From xboxdevwiki
Revision as of 20:50, 1 July 2017 by Codeasm (talk | contribs) (Voice Processor (VP): Source link for number of voices)
Jump to: navigation, search

The MCPX contains an APU (Audio Processing Unit).

  • SSL = Stream Segment List
  • SGE = Scatter Gather Entry
  • PRD = Physical Resource Descriptor (Same thing as SGE?!)

Frontend Engine (FE)

Voice Processor (VP)

A powerful voice processor. There can be up to 256 voices[1] and 64[2] of those can be 3D.

Per-voice settings:

  • Input type (8bit, 16bit, 24bit, ADPCM)
  • Head-related transfer function (HRTF)
  • Low-frequency oscillation (LFO)
  • Pitch (~187.5 Hz to ~12285920.7 Hz)
  • Optionally one of the following filters modes:
    • For 2D Mono:
      • DLS2 Low-Pass
      • Parametric Equalizer
      • DLS2 Low-Pass + Parametric Equalizer
    • For 2D Stereo:
      • DLS2 Low-Pass
      • Parametric Equalizer
    • For 3D:
      • DLS2 Low-Pass + I3DL2 Reverb
      • Parametric Equalizer + I3DL2 Reverb
      • I3DL2 Reverb
  • 2 Envelopes (DAHDSR: Delay, Attack, Hold, Decay, Sustain, Release)
    • Amplitude Envelope
    • Pitch / DLS2 Low-Pass Cutoff Envelope
  • 8 target bins, each with a custom volume for this voice

There are 32 bins which these voices will be mixed into.

Related APU memory

  • VPV = VP Voices
  • VPHT = VP HRTF Target
  • VPHC = VP HRTF Current

Voice lists

The voices are kept in a single-linked list. There are 3 voice lists:

  • 2D
  • 3D
  • MP (Multipass?)

Voice structure

This is 0x80 bytes

Pitch calculation

The 16 bit signed pitch value (p) can be converted to and from a unsigned frequency in Hz (f) using the following formulas:

p = 4096 * log2(f / 48000)
f = pow2(p / 4096) * 48000


There are seperate sections of the envelopes, 2 registers (CUR and COUNT) per envelope keeps track of this:

  • 0: Off = Envelope is not used
  • 1: Delay = Time where envelope stays at 0% until attack
    • COUNT register counts down.
  • 2: Attack = Rate at which the envelope goes from 0 to 100%
    • COUNT register counts up.
  • 3: Hold = Time the envelope stays at 100%
    • COUNT register counts down.
  • 4: Decay = Rate at which the envelope goes from 100% to 0%
    • COUNT register counts down.
    • When sustain level is reached the decay section is over
  • 5: Sustain = Level at which the envelope stays while the voice is being played
    • COUNT register is 0
  • 6: Release = Rate at which the envelope goes from current level to 0%
    • Can start at any time[FIXME]
    • COUNT register counts down[FIXME]
  • 7: Force Release = Unknown still[FIXME]

All durations are described using unsigned 12-bit times/rates. The level of sustain is stored unsigned in 8-bit. The COUNT register is stored in unsigned 16-bit.

The 12-bit times/rates are multiplied by 16 when loading them into the 16-bit COUNT register. The COUNT register counts at 1500 Hz[3]. A unit in the COUNT register is therefore 0.6 ms.

The 12-bit values of the envelope sections are given in units of 0.6 ms * 16 = 10.6 ms. This can also be written as 512 / (48000 Hz) = 10.6 ms. The maximum length of an envelope section is therefore 4095 * 10.6 ms = 43.68 seconds.

As the envelope counter runs at a fixed clock speed, it is independent of the voice pitch and duration.

If the Amplitude Envelope hits the zero level during release, DirectSound[FIXME] already deletes the voice, regardless of the Filter Envelope.

The sustain level can be changed during playback. Also the attack register can be changed to a lower value while the counter is counting up, however, if the COUNTER does not compare equal to the set value, it will keep counting, even after an overflow. It will not leave the attack phase and keep counting until it sees value COUNTER / 16 in the attack register. If the attack register is set to a higher value while counting, the volume is going down again. Also, if the attack register value is zero while counting, there won't be any audio output during the attack phase. This indicates that the COUNT register is used to calculate the actual value from the current rates.

Amplitude Envelope

The amplitude envelope is mixed with the volume during mixing. The volume registers are not modified. [FIXME] It is not yet known how many bits of the envelope state are used.

Filter Envelope


The pitch scale is multiplied with the current envelope state and added to the current pitch during mixing. The pitch registers are not modified.

f = 2^((signed_pitch+signed_pitch_mod*32*envelope_state_float)/4096)*48000 # envelope_state_float: [0, 1]

It is not yet known how many bits of the envelope state are used.



Formulas from DirectSound

FreqToHardwareCoeff(frequency): # Input in Hz
  if (frequency < 30) { return 0x8000 }
  if (frequency > 8000) { return 0x0000 }
  FC = 2 * sin(PI * frequency / 48000)
  octaves = 4096 * log2(FC)
  return octaves & 0xFFFF
hardware_coefficient[0] = FreqToHardwareCoeff(frequency)
dBToHardwareCoeff(resonance): # Input in dB
  resonance = min(resonance, 22.5)
  Q = pow(10, -0.05 * resonance)
  return min(0xFFFF)
hardware_coefficient[1] = dBToHardwareCoeff(resonance_in_db)

Stuff from the DLS2 spec:[FIXME]

There are 2 coeffiecents per channel:

  • F_c (Cutoff frequency)
  • resonance

From Page 8 of "DLS 2.2 Version 1.0"[4]

  • b_1 = -2 * r * cos(θ)
  • b_2 = r * r
  • K = g * (1 + b_1 + b_2)
y[i] = K * x[i] - b_1 * y[i-1] - b_2 * y[i-2]

Where y[i-2] and y[i-1] are the last two frames of the output and x[i] the current input.


Voices are stored in VPV. Input data (from the CPU) is loaded using VPSGE. Voices are then processed and written to the GP MIXBUF.

Global Processor (GP)

The GP is a DSP to do programmable audio processing on the voice bins.

The GP DSP seems to run at 160 MHz.


The MIXBUF is a 0x400 word (24-Bit, stored as 32-Bit) section. It is split into 32 * 0x20 words. Each 0x20 word block represents one of the 32 voice bins of the VP. The 0x20 words are 24-Bit PCM mono samples to be played back at 48kHz. The duration of each frame is hence 0.6ms.

Memory map

Related APU memory

  • GPS = GP Scratch (?)

Encode Processor (EP)

The EP is a DSP to encode the audio signal.

Memory map

Related APU memory

  • EPS = EP Scratch (?)

Usage in DirectSound

This topic deserves it's own article[FIXME]

The bins are used [FIXME] DirectSound allows to load custom GP DSP code for a filter / effects chain. [FIXME] The GP waits for the frame interrupt which signals that MIXBUF data is available. It then goes through the filter chain. At the end of the chain, the GP DSP will verify that the execution didn't take longer than the frame duration.

The GP will then issue 6 DMA requests to output the processed frames to a ringbuffer in scratch space. The frameformat will be the same format as the GP MIXBUF format (also 0x20 words per channel). Each ringbuffer is 0x200 words and therefore holds the last 16 frames. Therefore, the ringbuffer region is 6 * 0x800 Bytes = 0x3000 Bytes in physical memory.

The order of the channels in the ringbuffer is (also DMA order):

The EP maps the same data to its own scratch space. It is assumed that it will DMA this region to its own internal memory. The EP then AC3 encodes the audio data[citation needed] and writes it to the EP FIFO memory[FIXME]. [FIXME] The data is then send to the ACI AC97 using EP FIFO channels 0 (PCM) and 1 (SPDIF)[citation needed]. The EP code is loaded by DirectSound. The EP is not programmable using DirectSound.

Modifications for Boot Animation

During the Boot Animation a different version of DirectSound is used. The EP is disabled in this case. The data is send to the ACI AC97 using GP FIFO channel 0 (PCM). There is no AC3 / SPDIF during the boot animation[5][citation needed].

Related notes