + p[e] = 1.0; + } +#endif + return rv; +} + +/* Convert vector from simplex space to absolute cartesian space */ +static void simplex_to_abs( +simplex *x, +double *out, /* output in absolute space */ +double *in /* Input in simplex space */ +) { + rspl *s = x->s; + int e, di = s->di; + int *icomb = x->psxi->icomb; /* Coord combination order */ + + for (e = 0; e < di; e++) { /* For each absolute coord */ + double ov = x->p0[e]; /* Base value */ + int ee = icomb[e]; /* Simplex param index */ + if (ee >= 0) /* Simplex param value */ + ov += s->g.w[e] * in[ee]; + else if (ee == -2) /* 1 value */ + ov += s->g.w[e]; + /* Else 0 value */ + out[e] = ov; + } +} + +/* - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ +/* Given the parametric clip line equation, compute the */ +/* implicit equation in terms of the absolute output space. */ +/* Pad equation with target ink limit in case it is use */ +/* with CLIPSX sub-simplexes. */ +/* Note that no line equation values are returned if fdi = 1, */ +/* since there is no such thing as an implicit line equation. */ +static void +init_line_eq( +schbase *b, +double st[MXRO], /* Start point */ +double de[MXRO] /* Delta */ +) { + rspl *s = b->s; + int ff, f, fdi = s->fdi; + int i, p; + double lgst; + + DBG(("Computing clipping line implicit equation, dim = %d\n", fdi)); + + /* Pick a pivot element */ + for (lgst = -1.0, p = -1, f = 0; f < fdi; f++) { + double tt = de[f]; + b->cdir[f] = tt; /* Stash this away */ + tt = fabs(tt); + if (tt > lgst) { + lgst = tt; + p = f; + } + } + if (p < 0) /* Shouldn't happen */ + error("rspl rev, internal, trying to cope with zero length clip line\n"); + + if (b->cla == NULL) + b->cla = dmatrix(0, fdi-1, 0, fdi); /* Allow for ink limit supliment */ + + for (i = ff = 0; ff < fdi; ff++) { /* For the input rows */ + if (ff == p) { + continue; /* Skip pivot row */ + } + for (f = 0; f < fdi; f++) { /* For input & output columns */ + if (f == p) { + b->cla[i][f] = -de[ff]; /* Last column is -ve delta value */ + } else if (f == ff) { + b->cla[i][f] = de[p]; /* Diagonal is pivot value */ + } else { + b->cla[i][f] = 0.0; /* Else zero */ + } + } + b->clb[i] = de[p] * st[ff] - de[ff] * st[p]; + i++; + } + + /* Add ink limit target equation - */ + /* interpolated ink value == target */ + if (s->limitf != NULL) { + for (i = 0; i < (fdi-1); i++) + b->cla[i][fdi] = 0.0; + + for (f = 0; f < fdi; f++) + b->cla[fdi-1][f] = 0.0; + + b->cla[fdi-1][fdi] = 1.0; + b->clb[fdi-1] = s->limitv; + } + +#ifdef NEVER +/* Verify that the implicit equation is correct */ +{ + double pnt[MXRO], v[MXRO]; + double pa; /* Parameter */ + for (pa = 0.0; pa <= 1.0; pa += 0.125) { + for (f = 0; f < fdi; f++) { + pnt[f] = st[f] + pa * de[f]; + } + + /* Verify the implicit equation */ + for (ff = 0; ff < (fdi-1); ff++) { + v[ff] = 0.0; + for (f = 0; f < fdi; f++) { + v[ff] += b->cla[ff][f] * pnt[f]; + } + v[ff] -= b->clb[ff]; + if (v[ff] < 0.0) + v[ff] = -v[ff]; + if (v[ff] > 0.000001) { + printf("Point on clip line = %f %f %f\n",pnt[0],pnt[1],pnt[2]); + printf("Implicit %d error of = %f\n",ff, v[ff]); + } + } + } +} +#endif /* NEVER */ + +} + +/* - - - - - - */ +/* Simpex solution info #2 */ + +/* Create the LU or SVD decomp needed to compute solution or locus. */ +/* Return non-zero if it cannot be created */ +static int +add_lu_svd(simplex *x) { + + if (x->flags & SPLX_FLAG_2F) { /* Previously failed */ + return 1; + } + if (!(x->flags & SPLX_FLAG_2)) { + int ee, e, sdi = x->sdi; + int f, efdi = x->efdi; + int dof = sdi-efdi; /* Degree of freedom of locus, or -ve over specification */ + int adof = dof >= 0 ? dof : 0; /* Allocation dof */ + int i; + + if (x->aloc2 == NULL) { /* Allocate space for matricies and arrays */ + /* Do this in one hit to minimise malloc overhead */ + if (dof == 0) { + int i; + char *mem; + int asize = sizeof(double) * (efdi * sdi) + + sizeof(double *) * efdi + + sizeof(int) * sdi; + + if ((x->aloc2 = mem = (char *) rev_malloc(x->s, asize)) == NULL) + error("rspl malloc failed - reverse cell sub-simplex matricies"); + INCSZ(x->s, asize); + + /* Allocate biggest to smallest (double, pointers, ints) */ + /* to make sure that items lie on the natural boundaries. */ + + /* Reserve matrix doubles */ + mem += efdi * sdi * sizeof(double); + + /* Allocate pointers */ + x->d_u = (double **)mem, mem += efdi * sizeof(double *); + + /* Allocate ints */ + x->d_w = (double *)mem, mem += sdi * sizeof(int); + +#ifdef DEBUG + if (mem != (x->aloc2 + asize)) + error("~1 aloc2a assert failed! Is %d, should be %d\n",mem - x->aloc2,asize); +#endif /* DEBUG */ + + /* Reset and allocate matrix doubles */ + mem = x->aloc2; + for (i = 0; i < efdi; i++) + x->d_u[i] = (double *)mem, mem += sdi * sizeof(double); + + } else { + int i; + char *mem; + int asize = sizeof(double) * (sdi * (efdi + sdi + adof + 2) + efdi) + + sizeof(double *) * (efdi + 2 * sdi); + + if ((x->aloc2 = mem = (char *) rev_malloc(x->s, asize)) == NULL) + error("rspl malloc failed - reverse cell sub-simplex matricies"); + INCSZ(x->s, asize); + + /* Allocate biggest to smallest (double, pointers, ints) */ + /* to make sure that items lie on the natural boundaries. */ + + /* Reserve matrix doubles */ + mem += sdi * (efdi + sdi + adof) * sizeof(double); + + /* Allocate doubles */ + x->lo_xb = (double *)mem, mem += efdi * sizeof(double); + x->lo_bd = (double *)mem; mem += sdi * sizeof(double); + x->d_w = (double *)mem, mem += sdi * sizeof(double); + + /* Allocate pointers */ + x->d_u = (double **)mem, mem += efdi * sizeof(double *); + x->d_v = (double **)mem, mem += sdi * sizeof(double *); + x->lo_l = (double **)mem, mem += sdi * sizeof(double *); + +#ifdef DEBUG + if (mem != (x->aloc2 + asize)) + error("~1 aloc2b assert failed! Is %d, should be %d\n",mem - x->aloc2,asize); +#endif /* DEBUG */ + + /* Reset and allocate matrix doubles */ + mem = x->aloc2; + for (i = 0; i < efdi; i++) + x->d_u[i] = (double *)mem, mem += sdi * sizeof(double); + for (i = 0; i < sdi; i++) + x->d_v[i] = (double *)mem, mem += sdi * sizeof(double); + for (i = 0; i < sdi; i++) + x->lo_l[i] = (double *)mem, mem += adof * sizeof(double); + + /* Init any values that will be read before being written to. */ + for (f = 0; f < efdi; f++) + x->lo_xb[f] = 1e100; /* Silly value */ + } + } + + /* Setup matrix from vertex values */ + for (f = 0; f < efdi; f++) + for (e = 0; e < sdi; e++) + x->d_u[f][e] = x->v[e][f] - x->v[e+1][f]; + + if (dof == 0) { /* compute LU */ + double rip; +#ifdef STATS + x->s->rev.st[x->s->rev.sb->op].sinited2a++; +#endif /* STATS */ + if (lu_decomp(x->d_u, sdi, (int *)x->d_w, &rip)) { + x->flags |= SPLX_FLAG_2F; /* Failed */ + return 1; + } + } else { +//printf("~~ Creating SVD decomp, sdi = %d, efdi = %d\n", sdi, efdi); + +#ifdef STATS + x->s->rev.st[x->s->rev.sb->op].sinited2b++; +#endif /* STATS */ + if (svdecomp(x->d_u, x->d_w, x->d_v, efdi, sdi)) { + x->flags |= SPLX_FLAG_2F; /* Failed */ + return 1; + } + + /* Threshold the singular values W[] */ + svdthresh(x->d_w, sdi); + + if (dof >= 0) { /* If we expect a locus */ +//printf("~~ got dif %d locus from SVD\n",dof); + /* copy the locus direction coefficients out */ + for (i = e = 0; e < sdi; e++) { + if (x->d_w[e] == 0.0) { /* Found a zero W[] */ + if (i < dof) { + for (ee = 0; ee < sdi; ee++) { /* Copy column of V[][] */ + x->lo_l[ee][i] = x->d_v[ee][e]; + } + } + i++; + } + } + if (i != dof) { +//printf("~~ got unexpected dof in svd\n"); + x->flags |= SPLX_FLAG_2F; /* Failed */ + return 1; /* Didn't get expected d.o.f. */ + } + } + } + x->flags |= SPLX_FLAG_2; /* Set flag so that it isn't attempted again */ + +//if (x->s->rev.cache->nunlocked == 0 && x->s->rev.sz > x->s->rev.max_sz) +//printf("~1 unable to decrease_revcache 2\n"); + + /* keep memory in check */ + while (x->s->rev.cache->nunlocked > 0 && x->s->rev.sz > x->s->rev.max_sz) { + if (decrease_revcache(x->s->rev.cache) == 0) + break; + } + } + return 0; +} + +/* - - - - - - */ +/* Simplex solution info #4 */ + +/* Calculate the solution locus equation for this simplex and target */ +/* (The direction was calculated by add_svd(), but now calculate */ +/* the base solution point for this particular reverse lookup) */ +/* Return non-zero if this point canot be calculated */ +/* We are assuming that sdi > efdi */ +static int +add_locus( +schbase *b, +simplex *x +) { + int sdi = x->sdi; + int f, efdi = x->efdi; + int doback = 0; + +#ifdef STATS + x->s->rev.st[x->s->rev.sb->op].sinited4++; +#endif /* STATS */ + /* Use output of svdcmp() to solve overspecified and/or */ + /* singular equation A.x = b */ + + /* Init the RHS B[] vector, and check if it doesn't match */ + /* that used to compute base value last time. */ + for (f = 0; f < efdi; f++) { + double xb = b->v[f] - x->v[sdi][f]; + if (x->lo_xb[f] != xb) { + x->lo_xb[f] = xb; + doback = 1; /* RHS differs, so re-compute */ + } + } + +#ifdef STATS + if (doback && (x->flags & SPLX_FLAG_4)) + x->s->rev.st[x->s->rev.sb->op].sinited4i++; +#endif /* STATS */ + + /* Compute locus */ + if (doback || !(x->flags & SPLX_FLAG_4)) + svdbacksub(x->d_u, x->d_w, x->d_v, x->lo_xb, x->lo_bd, efdi, sdi); + + x->flags |= SPLX_FLAG_4; + +//if (x->s->rev.cache->nunlocked == 0 && x->s->rev.sz > x->s->rev.max_sz) +//printf("~1 unable to decrease_revcache 3\n"); + + /* keep memory in check */ + while (x->s->rev.cache->nunlocked > 0 && x->s->rev.sz > x->s->rev.max_sz) { + if (decrease_revcache(x->s->rev.cache) == 0) + break; + } + + return 0; +} + +/* - - - - - - */ +/* Simplex solution info #5 */ + +/* Compute LU or SVD decomp of lo_l */ +/* Allocates the memory for the various matricies */ +/* Return non-zero if this canot be calculated. */ +static int +add_auxil_lu_svd( +schbase *b, +simplex *x +) { + int ee, sdi = x->sdi; + int f, efdi = x->efdi; + int dof = sdi-efdi; /* Degree of freedom of locus */ + int naux = b->naux; /* Number of auxiliaries actually available */ + +#ifdef STATS + if (x->aaux != b->naux || x->auxbm != b->auxbm) + x->s->rev.st[x->s->rev.sb->op].sinited5i++; +#endif /* STATS */ + + if (x->aaux != b->naux) { /* Number of auxiliaries has changed */ + if (x->aloc5 != NULL) { + int asize; + if (x->naux == dof) + asize = sizeof(double *) * x->naux + + sizeof(double) * (x->naux * dof) + + sizeof(int) * dof; + else + asize = sizeof(double *) * (x->naux + dof) + + sizeof(double) * (dof * (x->naux + dof + 1)); + free(x->aloc5); + x->aloc5 = NULL; + DECSZ(x->s, asize); + } + x->flags &= ~(SPLX_FLAG_5 | SPLX_FLAG_5F); /* Force recompute */ + } + + if (x->auxbm != b->auxbm) { /* Different selection of auxiliaries */ + x->flags &= ~(SPLX_FLAG_5 | SPLX_FLAG_5F); /* Force recompute */ + } + + if (x->flags & SPLX_FLAG_5F) { /* Previously failed */ + return 1; + } + if (!(x->flags & SPLX_FLAG_5)) { + int *icomb = x->psxi->icomb; /* abs -> simplex coordinate translation */ + + if (x->aloc5 == NULL) { /* Allocate space for matricies and arrays */ + /* Do this in one hit to minimise malloc overhead */ + if (naux == dof) { + int i; + char *mem; + int asize = sizeof(double *) * naux + + sizeof(double) * (naux * dof) + + sizeof(int) * dof; + + if ((x->aloc5 = mem = (char *) rev_malloc(x->s, asize)) == NULL) + error("rspl malloc failed - reverse cell sub-simplex matricies"); + INCSZ(x->s, asize); + + /* Allocate biggest to smallest (double, pointers, ints) */ + /* to make sure that items lie on the natural boundaries. */ + + /* Reserve matrix doubles */ + mem += naux * dof * sizeof(double); + + /* Allocate pointers and ints */ + x->d_u = (double **)mem, mem += naux * sizeof(double *); + x->d_w = (double *)mem, mem += dof * sizeof(int); + +#ifdef DEBUG + if (mem != (x->aloc5 + asize)) + error("aloc5a assert failed! Is %d, should be %d\n",mem - x->aloc5,asize); +#endif /* DEBUG */ + + /* Reset and allocate matrix doubles */ + mem = x->aloc5; + for (i = 0; i < naux; i++) + x->d_u[i] = (double *)mem, mem += dof * sizeof(double); + } else { + int i; + char *mem; + int asize = sizeof(double *) * (naux + dof) + + sizeof(double) * (dof * (naux + dof + 1)); + + if ((x->aloc5 = mem = (char *) rev_malloc(x->s, asize)) == NULL) + error("rspl malloc failed - reverse cell sub-simplex matricies"); + INCSZ(x->s, asize); + + /* Allocate biggest to smallest (double, pointers, ints) */ + /* to make sure that items lie on the natural boundaries. */ + + /* Reserve matrix doubles */ + mem += dof * (naux + dof) * sizeof(double); + + /* Allocate doubles */ + x->ax_w = (double *)mem, mem += dof * sizeof(double); + + /* Allocate pointers, ints */ + x->ax_u = (double **)mem, mem += naux * sizeof(double *); + x->ax_v = (double **)mem, mem += dof * sizeof(double *); + +#ifdef DEBUG + if (mem != (x->aloc5 + asize)) + error("aloc5b assert failed! Is %d, should be %d\n",mem - x->aloc5,asize); +#endif /* DEBUG */ + + /* Reset and allocate matrix doubles */ + mem = x->aloc5; + for (i = 0; i < naux; i++) + x->ax_u[i] = (double *)mem, mem += dof * sizeof(double); + for (i = 0; i < dof; i++) + x->ax_v[i] = (double *)mem, mem += dof * sizeof(double); + } + x->aaux = naux; /* Number of auxiliaries allocated for */ + } + + /* Setup A[][] matrix to decompose, and figure number of auxiliaries actually needed */ + for (ee = naux = 0; ee < b->naux; ee++) { + int ei = icomb[b->auxi[ee]]; /* Simplex relative auxiliary index */ + if (ei < 0) + continue; /* aux corresponds with fixed input value for this simplex */ + for (f = 0; f < dof; f++) + x->ax_u[naux][f] = x->lo_l[ei][f]; + naux++; + } + x->naux = naux; /* Number of auxiliaries actually available */ + x->auxbm = b->auxbm; /* Mask of auxiliaries used */ + + if (naux == dof) { /* Use LU decomp to solve exactly */ + double rip; + +#ifdef STATS + x->s->rev.st[x->s->rev.sb->op].sinited5a++; +#endif /* STATS */ + if (lu_decomp(x->ax_u, dof, (int *)x->ax_w, &rip)) { + x->flags |= SPLX_FLAG_5F; + return 1; + } + + } else if (naux > 0) { /* Use SVD to solve least squares */ + +#ifdef STATS + x->s->rev.st[x->s->rev.sb->op].sinited5b++; +#endif /* STATS */ + if (svdecomp(x->ax_u, x->ax_w, x->ax_v, naux, dof)) { + x->flags |= SPLX_FLAG_5F; + return 1; + } + + /* Threshold the singular values W[] */ + svdthresh(x->ax_w, dof); + } /* else naux == 0, don't setup anything */ + + x->flags |= SPLX_FLAG_5; + +//if (x->s->rev.cache->nunlocked == 0 && x->s->rev.sz > x->s->rev.max_sz) +//printf("~1 unable to decrease_revcache 4\n"); + + /* keep memory in check */ + while (x->s->rev.cache->nunlocked > 0 && x->s->rev.sz > x->s->rev.max_sz) { + if (decrease_revcache(x->s->rev.cache) == 0) + break; + } + } + return 0; +} + +/* - - - - - - - - - - - - - - - - - - - - - - - - - - */ + +/* Initialise a static sub-simplex verticy information table */ +void +rspl_init_ssimplex_info( +rspl *s, +ssxinfo *xip, /* Pointer to sub-simplex info structure to init. */ +int sdi /* Sub-simplex dimensionality (range 0 - di) */ +) { + int e, di = s->di; /* Dimensionality */ + int vi, nospx; /* Number of sub-simplexes */ + XCOMBO(vcmb, MXDI, sdi+1, 1 << di);/* Simplex dimension sdi out of cube dimension di counter */ + + DBG(("init_ssimplex_info called with sdi = %d\n",sdi)); + /* First count the number of sub-simplexes */ + nospx = 0; + XCB_INIT(vcmb); + while (!XCB_DONE(vcmb)) { + nospx++; + XCB_INC(vcmb); + } + + xip->sdi = sdi; + xip->nospx = nospx; + if ((xip->spxi = (psxinfo *) rev_calloc(s, nospx, sizeof(psxinfo))) == NULL) + error("rspl malloc failed - reverse cell sub-simplex info array"); + INCSZ(s, nospx * sizeof(psxinfo)); + + DBG(("Number of subsimplex = %d\n",nospx)); + /* For all sub-simplexes */ + XCB_INIT(vcmb); + for (vi = 0; vi < nospx; vi++) { + psxinfo *x = &xip->spxi[vi]; + int i; + int andm, orm; + + /* XCOMB generates verticies in order from max to min offset */ + + /* Compute Absolute -> Parameter mapping */ + for (e = 0; e < di; e++) { /* For each absolute axis */ + + if ((vcmb[sdi] & (1<icomb[e] = -2; /* This abs is always '1' */ + + } else if ((vcmb[0] & (1<icomb[e] = -1; /* This abs is always '0' */ + + } else { + for (i = 0; i < sdi; i++) { /* For each verticy in large to small order (!first) */ + if ((vcmb[i] & (1<icomb[e] = i; /* This is parameter */ + break; + } + } + } + } + + /* Compute fwd grid offsets for each simplex vertex in baricentric order */ + for (i = 0; i <= sdi; i++) { /* For each verticy */ + int pmin[MXRI], pmax[MXRI]; + x->offs[i] = vcmb[i]; + x->goffs[i] = s->g.hi[vcmb[i]]; + x->foffs[i] = s->g.fhi[vcmb[i]]; + + /* Setup input coordinate bounding box value offsets */ + if (i == 0) { /* Init to first vertex of simplex */ + for (e = 0; e < di; e++) { /* Input space */ + x->pmino[e] = x->pmaxo[e] = vcmb[i]; + pmin[e] = pmax[e] = vcmb[i] & (1<pmino[e] = vcmb[i]; + pmin[e] = vv; + } else if (vv > pmax[e]) { + x->pmaxo[e] = vcmb[i]; + pmax[e] = vv; + } + } + } + } + + /* See if the sub-simplex lies on a cube face */ + andm = ~0; + orm = 0; + for (i = 0; i <= sdi; i++) { /* For each verticy */ + andm &= vcmb[i]; + orm |= vcmb[i]; + } + /* If one coordinate is common (all 0 or all 1) to the verticies, */ + /* they must all be on the same cube face. */ + if (andm != 0 || orm != ((1 << di)-1)) + x->face = 1; + else + x->face = 0; + +#ifdef DEBUG + printf("Verticies = "); + for (i = 0; i <= sdi; i++) + printf("%d ",vcmb[i]); + printf("\n"); + + printf("Face = %s\n",x->face ? "True" : "False"); + + printf("Abs -> Parm = "); + for (e = 0; e < di; e++) + printf("%d ",x->icomb[e]); + printf("\n"); + + printf("Grid Offset = "); + for (e = 0; e <= sdi; e++) + printf("%d ",x->goffs[e]); + printf("Float Offset = "); + for (e = 0; e <= sdi; e++) + printf("%d ",x->foffs[e]); + printf("\n"); + printf("\n"); +#endif /* DEBUG */ + + /* Increment the counter value */ + XCB_INC(vcmb); + } +} + +/* Free the given sub-simplex verticy information */ +void +rspl_free_ssimplex_info( +rspl *s, +ssxinfo *xip /* Pointer to sub-simplex info structure */ +) { + if (xip == NULL) /* Assert */ + return; + + free(xip->spxi); + DECSZ(s, xip->nospx * sizeof(psxinfo)); + xip->spxi = NULL; +} + +/* ====================================================== */ +/* Reverse cell cache code */ + +/* Allocate and initialise the reverse cell cache */ +static revcache * +alloc_revcache( +rspl *s +) { + revcache *rc; + + DBG(("alloc_revcache called\n")); + if ((rc = (revcache *) rev_calloc(s, 1, sizeof(revcache))) == NULL) + error("rspl malloc failed - reverse cell cache"); + INCSZ(s, sizeof(revcache)); + + rc->s = s; /* For stats */ + + /* Allocate an initial cell hash index */ + rc->cell_hash_size = primes[0]; + + if ((rc->hashtop = (cell **) rev_calloc(s, rc->cell_hash_size, sizeof(cell *))) == NULL) + error("rspl malloc failed - reverse cell cache index"); + INCSZ(s, rc->cell_hash_size * sizeof(cell *)); + + /* Allocate an initial simplex face match hash index */ + rc->spx_hash_size = primes[0]; + + if ((rc->spxhashtop = (simplex **) rev_calloc(s, rc->spx_hash_size, sizeof(simplex *))) == NULL) + error("rspl malloc failed - reverse simplex cache index"); + INCSZ(s, rc->spx_hash_size * sizeof(simplex *)); + + return rc; +} + +/* Free the reverse cell cache */ +static void +free_revcache(revcache *rc) { + int i; + cell *cp, *ncp; + + /* Free any stuff allocated in the cell contents, and the cell itself. */ + for (cp = rc->mrubot; cp != NULL; cp = ncp) { + ncp = cp->mruup; + free_cell_contents(cp); + free(cp); + DECSZ(rc->s, sizeof(cell)); + } + + /* Free the hash indexes */ + free(rc->hashtop); + DECSZ(rc->s, rc->cell_hash_size * sizeof(cell *)); + free(rc->spxhashtop); + DECSZ(rc->s, rc->spx_hash_size * sizeof(simplex *)); + + DECSZ(rc->s, sizeof(revcache)); + free(rc); +} + +/* Invalidate the whole cache */ +static void +invalidate_revcache( +revcache *rc) +{ + int i; + cell *cp; + + rc->nunlocked = 0; + + /* Free any stuff allocated in the cell contents */ + for (cp = rc->mrubot; cp != NULL; cp = cp->mruup) { + free_cell_contents(cp); + cp->refcount = 0; /* Make sure they can now be reused */ + cp->ix = 0; + cp->flags = 0; /* Contents needs re-initializing */ + rc->nunlocked++; + } + + /* Clear the hash table so they can't be hit */ + for (i = 0; i < rc->cell_hash_size; i++) { + rc->hashtop[i] = NULL; + } + +} + +#define HASH(xx, yy) ((yy) % xx->cell_hash_size) + +/* Allocate another cell, and add it to the cache. */ +/* This may re-size the hash index too. */ +/* Return the pointer to the new cell. */ +/* (Note it's not our job here to honour the memory limit) */ +static cell * +increase_revcache( +revcache *rc +) { + cell *nxcell; /* Newly allocated cell */ + int i; + + DBG(("Adding another chunk of cells to cache\n")); + +#ifdef NEVER /* We may be called with force != 0 */ + if (rc->s->rev.sz >= rc->s->rev.max_sz) + return NULL; +#endif + + if ((nxcell = (cell *) rev_calloc(rc->s, 1, sizeof(cell))) == NULL) + error("rspl malloc failed - reverse cache cells"); + INCSZ(rc->s, sizeof(cell)); + + nxcell->s = rc->s; + + /* Add cell to the bottom of the cache mru linked list */ + if (rc->mrutop == NULL) /* List was empty */ + rc->mrutop = nxcell; + else { + rc->mrubot->mrudown = nxcell; /* Splice into bottom */ + nxcell->mruup = rc->mrubot; + } + rc->mrubot = nxcell; + rc->nacells++; + rc->nunlocked++; + + DBG(("cache is now %d cells\n",rc->nacells)); + + /* See if the hash index should be re-sized */ + if (rc->nacells > (HASH_FILL_RATIO * rc->cell_hash_size)) { + for (i = 0; primes[i] > 0 && primes[i] <= rc->cell_hash_size; i++) + ; + if (primes[i] > 0) { + int cell_hash_size = rc->cell_hash_size; /* Old */ + cell **hashtop = rc->hashtop; + + rc->cell_hash_size = primes[i]; + + DBG(("Increasing cell cache hash index to %d\n",cell_hash_size)); + /* Allocate a new index */ + if ((rc->hashtop = (cell **) rev_calloc(rc->s, rc->cell_hash_size, sizeof(cell *))) == NULL) + error("rspl malloc failed - reverse cell cache index"); + INCSZ(rc->s, rc->cell_hash_size * sizeof(cell *)); + + /* Transfer all the cells to the new index */ + for (i = 0; i < cell_hash_size; i++) { + cell *c, *nc; + for (c = hashtop[i]; c != NULL; c = nc) { + int hash; + nc = c->hlink; + hash = HASH(rc, c->ix); /* New hash */ + c->hlink = rc->hashtop[hash]; /* Add to new hash index */ + rc->hashtop[hash] = c; + } + } + + /* Done with old index */ + free(hashtop); + DECSZ(rc->s, cell_hash_size * sizeof(cell *)); + } + } + + return nxcell; +} + +/* Reduce the cache memory usage by freeing the least recently used unlocked cell. */ +/* Return nz if we suceeeded in freeing some memory. */ +static int decrease_revcache( +revcache *rc /* Reverse cache structure */ +) { + int hit = 0; + int hash; + cell *cp; + + DBG(("Decreasing cell cache memory allocation by freeing a cell\n")); + + /* Use the least recently used unlocked cell */ + for (cp = rc->mrubot; cp != NULL && cp->refcount > 0; cp = cp->mruup) + ; + + /* Run out of unlocked cells */ + if (cp == NULL) { + DBG(("Failed to find unlocked cell to free\n")); +//printf("~1 failed to decrease memory\n"); + return 0; + } + + /* If it has been used before, free up the simplexes */ + free_cell_contents(cp); + + /* Remove from current hash index (if it is in it) */ + hash = HASH(rc,cp->ix); /* Old hash */ + if (rc->hashtop[hash] == cp) { + rc->hashtop[hash] = cp->hlink; + } else { + cell *c; + for (c = rc->hashtop[hash]; c != NULL && c->hlink != cp; c = c->hlink) + ; + if (c != NULL) + c->hlink = cp->hlink; + } + + /* Free up this cell - Remove it from LRU list */ + if (rc->mrutop == cp) + rc->mrutop = cp->mrudown; + if (rc->mrubot == cp) + rc->mrubot = cp->mruup; + if (cp->mruup != NULL) + cp->mruup->mrudown = cp->mrudown; + if (cp->mrudown != NULL) + cp->mrudown->mruup = cp->mruup; + cp->mruup = cp->mrudown = NULL; + free(cp); + DECSZ(rc->s, sizeof(cell)); + rc->nacells--; + rc->nunlocked--; + + DBG(("Freed a rev cache cell\n")); + return 1; +} + +/* Return a pointer to an appropriate reverse cell */ +/* cache structure. cell->flags will be 0 if the cell */ +/* has been reallocated. cell contents will be 0 if */ +/* never used before. */ +/* The cell reference count is incremented, so that it */ +/* can't be thrown out of the cache. The cell must be */ +/* released with uncache_rcell() when it's no longer needed. */ +/* return NULL if we ran out of room in the cache */ +static cell *cache_rcell( +revcache *rc, /* Reverse cache structure */ +int ix, /* fwd index of cell */ +int force /* if nz, force memory allocation, so that we have at least one cell */ +) { + int hit = 0; + int hash; + cell *cp; + + /* keep memory in check - fail if we're out of memory and can't free any */ + /* (Doesn't matter if it might be a hit, it will get picked up the next time) */ + if (!force && rc->s->rev.sz > rc->s->rev.max_sz && rc->nunlocked <= 0) { + return NULL; + } + +//if (rc->nunlocked == 0 && rc->s->rev.sz > rc->s->rev.max_sz) +//printf("~1 unable to decrease_revcache 5\n"); + + /* Free up memory to get below threshold */ + while (rc->nunlocked > 0 && rc->s->rev.sz > rc->s->rev.max_sz) { + if (decrease_revcache(rc) == 0) + break; + } + + hash = HASH(rc,ix); /* Compute hash of fwd cell index */ + + /* See if we get a cache hit */ + for (cp = rc->hashtop[hash]; cp != NULL; cp = cp->hlink) { + if (ix == cp->ix) { /* Hit */ + hit = 1; +#ifdef STATS + rc->s->rev.st[rc->s->rev.sb->op].chits++; +#endif /* STATS */ + break; + } + } + if (!hit) { /* No hit, use new cell or the least recently used cell */ + int ohash; + + /* If we haven't used all our memory, or if we are forced and have */ + /* no cell we can re-use, then noallocate another cell */ + if (rc->s->rev.sz < rc->s->rev.max_sz || (force && rc->nunlocked == 0)) { + cp = increase_revcache(rc); + hash = HASH(rc,ix); /* Re-compute hash in case hash size changed */ +//printf("~1 using new cell\n"); + } else { +//printf("~1 memory limit has been reached, using old cell\n"); + + for (;;) { + /* Use the least recently used unlocked cell */ + for (cp = rc->mrubot; cp != NULL && cp->refcount > 0; cp = cp->mruup) + ; + + /* Run out of unlocked cells */ + if (cp == NULL) { +//printf("~1 none available\n"); + return NULL; + } + + /* If it has been used before, free up the simplexes */ + free_cell_contents(cp); + + /* Remove from current hash index (if it is in it) */ + ohash = HASH(rc,cp->ix); /* Old hash */ + if (rc->hashtop[ohash] == cp) { + rc->hashtop[ohash] = cp->hlink; + } else { + cell *c; + for (c = rc->hashtop[ohash]; c != NULL && c->hlink != cp; c = c->hlink) + ; + if (c != NULL) + c->hlink = cp->hlink; + } + + /* If we're now under the memory limit, use this cell */ + if (rc->s->rev.sz < rc->s->rev.max_sz) { + break; + } + +//printf("~1 freeing a cell\n"); + /* Free up this cell and look for another one */ + /* Remove it from LRU list */ + if (rc->mrutop == cp) + rc->mrutop = cp->mrudown; + if (rc->mrubot == cp) + rc->mrubot = cp->mruup; + if (cp->mruup != NULL) + cp->mruup->mrudown = cp->mrudown; + if (cp->mrudown != NULL) + cp->mrudown->mruup = cp->mruup; + cp->mruup = cp->mrudown = NULL; + free(cp); + DECSZ(rc->s, sizeof(cell)); + rc->nacells--; + rc->nunlocked--; + } + } + +#ifdef STATS + rc->s->rev.st[rc->s->rev.sb->op].cmiss++; +#endif /* STATS */ + + /* Add this cell to hash index */ + cp->hlink = rc->hashtop[hash]; + rc->hashtop[hash] = cp; /* Add to hash table and list */ + + cp->ix = ix; + cp->flags = 0; /* Contents needs re-initializing */ +//printf("~1 returning fresh cell\n"); + } + + /* Move slected cell to the top of the mru list */ + if (cp->mruup != NULL) { /* This one wasn't already at top */ + cp->mruup->mrudown = cp->mrudown; + if (cp->mrudown == NULL) /* This was bottom */ + rc->mrubot = cp->mruup; /* New bottom */ + else + cp->mrudown->mruup = cp->mruup; + /* Put this one at the top */ + rc->mrutop->mruup = cp; + cp->mrudown = rc->mrutop; + rc->mrutop = cp; + cp->mruup = NULL; + } + if (cp->refcount == 0) { + rc->nunlocked--; + } + + cp->refcount++; + + return cp; +} + +/* Tell the cache that we aren't using this cell anymore, */ +/* but to keep it in case it is needed again. */ +static void uncache_rcell( +revcache *rc, /* Reverse cache structure */ +cell *cp +) { + if (cp->refcount > 0) { + cp->refcount--; + if (cp->refcount == 0) { + rc->nunlocked++; + } + } else + warning("rspl cell cache assert: refcount overdecremented!"); +} + +/* ====================================================== */ +/* Reverse rspl setup functions */ + +/* Called by rspl initialisation */ +/* Note that reverse cell lookup tables are not */ +/* allocated & created until the first call */ +/* to a reverse interpolation function. */ +void +init_rev(rspl *s) { + + /* First section */ + s->rev.inited = 0; + s->rev.res = 0; + s->rev.no = 0; + s->rev.rev = NULL; + + /* Second section */ + s->rev.rev_valid = 0; + s->rev.nnrev = NULL; + + /* Third section */ + s->rev.cache = NULL; + + /* Fourth section */ + s->rev.sb = NULL; + + /* Methods */ + s->rev_set_limit = rev_set_limit_rspl; + s->rev_get_limit = rev_get_limit_rspl; + s->rev_interp = rev_interp_rspl; + s->rev_locus = rev_locus_rspl; + s->rev_locus_segs = rev_locus_segs_rspl; +} + +/* Free up all the reverse interpolation info */ +void free_rev( +rspl *s /* Pointer to rspl grid */ +) { + int e, di = s->di; + int **rpp, *rp; + +#ifdef STATS + { + int i, totcalls = 0; + for (i = 0; i < 5; i++) { + totcalls += s->rev.st[i].searchcalls; + } + + printf("\n===============================\n"); + printf("di = %d, do = %d\n",s->di, s->fdi); + for (i = 0; i < 5; i++) { + int calls = s->rev.st[i].searchcalls; + if (calls == 0) + continue; + printf("\n- - - - - - - - - - - - - - - -\n"); + printf("Operation %s\n",opnames[i]); + printf("Search calls = %d = %f%%\n",s->rev.st[i].searchcalls, + 100.0 * s->rev.st[i].searchcalls/totcalls); + printf("Cells searched/call = %f\n",s->rev.st[i].csearched/(double)calls); + printf("Simplexes searched/call = %f\n",s->rev.st[i].ssearched/(double)calls); + printf("Simplexes inited level 1/call = %f\n",s->rev.st[i].sinited/(double)calls); + printf("Simplexes inited level 2 (LU)/call = %f\n",s->rev.st[i].sinited2a/(double)calls); + printf("Simplexes inited level 2 (SVD)/call = %f\n",s->rev.st[i].sinited2b/(double)calls); + printf("Simplexes invalidated level 4/call = %f\n",s->rev.st[i].sinited4i/(double)calls); + printf("Simplexes inited level 4/call = %f\n",s->rev.st[i].sinited4/(double)calls); + printf("Simplexes invalidated level 5/call = %f\n",s->rev.st[i].sinited5i/(double)calls); + printf("Simplexes inited level 5 (LU)/call = %f\n",s->rev.st[i].sinited5a/(double)calls); + printf("Simplexes inited level 5 (SVD)/call = %f\n",s->rev.st[i].sinited5b/(double)calls); + if ((s->rev.st[i].chits + s->rev.st[i].cmiss) == 0) + printf("No cache calls\n"); + else + printf("Cell hit rate = %f%%\n", + 100.0 * s->rev.st[i].chits/(double)(s->rev.st[i].chits + s->rev.st[i].cmiss)); + } + printf("\n===============================\n"); + } +#endif /* STATS */ + + /* Free up Fourth section */ + if (s->rev.sb != NULL) { + free_search(s->rev.sb); + s->rev.sb = NULL; + } + /* Free up the Third section */ + if (s->rev.cache != NULL) { + free_revcache(s->rev.cache); /* Reverse cell cache */ + s->rev.cache = NULL; + } + + /* Free up the Second section */ + if (s->rev.nnrev != NULL) { + /* Free arrays at grid points, taking care of reference count */ + for (rpp = s->rev.nnrev; rpp < (s->rev.nnrev + s->rev.no); rpp++) { + if ((rp = *rpp) != NULL && --rp[2] <= 0) { + DECSZ(s, rp[0] * sizeof(int)); + free(*rpp); + *rpp = NULL; + } + } + free(s->rev.nnrev); + DECSZ(s, s->rev.no * sizeof(int *)); + s->rev.nnrev = NULL; + } + + if (di > 1 && s->rev.rev_valid) { + rev_struct *rsi, **rsp; + size_t ram_portion = g_avail_ram; + + /* Remove it from the linked list */ + for (rsp = &g_rev_instances; *rsp != NULL; rsp = &((*rsp)->next)) { + if (*rsp == &s->rev) { + *rsp = (*rsp)->next; + break; + } + } + + /* Aportion the memory */ + g_no_rev_cache_instances--; + + if (g_no_rev_cache_instances > 0) { + ram_portion /= g_no_rev_cache_instances; + for (rsi = g_rev_instances; rsi != NULL; rsi = rsi->next) + rsi->max_sz = ram_portion; + if (s->verbose) + fprintf(stdout, "%cThere %s %d rev cache instance%s with %lu Mbytes limit\n", + cr_char, + g_no_rev_cache_instances > 1 ? "are" : "is", + g_no_rev_cache_instances, + g_no_rev_cache_instances > 1 ? "s" : "", + ram_portion/1000000); + } + } + + s->rev.rev_valid = 0; + + if (s->rev.rev != NULL) { + /* Free arrays at grid points, taking care of reference count */ + for (rpp = s->rev.rev; rpp < (s->rev.rev + s->rev.no); rpp++) { + if ((rp = *rpp) != NULL && --rp[2] <= 0) { + DECSZ(s, rp[0] * sizeof(int)); + free(*rpp); + *rpp = NULL; + } + } + free(s->rev.rev); + DECSZ(s, s->rev.no * sizeof(int *)); + s->rev.rev = NULL; + } + + /* If first section has been initialised */ + if (s->rev.inited != 0) { + + /* Sub-simplex information */ + for (e = 0; e <= di; e++) { + rspl_free_ssimplex_info(s, &s->rev.sspxi[e]); + } + s->rev.res = 0; + s->rev.no = 0; + s->rev.inited = 0; + } + DBG(("rev allocation left after free = %d bytes\n",s->rev.sz)); +} + +/* - - - - - - - - - - - - - - - - - - - - - - - - - - */ + +#ifdef NEVER /* Test code */ +/* Reverse closest find using exaustive pseudo hilbert search */ +static void debug_find_closest_rev( +rspl *s, +double *out, +double *in +) { + double best = 1e38; + int e, f; + rpsh counter; /* Pseudo-hilbert counter */ + int gc[MXDI]; /* Grid index value */ + double iv[MXDI]; + float *gp; /* Pointer to grid data */ + + rpsh_init(&counter, s->di, (unsigned int *)s->g.res, gc); /* Initialise counter */ + for (;;) { + double dist; + + /* Compute grid pointer and input sample values */ + gp = s->g.a; /* Base of grid data */ + for (e = 0; e < s->di; e++) { /* Input tables */ + gp += s->g.fci[e] * gc[e]; /* Grid value pointer */ + iv[e] = s->g.l[e] + gc[e] * s->g.w[e]; /* Input sample values */ + } + + dist = 0.0; + for (f = 0; f < s->fdi; f++) { + double tt = in[f] - (double)gp[f]; + dist += tt * tt; + } + if (dist < best) { + best = dist; + for (e = 0; e < s->di; e++) + out[e] = iv[e]; + } + + /* Increment counter */ + if (rpsh_inc(&counter, gc)) + break; + } +} +#endif /* NEVER */ + +/* ========================================================== */ +/* reverse lookup acceleration structure initialisation code */ + +/* The reverse lookup relies on a search of the fwd interpolation tables. */ +/* To eliminate out of gamut points quickly, to provide a starting point for */ +/* the search, and to guarantee that all possible reverse solutions are discovered, */ +/* a spatial indexing structure is used to provide a list of starting candidate */ +/* forward indexes for a given output value. (rev.rev[]) */ +/* The reverse structure contains an fdi dimensional cell grid, each element of the */ +/* cell grid holding the indexes of the forward interpolation grid, which intersect */ +/* that ranges of output values. A reverse cell will be empty if there is no */ +/* potential exact solution. */ +/* Note that unlike the forward grid which is composed of verticies, */ +/* this grid is composed of cells (there is an extra row allocated */ +/* during construction using verticies, that are not used when converted. */ +/* to cells) */ +/* For accelleration of the nearest lookup, a parallel reverse grid is */ +/* constructed that holds lists of forward grid cells that may hold the */ +/* nearest point within the gamut. These lists may be empty if we are within */ +/* gamut - ie. the rev.nnrev[] array is the complement of the rev.rev[] array. */ +/* During construction of rev.nnrev[], it is initially filled with lists for */ +/* the potential nearest cell list for each vertex (hence the extra rows allocated */ +/* for rev[] and nnrev[]), and these are then merged down to form the list */ +/* for each cell extent. The nnrev[] array is filled using a seed fill algorithm, */ +/* starting from the edges of the filled cells in rev[]. */ +/* Since many of the cells map to the same surface region, many of the fwd cell lists */ +/* are shared. */ + +/* NOTE: that the nnrev accuracy doesn't seem as good as fill_nncell() ! */ +/* Could we fix this with better geometry calculations ??? */ + +/* rev.nnrev[] cache entry record */ +struct _nncache{ + int min[MXRO]; /* bwd vertex extent covered by this list */ + int max[MXRO]; + int *rip; /* Fwd cell list */ + struct _nncache *next; /* Link list for this cache key */ +}; typedef struct _nncache nncache; + +/* Structure to hold prime seed vertex information */ +struct _primevx{ + int ix; /* Index of prime seed */ + int gc[MXRO]; /* coordinate of the prime seed vertex */ + struct _primevx *next; /* Linked list for final conversion */ + int *clist; /* Cell list generated for prime cell */ +}; typedef struct _primevx primevx; + +/* Structure to hold temporary nn reverse vertex propogation information */ +struct _propvx{ + int ix; /* Index of this secondary seed */ + int gc[MXRO]; /* coordinate of this secondary seed */ + int cix; /* Index of the closest surface nnrev vertex */ + double dsq; /* Distance to the closest point squared */ + int pass; /* Propogation pass */ + struct _propvx *next; /* Linked list for next seeds */ +}; typedef struct _propvx propvx; + +/* Initialise the rev Second section acceleration information. */ +/* This is called when it is discovered on a call that s->rev.rev_valid == 0 */ +static void init_revaccell( +rspl *s +) { + int i, j; /* Index of fwd grid point */ + int e, f, ee, ff; + int di = s->di; + int fdi = s->fdi; + int gno = s->g.no; + int rgno = s->rev.no; + int argres = s->rev.ares; /* Allocation rgres, = no bwd cells +1 */ + int rgres = s->rev.res; /* no bwd cells */ + int rgres_1 = rgres-1; /* rgres -1 == maximum base coord value */ + + schbase *b = s->rev.sb; /* Base search information */ + char *vflag = NULL; /* Per vertex flag used during construction of nnrev */ + float *gp; /* Pointer to fwd grid points */ + primevx *plist = NULL, *ptail = NULL; /* Prime seed list for last pass */ + propvx *alist = NULL; /* Linked list of active secondary seeds */ + propvx *nlist = NULL; /* Linked list of next secondary seeds */ + DCOUNT(gg, MXRO, fdi, 0, 0, argres);/* Track the prime seed coordinate */ + DCOUNT(cc, MXRO, fdi, -1, -1, 2); /* Neighborhood offset counter */ + int nn[MXRO]; /* Neighbor coordinate */ + int pass = 0; /* Propogation pass */ + int nncsize; /* Size of the rev.nnrev construction cache index */ + nncache **nnc; /* nn cache index, used during construction of nnrev */ + unsigned hashk; /* Hash key */ + nncache *ncp; /* Hash entry pointer */ + int nskcells = 0; /* Number of skiped cells (debug) */ +#ifdef DEBUG + int cellinrevlist = 0; + int fwdcells = 0; +#endif + + DBG(("init_revaccell called, di = %d, fdi = %d, mgres = %d\n",di,fdi,(int)s->g.mres)); + + if (!s->rev.fastsetup) { + /* Temporary per bwd vertex/cell flag */ + if ((vflag = (char *) calloc(rgno, sizeof(char))) == NULL) + error("rspl malloc failed - rev.vflag points"); + INCSZ(s, rgno * sizeof(char)); + } + + /* + * The rev[] and nnrev[] grids contain pointers to lists of grid cube base indexes. + * If the pointer is NULL, then there are no base indexes in that list. + * A non NULL list uses element [0] to indicate the alocation size of the list, + * [1] contains the index of the next free location, [2] contains the reference + * count (lists may be shared), the list starts at [3]. The last entry is marked with -1. + */ + + /* We won't include any fwd cells that are over the ink limit, */ + /* so makes sure that the fwd cell nodes all have an ink limit value. */ + if (b != NULL && s->limiten) { + ECOUNT(gc, MXDIDO, s->di, 0, s->g.res, 0); /* coordinates */ + double iv[MXDI]; /* Input value corresponding to grid */ + + DBG(("Looking up fwd vertex ink limit values\n")); + /* Calling the limit function for each fwd vertex could be bad */ + /* if the limit function is slow. Maybe an octree type algorithm */ + /* could be used if this is a problem ? */ + EC_INIT(gc); + for (i = 0, gp = s->g.a; i < s->g.no; i++, gp += s->g.pss) { + if (gp[-1] == L_UNINIT) { + for (e = 0; e < di; e++) + iv[e] = s->g.l[e] + gc[e] * s->g.w[e]; /* Input sample values */ + gp[-1] = (float)(INKSCALE * s->limitf(s->lcntx, iv)); + } + EC_INC(gc); + } + s->g.limitv_cached = 1; + } + + /* We then fill in the in-gamut reverse grid lookups, */ + /* and identify nnrev prime seed verticies */ + + DBG(("filling in rev.rev[] grid\n")); + + /* To create rev.rev[], for all fwd grid points, form the cube with that */ + /* point at its base, and determine the bounding box of the output values */ + /* that could intersect that cube. */ + /* As a start for creating rev.nnrevp[], flag which bwd verticies are */ + /* covered by the fwd grid output range. */ + for (gp = s->g.a, i = 0; i < gno; gp += s->g.pss, i++) { + datao min, max; + int imin[MXRO], imax[MXRO], gc[MXRO]; + int uil; /* One is under the ink limit */ + +//printf("~1 i = %d/%d\n",i,gno); + /* Skip grid points on the upper edge of the grid, since there */ + /* is no further grid point to form a cube range with. */ + for (e = 0; e < di; e++) { + if(G_FL(gp, e) == 0) /* At the top edge */ + break; + } + if (e < di) { /* Top edge - skip this cube */ + continue; + } + + /* Find the output value bounding box values for this grid cell */ + uil = 0; + for (f = 0; f < fdi; f++) /* Init output min/max */ + min[f] = max[f] = gp[f]; + if (b == NULL || !s->limiten || gp[-1] <= s->limitv) + uil = 1; + + /* For all other grid points in the cube */ + for (ee = 1; ee < (1 << di); ee++) { + float *gt = gp + s->g.fhi[ee]; /* Pointer to cube vertex */ + + if (b == NULL || !s->limiten || gt[-1] <= s->limitv) + uil = 1; + + /* Update bounding box for this grid point */ + for (f = 0; f < fdi; f++) { + if (min[f] > gt[f]) + min[f] = gt[f]; + if (max[f] < gt[f]) + max[f] = gt[f]; + } + } + + /* Skip any fwd cells that are over the ink limit */ + if (!uil) { + nskcells++; + continue; + } + + /* Figure out intersection range in reverse grid */ + for (f = 0; f < fdi; f++) { + double t; + int mi; + double gw = s->rev.gw[f]; + double gl = s->rev.gl[f]; + t = (min[f] - gl)/gw; + mi = (int)floor(t); /* Grid coordinate */ + if (mi < 0) /* Limit to valid cube base index range */ + mi = 0; + else if (mi > rgres_1) + mi = rgres_1; + imin[f] = mi; + t = (max[f] - gl)/gw; + mi = (int)floor(t); /* Grid coordinate */ + if (mi < 0) /* Limit to valid cube base index range */ + mi = 0; + else if (mi > rgres_1) + mi = rgres_1; + imax[f] = mi; + } + +//printf("Scanning over grid:\n"); +//for (f = 0; f < fdi; f++) +//printf("Min[%d] = %d -> Max[%d] = %d\n",f,imin[f],f,imax[f]); + + /* Now create forward index and vector with all the reverse grid cells */ + for (f = 0; f < fdi; f++) + gc[f] = imin[f]; /* init coords */ + + for (f = 0; f < fdi;) { /* For all of intersect cube */ + int **rpp, *rp; + + /* Compute pointer to grid cell */ + for (rpp = s->rev.rev, f = 0; f < fdi; f++) + rpp += gc[f] * s->rev.coi[f]; + rp = *rpp; + +//printf("Currently at grid:\n"); +//for (f = 0; f < fdi; f++) +//printf("gc[%d] = %d\n",f,gc[f]); + + if (rp == NULL) { + if ((rp = (int *) rev_malloc(s, 6 * sizeof(int))) == NULL) + error("rspl malloc failed - rev.grid entry"); + INCSZ(s, 6 * sizeof(int)); + *rpp = rp; + rp[0] = 6; /* Allocation */ + rp[1] = 4; /* Next free Cell */ + rp[2] = 1; /* Reference count */ + rp[3] = i; + rp[4] = -1; /* End marker */ + } else { + int z = rp[1], ll = rp[0]; + if (z >= (ll-1)) { /* Not enough space */ + INCSZ(s, ll * sizeof(int)); + ll *= 2; + if ((rp = (int *) rev_realloc(s, rp, sizeof(int) * ll)) == NULL) + error("rspl realloc failed - rev.grid entry"); + *rpp = rp; + rp[0] = ll; + } + rp[z++] = i; + rp[z] = -1; + rp[1] = z; + } + /* Increment index */ + for (f = 0; f < fdi; f++) { + gc[f]++; + if (gc[f] <= imax[f]) + break; /* No carry */ + gc[f] = imin[f]; + } + } /* Next reverse grid point in intersecting cube */ + + if (s->rev.fastsetup) + continue; /* Skip nnrev setup */ + + + /* Now also register which grid points are in-gamut and are part of cells */ + /* than have a rev.rev[] list. */ + + /* Figure out intersection range in reverse nn (construction) vertex grid */ + /* This range may be empty if a grid isn't stradled by the fwd cell output */ + /* range. */ + for (f = 0; f < fdi; f++) { + double t; + int mi; + double gw = s->rev.gw[f]; + double gl = s->rev.gl[f]; + t = (min[f] - gl)/gw; + mi = (int)ceil(t); /* Grid coordinate */ + if (mi < 0) /* Limit to valid cube base index range */ + mi = 0; + else if (mi >= argres) + mi = rgres; + imin[f] = mi; + t = (max[f] - gl)/gw; + mi = (int)floor(t); /* Grid coordinate */ + if (mi < 0) /* Limit to valid cube base index range */ + mi = 0; + else if (mi >= argres) + mi = rgres; + imax[f] = mi; + if (imax[f] < imin[f]) + break; /* Doesn't straddle any verticies */ + } + + if (f >= fdi) { /* There are seed verticies to mark */ + +//printf("~1 marking prime seed vertex %d\n",i); + + /* Mark an initial seed point nnrev vertex, and */ + /* create a surface point propogation record for it */ + for (f = 0; f < fdi; f++) + gc[f] = imin[f]; /* init coords */ + + for (f = 0; f < fdi;) { /* For all of intersect cube */ + int **rpp, *rp; + char *fpp; + + /* Compute pointer to grid cell */ + for (rpp = s->rev.nnrev, fpp = vflag, f = 0; f < fdi; f++) { + int inc = gc[f] * s->rev.coi[f]; + rpp += inc; + fpp += inc; + } + rp = *rpp; + + *fpp = 3; /* Initial seed point */ + + /* Increment index */ + for (f = 0; f < fdi; f++) { + gc[f]++; + if (gc[f] <= imax[f]) + break; /* No carry */ + gc[f] = imin[f]; + } + } + } + } /* Next base grid point */ + + DBG(("We skipped %d cells that were over the limit\n",nskcells)); + + /* Setup the nnrev array if we are not doing a fast setup. */ + /* (fastsetup will instead fill the nnrev array on demand, */ + /* using an exaustive search.) */ + if (!s->rev.fastsetup) { + + /* The next step is to use all the prime seed grid points to set and propogate */ + /* the index of the closest fwd vertex through the revnn[] array. */ + /* (This doesn't work perfectly. Sometimes a vertex is not linked to it's closest */ + /* prime. I'm not sure if this is due to a bug here, or is a quirk of geometry */ + /* that a prime that is closest to a vertex isn't closest for any of its neighbors.) */ + DBG(("filling in rev.nnrev[] grid\n")); + + /* For all the primary seed points */ + DC_INIT(gg); + for (i = 0; i < rgno; i++) { + int **rpp; + primevx *prime= NULL; /* prime cell information structure */ + + if (vflag[i] != 3) { /* Not a prime seed point */ + goto next_seed_point; + } + + rpp = s->rev.nnrev + i; + +//printf("~1 potential rev.nnrev[] prime seed %d, about to scan neibors\n",i); + /* For all the neigbors of this seed */ + DC_INIT(cc); + while (!DC_DONE(cc)) { + propvx *prop; /* neighor cell propogation structure */ + int nix = 0; /* Neighbor cell index */ + char *fpp = vflag; + int **nrpp = s->rev.nnrev; + double dsq; + + for (f = 0; f < fdi; f++) { + nn[f] = gg[f] + cc[f]; + if (nn[f] < 0 || nn[f] >= argres) + break; /* Out of bounds */ + nix += nn[f] * s->rev.coi[f]; + } + fpp = vflag + nix; + + /* If neighbor out of bounds, or is a prime seed point, skip it */ + if (f < fdi || *fpp == 3) { + goto next_neighbor; + } + +//printf("~1 identified prime seed %d with neighbor %d\n",i,nix); + /* We now know that this prime seed will propogate, */ + /* so get/create the temporary information record for it */ + if (prime == NULL) { + + /* If this prime seed hasn't be setup before */ + if (*rpp != NULL) { + prime = *((primevx **)rpp); + } else { + /* Allocate a primevx if there isn't one */ + if ((prime = (primevx *) calloc(1, sizeof(primevx))) == NULL) + error("rspl malloc failed - rev.nnrev prime info structs"); + *((primevx **)rpp) = prime; + prime->ix = i; + for (f = 0; f < fdi; f++) + prime->gc[f] = gg[f]; +//if (fdi > 1) printf("~1 setting prime %d, gc = %d, %d, %d\n", i, prime->gc[0], prime->gc[1], prime->gc[2]); + } + } + + /* Pointer to nnrev vertex neighbor point */ + nrpp = s->rev.nnrev + nix; + + /* Compute the distance squared from this prime seed to this neighbor */ + for (dsq = 0.0, f = 0; f < fdi; f++) { + double tt = (gg[f] - nn[f]) * s->rev.gw[f]; + dsq += tt * tt; + } + + /* Get or allocate a prop structure for it */ + if (*nrpp != NULL) { + prop = *((propvx **)nrpp); + if ((dsq + 1e-6) < prop->dsq) { /* This prime is closer than previous */ + prop->cix = i; /* The index of the closest prime */ + prop->dsq = dsq; /* Distance squared to closest prime */ + } + } else { + if ((prop = (propvx *) calloc(1, sizeof(propvx))) == NULL) + error("rspl malloc failed - rev.nnrev propogation structs"); + *((propvx **)nrpp) = prop; + prop->ix = nix; + for (f = 0; f < fdi; f++) + prop->gc[f] = nn[f]; /* This neighbors coord */ + prop->cix = i; + prop->dsq = dsq; + prop->pass = pass; + prop->next = nlist; /* Add new seed to list of next seeds */ + nlist = prop; + *fpp = 1; + + } + next_neighbor:; + DC_INC(cc); + } + + next_seed_point:; + DC_INC(gg); + } + +//printf("~1 about to propogate secondary seeds\n"); + /* Now we propogate the secondary seed points until there are no more left */ + while(nlist != NULL) { + propvx *next; + propvx *tlp; + + if ((pass += 2) < 0) + error("Assert rev: excessive propogation passes"); +//printf("~1 about to do a round of propogation pass %d\n",(pass+2)/2); + + /* Mark all seed points on the current list with pass-1 */ + for (tlp = nlist; tlp != NULL; tlp = tlp->next) { + *(vflag + tlp->ix) = 2; + tlp->pass = pass-1; + } + + /* Go through each secondary seed in the active list, propogating them */ + for (alist = nlist, nlist = NULL; alist != NULL; alist = next) { + int **rpp; + primevx *prime= NULL; /* prime cell information structure */ + + next = alist->next; /* Next unless we re-insert one */ + + /* Grab this seed points coodinate and index */ + for (i = f = 0; f < fdi; f++) { + gg[f] = alist->gc[f]; + i += gg[f] * s->rev.coi[f]; + } + +//printf("\n~1 propogating from seed %d\n",i); + /* rpp = s->rev.nnrev + i; */ + + /* Grab the corresponding prime seed information record */ + prime = *((primevx **)(s->rev.nnrev + alist->cix)); + + /* For all the neigbors of this seed */ + DC_INIT(cc); + while (!DC_DONE(cc)) { + propvx *prop; /* neighor cell propogation structure */ + int nix; /* Neighbor cell index */ + char *fpp = vflag; + int **nrpp = s->rev.nnrev; + double dsq; + + for (nix = f = 0; f < fdi; f++) { + nn[f] = gg[f] + cc[f]; + if (nn[f] < 0 || nn[f] >= argres) + break; /* Out of bounds */ + nix += nn[f] * s->rev.coi[f]; + } + fpp = vflag + nix; +//printf("~1 neighbor ix %d, flag %d\n",nix,*fpp); + + /* If neighbor out of bounds, current vertex or prime, skip it */ + if (f < fdi || i == nix || *fpp >= 3) { +//printf("~1 skipping neighbour %d\n",nix); + goto next_neighbor2; + } + + /* Pointer to nnrev vertex neighbor point */ + nrpp = s->rev.nnrev + nix; + + /* Compute the distance squared from the prime seed to this neighbor */ + for (dsq = 0.0, f = 0; f < fdi; f++) { + double tt = (prime->gc[f] - nn[f]) * s->rev.gw[f]; + dsq += tt * tt; + } + + /* Get or allocate a prop structure for it */ + if (*nrpp != NULL) { + prop = *((propvx **)nrpp); +//if (prop->ix != nix) error ("Assert: prop index %d doesn't match index %d",prop->ix, nix); + + if ((dsq + 1e-6) < prop->dsq) { /* This prime is closer than previous */ +//printf("~1 updating %d to prime %d, dsq = %f from %f\n",nix, prime->ix, dsq, prop->dsq); + prop->cix = prime->ix; /* The index of the new closest prime */ + prop->dsq = dsq; /* Distance squared to closest prime */ + /* If this is a vertex from previous pass that has changed, */ + /* and it's not ahead of us in the current list, */ + /* put it next on the current list. */ + if (*fpp == 2 && prop->pass != (pass-1)) { +//printf("~1 re-shedule %d (%d) for next propogate\n",nix,prop->ix); +//if (next == NULL) +//printf("Before insert, next = NULL\n"); +//else +//printf("Before insert, next = %d\n",next->ix); + prop->pass = pass-1; /* Re-shedule once only */ + prop->next = next; + next = prop; + } + } + } else { + if ((prop = (propvx *) calloc(1, sizeof(propvx))) == NULL) + error("rspl malloc failed - rev.nnrev propogation structs"); + *((propvx **)nrpp) = prop; + prop->ix = nix; + for (f = 0; f < fdi; f++) + prop->gc[f] = nn[f]; /* This neighbors coord */ + prop->cix = prime->ix; + prop->dsq = dsq; +//printf("~1 propogating to new, %d, dsq = %f, prime %d\n",nix, dsq, prime->ix); + prop->pass = pass; + prop->next = nlist; /* Add new seed to list of next seeds */ + nlist = prop; + *fpp = 1; + } + + next_neighbor2:; + DC_INC(cc); + } + alist->pass = pass; + } + } + +#ifdef DEBUG + DBG(("checking that every vertex is now touched\n")); + for (i = 0; i < rgno; i++) { + if (vflag[i] < 2) { + printf("~1 problem: vertex %d flag = %d\n",i, vflag[i]); + } + if (vflag[i] == 2 && *(s->rev.nnrev + i) == NULL) { + printf("~1 problem: vertex %d flag = %d and struct = NULL\n",i, vflag[i]); + } + } +#endif /* DEBUG */ + +#ifdef NEVER /* Check that all cells are closest to their primes than any other */ +DC_INIT(gg); +for (i = 0; i < rgno; i++) { /* For all the verticies */ + if (vflag[i] == 2) { + propvx *prop = (propvx *) *(s->rev.nnrev + i); + for (j = 0; j < rgno; j++) { /* For all the primes */ + if (vflag[j] == 3) { + primevx *prime = (primevx *) *(s->rev.nnrev + j); + double dsq; + if (prime == NULL) + continue; + for (dsq = 0.0, f = 0; f < fdi; f++) { + double tt = (prime->gc[f] - prop->gc[f]) * s->rev.gw[f]; + dsq += tt * tt; + } + if ((dsq + 1e-6) < prop->dsq) { + warning("~1 vertex %d prime %d, dsq = %f, is closer to prime %d, dsq %f\n", i,prop->cix, prop->dsq, j, dsq); + /* See if any of the neighbors have the closer prime */ + DC_INIT(cc); /* For all the neigbors of this seed */ + while (!DC_DONE(cc)) { + propvx *nprop; /* neighor cell propogation structure */ + int nix; /* Neighbor cell index */ + char *fpp = vflag; + int **nrpp = s->rev.nnrev; + double dsq; + + for (nix = f = 0; f < fdi; f++) { + nn[f] = gg[f] + cc[f]; + if (nn[f] < 0 || nn[f] >= argres) + break; /* Out of bounds */ + nix += nn[f] * s->rev.coi[f]; + } + fpp = vflag + nix; +//printf("~1 neighbor ix %d, flag %d\n",nix,*fpp); + + /* If neighbor out of bounds, current vertex or prime, skip it */ + if (f < fdi || i == nix || *fpp != 2) { +//printf("~1 skipping neighbour %d\n",nix); + goto next_neighbor3; + } + + /* Pointer to nnrev vertex neighbor point */ + nrpp = s->rev.nnrev + nix; + if ((nprop = *((propvx **)nrpp)) != NULL) { +//printf("~1 neighbor %d %d %d has prime %d dsq %f\n",cc[0],cc[1],cc[2],nprop->cix,nprop->dsq); + if (nprop->cix == j) { +//warning("~1 but neighbor has this prime point!\n"); + + } + } + next_neighbor3:; + DC_INC(cc); + } +// prop->cix = j; /* Fix it and see what happens */ +// prop->dsq = dsq; + } + } + } + } + DC_INC(gg); +} + +#endif /* NEVER */ + + + DBG(("about to do convert vertex values to cell lists\n")); + /* Setup a cache for the fwd cell lists created, so that we can */ + /* avoid the list creation and memory allocation for identical lists */ + nncsize = s->rev.ares * s->rev.ares; + if ((nnc = (nncache **) calloc(nncsize, sizeof(nncache *))) == NULL) + error("rspl malloc failed - rev.nnc cache entries"); + + /* Now convert the nnrev secondary vertex points to pointers to fwd cell lists */ + /* Do this in order, so that we don't need the verticies after */ + /* they are converted to cell lists. */ + DC_INIT(gg); + for (i = 0; i < rgno; i++) { + int **rpp, *rp; + propvx *prop = NULL; /* vertex information structure */ + primevx *prime= NULL; /* prime cell information structure */ + int imin[MXRO], imax[MXRO]; /* Prime vertex range for each axis */ + double rmin[MXRO], rmax[MXRO]; /* Float prime vertex value range */ + unsigned int tcount; /* grid touch count for this opperation */ + datao min, max; /* Fwd cell output range */ + int lpix; /* Last prime index seen */ + +//if (fdi > 1) printf("~1 converting vertex %d\n",i); +//if (fdi > 1) printf("~1 coord %d %d %d\n",gg[0],gg[1],gg[2]); + + rpp = s->rev.nnrev + i; + if (vflag[i] == 3) { /* Cell base is prime */ + prime = (primevx *) *rpp; + + if (prime != NULL) { /* It's a propogating prime */ + /* Add prime to the end of the ptime linked list */ + prime->next = NULL; + if (plist == NULL) { + plist = ptail = prime; + } else { + ptail->next = prime; + ptail = prime; + } + } + } else if (vflag[i] == 2) { /* Cell base is secondary */ + prop = (propvx *)*rpp; + } else { /* Hmm */ + /* This seems to happen if the space explored is not really 3D ? */ + if (s->rev.primsecwarn == 0) { + warning("rev: bwd vertex %d is not prime or secondary (vflag = %d)" + "(Check that your measurement data is sane!)",i,vflag[i]); + s->rev.primsecwarn = 1; + } + fill_nncell(s, gg, i); /* Is this valid to do ?? */ + continue; + } + + /* Setup to scan of cube corners, and check that base is within cube grid */ + for (f = 0; f < fdi; f++) { + if (gg[f] > rgres_1) { /* Vertex outside bwd cell range, */ + if (prop != NULL && prime == NULL) { + free(prop); + *rpp = NULL; + } +//printf("~1 done vertex %d because its out of cell bounds\n",i); + goto next_vertex; + } + imin[f] = 0x7fffffff; + imax[f] = -1; + } + + /* For all the vertex points in the nnrev cube starting at this base (i), */ + /* Check if any of them are secondary seed points */ + for (ff = 0; ff < (1 << fdi); ff++) { + if (vflag[i + s->rev.hoi[ff]] == 2) + break; + } + + /* If not a cell that we want to create a nearest fwd cell list for */ + if (ff >= (1 << fdi)) { + /* Don't free anything, because we leave a prime in place, */ + /* and it can't be a prop. */ + goto next_vertex; + } + + /* For all the vertex points in the nnrev cube starting at this base (i), */ + /* accumulate the range they cover */ + lpix = -1; + for (f = 0; f < fdi; f++) { + imin[f] = 0x7fffffff; + imax[f] = -1; + } + for (ff = 0; ff < (1 << fdi); ff++) { + int ii = i + s->rev.hoi[ff]; /* cube vertex index */ + primevx *tprime= NULL; + + /* Grab vertex info and corresponding prime vertex info */ + if (vflag[ii] == 3) { /* Corner is a prime */ + tprime = (primevx *) *(s->rev.nnrev + ii); /* Use itself */ + if (tprime == NULL) + continue; /* Not a propogated in-gamut vertex */ + } else if (vflag[ii] == 2) { + propvx *tprop = (propvx *) *(s->rev.nnrev + ii); /* Use propogated prime */ + tprime = (primevx *) *(s->rev.nnrev + tprop->cix); + } else { + continue; /* Hmm */ + } + if (tprime->ix == lpix) + continue; /* Don't waste time */ + +//if (fdi > 1) printf("~1 corner %d, ix %d, prime %d gc = %d, %d, %d\n", ff, ii, tprime->ix, tprime->gc[0], tprime->gc[1], tprime->gc[2]); + + /* Update bounding box for this prime grid point */ + for (f = 0; f < fdi; f++) { + if (tprime->gc[f] < imin[f]) + imin[f] = tprime->gc[f]; + if (tprime->gc[f] > imax[f]) + imax[f] = tprime->gc[f]; + } + lpix = tprime->ix; + } + +//if (fdi > 1) printf("~1 prime vertex index range = %d - %d, %d - %d, %d - %d\n", imin[0], imax[0], imin[1], imax[1], imin[2], imax[2]); + + /* See if a list matching this range is in the cache */ + hashk = 0; + for (hashk = f = 0; f < fdi; f++) + hashk = hashk * 97 + imin[f] + 43 * (imax[f] - imin[f]); + hashk = hashk % nncsize; +//if (fdi > 1) printf("~1 hashk = %d from %d - %d %d - %d %d - %d\n", hashk, imin[0], imax[0], imin[1], imax[1], imin[2], imax[2]); + + /* See if we can locate an existing list for this range */ + for (ncp = nnc[hashk]; ncp != NULL; ncp = ncp->next) { +//if (fdi > 1) printf("~1 checking %d - %d %d - %d %d - %d\n", ncp->min[0], ncp->max[0], ncp->min[1], ncp->max[1], ncp->min[2], ncp->max[2]); + for (f = 0; f < fdi; f++) { + if (ncp->min[f] != imin[f] + || ncp->max[f] != imax[f]) { +//if (fdi > 1) printf("~1 not a match\n"); + break; + } + } + if (f >= fdi) { +//if (fdi > 1) printf("~1 got a match\n"); + break; /* Found a matching cache entry */ + } + } + + if (ncp != NULL) { + rp = ncp->rip; +//if (fdi > 1) printf("~1 got cache hit hashk = %d, with ref count %d\n\n",hashk, rp[1]); + rp[2]++; /* Increase reference count */ + + } else { + /* This section seems to be the most time consuming part of the nnrev setup. */ + + /* Allocate a cache entry and place it */ + if ((ncp = (nncache *)calloc(1, sizeof(nncache))) == NULL) + error("rspl malloc failed - rev.nn cach record"); + + for (f = 0; f < fdi; f++) { + ncp->min[f] = imin[f]; + ncp->max[f] = imax[f]; + } + ncp->next = nnc[hashk]; + nnc[hashk] = ncp; + + /* Convert the nn destination vertex range into an output value range. */ + for (f = 0; f < fdi; f++) { + double gw = s->rev.gw[f]; + double gl = s->rev.gl[f]; + rmin[f] = gl + imin[f] * gw; + rmax[f] = gl + imax[f] * gw; + } + + /* Do any adjustment of the range needed to acount for the inacuracies */ + /* caused by the vertex quantization. */ + /* (I don't really understand the need for the extra avggw expansion, */ + /* but there are artefacts without this. This size of this sampling */ + /* expansion has a great effect on the performance.) */ + { + double avggw = 0.0; + for (f = 0; f < fdi; f++) + avggw += s->rev.gw[f]; + avggw /= (double)fdi; + for (f = 0; f < fdi; f++) { /* Quantizing range plus extra */ + double gw = s->rev.gw[f]; + rmin[f] -= (0.5 * gw + 0.99 * avggw); + rmax[f] += (0.5 * gw + 0.99 * avggw); + } + } +//if (fdi > 1) printf("~1 prime vertex value adjusted range = %f - %f, %f - %f, %f - %fn", rmin[0], rmax[0], rmin[1], rmax[1], rmin[2], rmax[2]); + + /* computue the rev.rev cell grid range we will need to cover to */ + /* get all the cell output ranges that could touch our nn reverse range */ + for (f = 0; f < fdi; f++) { + double gw = s->rev.gw[f]; + double gl = s->rev.gl[f]; + imin[f] = (int)floor((rmin[f] - gl)/gw); + if (imin[f] < 0) + imin[f] = 0; + else if (imin[f] > rgres_1) + imin[f] = rgres_1; + imax[f] = (int)floor((rmax[f] - gl)/gw); + if (imax[f] < 0) + imax[f] = 0; + else if (imax[f] > rgres_1) + imax[f] = rgres_1; + cc[f] = imin[f]; /* Set grid starting value */ + } + tcount = s->get_next_touch(s); /* Get next grid touched generation count */ + +//if (fdi > 1) printf("~1 Cells to scan = %d - %d, %d - %d, %d - %d\n", imin[0], imax[0], imin[1], imax[1], imin[2], imax[2]); + + rp = NULL; /* We always allocate a new list initially */ + for (f = 0; f < fdi;) { /* For all the cells in the min/max range */ + int ii; + int **nrpp, *nrp; /* Pointer to base of cell list, entry 0 = allocated space */ + + /* Get pointer to rev.rev[] cell list */ + for (nrpp = s->rev.rev, f = 0; f < fdi; f++) + nrpp += cc[f] * s->rev.coi[f]; + + if ((nrp = *nrpp) == NULL) + goto next_range_list; /* This rev.rev[] cell is empty */ + + +//if (fdi > 1) printf("~1 adding list from cell %d, list length %d\n",nrpp - s->rev.rev, nrp[0]); + /* For all the fwd cells in the rev.rev[] list */ + for(nrp += 3; *nrp != -1; nrp++) { + int ix = *nrp; /* Fwd cell index */ + float *fcb = s->g.a + ix * s->g.pss; /* Pntr to base float of fwd cell */ + + if (TOUCHF(fcb) >= tcount) { /* If we seen visited this fwd cell before */ +//if (fdi > 1) printf("~1 skipping cell %d because we alread have it\n",ix); + continue; + } + TOUCHF(fcb) = tcount; /* Touch it so we will skip it next time */ + + /* Compute the range of output values this cell covers */ + for (f = 0; f < fdi; f++) /* Init output min/max */ + min[f] = max[f] = fcb[f]; + + /* For all other grid points in the fwd cell cube */ + for (ee = 1; ee < (1 << di); ee++) { + float *gt = fcb + s->g.fhi[ee]; /* Pointer to cube vertex */ + + /* Update bounding box for this grid point */ + for (f = 0; f < fdi; f++) { + if (min[f] > gt[f]) + min[f] = gt[f]; + if (max[f] < gt[f]) + max[f] = gt[f]; + } + } + +//if (fdi > 1) printf("~1 cell %d range = %f - %f, %f - %f, %f - %f\n", ix, min[0], max[0], min[1], max[1], min[2], max[2]); + + /* See if this fwd cell output values overlaps our region of interest */ + for (f = 0; f < fdi; f++) { + if (min[f] > rmax[f] + || max[f] < rmin[f]) { + break; /* Doesn't overlap */ + } + } + + if (f < fdi) { +//if (fdi > 1) printf("~1 skipping cell %d because we doesn't overlap\n",ix); + continue; /* It doesn't overlap */ + } + +//if (fdi > 1) printf("~1 adding fwd index %d to list\n",ix); +//if (fdi > 1) printf("~1 cell %d range = %f - %f, %f - %f, %f - %f\n", ix, min[0], max[0], min[1], max[1], min[2], max[2]); +#ifdef DEBUG + fwdcells++; +#endif + /* It does, add it to our new list */ + if (rp == NULL) { + if ((rp = (int *) rev_malloc(s, 6 * sizeof(int))) == NULL) + error("rspl malloc failed - rev.nngrid entry"); + INCSZ(s, 6 * sizeof(int)); + rp[0] = 6; /* Allocation */ + rp[1] = 4; /* Next free Cell */ + rp[2] = 1; /* reference count */ + rp[3] = ix; + rp[4] = -1; + } else { + int z = rp[1], ll = rp[0]; + if (z >= (ll-1)) { /* Not enough space */ + INCSZ(s, ll * sizeof(int)); + ll *= 2; + if ((rp = (int *) rev_realloc(s, rp, sizeof(int) * ll)) == NULL) + error("rspl realloc failed - rev.grid entry"); + rp[0] = ll; + } + rp[z++] = ix; + rp[z] = -1; + rp[1] = z; + } + } /* Next fwd cell in list */ + + /* Increment index */ + next_range_list:; + for (f = 0; f < fdi; f++) { + if (++cc[f] <= imax[f]) + break; /* No carry */ + cc[f] = imin[f]; + } + } + + ncp->rip = rp; /* record nnrev cell in cache */ +#ifdef DEBUG + cellinrevlist++; +#endif +//if (fdi > 1) printf("~1 adding cache entry with hashk = %d\n\n",hashk); + } + + /* Put the resulting list in place */ + if (prime != NULL) + prime->clist = rp; /* Save it untill we get rid of the primes */ + else + *rpp = rp; + +//if (*rpp == NULL) printf("~1 problem: we ended up with no list or prime struct at cell %d\n",i); + +#ifdef NEVER +/* Sanity check the list, to see if the list cells corner contain an output value */ +/* that is at least closer to the target than the prime. */ +if (prop != NULL) { + int *tp = rp; + double bdist = 1e60; + double bdist2 = 1e60; + double vx[MXRO]; /* Vertex location */ + double px[MXRO]; /* Prime location */ + double cl[MXRO]; /* Closest output value from list */ + double acl[MXRO]; /* Absolute closest output value */ + double dst; /* Distance to prime */ + int ti; + + primevx *prm = (primevx *) *(s->rev.nnrev + prop->cix); + for (f = 0; f < fdi; f++) { + double gw = s->rev.gw[f]; + double gl = s->rev.gl[f]; + vx[f] = gl + prop->gc[f] * gw; + px[f] = gl + prm->gc[f] * gw; + } + + for(tp++; *tp != -1; tp++) { + int ix = *tp; /* Fwd cell index */ + float *fcb = s->g.a + ix * s->g.pss; /* Pntr to base float of fwd cell */ + + for (ee = 0; ee < (1 << di); ee++) { + double ss; + float *gt = fcb + s->g.fhi[ee]; /* Pointer to cube vertex */ + + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt = vx[f] - gt[f]; + ss += tt * tt; + } + if (ss < bdist) { + bdist = ss; + for (f = 0; f < fdi; f++) + cl[f] = gt[f]; + } + } + } + bdist = sqrt(bdist); + dst = sqrt(prop->dsq); + + /* Lookup best distance to any output value */ + if (dst < bdist) { + float *gt; + for (ti = 0, gt = s->g.a; ti < s->g.no; ti++, gt += s->g.pss) { + double ss; + + for (ss = 0.0, f = 0; f < fdi; f++) { + double tt = vx[f] - gt[f]; + ss += tt * tt; + } + if (ss < bdist2) { + bdist2 = ss; + for (f = 0; f < fdi; f++) + acl[f] = gt[f]; + } + } + } + bdist2 = sqrt(bdist2); + + if (dst < bdist) { + printf("~1 vertex %d has worse distance to values than prime\n",i); + printf("~1 vertex loc %f %f %f\n", vx[0], vx[1], vx[2]); + printf("~1 prime loc %f %f %f, dist %f\n", px[0], px[1], px[2],dst); + printf("~1 closest loc %f %f %f, dist %f\n", cl[0], cl[1], cl[2],bdist); + printf("~1 abs clst loc %f %f %f, dist %f\n", acl[0], acl[1], acl[2], bdist2); + } +} +#endif // NEVER + + if (prop != NULL && prime == NULL) { + free(prop); + } + + next_vertex:; + DC_INC(gg); + } + + DBG(("freeing up the prime seed structurs\n")); + /* Finaly convert all the prime verticies to cell lists */ + /* Free up all the prime seed structures */ + for (;plist != NULL; ) { + primevx *prime, *next = plist->next; + int **rpp; + + rpp = s->rev.nnrev + plist->ix; + if ((prime = (primevx *)(*rpp)) != NULL) { + if (prime->clist != NULL) /* There is a nn list for this cell */ + *rpp = prime->clist; + else + *rpp = NULL; + free(prime); + } else { + error("assert, prime cell %d was empty",plist->ix); + } + plist = next; + } + +#ifdef DEBUG + DBG(("sanity check that all rev accell cells are filled\n")); + DC_INIT(gg); + for (i = 0; i < rgno; i++) { + for (f = 0; f < fdi; f++) { + if (gg[f] > rgres_1) { /* Vertex outside bwd cell range, */ + goto next_vertex3; + } + } + + if (*(s->rev.nnrev + i) == NULL + && *(s->rev.rev + i) == NULL) { +// printf("~1 warning, cell %d [ %d %d %d] has a NULL list\n",i, gg[0],gg[1],gg[2]); + error("cell %d has a NULL list\n",i); + } + next_vertex3:; + DC_INC(gg); + } +#endif /* DEBUG */ + + /* Free up flag array used for construction */ + if (vflag != NULL) { + DECSZ(s, rgno * sizeof(char)); + free(vflag); + } + + /* Free up nn list cache indexing structure used in construction */ + if (nnc != 0) { + for (i = 0; i < nncsize; i++) { + nncache *nncp; + /* Run through linked list freeing entries */ + for (ncp = nnc[i]; ncp != NULL; ncp = nncp) { + nncp = ncp->next; + free(ncp); + } + } + free(nnc); + nnc = NULL; + } + } + + if (s->rev.rev_valid == 0 && di > 1) { + rev_struct *rsi; + size_t ram_portion = g_avail_ram; + + /* Add into linked list */ + s->rev.next = g_rev_instances; + g_rev_instances = &s->rev; + + /* Aportion the memory, and reduce cache if it is over new limit. */ + g_no_rev_cache_instances++; + ram_portion /= g_no_rev_cache_instances; + for (rsi = g_rev_instances; rsi != NULL; rsi = rsi->next) { + revcache *rc = rsi->cache; + + rsi->max_sz = ram_portion; + while (rc->nunlocked > 0 && rsi->sz > rsi->max_sz) { + if (decrease_revcache(rc) == 0) + break; + } +//printf("~1 rev instance ram = %d MB\n",rsi->sz/1000000); + } + + if (s->verbose) + fprintf(stdout, "%cThere %s %d rev cache instance%s with %lu Mbytes limit\n", + cr_char, + g_no_rev_cache_instances > 1 ? "are" : "is", + g_no_rev_cache_instances, + g_no_rev_cache_instances > 1 ? "s" : "", + ram_portion/1000000); + } + s->rev.rev_valid = 1; + +#ifdef DEBUG + if (fdi > 1) printf("%d cells in rev nn list\n",cellinrevlist); + if (fdi > 1) printf("%d fwd cells in rev nn list\n",fwdcells); + if (cellinrevlist > 1) printf("Avg list size = %f\n",(double)fwdcells/cellinrevlist); +#endif + + DBG(("init_revaccell finished\n")); +} + +/* Invalidate the reverse acceleration structures (section Two) */ +static void invalidate_revaccell( +rspl *s /* Pointer to rspl grid */ +) { + int e, di = s->di; + int **rpp, *rp; + + /* Invalidate the whole rev cache (Third section) */ + invalidate_revcache(s->rev.cache); + + /* Free up the contents of rev.rev[] and rev.nnrev[] */ + if (s->rev.rev != NULL) { + for (rpp = s->rev.rev; rpp < (s->rev.rev + s->rev.no); rpp++) { + if ((rp = *rpp) != NULL && --rp[2] <= 0) { + DECSZ(s, rp[0] * sizeof(int)); + free(*rpp); + *rpp = NULL; + } + } + } + if (s->rev.nnrev != NULL) { + for (rpp = s->rev.nnrev; rpp < (s->rev.nnrev + s->rev.no); rpp++) { + if ((rp = *rpp) != NULL && --rp[2] <= 0) { + DECSZ(s, rp[0] * sizeof(int)); + free(*rpp); + *rpp = NULL; + } + } + } + + if (di > 1 && s->rev.rev_valid) { + rev_struct *rsi, **rsp; + size_t ram_portion = g_avail_ram; + + /* Remove it from the linked list */ + for (rsp = &g_rev_instances; *rsp != NULL; rsp = &((*rsp)->next)) { + if (*rsp == &s->rev) { + *rsp = (*rsp)->next; + break; + } + } + + /* Aportion the memory */ + g_no_rev_cache_instances--; + + if (g_no_rev_cache_instances > 0) { + ram_portion /= g_no_rev_cache_instances; + for (rsi = g_rev_instances; rsi != NULL; rsi = rsi->next) + rsi->max_sz = ram_portion; + if (s->verbose) + fprintf(stdout, "%cThere %s %d rev cache instance%s with %lu Mbytes limit\n", + cr_char, + g_no_rev_cache_instances > 1 ? "are" : "is", + g_no_rev_cache_instances, + g_no_rev_cache_instances > 1 ? "s" : "", + ram_portion/1000000); + } + } + s->rev.rev_valid = 0; +} + +/* ====================================================== */ + +/* Initialise the rev First section, basic information that doesn't change */ +/* This is called on initial setup when s->rev.inited == 0 */ +static void make_rev_one( +rspl *s +) { + int i, j; /* Index of fwd grid point */ + int e, f, ee, ff; + int di = s->di; + int fdi = s->fdi; + int rgno, gno = s->g.no; + int argres; /* Allocation rgres, = no cells +1 */ + int rgres; + int rgres_1; /* rgres -1 == maximum base coord value */ + datao rgmin, rgmax; + + DBG(("make_rev_one called, di = %d, fdi = %d, mgres = %d\n",di,fdi,(int)s->g.mres)); + +//printf("~1 nnb = %d\n",nnb); + + s->get_out_range(s, rgmin, rgmax); /* overall output min/max */ + + /* Expand out range to encompass declared range */ + /* The declared range is assumed to be the range over which */ + /* we may want an reasonably accurate nearest reverse lookup. */ + for (f = 0; f < fdi; f++) { + if ((s->d.vl[f] + s->d.vw[f]) > rgmax[f]) + rgmax[f] = s->d.vl[f] + s->d.vw[f]; + if (s->d.vl[f] < rgmin[f]) + rgmin[f] = s->d.vl[f]; + } + + /* Expand out range slightly to allow for out of gamut points */ + for (f = 0; f < fdi; f++) { + double del = (rgmax[f] - rgmin[f]) * 0.10; /* Expand by +/- 10% */ + rgmax[f] += del; + rgmin[f] -= del; + } +//printf("~~got output range\n"); + + /* Heuristic - reverse grid acceleration resolution ? */ + /* Should this really be adapted to be constant in output space ? */ + /* (ie. make the gw aprox equal ?) Would complicate code rev accell */ + /* indexing though. */ + { + char *ev; + double gresmul = REV_ACC_GRES_MUL; /* Typically 2.0 */ + + if ((gresmul * s->g.mres) > (double)REV_ACC_GRES_LIMIT) { + gresmul = (double)REV_ACC_GRES_LIMIT/s->g.mres; /* Limit target res to typ. 43. */ + } + + /* Allow the user to override if it causes memory consumption problems */ + /* or to speed things up if more memory is available */ + if ((ev = getenv("ARGYLL_REV_ACC_GRID_RES_MULT")) != NULL) { + double mm; + mm = atof(ev); + if (mm > 0.1 && mm < 20.0) + gresmul *= mm; + } + /* Less than 4 is not functional */ + if ((rgres = (int) gresmul * s->g.mres) < 4) + rgres = 4; + } + argres = rgres+1; + s->rev.ares = argres; /* == number of verticies per side, used for construction */ + s->rev.res = rgres; /* == number of cells per side */ + rgres_1 = rgres-1; + + /* Number of elements in the rev.grid, including construction extra rows */ + for (rgno = 1, f = 0; f < fdi; f++, rgno *= argres); + s->rev.no = rgno; + +//printf("~1 argres = %d\n",argres); + /* Compute coordinate increments */ + s->rev.coi[0] = 1; +//printf("~1 coi[0] = %d\n",s->rev.coi[0]); + for (f = 1; f < fdi; f++) { + s->rev.coi[f] = s->rev.coi[f-1] * argres; +//printf("~1 coi[%d] = %d\n",f,s->rev.coi[f]); + } + + /* Compute index offsets from base of cube to other corners. */ + + for (s->rev.hoi[0] = f = 0, j = 1; f < fdi; j *= 2, f++) { + for (i = 0; i < j; i++) + s->rev.hoi[j+i] = s->rev.hoi[i] + s->rev.coi[f]; /* In grid points */ + } +//for (ff = 0; ff < (1 << fdi); ff++) +//printf("~1 hoi[%d] = %d\n",ff,s->rev.hoi[ff]); + + /* Conversion from output value to cell indexes */ + for (f = 0; f < fdi; f++) { + s->rev.gl[f] = rgmin[f]; + s->rev.gh[f] = rgmax[f]; + s->rev.gw[f] = (rgmax[f] - rgmin[f])/(double)rgres; + } + + if ((s->rev.rev = (int **) rev_calloc(s, rgno, sizeof(int *))) == NULL) + error("rspl malloc failed - rev.grid points"); + INCSZ(s, rgno * sizeof(int *)); + + if ((s->rev.nnrev = (int **) rev_calloc(s, rgno, sizeof(int *))) == NULL) + error("rspl malloc failed - rev.nngrid points"); + INCSZ(s, rgno * sizeof(int *)); + + s->rev.inited = 1; + + s->rev.stouch = 1; + + DBG(("make_rev_one finished\n")); +} + +/* ====================================================== */ + +/* First section of rev_struct init. */ +/* Initialise the reverse cell cache, sub simplex information */ +/* and reverse lookup acceleration structures. */ +/* This is called by a reverse interpolation call */ +/* that discovers that the reverse index list haven't */ +/* been initialised. */ +static void make_rev( +rspl *s +) { + int e, di = s->di; + char *ev; + size_t avail_ram = 256 * 1024 * 1024; /* Default assumed RAM in the system */ + size_t ram1, ram2; /* First Gig and rest */ + static int repsr = 0; /* Have we reported system RAM size ? */ + size_t max_vmem = 0; + + DBG(("make_rev called, di = %d, fdi = %d, mgres = %d\n",di,s->fdi,(int)s->g.mres)); + + /* Figure out how much RAM we can use for the rev cache. */ + /* (We compute this for each rev instance, to account for any VM */ + /* limit changes due to intervening allocations) */ + if (di > 1 || g_avail_ram == 0) { + #ifdef NT + { + BOOL (WINAPI* pGlobalMemoryStatusEx)(MEMORYSTATUSEX *) = NULL; + MEMORYSTATUSEX mstat; + + pGlobalMemoryStatusEx = (BOOL (WINAPI*)(MEMORYSTATUSEX *)) + GetProcAddress(LoadLibrary("KERNEL32"), "GlobalMemoryStatusEx"); + + if (pGlobalMemoryStatusEx == NULL) + error("Unable to link to GlobalMemoryStatusEx()"); + mstat.dwLength = sizeof(MEMORYSTATUSEX); + if ((*pGlobalMemoryStatusEx)(&mstat) != 0) { + if (sizeof(avail_ram) < 8 && mstat.ullTotalPhys > 0xffffffffL) + mstat.ullTotalPhys = 0xffffffffL; + avail_ram = mstat.ullTotalPhys; + } else { + warning("%cWarning - Unable to get system memory size",cr_char); + } + } + #else + #ifdef __APPLE__ + { + long long memsize; + size_t memsize_sz = sizeof(long long); + if (sysctlbyname("hw.memsize", &memsize, &memsize_sz, NULL, 0) == 0) { + if (sizeof(avail_ram) < 8 && memsize > 0xffffffffL) + memsize = 0xffffffff; + avail_ram = memsize; + } else { + warning("%cWarning - Unable to get system memory size",cr_char); + } + + } + #else /* Linux */ + { + long long total; + total = (long long)sysconf(_SC_PAGESIZE) * (long long)sysconf(_SC_PHYS_PAGES); + if (sizeof(avail_ram) < 8 && total > 0xffffffffL) + total = 0xffffffffL; + avail_ram = total; + } + #endif + #endif + DBG(("System RAM = %d Mbytes\n",avail_ram/1000000)); + + /* Make it sane */ + if (avail_ram < (256 * 1024 * 1024)) { + warning("%cWarning - System RAM size seems very small (%d MBytes)," + " assuming 256Mb instead",cr_char,avail_ram/1000000); + avail_ram = 256 * 1024 * 1024; + } + // avail_ram = -1; /* Fake 4GB of RAM. This will swap! */ + + ram1 = avail_ram; + ram2 = 0; + if (ram1 > (1024 * 1024 * 1024)) { + ram1 = 1024 * 1024 * 1024; + ram2 = avail_ram - ram1; + } + + /* Default maximum reverse memory (typically 50% of the first Gig, 75% of the rest) */ + g_avail_ram = (size_t)(REV_MAX_MEM_RATIO * ram1 + + REV_MAX_MEM_RATIO2 * ram2); + + /* Many 32 bit systems have a virtual memory limit, so we'd better stay under it. */ + /* This is slightly dodgy though, since we don't know how much memory other */ + /* software will need to malloc. A more sophisticated approach would be to */ + /* replace all malloc/calloc/realloc calls in the exe with a version that on failure, */ + /* sets the current memory usage as the new limit, and then */ + /* frees up some rev cache space before re-trying. This is a non-trivial change */ + /* to the source code though, and really has to include all user mode */ + /* libraries we're linked to, making implementation problematic. */ + /* Instead we do a simple test to see what the maximum allocation is, and */ + /* then use 75% of that for cache, and free cache and retry if */ + /* malloc failes in rev.c. Too bad if 25% isn't enough, and a malloc fails */ + /* outside rev.c... */ + if (sizeof(avail_ram) < 8) { + char *alocs[4 * 1024]; + size_t safe_max_vmem = 0; + int i; + +#ifdef __APPLE__ + int old_stderr, new_stderr; + + /* OS X malloc() blabs about a malloc failure. This */ + /* will confuse users, so we temporarily redirect stdout */ + fflush(stderr); + old_stderr = dup(fileno(stderr)); + new_stderr = open("/dev/null", O_WRONLY | O_APPEND); + dup2(new_stderr, fileno(stderr)); +#endif + for (i = 0; (i < 4 * 1024);i++) { + if ((alocs[i] = malloc(1024 * 1024)) == NULL) { + break; + } + max_vmem = (i+1) * 1024 * 1024; + } + for (--i; i >= 0; i--) { + free(alocs[i]); + } +#ifdef __APPLE__ + fflush(stderr); + dup2(old_stderr, fileno(stderr)); /* Restore stderr */ + close(new_stderr); + close(old_stderr); +#endif + /* To compute a true value, we need to allow for any VM already */ + /* used by any rev instances. */ + { + rev_struct *rsi; + + for (rsi = g_rev_instances; rsi != NULL; rsi = rsi->next) + max_vmem += rsi->sz; + } + +//fprintf(stdout,"~ Abs max VM = %d Mbytes\n",max_vmem/1000000); + safe_max_vmem = (size_t)(0.85 * max_vmem); + if (g_avail_ram > safe_max_vmem) { + g_avail_ram = safe_max_vmem; + if (s->verbose && repsr == 0) + fprintf(stdout,"%cTrimmed maximum cache RAM to %lu Mbytes to allow for VM limit\n",cr_char,g_avail_ram/1000000); + } + } + + /* Check for environment variable tweak */ + if ((ev = getenv("ARGYLL_REV_CACHE_MULT")) != NULL) { + double mm, gg; + mm = atof(ev); + if (mm < 0.01) /* Make it sane */ + mm = 0.01; + else if (mm > 100.0) + mm = 100.0; + gg = g_avail_ram * mm + 0.5; + if (gg > (double)(((size_t)0)-1)) + gg = (double)(((size_t)0)-1); + g_avail_ram = (size_t)(gg); + } + if (max_vmem != 0 && g_avail_ram > max_vmem && repsr == 0) { + g_avail_ram = (size_t)(0.95 * max_vmem); + fprintf(stdout,"%cARGYLL_REV_CACHE_MULT * RAM trimmed to %lu Mbytes to allow for VM limit\n",cr_char,g_avail_ram/1000000); + } + } + + /* Default - this will get aportioned as more instances appear */ + s->rev.max_sz = g_avail_ram; + + DBG(("reverse cache max memory = %d Mbytes\n",s->rev.max_sz/1000000)); + if (s->verbose && repsr == 0) { + fprintf(stdout, "%cRev cache RAM = %lu Mbytes\n",cr_char,g_avail_ram/1000000); + repsr = 1; + } + + /* Sub-simplex information for each sub dimension */ + for (e = 0; e <= di; e++) { + if (s->rev.sspxi[e].spxi != NULL) /* Assert */ + error("rspl rev, internal, init_ssimplex_info called on already init'd\n"); + + rspl_init_ssimplex_info(s, &s->rev.sspxi[e], e); + } + + make_rev_one(s); + + /* Reverse cell cache allocation */ + s->rev.cache = alloc_revcache(s); + + DBG(("make_rev finished\n")); +} + +/* ====================================================== */ + +#if defined(DEBUG1) || defined(DEBUG2) + +/* Utility - return a string containing a cells output value range */ +static char *pcellorange(cell *c) { + static char buf[5][300]; + static ix = 0; + char *bp; + rspl *s = c->s; + int di = s->di, fdi = s->fdi; + int ee, e, f; + + datao min, max; + +// double p[POW2MXRI][MXRI]; /* Vertex input positions for this cube. */ +// double v[POW2MXRI][MXRO+1]; /* Vertex data for this cube. Copied to x->v[] */ +// /* v[][fdi] is the ink limit values, if relevant */ + + for (f = 0; f < fdi; f++) { + min[f] = 1e60; + max[f] = -1e60; + } + + /* For all other grid points in the cube */ + for (ee = 0; ee < (1 << di); ee++) { + + /* Update bounding box for this grid point */ + for (f = 0; f < fdi; f++) { + if (min[f] > c->v[ee][f]) + min[f] = c->v[ee][f]; + if (max[f] < c->v[ee][f]) + max[f] = c->v[ee][f]; + } + } + if (++ix >= 5) + ix = 0; + bp = buf[ix]; + + for (e = 0; e < fdi; e++) { + if (e > 0) + *bp++ = ' '; + sprintf(bp, "%f:%f", min[e],max[e]); bp += strlen(bp); + } + return buf[ix]; +} + +#endif +/* ====================================================== */ + +#undef DEBUG +#undef DBGV +#undef DBG +#define DBGV(xxx) +#define DBG(xxx) + + + + + + -- cgit v1.2.3