Man Linux: Main Page and Category List


       Prima::codecs - How to write a codec for Prima image subsystem


       How to write a codec for Prima image subsystem

Start simple

       There are many graphical formats in the world, and yet more libraries,
       that depend on them. Writing a codec that supports particular library
       is a tedious task, especially if one wants many formats. Usually you
       never want to get into internal parts, the functionality comes first,
       and who needs all those funky options that format provides? We want to
       load a file and to show it. Everything else comes later - if ever. So,
       in a way to not scare you off, we start it simple.

       Define a callback function like:

          static Bool
          load( PImgCodec instance, PImgLoadFileInstance fi)

       Just that function is not enough for whole mechanism to work, but
       bindings will come later. Let us imagine we work with an imaginary
       library libduff, that we want to load files of .duf format.  [ To
       discern imaginary code from real, imaginary will be prepended with _  -
       like, _libduff_loadfile ]. So, we call _libduff_loadfile(), that loads
       black-and-white, 1-bits/pixel images, where 1 is white and 0 is black.

          static Bool
          load( PImgCodec instance, PImgLoadFileInstance fi)
             _LIBDUFF * _l = _libduff_load_file( fi-> fileName);
             if ( !_l) return false;

             // - create storage for our file
             CImage( fi-> object)-> create_empty( fi-> object,
               _l-> width, _l-> height, imBW);

             // Prima wants images aligned to 4-bytes boundary,
             // happily libduff has same considerations
             memcpy( PImage( fi-> object)-> data, _l-> bits,
               PImage( fi-> object)-> dataSize);

             _libduff_close_file( _l);

             return true;

       Prima keeps an open handle of the file; so we can use it if libduff
       trusts handles vs names:

            _LIBDUFF * _l = _libduff_load_file_from_handle( fi-> f);
          // In both cases, you don't need to close the handle -
          // however you might, it is ok:

             _libduff_close_file( _l);
             fclose( fi-> f);
          // You just assign it to null to indicate that you've closed it
             fi-> f = null;

       Together with load() you have to implement minimal open_load() and

       Simplest open_load() returns non-null pointer - it is enough to report

          static void *
          open_load( PImgCodec instance, PImgLoadFileInstance fi)
             return (void*)1;

       Its result will be available in "PImgLoadFileInstance-> instance", just
       in case. If it was dynamically allocated, free it in close_load().
       Dummy close_load() is doing simply nothing:

          static void
          close_load( PImgCodec instance, PImgLoadFileInstance fi)

   Writing to "PImage-> data"
       As mentioned above, Prima insists on keeping its image data in 32-bit
       aligned scanlines. If libduff allows reading from file by scanlines, we
       can use this possibility as well:

          PImage i = ( PImage) fi-> object;
          // note - since this notation is more convenient than
          // PImage( fi-> object)-> , instead i-> will be used

          Byte * dest = i-> data + ( _l-> height - 1) * i-> lineSize;
          while ( _l-> height--) {
             _libduff_read_next_scanline( _l, dest);
             dest -= i-> lineSize;

       Note that image is filled in reverse - Prima images are built like
       classical XY-coordinate grid, where Y ascends upwards.

       Here ends the simple part. You can skip down to "Registering with image
       subsystem" part, if you want it fast.

Single-frame loading

       Our libduff can be black-and-white in two ways - where 0 is black and 1
       is white and vice versa. While 0B/1W is perfectly corresponding to
       imbpp1 | imGrayScale and no palette operations are needed ( Image cares
       automatically about these), 0W/1B is although black-and-white grayscale
       but should be treated like general imbpp1 type.

            if ( l-> _reversed_BW) {
               i-> palette[0].r = i-> palette[0].g = i-> palette[0].b = 0xff;
               i-> palette[1].r = i-> palette[1].g = i-> palette[1].b = 0;

       NB. Image creates palette with size calculated by exponent of 2, since
       it can’t know beforehand of the actual palette size. If color palette
       for, say, 4-bit image contains 15 of 16 possible for 4-bit image
       colors, code like

            i-> palSize = 15;

       does the trick.

   Data conversion
       As mentioned before, Prima defines image scanline size to be aligned to
       32 bits, and the formula for lineSize calculation is

           lineSize = (( width * bits_per_pixel + 31) / 32) * 4;

       Prima defines number of converting routines between different data
       formats. Some of them can be applied to scanlines, and some to whole
       image ( due sampling algorithms ). These are defined in img_conv.h, and
       probably ones that you’ll need would be "bc_format1_format2", which
       work on scanlines and probably ibc_repad, which combines some
       "bc_XX_XX" with byte repadding.

       For those who are especially lucky, some libraries do not check between
       machine byte format and file byte format.  Prima unfortunately doesn’t
       provide easy method for determining this situation, but you have to
       convert your data in appropriate way to keep picture worthy of its
       name. Note the BYTEORDER symbol that is defined ( usually ) in

   Load with no data
       If a high-level code just needs image information rather than all its
       bits, codec can provide it in a smart way. Old code will work, but will
       eat memory and time. A flag "PImgLoadFileInstance-> noImageData" is
       indicating if image data is needed. On that condition, codec needs to
       report only dimensions of the image - but the type must be set anyway.
       Here comes full code:

          static Bool
          load( PImgCodec instance, PImgLoadFileInstance fi)
             _LIBDUFF * _l = _libduff_load_file( fi-> fileName);
             HV * profile = fi-> frameProperties;
             PImage i = ( PImage) fi-> frameProperties;
             if ( !_l) return false;

             CImage( fi-> object)-> create_empty( fi-> object, 1, 1,
                _l-> _reversed_BW ? imbpp1 : imBW);

             // copy palette, if any
             if ( _l-> _reversed_BW) {
                i-> palette[0].r = i-> palette[0].g = i-> palette[0].b = 0xff;
                i-> palette[1].r = i-> palette[1].g = i-> palette[1].b = 0;

             if ( fi-> noImageData) {
                // report dimensions
                pset_i( width,  _l-> width);
                pset_i( height, _l-> height);
                return true;

             // - create storage for our file
             CImage( fi-> object)-> create_empty( fi-> object,
                  _l-> width, _l-> height,
                  _l-> _reversed_BW ? imbpp1 : imBW);

             // Prima wants images aligned to 4-bytes boundary,
             // happily libduff has same considerations
             memcpy( PImage( fi-> object)-> data, _l-> bits,
               PImage( fi-> object)-> dataSize);

             _libduff_close_file( _l);

             return true;

       The newly introduced macro "pset_i" is a convenience operator,
       assigning integer (i) as a value to a hash key, given as a first
       parameter - it becomes string literal upon the expansion. Hash used for
       storage is a lexical of type "HV*".  Code

               HV * profile = fi-> frameProperties;
               pset_i( width, _l-> width);

       is a prettier way for

                   fi-> frameProperties,
                   "width", strlen( "width"),
                   newSViv( _l-> width),

       hv_store(), HV’s and SV’s along with other funny symbols are described
       in perlguts.pod in Perl installation.

   Return extra information
       Image attributes are dimensions, type, palette and data.  However, it
       is only Prima point of view - different formats can supply number of
       extra information, often irrelevant but sometimes useful. From perl
       code, Image has a hash reference ’extras’ on object, where comes all
       this stuff. Codec can report also such data, storing it in
       "PImgLoadFileInstance-> frameProperties".  Data should be stored in
       native perl format, so if you’re not familiar with perlguts, you better
       read it, especially if you want return arrays and hashes. But just in
       simple, you can return:

       1.  integers:       pset_i( integer, _l-> integer);

       2.  floats:         pset_f( float, _l-> float);

       3.  strings:        pset_c( string, _l-> charstar); - note - no malloc
           codec from you required

       4.  prima objects:  pset_H( Handle, _l-> primaHandle);

       5.  SV’s:           pset_sv_noinc( scalar, newSVsv(sv));

       6.  hashes:         pset_sv_noinc( scalar, ( SV *) newHV()); - hashes
           created through newHV() can be filled just in the same manner as
           described here

       7.  arrays:         pset_sv_noinc( scalar, ( SV *) newAV()); - arrays
           (AV) are described in perlguts also, but most useful function here
           is av_push. To push 4 values, for example, follow this code:

               AV * av = newAV();
               for ( i = 0;i < 4;i++) av_push( av, newSViv( i));
               pset_sv_noinc( myarray, newRV_noinc(( SV *) av);

           is a C equivalent to

                 ->{extras}-> {myarray} = [0,1,2,3];

       High level code can specify if the extra information should be loaded.
       This behavior is determined by flag "PImgLoadFileInstance->
       loadExtras". Codec may skip this flag, the extra information will not
       be returned, even if "PImgLoadFileInstance-> frameProperties" was
       changed. However, it is advisable to check for the flag, just for an
       efficiency.  All keys, possibly assigned to frameProperties should be
       enumerated for high-level code. These strings should be represented
       into "char ** PImgCodecInfo-> loadOutput" array.

          static char * loadOutput[] = {

          static ImgCodecInfo codec_info = {

          static void *
          init( PImgCodecInfo * info, void * param)
             *info = &codec_info;

       The code above is taken from codec_X11.c, where X11 bitmap can provide
       location of hot spot, two integers, X and Y. The type of the data is
       not specified.

   Loading to icons
       If high-level code wants an Icon instead of an Image, Prima takes care
       for producing and-mask automatically.  However, if codec knows
       explicitly about transparency mask stored in a file, it might change
       object in the way it fits better. Mask is stored on Icon in a "-> mask"

       a) Let us imagine, that 4-bit image always carries a transparent color
       index, in 0-15 range. In this case, following code will create
       desirable mask:

             if ( kind_of( fi-> object, CIcon) &&
                  ( _l-> transparent >= 0) &&
                  ( _l-> transparent < PIcon( fi-> object)-> palSize)) {
                PRGBColor p = PIcon( fi-> object)-> palette;
                p += _l-> transparent;
                PIcon( fi-> object)-> maskColor = ARGB( p->r, p-> g, p-> b);
                PIcon( fi-> object)-> autoMasking = amMaskColor;

       Of course,

             pset_i( transparentColorIndex, _l-> transparent);

       would be also helpful.

       b) if explicit bit mask is given, code will be like:

             if ( kind_of( fi-> object, CIcon) &&
                  ( _l-> maskData >= 0)) {
                memcpy( PIcon( fi-> object)-> mask, _l-> maskData, _l-> maskSize);
                PIcon( fi-> object)-> autoMasking = amNone;

       Note that mask is also subject to LSB/MSB and 32-bit alignment issues.
       Treat it as a regular imbpp1 data format.

       c) A format supports transparency information, but image does not
       contain any. In this case no action is required on the codec’s part;
       the high-level code specifies if the transparency mask is created (
       iconUnmask field ).

   open_load() and close_load()
       open_load() and close_load() are used as brackets for load requests,
       and although they come to full power in multiframe load requests, it is
       very probable that correctly written codec should use them. Codec that
       assigns "false" to "PImgCodecInfo-> canLoadMultiple" claims that it
       cannot load those images that have index different from zero. It may
       report total amount of frames, but still be incapable of loading them.
       There is also a load sequence, called null-load, when no load() calls
       are made, just open_load() and close_load().  These requests are made
       in case codec can provide some file information without loading frames
       at all. It can be any information, of whatever kind. It have to be
       stored into the hash "PImgLoadFileInstance-> fileProperties", to be
       filled once on open_load(). The only exception is
       "PImgLoadFileInstance-> frameCount", which can be filled on
       open_load(). Actually, frameCount could be filled on any load stage,
       except close_load(), to make sense in frame positioning. Even single
       frame codec is advised to fill this field, at least to tell whether
       file is empty ( frameCount == 0) or not ( frameCount == 1). More about
       frameCount comes into chapters dedicated to multiframe requests.  For
       strictly single-frame codecs it is therefore advised to care for
       open_load() and close_load().

   Load input
       So far codec is expected to respond for noImageData hint only, and it
       is possible to allow a high-level code to alter codec load behavior,
       passing specific parameters.  "PImgLoadFileInstance-> profile" is a
       hash, that contains these parameters. The data that should be applied
       to all frames and/or image file are set there when open_load() is
       called. These data, plus frame-specific keys passed to every load()
       call.  However, Prima passes only those hash keys, which are returned
       by load_defaults() function. This functions returns newly created ( by
       calling newHV()) hash, with accepted keys and their default ( and
       always valid ) value pairs.  Example below defines speed_vs_memory
       integer value, that should be 0, 1 or 2.

          static HV *
          load_defaults( PImgCodec c)
             HV * profile = newHV();
             pset_i( speed_vs_memory, 1);
             return profile;
          static Bool
          load( PImgCodec instance, PImgLoadFileInstance fi)
               HV * profile = fi-> profile;
               if ( pexist( speed_vs_memory)) {
                  int speed_vs_memory = pget_i( speed_vs_memory);
                  if ( speed_vs_memory < 0 || speed_vs_memory > 2) {
                       strcpy( fi-> errbuf, "speed_vs_memory should be 0, 1 or 2");
                       return false;
                  _libduff_set_load_optimization( speed_vs_memory);

       The latter code chunk can be applied to open_load() as well.

   Returning an error
       Image subsystem defines no severity gradation for codec errors.  If
       error occurs during load, codec returns false value, which is "null" on
       open_load() and "false" on load. It is advisable to explain the error,
       otherwise the user gets just "Loading error" string. To do so, error
       message is to be copied to "PImgLoadFileInstance-> errbuf", which is
       "char[256]".  On an extreme severe error codec may call croak(), which
       jumps to the closest G_EVAL block. If there is no G_EVAL blocks then
       program aborts. This condition could also happen if codec calls some
       Prima code that issues croak(). This condition is untrappable, - at
       least without calling perl functions.  Understanding that that behavior
       is not acceptable, it is still under design.

Multiple-frame load

       In order to indicate that a codec is ready to read multiframe images,
       it must set "PImgCodecInfo-> canLoadMultiple" flag to true. This only
       means, that codec should respond to the "PImgLoadFileInstance-> frame"
       field, which is integer that can be in range from 0 to
       "PImgLoadFileInstance-> frameCount - 1".  It is advised that codec
       should change the frameCount from its original value "-1" to actual
       one, to help Prima filter range requests before they go down to the
       codec. The only real problem that may happen to the codec which it
       strongly unwilling to initialize frameCount, is as follows.  If a
       loadAll request was made ( corresponding boolean
       "PImgLoadFileInstance-> loadAll" flag is set for codec’s information)
       and frameCount is not initialized, then Prima starts loading all
       frames, incrementing frame index until it receives an error. Assuming
       the first error it gets is an EOF, it reports no error, so there’s no
       way for a high-level code to tell whether there was an loading error or
       an end-of-file condition.  Codec may initialize frameCount at any time
       during open_load() or load(), even together with false return value.


       Approach for handling saving requests is very similar to a load ones.
       For the same reason and with same restrictions functions
       save_defaults() open_save(), save() and close_save() are defined. Below
       shown a typical saving code and highlighted differences from load.  As
       an example we’ll take existing codec_X11.c, which defines extra hot
       spot coordinates, x and y.

          static HV *
          save_defaults( PImgCodec c)
             HV * profile = newHV();
             pset_i( hotSpotX, 0);
             pset_i( hotSpotY, 0);
             return profile;

          static void *
          open_save( PImgCodec instance, PImgSaveFileInstance fi)
             return (void*)1;

          static Bool
          save( PImgCodec instance, PImgSaveFileInstance fi)
             PImage i = ( PImage) fi-> object;
             Byte * l;

             fprintf( fi-> f, "#define %s_width %d\n", name, i-> w);
             fprintf( fi-> f, "#define %s_height %d\n", name, i-> h);
             if ( pexist( hotSpotX))
                fprintf( fi-> f, "#define %s_x_hot %d\n", name, (int)pget_i( hotSpotX));
             if ( pexist( hotSpotY))
                fprintf( fi-> f, "#define %s_y_hot %d\n", name, (int)pget_i( hotSpotY));
             fprintf( fi-> f, "static char %s_bits[] = {\n  ", name);
             // printing of data bytes is omitted

          static void
          close_save( PImgCodec instance, PImgSaveFileInstance fi)

       Save request takes into account defined supported types, that are
       defined in "PImgCodecInfo-> saveTypes". Prima converts image to be
       saved into one of these formats, before actual save() call takes place.
       Another boolean flag, "PImgSaveFileInstance-> append" is summoned to
       govern appending to or rewriting a file, but this functionality is
       under design. Its current value is a hint, if true, for a codec not to
       rewrite but rather append the frames to an existing file. Due to
       increased complexity of the code, that should respond to the append
       hint, this behavior is not required.

       Codec may set two of PImgCodecInfo flags, canSave and canSaveMultiple.
       Save requests will never be called if canSave is false, and append
       requests along with multiframe save requests would be never invoked for
       a codec with canSaveMultiple set to false.  Scenario for a multiframe
       save request is the same as for a load one. All the issues concerning
       palette, data converting and saving extra information are actual,
       however there’s no corresponding flag like loadExtras - codec is
       expected to save all information what it can extract from
       "PImgSaveFileInstance-> objectExtras" hash.

Registering with image subsystem

       Finally, the code have to be registered. It is not as illustrative but
       this part better not to be oversimplified.  A codec’s callback
       functions are set into ImgCodecVMT structure.  Those function slots
       that are unused should not be defined as dummies - those are already
       defined and gathered under struct CNullImgCodecVMT. That’s why all
       functions in the illustration code were defined as static.  A codec
       have to provide some information that Prima uses to decide what codec
       should load this particular file.  If no explicit directions given,
       Prima asks those codecs whose file extensions match to file’s.  init()
       should return pointer to the filled struct, that describes codec’s

          // extensions to file - might be several, of course, thanks to dos...
          static char * myext[] = { "duf", "duff", nil };

          // we can work only with 1-bit/pixel
          static int    mybpp[] = {
              imbpp1 | imGrayScale, // 1st item is a default type
              0 };   // Zero means end-of-list. No type has zero value.

          // main structure
          static ImgCodecInfo codec_info = {
             "DUFF", // codec name
             "Numb & Number, Inc.", // vendor
             _LIBDUFF_VERS_MAJ, _LIBDUFF_VERS_MIN,    // version
             myext,    // extension
             "DUmb Format",     // file type
             "DUFF",     // file short type
             nil,    // features
             "",     // module
             true,   // canLoad
             false,  // canLoadMultiple
             false,  // canSave
             false,  // canSaveMultiple
             mybpp,  // save types
             nil,    // load output

          static void *
          init( PImgCodecInfo * info, void * param)
             *info = &codec_info;
             return (void*)1; // just non-null, to indicate success

       The result of init() is stored into "PImgCodec-> instance", and info
       into "PImgCodec-> info". If dynamic memory was allocated for these
       structs, it can be freed on done() invocation.  Finally, the function
       that is invoked from Prima, is the only that required to be exported,
       is responsible for registering a codec:

          apc_img_codec_duff( void )
             struct ImgCodecVMT vmt;
             memcpy( &vmt, &CNullImgCodecVMT, sizeof( CNullImgCodecVMT));
             vmt. init          = init;
             vmt. open_load     = open_load;
             vmt. load          = load;
             vmt. close_load    = close_load;
             apc_img_register( &vmt, nil);

       This procedure can register as many codecs as it wants to, but
       currently Prima is designed so that one codec_XX.c file should be
       connected to one library only.

       The name of the procedure is apc_img_codec_ plus library name, that is
       required for a compilation with Prima.  File with the codec should be
       called codec_duff.c ( is our case) and put into img directory in Prima
       source tree. Following these rules, Prima will be assembled with
       libduff.a ( or duff.lib, or whatever, the actual library name is system
       dependent) - if the library is present.


       Dmitry Karasik, <>.


       Prima, Prima::Image, Prima::internals, Prima::image-load