2.5: Texture Filtering

Patch by Alfredo de Greef with high quality image texture filters.
This adds 3 new filters:

* SAT: Summed Area Tables. This is like mipmaps, but using somewhat
  more memory avoids some artifacts.
* EWA: Ellipitical Weighted Average, anisotropic filter.
* FELINE: Fast elliptical lines for anisotropic texture mapping.

The one change I made to this was to try to fix an alpha/premul
problem, hopefully I didn't break anything, it looks compatible
with the existing filter now for me.

2 files changed, 88 insertions, 2 deletions

diff --git a/source/blender/imbuf/IMB_imbuf.h b/source/blender/imbuf/IMB_imbuf.h
index 1d8035a2358..d2f561438b9 100644
--- a/source/blender/imbuf/IMB_imbuf.h
+++ b/source/blender/imbuf/IMB_imbuf.h
@@ -319,7 +319,7 @@ void IMB_antialias(struct ImBuf * ibuf);
 void IMB_filter(struct ImBuf *ibuf);
 void IMB_filterN(struct ImBuf *out, struct ImBuf *in);
 void IMB_filter_extend(struct ImBuf *ibuf, char *mask);
-void IMB_makemipmap(struct ImBuf *ibuf, int use_filter);
+void IMB_makemipmap(struct ImBuf *ibuf, int use_filter, int SAT);
 
 /**
  *
diff --git a/source/blender/imbuf/intern/filter.c b/source/blender/imbuf/intern/filter.c
index 9802405fd8d..cc3315c7696 100644
--- a/source/blender/imbuf/intern/filter.c
+++ b/source/blender/imbuf/intern/filter.c
@@ -371,6 +371,7 @@ void IMB_filter_extend(struct ImBuf *ibuf, char *mask)
 	}
 }
 
+#if 0
 void IMB_makemipmap(ImBuf *ibuf, int use_filter)
 {
 	ImBuf *hbuf= ibuf;
@@ -394,5 +395,90 @@ void IMB_makemipmap(ImBuf *ibuf, int use_filter)
 		minsize= hbuf->x<hbuf->y?hbuf->x:hbuf->y;
 	}
 }
+#endif
 
-
+void IMB_makemipmap(ImBuf *ibuf, int use_filter, int SAT)
+{
+	if (SAT) {
+		// to maximize precision subtract image average, use intermediate double SAT,
+		// only convert to float at the end
+		const double dv = 1.0/255.0;
+		double avg[4] = {0, 0, 0, 0};
+		const int x4 = ibuf->x << 2;
+		int x, y, i;
+		ImBuf* sbuf = IMB_allocImBuf(ibuf->x, ibuf->y, 32, IB_rectfloat, 0);
+		double *satp, *satbuf = MEM_callocN(sizeof(double)*ibuf->x*ibuf->y*4, "tmp SAT buf");
+		const double mf = ibuf->x*ibuf->y;
+		float* fp;
+		ibuf->mipmap[0] = sbuf;
+		if (ibuf->rect_float) {
+			fp = ibuf->rect_float;
+			for (y=0; y<ibuf->y; ++y)
+				for (x=0; x<ibuf->x; ++x) {
+					avg[0] += *fp++;
+					avg[1] += *fp++;
+					avg[2] += *fp++;
+					avg[3] += *fp++;
+				}
+		}
+		else {
+			char* cp = (char*)ibuf->rect;
+			for (y=0; y<ibuf->y; ++y)
+				for (x=0; x<ibuf->x; ++x) {
+					avg[0] += *cp++ * dv;
+					avg[1] += *cp++ * dv;
+					avg[2] += *cp++ * dv;
+					avg[3] += *cp++ * dv;
+				}
+		}
+		avg[0] /= mf;
+		avg[1] /= mf;
+		avg[2] /= mf;
+		avg[3] /= mf;
+		for (y=0; y<ibuf->y; ++y)
+			for (x=0; x<ibuf->x; ++x) {
+				const unsigned int p = (x + y*ibuf->x) << 2;
+				char* cp = (char*)ibuf->rect + p;
+				fp = ibuf->rect_float + p;
+				satp = satbuf + p;
+				for (i=0; i<4; ++i, ++cp, ++fp, ++satp) {
+					double sv = (ibuf->rect_float ? (double)*fp : (double)(*cp)*dv) - avg[i];
+					if (x > 0) sv += satp[-4];
+					if (y > 0) sv += satp[-x4];
+					if (x > 0 && y > 0) sv -= satp[-x4 - 4];
+					*satp = sv;
+				}
+			}
+		fp = sbuf->rect_float;
+		satp = satbuf;
+		for (y=0; y<ibuf->y; ++y)
+			for (x=0; x<ibuf->x; ++x) {
+				*fp++ = (float)*satp++;
+				*fp++ = (float)*satp++;
+				*fp++ = (float)*satp++;
+				*fp++ = (float)*satp++;
+			}
+		MEM_freeN(satbuf);
+		fp = &sbuf->rect_float[(sbuf->x - 1 + (sbuf->y - 1)*sbuf->x) << 2];
+		fp[0] = avg[0];
+		fp[1] = avg[1];
+		fp[2] = avg[2];
+		fp[3] = avg[3];
+	}
+	else {
+		ImBuf *hbuf = ibuf;
+		int curmap = 0;
+		while (curmap < IB_MIPMAP_LEVELS) {
+			if (use_filter) {
+				ImBuf *nbuf= IMB_allocImBuf(hbuf->x, hbuf->y, 32, IB_rect, 0);
+				IMB_filterN(nbuf, hbuf);
+				ibuf->mipmap[curmap] = IMB_onehalf(nbuf);
+				IMB_freeImBuf(nbuf);
+			}
+			else ibuf->mipmap[curmap] = IMB_onehalf(hbuf);
+			hbuf = ibuf->mipmap[curmap];
+			if (hbuf->x == 1 && hbuf->y == 1) break;
+			curmap++;
+		}
+	}
+}