Picture format

Many of the Doom engine graphics, including wall patches, sprites, and all menu graphics, are stored in the WAD files in a special picture format, known as the patch format. This format is unique to Doom. Notably excepted are the textures for floors and ceilings, which are known as flats. There is also a different picture format used for the starting screens, help pages, and the Heretic picture format.

Details of the patch format in Doom are given in the Unofficial Doom Specs. (Note: the referenced copy of the UDF is broken, and the graphics section seems to be inaccessible).

Patch format patch_t: uint16_t width; uint16_t height; int16_t  leftoffset;  // pixels to the left of the origin int16_t  topoffset;   // pixels below the origin uint32_t columnofs[]; // array[width] of column offset // locates the column data as an offset from the first byte of this header // Each column is an array of post_t, of unknown length, terminated by a byte 0xFF.

Patch column format post_t: byte     topdelta;    // the y offset of this post in this patch // if topdelta is 0xFF then are at end-of-column (not a valid post) byte     length;      // length of data in this post byte     data[];      // array[length] of pixels // each data pixel is an index into the Doom palette

A patch header gives its width and height, and offset values. Following the header are pointers to data for each column of pixels; the number of these pointers is equal to the picture width.

The data for each column is divided into posts, which are lines of colored pixels going downward on the screen. Each post is described by its starting height (relative to the top of the picture) and number of pixels, followed by a value for each of the pixels. Picture descriptions can (and do) skip over some pixel positions; these pixels are transparent. (Since transparent pixels are not changed when drawing a particular picture, whatever was drawn into the frame buffer previously will show through.)

Each pixel is given as an unsigned byte (and thus is valued from 0 to 255). The pixel value is first used as an index into the current COLORMAP, which gives a new pixel value (from 0 to 255) adjusted for the desired light level. (At full brightness, the pixel value is unchanged.) Then this new pixel value is written into the frame buffer. The actual red, green, and blue values corresponding to the palette index in the current palette are stored in the VGA graphics card's 8-bit hardware palette.

Note that gamma correction, a user-adjustable setting that can lighten the colors for dark-looking monitors, is handled when setting the game's palette and not when actually drawing the graphics themselves. This avoids an additional indirection.

Converting to a doom picture
This algorithm will convert some pixel data (eg, from a windows bitmap file) to a doom picture.

Notes: --

Byte = 0 - 255 Word = 0 - 65535 DWord = 0 - 4294967295

dummy_value = 	Byte, those unused bytes in the file (excerpt from UDS: "..left overs from NeXT machines?..") picture_* = 	Word, the maximum width for an image in doom picture format is 256 pixels pixel_count = 	Byte, the number of pixels in a post Pixel = 	Byte, the pixel colour column_array =	array of DWord, this holds all the post start offsets for each column

Algorithm: --

begin

write picture_width to file write picture_height to file write picture_top to file write picture_left to file

while loop, exit on x = picture width increase column_array by 1

write memory buffer position to end of column_array

y = 0

operator = true

do while loop, until y = picture height get Pixel value if the pixel = transparent_colour and operator = false then dummy_value = 0 write dummy_value to memory buffer

operator = true

otherwise, if pixel != transparent_colour and operator = true then row_start = y			pixel_count = 0

dummy value = 0

write above post data to memory buffer

offset = current post position in memory buffer

operator = false

otherwise, if pixel != transparent_colour and operator = false then increment current post pixel_count

if offset > 0 and pixel count > 0 then previous_offset = current post position

seek back in memory buffer by offset - 2

write pixel_count to memory buffer

seek back to previous_offset end block write pixel to memory buffer end block

increment y by 1

end block

if operator = true or y = height then Pixel = 0

write Pixel to memory buffer

rowstart = 255 write rowstart to memory buffer end block

increment x by 1

end block

seek memory buffer position to 0

block_size = picture_width * size of dword

allocate block_memory, filled with 0's, with block_size

write block_memory to file, using block_size as size

offset = current file_position

free block_memory

seek to position 8 in file from start

for loop, count = 0, break on count = number of elements in column_array column_offset = column_array[count] + offset

write column_offset to file end block

write memory buffer to file

Converting from a doom picture
This algorithm will convert a doom picture to some pixel data.

Notes: --

Byte = 0 - 255 Word = 0 - 65535 DWord = 0 - 4294967295

dummy_value = 	Byte, those unused bytes in the file (excerpt from UDS: "..left overs from NeXT machines?..") picture_* = 	Word, the maximum width for an image in doom picture format is 256 pixels pixel_count = 	Byte, the number of pixels in a post Pixel = 	Byte, the pixel colour column_array =	array of DWord, this holds all the post start offsets for each column

doom image = could be a file or memory stream

Algorithm: --

create a image with a pixel format of 8bit and the doom palette, set the background colour to a contrasting colour (cyan).

read width from doom image (word) read height from doom image (word) read left from doom image (word) read top from doom image (word)

create column_array with width number of elements

for loop, i = 0, break on i = width - 1 column_array[i] = read from doom image, 4 bytes end block

for loop, i = 0, break on i = width - 1 seek doom image to column_array[i] from beginning of doom image

rowstart = 0

while loop, rowstart != 255 read rowstart from doom image, 1 byte if rowstart = 255, break from this loop

read pixel_count from doom image, 1 byte

read dummy_value from doom image, 1 byte

for loop, j = 0, break on j = pixel_count - 1 read Pixel from doom image, 1 byte write Pixel to image, j + rowstart = row, i = column end block read dummy_value from doom image, 1 byte end block end block

Earlier versions of the format
The patch format originally (in the alpha and beta version) was more compact but limited. The alpha used bytes (instead of words) for dimension and starting offsets in the picture header. Both alpha and beta use words (instead of dwords) for locating column offsets, and neither features the dummy bytes that were added to the column spans in the final version.

For comparison, a table of the old and the new inserted in the patch_t struct from the Doom source code:

A minor variant on the beta format is also used by Rise of the Triad:

typedef struct {   short          origsize;         // the orig size of "grabbed" gfx short         width;            // bounding box size short         height; short         leftoffset;       // pixels to the left of origin short         topoffset;        // pixels above the origin unsigned short collumnofs[320]; // only [width] used, the [0] is &collumnofs[width] } patch_t;

Other early picture formats
Since the alpha version of the picture format used a byte for height and width, it did not support images larger than 255 pixels in any dimension. Other formats were therefore needed for wide graphics. The following formats have been used in alpha versions before being discarded: SLADE3 can display all these graphic formats.
 * Alpha raw-and-header: Uses a header similar to that of final patches (four shorts for width, height, and presumably left and top offsets) followed by a raw dump of pixel data technically identical to flats (one byte per pixel, row-major). Index 255 is transparent. This is used notably for the Doom v0.2 PLAYSCREEN lump.
 * GNUM: A simple raw dump of pixel data technically identical to flats, but with a size 10x12 pixels. This is used in Doom v0.4 for a series of lumps called GNUM0 to GNUM9.
 * Snea: Uses a two-byte header indicating the quarter of the width and the full height. This is followed by an interleaved, column-major dump of pixel data: columns are described in the sequence 0, 4, 8, etc., then 1, 5, 9, etc. and so on. This format was introduced as a replacement of the raw-and-header one in Doom v0.4, and remained until the Beta where it was still used for the TITLEPIC (but nothing else).
 * HUFONT: Uses a 770-byte header indicating first the character height on a short, then the character width of each 256 characters in sequence, each on a single byte, then the start offset of each character data on a short. Character data is in column-major format, color index 0 is transparent. This was used in Doom v0.4 and Doom v0.5 for the system font.

Tall patches
A post starts with a two-byte header, the first byte containing the post's "topdelta", or its vertical offset from the top of the picture, and the second containing its length. This means that a post has severe limitations. Since the format uses a single byte to store a post's offset from the top of the image, if a picture is tall (more than 255 pixels high) it is possible that a post's topdelta becomes impossible to express normally. A topdelta of 255 is reserved to signal that there are no more posts in the column, and since it is stored in a single byte, a topdelta therefore cannot be larger than 254.

A workaround was designed, which is known as DeePsea tall patches. A post's topdelta is compared to the previous post's topdelta (or to -1 if there are no previous posts in the current column). If the new topdelta is less than the previous topdelta, it is interpreted as representing a tall patch, and the two values are added together, with the sum serving as the current post's offset. Since a post is normally supposed to go below its predecessor, this trick is usually non-ambiguous. However, if support for tall patches is not implemented, then these posts end up overwriting previous posts by being put at their literal offset, rather than the derived one.