A description on how to use and modify libpng


libpng version 1.0.3 - January 14, 1999
Updated and distributed by Glenn Randers-Pehrson
glennrp@users.sf.net
Copyright (c) 1998, 1999 Glenn Randers-Pehrson
For conditions of distribution and use, see copyright notice in png.h.

based on:

libpng 1.0 beta 6 version 0.96 May 28, 1997
Updated and distributed by Andreas Dilger
Copyright (c) 1996, 1997 Andreas Dilger

libpng 1.0 beta 2 - version 0.88 January 26, 1996
For conditions of distribution and use, see copyright notice in png.h.
Copyright (c) 1995, 1996 Guy Eric
Schalnat, Group 42, Inc.

Updated/rewritten per request in the libpng FAQ
Copyright (c) 1995 Frank J. T. Wojcik
December 18, 1995 & January 20, 1996

I. INTRODUCTION

This file describes how to use and modify the PNG reference library (known as libpng) for your own use. There are five sections to this file: introduction, structures, reading, writing, and modification and configuration notes for various special platforms. In addition to this file, example.c is a good starting point for using the library, as it is heavily commented and should include everything most people will need. We assume that libpng is already installed; see the INSTALL file for instructions on how to install libpng.

Libpng was written as a companion to the PNG specification, as a way of reducing the amount of time and effort it takes to support the PNG file format in application programs. The PNG specification is available as RFC 2083 ftp://ftp.uu.net/graphics/png/documents/ and as a W3C Recommendation http://www.w3.org/TR/REC.png.html. Some additional chunks are described in the special-purpose public chunks documents at ftp://ftp.uu.net/graphics/png/documents/. Other information about PNG, and the latest version of libpng, can be found at the PNG home page, http://www.cdrom.com/pub/png/.

Most users will not have to modify the library significantly; advanced users may want to modify it more. All attempts were made to make it as complete as possible, while keeping the code easy to understand. Currently, this library only supports C. Support for other languages is being considered.

Libpng has been designed to handle multiple sessions at one time, to be easily modifiable, to be portable to the vast majority of machines (ANSI, K&R, 16-, 32-, and 64-bit) available, and to be easy to use. The ultimate goal of libpng is to promote the acceptance of the PNG file format in whatever way possible. While there is still work to be done (see the TODO file), libpng should cover the majority of the needs of its users.

Libpng uses zlib for its compression and decompression of PNG files. Further information about zlib, and the latest version of zlib, can be found at the zlib home page, http://www.cdrom.com/pub/infozip/zlib/. The zlib compression utility is a general purpose utility that is useful for more than PNG files, and can be used without libpng. See the documentation delivered with zlib for more details. You can usually find the source files for the zlib utility wherever you find the libpng source files.

Libpng is thread safe, provided the threads are using different instances of the structures. Each thread should have its own png_struct and png_info instances, and thus its own image. Libpng does not protect itself against two threads using the same instance of a structure.

II. STRUCTURES

There are two main structures that are important to libpng, png_struct and png_info. The first, png_struct, is an internal structure that will not, for the most part, be used by a user except as the first variable passed to every libpng function call.

The png_info structure is designed to provide information about the PNG file. At one time, the fields of png_info were intended to be directly accessible to the user. However, this tended to cause problems with applications using dynamically loaded libraries, and as a result a set of interface functions for png_info was developed. The fields of png_info are still available for older applications, but it is suggested that applications use the new interfaces if at all possible.

The png.h header file is an invaluable reference for programming with libpng. And while I'm on the topic, make sure you include the libpng header file:

#include <png.h>

III. READING

Reading PNG files:

We'll now walk you through the possible functions to call when reading in a PNG file, briefly explaining the syntax and purpose of each one. See example.c and png.h for more detail. While Progressive reading is covered in the next section, you will still need some of the functions discussed in this section to read a PNG file.

You will want to do the I/O initialization(*) before you get into libpng, so if it doesn't work, you don't have much to undo. Of course, you will also want to insure that you are, in fact, dealing with a PNG file. Libpng provides a simple check to see if a file is a PNG file. To use it, pass in the first 1 to 8 bytes of the file, and it will return true or false (1 or 0) depending on whether the bytes could be part of a PNG file. Of course, the more bytes you pass in, the greater the accuracy of the prediction.

If you are intending to keep the file pointer open for use in libpng, you must ensure you don't read more than 8 bytes from the beginning of the file, and you also have to make a call to png_set_sig_bytes_read() with the number of bytes you read from the beginning. Libpng will then only check the bytes (if any) that your program didn't read.

(*): If you are not using the standard I/O functions, you will need to replace them with custom functions. See the discussion under Customizing libpng.

FILE *fp = fopen(file_name, "rb");
    if (!fp) return;
    fread(header, 1, number, fp);
    is_png = !png_sig_cmp(header, 0, number);
    if (!is_png) return;

Next, png_struct and png_info need to be allocated and initialized. In order to ensure that the size of these structures is correct even with a dynamically linked libpng, there are functions to initialize and allocate the structures. We also pass the library version, optional pointers to error handling functions, and a pointer to a data struct for use by the error functions, if necessary (the pointer and functions can be NULL if the default error handlers are to be used). See the section on Changes to Libpng below regarding the old initialization functions.

png_structp png_ptr = png_create_read_struct
       (PNG_LIBPNG_VER_STRING, (png_voidp)user_error_ptr,
        user_error_fn, user_warning_fn);
    if (!png_ptr)
        return;

    png_infop info_ptr = png_create_info_struct(png_ptr);
    if (!info_ptr) {
        png_destroy_read_struct(&png_ptr,
           (png_infopp)NULL, (png_infopp)NULL);
        return;
    }

    png_infop end_info = png_create_info_struct(png_ptr);
    if (!end_info) {
        png_destroy_read_struct(&png_ptr, &info_ptr,
          (png_infopp)NULL);
        return;
    }

If you want to use your own memory allocation routines, define PNG_USER_MEM_SUPPORTED and use png_create_read_struct_2() instead of png_create_read_struct():

png_structp png_ptr = png_create_read_struct_2
       (PNG_LIBPNG_VER_STRING, (png_voidp)user_error_ptr,
        user_error_fn, user_warning_fn, (png_voidp)
        user_mem_ptr, user_malloc_fn, user_free_fn);

The error handling routines passed to png_create_read_struct() and the memory alloc/free routines passed to png_create_struct_2() are only necessary if you are not using the libpng supplied error handling and memory alloc/free functions.

When libpng encounters an error, it expects to longjmp back to your routine. Therefore, you will need to call setjmp and pass your png_ptr->jmpbuf. If you read the file from different routines, you will need to update the jmpbuf field every time you enter a new routine that will call a png_ function.

See your documentation of setjmp/longjmp for your compiler for more handling in the Customizing Libpng section below for more information on the libpng error handling. If an error occurs, and libpng longjmp's back to your setjmp, you will want to call png_destroy_read_struct() to free any memory.

if (setjmp(png_ptr->jmpbuf)) {
        png_destroy_read_struct(&png_ptr, &info_ptr,
           &end_info);
        fclose(fp);
        return;
    }

Now you need to set up the input code. The default for libpng is to use the C function fread(). If you use this, you will need to pass a valid FILE * in the function png_init_io(). Be sure that the file is opened in binary mode. If you wish to handle reading data in another way, you need not call the png_init_io() function, but you must then implement the libpng I/O methods discussed in the Customizing Libpng section below.

png_init_io(png_ptr, fp);

If you had previously opened the file and read any of the signature from the beginning in order to see if this was a PNG file, you need to let libpng know that there are some bytes missing from the start of the file.

png_set_sig_bytes(png_ptr, number);

At this point, you can set up a callback function that will be called after each row has been read, which you can use to control a progress meter or the like. It's demonstrated in pngtest.c. You must supply a function

void
    read_row_callback(png_ptr, png_uint_32 row, int pass) {
      /* put your code here */
    }

(You can give it another name that you like instead of "read_row_callback")

To inform libpng about your function, use

png_set_read_status_fn(png_ptr, read_row_callback);

In PNG files, the alpha channel in an image is the level of opacity. If you need the alpha channel in an image to be the level of transparency instead of opacity, you can invert the alpha channel (or the tRNS chunk data) after it's read, so that 0 is fully opaque and 255 (in 8-bit or paletted images) or 65535 (in 16-bit images) is fully transparent, with

png_set_invert_alpha(png_ptr);

This has to appear here rather than later with the other transformations because the tRNS chunk data must be modified in the case of paletted images. If your image is not a paletted image, the tRNS data (which in such cases represents a single color to be rendered as transparent) won't be changed.

Finally, you can write your own transformation function if none of the existing ones meets your needs. This is done by setting a callback with

png_set_read_user_transform_fn(png_ptr,
       read_transform_fn);

You must supply the function

void read_transform_fn(png_ptr ptr, row_info_ptr
       row_info, png_bytep data)

See pngtest.c for a working example. Your function will be called after all of the other transformations have been processed.

You are now ready to read all the file information up to the actual image data. You do this with a call to png_read_info().

png_read_info(png_ptr, info_ptr);

Functions are used to get the information from the info_ptr:

png_get_IHDR(png_ptr, info_ptr, &width, &height,
       &bit_depth, &color_type, &interlace_type,
       &compression_type, &filter_type);

    width          - holds the width of the image
                     in pixels (up to 2^31).
    height         - holds the height of the image
                     in pixels (up to 2^31).
    bit_depth      - holds the bit depth of one of the
                     image channels.  (valid values are
                     1, 2, 4, 8, 16 and depend also on
                     the color_type.  See also
                     significant bits (sBIT) below).
    color_type     - describes which color/alpha channels
                         are present.
                     PNG_COLOR_TYPE_GRAY
                        (bit depths 1, 2, 4, 8, 16)
                     PNG_COLOR_TYPE_GRAY_ALPHA
                        (bit depths 8, 16)
                     PNG_COLOR_TYPE_PALETTE
                        (bit depths 1, 2, 4, 8)
                     PNG_COLOR_TYPE_RGB
                        (bit_depths 8, 16)
                     PNG_COLOR_TYPE_RGB_ALPHA
                        (bit_depths 8, 16)

                     PNG_COLOR_MASK_PALETTE
                     PNG_COLOR_MASK_COLOR
                     PNG_COLOR_MASK_ALPHA

    filter_type    - (must be PNG_FILTER_TYPE_BASE
                     for PNG 1.0)
    compression_type - (must be PNG_COMPRESSION_TYPE_BASE
                     for PNG 1.0)
    interlace_type - (PNG_INTERLACE_NONE or
                     PNG_INTERLACE_ADAM7)
    Any or all of interlace_type, compression_type, of
                     filter_type can be
    NULL if you are not interested in their values.

    channels = png_get_channels(png_ptr, info_ptr);
    channels       - number of channels of info for the
                     color type (valid values are 1 (GRAY,
                     PALETTE), 2 (GRAY_ALPHA), 3 (RGB),
                     4 (RGB_ALPHA or RGB + filler byte))
    rowbytes = png_get_rowbytes(png_ptr, info_ptr);
    rowbytes       - number of bytes needed to hold a row

    signature = png_get_signature(png_ptr, info_ptr);
    signature      - holds the signature read from the
                     file (if any).  The data is kept in
                     the same offset it would be if the
                     whole signature were read (i.e. if an
                     application had already read in 4
                     bytes of signature before starting
                     libpng, the remaining 4 bytes would
                     be in signature[4] through signature[7]
                     (see png_set_sig_bytes())).


    width            = png_get_image_width(png_ptr,
                         info_ptr);
    height           = png_get_image_height(png_ptr,
                         info_ptr);
    bit_depth        = png_get_bit_depth(png_ptr,
                         info_ptr);
    color_type       = png_get_color_type(png_ptr,
                         info_ptr);
    filter_type      = png_get_filter_type(png_ptr,
                         info_ptr);
    compression_type = png_get_compression_type(png_ptr,
                         info_ptr);
    interlace_type   = png_get_interlace_type(png_ptr,
                         info_ptr);

These are also important, but their validity depends on whether the chunk has been read. The png_get_valid(png_ptr, info_ptr, PNG_INFO_<chunk>) and png_get_<chunk>(png_ptr, info_ptr, ...) functions return non-zero if the data has been read, or zero if it is missing. The parameters to the png_get_<chunk> are set directly if they are simple data types, or a pointer into the info_ptr is returned for any complex types.

png_get_PLTE(png_ptr, info_ptr, &palette,
                     &num_palette);
    palette        - the palette for the file
                     (array of png_color)
    num_palette    - number of entries in the palette

    png_get_gAMA(png_ptr, info_ptr, &gamma);
    gamma          - the gamma the file is written
                     at (PNG_INFO_gAMA)

    png_get_sRGB(png_ptr, info_ptr, &srgb_intent);
    srgb_intent    - the rendering intent (PNG_INFO_sRGB)
                     The presence of the sRGB chunk
                     means that the pixel data is in the
                     sRGB color space.  This chunk also
                     implies specific values of gAMA and
                     cHRM.

    png_get_sBIT(png_ptr, info_ptr, &sig_bit);
    sig_bit        - the number of significant bits for
                     (PNG_INFO_sBIT) each of the gray,
                     red, green, and blue channels,
                     whichever are appropriate for the
                     given color type (png_color_16)

    png_get_tRNS(png_ptr, info_ptr, &trans, &num_trans,
                     &trans_values);
    trans          - array of transparent entries for
                     palette (PNG_INFO_tRNS)
    trans_values   - transparent pixel for non-paletted
                     images (PNG_INFO_tRNS)
    num_trans      - number of transparent entries
                     (PNG_INFO_tRNS)

    png_get_hIST(png_ptr, info_ptr, &hist);
                     (PNG_INFO_hIST)
    hist           - histogram of palette (array of
                     png_color_16)

    png_get_tIME(png_ptr, info_ptr, &mod_time);
    mod_time       - time image was last modified
                    (PNG_VALID_tIME)

    png_get_bKGD(png_ptr, info_ptr, &background);
    background     - background color (PNG_VALID_bKGD)

    num_text = png_get_text(png_ptr, info_ptr, &text_ptr);
    text_ptr       - array of png_text holding image
                     comments
    text_ptr[i]->key   - keyword for comment.
    text_ptr[i]->text  - text comments for current
                         keyword.
    text_ptr[i]->compression - type of compression used
                     on "text" PNG_TEXT_COMPRESSION_NONE
                     or PNG_TEXT_COMPRESSION_zTXt
    num_text       - number of comments

    png_get_oFFs(png_ptr, info_ptr, &offset_x, &offset_y,
                     &unit_type);
    offset_x       - positive offset from the left edge
                     of the screen
    offset_y       - positive offset from the top edge
                     of the screen
    unit_type      - PNG_OFFSET_PIXEL, PNG_OFFSET_MICROMETER

    png_get_pHYs(png_ptr, info_ptr, &res_x, &res_y,
                     &unit_type);
    res_x          - pixels/unit physical resolution in
                     x direction
    res_y          - pixels/unit physical resolution in
                     x direction
    unit_type      - PNG_RESOLUTION_UNKNOWN,
                     PNG_RESOLUTION_METER

The data from the pHYs chunk can be retrieved in several convenient forms:

res_x = png_get_x_pixels_per_meter(png_ptr,
                  info_ptr)
    res_y = png_get_y_pixels_per_meter(png_ptr,
                  info_ptr)
    res_x_and_y = png_get_pixels_per_meter(png_ptr,
                  info_ptr)
    aspect_ratio = png_get_pixel_aspect_ratio(png_ptr,
                  info_ptr)

   (Each of these returns 0 [signifying "unknown"] if
       the data is not present or if res_x is 0;
       res_x_and_y is 0 if res_x != res_y)

For more information, see the png_info definition in png.h and the PNG specification for chunk contents. Be careful with trusting rowbytes, as some of the transformations could increase the space needed to hold a row (expand, filler, gray_to_rgb, etc.). See png_read_update_info(), below.

A quick word about text_ptr and num_text. PNG stores comments in keyword/text pairs, one pair per chunk, with no limit on the number of text chunks, and a 2^31 byte limit on their size. While there are suggested keywords, there is no requirement to restrict the use to these strings. It is strongly suggested that keywords and text be sensible to humans (that's the point), so don't use abbreviations. Non-printing symbols are not allowed. See the PNG specification for more details. There is also no requirement to have text after the keyword.

Keywords should be limited to 79 Latin-1 characters without leading or trailing spaces, but non-consecutive spaces are allowed within the keyword. It is possible to have the same keyword any number of times. The text_ptr is an array of png_text structures, each holding pointer to a keyword and a pointer to a text string. Only the text string may be null. The keyword/text pairs are put into the array in the order that they are received. However, some or all of the text chunks may be after the image, so, to make sure you have read all the text chunks, don't mess with these until after you read the stuff after the image. This will be mentioned again below in the discussion that goes with png_read_end().

After you've read the header information, you can set up the library to handle any special transformations of the image data. The various ways to transform the data will be described in the order that they should occur. This is important, as some of these change the color type and/or bit depth of the data, and some others only work on certain color types and bit depths. Even though each transformation checks to see if it has data that it can do something with, you should make sure to only enable a transformation if it will be valid for the data. For example, don't swap red and blue on grayscale data.

The colors used for the background and transparency values should be supplied in the same format/depth as the current image data. They are stored in the same format/depth as the image data in a bKGD or tRNS chunk, so this is what libpng expects for this data. The colors are transformed to keep in sync with the image data when an application calls the png_read_update_info() routine (see below).

Data will be decoded into the supplied row buffers packed into bytes unless the library has been told to transform it into another format. For example, 4 bit/pixel paletted or grayscale data will be returned 2 pixels/byte with the leftmost pixel in the high-order bits of the byte, unless png_set_packing() is called. 8-bit RGB data will be stored in RGB RGB RGB format unless png_set_filler() is called to insert filler bytes, either before or after each RGB triplet. 16-bit RGB data will be returned RRGGBB RRGGBB, with the most significant byte of the color value first, unless png_set_strip_16() is called to transform it to regular RGB RGB triplets, or png_set_filler() is called to insert filler bytes, either before or after each RRGGBB triplet. Similarly, 8-bit or 16-bit grayscale data can be modified with png_set_filler() or png_set_strip_16().

The following code transforms grayscale images of less than 8 to 8 bits, changes paletted images to RGB, and adds a full alpha channel if there is transparency information in a tRNS chunk. This is most useful on grayscale images with bit depths of 2 or 4 or if there is a multiple-image viewing application that wishes to treat all images in the same way.

if (color_type == PNG_COLOR_TYPE_PALETTE &&
        bit_depth <= 8) png_set_expand(png_ptr);

    if (color_type == PNG_COLOR_TYPE_GRAY &&
        bit_depth < 8) png_set_expand(png_ptr);

    if (png_get_valid(png_ptr, info_ptr,
        PNG_INFO_tRNS)) png_set_expand(png_ptr);

PNG can have files with 16 bits per channel. If you only can handle 8 bits per channel, this will strip the pixels down to 8 bit.

if (bit_depth == 16)
        png_set_strip_16(png_ptr);

The png_set_background() function tells libpng to composite images with alpha or simple transparency against the supplied background color. If the PNG file contains a bKGD chunk (PNG_INFO_bKGD valid), you may use this color, or supply another color more suitable for the current display (e.g., the background color from a web page). You need to tell libpng whether the color is in the gamma space of the display (PNG_BACKGROUND_GAMMA_SCREEN for colors you supply), the file (PNG_BACKGROUND_GAMMA_FILE for colors from the bKGD chunk), or one that is neither of these gammas (PNG_BACKGROUND_GAMMA_UNIQUE - I don't know why anyone would use this, but it's here).

If, for some reason, you don't need the alpha channel on an image, and you want to remove it rather than combining it with the background (but the image author certainly had in mind that you *would* combine it with the background, so that's what you should probably do):

if (color_type & PNG_COLOR_MASK_ALPHA)
        png_set_strip_alpha(png_ptr);

PNG files pack pixels of bit depths 1, 2, and 4 into bytes as small as they can, resulting in, for example, 8 pixels per byte for 1 bit files. This code expands to 1 pixel per byte without changing the values of the pixels:

if (bit_depth < 8)
        png_set_packing(png_ptr);

PNG files have possible bit depths of 1, 2, 4, 8, and 16. All pixels stored in a PNG image have been "scaled" or "shifted" up to the next higher possible bit depth (e.g. from 5 bits/sample in the range [0,31] to 8 bits/sample in the range [0, 255]). However, it is also possible to convert the PNG pixel data back to the original bit depth of the image. This call reduces the pixels back down to the original bit depth:

png_color_16p sig_bit;

    if (png_get_sBIT(png_ptr, info_ptr, &sig_bit))
        png_set_shift(png_ptr, sig_bit);

PNG files store 3-color pixels in red, green, blue order. This code changes the storage of the pixels to blue, green, red:

if (color_type == PNG_COLOR_TYPE_RGB ||
        color_type == PNG_COLOR_TYPE_RGB_ALPHA)
        png_set_bgr(png_ptr);

PNG files store RGB pixels packed into 3 bytes. This code expands them into 4 bytes for windowing systems that need them in this format:

if (bit_depth == 8 && color_type ==
        PNG_COLOR_TYPE_RGB) png_set_filler(png_ptr,
        filler, PNG_FILLER_BEFORE);

where "filler" is the 8 or 16-bit number to fill with, and the location is either PNG_FILLER_BEFORE or PNG_FILLER_AFTER, depending upon whether you want the filler before the RGB or after. This transformation does not affect images that already have full alpha channels.

If you are reading an image with an alpha channel, and you need the data as ARGB instead of the normal PNG format RGBA:

if (color_type == PNG_COLOR_TYPE_RGB_ALPHA)
        png_set_swap_alpha(png_ptr);

For some uses, you may want a grayscale image to be represented as RGB. This code will do that conversion:

if (color_type == PNG_COLOR_TYPE_GRAY ||
        color_type == PNG_COLOR_TYPE_GRAY_ALPHA)
          png_set_gray_to_rgb(png_ptr);

Conversely, you can convert an RGB or RGBA image to grayscale or grayscale with alpha. This is intended for conversion of images that really are gray (red == green == blue), so the function simply strips out the red and blue channels, leaving the green channel in the gray position.

if (color_type == PNG_COLOR_TYPE_RGB ||
        color_type == PNG_COLOR_TYPE_RGB_ALPHA)
          png_set_rgb_to_gray(png_ptr, error_action,
             float red_weight, float green_weight);

    error_action = 1: silently do the conversion
    error_action = 2: issue a warning if the original
                      image has any pixel where
                      red != green or red != blue
    error_action = 3: issue an error and abort the
                      conversion if the original
                      image has any pixel where
                      red != green or red != blue

    red_weight:       weight of red component
                      (NULL -> default 54/256)
    green_weight:     weight of green component
                      (NULL -> default 183/256)

If you have set error_action = 1 or 2, you can later check whether the image really was gray, after processing the image rows, with the png_get_rgb_to_gray_status(png_ptr) function. It will return a png_byte that is zero if the image was gray or 1 if there were any non-gray pixels. bKGD and sBIT data will be silently converted to grayscale, using the green channel data, regardless of the error_action setting.

With 0.0<=red_weight+green_weight<=1.0, the normalized graylevel is computed:

int rw = red_weight * 256;
    int gw = green_weight * 256;
    int bw = 256 - (rw + gw);
    gray = (rw*red + gw*green + bw*blue)/256;

The default values approximate those recommended in the Charles Poynton's Color FAQ, http://www.inforamp.net/~poynton/ Copyright (c) 1998-01-04 Charles Poynton poynton@inforamp.net

Y = 0.212671 * R + 0.715160 * G + 0.072169 * B

Libpng approximates this with

Y = 0.211 * R + 0.715 * G + 0.074 * B

which can be expressed with integers as

Y = (54 * R + 183 * G + 19 * B)/256

The calculation is done in a linear colorspace, if the image gamma is known.

If you have a grayscale and you are using png_set_expand() to change to a higher bit-depth, you must either supply the background color as a gray value at the original file bit-depth (need_expand = 1) or else supply the background color as an RGB triplet at the final, expanded bit depth (need_expand = 0). Similarly, if you are reading a paletted image, you must either supply the background color as a palette index (need_expand = 1) or as an RGB triplet that may or may not be in the palette (need_expand = 0).

png_color_16 my_background;
    png_color_16p image_background;

    if (png_get_bKGD(png_ptr, info_ptr, &image_background))
        png_set_background(png_ptr, image_background,
          PNG_BACKGROUND_GAMMA_FILE, 1, 1.0);
    else
        png_set_background(png_ptr, &my_background,
          PNG_BACKGROUND_GAMMA_SCREEN, 0, 1.0);

To properly display PNG images on any kind of system, the application needs to know what the display gamma is. Ideally, the user will know this, and the application will allow them to set it. One method of allowing the user to set the display gamma separately for each system is to check for the DISPLAY_GAMMA and VIEWING_GAMMA environment variables or for a SCREEN_GAMMA environment variable, which will hopefully be correctly set.

Note that display_gamma is the gamma of your display, while screen_gamma is the overall gamma correction required to produce pleasing results, which depends on the lighting conditions in the surrounding environment. Screen_gamma is display_gamma/viewing_gamma, where viewing_gamma is the amount of additional gamma correction needed to compensate for a (viewing_gamma=1.25) environment. In a dim or brightly lit room, no compensation other than the display_gamma is needed (viewing_gamma=1.0).

if (/* We have a user-defined screen
       gamma value */) {
      screen_gamma = user_defined_screen_gamma;
   }
   /* One way that applications can share the same
      screen gamma value */
   else if ((gamma_str = getenv("SCREEN_GAMMA"))!= NULL) {
      screen_gamma = atof(gamma_str);
   }
   /* If we don't have another value */
   else {
      screen_gamma = 2.2; /* A good guess for a
           PC monitor in a bright office or a dim room */
      screen_gamma = 2.0; /* A good guess for a
           PC monitor in a dark room */
      screen_gamma = 1.7 or 1.0;  /* A good
           guess for Mac systems */
   }

The png_set_gamma() function handles gamma transformations of the data. Pass both the file gamma and the current screen_gamma. If the file does not have a gamma value, you can pass one anyway if you have an idea what it is (usually 0.45455 is a good guess for GIF images on PCs). Note that file gammas are inverted from screen gammas. See the discussions on gamma in the PNG specification for an excellent description of what gamma is, and why all applications should support it. It is strongly recommended that PNG viewers support gamma correction.

if (png_get_gAMA(png_ptr, info_ptr, &gamma))
      png_set_gamma(png_ptr, screen_gamma, gamma);
   else
      png_set_gamma(png_ptr, screen_gamma, 0.45455);

If you need to reduce an RGB file to a paletted file, or if a paletted file has more entries then will fit on your screen, png_set_dither() will do that. Note that this is a simple match dither that merely finds the closest color available. This should work fairly well with optimized palettes, and fairly badly with linear color cubes. If you pass a palette that is larger then maximum_colors, the file will reduce the number of colors in the palette so it will fit into maximum_colors. If there is a histogram, it will use it to make more intelligent choices when reducing the palette. If there is no histogram, it may not do as good a job.

if (color_type & PNG_COLOR_MASK_COLOR) {
      if (png_get_valid(png_ptr, info_ptr,
         PNG_INFO_PLTE)) {
         png_color_16p histogram;

         png_get_hIST(png_ptr, info_ptr,
            &histogram);
         png_set_dither(png_ptr, palette, num_palette,
            max_screen_colors, histogram, 1);
      } else {
         png_color std_color_cube[MAX_SCREEN_COLORS] =
            { ... colors ... };

         png_set_dither(png_ptr, std_color_cube,
            MAX_SCREEN_COLORS, MAX_SCREEN_COLORS,
            NULL,0);
      }
   }

PNG files describe monochrome as black being zero and white being one. The following code will reverse this (make black be one and white be zero):

if (bit_depth == 1 && color_type == PNG_COLOR_GRAY)
      png_set_invert_mono(png_ptr);

PNG files store 16 bit pixels in network byte order (big-endian, ie. most significant bits first). This code changes the storage to the other way (little-endian, i.e. least significant bits first, the way PCs store them):

if (bit_depth == 16)
        png_set_swap(png_ptr);

If you are using packed-pixel images (1, 2, or 4 bits/pixel), and you need to change the order the pixels are packed into bytes, you can use:

if (bit_depth < 8)
       png_set_packswap(png_ptr);

The last thing to handle is interlacing; this is covered in detail below, but you must call the function here if you want libpng to handle expansion of the interlaced image.

number_of_passes = png_set_interlace_handling(png_ptr);

After setting the transformations, libpng can update your png_info structure to reflect any transformations you've requested with this call. This is most useful to update the info structure's rowbytes field so you can use it to allocate your image memory. This function will also update your palette with the correct screen_gamma and background if these have been given with the calls above.

png_read_update_info(png_ptr, info_ptr);

After you call png_read_update_info(), you can allocate any memory you need to hold the image. The row data is simply raw byte data for all forms of images. As the actual allocation varies among applications, no example will be given. If you are allocating one large chunk, you will need to build an array of pointers to each row, as it will be needed for some of the functions below.

After you've allocated memory, you can read the image data. The simplest way to do this is in one function call. If you are allocating enough memory to hold the whole image, you can just call png_read_image() and libpng will read in all the image data and put it in the memory area supplied. You will need to pass in an array of pointers to each row.

This function automatically handles interlacing, so you don't need to call png_set_interlace_handling() or call this function multiple times, or any of that other stuff necessary with png_read_rows().

png_read_image(png_ptr, row_pointers);

where row_pointers is:

png_bytep row_pointers[height];

You can point to void or char or whatever you use for pixels.

If you don't want to read in the whole image at once, you can use png_read_rows() instead. If there is no interlacing (check interlace_type == PNG_INTERLACE_NONE), this is simple:

png_read_rows(png_ptr, row_pointers, NULL,
       number_of_rows);

where row_pointers is the same as in the png_read_image() call.

If you are doing this just one row at a time, you can do this with row_pointers:

png_bytep row_pointers = row;
    png_read_row(png_ptr, &row_pointers, NULL);

If the file is interlaced (info_ptr->interlace_type != 0), things get somewhat harder. The only current (PNG Specification version 1.0) interlacing type for PNG is (interlace_type == PNG_INTERLACE_ADAM7) is a somewhat complicated 2D interlace scheme, known as Adam7, that breaks down an image into seven smaller images of varying size, based on an 8x8 grid.

libpng can fill out those images or it can give them to you "as is". If you want them filled out, there are two ways to do that. The one mentioned in the PNG specification is to expand each pixel to cover those pixels that have not been read yet (the "rectangle" method). This results in a blocky image for the first pass, which gradually smooths out as more pixels are read. The other method is the "sparkle" method, where pixels are drawn only in their final locations, with the rest of the image remaining whatever colors they were initialized to before the start of the read. The first method usually looks better, but tends to be slower, as there are more pixels to put in the rows.

If you don't want libpng to handle the interlacing details, just call png_read_rows() seven times to read in all seven images. Each of the images is a valid image by itself, or they can all be combined on an 8x8 grid to form a single image (although if you intend to combine them you would be far better off using the libpng interlace handling).

The first pass will return an image 1/8 as wide as the entire image (every 8th column starting in column 0) and 1/8 as high as the original (every 8th row starting in row 0), the second will be 1/8 as wide (starting in column 4) and 1/8 as high (also starting in row 0). The third pass will be 1/4 as wide (every 4th pixel starting in column 0) and 1/8 as high (every 8th row starting in row 4), and the fourth pass will be 1/4 as wide and 1/4 as high (every 4th column starting in column 2, and every 4th row starting in row 0). The fifth pass will return an image 1/2 as wide, and 1/4 as high (starting at column 0 and row 2), while the sixth pass will be 1/2 as wide and 1/2 as high as the original (starting in column 1 and row 0). The seventh and final pass will be as wide as the original, and 1/2 as high, containing all of the odd numbered scanlines. Phew!

If you want libpng to expand the images, call this before calling png_start_read_image() or png_read_update_info():

if (interlace_type == PNG_INTERLACE_ADAM7)
        number_of_passes
           = png_set_interlace_handling(png_ptr);

This will return the number of passes needed. Currently, this is seven, but may change if another interlace type is added. This function can be called even if the file is not interlaced, where it will return one pass.

If you are not going to display the image after each pass, but are going to wait until the entire image is read in, use the sparkle effect. This effect is faster and the end result of either method is exactly the same. If you are planning on displaying the image after each pass, the "rectangle" effect is generally considered the better looking one.

If you only want the "sparkle" effect, just call png_read_rows() as normal, with the third parameter NULL. Make sure you make pass over the image number_of_passes times, and you don't change the data in the rows between calls. You can change the locations of the data, just not the data. Each pass only writes the pixels appropriate for that pass, and assumes the data from previous passes is still valid.

png_read_rows(png_ptr, row_pointers, NULL,
       number_of_rows);

If you only want the first effect (the rectangles), do the same as before except pass the row buffer in the third parameter, and leave the second parameter NULL.

png_read_rows(png_ptr, NULL, row_pointers,
       number_of_rows);

After you are finished reading the image, you can finish reading the file. If you are interested in comments or time, which may be stored either before or after the image data, you should pass the separate png_info struct if you want to keep the comments from before and after the image separate. If you are not interested, you can pass NULL.

png_read_end(png_ptr, end_info);

When you are done, you can free all memory allocated by libpng like this:

png_destroy_read_struct(&png_ptr, &info_ptr,
       &end_info);

For a more compact example of reading a PNG image, see the file example.c.

Reading PNG files progressively:

The progressive reader is slightly different then the non-progressive reader. Instead of calling png_read_info(), png_read_rows(), and png_read_end(), you make one call to png_process_data(), which calls callbacks when it has the info, a row, or the end of the image. You set up these callbacks with png_set_progressive_read_fn(). You don't have to worry about the input/output functions of libpng, as you are giving the library the data directly in png_process_data(). I will assume that you have read the section on reading PNG files above, so I will only highlight the differences (although I will show all of the code).

png_structp png_ptr;
png_infop info_ptr;

 /*  An example code fragment of how you would
     initialize the progressive reader in your
     application. */
 int
 initialize_png_reader() {
    png_ptr = png_create_read_struct
        (PNG_LIBPNG_VER_STRING, (png_voidp)user_error_ptr,
         user_error_fn, user_warning_fn);
    if (!png_ptr)
        return -1;
    info_ptr = png_create_info_struct(png_ptr);
    if (!info_ptr) {
        png_destroy_read_struct(&png_ptr, (png_infopp)NULL,
           (png_infopp)NULL);
        return -1;
    }

    if (setjmp(png_ptr->jmpbuf)) {
        png_destroy_read_struct(&png_ptr, &info_ptr,
           (png_infopp)NULL);
        return -1;
    }

    /* This one's new.  You can provide functions
       to be called when the header info is valid,
       when each row is completed, and when the image
       is finished.  If you aren't using all functions,
       you can specify NULL parameters.  Even when all
       three functions are NULL, you need to call
       png_set_progressive_read_fn().  You can use
       any struct as the user_ptr (cast to a void pointer
       for the function call), and retrieve the pointer
       from inside the callbacks using the function

          png_get_progressive_ptr(png_ptr);

       which will return a void pointer, which you have
       to cast appropriately.
     */
    png_set_progressive_read_fn(png_ptr, (void *)user_ptr,
        info_callback, row_callback, end_callback);

    return 0;
 }

 /* A code fragment that you call as you receive blocks
   of data */
 int
 process_data(png_bytep buffer, png_uint_32 length) {
    if (setjmp(png_ptr->jmpbuf)) {
        png_destroy_read_struct(&png_ptr, &info_ptr,
           (png_infopp)NULL);
        return -1;
    }

    /* This one's new also.  Simply give it a chunk
       of data from the file stream (in order, of
       course).  On machines with segmented memory
       models machines, don't give it any more than
       64K.  The library seems to run fine with sizes
       of 4K. Although you can give it much less if
       necessary (I assume you can give it chunks of
       1 byte, I haven't tried less then 256 bytes
       yet).  When this function returns, you may
       want to display any rows that were generated
       in the row callback if you don't already do
       so there.
     */
    png_process_data(png_ptr, info_ptr, buffer, length);
    return 0;
 }

 /* This function is called (as set by
    png_set_progressive_read_fn() above) when enough data
    has been supplied so all of the header has been
    read.
 */
 void
 info_callback(png_structp png_ptr, png_infop info) {
    /* Do any setup here, including setting any of
       the transformations mentioned in the Reading
       PNG files section.  For now, you _must_ call
       either png_start_read_image() or
       png_read_update_info() after all the
       transformations are set (even if you don't set
       any).  You may start getting rows before
       png_process_data() returns, so this is your
       last chance to prepare for that.
     */
 }

 /* This function is called when each row of image
    data is complete */
 void
 row_callback(png_structp png_ptr, png_bytep new_row,
    png_uint_32 row_num, int pass) {
    /* If the image is interlaced, and you turned
       on the interlace handler, this function will
       be called for every row in every pass.  Some
       of these rows will not be changed from the
       previous pass.  When the row is not changed,
       the new_row variable will be NULL.  The rows
       and passes are called in order, so you don't
       really need the row_num and pass, but I'm
       supplying them because it may make your life
       easier.

       For the non-NULL rows of interlaced images,
       you must call png_progressive_combine_row()
       passing in the row and the old row.  You can
       call this function for NULL rows (it will just
       return) and for non-interlaced images (it just
       does the memcpy for you) if it will make the
       code easier.  Thus, you can just do this for
       all cases:
     */

        png_progressive_combine_row(png_ptr, old_row,
          new_row);

    /* where old_row is what was displayed for
       previously for the row.  Note that the first
       pass (pass == 0, really) will completely cover
       the old row, so the rows do not have to be
       initialized.  After the first pass (and only
       for interlaced images), you will have to pass
       the current row, and the function will combine
       the old row and the new row.
    */
 }

 void
 end_callback(png_structp png_ptr, png_infop info) {
    /* This function is called after the whole image
       has been read, including any chunks after the
       image (up to and including the IEND).  You
       will usually have the same info chunk as you
       had in the header, although some data may have
       been added to the comments and time fields.

       Most people won't do much here, perhaps setting
       a flag that marks the image as finished.
     */
 }

IV. WRITING

Much of this is very similar to reading. However, everything of importance is repeated here, so you won't have to constantly look back up in the reading section to understand writing.

You will want to do the I/O initialization before you get into libpng, so if it doesn't work, you don't have anything to undo. If you are not using the standard I/O functions, you will need to replace them with custom writing functions. See the discussion under Customizing libpng.

FILE *fp = fopen(file_name, "wb");
    if (!fp) return;

Next, png_struct and png_info need to be allocated and initialized. As these can be both relatively large, you may not want to store these on the stack, unless you have stack space to spare. Of course, you will want to check if they return NULL. If you are also reading, you won't want to name your read structure and your write structure both "png_ptr"; you can call them anything you like, such as "read_ptr" and "write_ptr". Look at pngtest.c, for example.

png_structp png_ptr = png_create_write_struct
       (PNG_LIBPNG_VER_STRING, (png_voidp)user_error_ptr,
        user_error_fn, user_warning_fn);
    if (!png_ptr)
       return;

    png_infop info_ptr = png_create_info_struct(png_ptr);
    if (!info_ptr) {
       png_destroy_write_struct(&png_ptr,
         (png_infopp)NULL);
       return;
    }

If you want to use your own memory allocation routines, define PNG_USER_MEM_SUPPORTED and use png_create_write_struct_2() instead of png_create_read_struct():

png_structp png_ptr = png_create_write_struct_2
       (PNG_LIBPNG_VER_STRING, (png_voidp)user_error_ptr,
        user_error_fn, user_warning_fn, (png_voidp)
        user_mem_ptr, user_malloc_fn, user_free_fn);

After you have these structures, you will need to set up the error handling. When libpng encounters an error, it expects to longjmp() back to your routine. Therefore, you will need to call setjmp() and pass the png_ptr->jmpbuf. If you write the file from different routines, you will need to update the jmpbuf field every time you enter a new routine that will call a png_ function. See your documentation of setjmp/longjmp for your compiler for more information on setjmp/longjmp. See the discussion on libpng error handling in the Customizing Libpng section below for more information on the libpng error handling.

if (setjmp(png_ptr->jmpbuf)) {
        png_destroy_write_struct(&png_ptr, &info_ptr);
        fclose(fp);
        return;
    }
    ...
    return;

Now you need to set up the output code. The default for libpng is to use the C function fwrite(). If you use this, you will need to pass a valid FILE * in the function png_init_io(). Be sure that the file is opened in binary mode. Again, if you wish to handle writing data in another way, see the discussion on libpng I/O handling in the Customizing Libpng section below.

png_init_io(png_ptr, fp);

At this point, you can set up a callback function that will be called after each row has been written, which you can use to control a progress meter or the like. It's demonstrated in pngtest.c. You must supply a function

void
    write_row_callback(png_ptr, png_uint_32 row, int pass) {
      /* put your code here */
    }

(You can give it another name that you like instead of "write_row_callback")

To inform libpng about your function, use

png_set_write_status_fn(png_ptr, write_row_callback);

You now have the option of modifying how the compression library will run. The following functions are mainly for testing, but may be useful in some cases, like if you need to write PNG files extremely fast and are willing to give up some compression, or if you want to get the maximum possible compression at the expense of slower writing. If you have no special needs in this area, let the library do what it wants by not calling this function at all, as it has been tuned to deliver a good speed/compression ratio. The second parameter to png_set_filter() is the filter method, for which the only valid value is '0' (as of the October 1996 PNG specification, version 1.0). The third parameter is a flag that indicates which filter type(s) are to be tested for each scanline. See the Compression Library for details on the specific filter types.

/* turn on or off filtering, and/or choose
       specific filters */
    png_set_filter(png_ptr, 0,
       PNG_FILTER_NONE | PNG_FILTER_SUB |
       PNG_FILTER_PAETH);

The png_set_compression_???() functions interface to the zlib compression library, and should mostly be ignored unless you really know what you are doing. The only generally useful call is png_set_compression_level() which changes how much time zlib spends on trying to compress the image data. See the Compression Library for details on the compression levels.

/* set the zlib compression level */
    png_set_compression_level(png_ptr,
        Z_BEST_COMPRESSION);

    /* set other zlib parameters */
    png_set_compression_mem_level(png_ptr, 8);
    png_set_compression_strategy(png_ptr,
        Z_DEFAULT_STRATEGY);
    png_set_compression_window_bits(png_ptr, 15);
    png_set_compression_method(png_ptr, 8);

You now need to fill in the png_info structure with all the data you wish to write before the actual image. Note that the only thing you are allowed to write after the image is the text chunks and the time chunk (as of PNG Specification 1.0, anyway). See png_write_end() and the latest PNG specification for more information on that. If you wish to write them before the image, fill them in now, and flag that data as being valid. If you want to wait until after the data, don't fill them until png_write_end(). For all the fields in png_info and their data types, see png.h. For explanations of what the fields contain, see the PNG specification.

Some of the more important parts of the png_info are:

png_set_IHDR(png_ptr, info_ptr, width, height,
       bit_depth, color_type, interlace_type,
       compression_type, filter_type)
    width          - holds the width of the image
                     in pixels (up to 2^31).
    height         - holds the height of the image
                     in pixels (up to 2^31).
    bit_depth      - holds the bit depth of one of the
                     image channels.
                     (valid values are 1, 2, 4, 8, 16
                     and depend also on the
                     color_type.  See also significant
                     bits (sBIT) below).
    color_type     - describes which color/alpha
                     channels are present.
                     PNG_COLOR_TYPE_GRAY
                        (bit depths 1, 2, 4, 8, 16)
                     PNG_COLOR_TYPE_GRAY_ALPHA
                        (bit depths 8, 16)
                     PNG_COLOR_TYPE_PALETTE
                        (bit depths 1, 2, 4, 8)
                     PNG_COLOR_TYPE_RGB
                        (bit_depths 8, 16)
                     PNG_COLOR_TYPE_RGB_ALPHA
                        (bit_depths 8, 16)

                     PNG_COLOR_MASK_PALETTE
                     PNG_COLOR_MASK_COLOR
                     PNG_COLOR_MASK_ALPHA

    interlace_type - PNG_INTERLACE_NONE or
                     PNG_INTERLACE_ADAM7
    compression_type - (must be
                     PNG_COMPRESSION_TYPE_DEFAULT)
    filter_type    - (must be PNG_FILTER_TYPE_DEFAULT)

    png_set_PLTE(png_ptr, info_ptr, palette,
       num_palette);
    palette        - the palette for the file
                     (array of png_color)
    num_palette    - number of entries in the palette

    png_set_gAMA(png_ptr, info_ptr, gamma);
    gamma          - the gamma the image was created
                     at (PNG_INFO_gAMA)

    png_set_sRGB(png_ptr, info_ptr, srgb_intent);
    srgb_intent    - the rendering intent
                     (PNG_INFO_sRGB) The presence of
                     the sRGB chunk means that the pixel
                     data is in the sRGB color space.
                     This chunk also implies specific
                     values of gAMA and cHRM.  Rendering
                     intent is the CSS-1 property that
                     has been defined by the International
                     Color Consortium
                     (http://www.color.org).
                     It can be one of
                     PNG_SRGB_INTENT_SATURATION,
                     PNG_SRGB_INTENT_PERCEPTUAL,
                     PNG_SRGB_INTENT_ABSOLUTE, or
                     PNG_SRGB_INTENT_RELATIVE.


    png_set_sRGB_gAMA_and_cHRM(png_ptr, info_ptr,
       srgb_intent);
    srgb_intent    - the rendering intent
                     (PNG_INFO_sRGB) The presence of the
                     sRGB chunk means that the pixel
                     data is in the sRGB color space.
                     This function also causes gAMA and
                     cHRM chunks with the specific values
                     that are consistent with sRGB to be
                     written.

    png_set_sBIT(png_ptr, info_ptr, sig_bit);
    sig_bit        - the number of significant bits for
                     (PNG_INFO_sBIT) each of the gray, red,
                     green, and blue channels, whichever are
                     appropriate for the given color type
                     (png_color_16)

    png_set_tRNS(png_ptr, info_ptr, trans, num_trans,
       trans_values);
    trans          - array of transparent entries for
                     palette (PNG_INFO_tRNS)
    trans_values   - transparent pixel for non-paletted
                     images (PNG_INFO_tRNS)
    num_trans      - number of transparent entries
                     (PNG_INFO_tRNS)

    png_set_hIST(png_ptr, info_ptr, hist);
                    (PNG_INFO_hIST)
    hist           - histogram of palette (array of
                     png_color_16)

    png_set_tIME(png_ptr, info_ptr, mod_time);
    mod_time       - time image was last modified
                     (PNG_VALID_tIME)

    png_set_bKGD(png_ptr, info_ptr, background);
    background     - background color (PNG_VALID_bKGD)

    png_set_text(png_ptr, info_ptr, text_ptr, num_text);
    text_ptr       - array of png_text holding image
                     comments
    text_ptr[i]->key   - keyword for comment.
    text_ptr[i]->text  - text comments for current
                         keyword.
    text_ptr[i]->compression - type of compression used
         on "text" PNG_TEXT_COMPRESSION_NONE or
         PNG_TEXT_COMPRESSION_zTXt
    num_text    - number of comments in text_ptr

    png_set_oFFs(png_ptr, info_ptr, offset_x, offset_y,
        unit_type);
    offset_x  - positive offset from the left
                     edge of the screen
    offset_y  - positive offset from the top
                     edge of the screen
    unit_type - PNG_OFFSET_PIXEL, PNG_OFFSET_MICROMETER

    png_set_pHYs(png_ptr, info_ptr, res_x, res_y,
        unit_type);
    res_x       - pixels/unit physical resolution
                  in x direction
    res_y       - pixels/unit physical resolution
                  in y direction
    unit_type   - PNG_RESOLUTION_UNKNOWN,
                  PNG_RESOLUTION_METER

In PNG files, the alpha channel in an image is the level of opacity. If your data is supplied as a level of transparency, you can invert the alpha channel before you write it, so that 0 is fully transparent and 255 (in 8-bit or paletted images) or 65535 (in 16-bit images) is fully opaque, with

png_set_invert_alpha(png_ptr);

This must appear here instead of later with the other transformations because in the case of paletted images the tRNS chunk data has to be inverted before the tRNS chunk is written. If your image is not a paletted image, the tRNS data (which in such cases represents a single color to be rendered as transparent) won't be changed.

A quick word about text and num_text. text is an array of png_text structures. num_text is the number of valid structures in the array. If you want, you can use max_text to hold the size of the array, but libpng ignores it for writing (it does use it for reading). Each png_text structure holds a keyword-text value, and a compression type. The compression types have the same valid numbers as the compression types of the image data. Currently, the only valid number is zero. However, you can store text either compressed or uncompressed, unlike images, which always have to be compressed. So if you don't want the text compressed, set the compression type to PNG_TEXT_COMPRESSION_NONE. Until text gets around 1000 bytes, it is not worth compressing it. After the text has been written out to the file, the compression type is set to PNG_TEXT_COMPRESSION_NONE_WR or PNG_TEXT_COMPRESSION_zTXt_WR, so that it isn't written out again at the end (in case you are calling png_write_end() with the same struct.

The keywords that are given in the PNG Specification are:

Title            Short (one line) title or
                     caption for image
    Author           Name of image's creator
    Description      Description of image (possibly long)
    Copyright        Copyright notice
    Creation Time    Time of original image creation
                     (usually RFC 1123 format, see below)
    Software         Software used to create the image
    Disclaimer       Legal disclaimer
    Warning          Warning of nature of content
    Source           Device used to create the image
    Comment          Miscellaneous comment; conversion
                     from other image format

The keyword-text pairs work like this. Keywords should be short simple descriptions of what the comment is about. Some typical keywords are found in the PNG specification, as is some recommendations on keywords. You can repeat keywords in a file. You can even write some text before the image and some after. For example, you may want to put a description of the image before the image, but leave the disclaimer until after, so viewers working over modem connections don't have to wait for the disclaimer to go over the modem before they start seeing the image. Finally, keywords should be full words, not abbreviations. Keywords and text are in the ISO 8859-1 (Latin-1) character set (a superset of regular ASCII) and can not contain NUL characters, and should not contain control or other unprintable characters. To make the comments widely readable, stick with basic ASCII, and avoid machine specific character set extensions like the IBM-PC character set. The keyword must be present, but you can leave off the text string on non-compressed pairs. Compressed pairs must have a text string, as only the text string is compressed anyway, so the compression would be meaningless.

PNG supports modification time via the png_time structure. Two conversion routines are provided, png_convert_from_time_t() for time_t and png_convert_from_struct_tm() for struct tm. The time_t routine uses gmtime(). You don't have to use either of these, but if you wish to fill in the png_time structure directly, you should provide the time in universal time (GMT) if possible instead of your local time. Note that the year number is the full year (e.g. 1998, rather than 98 - PNG is year 2000 compliant!), and that months start with 1.

If you want to store the time of the original image creation, you should use a plain tEXt chunk with the "Creation Time" keyword. This is necessary because the "creation time" of a PNG image is somewhat vague, depending on whether you mean the PNG file, the time the image was created in a non-PNG format, a still photo from which the image was scanned, or possibly the subject matter itself. In order to facilitate machine-readable dates, it is recommended that the "Creation Time" tEXt chunk use RFC 1123 format dates (e.g. "22 May 1997 18:07:10 GMT"), although this isn't a requirement. Unlike the tIME chunk, the "Creation Time" tEXt chunk is not expected to be automatically changed by the software. To facilitate the use of RFC 1123 dates, a function png_convert_to_rfc1123(png_timep) is provided to convert from PNG time to an RFC 1123 format string.

You are now ready to write all the file information up to the actual image data. You do this with a call to png_write_info().

png_write_info(png_ptr, info_ptr);

After you've written the file information, you can set up the library to handle any special transformations of the image data. The various ways to transform the data will be described in the order that they should occur. This is important, as some of these change the color type and/or bit depth of the data, and some others only work on certain color types and bit depths. Even though each transformation checks to see if it has data that it can do something with, you should make sure to only enable a transformation if it will be valid for the data. For example, don't swap red and blue on grayscale data.

PNG files store RGB pixels packed into 3 or 6 bytes. This code tells the library to expand the input data to 4 or 8 bytes per pixel (or expand 1 or 2-byte grayscale data to 2 or 4 bytes per pixel).

png_set_filler(png_ptr, 0, PNG_FILLER_BEFORE);

where the 0 is the value that will be put in the 4th byte, and the location is either PNG_FILLER_BEFORE or PNG_FILLER_AFTER, depending upon whether the filler byte is stored XRGB or RGBX.

PNG files pack pixels of bit depths 1, 2, and 4 into bytes as small as they can, resulting in, for example, 8 pixels per byte for 1 bit files. If the data is supplied at 1 pixel per byte, use this code, which will correctly pack the pixels into a single byte:

png_set_packing(png_ptr);

PNG files reduce possible bit depths to 1, 2, 4, 8, and 16. If your data is of another bit depth, you can write an sBIT chunk into the file so that decoders can get the original data if desired.

/* Set the true bit depth of the image data */
    if (color_type & PNG_COLOR_MASK_COLOR) {
        sig_bit.red = true_bit_depth;
        sig_bit.green = true_bit_depth;
        sig_bit.blue = true_bit_depth;
    } else {
        sig_bit.gray = true_bit_depth;
    }
    if (color_type & PNG_COLOR_MASK_ALPHA) {
        sig_bit.alpha = true_bit_depth;
    }

    png_set_sBIT(png_ptr, info_ptr, &sig_bit);

If the data is stored in the row buffer in a bit depth other than one supported by PNG (e.g. 3 bit data in the range 0-7 for a 4-bit PNG), this will scale the values to appear to be the correct bit depth as is required by PNG.

png_set_shift(png_ptr, &sig_bit);

PNG files store 16 bit pixels in network byte order (big-endian, ie. most significant bits first). This code would be used if they are supplied the other way (little-endian, i.e. least significant bits first, the way PCs store them):

if (bit_depth > 8)
       png_set_swap(png_ptr);

If you are using packed-pixel images (1, 2, or 4 bits/pixel), and you need to change the order the pixels are packed into bytes, you can use:

if (bit_depth < 8)
       png_set_packswap(png_ptr);

PNG files store 3 color pixels in red, green, blue order. This code would be used if they are supplied as blue, green, red:

png_set_bgr(png_ptr);

PNG files describe monochrome as black being zero and white being one. This code would be used if the pixels are supplied with this reversed (black being one and white being zero):

png_set_invert_mono(png_ptr);

Finally, you can write your own transformation function if none of the existing ones meets your needs. This is done by setting a callback with

png_set_write_user_transform_fn(png_ptr,
       write_transform_fn);

You must supply the function

void write_transform_fn(png_ptr ptr, row_info_ptr
       row_info, png_bytep data)

See pngtest.c for a working example. Your function will be called before any of the other transformations have been processed.

It is possible to have libpng flush any pending output, either manually, or automatically after a certain number of lines have been written. To flush the output stream a single time call:

png_write_flush(png_ptr);

and to have libpng flush the output stream periodically after a certain number of scanlines have been written, call:

png_set_flush(png_ptr, nrows);

Note that the distance between rows is from the last time png_write_flush() was called, or the first row of the image if it has never been called. So if you write 50 lines, and then png_set_flush 25, it will flush the output on the next scanline, and every 25 lines thereafter, unless png_write_flush() is called before 25 more lines have been written. If nrows is too small (less than about 10 lines for a 640 pixel wide RGB image) the image compression may decrease noticeably (although this may be acceptable for real-time applications). Infrequent flushing will only degrade the compression performance by a few percent over images that do not use flushing.

That's it for the transformations. Now you can write the image data. The simplest way to do this is in one function call. If have the whole image in memory, you can just call png_write_image() and libpng will write the image. You will need to pass in an array of pointers to each row. This function automatically handles interlacing, so you don't need to call png_set_interlace_handling() or call this function multiple times, or any of that other stuff necessary with png_write_rows().

png_write_image(png_ptr, row_pointers);

where row_pointers is:

png_byte *row_pointers[height];

You can point to void or char or whatever you use for pixels.

If you don't want to write the whole image at once, you can use png_write_rows() instead. If the file is not interlaced, this is simple:

png_write_rows(png_ptr, row_pointers,
       number_of_rows);

row_pointers is the same as in the png_write_image() call.

If you are just writing one row at a time, you can do this with row_pointers:

png_bytep row_pointer = row;

    png_write_row(png_ptr, &row_pointer);

When the file is interlaced, things can get a good deal more complicated. The only currently (as of February 1998 -- PNG Specification version 1.0, dated October 1996) defined interlacing scheme for PNG files is the "Adam7" interlace scheme, that breaks down an image into seven smaller images of varying size. libpng will build these images for you, or you can do them yourself. If you want to build them yourself, see the PNG specification for details of which pixels to write when.

If you don't want libpng to handle the interlacing details, just use png_set_interlace_handling() and call png_write_rows() the correct number of times to write all seven sub-images.

If you want libpng to build the sub-images, call this before you start writing any rows:

number_of_passes =
       png_set_interlace_handling(png_ptr);

This will return the number of passes needed. Currently, this is seven, but may change if another interlace type is added.

Then write the complete image number_of_passes times.

png_write_rows(png_ptr, row_pointers,
       number_of_rows);

As some of these rows are not used, and thus return immediately, you may want to read about interlacing in the PNG specification, and only update the rows that are actually used.

After you are finished writing the image, you should finish writing the file. If you are interested in writing comments or time, you should pass an appropriately filled png_info pointer. If you are not interested, you can pass NULL.

png_write_end(png_ptr, info_ptr);

When you are done, you can free all memory used by libpng like this:

png_destroy_write_struct(&png_ptr, &info_ptr);

You must free any data you allocated for info_ptr, such as comments, palette, or histogram, before the call to png_destroy_write_struct();

For a more compact example of writing a PNG image, see the file example.c.

V. MODIFYING/CUSTOMIZING LIBPNG:

There are two issues here. The first is changing how libpng does standard things like memory allocation, input/output, and error handling. The second deals with more complicated things like adding new chunks, adding new transformations, and generally changing how libpng works.

All of the memory allocation, input/output, and error handling in libpng goes through callbacks that are user settable. The default routines are in pngmem.c, pngrio.c, pngwio.c, and pngerror.c respectively. To change these functions, call the appropriate png_set_???_fn() function.

Memory allocation is done through the functions png_large_malloc(), png_malloc(), png_realloc(), png_large_free(), and png_free(). These currently just call the standard C functions. The large functions must handle exactly 64K, but they don't have to handle more than that. If your pointers can't access more then 64K at a time, you will want to set MAXSEG_64K in zlib.h. Since it is unlikely that the method of handling memory allocation on a platform will change between applications, these functions must be modified in the library at compile time.

Input/Output in libpng is done through png_read() and png_write(), which currently just call fread() and fwrite(). The FILE * is stored in png_struct and is initialized via png_init_io(). If you wish to change the method of I/O, the library supplies callbacks that you can set through the function png_set_read_fn() and png_set_write_fn() at run time, instead of calling the png_init_io() function. These functions also provide a void pointer that can be retrieved via the function png_get_io_ptr(). For example:

png_set_read_fn(png_structp read_ptr,
        voidp read_io_ptr, png_rw_ptr read_data_fn)

    png_set_write_fn(png_structp write_ptr,
        voidp write_io_ptr, png_rw_ptr write_data_fn,
        png_flush_ptr output_flush_fn);

    voidp read_io_ptr = png_get_io_ptr(read_ptr);
    voidp write_io_ptr = png_get_io_ptr(write_ptr);

The replacement I/O functions should have prototypes as follows:

void user_read_data(png_structp png_ptr,
        png_bytep data, png_uint_32 length);
    void user_write_data(png_structp png_ptr,
        png_bytep data, png_uint_32 length);
    void user_flush_data(png_structp png_ptr);

Supplying NULL for the read, write, or flush functions sets them back to using the default C stream functions. It is an error to read from a write stream, and vice versa.

Error handling in libpng is done through png_error() and png_warning(). Errors handled through png_error() are fatal, meaning that png_error() should never return to its caller. Currently, this is handled via setjmp() and longjmp(), but you could change this to do things like exit() if you should wish. On non-fatal errors, png_warning() is called to print a warning message, and then control returns to the calling code. By default png_error() and png_warning() print a message on stderr via fprintf() unless the library is compiled with PNG_NO_STDIO defined. If you wish to change the behavior of the error functions, you will need to set up your own message callbacks. These functions are normally supplied at the time that the png_struct is created. It is also possible to change these functions after png_create_???_struct() has been called by calling:

png_set_error_fn(png_structp png_ptr,
        png_voidp error_ptr, png_error_ptr error_fn,
        png_error_ptr warning_fn);

    png_voidp error_ptr = png_get_error_ptr(png_ptr);

If NULL is supplied for either error_fn or warning_fn, then the libpng default function will be used, calling fprintf() and/or longjmp() if a problem is encountered. The replacement error functions should have parameters as follows:

void user_error_fn(png_structp png_ptr,
        png_const_charp error_msg);
    void user_warning_fn(png_structp png_ptr,
        png_const_charp warning_msg);

The motivation behind using setjmp() and longjmp() is the C++ throw and catch exception handling methods. This makes the code much easier to write, as there is no need to check every return code of every function call. However, there are some uncertainties about the status of local variables after a longjmp, so the user may want to be careful about doing anything after setjmp returns non-zero besides returning itself. Consult your compiler documentation for more details.

If you need to read or write custom chunks, you will need to get deeper into the libpng code, as a mechanism has not yet been supplied for user callbacks with custom chunks. First, read the PNG specification, and have a first level of understanding of how it works. Pay particular attention to the sections that describe chunk names, and look at how other chunks were designed, so you can do things similarly. Second, check out the sections of libpng that read and write chunks. Try to find a chunk that is similar to yours and copy off of it. More details can be found in the comments inside the code. A way of handling unknown chunks in a generic method, potentially via callback functions, would be best.

If you wish to write your own transformation for the data, look through the part of the code that does the transformations, and check out some of the simpler ones to get an idea of how they work. Try to find a similar transformation to the one you want to add and copy off of it. More details can be found in the comments inside the code itself.

Configuring for 16 bit platforms:

You may need to change the png_large_malloc() and png_large_free() routines in pngmem.c, as these are required to allocate 64K, although there is already support for many of the common DOS compilers. Also, you will want to look into zconf.h to tell zlib (and thus libpng) that it cannot allocate more then 64K at a time. Even if you can, the memory won't be accessible. So limit zlib and libpng to 64K by defining MAXSEG_64K.

Configuring for DOS:

For DOS users who only have access to the lower 640K, you will have to limit zlib's memory usage via a png_set_compression_mem_level() call. See zlib.h or zconf.h in the zlib library for more information.

Configuring for Medium Model:

Libpng's support for medium model has been tested on most of the popular compilers. Make sure MAXSEG_64K gets defined, USE_FAR_KEYWORD gets defined, and FAR gets defined to far in pngconf.h, and you should be all set. Everything in the library (except for zlib's structure) is expecting far data. You must use the typedefs with the p or pp on the end for pointers (or at least look at them and be careful). Make note that the row's of data are defined as png_bytepp, which is an unsigned char far * far *.

Configuring for gui/windowing platforms:

You will need to write new error and warning functions that use the GUI interface, as described previously, and set them to be the error and warning functions at the time that png_create_???_struct() is called, in order to have them available during the structure initialization. They can be changed later via png_set_error_fn(). On some compilers, you may also have to change the memory allocators (png_malloc, etc.).

Configuring for compiler xxx:

All includes for libpng are in pngconf.h. If you need to add/change/delete an include, this is the place to do it. The includes that are not needed outside libpng are protected by the PNG_INTERNAL definition, which is only defined for those routines inside libpng itself. The files in libpng proper only include png.h, which includes pngconf.h.

Configuring zlib:

There are special functions to configure the compression. Perhaps the most useful one changes the compression level, which currently uses input compression values in the range 0 - 9. The library normally uses the default compression level (Z_DEFAULT_COMPRESSION = 6). Tests have shown that for a large majority of images, compression values in the range 3-6 compress nearly as well as higher levels, and do so much faster. For online applications it may be desirable to have maximum speed (Z_BEST_SPEED = 1). With versions of zlib after v0.99, you can also specify no compression (Z_NO_COMPRESSION = 0), but this would create files larger than just storing the raw bitmap. You can specify the compression level by calling:

png_set_compression_level(png_ptr, level);

Another useful one is to reduce the memory level used by the library. The memory level defaults to 8, but it can be lowered if you are short on memory (running DOS, for example, where you only have 640K).

png_set_compression_mem_level(png_ptr, level);

The other functions are for configuring zlib. They are not recommended for normal use and may result in writing an invalid PNG file. See zlib.h for more information on what these mean.

png_set_compression_strategy(png_ptr,
        strategy);
    png_set_compression_window_bits(png_ptr,
        window_bits);
    png_set_compression_method(png_ptr, method);

Controlling row filtering:

If you want to control whether libpng uses filtering or not, which filters are used, and how it goes about picking row filters, you can call one of these functions. The selection and configuration of row filters can have a significant impact on the size and encoding speed and a somewhat lesser impact on the decoding speed of an image. Filtering is enabled by default for RGB and grayscale images (with and without alpha), but not for paletted images nor for any images with bit depths less than 8 bits/pixel.

The 'method' parameter sets the main filtering method, which is currently only '0' in the PNG 1.0 specification. The 'filters' parameter sets which filter(s), if any, should be used for each scanline. Possible values are PNG_ALL_FILTERS and PNG_NO_FILTERS to turn filtering on and off, respectively.

Individual filter types are PNG_FILTER_NONE, PNG_FILTER_SUB, PNG_FILTER_UP, PNG_FILTER_AVG, PNG_FILTER_PAETH, which can be bitwise ORed together '|' to specify one or more filters to use. These filters are described in more detail in the PNG specification. If you intend to change the filter type during the course of writing the image, you should start with flags set for all of the filters you intend to use so that libpng can initialize its internal structures appropriately for all of the filter types.

filters = PNG_FILTER_NONE | PNG_FILTER_SUB
       | PNG_FILTER_UP;
    png_set_filter(png_ptr, PNG_FILTER_TYPE_BASE,
       filters);

It is also possible to influence how libpng chooses from among the available filters. This is done in two ways - by telling it how important it is to keep the same filter for successive rows, and by telling it the relative computational costs of the filters.

double weights[3] = {1.5, 1.3, 1.1},
       costs[PNG_FILTER_VALUE_LAST] =
       {1.0, 1.3, 1.3, 1.5, 1.7};

    png_set_filter_selection(png_ptr,
       PNG_FILTER_SELECTION_WEIGHTED, 3,
       weights, costs);

The weights are multiplying factors that indicate to libpng that the row filter should be the same for successive rows unless another row filter is that many times better than the previous filter. In the above example, if the previous 3 filters were SUB, SUB, NONE, the SUB filter could have a "sum of absolute differences" 1.5 x 1.3 times higher than other filters and still be chosen, while the NONE filter could have a sum 1.1 times higher than other filters and still be chosen. Unspecified weights are taken to be 1.0, and the specified weights should probably be declining like those above in order to emphasize recent filters over older filters.

The filter costs specify for each filter type a relative decoding cost to be considered when selecting row filters. This means that filters with higher costs are less likely to be chosen over filters with lower costs, unless their "sum of absolute differences" is that much smaller. The costs do not necessarily reflect the exact computational speeds of the various filters, since this would unduly influence the final image size.

Note that the numbers above were invented purely for this example and are given only to help explain the function usage. Little testing has been done to find optimum values for either the costs or the weights.

Removing unwanted object code:

There are a bunch of #define's in pngconf.h that control what parts of libpng are compiled. All the defines end in _SUPPORTED. If you are never going to use a capability, you can change the #define to #undef before recompiling libpng and save yourself code and data space, or you can turn off individual capabilities with defines that begin with PNG_NO_.

You can also turn all of the transforms and ancillary chunk capabilities off en masse with compiler directives that define PNG_NO_READ[or WRITE]_TRANSFORMS, or PNG_NO_READ[or WRITE]_ANCILLARY_CHUNKS, or all four, along with directives to turn on any of the capabilities that you do want. The PNG_NO_READ[or WRITE]_TRANSFORMS directives disable the extra transformations but still leave the library fully capable of reading and writing PNG files with all known public chunks [except for sPLT]. Use of the PNG_NO_READ[or WRITE]_ANCILLARY_CHUNKS directive produces a library that is incapable of reading or writing ancillary chunks. If you are not using the progressive reading capability, you can turn that off with PNG_NO_PROGRESSIVE_READ (don't confuse this with the INTERLACING capability, which you'll still have).

All the reading and writing specific code are in separate files, so the linker should only grab the files it needs. However, if you want to make sure, or if you are building a stand alone library, all the reading files start with pngr and all the writing files start with pngw. The files that don't match either (like png.c, pngtrans.c, etc.) are used for both reading and writing, and always need to be included. The progressive reader is in pngpread.c

If you are creating or distributing a dynamically linked library (a .so or DLL file), you should not remove or disable any parts of the library, as this will cause applications linked with different versions of the library to fail if they call functions not available in your library. The size of the library itself should not be an issue, because only those sections that are actually used will be loaded into memory.

Requesting debug printout:

The macro definition PNG_DEBUG can be used to request debugging printout. Set it to an integer value in the range 0 to 3. Higher numbers result in increasing amounts of debugging information. The information is printed to the "stderr" file, unless another file name is specified in the PNG_DEBUG_FILE macro definition.

When PNG_DEBUG > 0, the following functions (macros) become available:

png_debug(level, message)
   png_debug1(level, message, p1)
   png_debug2(level, message, p1, p2)

in which "level" is compared to PNG_DEBUG to decide whether to print the message, "message" is the formatted string to be printed, and p1 and p2 are parameters that are to be embedded in the string according to printf-style formatting directives. For example,

png_debug1(2, "foo=%d\n", foo);

is expanded to

if(PNG_DEBUG > 2)
     fprintf(PNG_DEBUG_FILE, "foo=%d\n", foo);

When PNG_DEBUG is defined but is zero, the macros aren't defined, but you can still use PNG_DEBUG to control your own debugging:

#ifdef PNG_DEBUG fprintf(stderr, ... #endif

When PNG_DEBUG = 1, the macros are defined, but only png_debug statements having level = 0 will be printed. There aren't any such statements in this version of libpng, but if you insert some they will be printed.

VI. CHANGES TO LIBPNG FROM VERSION 0.88

It should be noted that versions of libpng later than 0.96 are not distributed by the original libpng author, Guy Schalnat, nor by Andreas Dilger, who had taken over from Guy during 1996 and 1997, and distributed versions 0.89 through 0.96, but rather by another member of the original PNG Group, Glenn Randers-Pehrson. Guy and Andreas are still alive and well, but they have moved on to other things.

The old libpng functions png_read_init(), png_write_init(), png_info_init(), png_read_destroy(), and png_write_destory() have been moved to PNG_INTERNAL in version 0.95 to discourage their use. The preferred method of creating and initializing the libpng structures is via the png_create_read_struct(), png_create_write_struct(), and png_create_info_struct() because they isolate the size of the structures from the application, allow version error checking, and also allow the use of custom error handling routines during the initialization, which the old functions do not. The functions png_read_destroy() and png_write_destroy() do not actually free the memory that libpng allocated for these structs, but just reset the data structures, so they can be used instead of png_destroy_read_struct() and png_destroy_write_struct() if you feel there is too much system overhead allocating and freeing the png_struct for each image read.

Setting the error callbacks via png_set_message_fn() before png_read_init() as was suggested in libpng-0.88 is no longer supported because this caused applications that do not use custom error functions to fail if the png_ptr was not initialized to zero. It is still possible to set the error callbacks AFTER png_read_init(), or to change them with png_set_error_fn(), which is essentially the same function, but with a new name to force compilation errors with applications that try to use the old method.

VII. Y2K COMPLIANCE IN LIBPNG

January 13, 1999

Since the PNG Development group is an ad-hoc body, we can't make an official declaration.

This is your unofficial assurance that libpng from version 0.81 and upward are Y2K compliant. It is my belief that earlier versions were also Y2K compliant.

Libpng only has three year fields. One is a 2-byte unsigned integer that will hold years up to 65535. The other two hold the date in text format, and will hold years up to 9999.

The integer is

"png_uint_16 year" in png_time_struct.

The strings are

"png_charp time_buffer" in png_struct and
    "near_time_buffer", which is a local character string in png.c.

There are seven time-related functions:

png_convert_to_rfc_1123() in png.c 
      (formerly png_convert_to_rfc_1152() in error)
    png_convert_from_struct_tm() in pngwrite.c, called in pngwrite.c
    png_convert_from_time_t() in pngwrite.c
    png_get_tIME() in pngget.c
    png_handle_tIME() in pngrutil.c, called in pngread.c
    png_set_tIME() in pngset.c
    png_write_tIME() in pngwutil.c, called in pngwrite.c

All appear to handle dates properly in a Y2K environment. The png_convert_from_time_t() function calls gmtime() to convert from system clock time, which returns (year - 1900), which we properly convert to the full 4-digit year. There is a possibility that applications using libpng are not passing 4-digit years into the png_convert_to_rfc_1123() function, or incorrectly passing only a 2-digit year instead of "year - 1900" into the png_convert_from_struct_tm() function, but this is not under our control. The libpng documentation has always stated that it works with 4-digit years, and the APIs have been documented as such.

The tIME chunk itself is also Y2K compliant. It uses a 2-byte unsigned integer to hold the year, and can hold years as large as 65535.


Glenn Randers-Pehrson
libpng maintainer
PNG Development Group