X-Git-Url: https://git.cinelerra-gg.org/git/?p=goodguy%2Fcinelerra.git;a=blobdiff_plain;f=cinelerra-5.1%2Fcinelerra%2Fmaskengine.C;h=589b7938a875b9dc27a1aba95ee193d91dfa5456;hp=fd7458e1bb54d9eb9bcbb44b8e5a00bc926f5350;hb=592d502d68e409919fb36a31ba60c23f57b00f35;hpb=c857b2fb7965d27d86d5785fb9f1b8957a871a1a

diff --git a/cinelerra-5.1/cinelerra/maskengine.C b/cinelerra-5.1/cinelerra/maskengine.C
index fd7458e1..589b7938 100644
--- a/cinelerra-5.1/cinelerra/maskengine.C
+++ b/cinelerra-5.1/cinelerra/maskengine.C
@@ -33,6 +33,22 @@
 #include <stdint.h>
 #include <string.h>
 
+void write_mask(VFrame *vfrm, const char *fmt, ...)
+{
+  va_list ap;    va_start(ap, fmt);
+  char fn[256];  vsnprintf(fn, sizeof(fn), fmt, ap);
+  va_end(ap);
+  FILE *fp = !strcmp(fn,"-") ? stdout : fopen(fn,"w");
+  if( fp ) {
+    int w = vfrm->get_w(), h = vfrm->get_h();
+    int m = vfrm->get_color_model();
+    fprintf(fp,"P5\n%d %d\n%d\n",w,h,m==BC_A8? 0xff : 0xffff);
+    int bpp = m==BC_A8? 1 : 2;
+    fwrite(vfrm->get_data(),bpp*w,h,fp);  fflush(fp);
+    if( fp != stdout ) fclose(fp);
+  }
+}
+
 MaskPackage::MaskPackage()
 {
 }
@@ -45,469 +61,253 @@ MaskUnit::MaskUnit(MaskEngine *engine)
  : LoadClient(engine)
 {
 	this->engine = engine;
-	this->temp = 0;
+	spot = 0;
+	r = 0;
 }
 
 MaskUnit::~MaskUnit()
 {
-	if( temp ) delete temp;
 }
 
-
-#define OVERSAMPLE 8
-
-void MaskUnit::draw_line_clamped(VFrame *frame,
-	int x1, int y1, int x2, int y2, unsigned char k)
+void MaskUnit::draw_line(int v, int ix1, int iy1, int ix2, int iy2)
 {
-	int draw_x1, draw_y1;
-	int draw_x2, draw_y2;
-
-	if( y2 < y1 ) {
-		draw_x1 = x2;  draw_y1 = y2;
-		draw_x2 = x1;  draw_y2 = y1;
-	}
-	else {
-		draw_x1 = x1;  draw_y1 = y1;
-		draw_x2 = x2;  draw_y2 = y2;
-	}
-
-	unsigned char **rows = (unsigned char**)frame->get_rows();
-
-	if( draw_y2 != draw_y1 ) {
-		float slope = ((float)draw_x2 - draw_x1) / ((float)draw_y2 - draw_y1);
-		int w = frame->get_w() - 1;
-		int h = frame->get_h();
-
-		for( float y = draw_y1; y < draw_y2; y++ ) {
-			if( y >= 0 && y < h ) {
-				int x = (int)((y - draw_y1) * slope + draw_x1);
-				int y_i = (int)y;
-				int x_i = CLIP(x, 0, w);
-
-				if( rows[y_i][x_i] == k )
-					rows[y_i][x_i] = 0;
-				else
-					rows[y_i][x_i] = k;
-			}
-		}
+	if( iy1 == iy2 ) return;
+	int x1 = iy1 < iy2 ? ix1 : ix2;
+	int y1 = iy1 < iy2 ? iy1 : iy2;
+	int x2 = iy1 < iy2 ? ix2 : ix1;
+	int y2 = iy1 < iy2 ? iy2 : iy1;
+	float slope = (float)(x2-x1) / (y2-y1);
+	int dy = y1 - start_y;
+	int i = dy < 0 ? (y1=start_y, -dy) : 0;
+	if( y2 > end_y ) y2 = end_y;
+	if( y2 < start_y || y1 >= end_y ) return;
+
+	VFrame *temp = engine->temp;
+	int w1 = temp->get_w()-1;
+	temp_t **rows = (temp_t **)temp->get_rows();
+	for( int y=y1; y<y2; ++i,++y ) {
+		int x = (int)(i*slope + x1);
+		bclamp(x, 0, w1);
+		rows[y][x] = rows[y][x] == v ? 0 : v;
 	}
 }
 
-void MaskUnit::blur_strip(double *val_p, double *val_m,
-	double *dst, double *src, int size, int max)
+void MaskUnit::draw_fill(int v)
 {
-	double *sp_p = src;
-	double *sp_m = src + size - 1;
-	double *vp = val_p;
-	double *vm = val_m + size - 1;
-	double initial_p = sp_p[0];
-	double initial_m = sp_m[0];
-
-//printf("MaskUnit::blur_strip %d\n", size);
-	for( int k = 0; k < size; k++ ) {
-		int terms = (k < 4) ? k : 4;
-		int l;
-		for( l = 0; l <= terms; l++ ) {
-			*vp += n_p[l] * sp_p[-l] - d_p[l] * vp[-l];
-			*vm += n_m[l] * sp_m[l] - d_m[l] * vm[l];
-		}
-
-		for( ; l <= 4; l++) {
-			*vp += (n_p[l] - bd_p[l]) * initial_p;
-			*vm += (n_m[l] - bd_m[l]) * initial_m;
+	VFrame *temp = engine->temp;
+	int temp_w = temp->get_w();
+	temp_t **rows = (temp_t**)temp->get_rows();
+
+	for( int y=start_y; y<end_y; ++y ) {
+		temp_t *row = rows[y];
+		int value = 0, total = 0;
+		for( int x=0; x<temp_w; ++x )
+			if( row[x] == v ) ++total;
+		if( total < 2 ) continue;
+		if( total & 0x1 ) --total;
+		for( int x=0; x<temp_w; ++x ) {
+			if( row[x]==v && total>0 ) {
+				--total;
+				value = value ? 0 : v;
+			}
+			else if( value )
+				row[x] = value;
 		}
-		sp_p++;  sp_m--;
-		vp++;    vm--;
-	}
-
-	for( int i = 0; i < size; i++ ) {
-		double sum = val_p[i] + val_m[i];
-		CLAMP(sum, 0, max);
-		dst[i] = sum;
 	}
 }
 
-void MaskUnit::do_feather(VFrame *output, VFrame *input, double feather,
-		int start_y, int end_y, int start_x, int end_x)
+void MaskUnit::draw_feather(int ix1,int iy1, int ix2,int iy2)
 {
-//printf("MaskUnit::do_feather %f\n", feather);
-// Get constants
-	double constants[8];
-	double div;
-	double std_dev = sqrt(-(double)(feather * feather) / (2 * log(1.0 / 255.0)));
-	div = sqrt(2 * M_PI) * std_dev;
-	constants[0] = -1.783 / std_dev;
-	constants[1] = -1.723 / std_dev;
-	constants[2] = 0.6318 / std_dev;
-	constants[3] = 1.997  / std_dev;
-	constants[4] = 1.6803 / div;
-	constants[5] = 3.735 / div;
-	constants[6] = -0.6803 / div;
-	constants[7] = -0.2598 / div;
-
-	n_p[0] = constants[4] + constants[6];
-	n_p[1] = exp(constants[1]) *
-				(constants[7] * sin(constants[3]) -
-				(constants[6] + 2 * constants[4]) * cos(constants[3])) +
-				exp(constants[0]) *
-				(constants[5] * sin(constants[2]) -
-				(2 * constants[6] + constants[4]) * cos(constants[2]));
-
-	n_p[2] = 2 * exp(constants[0] + constants[1]) *
-				((constants[4] + constants[6]) * cos(constants[3]) *
-				cos(constants[2]) - constants[5] *
-				cos(constants[3]) * sin(constants[2]) -
-				constants[7] * cos(constants[2]) * sin(constants[3])) +
-				constants[6] * exp(2 * constants[0]) +
-				constants[4] * exp(2 * constants[1]);
-
-	n_p[3] = exp(constants[1] + 2 * constants[0]) *
-				(constants[7] * sin(constants[3]) -
-				constants[6] * cos(constants[3])) +
-				exp(constants[0] + 2 * constants[1]) *
-				(constants[5] * sin(constants[2]) - constants[4] *
-				cos(constants[2]));
-	n_p[4] = 0.0;
-
-	d_p[0] = 0.0;
-	d_p[1] = -2 * exp(constants[1]) * cos(constants[3]) -
-				2 * exp(constants[0]) * cos(constants[2]);
-
-	d_p[2] = 4 * cos(constants[3]) * cos(constants[2]) *
-				exp(constants[0] + constants[1]) +
-				exp(2 * constants[1]) + exp (2 * constants[0]);
-
-	d_p[3] = -2 * cos(constants[2]) * exp(constants[0] + 2 * constants[1]) -
-				2 * cos(constants[3]) * exp(constants[1] + 2 * constants[0]);
-
-	d_p[4] = exp(2 * constants[0] + 2 * constants[1]);
-
-	for( int i = 0; i < 5; i++ ) d_m[i] = d_p[i];
-
-	n_m[0] = 0.0;
-	for( int i = 1; i <= 4; i++ )
-		n_m[i] = n_p[i] - d_p[i] * n_p[0];
-
-	double sum_n_p, sum_n_m, sum_d;
-	double a, b;
-
-	sum_n_p = 0.0;
-	sum_n_m = 0.0;
-	sum_d = 0.0;
-	for( int i = 0; i < 5; i++ ) {
-		sum_n_p += n_p[i];
-		sum_n_m += n_m[i];
-		sum_d += d_p[i];
+	int x1 = iy1 < iy2 ? ix1 : ix2;
+	int y1 = iy1 < iy2 ? iy1 : iy2;
+	int x2 = iy1 < iy2 ? ix2 : ix1;
+	int y2 = iy1 < iy2 ? iy2 : iy1;
+	VFrame *temp = engine->temp;
+	int h = temp->get_h();
+	if( y2 < 0 || y1 >= h ) return;
+
+	int x = x1, y = y1;
+	int dx = x2-x1, dy = y2-y1;
+	int dx2 = 2*dx, dy2 = 2*dy;
+	if( dx < 0 ) dx = -dx;
+	int m = dx > dy ? dx : dy, n = m;
+	if( dy >= dx ) {
+		if( dx2 >= 0 ) do {     /* +Y, +X */
+			draw_spot(x, y++);
+			if( (m -= dx2) < 0 ) { m += dy2;  ++x; }
+		} while( --n >= 0 );
+		else do {	       /* +Y, -X */
+			draw_spot(x, y++);
+			if( (m += dx2) < 0 ) { m += dy2;  --x; }
+		} while( --n >= 0 );
 	}
-
-	a = sum_n_p / (1 + sum_d);
-	b = sum_n_m / (1 + sum_d);
-
-	for( int i = 0; i < 5; i++ ) {
-		bd_p[i] = d_p[i] * a;
-		bd_m[i] = d_m[i] * b;
+	else {
+		if( dx2 >= 0 ) do {     /* +X, +Y */
+			draw_spot(x++, y);
+			if( (m -= dy2) < 0 ) { m += dx2;  ++y; }
+		} while( --n >= 0 );
+		else do {	       /* -X, +Y */
+			draw_spot(x--, y);
+			if( (m -= dy2) < 0 ) { m -= dx2;  ++y; }
+		} while( --n >= 0 );
 	}
-#define DO_FEATHER(type, max) { \
-	int frame_w = input->get_w(); \
-	int frame_h = input->get_h(); \
-	int size = MAX(frame_w, frame_h); \
-	double *src = new double[size]; \
-	double *dst = new double[size]; \
-	double *val_p = new double[size]; \
-	double *val_m = new double[size]; \
-	type **in_rows = (type**)input->get_rows(); \
-	type **out_rows = (type**)output->get_rows(); \
-	int j; \
- \
-/* printf("DO_FEATHER 1\n"); */ \
-	if( end_x > start_x ) { \
-		for( j = start_x; j < end_x; j++ ) { \
-	/* printf("DO_FEATHER 1.1 %d\n", j); */ \
-			bzero(val_p, sizeof(double) * frame_h); \
-			bzero(val_m, sizeof(double) * frame_h); \
-			for( int k = 0; k < frame_h; k++ ) { \
-				src[k] = (double)in_rows[k][j]; \
-			} \
-			blur_strip(val_p, val_m, dst, src, frame_h, max); \
-			for( int k = 0; k < frame_h; k++ ) { \
-				out_rows[k][j] = (type)dst[k]; \
-			} \
-		} \
-	} \
- 	if( end_y > start_y ) { \
-		for( j = start_y; j < end_y; j++ ) { \
-	/* printf("DO_FEATHER 2 %d\n", j); */ \
-			bzero(val_p, sizeof(double) * frame_w); \
-			bzero(val_m, sizeof(double) * frame_w); \
-			for( int k = 0; k < frame_w; k++ ) { \
-				src[k] = (double)out_rows[j][k]; \
-			} \
-			blur_strip(val_p, val_m, dst, src, frame_w, max); \
-			for( int k = 0; k < frame_w; k++ ) { \
-				out_rows[j][k] = (type)dst[k]; \
-			} \
-		} \
-	} \
-/* printf("DO_FEATHER 3\n"); */ \
-	delete [] src; \
-	delete [] dst; \
-	delete [] val_p; \
-	delete [] val_m; \
-/* printf("DO_FEATHER 4\n"); */ \
 }
 
-//printf("do_feather %d\n", frame->get_color_model());
-	switch( input->get_color_model() ) {
-	case BC_A8:
-		DO_FEATHER(unsigned char, 0xff);
-		break;
-	case BC_A16:
-		DO_FEATHER(uint16_t, 0xffff);
-		break;
-	case BC_A_FLOAT:
-		DO_FEATHER(float, 1);
-		break;
+void MaskUnit::draw_spot(int ix, int iy)
+{
+	int rr = r * r, n = abs(r), rv = r * v;
+	if( iy < start_y-n || iy >= end_y+n ) return;
+	VFrame *temp = engine->temp;
+	int w1 = temp->get_w()-1, h1 = temp->get_h()-1;
+	int xs = ix - n;  bclamp(xs, 0, w1);
+	int xn = ix + n;  bclamp(xn, 0, w1);
+	int ys = iy - n;  bclamp(ys, 0, h1);
+	int yn = iy + n;  bclamp(yn, 0, h1);
+
+	temp_t **rows = (temp_t**)temp->get_rows();
+	for( int y=ys ; y<=yn; ++y ) {
+		temp_t *row = rows[y];
+		for( int x=xs; x<=xn; ++x ) {
+			int dx = x-ix, dy = y-iy;
+			int dd = dx*dx + dy*dy;
+			if( dd >= rr ) continue;
+			temp_t *rp = &row[x], a = spot[dd];
+			if( rv*(*rp-a) < 0 ) *rp = a;
+		}
 	}
 }
 
 void MaskUnit::process_package(LoadPackage *package)
 {
 	MaskPackage *ptr = (MaskPackage*)package;
-	float engine_value = engine->value;
-	int engine_mode = engine->mode;
-
+	start_y = ptr->start_y;
+	end_y = ptr->end_y;
+	if( start_y >= end_y ) return;
+	mask_model = engine->mask->get_color_model();
+	VFrame *temp = engine->temp;
 	if( engine->recalculate && engine->step == DO_MASK ) {
-		VFrame *mask = engine->feather > 0 ? engine->temp_mask : engine->mask;
-// Generated oversampling frame
-		int mask_w = mask->get_w();
-		//int mask_h = mask->get_h();
-		int oversampled_package_w = mask_w * OVERSAMPLE;
-		int oversampled_package_h = (ptr->end_y - ptr->start_y) * OVERSAMPLE;
-		if( temp &&
-		    (temp->get_w() != oversampled_package_w ||
-		     temp->get_h() != oversampled_package_h) ) {
-			delete temp;  temp = 0;
+// Draw masked region of polygons on temp
+		for( int k=0; k<engine->edges.size(); ++k ) {
+			if( !engine->edges[k] ) continue;
+			MaskEdge &edge = *engine->edges[k];
+			if( edge.size() < 3 ) continue;
+			int v = k + 1;
+			for( int i=0; i<edge.size(); ++i ) {
+				MaskCoord a = edge[i];
+				MaskCoord b = i<edge.size()-1 ? edge[i+1] : edge[0];
+				draw_line(v, a.x,a.y, b.x,b.y);
+			}
+			draw_fill(v);
 		}
-		if( !temp ) {
-			temp = new VFrame(oversampled_package_w, oversampled_package_h, BC_A8, 0);
+// map temp to fader alpha
+		int temp_w = temp->get_w();
+		temp_t **rows = (temp_t**)temp->get_rows();
+		temp_t *fade = engine->fade;
+		for( int y=start_y; y<end_y; ++y ) {
+			temp_t *tp = rows[y];
+			for( int i=temp_w; --i>=0; ++tp ) *tp = fade[*tp];
 		}
-		temp->clear_frame();
-// Draw oversampled region of polygons on temp
-		for( int k=0; k<engine->point_sets.total; ++k ) {
-			int old_x, old_y;
-			unsigned char max = k + 1;
-			ArrayList<MaskPoint*> *points = engine->point_sets.values[k];
-
-			if( points->total < 3 ) continue;
-//printf("MaskUnit::process_package 2 %d %d\n", k, points->total);
-			for( int i=0; i<points->total; ++i ) {
-				MaskPoint *point1 = points->values[i];
-				MaskPoint *point2 = (i >= points->total - 1) ?
-					points->values[0] :
-					points->values[i + 1];
-
-				float x, y;
-				int segments = (int)(sqrt(SQR(point1->x - point2->x) + SQR(point1->y - point2->y)));
-				if( point1->control_x2 == 0 &&
-					point1->control_y2 == 0 &&
-					point2->control_x1 == 0 &&
-					point2->control_y1 == 0 )
-					segments = 1;
-				float x0 = point1->x;
-				float y0 = point1->y;
-				float x1 = point1->x + point1->control_x2;
-				float y1 = point1->y + point1->control_y2;
-				float x2 = point2->x + point2->control_x1;
-				float y2 = point2->y + point2->control_y1;
-				float x3 = point2->x;
-				float y3 = point2->y;
-
-				for( int j = 0; j <= segments; j++ ) {
-					float t = (float)j / segments;
-					float tpow2 = t * t;
-					float tpow3 = t * t * t;
-					float invt = 1 - t;
-					float invtpow2 = invt * invt;
-					float invtpow3 = invt * invt * invt;
-
-					x = (        invtpow3 * x0
-						+ 3 * t     * invtpow2 * x1
-						+ 3 * tpow2 * invt     * x2
-						+     tpow3            * x3);
-					y = (        invtpow3 * y0
-						+ 3 * t     * invtpow2 * y1
-						+ 3 * tpow2 * invt     * y2
-						+     tpow3            * y3);
-
-					y -= ptr->start_y;
-					x *= OVERSAMPLE;
-					y *= OVERSAMPLE;
-
-					if( j > 0 ) {
-						draw_line_clamped(temp, old_x, old_y, (int)x, (int)y, max);
-					}
-
-					old_x = (int)x;
-					old_y = (int)y;
-				}
+	}
+	if( engine->recalculate && engine->step == DO_FEATHER ) {
+// draw feather
+		for( int k=0; k<engine->edges.size(); ++k ) {
+			if( !(v = engine->faders[k]) ) continue;
+			if( !(r = engine->feathers[k]) ) continue;
+			MaskEdge &edge = *engine->edges[k];
+			if( !edge.size() ) continue;
+			float rv = r * v, vv = fabs(v);
+			int fg = 0xffff * (rv >= 0 ? vv : 0);
+			int bg = 0xffff * (rv >= 0 ? 0 : vv);
+			int rr = r*r;  double dr = 1./rr;
+			temp_t psf[rr+1];  spot = psf;
+			for( int i=0; i<=rr; ++i ) {
+				double d = i*dr;
+				psf[i] = (1-d)*fg + d*bg;
 			}
-// Fill in the polygon in the horizontal direction
-			for( int i=0; i<oversampled_package_h; ++i ) {
-				unsigned char *row = (unsigned char*)temp->get_rows()[i];
-				int value = 0, total = 0;
-
- 				for( int j=0; j<oversampled_package_w; ++j )
-					if( row[j] == max ) ++total;
-
- 				if( total > 1 ) {
-					if( total & 0x1 ) --total;
-					for( int j=0; j<oversampled_package_w; ++j ) {
-						if( row[j]==max && total>0 ) {
-							--total;
-							value = value ? 0 : max;
-						}
-						else if( value )
-							row[j] = value;
-					}
-				}
+			int n = edge.size();
+			for( int i=0; i<n; ++i ) {
+				MaskCoord &a = edge[i];
+				MaskCoord &b = i<edge.size()-1 ? edge[i+1] : edge[0];
+				draw_feather(a.x,a.y, b.x,b.y);
 			}
 		}
-#define DOWNSAMPLE(type, temp_type, value, v) do { \
-for( int y=0; y<ptr->end_y-ptr->start_y; ++y ) { \
-	type *output_row = (type*)mask->get_rows()[y + ptr->start_y]; \
-	unsigned char **input_rows = (unsigned char**)temp->get_rows() + y * OVERSAMPLE; \
-	for( int x=0; x<mask_w; ++x ) { \
-		temp_type total = 0; \
-/* Accumulate pixel */ \
-		for( int k=0; k<OVERSAMPLE; ++k ) { \
-			unsigned char *input_vector = input_rows[k] + x * OVERSAMPLE; \
-			for( int l=0; l<OVERSAMPLE; ++l ) { \
-				total += (input_vector[l] ? value : 0); \
-			} \
-		} \
-/* Divide pixel */ \
-		total /= OVERSAMPLE * OVERSAMPLE; \
-		output_row[x] = v; \
-	} \
-} } while(0)
 
-		if( engine_value < 0 ) {
-			engine_value = -engine_value;
-			engine_mode = 1-engine_mode;
+#define REMAP(cmodel, type, expr) case cmodel: { \
+type **msk_rows = (type**)engine->mask->get_rows(); \
+for( int y=start_y; y<end_y; ++y ) { \
+	temp_t *rp = rows[y]; \
+	type *mp = msk_rows[y]; \
+	for( int i=temp_w; --i>=0; ++rp,++mp ) *mp = expr; \
+} } break
+// map alpha to mask
+		const float to_flt = 1/65535.;
+		int temp_w = temp->get_w();
+		temp_t **rows = (temp_t**)temp->get_rows();
+		switch( mask_model ) {
+		REMAP(BC_A8, uint8_t, *rp >> 8);
+		REMAP(BC_A16, uint16_t, *rp);
+		REMAP(BC_A_FLOAT, float, *rp * to_flt);
 		}
-// Downsample polygon
-		switch( mask->get_color_model() ) {
-		case BC_A8: {
-			unsigned char value;
-			value = (int)(engine_value / 100 * 0xff);
-			if( engine->value >= 0 )
-				DOWNSAMPLE(unsigned char, int64_t, value, total);
-			else
-				DOWNSAMPLE(unsigned char, int64_t, value, value-total);
-			break; }
-
-		case BC_A16: {
-			uint16_t value;
-			value = (int)(engine_value / 100 * 0xffff);
-			if( engine->value >= 0 )
-				DOWNSAMPLE(uint16_t, int64_t, value, total);
-			else
-				DOWNSAMPLE(uint16_t, int64_t, value, value-total);
-			break; }
-
-		case BC_A_FLOAT: {
-			float value;
-			value = engine_value / 100;
-			if( engine->value >= 0 )
-				DOWNSAMPLE(float, double, value, total);
-			else
-				DOWNSAMPLE(float, double, value, value-total);
-			break; }
-		}
-	}
-
-	if( engine->step == DO_X_FEATHER && engine->recalculate && // Feather polygon
-	    engine->feather > 0 ) {
-		do_feather(engine->mask,
-			engine->temp_mask, engine->feather,
-			ptr->start_y, ptr->end_y, 0, 0);
-//printf("MaskUnit::process_package 3 %f\n", engine->feather);
 	}
 
-	if( engine->step == DO_Y_FEATHER && engine->recalculate && // Feather polygon
-	    engine->feather > 0 ) {
-		do_feather(engine->mask,
-			engine->temp_mask, engine->feather,
-			0, 0, ptr->start_x, ptr->end_x);
-	}
-
-	if( engine->step == DO_APPLY ) { // Apply mask
-#define APPLY_MASK_ALPHA(cmodel, type, max, components, do_yuv, a, b) \
+// Apply mask
+	if( engine->step == DO_APPLY ) {
+		int mask_w = engine->mask->get_w();
+		uint8_t **out_rows = engine->output->get_rows();
+		uint8_t **msk_rows = engine->mask->get_rows();
+#define APPLY_MASK_ALPHA(cmodel, type, max, components, do_yuv) \
 case cmodel: \
 for( int y=ptr->start_y; y<ptr->end_y; ++y ) { \
-	type *output_row = (type*)engine->output->get_rows()[y]; \
-	type *mask_row = (type*)engine->mask->get_rows()[y]; \
-	int chroma_offset = (int)(max + 1) / 2; \
+	type *out_row = (type*)out_rows[y]; \
+	type *msk_row = (type*)msk_rows[y]; \
+	type chroma_offset = (int)(max + 1) / 2; \
 	for( int x=0; x<mask_w; ++x ) { \
-		int m = mask_row[x], n = max-m; \
+		type a = msk_row[x], b = max-a; \
 		if( components == 4 ) { \
-			output_row[x*4 + 3] = output_row[x*4 + 3]*a / max; \
+			out_row[x*4 + 3] = out_row[x*4 + 3]*b / max; \
 		} \
 		else { \
-			output_row[x*3 + 0] = output_row[x*3 + 0]*a / max; \
-			output_row[x*3 + 1] = output_row[x*3 + 1]*a / max; \
-			output_row[x*3 + 2] = output_row[x*3 + 2]*a / max; \
+			out_row[x*3 + 0] = out_row[x*3 + 0]*b / max; \
+			out_row[x*3 + 1] = out_row[x*3 + 1]*b / max; \
+			out_row[x*3 + 2] = out_row[x*3 + 2]*b / max; \
 			if( do_yuv ) { \
-				output_row[x*3 + 1] += chroma_offset*b / max; \
-				output_row[x*3 + 2] += chroma_offset*b / max; \
+				out_row[x*3 + 1] += chroma_offset*a / max; \
+				out_row[x*3 + 2] += chroma_offset*a / max; \
 			} \
 		} \
 	} \
 } break
 
-#define MASK_ALPHA(mode, a, b) \
-case mode: \
-	switch( engine->output->get_color_model() ) { \
-	APPLY_MASK_ALPHA(BC_RGB888, unsigned char, 0xff, 3, 0, a, b); \
-	APPLY_MASK_ALPHA(BC_RGB_FLOAT, float, 1.0, 3, 0, a, b); \
-	APPLY_MASK_ALPHA(BC_YUV888, unsigned char, 0xff, 3, 1, a, b); \
-	APPLY_MASK_ALPHA(BC_RGBA_FLOAT, float, 1.0, 4, 0, a, b); \
-	APPLY_MASK_ALPHA(BC_YUVA8888, unsigned char, 0xff, 4, 1, a, b); \
-	APPLY_MASK_ALPHA(BC_RGBA8888, unsigned char, 0xff, 4, 0, a, b); \
-	APPLY_MASK_ALPHA(BC_RGB161616, uint16_t, 0xffff, 3, 0, a, b); \
-	APPLY_MASK_ALPHA(BC_YUV161616, uint16_t, 0xffff, 3, 1, a, b); \
-	APPLY_MASK_ALPHA(BC_YUVA16161616, uint16_t, 0xffff, 4, 1, a, b); \
-	APPLY_MASK_ALPHA(BC_RGBA16161616, uint16_t, 0xffff, 4, 0, a, b); \
-} break
-
-//printf("MaskUnit::process_package 1 %d\n", engine->mode);
-		int mask_w = engine->mask->get_w();
-		switch( engine_mode ) {
-		MASK_ALPHA(MASK_MULTIPLY_ALPHA, m, n);
-		MASK_ALPHA(MASK_SUBTRACT_ALPHA, n, m);
+		switch( engine->output->get_color_model() ) { \
+		APPLY_MASK_ALPHA(BC_RGB888, uint8_t, 0xff, 3, 0); \
+		APPLY_MASK_ALPHA(BC_RGB_FLOAT, float, 1.0, 3, 0); \
+		APPLY_MASK_ALPHA(BC_YUV888, uint8_t, 0xff, 3, 1); \
+		APPLY_MASK_ALPHA(BC_RGBA_FLOAT, float, 1.0, 4, 0); \
+		APPLY_MASK_ALPHA(BC_YUVA8888, uint8_t, 0xff, 4, 1); \
+		APPLY_MASK_ALPHA(BC_RGBA8888, uint8_t, 0xff, 4, 0); \
+		APPLY_MASK_ALPHA(BC_RGB161616, uint16_t, 0xffff, 3, 0); \
+		APPLY_MASK_ALPHA(BC_YUV161616, uint16_t, 0xffff, 3, 1); \
+		APPLY_MASK_ALPHA(BC_YUVA16161616, uint16_t, 0xffff, 4, 1); \
+		APPLY_MASK_ALPHA(BC_RGBA16161616, uint16_t, 0xffff, 4, 0); \
 		}
 	}
 }
 
 
 MaskEngine::MaskEngine(int cpus)
- : LoadServer(cpus, cpus * OVERSAMPLE * 2)
-// : LoadServer(1, OVERSAMPLE * 2)
+ : LoadServer(cpus, 2*cpus)
+// : LoadServer(1, 1)
 {
 	mask = 0;
+	temp = 0;
 }
 
 MaskEngine::~MaskEngine()
 {
-	if( mask ) {
-		delete mask;
-		delete temp_mask;
-	}
-
-	for( int i = 0; i < point_sets.total; i++ ) {
-		ArrayList<MaskPoint*> *points = point_sets.values[i];
-		points->remove_all_objects();
-	}
+	delete mask;
+	delete temp;
+	for( int i = 0; i < point_sets.total; i++ )
+		point_sets[i]->remove_all_objects();
 	point_sets.remove_all_objects();
 }
 
@@ -518,133 +318,164 @@ int MaskEngine::points_equivalent(ArrayList<MaskPoint*> *new_points,
 	if( new_points->total != points->total ) return 0;
 
 	for( int i = 0; i < new_points->total; i++ ) {
-		if( !(*new_points->values[i] == *points->values[i]) ) return 0;
+		if( !(*new_points->get(i) == *points->get(i)) ) return 0;
 	}
 
 	return 1;
 }
 
+void MaskEngine::draw_edge(MaskEdge &edge, MaskPointSet &points)
+{
+	if( points.size() < 2 ) return;
+	edge.remove_all();
+	for( int i=0; i<points.size(); ++i ) {
+		MaskPoint *ap = points[i];
+		MaskPoint *bp = (i>=points.size()-1) ?
+				points[0] : points[i+1];
+		int segments = 0;
+		if( ap->control_x2 == 0 && ap->control_y2 == 0 &&
+		    bp->control_x1 == 0 && bp->control_y1 == 0 )
+			segments = 1;
+		float x0 = ap->x, y0 = ap->y;
+		float x1 = ap->x + ap->control_x2;
+		float y1 = ap->y + ap->control_y2;
+		float x2 = bp->x + bp->control_x1;
+		float y2 = bp->y + bp->control_y1;
+		float x3 = bp->x, y3 = bp->y;
+
+// from Playback3D::do_mask_sync
+		float cx3 = -  x0 + 3*x1 - 3*x2 + x3;
+		float cx2 =  3*x0 - 6*x1 + 3*x2;
+		float cx1 = -3*x0 + 3*x1;
+		float cx0 =    x0;
+
+		float cy3 = -  y0 + 3*y1 - 3*y2 + y3;
+		float cy2 =  3*y0 - 6*y1 + 3*y2;
+		float cy1 = -3*y0 + 3*y1;
+		float cy0 =    y0;
+
+		if( segments == 0 ) {
+			float maxaccel1 = fabs(2*cy2) + fabs(6*cy3);
+			float maxaccel2 = fabs(2*cx2) + fabs(6*cx3);
+			float maxaccel = maxaccel1 > maxaccel2 ? maxaccel1 : maxaccel2;
+			float h = 1.0;
+			if( maxaccel > 8.0 ) h = sqrt((8.0) / maxaccel);
+			segments = int(1/h);
+		}
+
+		for( int j = 0; j <= segments; ++j ) {
+			float t = (float)j / segments;
+			float x = cx0 + t*(cx1 + t*(cx2 + t*cx3));
+			float y = cy0 + t*(cy1 + t*(cy2 + t*cy3));
+			edge.append(x, y);
+		}
+	}
+}
+
 void MaskEngine::do_mask(VFrame *output,
 	int64_t start_position_project,
 	MaskAutos *keyframe_set,
 	MaskAuto *keyframe,
 	MaskAuto *default_auto)
 {
-	int new_color_model = 0;
+	this->output = output;
 	recalculate = 0;
+	int mask_model = 0;
 
 	switch( output->get_color_model() ) {
 	case BC_RGB_FLOAT:
 	case BC_RGBA_FLOAT:
-		new_color_model = BC_A_FLOAT;
+		mask_model = BC_A_FLOAT;
 		break;
 
 	case BC_RGB888:
 	case BC_RGBA8888:
 	case BC_YUV888:
 	case BC_YUVA8888:
-		new_color_model = BC_A8;
+		mask_model = BC_A8;
 		break;
 
 	case BC_RGB161616:
 	case BC_RGBA16161616:
 	case BC_YUV161616:
 	case BC_YUVA16161616:
-		new_color_model = BC_A16;
+		mask_model = BC_A16;
 		break;
 	}
 
 // Determine if recalculation is needed
 SET_TRACE
 
-	if( mask &&
-	    (mask->get_w() != output->get_w() ||
-	     mask->get_h() != output->get_h() ||
-	     mask->get_color_model() != new_color_model) ) {
-		delete mask;
-		delete temp_mask;
-		mask = 0;
+	int mask_w = output->get_w(), mask_h = output->get_h();
+	if( mask && ( mask->get_color_model() != mask_model ||
+	    mask->get_w() != mask_w || mask->get_h() != mask_h ) ) {
+		delete mask;  mask = 0;
 		recalculate = 1;
 	}
-
-	if( !recalculate ) {
-		if( point_sets.total != keyframe_set->total_submasks(start_position_project,
-			PLAY_FORWARD) )
-			recalculate = 1;
+	if( temp && ( temp->get_w() != mask_w || temp->get_h() != mask_h ) ) {
+		delete temp;  temp = 0;
 	}
 
-	if( !recalculate ) {
-		for( int i=0,n=keyframe_set->total_submasks(start_position_project, PLAY_FORWARD);
-				i<n && !recalculate; ++i ) {
-			ArrayList<MaskPoint*> new_points;
-			keyframe_set->get_points(&new_points, i,
+	total_submasks = keyframe_set->total_submasks(start_position_project, PLAY_FORWARD);
+	if( total_submasks != point_sets.size() )
+		recalculate = 1;
+
+	for( int i=0; i<total_submasks && !recalculate; ++i ) {
+		float new_fader = keyframe_set->get_fader(start_position_project, i, PLAY_FORWARD);
+		if( new_fader != faders[i] ) { recalculate = 1;  break; }
+		float new_feather = keyframe_set->get_feather(start_position_project, i, PLAY_FORWARD);
+		if( new_feather != feathers[i] ) { recalculate = 1;  break; }
+		ArrayList<MaskPoint*> new_points;
+		keyframe_set->get_points(&new_points, i,
 				start_position_project, PLAY_FORWARD);
-			if( !points_equivalent(&new_points, point_sets.values[i]) )
-				recalculate = 1;
-			new_points.remove_all_objects();
-		}
+		if( !points_equivalent(&new_points, point_sets[i]) )
+			recalculate = 1;
+		new_points.remove_all_objects();
 	}
 
-	int new_value = keyframe_set->get_value(start_position_project,
-		PLAY_FORWARD);
-	float new_feather = keyframe_set->get_feather(start_position_project,
-		PLAY_FORWARD);
-
-	if( recalculate ||
-		!EQUIV(new_feather, feather) ||
-		!EQUIV(new_value, value) ) {
-		recalculate = 1;
-		if( !mask ) {
-			mask = new VFrame(output->get_w(), output->get_h(),
-					new_color_model, 0);
-			temp_mask = new VFrame(output->get_w(), output->get_h(),
-					new_color_model, 0);
-		}
-		if( new_feather > 0 )
-			temp_mask->clear_frame();
-		else
-			mask->clear_frame();
-
+	if( recalculate ) {
 		for( int i = 0; i < point_sets.total; i++ ) {
-			ArrayList<MaskPoint*> *points = point_sets.values[i];
+			ArrayList<MaskPoint*> *points = point_sets[i];
 			points->remove_all_objects();
 		}
 		point_sets.remove_all_objects();
-
-		for( int i = 0;
-			i < keyframe_set->total_submasks(start_position_project,
-				PLAY_FORWARD);
-			i++ ) {
-			ArrayList<MaskPoint*> *new_points = new ArrayList<MaskPoint*>;
-			keyframe_set->get_points(new_points,
-				i,
-				start_position_project,
-				PLAY_FORWARD);
+		edges.remove_all_objects();
+		faders.remove_all();
+		feathers.remove_all();
+		fade[0] = 0;
+
+		for( int i=0; i<total_submasks; ++i ) {
+			float fader = keyframe_set->get_fader(start_position_project, i, PLAY_FORWARD);
+			float v = fader / 100;
+			faders.append(v);
+			temp_t t = fabs(v) * 0xffff;
+			if( fader < 0 ) {
+				if( fade[0] < t ) fade[0] = t;
+				t = 0;
+			}
+			fade[i+1] = t;
+			float feather = keyframe_set->get_feather(start_position_project, i, PLAY_FORWARD);
+			feathers.append(feather);
+			MaskPointSet *new_points = new MaskPointSet();
+			keyframe_set->get_points(new_points, i, start_position_project, PLAY_FORWARD);
 			point_sets.append(new_points);
+			draw_edge(*edges.append(new MaskEdge()), *new_points);
 		}
+// draw mask
+		if( !mask ) mask = new VFrame(mask_w, mask_h, mask_model, 0);
+		if( !temp ) temp = new VFrame(mask_w, mask_h, BC_A16, 0);
+		mask->clear_frame();
+		temp->clear_frame();
+		step = DO_MASK;
+		process_packages();
+		step = DO_FEATHER;
+		process_packages();
 	}
-
-
-
-	this->output = output;
-	this->mode = default_auto->mode;
-	this->feather = new_feather;
-	this->value = new_value;
-
-
 // Run units
 SET_TRACE
-	step = DO_MASK;
-	process_packages();
-	step = DO_Y_FEATHER;
-	process_packages();
-	step = DO_X_FEATHER;
-	process_packages();
 	step = DO_APPLY;
 	process_packages();
 SET_TRACE
-
-
 }
 
 void MaskEngine::init_packages()