Sundry minor fixes and optimisations

[melted] / src / modules / core / transition_composite.c

/*
 * transition_composite.c -- compose one image over another using alpha channel
 * Copyright (C) 2003-2004 Ushodaya Enterprises Limited
 * Author: Dan Dennedy <dan@dennedy.org>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software Foundation,
 * Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
 */

#include "transition_composite.h"
#include <framework/mlt.h>

#include <stdio.h>
#include <stdlib.h>
#include <ctype.h>
#include <string.h>
#include <math.h>

typedef void ( *composite_line_fn )( uint8_t *dest, uint8_t *src, int width_src, uint8_t *alpha, int weight, uint16_t *luma, int softness );

/* mmx function declarations */
#ifdef USE_MMX
        void composite_line_yuv_mmx( uint8_t *dest, uint8_t *src, int width_src, uint8_t *alpha, int weight, uint16_t *luma, int softness );
        int composite_have_mmx( void );
#endif

/** Geometry struct.
*/

struct geometry_s
{
        struct mlt_geometry_item_s item;
        int nw; // normalised width
        int nh; // normalised height
        int sw; // scaled width, not including consumer scale based upon w/nw
        int sh; // scaled height, not including consumer scale based upon h/nh
        int halign; // horizontal alignment: 0=left, 1=center, 2=right
        int valign; // vertical alignment: 0=top, 1=middle, 2=bottom
};

/** Parse the alignment properties into the geometry.
*/

static int alignment_parse( char* align )
{
        int ret = 0;
        
        if ( align == NULL );
        else if ( isdigit( align[ 0 ] ) )
                ret = atoi( align );
        else if ( align[ 0 ] == 'c' || align[ 0 ] == 'm' )
                ret = 1;
        else if ( align[ 0 ] == 'r' || align[ 0 ] == 'b' )
                ret = 2;

        return ret;
}

/** Calculate real geometry.
*/

static void geometry_calculate( mlt_transition this, struct geometry_s *output, float position )
{
        mlt_properties properties = MLT_TRANSITION_PROPERTIES( this );
        mlt_geometry geometry = mlt_properties_get_data( properties, "geometries", NULL );
        int length = mlt_geometry_get_length( geometry );

        // Allow wrapping
        if ( position >= length && length != 0 )
        {
                int section = position / length;
                position -= section * length;
                if ( section % 2 == 1 )
                        position = length - position;
        }

        // Fetch the key for the position
        mlt_geometry_fetch( geometry, &output->item, position );
}

static mlt_geometry transition_parse_keys( mlt_transition this, int normalised_width, int normalised_height )
{
        // Loop variable for property interrogation
        int i = 0;

        // Get the properties of the transition
        mlt_properties properties = MLT_TRANSITION_PROPERTIES( this );

        // Create an empty geometries object
        mlt_geometry geometry = mlt_geometry_init( );

        // Get the in and out position
        mlt_position in = mlt_transition_get_in( this );
        mlt_position out = mlt_transition_get_out( this );
        int length = out - in + 1;

        // Get the new style geometry string
        char *property = mlt_properties_get( properties, "geometry" );

        // Allow a geometry repeat cycle
        if ( mlt_properties_get_int( properties, "cycle" ) )
                length = mlt_properties_get_int( properties, "cycle" );

        // Parse the geometry if we have one
        mlt_geometry_parse( geometry, property, length, normalised_width, normalised_height );

        // Check if we're using the old style geometry
        if ( property == NULL )
        {
                // DEPRECATED: Multiple keys for geometry information is inefficient and too rigid for 
                // practical use - while deprecated, it has been slightly extended too - keys can now
                // be specified out of order, and can be blanked or NULL to simulate removal

                // Structure to use for parsing and inserting
                struct mlt_geometry_item_s item;

                // Parse the start property
                item.frame = 0;
                if ( mlt_geometry_parse_item( geometry, &item, mlt_properties_get( properties, "start" ) ) == 0 )
                        mlt_geometry_insert( geometry, &item );

                // Parse the keys in between
                for ( i = 0; i < mlt_properties_count( properties ); i ++ )
                {
                        // Get the name of the property
                        char *name = mlt_properties_get_name( properties, i );
        
                        // Check that it's valid
                        if ( !strncmp( name, "key[", 4 ) )
                        {
                                // Get the value of the property
                                char *value = mlt_properties_get_value( properties, i );
        
                                // Determine the frame number
                                item.frame = atoi( name + 4 );
        
                                // Parse and add to the list
                                if ( mlt_geometry_parse_item( geometry, &item, value ) == 0 )
                                        mlt_geometry_insert( geometry, &item );
                                else
                                        fprintf( stderr, "Invalid Key - skipping %s = %s\n", name, value );
                        }
                }

                // Parse the end
                item.frame = -1;
                if ( mlt_geometry_parse_item( geometry, &item, mlt_properties_get( properties, "end" ) ) == 0 )
                        mlt_geometry_insert( geometry, &item );
        }
        
        return geometry;
}

/** Adjust position according to scaled size and alignment properties.
*/

static void alignment_calculate( struct geometry_s *geometry )
{
        geometry->item.x += ( geometry->item.w - geometry->sw ) * geometry->halign / 2;
        geometry->item.y += ( geometry->item.h - geometry->sh ) * geometry->valign / 2;
}

/** Calculate the position for this frame.
*/

static int position_calculate( mlt_transition this, mlt_position position )
{
        // Get the in and out position
        mlt_position in = mlt_transition_get_in( this );

        // Now do the calcs
        return position - in;
}

/** Calculate the field delta for this frame - position between two frames.
*/

static inline float delta_calculate( mlt_transition this, mlt_frame frame )
{
        // Get the in and out position
        mlt_position in = mlt_transition_get_in( this );
        mlt_position out = mlt_transition_get_out( this );
        float length = out - in + 1;

        // Get the position of the frame
        char *name = mlt_properties_get( MLT_TRANSITION_PROPERTIES( this ), "_unique_id" );
        mlt_position position = mlt_properties_get_position( MLT_FRAME_PROPERTIES( frame ), name );

        // Now do the calcs
        float x = ( float )( position - in ) / length;
        float y = ( float )( position + 1 - in ) / length;

        return length * ( y - x ) / 2.0;
}

static int get_value( mlt_properties properties, char *preferred, char *fallback )
{
        int value = mlt_properties_get_int( properties, preferred );
        if ( value == 0 )
                value = mlt_properties_get_int( properties, fallback );
        return value;
}

/** A linear threshold determination function.
*/

static inline int32_t linearstep( int32_t edge1, int32_t edge2, int32_t a )
{
        if ( a < edge1 )
                return 0;

        if ( a >= edge2 )
                return 0x10000;

        return ( ( a - edge1 ) << 16 ) / ( edge2 - edge1 );
}

/** A smoother, non-linear threshold determination function.
*/

static inline int32_t smoothstep( int32_t edge1, int32_t edge2, uint32_t a )
{
        if ( a < edge1 )
                return 0;

        if ( a >= edge2 )
                return 0x10000;

        a = ( ( a - edge1 ) << 16 ) / ( edge2 - edge1 );

        return ( ( ( a * a ) >> 16 )  * ( ( 3 << 16 ) - ( 2 * a ) ) ) >> 16;
}

/** Load the luma map from PGM stream.
*/

static void luma_read_pgm( FILE *f, uint16_t **map, int *width, int *height )
{
        uint8_t *data = NULL;
        while (1)
        {
                char line[128];
                char comment[128];
                int i = 2;
                int maxval;
                int bpp;
                uint16_t *p;

                line[127] = '\0';

                // get the magic code
                if ( fgets( line, 127, f ) == NULL )
                        break;

                // skip comments
                while ( sscanf( line, " #%s", comment ) > 0 )
                        if ( fgets( line, 127, f ) == NULL )
                                break;

                if ( line[0] != 'P' || line[1] != '5' )
                        break;

                // skip white space and see if a new line must be fetched
                for ( i = 2; i < 127 && line[i] != '\0' && isspace( line[i] ); i++ );
                if ( ( line[i] == '\0' || line[i] == '#' ) && fgets( line, 127, f ) == NULL )
                        break;

                // skip comments
                while ( sscanf( line, " #%s", comment ) > 0 )
                        if ( fgets( line, 127, f ) == NULL )
                                break;

                // get the dimensions
                if ( line[0] == 'P' )
                        i = sscanf( line, "P5 %d %d %d", width, height, &maxval );
                else
                        i = sscanf( line, "%d %d %d", width, height, &maxval );

                // get the height value, if not yet
                if ( i < 2 )
                {
                        if ( fgets( line, 127, f ) == NULL )
                                break;

                        // skip comments
                        while ( sscanf( line, " #%s", comment ) > 0 )
                                if ( fgets( line, 127, f ) == NULL )
                                        break;

                        i = sscanf( line, "%d", height );
                        if ( i == 0 )
                                break;
                        else
                                i = 2;
                }

                // get the maximum gray value, if not yet
                if ( i < 3 )
                {
                        if ( fgets( line, 127, f ) == NULL )
                                break;

                        // skip comments
                        while ( sscanf( line, " #%s", comment ) > 0 )
                                if ( fgets( line, 127, f ) == NULL )
                                        break;

                        i = sscanf( line, "%d", &maxval );
                        if ( i == 0 )
                                break;
                }

                // determine if this is one or two bytes per pixel
                bpp = maxval > 255 ? 2 : 1;

                // allocate temporary storage for the raw data
                data = mlt_pool_alloc( *width * *height * bpp );
                if ( data == NULL )
                        break;

                // read the raw data
                if ( fread( data, *width * *height * bpp, 1, f ) != 1 )
                        break;

                // allocate the luma bitmap
                *map = p = (uint16_t*)mlt_pool_alloc( *width * *height * sizeof( uint16_t ) );
                if ( *map == NULL )
                        break;

                // proces the raw data into the luma bitmap
                for ( i = 0; i < *width * *height * bpp; i += bpp )
                {
                        if ( bpp == 1 )
                                *p++ = data[ i ] << 8;
                        else
                                *p++ = ( data[ i ] << 8 ) + data[ i + 1 ];
                }

                break;
        }

        if ( data != NULL )
                mlt_pool_release( data );
}

/** Generate a luma map from any YUV image.
*/

static void luma_read_yuv422( uint8_t *image, uint16_t **map, int width, int height )
{
        int i;
        
        // allocate the luma bitmap
        uint16_t *p = *map = ( uint16_t* )mlt_pool_alloc( width * height * sizeof( uint16_t ) );
        if ( *map == NULL )
                return;

        // proces the image data into the luma bitmap
        for ( i = 0; i < width * height * 2; i += 2 )
                *p++ = ( image[ i ] - 16 ) * 299; // 299 = 65535 / 219
}


/** Composite a source line over a destination line
*/

static inline
void composite_line_yuv( uint8_t *dest, uint8_t *src, int width_src, uint8_t *alpha, int weight, uint16_t *luma, int softness )
{
        register int j;
        int a, mix;
        
        for ( j = 0; j < width_src; j ++ )
        {
                a = ( alpha == NULL ) ? 255 : *alpha ++;
                mix = ( luma == NULL ) ? weight : smoothstep( luma[ j ], luma[ j ] + softness, weight + softness );
                mix = ( mix * a ) >> 8;
                *dest = ( *src++ * mix + *dest * ( ( 1 << 16 ) - mix ) ) >> 16;
                dest++;
                *dest = ( *src++ * mix + *dest * ( ( 1 << 16 ) - mix ) ) >> 16;
                dest++;
        }
}

/** Composite function.
*/

static int composite_yuv( uint8_t *p_dest, int width_dest, int height_dest, uint8_t *p_src, int width_src, int height_src, uint8_t *p_alpha, struct geometry_s geometry, int field, uint16_t *p_luma, int32_t softness, composite_line_fn line_fn )
{
        int ret = 0;
        int i;
        int x_src = 0, y_src = 0;
        int32_t weight = ( 1 << 16 ) * ( geometry.item.mix / 100 );
        int step = ( field > -1 ) ? 2 : 1;
        int bpp = 2;
        int stride_src = width_src * bpp;
        int stride_dest = width_dest * bpp;
        
        // Adjust to consumer scale
        int x = rint( 0.5 + geometry.item.x * width_dest / geometry.nw );
        int y = rint( 0.5 + geometry.item.y * height_dest / geometry.nh );
        int x_uneven = x & 1;

        // optimization points - no work to do
        if ( width_src <= 0 || height_src <= 0 )
                return ret;

        if ( ( x < 0 && -x >= width_src ) || ( y < 0 && -y >= height_src ) )
                return ret;

        // crop overlay off the left edge of frame
        if ( x < 0 )
        {
                x_src = -x;
                width_src -= x_src;
                x = 0;
        }
        
        // crop overlay beyond right edge of frame
        if ( x + width_src > width_dest )
                width_src = width_dest - x;

        // crop overlay off the top edge of the frame
        if ( y < 0 )
        {
                y_src = -y;
                height_src -= y_src;
                y = 0;
        }
        
        // crop overlay below bottom edge of frame
        if ( y + height_src > height_dest )
                height_src = height_dest - y;

        // offset pointer into overlay buffer based on cropping
        p_src += x_src * bpp + y_src * stride_src;

        // offset pointer into frame buffer based upon positive coordinates only!
        p_dest += ( x < 0 ? 0 : x ) * bpp + ( y < 0 ? 0 : y ) * stride_dest;

        // offset pointer into alpha channel based upon cropping
        if ( p_alpha )
                p_alpha += x_src + y_src * stride_src / bpp;

        // offset pointer into luma channel based upon cropping
        if ( p_luma )
                p_luma += x_src + y_src * stride_src / bpp;
        
        // Assuming lower field first
        // Special care is taken to make sure the b_frame is aligned to the correct field.
        // field 0 = lower field and y should be odd (y is 0-based).
        // field 1 = upper field and y should be even.
        if ( ( field > -1 ) && ( y % 2 == field ) )
        {
                if ( ( field == 1 && y < height_dest - 1 ) || ( field == 0 && y == 0 ) )
                        p_dest += stride_dest;
                else
                        p_dest -= stride_dest;
        }

        // On the second field, use the other lines from b_frame
        if ( field == 1 )
        {
                p_src += stride_src;
                if ( p_alpha )
                        p_alpha += stride_src / bpp;
                height_src--;
        }

        stride_src *= step;
        stride_dest *= step;
        int alpha_stride = stride_src / bpp;

        // Make sure than x and w are even
        if ( x_uneven )
        {
                p_src += 2;
                width_src --;
        }

        // now do the compositing only to cropped extents
        if ( line_fn != NULL )
        {
                for ( i = 0; i < height_src; i += step )
                {
                        line_fn( p_dest, p_src, width_src, p_alpha, weight, p_luma, softness );
        
                        p_src += stride_src;
                        p_dest += stride_dest;
                        if ( p_alpha )
                                p_alpha += alpha_stride;
                        if ( p_luma )
                                p_luma += alpha_stride;
                }
        }
        else
        {
                for ( i = 0; i < height_src; i += step )
                {
                        composite_line_yuv( p_dest, p_src, width_src, p_alpha, weight, p_luma, softness );
        
                        p_src += stride_src;
                        p_dest += stride_dest;
                        if ( p_alpha )
                                p_alpha += alpha_stride;
                        if ( p_luma )
                                p_luma += alpha_stride;
                }
        }

        return ret;
}


/** Scale 16bit greyscale luma map using nearest neighbor.
*/

static inline void
scale_luma ( uint16_t *dest_buf, int dest_width, int dest_height, const uint16_t *src_buf, int src_width, int src_height, int invert )
{
        register int i, j;
        register int x_step = ( src_width << 16 ) / dest_width;
        register int y_step = ( src_height << 16 ) / dest_height;
        register int x, y = 0;

        for ( i = 0; i < dest_height; i++ )
        {
                const uint16_t *src = src_buf + ( y >> 16 ) * src_width;
                x = 0;
                
                for ( j = 0; j < dest_width; j++ )
                {
                        *dest_buf++ = src[ x >> 16 ] ^ invert;
                        x += x_step;
                }
                y += y_step;
        }
}

static uint16_t* get_luma( mlt_properties properties, int width, int height )
{
        // The cached luma map information
        int luma_width = mlt_properties_get_int( properties, "_luma.width" );
        int luma_height = mlt_properties_get_int( properties, "_luma.height" );
        uint16_t *luma_bitmap = mlt_properties_get_data( properties, "_luma.bitmap", NULL );
        int invert = mlt_properties_get_int( properties, "luma_invert" );
        
        // If the filename property changed, reload the map
        char *resource = mlt_properties_get( properties, "luma" );

        char temp[ 512 ];

        if ( resource != NULL && strchr( resource, '%' ) )
        {
                sprintf( temp, "%s/lumas/%s/%s", mlt_factory_prefix( ), mlt_environment( "MLT_NORMALISATION" ), strchr( resource, '%' ) + 1 );
                resource = temp;
        }

        if ( resource != NULL && ( luma_bitmap == NULL || luma_width != width || luma_height != height ) )
        {
                uint16_t *orig_bitmap = mlt_properties_get_data( properties, "_luma.orig_bitmap", NULL );
                luma_width = mlt_properties_get_int( properties, "_luma.orig_width" );
                luma_height = mlt_properties_get_int( properties, "_luma.orig_height" );

                // Load the original luma once
                if ( orig_bitmap == NULL )
                {
                        char *extension = extension = strrchr( resource, '.' );
                        
                        // See if it is a PGM
                        if ( extension != NULL && strcmp( extension, ".pgm" ) == 0 )
                        {
                                // Open PGM
                                FILE *f = fopen( resource, "r" );
                                if ( f != NULL )
                                {
                                        // Load from PGM
                                        luma_read_pgm( f, &orig_bitmap, &luma_width, &luma_height );
                                        fclose( f );
                                        
                                        // Remember the original size for subsequent scaling
                                        mlt_properties_set_data( properties, "_luma.orig_bitmap", orig_bitmap, luma_width * luma_height * 2, mlt_pool_release, NULL );
                                        mlt_properties_set_int( properties, "_luma.orig_width", luma_width );
                                        mlt_properties_set_int( properties, "_luma.orig_height", luma_height );
                                }
                        }
                        else
                        {
                                // Get the factory producer service
                                char *factory = mlt_properties_get( properties, "factory" );
        
                                // Create the producer
                                mlt_producer producer = mlt_factory_producer( factory, resource );
        
                                // If we have one
                                if ( producer != NULL )
                                {
                                        // Get the producer properties
                                        mlt_properties producer_properties = MLT_PRODUCER_PROPERTIES( producer );
        
                                        // Ensure that we loop
                                        mlt_properties_set( producer_properties, "eof", "loop" );
        
                                        // Now pass all producer. properties on the transition down
                                        mlt_properties_pass( producer_properties, properties, "luma." );
        
                                        // We will get the alpha frame from the producer
                                        mlt_frame luma_frame = NULL;
        
                                        // Get the luma frame
                                        if ( mlt_service_get_frame( MLT_PRODUCER_SERVICE( producer ), &luma_frame, 0 ) == 0 )
                                        {
                                                uint8_t *luma_image;
                                                mlt_image_format luma_format = mlt_image_yuv422;
        
                                                // Get image from the luma producer
                                                mlt_properties_set( MLT_FRAME_PROPERTIES( luma_frame ), "rescale.interp", "none" );
                                                mlt_frame_get_image( luma_frame, &luma_image, &luma_format, &luma_width, &luma_height, 0 );
        
                                                // Generate the luma map
                                                if ( luma_image != NULL && luma_format == mlt_image_yuv422 )
                                                        luma_read_yuv422( luma_image, &orig_bitmap, luma_width, luma_height );
        
                                                // Remember the original size for subsequent scaling
                                                mlt_properties_set_data( properties, "_luma.orig_bitmap", orig_bitmap, luma_width * luma_height * 2, mlt_pool_release, NULL );
                                                mlt_properties_set_int( properties, "_luma.orig_width", luma_width );
                                                mlt_properties_set_int( properties, "_luma.orig_height", luma_height );
                                                
                                                // Cleanup the luma frame
                                                mlt_frame_close( luma_frame );
                                        }
        
                                        // Cleanup the luma producer
                                        mlt_producer_close( producer );
                                }
                        }
                }
                // Scale luma map
                luma_bitmap = mlt_pool_alloc( width * height * sizeof( uint16_t ) );
                scale_luma( luma_bitmap, width, height, orig_bitmap, luma_width, luma_height, invert * ( ( 1 << 16 ) - 1 ) );

                // Remember the scaled luma size to prevent unnecessary scaling
                mlt_properties_set_int( properties, "_luma.width", width );
                mlt_properties_set_int( properties, "_luma.height", height );
                mlt_properties_set_data( properties, "_luma.bitmap", luma_bitmap, width * height * 2, mlt_pool_release, NULL );
        }
        return luma_bitmap;
}

/** Get the properly sized image from b_frame.
*/

static int get_b_frame_image( mlt_transition this, mlt_frame b_frame, uint8_t **image, int *width, int *height, struct geometry_s *geometry )
{
        int ret = 0;
        mlt_image_format format = mlt_image_yuv422;

        // Get the properties objects
        mlt_properties b_props = MLT_FRAME_PROPERTIES( b_frame );
        mlt_properties properties = MLT_TRANSITION_PROPERTIES( this );

        if ( mlt_properties_get_int( properties, "distort" ) == 0 && mlt_properties_get_int( b_props, "distort" ) == 0 && geometry->item.distort == 0 )
        {
                // Adjust b_frame pixel aspect
                int normalised_width = geometry->item.w;
                int normalised_height = geometry->item.h;
                int real_width = get_value( b_props, "real_width", "width" );
                int real_height = get_value( b_props, "real_height", "height" );
                double input_ar = mlt_frame_get_aspect_ratio( b_frame );
                double output_ar = mlt_properties_get_double( b_props, "consumer_aspect_ratio" );
                int scaled_width = input_ar / output_ar * real_width;
                int scaled_height = real_height;
                        
                // Now ensure that our images fit in the normalised frame
                if ( scaled_width > normalised_width )
                {
                        scaled_height = scaled_height * normalised_width / scaled_width;
                        scaled_width = normalised_width;
                }
                if ( scaled_height > normalised_height )
                {
                        scaled_width = scaled_width * normalised_height / scaled_height;
                        scaled_height = normalised_height;
                }

                // Honour the fill request - this will scale the image to fill width or height while maintaining a/r
                // ????: Shouln't this be the default behaviour?
                if ( mlt_properties_get_int( properties, "fill" ) )
                {
                        if ( scaled_height < normalised_height && scaled_width * normalised_height / scaled_height < normalised_width )
                        {
                                scaled_width = scaled_width * normalised_height / scaled_height;
                                scaled_height = normalised_height;
                        }
                        else if ( scaled_width < normalised_width && scaled_height * normalised_width / scaled_width < normalised_height )
                        {
                                scaled_height = scaled_height * normalised_width / scaled_width;
                                scaled_width = normalised_width;
                        }
                }

                // Save the new scaled dimensions
                geometry->sw = scaled_width;
                geometry->sh = scaled_height;
        }
        else
        {
                geometry->sw = geometry->item.w;
                geometry->sh = geometry->item.h;
        }

        // We want to ensure that we bypass resize now...
        mlt_properties_set_int( b_props, "distort", 1 );

        // Take into consideration alignment for optimisation
        if ( !mlt_properties_get_int( properties, "titles" ) )
                alignment_calculate( geometry );

        // Adjust to consumer scale
        int x = geometry->item.x * *width / geometry->nw;
        int y = geometry->item.y * *height / geometry->nh;
        *width = geometry->sw * *width / geometry->nw;
        *height = geometry->sh * *height / geometry->nh;

        //x = ( x | 1 ) ^ 1;

        // optimization points - no work to do
        if ( *width < 1 || *height < 1 )
                return 1;

        if ( ( x < 0 && -x >= *width ) || ( y < 0 && -y >= *height ) )
                return 1;

        ret = mlt_frame_get_image( b_frame, image, &format, width, height, 1 );

        return ret;
}


static mlt_geometry composite_calculate( mlt_transition this, struct geometry_s *result, mlt_frame a_frame, float position )
{
        // Get the properties from the transition
        mlt_properties properties = MLT_TRANSITION_PROPERTIES( this );

        // Get the properties from the frame
        mlt_properties a_props = MLT_FRAME_PROPERTIES( a_frame );
        
        // Structures for geometry
        mlt_geometry start = mlt_properties_get_data( properties, "geometries", NULL );

        // Obtain the normalised width and height from the a_frame
        int normalised_width = mlt_properties_get_int( a_props, "normalised_width" );
        int normalised_height = mlt_properties_get_int( a_props, "normalised_height" );

        // Now parse the geometries
        if ( start == NULL )
        {
                // Parse the transitions properties
                start = transition_parse_keys( this, normalised_width, normalised_height );

                // Assign to properties to ensure we get destroyed
                mlt_properties_set_data( properties, "geometries", start, 0, ( mlt_destructor )mlt_geometry_close, NULL );
        }
        else
        {
                int length = mlt_transition_get_out( this ) - mlt_transition_get_in( this ) + 1;
                if ( mlt_properties_get_int( properties, "cycle" ) )
                        length = mlt_properties_get_int( properties, "cycle" );
                mlt_geometry_refresh( start, mlt_properties_get( properties, "geometry" ), length, normalised_width, normalised_height );
        }

        // Do the calculation
        geometry_calculate( this, result, position );

        // Assign normalised info
        result->nw = normalised_width;
        result->nh = normalised_height;

        // Now parse the alignment
        result->halign = alignment_parse( mlt_properties_get( properties, "halign" ) );
        result->valign = alignment_parse( mlt_properties_get( properties, "valign" ) );

        return start;
}

static inline void inline_memcpy( uint8_t *dest, uint8_t *src, int length )
{
        uint8_t *end = src + length;
        while ( src < end )
        {
                *dest ++ = *src ++;
                *dest ++ = *src ++;
        }
}

mlt_frame composite_copy_region( mlt_transition this, mlt_frame a_frame, mlt_position frame_position )
{
        // Create a frame to return
        mlt_frame b_frame = mlt_frame_init( );

        // Get the properties of the a frame
        mlt_properties a_props = MLT_FRAME_PROPERTIES( a_frame );

        // Get the properties of the b frame
        mlt_properties b_props = MLT_FRAME_PROPERTIES( b_frame );

        // Get the position
        int position = position_calculate( this, frame_position );

        // Destination image
        uint8_t *dest = NULL;

        // Get the image and dimensions
        uint8_t *image = mlt_properties_get_data( a_props, "image", NULL );
        int width = mlt_properties_get_int( a_props, "width" );
        int height = mlt_properties_get_int( a_props, "height" );

        // Pointers for copy operation
        uint8_t *p;

        // Coordinates
        int w = 0;
        int h = 0;
        int x = 0;
        int y = 0;

        int ss = 0;
        int ds = 0;

        // Will need to know region to copy
        struct geometry_s result;

        float delta = delta_calculate( this, a_frame );

        // Calculate the region now
        composite_calculate( this, &result, a_frame, position + delta / 2 );

        // Need to scale down to actual dimensions
        x = rint( 0.5 + result.item.x * width / result.nw );
        y = rint( 0.5 + result.item.y * height / result.nh );
        w = rint( 0.5 + result.item.w * width / result.nw );
        h = rint( 0.5 + result.item.h * height / result.nh );

        // Make sure that x and w are even
        if ( x & 1 )
        {
                x --;
                w += 2;
                if ( w & 1 )
                        w --;
        }
        else if ( w & 1 )
        {
                w ++;
        }

        ds = w * 2;
        ss = width * 2;

        // Now we need to create a new destination image
        dest = mlt_pool_alloc( w * h * 2 );

        // Assign to the new frame
        mlt_properties_set_data( b_props, "image", dest, w * h * 2, mlt_pool_release, NULL );
        mlt_properties_set_int( b_props, "width", w );
        mlt_properties_set_int( b_props, "height", h );

        if ( y < 0 )
        {
                dest += ( ds * -y );
                h += y;
                y = 0;
        }

        if ( y + h > height )
                h -= ( y + h - height );

        if ( x < 0 )
        {
                dest += -x * 2;
                w += x;
                x = 0;
        }

        if ( w > 0 && h > 0 )
        {
                // Copy the region of the image
                p = image + y * ss + x * 2;

                while ( h -- )
                {
                        inline_memcpy( dest, p, w * 2 );
                        dest += ds;
                        p += ss;
                }
        }

        // Assign this position to the b frame
        mlt_frame_set_position( b_frame, frame_position );
        mlt_properties_set_int( b_props, "distort", 1 );

        // Return the frame
        return b_frame;
}

/** Get the image.
*/

static int transition_get_image( mlt_frame a_frame, uint8_t **image, mlt_image_format *format, int *width, int *height, int writable )
{
        // Get the b frame from the stack
        mlt_frame b_frame = mlt_frame_pop_frame( a_frame );

        // Get the transition from the a frame
        mlt_transition this = mlt_frame_pop_service( a_frame );

        // This compositer is yuv422 only
        *format = mlt_image_yuv422;

        // Get the image from the a frame
        mlt_frame_get_image( a_frame, image, format, width, height, 1 );

        // Get the properties from the transition
        mlt_properties properties = MLT_TRANSITION_PROPERTIES( this );

        if ( b_frame != NULL )
        {
                // Get the properties of the a frame
                mlt_properties a_props = MLT_FRAME_PROPERTIES( a_frame );

                // Get the properties of the b frame
                mlt_properties b_props = MLT_FRAME_PROPERTIES( b_frame );

                // Structures for geometry
                struct geometry_s result;

                // Calculate the position
                float position = mlt_properties_get_double( b_props, "relative_position" );
                float delta = delta_calculate( this, a_frame );

                // Get the image from the b frame
                uint8_t *image_b = NULL;
                int width_b = *width;
                int height_b = *height;
        
                // Do the calculation
                composite_calculate( this, &result, a_frame, position );

                // Optimisation - no compositing required
                if ( result.item.mix == 0 || ( result.item.w == 0 && result.item.h == 0 ) )
                        return 0;

                // Need to keep the width/height of the a_frame on the b_frame for titling
                if ( mlt_properties_get( a_props, "dest_width" ) == NULL )
                {
                        mlt_properties_set_int( a_props, "dest_width", *width );
                        mlt_properties_set_int( a_props, "dest_height", *height );
                        mlt_properties_set_int( b_props, "dest_width", *width );
                        mlt_properties_set_int( b_props, "dest_height", *height );
                }
                else
                {
                        mlt_properties_set_int( b_props, "dest_width", mlt_properties_get_int( a_props, "dest_width" ) );
                        mlt_properties_set_int( b_props, "dest_height", mlt_properties_get_int( a_props, "dest_height" ) );
                }

                // Since we are the consumer of the b_frame, we must pass along these
                // consumer properties from the a_frame
                mlt_properties_set_double( b_props, "consumer_deinterlace", mlt_properties_get_double( a_props, "consumer_deinterlace" ) );
                mlt_properties_set_double( b_props, "consumer_aspect_ratio", mlt_properties_get_double( a_props, "consumer_aspect_ratio" ) );
                mlt_properties_set_int( b_props, "normalised_width", mlt_properties_get_double( a_props, "normalised_width" ) );
                mlt_properties_set_int( b_props, "normalised_height", mlt_properties_get_double( a_props, "normalised_height" ) );

                // Special case for titling...
                if ( mlt_properties_get_int( properties, "titles" ) )
                {
                        if ( mlt_properties_get( b_props, "rescale.interp" ) == NULL )
                                mlt_properties_set( b_props, "rescale.interp", "hyper" );
                        width_b = mlt_properties_get_int( a_props, "dest_width" );
                        height_b = mlt_properties_get_int( a_props, "dest_height" );
                }

                if ( get_b_frame_image( this, b_frame, &image_b, &width_b, &height_b, &result ) == 0 )
                {
                        uint8_t *dest = *image;
                        uint8_t *src = image_b;
                        uint8_t *alpha = mlt_frame_get_alpha_mask( b_frame );
                        int progressive = 
                                        mlt_properties_get_int( a_props, "consumer_deinterlace" ) ||
                                        mlt_properties_get_int( properties, "progressive" );
                        int field;
                        
                        int32_t luma_softness = mlt_properties_get_double( properties, "softness" ) * ( 1 << 16 );
                        uint16_t *luma_bitmap = get_luma( properties, width_b, height_b );
                        //composite_line_fn line_fn = mlt_properties_get_int( properties, "_MMX" ) ? composite_line_yuv_mmx : NULL;
                        composite_line_fn line_fn = NULL;

                        for ( field = 0; field < ( progressive ? 1 : 2 ); field++ )
                        {
                                // Assume lower field (0) first
                                float field_position = position + field * delta;
                                
                                // Do the calculation if we need to
                                composite_calculate( this, &result, a_frame, field_position );

                                if ( mlt_properties_get_int( properties, "titles" ) )
                                {
                                        result.item.w = *width * ( result.item.w / result.nw );
                                        result.nw = result.item.w;
                                        result.item.h = *height * ( result.item.h / result.nh );
                                        result.nh = *height;
                                        result.sw = width_b;
                                        result.sh = height_b;
                                }

                                // Align
                                alignment_calculate( &result );

                                // Composite the b_frame on the a_frame
                                composite_yuv( dest, *width, *height, src, width_b, height_b, alpha, result, progressive ? -1 : field, luma_bitmap, luma_softness, line_fn );
                        }
                }
        }

        return 0;
}

/** Composition transition processing.
*/

static mlt_frame composite_process( mlt_transition this, mlt_frame a_frame, mlt_frame b_frame )
{
        // Get a unique name to store the frame position
        char *name = mlt_properties_get( MLT_TRANSITION_PROPERTIES( this ), "_unique_id" );

        // Assign the current position to the name
        mlt_properties_set_position( MLT_FRAME_PROPERTIES( a_frame ), name, mlt_frame_get_position( a_frame ) );

        // Propogate the transition properties to the b frame
        mlt_properties_set_double( MLT_FRAME_PROPERTIES( b_frame ), "relative_position", position_calculate( this, mlt_frame_get_position( a_frame ) ) );
        
        mlt_frame_push_service( a_frame, this );
        mlt_frame_push_frame( a_frame, b_frame );
        mlt_frame_push_get_image( a_frame, transition_get_image );
        return a_frame;
}

/** Constructor for the filter.
*/

mlt_transition transition_composite_init( char *arg )
{
        mlt_transition this = calloc( sizeof( struct mlt_transition_s ), 1 );
        if ( this != NULL && mlt_transition_init( this, NULL ) == 0 )
        {
                mlt_properties properties = MLT_TRANSITION_PROPERTIES( this );
                
                this->process = composite_process;
                
                // Default starting motion and zoom
                mlt_properties_set( properties, "start", arg != NULL ? arg : "0,0:100%x100%" );
                
                // Default factory
                mlt_properties_set( properties, "factory", "fezzik" );

#ifdef USE_MMX
                //mlt_properties_set_int( properties, "_MMX", composite_have_mmx() );
#endif
        }
        return this;
}