The PSP's (hidden) media engine processor handles decoding of formats like Atrac3+, Aac and so forth. This processor is not directly accessed by games, instead they go through libraries like sceAtrac, sceAac and so on. Atrac3+ is Sony's proprietary format, and the format used for background music in 99% of all games (except those that use MIDI/sequenced music). Mainly "minis" use other formats like MP3 and AAC.
Emulating the sceAtrac library isn't easy, it's a complex high-level library with many different modes and behaviors.
Below is my attempt at documenting how the library is used by games, in the hopes of improving emulation (and who knows, by documenting it, maybe LLMs will one day understand this library, heh).
A lot of the below applies to both Atrac3+ and Atrac3 - the same library is used. Some details differ though like frame sizes, auxiliary data in the file, etc (Atrac3+ doesn't have any).
We first started by using an external library, which required buffering up audio in memory in order to decode, rather than feeding a decoder individual packets. This is bad because we had to fake a lot of the behavior.
When ffmpeg finally added support, we never switched over to using the internal buffering. Mostly because I didn't realize back then that the entire context is conveniently exposed (see below).
But now, finally, we're getting accurate emulation of the buffering, which is critical for some games that use the library in slightly unusual ways, such as Flatout.
An Atrac3+ stream is usually stored in a standard WAVE file, which uses the RIFF container format.
It contains looping information.
There are multiple "coordinate systems" for offsets to be aware of.
ALL_DATA_LOADED and HALFWAY_BUFFER modes, these correspond exactly to file offsets. However, when streaming, there's no simple mapping, it's treated like a circular buffer of "frames", and there's a second buffer for the "tail" when looping an internal segment of a file.The internal state variables uses all these coordinate systems for various things.
NO_DATA = 1: Not yet initializedALL_DATA_LOADED = 2: The entire size of the audio data, and filled with audio dataHALFWAY_BUFFER = 3: The entire size is allocated in memory, but only partially filled so farSTREAMED_WITHOUT_LOOP = 4: Streaming through a small buffer, without any loopSTREAMED_WITH_LOOP_AT_END = 5: Streaming through a small buffer, sliding with a loop at the endSTREAMED_WITH_SECOND_BUF = 6: Smaller with a second buffer to help with a loop in the middleLOW_LEVEL = 8: Not managed, decoding using "low level" API (one packet at a time). Manual looping if needed. By far the easiest to emulate.RESERVED = 16: Not managed, reserved externally - possibly by sceSassceAtracSetHalfwayBufferAndGetID(bufferPtr, validDataSize, bufferSize)Tells the library that you'll supply a partially filled buffer.
Use the combination sceAtracGetRemainFrame/(last output of sceAtracDecodeData)/sceAtracGetStreamDataInfo/sceAtracAddStreamData to progressively fill it up while it's being played.
This is most useful when you have spare memory, and want to loop the audio without extra I/O later.
If the buffer can fit the whole file, the internal state is set to HALFWAY_BUFFER (see above).
You can also use it to only partially initialize a buffer for streaming, GTA does this. In this case, the state will end up in one of the streaming modes, depending on the loop parameters stored in the WAVE header.
Possible resulting states:
ALL_DATA_LOADEDHALFWAY_BUFFERSTREAMED_WITHOUT_LOOPSTREAMED_WITH_LOOP_AT_ENDSTREAMED_WITH_SECOND_BUFsceAtracSetDataAndGetID(data, size)Like sceAtracSetHalfwayBuffer, but fills the entire specified buffer.
Size can be any value (down to certain limitations), even byte-aligned! 0x4301 works and doesn't crash. Games generally uses sizes like 0x18000, 0x10000, etc.
Depending on the file contents and the buffer size, the internal state can end up as any of:
ALL_DATA_LOADEDSTREAMED_WITHOUT_LOOPSTREAMED_WITH_LOOP_AT_ENDSTREAMED_WITH_SECOND_BUFsceAtracGetNextSample(id, *outPtr)Returns how many audio samples will be written by the next call to sceAtracDecodeData.
sceAtracGetMaxSample(id, *outPtr)Returns the maximum possible value that sceAtracGetNextSample can return (use it to size your buffer, maybe?).
sceAtracGetNextDecodePosition(id, *outPtr)Returns the sample position (NOT frame position) in the decoded audio track of the next frame to be decoded. Usually not very useful, although is quite suitable for synchronizing events to the audio playback.
Note that at the start of the song, this will have some value like 0x970 because the decoder always skips a certain number of samples at the start before writing them out.
sceAtracDecodeData(id, *outPtr, *outNumSamples, *outFinishFlag, *outRemainFrames)Decodes some audio data to memory, reading from the previously provided data (if outPtr is null, the decoded data is simply discarded, processing happens as normal).
*outNumSamples is set to the number of (mono or stereo) samples written. The first and last frames, and also possibly frames when looping, are usually short (unless the size and/or loop points are carefully aligned), while all other frames are 0x800 samples (2048), at least for Atrac3+. If you're playing back the result by directly pushing decoded buffers to sceAudioOutputBlocking, you're gonna need to move the frame forward within the buffer size, to avoid a glitch. Example:
// (dec_frame == outPtr, maxSamples == 2048, and a pointer to `samples` is passed in to the `outNumSamples` argument)
if (firstFrame && samples < maxSamples) {
printf("Deglitching first frame\n");
memmove(dec_frame + (maxSamples - samples) * 4, dec_frame, samples * 4);
memset(dec_frame, 0, (maxSamples - samples) * 4);
}
Note that we need to check for firstFrame, otherwise we might do this to the last frame, which can also be short! In the latter case, you might want to zero the rest of the buffer, or maintain your own ring buffers and feed that to sceAudioOutputBlocking (this is necessary for smooth looping, too).
Although of course you don't actually need a bool to determine the first frame, you can also just call sceAtracGetNextDecodePosition, though the return value isn't super easy to interpret.
If the end is reached and looping is off, *outFinishFlag is set to 1, otherwise 0 is written (including when looping).
This call is blocking, so while decoding is being processed in the background (by the ME), the main CPU will switch over to other available threads to perform work.
The number of frames remaining (left to decode) in the buffer is returned through *outRemainFrames (exactly the same as calling sceAtracGetRemainFrame, so that function is redundant, really).
sceAtracGetRemainFrame(id, *outPtr)Returns the number of frames that can be decoded from the currently buffered data, or -1 or -2 if you're not gonna need to load more before the loop (depending on if looping is on). If streaming or half-way filled, when the returned value goes below a game-specific threshold, the game will load more data into memory so that it doesn't run out.
Note, this same value is also returned through the last parameter to sceAtracDecodeData, so it's usually not necessary
to call this one on its own.
sceAtracResetPlayPosition(id, buffer, bytesWrittenFirstBuffer, bytesWrittenSecondBuffer)This is called after GetBufferInfoForResetting and then filling the buffers accordingly.
sceAtracGetStreamDataInfoCall this to figure out what data to read from the file and where in the main buffer to put it.
Then call sceAtracAddStreamData to notify sceAtrac about how much data you actually added.
sceAtracAddStreamDataTells the engine that you loaded the data it wanted.
Note: It's OK to "add" a smaller amount of data than sceAtracGetStreamDataInfo asked you to.
sceAtracSetSecondBuffer(id, buffer, bufSize)This informs sceAtrac about where we want a second buffer to be placed. This is read from in a few circumstances like when looping with a tail (that gets played after the last loop).
sceAtracGetChannel(id, *outPtr)Returns the number of channels the audio track has.
sceAtracGetSoundSample(id, *outEndSample, *outLoopStartSample, *outLoopEndSample)Gets the length of the track and the extents of the loop (if any, otherwise -1), counted in output audio sample positions.
sceAtracReleaseAtracID(id)Releases the Atrac context, free-ing up a bit of memory.
_sceAtracGetContextAddress(id)Returns a pointer to the internal context data.
The previous Atrac3+ implementation only copied some variables back to this. The new one uses it directly to hold most state.
During init, you should call sceAtracGetMaxSample to figure out your audio channel buffer size. You need to round-robin these, not sure if 2 or 3 are needed,
I've seen games use both I think. This gets passed into the second argument of sceAudioChReserve.
sceAtracGetNextSample returns how many audio samples will be written by the next call to sceAtracDecodeData. If it's less than sceAtracGetMaxSample() you probably want to offset the output a bit to line up with the next frame to avoid a small glitch.
The game calls sceAtracDecodeData with a pointer to a destination buffer to get the next frame(s).
To know when to top up the buffer (if streaming), a game calls sceAtracGetRemainFrame(id, *u32) or just checks the last argument to sceAtracDecodeData.
When that value goes below a specific threshold, call sceAtracGetStreamDataInfo to figure out what data you need to read, and where in memory to read it to. It will be somewhere inside the specified buffer (in SetData). After adding the data, you call sceAtracAddStreamData to notify the library of exactly how much data you added.
To seek when streaming, you call sceAtracGetBufferInfoForResetting to figure out what to write, then you call
sceAtracResetPlayPosition.
It just can't work! Well, that was my first reaction. But there's a nasty hack going on.
The first time around when we set up a buffer for streaming using sceAtracSetDataAndGetID(), sceAtrac isn't able to give us any advice on how much data to write, in order to avoid splitting packets. So, what it does is, it'll accept say 0x8000 bytes. As it turns out, at that size, offset 0x8000 is right in the middle of a packet, so it simply copies the first half of the packet right on top of the headers it just read, at the beginning of the buffer.
Subsequently, the next time we call sceAtracGetStreamInfo, it'll apply the appropriate offset, and we'll end up completing the packet! So when decoding comes along, it won't have to read a split packet.
However, if you supply less data than asked, you can still create split packets temporarily, though you should complete them before decoding gets there.
Starts with a halfway initialized buffer:
Good ones for testing (early)
Others
In this case, the game just submits packets to decode to the decoder, sceAtrac doesn't really do anything. This is the easiest mode to emulate and has worked reliably for a long time.
There's a way to use Atrac3+ audio as a sound channel in the sceSas software mixer, by using the __sceSasSetVoiceATRAC3 function.
This is rare, known game using it:
NOTE: Logs below are from PPSSPP (and the old sceAtrac implementation), not hardware.
They will be replaced later.
STREAMED_WITHOUT_LOOP
0=sceAtracSetDataAndGetID(09ef9f00, 00040000)
0=sceAtracGetRemainFrame(0, 09ffee24[0000015f])
0=sceAtracGetMaxSample(0, 09ffee3c[00000800])
loop:
0=sceAtracDecodeData(0, 09ef1f00, 09ffee28[00000690], 09ffee38[00000000], 09ffee24[350])
Until remainFrame (last param) <= 175. Then:
0=sceAtracGetStreamDataInfo(0, 09ffee2c[09ef9fa0], 09ffee30[00020228], 09ffee34[00040000])
0=sceAtracAddStreamData(0, 00020000)
goto loop
ALL_DATA_LOADED
0=sceAtracSetData(0, 0952e700, 00177dac)
0=sceAtracSetLoopNum(0, -1)
0=sceAtracDecodeData(0, 09fe1680, 09fdb664[00000180], 09fdb660[00000000], 09fdb668[-1])
loop:
0=sceAtracDecodeData(0, 09fe1680, 09fdb664[00000800], 09fdb660[00000000], 09fdb668[-1])
goto loop
Uses an unusually small buffer. We match this in our stream.prx test.
0=sceAtracSetDataAndGetID(092bd940, 00010000)
0=sceAtracSetLoopNum(0, -1)
0=sceAtracGetRemainFrame(0, 09fdea78[00000056])
0=sceAtracIsSecondBufferNeeded(0)
0=sceAtracDecodeData(0, 092cfa00, 09fdea70[000003b4], 09fdea74[00000000], 09fdea78[85])
0=sceAtracDecodeData(0, 092d08d0, 09fdea70[00000800], 09fdea74[00000000], 09fdea78[84])
0=sceAtracDecodeData(0, 092d28d0, 09fdea70[00000800], 09fdea74[00000000], 09fdea78[83])
0=sceAtracGetStreamDataInfo(0, 093241e0[092bdbc4], 093241e4[000080e8], 093241e8[00010000])
0=sceAtracDecodeData(0, 092d1ad0, 09fdea70[00000800], 09fdea74[00000000], 09fdea78[41])
0=sceAtracAddStreamData(0, 000080e8)
0=sceAtracDecodeData(0, 092d3ad0, 09fdea70[00000800], 09fdea74[00000000], 09fdea78[85])
Uses the FMOD library.
NOTE: Does two setdata! First one just with the header I guess. Then the proper one.
Later, it seems to run out of data just before refilling, even though there should be ~88 buffered frames left. This makes little sense.
This never worked before (see below), but the new implementation fixes this just fine!
0=sceAtracSetDataAndGetID(08b89d40, 00001000)
0=sceAtracGetChannel(0, 09fdee70[00000002])
0=sceAtracGetBitrate(0, 09fdee74[00000060])
0=sceAtracGetSoundSample(0, 09fdee78[007946db], 09fdee7c[ffffffff], 09fdee80[ffffffff])
0=sceAtracGetBufferInfoForResetting(0, 0, 09fdee40) << maybe important that this is called before SetData?
0=sceAtracSetData(0, 08fe93c0, 00018000)
SCE_KERNEL_ERROR_SCE_ERROR_ATRAC_NO_LOOP_INFORMATION=sceAtracSetLoopNum(0, 0): no loop information to write to!
0=sceAtracGetRemainFrame(0, 09fdedb0[000000ae])
0=sceAtracDecodeData(0, 08fd5bd0, 09fdedcc[00000690], 09fdedd0[00000000], 09fdedb0[173])
0=sceAtracGetRemainFrame(0, 09fdedb0[000000ad])
...
0=sceAtracDecodeData(0, 090014c0, 09fe199c[00000800], 09fe19a0[00000000], 09fe1980[87])
0=sceAtracGetRemainFrame(0, 09fe1980[00000057])
0=sceAtracGetStreamDataInfo(0, 09fe1990[08fe9490], 09fe1994[0000c010], 09fe1998[00018000])
0=sceAtracAddStreamData(0, 0000c010)
0=sceAtracDecodeData(0, 08fdb610, 09fe199c[00000800], 09fe19a0[00000000], 09fe1980[174])
0=sceAtracGetRemainFrame(0, 09fe1980[000000ae])
Uses sceAtracSetHalfwayBufferAndGetID with a very small readSize, and a buffer much smaller than the size of the file. This results in a streaming state, not a "halfway" state, but differently initialized. This is now implemented and working in Atrac2.