MP3

MP3

I am looking into the prospects for playing MP3 music files with the Propeller. At this point, it looks futile, but I'm looking at it closely anyway... The main issues are amount of RAM and processing speed.

	Goal: Playback of stereo, 44100-kHz 128-kbps MP3 file from SD card
	Fallback 1: Play mono or single channel with the option of using 2 Props to play stereo.
	Fallback 2: Wait for Prop II (coming soon?)

I've found that the LIBMAD software is well suited for the task because it uses integer (fixed-point) math whereas most others use floating-point routines. LIBMAD appears to have been used in many embedded projects.

MP3->WAV converter in SPIN.

There's no way that real-time MP3 playback could be achieved in SPIN. But, code in SPIN and PASM (Propeller Assembly) seem to require about the same amount of RAM and the relative speed difference is know to be between 40 and 200. So, if an MP3 file can be decoded in SPIN in ~100x real time, then the prospects would be good for real-time playback when converted to PASM.

So far, I am able to decode the first 576 samples of a test MP3 file with two SPIN programs (because it wouldn't fit into 32k). This represents the first channel of the first granule of the first frame. The first SPIN program decodes the MP3 data stream and produces time domain samples in the XR array. The second SPIN program converts the XR array to frequency domain and then synthesizes the PCM samples.

Program 1: MP32Wav2A.SPIN

Requires these files:

	test.mp3
	rq_table.dat
	Huffman Tables

Program 2: MP32Wav1B.SPIN

Requires these files:

	frames.tmp
	d.dat

Memory requirements for MP3 playback.

This is the first hurdle to overcome. There are several arrays used in LIBMAD that are much to big to fit into the 32kB of the Prop. Most of them can be reduced by going from 32 to 16 bits. Also, there are other tricks for some arrays. Here's a table of big arrays used in LIBMAD and the minimum size I think they could be for the Prop:

Array Name	Usage	Size in LIBMAD	Min. Size for with Prop
InputBuffer	MP3 file input buffer	40 kB	perhaps 512 B since using flash
OutputBuffer	PCM sample output buffer	8 kB	2.4 kB {word [2*576] half this for mono}
Main_Data	buffer for main data of stream	2.5 kB	~2 kB {maybe a bit less for mono}
xr	decode frequency domain data	9.2 kB	2.4 kB {word [2*576] half this for mono}
rq_table	requantization tables	32.8 kB	1.0 kB {maybe half this for 16 bits?}
Layer3 data		~4 kB	906 B {much less if only supporting 44100 Hz}
Huffman tables	info to uncompress main data	~8 kB	4.4 kB
t	temp storage for DCT32	1 kB	354 B
D	constants for synthesis	2.2 kB	1.1 kB {made word sized with no loss}
pcm_samples	output samples	9.2 kB	0 {use outputBuffer instead}
sideinfo	frame sideinfo storage		234 B
frame_sbsample	buffer for time domain data	9.2 kB	1.2 kB {doing one channel at a time with words}
frame_overlap	time domain overlapping data	4.6 kB	2.3 kB (half for mono)
window_l	windowing constants	144 B	72 B
synth_filter	data for synthesis	4 kB	2 kB (half for mono)

TOTAL		a lot !	21 kB

Here's a look at the big procedure sizes

Function SPIN Size (B)

III_huffdecode 1500

synth_full 1411

dct32 4308

imdct36 2894

total 8.6 kB