cgen.c

这是一个高速多维插值算法。当我们建模以后

	/* Declare the input pointers and init them */
	for (e = 0; e < f->nip; e++) {
		line(f, "%s *ip%d = (%s *)inp[%d];",
		     a->ords[f->ipt[e]].name, e, a->ords[f->ipt[e]].name, e);
	}

	/* Declare the output pointers and init them */
	for (e = 0; e < f->nop; e++) {
		line(f, "%s *op%d = (%s *)outp[%d];",
		     a->ords[f->opt[e]].name, e, a->ords[f->opt[e]].name, e);
	}

	/* Declare and intialise the end pointer */
	line(f, "%s *ep = ip0 + npix * %d ;",
	        a->ords[f->ipt[0]].name, g->in.chi[0]);

	/* Declare and initialise the input table pointers */
	for (e = 0; e < g->id; e++)
		line(f,"pointer it%d = (pointer)p->in_tables[%d];",e,e);

	/* Declare and initialise the output table pointers */
	for (e = 0; e < g->od; e++)
		line(f,"pointer ot%d = (pointer)p->out_tables[%d];",e,e);

	if (!t->sort) {
		/* Declare and initialise the Simplex weighting base pointer */
		line(f,"pointer sw_base = (pointer)p->sw_table;");
	}

	/* Declare and initialise the Interpolation multidim base pointer */
	line(f,"pointer im_base = (pointer)p->im_table;");

	/* Figure out whether input channel reads can be used directly as table offsets */
	t->it_ix = 1;				/* Default use input table lookup to extract value */

	if (g->in.packed != 0)
		t->it_ix = 0;				/* Extract will be done explicitly */

	for (e = 0; e < g->id; e++) {
		int ee = (g->in.pint != 0) ? 0 : e;		/* bpch index */

		if ((g->in.bov[e] + g->in.bpv[e]) <= 12)
			continue;							/* Table can do extract */

		if (g->in.bov[e] != 0 || g->in.bpv[e] != g->in.bpch[ee]) {
			t->it_ix = 0;						/* Extract will be done explicitly */
			break;
		}
	}
	
	/* ------------------------------- */
#ifdef VERBOSE
	printf("Starting pixel processing loop\n"); fflush(stdout);
#endif /* VERBOSE */

	/* Start the pixel processing loop */
	cr(f);
	sline(f, "for(;ip0 < ep;");
	for (e = 0; e < f->nip; e++)
		mline(f, " ip%d += %d,", e, g->in.chi[e]);
	for (e = 0; e < f->nop; e++)
		mline(f, " op%d += %d%s", e, g->out.chi[e], ((e+1) < f->nop) ? "," : "");
	eline(f, ") {");
	inc(f);

	/* Declare output value accumulator(s) */
	for (i = 0; i < t->im_fn; i++) {
		line(f,"%s ova%d;	/* Output value accumulator */",a->ords[f->iafvt].name,i);
	}
	for (; i < f->ian; i++) {
		line(f,"%s ova%d;	/* Output value partial accumulator */",a->ords[f->iapvt].name,i);
	}

	/* Context around interp/Simplex table lookup */
	line(f, "{");
	inc(f);

	if (!t->sort)
		line(f,"pointer swp;");		/* Declare Simplex weighting pointer */
	line(f,"pointer imp;");			/* Declare Interpolation multidim pointer */

	/* Declare the input weighting/vertex offset variables */
	if (t->sort) {
		for (e = 0; e < g->id; e++) {
			if (t->wo_xs) {
				line(f,"%s we%d;	/* Weighting value variable */",
				       a->ords[f->wevt].name, e);
				line(f,"%s vo%d;	/* Vertex offset variable */",
				       a->ords[f->vovt].name, e);
			} else {
				line(f,"%s wo%d;	/* Weighting value and vertex offset variable */",
				       a->ords[f->wovt].name, e);
			}
		}
	}

	/* Context around input table processing */
	line(f, "{");
	inc(f);

	/* Declare the table index variables/input weighting/vertex offset variables */
	if (t->sort) {
		if (!t->it_xs)
			line(f,"%s ti;		/* Input table entry variable */",a->ords[f->itvt].name);
		line(f,"%s ti_i;	/* Interpolation index variable */",a->ords[f->ixvt].name);
	} else {
		if (t->it_xs) {
			line(f,"%s ti_s;	/* Simplex index variable */",a->ords[f->sxvt].name);
			line(f,"%s ti_i;	/* Interpolation index variable */",a->ords[f->ixvt].name);
		} else {
			line(f,"%s ti;	/* Simplex+Interpolation index variable */",a->ords[f->itvt].name);
		}
	}

	if (g->in.packed != 0)	/* We need to unpack from a single read */
		line(f,"%s rdv;		/* Read value */",a->ords[f->ipt[0]].name);

	if (t->it_ix == 0) {
		int bv = 0;
		for (e = 0; e < f->nip; e++) {	/* Find largest input type */
			if (f->ipt[e] > bv)
				bv = f->ipt[e];
		}
		bv = nord(f, bv);
		line(f,"%s chv;	/* Channel value */",a->ords[bv].name);
		f->chv_bits = a->ords[bv].bits;
	}
	cr(f);

#ifdef VERBOSE
	printf("Read code\n"); fflush(stdout);
#endif /* VERBOSE */

	/* For all the input channels */
	for (e = 0; e < g->id; e++) {
		char rde[50];		/* Read expression */
		char toff[50];		/* Table offset expression */
		int ee = (g->in.pint != 0) ? 0 : e;		/* bpch index */

		if (g->in.pint != 0) 	/* Pixel interleaved */
			sprintf(rde,"ip0[%d]",e);	/* Offset from single pointer */
		else
			sprintf(rde,"*ip%d",e);		/* Pointer per channel */

		if (g->in.packed != 0) {
			if (e == 0)
				line(f,"rdv = %s;",rde);	/* Do single read */
			sprintf(rde,"rdv");				/* Use read value for extraction */
		}

		if (t->it_ix == 0) {
			if (g->in.bov[e] == 0 ) {				/* No offset */
				if (g->in.bpv[e] == g->in.bpch[ee])	/* No mask */
					line(f,"chv = %s;",rde);
				else								/* Just mask  */
					line(f,"chv = (%s & %s);",rde, hmask(g->in.bpv[e]));
			} else {								/* Offset */
				if ((g->in.bov[e] + g->in.bpv[e]) == g->in.bpch[ee])
					line(f,"chv = (%s >> %d);",rde, g->in.bov[e]);
				else {								/* Offset and mask */
					if (a->shfm || g->in.bpv[e] > 32) {
						/* Extract using just shifts */
						line(f,"chv = ((%s << %d) >> %d);", rde,
						        f->chv_bits - g->in.bpv[e] - g->in.bov[e],
						        f->chv_bits - g->in.bpv[e]);
					} else {
						/* Extract using shift and mask */
						line(f,"chv = ((%s >> %d) & %s);",
						        rde, g->in.bov[e], hmask(g->in.bpv[e]));
					}
				}
			}
			sprintf(toff,"chv");
		} else {									/* No extraction */
			sprintf(toff,"%s",rde);
		}

		if (t->sort) {
			if (t->it_xs) {
				line(f,"ti_i %s= IT_IX(it%d, %s);", e ? "+" : " ", e, toff);
				if (t->wo_xs) {
					line(f,"we%d   = IT_WE(it%d, %s);", e, e, toff);
					line(f,"vo%d   = IT_VO(it%d, %s);", e, e, toff);
				} else {
					line(f,"wo%d   = IT_WO(it%d, %s);", e, e, toff);
				}
			} else {	/* All three combined */
				line(f,"ti = IT_IT(it%d, %s);", e, toff);
				if (a->shfm || t->wo_ab > 32) {
					/* Extract using just shifts */
					line(f,"wo%d   = ((ti << %d) >> %d);	"
					     "/* Extract weighting/vertex offset value */",
					     e, a->ords[f->wovt].bits - t->wo_ab, a->ords[f->wovt].bits - t->wo_ab);
					line(f,"ti_i %s= (ti >> %d);	"
					     "/* Extract interpolation table value */",
					     e ? "+" : " ", t->wo_ab);
				} else {
					/* Extract using shift and mask */
					line(f,"wo%d   = (ti & %s);	"
					     "/* Extract weighting/vertex offset value */",
					     e, hmask(t->wo_ab));
					line(f,"ti_i %s= (ti >> %d);	"
					     "/* Extract interpolation table value */",
					     e ? "+" : " ", t->wo_ab);
				}
			}

		} else {	/* Simplex */
			if (t->it_xs) {
				/* ~~~~ should toff be forced to be a temp variable ?? */
				/* (ie. force use of rde (above) if t->it_xs is nonz) */ 
				line(f,"ti_i %s= IT_IX(it%d, %s);", e ? "+" : " ", e, toff);
				line(f,"ti_s %s= IT_SX(it%d, %s);", e ? "+" : " ", e, toff);
			} else {
				line(f,"ti %s= IT_IT(it%d, %s);", e ? "+" : " ", e, toff);
			}
		}
	}

#ifdef VERBOSE
	printf("Index extraction code\n"); fflush(stdout);
#endif /* VERBOSE */

	cr(f);

	if (t->sort) {
		/* Extract Simplex and Interpolation indexes from accumulator */
		line(f,"imp = im_base + IM_O(ti_i);		/* Compute interp. table entry pointer */");
	} else {
		if (t->it_xs) {		/* Extract Simplex and Interpolation indexes from accumulator */
			line(f,"swp = sw_base + SW_O(ti_s);		/* Compute simplex table entry pointer */");
			line(f,"imp = im_base + IM_O(ti_i);		/* Compute interp. table entry pointer */");
		} else {
			line(f,"imp = im_base + IM_O(ti >> %d);		"
			     "/* Extract interp. index and comp. entry */",
			     t->sx_ab);
			if (a->shfm || t->sx_ab > 32) {
				/* Extract using just shifts */
				line(f,"swp = sw_base + SW_O((ti << %d) >> %d);	"
				     "/* Extract simplex index & comp. entry */",
				     a->ords[f->itvt].bits - t->sx_ab, a->ords[f->itvt].bits - t->sx_ab);
			} else {
				/* Extract using shift and mask */
				line(f,"swp = sw_base + SW_O(ti & %s);	"
				     "/* Extract simplex index and comp. entry */",
				     hmask(t->sx_ab));
			}
		}
	}

	/* Do the explicit sort now */
	if (t->sort) {
		cr(f);
		/* Sort from largest to smallest using a selection sort */
		/* Use simple sequence for the moment. More elaborate sequence */
		/* may allow other optimisations. */
		line(f,"/* Sort weighting values and vertex offset values */");
		for (i = 0; i < (g->id-1); i++) {
			for (e = i+1; e < g->id; e++) {
				if (t->wo_xs)
					line(f,"CEX(we%d, vo%d, we%d, vo%d);",i,i,e,e);
				else
					line(f,"CEX(wo%d, wo%d);",i,e);
			}
		}
	}

	/* End of input table processing context */
	dec(f);
	line(f,"}");

	line(f,"{");	/* Context around vertex lookup and accumulation */
	inc(f);

	/* Declare vertex offset and weight variables */
	if (t->sort && t->wo_xs == 0) {
		line(f,"%s nvof;	/* Next vertex offset value */",a->ords[f->vovt].name);
	} else {
		if (!t->wo_xs)	/* If combined in table */
			line(f,"%s vowr;	/* Vertex offset/weight value */",a->ords[f->wovt].name);
	}
	line(f,"%s vof;	/* Vertex offset value */",a->ords[f->vovt].name);
	line(f,"%s vwe;	/* Vertex weighting */",a->ords[f->wevt].name);
	if (timp && t->im_fn > 1) 
		line(f,"pointer timp;		/* Temporary interpolation table pointer */");
	cr(f);

#ifdef VERBOSE
	printf("Vertex offset and weight code\n"); fflush(stdout);
#endif /* VERBOSE */

	/* For each vertex in the simplex */
	for (e = 0; e < (g->id +1); e++) {

		if (t->sort) {

			if (e == 0) {
				line(f,"vof = 0;				/* First vertex offset is 0 */");
			} else {
				if (t->wo_xs)
					line(f,"vof += vo%d;			/* Move to next vertex */",e-1);
				else
					line(f,"vof += nvof;			/* Move to next vertex */");
			}

			/* Extract the vertex offset and weight values from the sorted input values */
			if (e < g->id && !t->wo_xs) {
				if (a->shfm || t->vo_ab > 32) {
					/* Extract using just shifts */
					line(f,"nvof = ((wo%d << %d) >> %d);	"
					     "/* Extract offset value */",
					     e, a->ords[f->vovt].bits - t->vo_ab, a->ords[f->vovt].bits - t->vo_ab);
					line(f,"wo%d = (wo%d >> %d);	"
					     "	/* Extract weighting value */",
					     e, e, t->vo_ab);
				} else {
					/* Extract using shift and mask */
					line(f,"nvof = (wo%d & %s);	"
					     "/* Extract offset value */",
					     e, hmask(t->vo_ab));
					line(f,"wo%d = (wo%d >> %d);	"
					     "	/* Extract weighting value */",
					     e, e, t->vo_ab);
				}
			}
			/* Compute the weighting value */
			if (!t->wo_xs) {
				if (e == 0) {
					line(f,"vwe = %d - wo%d;		/* Baricentric weighting */", 1 << g->prec, e);
				} else if (e < g->id) {
					line(f,"vwe = wo%d - wo%d;		/* Baricentric weighting */", e-1, e);
				} else {
					line(f,"vwe = wo%d;				/* Baricentric weighting */", e-1);
				}
			} else {
				if (e == 0) {
					line(f,"vwe = %d - we%d;		/* Baricentric weighting */", 1 << g->prec, e);
				} else if (e < g->id) {
					line(f,"vwe = we%d - we%d;		/* Baricentric weighting */", e-1, e);
				} else {
					line(f,"vwe = we%d;				/* Baricentric weighting */", e-1);
				}
			}

		} else {	/* Not sort */
			/* Read the vertex offset and weight values from the simplex table */
			if (t->wo_xs) { 			/* If separate */
				line(f,"vof = SX_VO(swp, %d);	/* Read vertex offset value */", e);
				line(f,"vwe = SX_WE(swp, %d);	/* Read vertex weighting value */", e);
			} else { 			/* If combined in table */
				line(f,"vowr = SX_WO(swp, %d);	/* Read vertex offset+weighting values */", e);
				if (a->shfm || t->vo_ab > 32) {
					/* Extract using just shifts */
					line(f,"vof = ((vowr << %d) >> %d);	"
					     "/* Extract offset value */",
					     a->ords[f->vovt].bits - t->vo_ab, a->ords[f->vovt].bits - t->vo_ab);
					line(f,"vwe = (vowr >> %d);	"
					     "/* Extract weighting value */",
					     t->vo_ab);
				} else {
					/* Extract using shift and mask */
					line(f,"vof = (vowr & %s);	"
					     "/* Extract offset value */",
					     hmask(t->vo_ab));
					line(f,"vwe = (vowr >> %d);	"
					     "/* Extract weighting value */",
					     t->vo_ab);
				}
			}
		}

		/* Lookup the vertex value, weight it, and accumulate it into output value */
		if (timp && t->im_fn > 1) 
			line(f,"timp = IM_TP(imp, vof);	/* Vertex address */");
		for (i = 0; i < f->ian; i++) {		/* For each output accumulation chunk */
			if (i < t->im_fn) { 	/* Full entry */
				if (!timp || t->im_fn == 1) 
					line(f,"ova%d %s= IM_FE(imp, vof, %d) * vwe;	"
					     "/* Accumulate weighted output values */",
					     i, e ? "+" : " ", i);
				else
					line(f,"ova%d %s= IM_FE(timp, %d) * vwe;	"
					     "/* Accumulate weighted output values */",
					     i, e ? "+" : " ", i);
			} else				/* One partial entry */
				line(f,"ova%d %s= IM_PE(imp, vof) * vwe;	"
				     "/* Accumulate last weighted output values */",
				     i, e ? "+" : " ");
		}
	}

	dec(f);
	line(f, "}"); 	/* End of output value lookup context */

	dec(f);
	line(f, "}"); 	/* End of output value accumulation context */

	/* Start of output lookup and write */
	line(f,"{");
	inc(f);

#ifdef VERBOSE
	printf("Output table code\n"); fflush(stdout);
#endif /* VERBOSE */

	{
		char wre[50];		/* Write destination expression */

		if (g->out.packed != 0)	/* We need to pack results into a single write */
			line(f,"%s wrv;		/* Write value */",a->ords[f->ipt[0]].name);
	
		/* Declare temporary to hold index into output lookup table */
		line(f,"%s oti;	/* Vertex offset value */",a->ords[f->otit].name);
	
		/* For each accumulator value */
		/* (Assume they are in output order for the moment ?) */
		for (e = i = 0; i < f->ian; i++) {		/* For each output accumulation chunk */
			int vpa = i < t->im_fn ? t->im_fv : t->im_pv;		/* Chanel values per accumulator */
			int oat = i < t->im_fn ? f->iafvt : f->iapvt;		/* Output accumulator type */
			int ee;		/* Relative e to this accumulator */
	
			/* For each output value in this accumulator */
			for (ee = 0; ee < vpa && e < g->od; ee++, e++) {
				int off, size;		/* Bits to be extracted */
		
				/* Extract wanted 8 bits from the 8.8 bit result in accumulator */
				off = ee * f->iaovb + (f->iaovb - g->prec);	
				size = g->prec;
	
				if (e == 0 || g->out.packed == 0) {
					if (g->out.pint != 0) 	/* Pixel interleaved */
						sprintf(wre,"op0[%d]",e);	/* Offset from single pointer */
					else
						sprintf(wre,"*op%d",e);		/* Pointer per channel */
				}
	
				if (a->shfm || size > 32) {
					/* Extract using just shifts */
					line(f,"oti = ((ova%d << %d) >> %d);	"
					     "/* Extract integer part of result */",
					     i, a->ords[oat].bits - off - size, a->ords[oat].bits - size);
				} else {
					/* Extract using shift and mask */
					line(f,"oti = ((ova%d >> %d) & %s);	"
					     "/* Extract integer part of result */",
					     i, off, hmask(size));
				}
	
				/* Lookup in output table and write to destination */
				if (g->out.packed != 0) {
					line(f,"wrv %s= OT_E(ot%d, oti);", e ? "+" : "", e);
				} else {
					line(f,"%s = OT_E(ot%d, oti);	/* Write result */", wre, e);
				}
			}
		}
	
		if (g->out.packed != 0) {	/* Write out the accumulated value */
			line(f,"%s = wrv;	/* Write result */", wre);
		}
	}

	/* The end of the output lookup and write */
	dec(f);
	line(f, "}");