Electronic Arts MAD

• Extension: mad
• Company: Electronic Arts
• Samples: http://samples.mplayerhq.hu/game-formats/ea-mad/

MAD is a video codec used in various games published by companies under the Electronic Arts umbrella. The codec is an evolution of the earlier TGQ and TQI codecs, adding motion compensation and low-quality frame concepts. MAD is an abbreviation for "Madcow Movie".

Chunk Types

See Electronic Arts Formats for file format description. Each MADk, MADm and MADe chunk corresponds to a video frame, where MADk is an intra-frame, MADm is an inter-frame and MADe is a low-quality-inter frame. Low-quality-inter frames are identical to inter frames, but are not used for motion compensation and are discarded immediately after presentation. Each frame is divided into 16x16 pixel macroblocks, with each macroblock further divided into with four 8x8 luma and two half-resolution 8x8 chroma blocks. Standard MPEG-1 block ordering is used.

Each chunk begins with a 16 byte header, followed by a bitstream. Symbols within the header are encoded using 16-bit little-endian most-to-least-significant bit packing. <FIXME: explain bit packing>. Although frame type is reported within the header, the field is not valid in some samples, and is therefore unreliable. Frame type should be inferred from the chunk type.

   Offset   Data_type   Name         Description
   0x0000   four bytes  unknown      always zero
   0x0000   le_u16      unknown
   0x0006   le_u16      frame_rate   milliseconds per frame
   0x0008   le_u16      width        image width (pixels)
   0x000a   le_u16      height       image height (pixels)
   0x000c   u8          unknown
   0x000d   u8          quantizer    frame quantizer
   0x000e   u8          frame_type   not valid for all samples
   0x000f   u8          unknown
   for each vertical macroblock
       for each horizontal macroblock
           decode_mb

Macroblock Decoding

Macroblock composition is based on a 6-bit block type, where each bit indicates the type of an 8x8 block within macroblock. The least-significant-bit corresponds to the first block. A clear bit indicates an intra-block, and a set bit indicates a motion-compensated block.

For intra frames, the block type is assumed to be all clear. For inter and low-quality inter frames, the block type is derived from a VLC. If the block type is none zero, then a full-pel motion vector is read from the bitstream. A description of the block_type VLC and decoding algorithm is provided below.

  if (intra frame) 
      block_type = 000000b
  else // inter or low-quality-inter frame
      switch get_vlc()
      case 1b  : block_type = 111111b
      case 01b : block_type = get_bits(6);
      case 00b : block_type = 000000b
  
      if (block_type != 0)
          x = GetVector();
          y = GetVector();
  
  for i = 0 through 5	// for each block
      if block_type & (1<<i)
          mc_offset = GetVector()
          motion_compensation
      else
          decode_coeffs

Full-pel 8x8 block motion compensation is performed using the motion vector. Following the compensation step, each pixel in the block is offset by the mc_offset value and clipped to 0, 255.

GetVector Definition

Vector values are read from the bitstream using the following algorithm.

   value = 0
   if (get_bits(1))
       if (get_bits(1))
           value = -17
       value += get_bits(4) + 1

Coefficient Decoding

All coefficients, including DC, are decoded using the MPEG-1 AC run-length VLC technique. The RLVLC table is identical to that used for the TGQ codec and coefficients are ordered using the EA zigzag scan pattern. Dequantization is performed by multiplying the coefficients by a dequantization table. This table is calculated for each frame using the frame quantizer.

   dequant_table[0] = base_table[0] * base_table2[0]
   for i equals 0 through 63
       dequant_table[i] = (base_table[i]*base_table2[i]*frame_quant + 4) / 8

   where base_table and base_table2 are defined in TGQ and TQI respectively.

Inverse DCT and Clipping

An IDCT-based algorithm is applied to each block. <FIXME: hyperlink to implementation>. The resulting values are offset by 128, and clipped to 0, 255 to give 8-bit luma and chroma pixel levels.

Games Using MAD

See Electronic Arts Formats for listing.