XA Audio

CD-ROM XA ADPCM playback through the CD-ROM decoder.

Summary

ADPCM (Adaptive Differential PCM) compresses audio data encoded as 16-bit straight PCM to 1/4 of its size. A sector that stores ADPCM data is called an ADPCM sector, and strictly it is a real-time or mode 2 form 2 sector.¹² The CD-ROM decoder decodes the sectors itself and sends the audio to the SPU mixer without using main memory, the sound buffer or an SPU voice.³ The XA data is held in the local memory of the CD subsystem and decoded there.⁴ The file format and the four audio types are described in XA.

Interleave

An ADPCM sector holds audio compressed to 1/4, so the disc delivers audio faster than it plays. To play an audio series without seeking for each sector, ADPCM sectors are recorded on the disc at every fourth sector for normal speed play and at every eighth sector for double speed play. Sony calls this interleave, and it makes it possible to read data from the other sectors while ADPCM plays.¹²

The interleave depends on the speed, the sample rate and the number of channels. psx-spx lists the usual values at double speed:⁵

The other sectors can also hold other ADPCM channels. Sony's example interleaves 8 channels, A0 to A7, one after another, so the program can switch between 8 audio channels without a seek.⁶ psx-spx notes that it is not known whether the 1/16 and 1/32 interleaves always work, because the controller overwrites its buffer entries once every eight sectors. However, Crash Team Racing plays 37800 Hz mono at double speed, and 1/16 therefore works.⁵

Channel Selection

The channel to play is chosen by the file number and channel number in the subheader of each sector. CdlSetfilter sets the numbers to play, and CdlModeSF must be set in the mode for the filter to take effect.⁶ The file number selects the playing channel of a multichannel XA-ADPCM disc. The default is file 1 and channel 1.⁷⁸

psx-spx gives the order in which the controller handles a sector. It first tries to send the sector to the ADPCM decoder, and if that fails it tries to send it to the CPU.⁹

A sector that is not played as ADPCM goes to the CPU with INT1, unless CdlModeSF is set and the sector is an audio and real-time sector. With CdlModeRT or CdlModeSF set, INT1 therefore comes only for the data sectors.⁹¹⁰ Since the sector size in the mode does not apply to ADPCM, the decoder always decodes the full 900h bytes.¹⁰

Playing XA Audio

XA-ADPCM sectors are organized in files instead of tracks. They are therefore played with a read command instead of CdlPlay.¹⁰ The SPU is first set up for CD input. The program then sets CdlModeRT (and CdlModeSF if needed) with CdlSetmode, sets the start sector with CdlSetloc and issues a read, normally CdlReadS.¹⁰ CdlReadN retries even in CdlModeRT, so CdlReadS is used to inhibit retries.¹¹ There is no way to detect the end of XA playback automatically. The program polls the sector ID periodically to find it.¹²

Sector Layout

An ADPCM sector carries 12h portions of 128 bytes (900h bytes), and the rest of its 914h-byte data area is zero. Each portion holds a 16-byte header and 28 data words of 4 bytes.¹³

The copies let the controller replace a header that it read with an error.¹³ Each header byte holds the shift in bits 0 to 3 (0 to 12, where 0 is loudest) and the filter in bits 4 and 5. Bits 6 and 7 are unused.¹⁴

The 28 data words follow at 10h to 7Fh. Word N holds sample N of every block: in 4-bit mode eight nibbles for blocks 1 to 8, and in 8-bit mode four bytes for blocks 1 to 4. In stereo the entries alternate between left and right, and the first entry is left.¹⁵

Decoding

The decoding is the same as for SPU-ADPCM in SPU ADPCM. Each 4-bit sample is shifted left by 12 (or an 8-bit sample by 8), then right by the shift, and the filter adds a weighted sum of the two previous outputs.¹⁴¹⁶ XA uses only filters 0 to 3, with weights (0, 0), (60, 0), (115, -52) and (98, -55) in 64ths. The SPU also has filter 4.¹⁷ For 8-bit data the useful shifts are 0 to 8, and the reserved shifts 13 to 15 act as 9.¹⁴

The decoder converts the 37800 Hz output to 44100 Hz with a 25-point zigzag interpolation. The SPU uses a 4-point Gaussian interpolation instead. A six-step counter makes seven outputs from every six samples (37800 x 7 / 6 = 44100). Each 900h-byte sector holds a multiple of six samples, so the counter is the same at the start and end of each sector.¹⁸¹⁹ The counter starts at a random value at power on.¹⁹ Emphasis is supported but no known game uses it.²⁰

Sound Map

In sound map mode the host sends ADPCM data to the decoder through WRDATA in place of the disc. This is useful for testing XA-ADPCM from an executable, and it supports both 4-bit and 8-bit data, while the SPU supports only 4-bit.²¹ The procedure from psx-spx is as follows:²¹

Output starts after the last of the 900h bytes is written. The decoder stores the block in one of three 900h-byte slots, but it does not always play the slot that was written last.²¹ The data sheet adds that the BFWRDY interrupt shows room for one more sector and BFEMPT shows that the decoder has run out of sectors. SMADPCLR in HCLRCTL ends sound map decoding at once.²²²³

References