The interface between Ghostscript and device drivers

Table of contents

For other information, see the Ghostscript overview and the documentation on how to build Ghostscript.

Adding a driver

To add a driver to Ghostscript, first pick a name for your device, say "smurf". (Device names must be 1 to 8 characters, begin with a letter, and consist only of letters, digits, and underscores. Case is significant: all current device names are lower case.) Then all you need do is edit contrib.mak in two places.

  1. The list of devices, in the section headed "Catalog". Add smurf to the list.
  2. The section headed "Device drivers".

    Suppose the files containing the smurf driver are called "joe" and "fred". Then you should add the following lines:

    # ------ The SMURF device ------ #
    smurf_=$(GLOBJ)joe.$(OBJ) $(GLOBJ)fred.$(OBJ)
    $(DD) $(smurf_)
            $(SETDEV) $(DD)smurf $(smurf_)
    $(GLOBJ)joe.$(OBJ) : $(GLSRC)joe.c
    	$(GLCC) $(GLO_)joe.$(OBJ) $(C_) $(GLSRC)joe.c
    $(GLOBJ)fred.$(OBJ) : $(GLSRC)fred.c
    	$(GLCC) $(GLO_)fred.$(OBJ) $(C_) $(GLSRC)fred.c

    and whatever joe.c and fred.c depend on. If the smurf driver also needs special libraries, for instance a library named "gorf", then the entry should look like this:

    $(DD) : $(smurf_)
            $(SETDEV) $(DD)smurf $(smurf_)
            $(ADDMOD) $(DD)smurf -lib gorf

    If, as will usually be the case, your driver is a printer driver (as discussed below), the device entry should look like this:

    $(DD) : $(smurf_) $(GLD)
            $(SETPDEV) $(DD)smurf $(smurf_)


    $(DD) : $(smurf_) $(GLD)
            $(SETPDEV) $(DD)smurf $(smurf_)
            $(ADDMOD) $(DD)smurf -lib gorf

    Note that the space before the :, and the explicit compilation rules for the .c files, are required for portability,

Keeping things simple

If you want to add a simple device (specifically, a monochrome printer), you probably don't need to read the rest of this document; just use the code in an existing driver as a guide. The Epson and Canon BubbleJet drivers gdevepsn.c and gdevbj10.c are good models for dot-matrix printers, which require presenting the data for many scan lines at once; the DeskJet/LaserJet drivers in gdevdjet.c are good models for laser printers, which take a single scan line at a time but support data compression. For color printers, there are unfortunately no good models: the two major color inkjet printer drivers, gdevcdj.c and gdevstc.c, are far too complex to read.

On the other hand, if you're writing a driver for some more esoteric device, you probably do need at least some of the information in the rest of this document. It might be a good idea for you to read it in conjunction with one of the existing drivers.

Duplication of code, and sheer volume of code, is a serious maintenance and distribution problem for Ghostscript. If your device is similar to an existing one, try to implement your driver by adding some parameterization to an existing driver rather than by copying code to create an entirely new source module. gdevepsn.c and gdevdjet.c are good examples of this approach.

Driver structure

A device is represented by a structure divided into three parts:

Normally the procedure structure is defined and initialized at compile time. A prototype of the parameter structure (including both generic and device-specific parameters) is defined and initialized at compile time, but is copied and filled in when an instance of the device is created. Both of these structures should be declared as const, but for backward compatibility reasons the latter is not.

The gx_device_common macro defines the common structure elements, with the intent that devices define and export a structure along the following lines. Do not fill in the individual generic parameter values in the usual way for C structures: use the macros defined for this purpose in gxdevice.h or, if applicable, gdevprn.h.

typedef struct smurf_device_s {
        ... device-specific parameters ...
} smurf_device;
smurf_device gs_smurf_device = {
        ... macro for generic parameter values ...,
        { ... procedures ... },         /* std_procs */
        ... device-specific parameter values if any ...

The device structure instance must have the name gs_smurf_device, where smurf is the device name used in contrib.mak. gx_device_common is a macro consisting only of the element definitions.

All the device procedures are called with the device as the first argument. Since each device type is actually a different structure type, the device procedures must be declared as taking a gx_device * as their first argument, and must cast it to smurf_device * internally. For example, in the code for the "memory" device, the first argument to all routines is called dev, but the routines actually use mdev to refer to elements of the full structure, using the following standard initialization statement at the beginning of each procedure:

gx_memory_device *const mdev = (gx_device_memory *)dev;

(This is a cheap version of "object-oriented" programming: in C++, for example, the cast would be unnecessary, and in fact the procedure table would be constructed by the compiler.)

Structure definition

You should consult the definition of struct gx_device_s in gxdevice.h for the complete details of the generic device structure. Some of the most important members of this structure for ordinary drivers are:

const char *dname;      The device name
bool is_open;   True if device has been opened
gx_device_color_info color_info;   Color information
int width;   Width in pixels
int height;   Height in pixels

The name in the structure (dname) should be the same as the name in contrib.mak.

For sophisticated developers only

If for any reason you need to change the definition of the basic device structure, or to add procedures, you must change the following places:

You may also have to change the code for gx_default_get_params or gx_default_put_params in gsdparam.c.

You should not have to change any of the real devices in the standard Ghostscript distribution (listed in devs.mak and contrib.mak) or any of your own devices, because all of them are supposed to use the macros in gxdevice.h or gdevprn.h to define and initialize their state.

Coordinates and types

Coordinate system

Since each driver specifies the initial transformation from user coordinates to device coordinates, the driver can use any coordinate system it wants, as long as a device coordinate will fit in an int. (This is only an issue on DOS systems, where ints are only 16 bits. User coordinates are represented as floats.) Most current drivers use a coordinate system with (0,0) in the upper left corner, with X increasing to the right and Y increasing toward the bottom. However, there is supposed to be nothing in the rest of Ghostscript that assumes this, and indeed some drivers use a coordinate system with (0,0) in the lower left corner.

Drivers must check (and, if necessary, clip) the coordinate parameters given to them: they should not assume the coordinates will be in bounds. The fit_fill and fit_copy macros in gxdevice.h are very helpful in doing this.

Color definition

Between the Ghostscript graphics library and the device, colors are represented in three forms. Color components in a color space (Gray, RGB, DeviceN, etc.) represented as frac values. Device colorants are represented as gx_color_value values. For many procedures, colors are represented in a type called gx_color_index. All three types are described in more detail in Types

The color_info member of the device structure defines the color and gray-scale capabilities of the device. Its type is defined as follows:

 * The enlarged color model information structure: Some of the
 * information that was implicit in the component number in
 * the earlier conventions (component names, polarity, mapping
 * functions) are now explicitly provided.
 * Also included is some information regarding the encoding of
 * color information into gx_color_index. Some of this information
 * was previously gathered indirectly from the mapping
 * functions in the existing code, specifically to speed up the
 * halftoned color rendering operator (see
 * gx_dc_ht_colored_fill_rectangle in gxcht.c). The information
 * is now provided explicitly because such optimizations are
 * more critical when the number of color components is large.
 * Note: no pointers have been added to this structure, so there
 *       is no requirement for a structure descriptor.
typedef struct gx_device_color_info_s {

     * max_components is the maximum number of components for all
     * color models supported by this device. This does not include
     * any alpha components.
    int max_components;

     * The number of color components. This does not include any
     * alpha-channel information, which may be integrated into
     * the gx_color_index but is otherwise passed as a separate
     * component.
    int num_components;

     * Polarity of the components of the color space, either
     * additive or subtractive. This is used to interpret transfer
     * functions and halftone threshold arrays. Possible values
    gx_color_polarity_t polarity;

     * The number of bits of gx_color_index actually used. 
     * This must be <= sizeof(gx_color_index), which is usually 64.
    byte depth;

     * Index of the gray color component, if any. The max_gray and
     * dither_gray values apply to this component only; all other
     * components use the max_color and dither_color values.
     * This will be GX_CINFO_COMP_NO_INDEX if there is no gray 
     * component.
    byte gray_index;

     * max_gray and max_color are the number of distinct native
     * intensity levels, less 1, for the gray and all other color
     * components, respectively. For nearly all current devices
     * that support both gray and non-gray components, the two
     * parameters have the same value.
     * dither_grays and dither_colors are the number of intensity
     * levels between which halftoning can occur, for the gray and
     * all other color components, respectively. This is
     * essentially redundant information: in all reasonable cases,
     * dither_grays = max_gray + 1 and dither_colors = max_color + 1.
     * These parameters are, however, extensively used in the
     * current code, and thus have been retained.
     * Note that the non-gray values may now be relevant even if
     * num_components == 1. This simplifies the handling of devices
     * with configurable color models which may be set for a single
     * non-gray color model.
    gx_color_value max_gray;	/* # of distinct color levels -1 */
    gx_color_value max_color;

    gx_color_value dither_grays;
    gx_color_value dither_colors;

     * Information to control super-sampling of objects to support
     * anti-aliasing.
    gx_device_anti_alias_info anti_alias;

     * Flag to indicate if gx_color_index for this device may be divided
     * into individual fields for each component. This is almost always
     * the case for printers, and is the case for most modern displays
     * as well. When this is the case, halftoning may be performed
     * separately for each component, which greatly simplifies processing
     * when the number of color components is large.
     * If the gx_color_index is separable in this manner, the comp_shift
     * array provides the location of the low-order bit for each
     * component. This may be filled in by the client, but need not be.
     * If it is not provided, it will be calculated based on the values
     * in the max_gray and max_color fields as follows:
     *     comp_shift[num_components - 1] = 0,
     *     comp_shift[i] = comp_shift[i + 1]
     *                      + ( i == gray_index ? ceil(log2(max_gray + 1))
     *                                          : ceil(log2(max_color + 1)) )
     * The comp_mask and comp_bits fields should be left empty by the client.
     * They will be filled in during initialization using the following
     * mechanism:
     *     comp_bits[i] = ( i == gray_index ? ceil(log2(max_gray + 1))
     *                                      : ceil(log2(max_color + 1)) )
     *     comp_mask[i] = (((gx_color_index)1 << comp_bits[i]) - 1)
     *                       << comp_shift[i]
     * (For current devices, it is almost always the case that
     * max_gray == max_color, if the color model contains both gray and
     * non-gray components.)
     * If separable_and_linear is not set, the data in the other fields
     * is unpredictable and should be ignored.
    gx_color_enc_sep_lin_t separable_and_linear;
    byte                   comp_shift[GX_DEVICE_COLOR_MAX_COMPONENTS];
    byte                   comp_bits[GX_DEVICE_COLOR_MAX_COMPONENTS];
    gx_color_index         comp_mask[GX_DEVICE_COLOR_MAX_COMPONENTS];
     * Pointer to name for the process color model.
    const char * cm_name;

} gx_device_color_info;

Note: See Changing color_info data before changing any information in the color_info structure for a device.

It is recommended that the values for this structure be defined using one of the standard macros provided for this purpose. This allows for future changes to be made to the structure without changes being required in the actual device code.

The following macros (in gxdevcli.h) provide convenient shorthands for initializing this structure for ordinary black-and-white or color devices:

#define dci_black_and_white ...
#define dci_color(depth,maxv,dither) ...

The #define dci_black_and_white macro defines a single bit monochrome device (For example: a typical monochrome printer device.)

The #define dci_color(depth,maxv,dither) macro can be used to define a 24 bit RGB device or a 4 or 32 bit CMYK device.

The #define dci_extended_alpha_values macro (in gxdevcli.h) specifies most of the current fields in the structure. However this macro allows only the default setting for the comp_shift, comp_bits, and comp_mask fields to be set. Any device which requires a non-default setting for these fields has to correctly these fields during the device open procedure. See Separable and linear fields> and Changing color_info data.

The idea is that a device has a certain number of gray levels (max_gray+1) and a certain number of colors (max_rgb+1) that it can produce directly. When Ghostscript wants to render a given color space color value as a device color, it first tests whether the color is a gray level and if so:

If max_gray is large (>= 31), Ghostscript asks the device to approximate the gray level directly. If the device returns a valid gx_color_index, Ghostscript uses it. Otherwise, Ghostscript assumes that the device can represent dither_gray distinct gray levels, equally spaced along the diagonal of the color cube, and uses the two nearest ones to the desired color for halftoning.

If the color is not a gray level:

If max_rgb is large (>= 31), Ghostscript asks the device to approximate the color directly. If the device returns a valid gx_color_index, Ghostscript uses it. Otherwise, Ghostscript assumes that the device can represent
dither_rgb × dither_rgb × dither_rgb

distinct colors, equally spaced throughout the color cube, and uses two of the nearest ones to the desired color for halftoning.

Separable and linear fields

The three fields comp_shift, comp_bits, and comp_mask are only used if the separable_and_linear field is set to GX_CINFO_SEP_LIN. In this situation a gx_color_index value must represent a combination created by or'ing bits for each of the devices's output colorants. The comp_shift array defines the location (shift count) of each colorants bits in the output gx_color_index value. The comp_bits array defines the number of bits for each colorant. The comp_mask array contains a mask which can be used to isolate the bits for each colorant. These fields must be set if the device supports more than four colorants.

Changing color_info data

For most devices, the information in the device's color_info structure is defined by the various device definition macros and the data remains constant during the entire existence of the device. In general the Ghostscript graphics assumes that the information is constant. However some devices want to modify the data in this structure.

The device's put_params procedure may change color_info field values. After the data has been modified then the device should be closed (via a call to gs_closedevice). Closing the device will erase the current page so these changes should only be made before anything has been drawn on a page.

The device's open_device procedure may change color_info field values. These changes should be done before any other procedures are called.

The Ghostscript graphics library uses some of the data in color_info to set the default procedures for the get_color_mapping_procs, get_color_comp_index, encode_color, and decode_color procedures. These default procedures are set when the device is originally created. If any changes are made to the color_info fields then the device's open_device procedure has responsibility for insuring that the correct procedures are contained in the device structure. (For an example, see the display device open procedure display_open and its subroutine display_set_color_format (in gdevdisp).


Here is a brief explanation of the various types that appear as parameters or results of the drivers.

frac (defined in gxfrac.h)
This is the type used to represent color values for the input to the color model mapping procedures. It is currently defined as a short. It has a range of frac_0 to frac_1.
gx_color_value (defined in gxdevice.h)
This is the type used to represent RGB or CMYK color values. It is currently equivalent to unsigned short. However, Ghostscript may use less than the full range of the type to represent color values: gx_color_value_bits is the number of bits actually used, and gx_max_color_value is the maximum value, equal to (2^gx_max_color_value_bits)-1.
gx_device (defined in gxdevice.h)
This is the device structure, as explained above.
gs_matrix (defined in gsmatrix.h)
This is a 2-D homogeneous coordinate transformation matrix, used by many Ghostscript operators.
gx_color_index (defined in gxcindex.h)
This is meant to be whatever the driver uses to represent a device color. For example, it might be an index in a color map, or it might be R, G, and B values packed into a single integer. The Ghostscript graphics library gets gx_color_index values from the device's encode_color and hands them back as arguments to several other procedures. If the separable_and_linear field in the device's color_info structure is not set to GX_CINFO_SEP_LIN then Ghostscript does not do any computations with gx_color_index values.

The special value gx_no_color_index (defined as (~(gx_color_index)(0)) ) means "transparent" for some of the procedures.

The size of gx_color_index can be either 32 or 64 bits. The choice depends upon the architecture of the CPU and the compiler. The default type definition is simply:

typedef unsigned long gx_color_index;
However if GX_COLOR_INDEX_TYPE is defined, then it is used as the type for gx_color_index.
typedef GX_COLOR_INDEX_TYPE gx_color_index;
The smaller size (32 bits) may produce more efficient or faster executing code. The larger size (64 bits) is needed for representing either more bits per component or more components. An example of the later case is a device that supports 8 bit contone colorants using a DeviceCMYK process color model with its four colorants and also supports additional spot colorants.

Currently autoconf attempts to find a 64 bit type definition for the compiler being used, and if a 64 bit type is found then GX_COLOR_INDEX_TYPE is set to the type.

For Microsoft and the MSVC compiler, GX_COLOR_INDEX_TYPE will be set to unsigned _int64 if USE_LARGE_COLOR_INDEX is set to 1 either on the make command line or by editing the definition in msvc32.mak

gs_param_list (defined in gsparam.h)
This is a parameter list, which is used to read and set attributes in a device. See the comments in gsparam.h, and the description of the get_params and put_params procedures below, for more detail.
gx_tile_bitmap (defined in gxbitmap.h)
gx_strip_bitmap (defined in gxbitmap.h)
These structure types represent bitmaps to be used as a tile for filling a region (rectangle). gx_tile_bitmap is an older, deprecated type lacking shift and rep_shift; gx_strip_bitmap has superseded it, and should be used in new code. Here is a copy of the relevant part of the file:
 * Structure for describing stored bitmaps.
 * Bitmaps are stored bit-big-endian (i.e., the 2^7 bit of the first
 * byte corresponds to x=0), as a sequence of bytes (i.e., you can't
 * do word-oriented operations on them if you're on a little-endian
 * platform like the Intel 80x86 or VAX).  Each scan line must start on
 * a (32-bit) word boundary, and hence is padded to a word boundary,
 * although this should rarely be of concern, since the raster and width
 * are specified individually.  The first scan line corresponds to y=0
 * in whatever coordinate system is relevant.
 * For bitmaps used as halftone tiles, we may replicate the tile in
 * X and/or Y, but it is still valuable to know the true tile dimensions
 * (i.e., the dimensions prior to replication).  Requirements:
 *      width % rep_width = 0
 *      height % rep_height = 0
 * For halftones at arbitrary angles, we provide for storing the halftone
 * data as a strip that must be shifted in X for different values of Y.
 * For an ordinary (non-shifted) halftone that has a repetition width of
 * W and a repetition height of H, the pixel at coordinate (X,Y)
 * corresponds to halftone pixel (X mod W, Y mod H), ignoring phase;
 * for a shifted halftone with shift S, the pixel at (X,Y) corresponds
 * to halftone pixel ((X + S * floor(Y/H)) mod W, Y mod H).  Requirements:
 *      strip_shift < rep_width
 *      strip_height % rep_height = 0
 *      shift = (strip_shift * (size.y / strip_height)) % rep_width
typedef struct gx_strip_bitmap_s {
        byte *data;
        int raster;                     /* bytes per scan line */
        gs_int_point size;              /* width, height */
        gx_bitmap_id id;
        ushort rep_width, rep_height;   /* true size of tile */
        ushort strip_height;
        ushort strip_shift;
        ushort shift;
} gx_strip_bitmap;

Coding conventions

All the driver procedures defined below that return int results return 0 on success, or an appropriate negative error code in the case of error conditions. The error codes are defined in gserrors.h; they correspond directly to the errors defined in the PostScript language reference manuals. The most common ones for drivers are:

An attempt to open a file failed.
An error occurred in reading or writing a file.
An otherwise valid parameter value was too large for the implementation.
A parameter was outside the valid range.
An attempt to allocate memory failed. (If this happens, the procedure should release all memory it allocated before it returns.)

If a driver does return an error, rather than a simple return statement it should use the return_error macro defined in gx.h, which is automatically included by gdevprn.h but not by gserrors.h. For example


Allocating storage

While most drivers (especially printer drivers) follow a very similar template, there is one important coding convention that is not obvious from reading the code for existing drivers: driver procedures must not use malloc to allocate any storage that stays around after the procedure returns. Instead, they must use gs_malloc and gs_free, which have slightly different calling conventions. (The prototypes for these are in gsmemory.h, which is included in gx.h, which is included in gdevprn.h.) This is necessary so that Ghostscript can clean up all allocated memory before exiting, which is essential in environments that provide only single-address-space multi-tasking (some versions of Microsoft Windows).

char *gs_malloc(uint num_elements, uint element_size,
  const char *client_name);

Like calloc, but unlike malloc, gs_malloc takes an element count and an element size. For structures, num_elements is 1 andi element_size is sizeof the structure; for byte arrays, num_elements is the number of bytes and element_size is 1. Unlike calloc, gs_malloc does not clear the block of storage.

The client_name is used for tracing and debugging. It must be a real string, not NULL. Normally it is the name of the procedure in which the call occurs.

void gs_free(char *data, uint num_elements, uint element_size,
  const char *client_name);

Unlike free, gs_free demands that num_elements and element_size be supplied. It also requires a client name, like gs_malloc.

Driver instance allocation

All driver instances allocated by Ghostscript's standard allocator must point to a "structure descriptor" that tells the garbage collector how to trace pointers in the structure. For drivers registered in the normal way (using the makefile approach described above), no special care is needed as long as instances are created only by calling the gs_copydevice procedure defined in gsdevice.h. If you have a need to define devices that are not registered in this way, you must fill in the stype member in any dynamically allocated instances with a pointer to the same structure descriptor used to allocate the instance. For more information about structure descriptors, see gsmemory.h and gsstruct.h.

Printer drivers

Printer drivers (which include drivers that write some kind of raster file) are especially simple to implement. The printer driver must implement a print_page or print_page_copies procedure. There are macros in gdevprn.h that generate the device structure for such devices, of which the simplest is prn_device; for an example, see gdevbj10.c. If you are writing a printer driver, we suggest you start by reading gdevprn.h and the subsection on "Color mapping" below; you may be able to ignore all the rest of the driver procedures.

The print_page procedures are defined as follows:

int (*print_page)(gx_device_printer *, FILE *)
int (*print_page_copies)(gx_device_printer *, FILE *, int)

This procedure must read out the rendered image from the device and write whatever is appropriate to the file. To read back one or more scan lines of the image, the print_page procedure must call one of the following procedures:

int gdev_prn_copy_scan_lines(gx_device_printer *pdev, int y, byte *str,
    uint size)

For this procedure, str is where the data should be copied to, and size is the size of the buffer starting at str. This procedure returns the number of scan lines copied, or <0 for an error. str need not be aligned.

int gdev_prn_get_bits(gx_device_printer *pdev, int y, byte *str,
  byte **actual_data)

This procedure reads out exactly one scan line. If the scan line is available in the correct format already, *actual_data is set to point to it; otherwise, the scan line is copied to the buffer starting at str, and *actual_data is set to str. This saves a copying step most of the time. str need not be aligned; however, if *actual_data is set to point to an existing scan line, it will be aligned. (See the description of the get_bits procedure below for more details.)

In either case, each row of the image is stored in the form described in the comment under gx_tile_bitmap above; each pixel takes the number of bits specified as color_info.depth in the device structure, and holds values returned by the device's encode_color procedure.

The print_page procedure can determine the number of bytes required to hold a scan line by calling:

uint gdev_prn_raster(gx_device_printer *)

For a very simple concrete example, we suggest reading the code in bit_print_page in gdevbit.c.

If the device provides print_page, Ghostscript will call print_page the requisite number of times to print the desired number of copies; if the device provides print_page_copies, Ghostscript will call print_page_copies once per page, passing it the desired number of copies.

Driver procedures

Most of the procedures that a driver may implement are optional. If a device doesn't supply an optional procedure WXYZ, the entry in the procedure structure may be either gx_default_WXYZ, for instance gx_default_tile_rectangle, or NULL or 0. (The device procedure must also call the gx_default_ procedure if it doesn't implement the function for particular values of the arguments.) Since C compilers supply 0 as the value for omitted structure elements, this convention means that statically initialized procedure structures continue to work even if new (optional) members are added.

Life cycle

A device instance begins life in a closed state. In this state, no output operations will occur. Only the following procedures may be called:


When setdevice installs a device instance in the graphics state, it checks whether the instance is closed or open. If the instance is closed, setdevice calls the open routine, and then sets the state to open.

There is no user-accessible operation to close a device instance. This is not an oversight -- it is required in order to enforce the following invariant:

If a device instance is the current device in any graphics state, it must be open (have is_open set to true).

Device instances are only closed when they are about to be freed, which occurs in three situations:

Open, close, sync, copy

int (*open_device)(gx_device *) [OPTIONAL]
Open the device: do any initialization associated with making the device instance valid. This must be done before any output to the device. The default implementation does nothing. NOTE: Clients should never call a device's open_device procedure directly: they should always call gs_opendevice instead.
int (*finish_copydevice)(gx_device *dev, const gx_device *from_dev) [OPTIONAL]
Perform any cleanup required after copydevice has created a new device instance by copying from_dev. If the copy operation should not be allowed, this procedure should return an error; the copy will be freed. The default implementation allows copying the device prototype, but does not allow copying device instances, because instances may contain internal pointers that should not be shared between copies, and there is no way to determine this from outside the device. NOTE: Clients should never call a device's finish_copydevice procedure: this procedure is only intended for use by gs_copydevice[2].
void (*get_initial_matrix)(gx_device *, gs_matrix *) [OPTIONAL]
Construct the initial transformation matrix mapping user coordinates (nominally 1/72 inch per unit) to device coordinates. The default procedure computes this from width, height, and [xy]_pixels_per_inch on the assumption that the origin is in the upper left corner, that is
xx = x_pixels_per_inch/72, xy = 0,
yx = 0, yy = -y_pixels_per_inch/72,
tx = 0, ty = height.
int (*sync_output)(gx_device *) [OPTIONAL]
Synchronize the device. If any output to the device has been buffered, send or write it now. Note that this may be called several times in the process of constructing a page, so printer drivers should not implement this by printing the page. The default implementation does nothing.
int (*output_page)(gx_device *, int num_copies, int flush) [OPTIONAL]
Output a fully composed page to the device. The num_copies argument is the number of copies that should be produced for a hardcopy device. (This may be ignored if the driver has some other way to specify the number of copies.) The flush argument is true for showpage, false for copypage. The default definition just calls sync_output. Printer drivers should implement this by printing and ejecting the page.
int (*close_device)(gx_device *) [OPTIONAL]
Close the device: release any associated resources. After this, output to the device is no longer allowed. The default implementation does nothing. NOTE: Clients should never call a device's close_device procedure directly: they should always call gs_closedevice instead.

Color and alpha mapping

Note that code in the Ghostscript library may cache the results of calling one or more of the color mapping procedures. If the result returned by any of these procedures would change (other than as a result of a change made by the driver's put_params procedure), the driver must call gx_device_decache_colors(dev).

The map_rgb_color, map_color_rgb, and map_cmyk_color are obsolete. They have been left in the device procedure list for backward compatibility. See the encode_color and decode_color procedures below. To insure that older device drivers are changed to use the new encode_color and decode_color procedures, the parameters for the older procedures have been changed to match the new procedures. To minimize changes in devices that have already been written, the map_rgb_color and map_cmyk_color routines are used as the default value for the encode_color routine. The map_cmyk_color routine is used if the number of components is four. The map_rgb_color routine is used if the number of components is one or three. This works okay for RGB and CMYK process color model devices. However this does not work properly for gray devices. The encode_color routine for a gray device is only passed one component. Thus the map_rgb_color routine must be modified to only use a single input (instead of three). (See the encode_color and decode_color routines below.)

Colors can be specified to the Ghostscript graphics library in a variety of forms. For example, there are a wide variety of color spaces that can be used such as Gray, RGB, CMYK, DeviceN, Separation, Indexed, CIEbasedABC, etc. The graphics library converts the various input color space values into four base color spaces: Gray, RGB, CMYK, and DeviceN. The DeviceN color space allows for specifying values for individual device colorants or spot colors.

Colors are converted by the device in a two step process. The first step is to convert a color in one of the base color spaces (Gray, RGB, CMYK, or DeviceN) into values for each device colorant. This transformation is done via a set of procedures provided by the device. These procedures are provided by the get_color_mapping_procs device procedure.

Between the first and second steps, the graphics library applies transfer functions to the device colorants. Where needed, the output of the results after the transfer functions is used by the graphics library for halftoning.

In the second step, the device procedure encode_color is used to convert the transfer function results into a gx_color_index value. The gx_color_index values are passed to specify colors to various routines. The choice of the encoding for a gx_color_index is up to the device. Common choices are indexes into a color palette or several integers packed together into a single value. The manner of this encoding is usually opaque to the graphics library. The only exception to this statement occurs when halftoning 5 or more colorants. In this case the graphics library assumes that if a colorant values is zero then the bits associated with the colorant in the gx_color_index value are zero.

int get_color_comp_index(const gx_device * dev, const char * pname, int name_size, int src_index) [OPTIONAL]
This procedure returns the device colorant number of the given name. The possible return values are -1, 0 to GX_DEVICE_COLOR_MAX_COMPONENTS - 1, or GX_DEVICE_COLOR_MAX_COMPONENTS. A value of -1 indicates that the specified name is not a colorant for the device. A value of 0 to GX_DEVICE_COLOR_MAX_COMPONENTS - 1 indicates the colorant number of the given name. A value of GX_DEVICE_COLOR_MAX_COMPONENTS indicates that the given name is a valid colorant name for the device but the colorant is not currently being used. This is used for implementing names which are in SeparationColorNames but not in SeparationOrder.

The default procedure returns results based upon process color model of DeviceGray, DeviceRGB, or DeviceCMYK selected by color_info.num_components. This procedure must be defined if another process color model is used by the device or spot colors are supported by the device.

const gx_cm_color_map_procs * get_color_mapping_procs(const gx_device * dev) [OPTIONAL]
This procedure returns a list of three procedures. These procedures are used to translate values in either Gray, RGB, or CMYK color spaces into device colorant values. A separate procedure is not required for the DeviceN and Separation color spaces since these already represent device colorants.

The default procedure returns a list of procedures based upon color_info.num_components. These procedures are appropriate for DeviceGray, DeviceRGB, or DeviceCMYK process color model devices. A procedure must be defined if another process color model is used by the device or spot colors are to be supported.

gx_color_index (*encode_color)(gx_device * dev, gx_color_value * cv) [OPTIONAL]
Map a set of device color values into a gx_color_index value. The range of legal values of the arguments is 0 to gx_max_color_value. The default procedure packs bits into a gx_color_index value based upon the values in color_info.depth and color_info.num_components.

Note that the encode_color procedure must not return gx_no_color_index (all 1s).

int (*decode_color)(gx_device *, gx_color_index color, gx_color_value * CV) [OPTIONAL]
This is the inverse of the encode_color procedure. Map a gx_color_index value to color values. The default procedure unpacks bits from the gx_color_index value based upon the values in color_info.depth and color_info.num_components.
gx_color_index (*map_rgb_alpha_color)(gx_device *, gx_color_value red, gx_color_value green, gx_color_value blue, gx_color_value alpha) [OPTIONAL]
Map a RGB color and an opacity value to a device color. The range of legal values of the RGB and alpha arguments is 0 to gx_max_color_value; alpha = 0 means transparent, alpha = gx_max_color_value means fully opaque. The default is to use the encode_color procedure and ignore alpha.

Note that if a driver implements map_rgb_alpha_color, it must also implement encode_color, and must implement them in such a way that map_rgb_alpha_color(dev, r, g, b, gx_max_color_value) returns the same value as encode_color(dev, CV).

int (*map_color_rgb_alpha)(gx_device *, gx_color_index color, gx_color_value rgba[4]) [OPTIONAL]
Map a device color code to RGB and alpha values. The default implementation calls map_color_rgb and fills in gx_max_color_value for alpha.

Note that if a driver implements map_color_rgb_alpha, it must also implement decode_color, and must implement them in such a way that the first 3 values returned by map_color_rgb_alpha are the same as the values returned by decode_color.

Note that only RGB devices currently support variable opacity; alpha is ignored on other devices. The PDF 1.4 transparency features are supported on all devices.

typedef enum { go_text, go_graphics } graphic_object_type; int (*get_alpha_bits)(gx_device *dev, graphic_object_type type) [OPTIONAL] [OBSOLETE]
This procedure is no longer used: it is replaced by the color_info.anti_alias member of the driver structure. However, it still appears in the driver procedure vector for backward compatibility. It should never be called, and drivers should not implement it.
void (*update_spot_equivalent_colors)(gx_device *, const gs_state *) [OPTIONAL]
This routine provides a method for the device to gather an equivalent color for spot colorants. This routine is called when a Separation or DeviceN color space is installed. See comments at the start of gsequivc.c. Note: This procedure is only needed for devices that support spot colorants and also need to have an equivalent color for simulating the appearance of the spot colorants.

Pixel-level drawing

This group of drawing operations specifies data at the pixel level. All drawing operations use device coordinates and device color values.

int (*fill_rectangle)(gx_device *, int x, int y, int width, int height, gx_color_index color)
Fill a rectangle with a color. The set of pixels filled is {(px,py) | x <= px < x + width and y <= py < y + height}. In other words, the point (x,y) is included in the rectangle, as are (x+w-1,y), (x,y+h-1), and (x+w-1,y+h-1), but not (x+w,y), (x,y+h), or (x+w,y+h). If width <= 0 or height <= 0, fill_rectangle should return 0 without drawing anything.

Note that fill_rectangle is the only non-optional procedure in the driver interface.

Bitmap imaging

Bitmap (or pixmap) images are stored in memory in a nearly standard way. The first byte corresponds to (0,0) in the image coordinate system: bits (or polybit color values) are packed into it left to right. There may be padding at the end of each scan line: the distance from one scan line to the next is always passed as an explicit argument.

int (*copy_mono)(gx_device *, const unsigned char *data, int data_x, int raster, gx_bitmap_id id, int x, int y, int width, int height, gx_color_index color0, gx_color_index color1) [OPTIONAL]
Copy a monochrome image (similar to the PostScript image operator). Each scan line is raster bytes wide. Copying begins at (data_x,0) and transfers a rectangle of the given width and height to the device at device coordinate (x,y). (If the transfer should start at some non-zero y value in the data, the caller can adjust the data address by the appropriate multiple of the raster.) The copying operation writes device color color0 at each 0-bit, and color1 at each 1-bit: if color0 or color1 is gx_no_color_index, the device pixel is unaffected if the image bit is 0 or 1 respectively. If id is different from gx_no_bitmap_id, it identifies the bitmap contents unambiguously; a call with the same id will always have the same data, raster, and data contents.

This operation, with color0 = gx_no_color_index, is the workhorse for text display in Ghostscript, so implementing it efficiently is very important.

int (*tile_rectangle)(gx_device *, const gx_tile_bitmap *tile, int x, int y, int width, int height, gx_color_index color0, gx_color_index color1, int phase_x, int phase_y) [OPTIONAL] [OBSOLETE]
This procedure is still supported, but has been superseded by strip_tile_rectangle. New drivers should implement strip_tile_rectangle; if they cannot cope with non-zero shift values, they should test for this explicitly and call the default implementation (gx_default_strip_tile_rectangle) if shift != 0. Clients should call strip_tile_rectangle, not tile_rectangle.
int (*strip_tile_rectangle)(gx_device *, const gx_strip_bitmap *tile, int x, int y, int width, int height, gx_color_index color0, gx_color_index color1, int phase_x, int phase_y) [OPTIONAL]
Tile a rectangle. Tiling consists of doing multiple copy_mono operations to fill the rectangle with copies of the tile. The tiles are aligned with the device coordinate system, to avoid "seams". Specifically, the (phase_x, phase_y) point of the tile is aligned with the origin of the device coordinate system. (Note that this is backwards from the PostScript definition of halftone phase.) phase_x and phase_y are guaranteed to be in the range [0..tile->width) and [0..tile->height) respectively.

If color0 and color1 are both gx_no_color_index, then the tile is a color pixmap, not a bitmap: see the next section.

This operation is the workhorse for halftone filling in Ghostscript, so implementing it efficiently for solid tiles (that is, where either color0 and color1 are both gx_no_color_index, for colored halftones, or neither one is gx_no_color_index, for monochrome halftones) is very important.

Pixmap imaging

Pixmaps are just like bitmaps, except that each pixel occupies more than one bit. All the bits for each pixel are grouped together (this is sometimes called "chunky" or "Z" format). For copy_color, the number of bits per pixel is given by the color_info.depth parameter in the device structure: the legal values are 1, 2, 4, 8, 16, 24, 32, 40, 48, 56, or 64. The pixel values are device color codes (that is, whatever it is that encode_color returns).

int (*copy_color)(gx_device *, const unsigned char *data, int data_x, int raster, gx_bitmap_id id, int x, int y, int width, int height) [OPTIONAL]
Copy a color image with multiple bits per pixel. The raster is in bytes, but x and width are in pixels, not bits. If id is different from gx_no_bitmap_id, it identifies the bitmap contents unambiguously; a call with the same id will always have the same data, raster, and data contents.

We do not provide a separate procedure for tiling with a pixmap; instead, tile_rectangle can also take colored tiles. This is indicated by the color0 and color1 arguments' both being gx_no_color_index. In this case, as for copy_color, the raster and height in the "bitmap" are interpreted as for real bitmaps, but the x and width are in pixels, not bits.


In addition to direct writing of opaque pixels, devices must also support compositing. Currently two kinds of compositing are defined (RasterOp and alpha-based), but more may be added in the future.

int (*copy_alpha)(gx_device *dev, const unsigned char *data, int data_x, int raster, gx_bitmap_id id, int x, int y, int width, int height, gx_color_index color, int depth) [OPTIONAL]
This procedure is somewhat misnamed: it was added to the interface before we really understood alpha channel and compositing.

Fill a given region with a given color modified by an individual alpha value for each pixel. For each pixel, this is equivalent to alpha-compositing with a source pixel whose alpha value is obtained from the pixmap (data, data_x, and raster) and whose color is the given color (which has not been premultiplied by the alpha value), using the Sover rule. depth, the number of bits per alpha value, is either 2 or 4, and in any case is always a value returned by a previous call on the get_alpha_bits procedure. Note that if get_alpha_bits always returns 1, this procedure will never be called.

int (*create_compositor)(dev_t *dev, gx_device_t **pcdev, const gs_composite_t *pcte, const gs_imager_state *pis, gs_memory_t *memory) [OPTIONAL]
Create a new device (called a "compositing device" or "compositor") that will composite data written to it with the device's existing data, according to the compositing function defined by *pcte. Devices will normally implement this in one of the following standard ways:

Other kinds of forwarding devices, which don't fall into any of these categories, require special treatment. In principle, what they do is ask their target to create a compositor, and then create and return a copy of themselves with the target's new compositor as the target of the copy. There is a possible default implementation of this approach: if the original device was D with target T, and T creates a compositor C, then the default implementation creates a device F that for each operation temporarily changes D's target to C, forwards the operation to D, and then changes D's target back to T. However, the Ghostscript library currently only creates a compositor with an imaging forwarding device as target in a few specialized situations (banding, and bounding box computation), and these are handled as special cases.

Note that the compositor may have a different color space, color representation, or bit depth from the device to which it is compositing. For example, alpha-compositing devices use standard-format chunky color even if the underlying device doesn't.

Closing a compositor frees all of its storage, including the compositor itself. However, since the create_compositor call may return the same device, clients must check for this case, and only call the close procedure if a separate device was created.


int (*copy_rop)(gx_device *dev, const byte *sdata, int sourcex, uint sraster, gx_bitmap_id id, const gx_color_index *scolors, const gx_tile_bitmap *texture, const gx_color_index *tcolors, int x, int y, int width, int height, int phase_x, int phase_y, int command) [OPTIONAL]
This procedure is still supported, but has been superseded by strip_copy_rop. New drivers should implement strip_copy_rop; if they cannot cope with non-zero shift values in the texture, they should test for this explicitly and call the default implementation (gx_default_strip_copy_rop) if shift != 0. Clients should call strip_copy_rop, not copy_rop.
int (*strip_copy_rop)(gx_device *dev, const byte *sdata, int sourcex, uint sraster, gx_bitmap_id id, const gx_color_index *scolors, const gx_strip_bitmap *texture, const gx_color_index *tcolors, int x, int y, int width, int height, int phase_x, int phase_y, int command) [OPTIONAL]
Combine an optional source image S (as for copy_mono or copy_color) and an optional texture T (a tile, as for tile_rectangle) with the existing bitmap or pixmap D held by the driver, pixel by pixel, using any 3-input Boolean operation as modified by "transparency" flags: schematically, set D = f(D,S,T), computing f in RGB space rather than using actual device pixel values. S and T may each (independently) be a solid color, a bitmap with "foreground" and "background" colors, or a pixmap. This is a complex (and currently rather slow) operation. The arguments are as follows:
dev   the device, as for all driver procedures
sdata, sourcex, sraster, id, scolors   specify S, see below
texture, tcolors   specify T, see below
x, y, width, height   as for the other copy and fill procedures
phase_x, phase_y   part of T specification, see below
command   see below
The source specification S

As noted above, the source S may be a solid color, a bitmap, or a pixmap. If S is a solid color:

If S is a bitmap:

If S is a pixmap:

Note that if the source is a bitmap with background=0 and foreground=1, and the destination is 1 bit deep, then the source can be treated as a pixmap (scolors=NULL).

The texture specification T

Similar to the source, the texture T may be a solid color, a bitmap, or a pixmap. If T is a solid color:

If T is a bitmap:

If T is a pixmap:

Again, if the texture is a bitmap with background=0 and foreground=1, and the destination depth is 1, the texture bitmap can be treated as a pixmap (tcolors=NULL).

Note that while a source bitmap or pixmap has the same width and height as the destination, a texture bitmap or pixmap has its own width and height specified in the gx_tile_bitmap structure, and is replicated or clipped as needed.

The function specification f

"Command" indicates the raster operation and transparency as follows:

7-0   raster op
8   0 if source opaque, 1 if source transparent
9   0 if texture opaque, 1 if texture transparent
?-10   unused, must be 0

The raster operation follows the Microsoft and H-P specification. It is an 8-element truth table that specifies the output value for each of the possible 2×2×2 input values as follows:

Bit   Texture   Source   Destination

7   1   1   1
6   1   1   0
5   1   0   1
4   1   0   0
3   0   1   1
2   0   1   0
1   0   0   1
0   0   0   0

Transparency affects the output in the following way. A source or texture pixel is considered transparent if its value is all 1s (for instance, 1 for bitmaps, 0xffffff for 24-bit RGB pixmaps) and the corresponding transparency bit is set in the command. For each pixel, the result of the Boolean operation is written into the destination iff neither the source nor the texture pixel is transparent. (Note that the HP RasterOp specification, on which this is based, specifies that if the source and texture are both all 1s and the command specifies transparent source and opaque texture, the result should be written in the output. We think this is an error in the documentation.)


copy_rop is defined to operate on pixels in RGB space, again following the HP and Microsoft specification. For devices that don't use RGB (or gray-scale with black = 0, white = all 1s) as their native color representation, the implementation of copy_rop must convert to RGB or gray space, do the operation, and convert back (or do the equivalent of this). Here are the copy_rop equivalents of the most important previous imaging calls. We assume the declaration:

static const gx_color_index white2[2] = { 1, 1 };

Note that rop3_S may be replaced by any other Boolean operation. For monobit devices, we assume that black = 1.

/* For all devices: */
(*fill_rectangle)(dev, x, y, w, h, color) ==>

        { gx_color_index colors[2];
          colors[0] = colors[1] = color;
          (*dev_proc(dev, copy_rop))(dev, NULL, 0, 0, gx_no_bitmap_id, colors,
                                     NULL, colors /*irrelevant*/,
                                     x, y, w, h, 0, 0, rop3_S);

/* For black-and-white devices only: */
(*copy_mono)(dev, base, sourcex, sraster, id,
             x, y, w, h, (gx_color_index)0, (gx_color_index)1) ==>

        (*dev_proc(dev, copy_rop))(dev, base, sourcex, sraster, id, NULL,
                                   NULL, white2 /*irrelevant*/,
                                   x, y, w, h, 0, 0, rop3_S);

/* For color devices, where neither color0 nor color1 is gx_no_color_index: */
(*copy_mono)(dev, base, sourcex, sraster, id,
             x, y, w, h, color0, color1) ==>

        { gx_color_index colors[2];
          colors[0] = color0, colors[1] = color1;
          (*dev_proc(dev, copy_rop))(dev, base, sourcex, sraster, id, colors,
                                     NULL, white2 /*irrelevant*/,
                                     x, y, w, h, 0, 0, rop3_S);

/* For black-and-white devices only: */
(*copy_mono)(dev, base, sourcex, sraster, id,
             x, y, w, h, gx_no_color_index, (gx_color_index)1) ==>

        (*dev_proc(dev, copy_rop))(dev, base, sourcex, sraster, id, NULL,
                                   NULL, white2 /*irrelevant*/,
                                   x, y, w, h, 0, 0,
                                   rop3_S | lop_S_transparent);

/* For all devices: */
(*copy_color)(dev, base, sourcex, sraster, id,
              x, y, w, h) ==> [same as first copy_mono above]

/* For black-and-white devices only: */
(*tile_rectangle)(dev, tile, x, y, w, h,
                  (gx_color_index)0, (gx_color_index)1, px, py) ==>

        (*dev_proc(dev, copy_rop))(dev, NULL, 0, 0, gx_no_bitmap_id,
                                   white2 /*irrelevant*/,
                                   tile, NULL,
                                   x, y, w, h, px, py, rop3_T)

Polygon-level drawing

In addition to the pixel-level drawing operations that take integer device coordinates and pure device colors, the driver interface includes higher-level operations that draw polygons using fixed-point coordinates, possibly halftoned colors, and possibly a non-default logical operation.

The fill_* drawing operations all use the center-of-pixel rule: a pixel is colored iff its center falls within the polygonal region being filled. If a pixel center (X+0.5,Y+0.5) falls exactly on the boundary, the pixel is filled iff the boundary is horizontal and the filled region is above it, or the boundary is not horizontal and the filled region is to the right of it.

int (*fill_trapezoid)(gx_device *dev, const  gs_fixed_edge *left, const gs_fixed_edge *right, fixed ybot, fixed ytop, bool swap_axes, const gx_drawing_color *pdcolor, gs_logical_operation_t lop) [OPTIONAL]
Fill a trapezoid. The bottom and top edges are parallel to the x axis, and are defined by ybot and ytop, respectively. The left and right edges are defined by left and right. Both of these represent lines (gs_fixed_edge is defined in gxdevcli.h and consists of gs_fixed_point start and end points). The y coordinates of these lines need not have any specific relation to ybot and ytop. The routine is defined this way so that the filling algorithm can subdivide edges and still guarantee that the exact same pixels will be filled. If swap_axes is set, the meanings of X and Y are interchanged.
int (*fill_parallelogram)(gx_device *dev, fixed px, fixed py, fixed ax, fixed ay, fixed bx, fixed by, const gx_drawing_color *pdcolor, gs_logical_operation_t lop) [OPTIONAL]
Fill a parallelogram whose corners are (px,py), (px+ax,py+ay), (px+bx,py+by), and (px+ax+bx,py+ay+by). There are no constraints on the values of any of the parameters, so the parallelogram may have any orientation relative to the coordinate axes.
int (*fill_triangle)(gx_device *dev, fixed px, fixed py, fixed ax, fixed ay, fixed bx, fixed by, const gx_drawing_color *pdcolor, gs_logical_operation_t lop) [OPTIONAL]
Fill a triangle whose corners are (px,py), (px+ax,py+ay), and (px+bx,py+by).
int (*draw_thin_line)(gx_device *dev, fixed fx0, fixed fy0, fixed fx1, fixed fy1, const gx_drawing_color *pdcolor, gs_logical_operation_t lop) [OPTIONAL]
Draw a one-pixel-wide line from (fx0,fy0) to (fx1,fy1).
int (*draw_line)(gx_device *dev, int x0, int y0, int x1, int y1, gx_color_index color) [OPTIONAL] [OBSOLETE]
This procedure is no longer used: it is replaced by the draw_thin_line procedure. However, still appears in the driver procedure vector for backward compatibility. It should never be called, and drivers should not implement it.

Linear color drawing

Linear color functions allow fast high quality rendering of shadings on continuous tone devices. They implement filling simple areas with a lineary varying color. These functions are not called if the device applies halftones, or uses a non-separable or a non-linear color model.

int (*fill_linear_color_triangle) (dev_t *dev, const gs_fill_attributes *fa, const gs_fixed_point *p0, const gs_fixed_point *p1, const gs_fixed_point *p2, const frac31 *c0, const frac31 *c1, const frac31 *c2) [OPTIONAL]
This function is the highest level one within the linear color function group. It fills a triangle with a linearly varying color. Arguments specify 3 points in the device space - vertices of a triangle, and their colors. The colors are represented as vectors of positive fractional numbers, each of which represents a color component value in the interval [0,1]. The number of components in a vector in the number of color components in the device (process) color model.
The implementation fills entire triangle. The filling rule is same as for Polygon-level drawing. A color for each pixel within the triangle to be computed as a linear interpolation of vertex colors.
The implementation may reject the request if the area or the color appears too complex for filling in a single action. For doing that the implementation returns 0 and must not paint any pixel. In this case the graphics library will perform a subdivision of the area into smaller triangles and call the function again with smaller areas.
Important note : Do not try to decompose the area within the implementation of fill_linear_color_triangle, because it can break the plane coverage contiguity and cause a dropout. Instead request that graphics library to perform the decomposition. The graphics libary is smart enough to do that properly.
Important note : The implementation must handle a special case, when only 2 colors are specified. It happens if p3 one is NULL. This means that the color does not depend on the X coordinate, i.e. it forms a linear gradient along the Y axis. The implementation must not reject (return 0) such cases.
Important note :The device color component value 1 may be represented with several hexadecimal values : 0x7FFF0000, 0x7FFFF000, 0x7FFFFF00, etc., because the precision here exceeds the color precision of the device. To convert a frac31 value into a device color component value, fist drop (ignore) the sign bit, then drop least significant bits - so many ones as you need to fit the device color precision.
Important note : The fa argument may contain the swap_axes bit set. In this case the implementation must swap (transpoze) X and Y axes.
Important note : The implementation must not paint outside the clipping rectangle specified in the fa argument. If fa->swap_axes is true, the clipping rectangle is transposed.
See gx_default_fill_linear_color_triangle in gdevddrw.c as a sample code.
int (*fill_linear_color_trapezoid) (dev_t *dev, const gs_fill_attributes *fa, const gs_fixed_point *p0, const gs_fixed_point *p1, const gs_fixed_point *p2, const gs_fixed_point *p3, const frac31 *c0, const frac31 *c1, const frac31 *c2, const frac31 *c2) [OPTIONAL]
This function is a lower level one within the linear color function group. The default implementation of fill_linear_color_triangle calls this function 1-2 times per triangle. Besides that, this function may be called by the graphics library for other special cases, when a decomposition into triangles appears undiserable.
Rather the prototype can specify a bilinear color, we assume that the implementation handles linear colors only. This means that the implementation can ignore any of c0, c1, c2, c3 . The graphics library takes a special care of the color linearity when calling this function. The reason for passing all 4 color arguments is to avoid color precision problems.
Similarly to fill_linear_color_triangle , this function may be called with only 2 colors, and may reject too comple areas. All those important notes are applicable here.
A sample code may be found in in gxdtfill.h, rather it's a kind of complicated. A linear color function is generated from it as gx_fill_trapezoid_ns_lc with the following template parametres :
#define LINEAR_COLOR 1 
#define EDGE_TYPE gs_linear_color_edge
#define FILL_ATTRS const gs_fill_attributes *
#define SWAP_AXES 0
#define FILL_DIRECT 1
See the helplers init_gradient, step_gradient (defined in in gdevddrw.c), how to manage colors. See check_gradient_overflow (defined in in gdevddrw.c), as an example of an area that can't be painted in a single action due to 64-bits fixed overflows.
int (*fill_linear_color_scanline) (dev_t *dev, const gs_fill_attributes *fa, int i, int j, int w, const frac31 *c0, const int32_t *c0_f, const int32_t *cg_num, int32_t cg_den) [OPTIONAL]
This function is the lowest level one within the linear color function group. It implements filling a scanline with a linearly varying color. The default implementation for fill_linear_color_trapezoid calls this function, and there are no other calls to it from the graphics libary. Thus if the device implements fill_linear_color_triangle and fill_linear_color_trapezoid by own means, this function may be left unimplemented.
i and j specify device coordinates (indices) of the starting pixel of the scanline, w specifies the width of the scanline, i.e. the number of pixels to be painted to the right from the starting pixel, including the starting pixel.
c0 specifies the color for the starting pixel as a vector of fraction values, each of which represents a color value in the interval [0,1].
c0_f specify a fraction part of the color for the starting pixel. See the formula below about using it.
cg_num specify a numerator for the color gradient - a vector of values in [-1,1], each of which correspond to a color component.
cg_den specify the denominator for the color gradient - a value in [-1,1].

The color for the pixel [i + k, j] to be computed like this :

         (double)(c0[n] + (c0_f[n] + cg_num[n] * k) / cg_den) / (1 ^ 31 - 1)
where 0 <= k <= w , and n is a device color component index.
Important note : The fa argument may contain the swap_axes bit set. In this case the implementation must swap (transpose) X and Y axes.
Important note : The implementation must not paint outside the clipping rectangle specified in the fa argument. If fa->swap_axes is true, the clipping rectangle is transposed.
See gx_default_fill_linear_color_scanline in gdevdsha.c as a sample code.

High-level drawing

In addition to the lower-level drawing operations described above, the driver interface provides a set of high-level operations. Normally these will have their default implementation, which converts the high-level operation to the low-level ones just described; however, drivers that generate high-level output formats such as CGM, or communicate with devices that have firmware for higher-level operations such as polygon fills, may implement these high-level operations directly. For more details, please consult the source code, specifically:

Header     Defines
gxpaint.h   gx_fill_params, gx_stroke_params
gxfixed.h   fixed, gs_fixed_point (used by gx_*_params)
gxistate.h   gs_imager_state (used by gx_*_params)
gxline.h   gx_line_params (used by gs_imager_state)
gslparam.h   line cap/join values (used by gx_line_params)
gxmatrix.h   gs_matrix_fixed (used by gs_imager_state)
gspath.h, gxpath.h, gzpath.h   gx_path
gxcpath.h, gzcpath.h   gx_clip_path

For a minimal example of how to implement the high-level drawing operations, see gdevtrac.c.


int (*fill_path)(gx_device *dev, const gs_imager_state *pis, gx_path *ppath, const gx_fill_params *params, const gx_drawing_color *pdcolor, const gx_clip_path *pcpath) [OPTIONAL]
Fill the given path, clipped by the given clip path, according to the given parameters, with the given color. The clip path pointer may be NULL, meaning do not clip.
The implementation must paint the path with the specified device color, which may be either a pure color, or a pattern. If the device can't handle non-pure colors, it should check the color type and call the default implementation gx_default_fill_path for cases which it can't handle. The default implementation will perform a subdivision of the area to be painted, and will call other device virtual functions (such as fill_linear_color_triangle) with simpler areas.
int (*stroke_path)(gx_device *dev, const gs_imager_state *pis, gx_path *ppath, const gx_stroke_params *params, const gx_drawing_color *pdcolor, const gx_clip_path *pcpath) [OPTIONAL]
Stroke the given path, clipped by the given clip path, according to the given parameters, with the given color. The clip path pointer may be NULL, meaning not to clip.
int (*fill_mask)(gx_device *dev, const byte *data, int data_x, int raster, gx_bitmap_id id, int x, int y, int width, int height, const gx_drawing_color *pdcolor, int depth, int command, const gx_clip_path *pcpath) [OPTIONAL]
Color the 1-bits in the given mask (or according to the alpha values, if depth > 1), clipped by the given clip path, with the given color and logical operation. The clip path pointer may be NULL, meaning do not clip. The parameters data, ..., height are as for copy_mono; depth is as for copy_alpha; command is as for copy_rop.


Similar to the high-level interface for fill and stroke graphics, a high-level interface exists for bitmap images. The procedures in this part of the interface are optional.

Bitmap images come in a variety of types, corresponding closely (but not precisely) to the PostScript ImageTypes. The generic or common part of all bitmap images is defined by:

typedef struct {
	const gx_image_type_t *type;
        gs_matrix ImageMatrix;
} gs_image_common_t;

Bitmap images that supply data (all image types except image_type_from_device (2)) are defined by:

#define gs_image_max_components 5
typedef struct {
        << gs_image_common_t >>
        int Width;
        int Height;
        int BitsPerComponent;
        float Decode[gs_image_max_components * 2];
        bool Interpolate;
} gs_data_image_t;

Images that supply pixel (as opposed to mask) data are defined by:

typedef enum {
	/* Single plane, chunky pixels. */
	gs_image_format_chunky = 0,
	/* num_components planes, chunky components. */
	gs_image_format_component_planar = 1,
	/* BitsPerComponent * num_components planes, 1 bit per plane */
	gs_image_format_bit_planar = 2
} gs_image_format_t;
typedef struct {
        << gs_data_image_t >>
        const gs_color_space *ColorSpace;
        bool CombineWithColor;
} gs_pixel_image_t;

Ordinary PostScript Level 1 or Level 2 (ImageType 1) images are defined by:

typedef enum {
	/* No alpha. */
	gs_image_alpha_none = 0,
	/* Alpha precedes color components. */
	/* Alpha follows color components. */
} gs_image_alpha_t;
typedef struct {
        << gs_pixel_image_t >>
        bool ImageMask;
        bool adjust;
	gs_image_alpha_t Alpha;
} gs_image1_t;
typedef gs_image1_t gs_image_t;

Of course, standard PostScript images don't have an alpha component. For more details, consult the source code in gsiparam.h and gsiparm*.h, which define parameters for an image.

The begin[_typed_]image driver procedures create image enumeration structures. The common part of these structures consists of:

typedef struct gx_image_enum_common_s {
        const gx_image_type_t *image_type;
	const gx_image_enum_procs_t *procs;
	gx_device *dev;
	gs_id id;
        int num_planes;
        int plane_depths[gs_image_max_planes];  /* [num_planes] */
	int plane_widths[gs_image_max_planes]	/* [num_planes] */
} gx_image_enum_common_t;

where procs consists of:

typedef struct gx_image_enum_procs_s {

         * Pass the next batch of data for processing.
#define image_enum_proc_plane_data(proc)\
  int proc(gx_device *dev,\
    gx_image_enum_common_t *info, const gx_image_plane_t *planes,\
    int height)


         * End processing an image, freeing the enumerator.
#define image_enum_proc_end_image(proc)\
  int proc(gx_device *dev,\
    gx_image_enum_common_t *info, bool draw_last)


	 * Flush any intermediate buffers to the target device.
	 * We need this for situations where two images interact
	 * (currently, only the mask and the data of ImageType 3).
	 * This procedure is optional (may be 0).
#define image_enum_proc_flush(proc)\
  int proc(gx_image_enum_common_t *info)


} gx_image_enum_procs_t;

In other words, begin[_typed]_image sets up an enumeration structure that contains the procedures that will process the image data, together with all variables needed to maintain the state of the process. Since this is somewhat tricky to get right, if you plan to create one of your own you should probably read an existing implementation of begin[_typed]_image, such as the one in gdevbbox.c or gdevps.c.

The data passed at each call of image_plane_data consists of one or more planes, as appropriate for the type of image. begin[_typed]_image must initialize the plane_depths array in the enumeration structure with the depths (bits per element) of the planes. The array of gx_image_plane_t structures passed to each call of image_plane_data then defines where the data are stored, as follows:

typedef struct gx_image_plane_s {
  const byte *data;
  int data_x;
  uint raster;
} gx_image_plane_t;
int (*begin_image)(gx_device *dev, const gs_imager_state *pis, const gs_image_t *pim, gs_image_format_t format, gs_int_rect *prect, const gx_drawing_color *pdcolor, const gx_clip_path *pcpath, gs_memory_t *memory, gx_image_enum_common_t **pinfo) [OPTIONAL]
Begin the transmission of an image. Zero or more calls of image_plane_data will follow, and then a call of end_image. The parameters of begin_image are as follows:
pis     pointer to an imager state. The only relevant elements of the imager state are the CTM (coordinate transformation matrix), the logical operation (RasterOp or transparency), and the color rendering information.
pim   pointer to the gs_image_t structure that defines the image parameters
format   defines how pixels are represented for image_plane_data. See the description of image_plane_data below
prect   if not NULL, defines a subrectangle of the image; only the data for this subrectangle will be passed to image_plane_data, and only this subrectangle should be drawn
pdcolor   defines a drawing color, only needed for masks or if CombineWithColor is true
pcpath   if not NULL, defines an optional clipping path
memory   defines the allocator to be used for allocating bookkeeping information
pinfo   the implementation should return a pointer to its state structure here

begin_image is expected to allocate a structure for its bookkeeping needs, using the allocator defined by the memory parameter, and return it in *pinfo. begin_image should not assume that the structures in *pim, *prect, or *pdcolor will survive the call on begin_image (except for the color space in *pim->ColorSpace): it should copy any necessary parts of them into its own bookkeeping structure. It may, however, assume that *pis, *pcpath, and of course *memory will live at least until end_image is called.

begin_image returns 0 normally, or 1 if the image does not need any data. In the latter case, begin_image does not allocate an enumeration structure.

int (*begin_typed_image)(gx_device *dev, const gs_imager_state *pis, const gs_matrix *pmat, const gs_image_common_t *pim, gs_int_rect *prect, const gx_drawing_color *pdcolor, const gx_clip_path *pcpath, gs_memory_t *memory, gx_image_enum_common_t **pinfo) [OPTIONAL]
This has the same function as begin_image, except
  • The image may be of any ImageType, not only image_type_simple (1);
  • The image format is included in the image structure, not supplied as a separate argument;
  • The optional pmat argument provides a matrix that substitutes for the one in the imager state;
  • For mask images, if pmat is not NULL and the color is pure, pis may be NULL.

The actual transmission of data uses the procedures in the enumeration structure, not driver procedures, since the handling of the data usually depends on the image type and parameters rather than the device. These procedures are specified as follows.

int (*image_plane_data)(gx_device *dev, gx_image_enum_common_t *info, const gx_image_plane_t *planes, int height)
This call provides more of the image source data: specifically, height rows, with Width pixels supplied for each row.

The data for each row are packed big-endian within each byte, as for copy_color. The data_x (starting X position within the row) and raster (number of bytes per row) are specified separately for each plane, and may include some padding at the beginning or end of each row. Note that for non-mask images, the input data may be in any color space and may have any number of bits per component (1, 2, 4, 8, 12); currently mask images always have 1 bit per component, but in the future, they might allow multiple bits of alpha. Note also that each call of image_plane_data passes complete pixels: for example, for a chunky image with 24 bits per pixel, each call of image_plane_data passes 3N bytes of data (specifically, 3 × Width × height).

The interpretation of planes depends on the format member of the gs_image[_common]_t structure:

  • If the format is gs_image_format_chunky, planes[0].data points to data in "chunky" format, in which the components follow each other (for instance, RGBRGBRGB....)
  • If the format is gs_image_format_component_planar, planes[0 .. N-1].data point to data for the N components (for example, N=3 for RGB data); each plane contains samples for a single component, for instance, RR..., GG..., BB.... Note that the planes are divided by component, not by bit: for example, for 24-bit RGB data, N=3, with 8-bit values in each plane of data.
  • If the format is gs_image_format_bit_planar, planes[0 .. N*B-1].data point to data for the N components of B bits each (for example, N=3 and B=4 for RGB data with 4 bits per component); each plane contains samples for a single bit, for instance, R0 R1 R2 R3 G0 G1 G2 G3 B0 B1 B2 B3. Note that the most significant bit of each plane comes first.

If, as a result of this call, image_plane_data has been called with all the data for the (sub-)image, it returns 1; otherwise, it returns 0 or an error code as usual.

image_plane_data, unlike most other procedures that take bitmaps as arguments, does not require the data to be aligned in any way.

Note that for some image types, different planes may have different numbers of bits per pixel, as defined in the plane_depths array.

int (*end_image)(gx_device *dev, void *info, bool draw_last)
Finish processing an image, either because all data have been supplied or because the caller has decided to abandon this image. end_image may be called at any time after begin_image. It should free the info structure and any subsidiary structures. If draw_last is true, it should finish drawing any buffered lines of the image.

While there will almost never be more than one image enumeration in progress -- that is, after a begin_image, end_image will almost always be called before the next begin_image -- driver code should not rely on this property; in particular, it should store all information regarding the image in the info structure, not in the driver structure.

Note that if begin_[typed_]image saves its parameters in the info structure, it can decide on each call whether to use its own algorithms or to use the default implementation. (It may need to call gx_default_begin/end_image partway through.) [A later revision of this document may include an example here.]


The third high-level interface handles text. As for images, the interface is based on creating an enumerator which then may execute the operation in multiple steps. As for the other high-level interfaces, the procedures are optional.

int (*text_begin)(gx_device *dev, gs_imager_state *pis, const gs_text_params_t *text, gs_font *font, gx_path *path, const gx_device_color *pdcolor, const gx_clip_path *pcpath, gs_memory_t *memory, gs_text_enum_t **ppte) [OPTIONAL]
Begin processing text, by creating a state structure and storing it in *ppte. The parameters of text_begin are as follows:
dev     The usual pointer to the device.
pis     A pointer to an imager state. All elements may be relevant, depending on how the text is rendered.
text   A pointer to the structure that defines the text operation and parameters. See gstext.h for details.
font   Defines the font for drawing.
path   Defines the path where the character outline will be appended (if the text operation includes TEXT_DO_...PATH), and whose current point indicates where drawing should occur and will be updated by the string width (unless the text operation includes TEXT_DO_NONE).
pdcolor   Defines the drawing color for the text. Only relevant if the text operation includes TEXT_DO_DRAW.
pcpath   If not NULL, defines an optional clipping path. Only relevant if the text operation includes TEXT_DO_DRAW.
memory   Defines the allocator to be used for allocating bookkeeping information.
ppte   The implementation should return a pointer to its state structure here.

text_begin must allocate a structure for its bookkeeping needs, using the allocator defined by the memory parameter, and return it in *ppte. text_begin may assume that the structures passed as parameters will survive until text processing is complete.

Clients should not call the driver text_begin procedure directly. Instead, they should call gx_device_text_begin, which takes the same parameters and also initializes certain common elements of the text enumeration structure, or gs_text_begin, which takes many of the parameters from a graphics state structure. For details, see gstext.h.

The actual processing of text uses the procedures in the enumeration structure, not driver procedures, since the handling of the text may depend on the font and parameters rather than the device. Text processing may also require the client to take action between characters, either because the client requested it (TEXT_INTERVENE in the operation) or because rendering a character requires suspending text processing to call an external package such as the PostScript interpreter. (It is a deliberate design decision to handle this by returning to the client, rather than calling out of the text renderer, in order to avoid potentially unknown stack requirements.) Specifically, the client must call the following procedures, which in turn call the procedures in the text enumerator.

int gs_text_process(gs_text_enum_t *pte)
Continue processing text. This procedure may return 0 or a negative error code as usual, or one of the following values (see gstext.h for details).
TEXT_PROCESS_RENDER The client must cause the current character to be rendered. This currently only is used for PostScript Type 0-4 fonts and their CID-keyed relatives.
TEXT_PROCESS_INTERVENE The client has asked to intervene between characters. This is used for cshow and kshow.
int gs_text_release(gs_text_enum_t *pte, client_name_t cname)
Finish processing text and release all associated structures. Clients must call this procedure after gs_text_process returns 0 or an error, and may call it at any time.

There are numerous other procedures that clients may call during text processing. See gstext.h for details.


Note that unlike many other optional procedures, the default implementation of text_begin cannot simply return: like the default implementation of begin[_typed]_image, it must create and return an enumerator. Furthermore, the implementation of the process procedure (in the enumerator structure, called by gs_text_process) cannot simply return without doing anything, even if it doesn't want to draw anything on the output. See the comments in gxtext.h for details.

Unicode support for high level devices

Implementing a new high level device, one may need to translate Postscript character codes into Unicode. This can be done pretty simply.

For translating a Postscript text you need to inplement the device virtual function text_begin. It should create a new instance of gs_text_enum_t in the heap (let its pointer be pte), and assign a special function to gs_text_enum_t::procs.process. The function will receive pte. It should take the top level font from pte->orig_font, and iterate with font->procs.next_char_glyph(pte, ..., &glyph). The last argument receives a gs_glyph value, which encodes a Postscript character name or CID (and also stores it into pte->returned.current_glyph). Then obtain the current subfont with gs_text_current_font(pte) (it can differ from the font) and call subfont->procs.decode_glyph(subfont, glyph). The return value will be an Unicode code, or GS_NO_CHAR if the glyph can't be translated to Unicode.

Reading bits back

int (*get_bits_rectangle)(gx_device *dev, const gs_int_rect *prect, gs_get_bits_params_t *params, gs_int_rect **unread) [OPTIONAL]
Read a rectangle of bits back from the device. The params structure consists of:
options   the allowable formats for returning the data
data[32]   pointers to the returned data
x_offset   the X offset of the first returned pixel in data
raster   the distance between scan lines in the returned data

options is a bit mask specifying what formats the client is willing to accept. (If the client has more flexibility, the implementation may be able to return the data more efficiently, by avoiding representation conversions.) The options are divided into groups.

Specifies whether the returned data must be aligned in the normal manner for bitmaps, or whether unaligned data are acceptable.
pointer or copy
Specifies whether the data may be copied into storage provided by the client and/or returned as pointers to existing storage. (Note that if copying is not allowed, it is much more likely that the implementation will return an error, since this requires that the client accept the data in the implementation's internal format.)
X offset
Specifies whether the returned data must have a specific X offset (usually zero, but possibly other values to avoid skew at some later stage of processing) or whether it may have any X offset (which may avoid skew in the get_bits_rectangle operation itself).
Specifies whether the raster (distance between returned scan lines) must have its standard value, must have some other specific value, or may have any value. The standard value for the raster is the device width padded out to the alignment modulus when using pointers, or the minimum raster to accommodate the X offset + width when copying (padded out to the alignment modulus if standard alignment is required).
Specifies whether the data are returned in chunky (all components of a single pixel together), component-planar (each component has its own scan lines), or bit-planar (each bit has its own scan lines) format.
color space
Specifies whether the data are returned as native device pixels, or in a standard color space. Currently the only supported standard space is RGB.
standard component depth
Specifies the number of bits per component if the data are returned in the standard color space. (Native device pixels use dev->color_info.depth bits per pixel.)
Specifies whether alpha channel information should be returned as the first component, the last component, or not at all. Note that for devices that have no alpha capability, the returned alpha values will be all 1s.

The client may set more than one option in each of the above groups; the implementation will choose one of the selected options in each group to determine the actual form of the returned data, and will update params[].options to indicate the form. The returned params[].options will normally have only one option set per group.

For further details on params, see gxgetbit.h. For further details on options, see gxbitfmt.h.

Define w = prect->q.x - prect->p.x, h = prect->q.y - prect->p.y. If the bits cannot be read back (for example, from a printer), return gs_error_unknownerror; if raster bytes is not enough space to hold offset_x + w pixels, or if the source rectangle goes outside the device dimensions (p.x < 0 || p.y < 0 || q.x > dev->width || q.y > dev->height), return gs_error_rangecheck; if any regions could not be read, return gs_error_ioerror if unpainted is NULL, otherwise the number of rectangles (see below); otherwise return 0.

The caller supplies a buffer of raster × h bytes starting at data[0] for the returned data in chunky format, or N buffers of raster × h bytes starting at data[0] through data[N-1] in planar format where N is the number of components or bits. The contents of the bits beyond the last valid bit in each scan line (as defined by w) are unpredictable. data need not be aligned in any way. If x_offset is non-zero, the bits before the first valid bit in each scan line are undefined. If the implementation returns pointers to the data, it stores them into data[0] or data[0..N-1].

If not all the source data are available (for example, because the source was a partially obscured window and backing store was not available or not used), or if the rectangle does not fall completely within the device's coordinate system, any unread bits are undefined, and the value returned depends on whether unread is NULL. If unread is NULL, return gs_error_ioerror; in this case, some bits may or may not have been read. If unread is not NULL, allocate (using dev->memory) and fill in a list of rectangles that could not be read, store the pointer to the list in *unread, and return the number of rectangles; in this case, all bits not listed in the rectangle list have been read back properly. The list is not sorted in any particular order, but the rectangles do not overlap. Note that the rectangle list may cover a superset of the region actually obscured: for example, a lazy implementation could return a single rectangle that was the bounding box of the region.

int (*get_bits)(gx_device *dev, int y, byte *data, byte **actual_data) [OPTIONAL]
Read scan line y of bits back from the device into the area starting at data. This call is functionally equivalent to
  (dev, {0, y, dev->width, y+1},

with the returned value of params->data[0] stored in *actual_data, and will in fact be implemented this way if the device defines a get_bits_rectangle procedure and does not define one for get_bits. (If actual_data is NULL, GB_RETURN_POINTER is omitted from the options.)


Devices may have an open-ended set of parameters, which are simply pairs consisting of a name and a value. The value may be of various types: integer (int or long), boolean, float, string, name, NULL, array of integer, array of float, or arrays or dictionaries of mixed types. For example, the Name of a device is a string; the Margins of a device is an array of two floats. See gsparam.h for more details.

If a device has parameters other than the ones applicable to all devices (or, in the case of printer devices, all printer devices), it must provide get_params and put_params procedures. If your device has parameters beyond those of a straightforward display or printer, we strongly advise using the _get_params and _put_params procedures in an existing device (for example, gdevcdj.c or gdevbit.c) as a model for your own code.

int (*get_params)(gx_device *dev, gs_param_list *plist) [OPTIONAL]
Read the parameters of the device into the parameter list at plist, using the param_write_* macros or procedures defined in gsparam.h.
int (*get_hardware_params)(gx_device *dev, gs_param_list *plist) [OPTIONAL]
Read the hardware-related parameters of the device into the parameter list at plist. These are any parameters whose values are under control of external forces rather than the program -- for example, front panel switches, paper jam or tray empty sensors, etc. If a parameter involves significant delay or hardware action, the driver should only determine the value of the parameter if it is "requested" by the gs_param_list [param_requested(plist, key_name)]. This function may cause the asynchronous rendering pipeline (if enabled) to be drained, so it should be used sparingly.
int (*put_params)(gx_device *dev, gs_param_list *plist) [OPTIONAL]
Set the parameters of the device from the parameter list at plist, using the param_read_* macros/procedures defined in gsparam.h. All put_params procedures must use a "two-phase commit" algorithm; see gsparam.h for details.

Default color rendering dictionary (CRD) parameters

Drivers that want to provide one or more default CIE color rendering dictionaries (CRDs) can do so through get_params. To do this, they create the CRD in the usual way (normally using the gs_cie_render1_build and _initialize procedures defined in gscrd.h), and then write it as a parameter using param_write_cie_render1 defined in gscrdp.h. However, the TransformPQR procedure requires special handling. If the CRD uses a TransformPQR procedure different from the default (identity), the driver must do the following:

For a complete example, see the bit_get_params procedure in gdevbit.c. Note that it is essential that the driver return the CRD or the procedure address only if specifically requested (param_requested(...) > 0); otherwise, errors will occur.

External fonts

Drivers may include the ability to display text. More precisely, they may supply a set of procedures that in turn implement some font and text handling capabilities, described in a separate document. The link between the two is the driver procedure that supplies the font and text procedures:

xfont_procs *(*get_xfont_procs)(gx_device *dev) [OPTIONAL]
Return a structure of procedures for handling external fonts and text display. A NULL value means that this driver doesn't provide this capability.

For technical reasons, a second procedure is also needed:

gx_device *(*get_xfont_device)(gx_device *dev) [OPTIONAL]
Return the device that implements get_xfont_procs in a non-default way for this device, if any. Except for certain special internal devices, this is always the device argument.

Page devices

gx_device *(*get_page_device)(gx_device *dev) [OPTIONAL]
According to the Adobe specifications, some devices are "page devices" and some are not. This procedure returns NULL if the device is not a page device, or the device itself if it is a page device. In the case of forwarding devices, get_page_device returns the underlying page device (or NULL if the underlying device is not a page device).


int (*get_band)(gx_device *dev, int y, int *band_start) [OPTIONAL]
If the device is a band device, this procedure stores in *band_start the scan line (device Y coordinate) of the band that includes the given Y coordinate, and returns the number of scan lines in the band. If the device is not a band device, this procedure returns 0. The latter is the default implementation.
void (*get_clipping_box)(gx_device *dev, gs_fixed_rect *pbox)) [OPTIONAL]
Stores in *pbox a rectangle that defines the device's clipping region. For all but a few specialized devices, this is ((0,0),(width,height)).

Tray selection

The logic for selecting input trays, and modifying other parameters based on tray selection, can be complex and subtle, largely thanks to the requirement to be compatible with the PostScript language setpagedevice mechanism. This section will describe recipes for several common scenarios for tray selection, with special attention to the how the overall task factors into configuration options, generic logic provided by the PostScript language (or not, if the device is used with other PDL's), and implementation of the put_param / get_param device functions within the device.

In general, tray selection is determined primarily through the setpagedevice operator, which is part of the PostScript runtime. Ghostscript attempts to be as compatible as is reasonable with the PostScript standard, so for more details, see the description in the PostScript language specifications, including the "supplements", which tend to have more detail about setpagedevice behavior than the PLRM book itself.

The first step is to set up an /InputAttributes dictionary matching the trays and so on available in the device. The standard Ghostscript initialization files set up a large InputAttributes dictionary with many "known" page sizes (the full list is in, under .setpagesize). It's possible to edit this list in the Ghostscript source, of course, but most of the time it is better to execute a snippet of PostScript code after the default initialization but before sending any actual jobs.

Simply setting a new /InputAttributes dictionary with setpagedevice will not work, because the the language specification for setpagedevice demands a "merging" behavior - paper tray keys present in the old dictionary will be preserved even if the key is not present in the new /InputAttributes dictionary. Here is a sample invocation that clears out all existing keys, and installs three new ones: a US letter page size for trays 0 and 1, and 11x17 for tray 1. Note that you must add at least one valid entry into the /InputAttributes dictionary; if all are null, then the setpagedevice will fail with a /configurationerror.

<< /InputAttributes
  currentpagedevice /InputAttributes get
  dup { pop 1 index exch null put } forall

  dup 0 << /PageSize [612 792] >> put
  dup 1 << /PageSize [612 792] >> put
  dup 2 << /PageSize [792 1224] >> put
>> setpagedevice

After this code runs, then requesting a letter page size (612x792 points) from setpagedevice will select tray 0, and requesting an 11x17 size will select tray 2. To explicitly request tray 1, run:

<< /PageSize [612 792] /MediaPosition 1 >> setpagedevice

At this point, the chosen tray is sent to the device as the (nonstandard) %MediaSource device parameter. Devices with switchable trays should implement this device parameter in the put_params procedure. Unlike the usual protocol for device parameters, it is not necessary for devices to also implement get_params querying of this paramter; it is effectively a write-only communication from the language to the device. Currently, among the devices that ship with Ghostscript, only PCL (gdevdjet.c) and PCL/XL (gdevpx.c) implement this parameter, but that list may well grow over time. If the device has dynamic configuration of trays, etc., then the easiest way to get that information into the tray selection logic is to send a setpagedevice request (if using the standard API, then using gsapi_run_string_continue) to update the /InputAttributes dictionary immediately before beginning a job.

Tray rotation and the LeadingEdge parameter

Large, sophisticated printers often have multiple trays supporting both short-edge and long-edge feed. For example, if the paper path is 11 inches wide, then 11x17 pages must always print short-edge, but letter size pages print with higher throughput if fed from long-edge trays. Generally, the device will expect the rasterized bitmap image to be rotated with respect to the page, so that it's always the same orientation with respect to the paper feed direction.

The simplest way to achieve this behavior is to call gx_device_request_leadingedge to request a LeadingEdge value LeadingEdge field in the device structure based on the %MediaSource tray selection index and knowledge of the device's trays. The default put_params implementation will then handle this request (it's done this way to preserve the transactional semantics of put_params; it needs the new value, but the changes can't actually be made until all params succeed). For example, if tray 0 is long-edge, while trays 1 and 2 are short-edge, the following code outline should select the appropriate rotation:

my_put_params(gx_device *pdev, gs_param_list *plist) {
    my_device *dev = (my_device *)pdev;
    int MediaSource = dev->myMediaSource;

    code = param_read_int(plist, "%MediaSource", &MediaSource);

    switch (MediaSource) {
    case 0:
        gx_device_req_leadingedge(dev, 1);
    case 1:
    case 2:
        gx_device_req_leadingedge(dev, 0);
    } default put_params, which makes the change...

    dev->myMediaSource = MediaSource;
    return 0;

Ghostscript also supports explicit rotation of the page through setting the /LeadingEdge parameter with setpagedevice. The above code snippet will simply override this request. To give manual setting through setpagedevice priority, don't change the LeadingEdge field in the device if its LEADINGEDGE_SET_MASK bit is set. In other words, simply enclose the above switch statement inside an if (!(dev->LeadingEdge & LEADINGEDGE_SET_MASK) { ... } statement.

Interaction between LeadingEdge and PageSize

As of LanguageLevel 3, PostScript now has two mechanisms for rotating the imaging of the page: the LeadingEdge parameter described in detail above, and the automatic rotation as enabled by the /PageSize page device parameter (described in detail in Table 6.2 of the PLRM3). Briefly, the PageSize autorotation handles the case where the page size requested in setpagedevice matches the swapped size of the paper source (as set in the InputAttributesDictionary). This mechanism can be, and has been, used to implement long-edge feed, but has several disadvantages. Among other things, it's overly tied to the PostScript language, while the device code above will work with other languages. Also, it only specifies one direction of rotation (90 degrees counterclockwise). Thus, given the choice, LeadingEdge is to be preferred.

If PageSize is used, the following things are different:

Copyright © 2000-2007 Artifex Software, Inc. All rights reserved.

This software is provided AS-IS with no warranty, either express or implied. This software is distributed under license and may not be copied, modified or distributed except as expressly authorized under the terms of that license. Refer to licensing information at or contact Artifex Software, Inc., 7 Mt. Lassen Drive - Suite A-134, San Rafael, CA 94903, U.S.A., +1(415)492-9861, for further information.

Ghostscript version 8.62, 29 February 2008