Developement and Implementation of an MPEG1 Layer III Decoder on x86 and TMS320C6711 platforms Farina Simone (Braidotti Enrico)

DECODING PROCESS

ALIAS RECONSTRUCTION It is performed only when using long blocks: this means only when using pure long blocks or mixed blocks. Lets see what long/short blocks are Not encoded signal Same signal encoded using long blocks Same signal encoded using short blocks

HYBRID SYNTHESIS IMDCT (Inverse Modified Discrete Cosine Transform) Subbands are backward transformed separately depending on block length. 6-point IMDCT When short blocks are used (pre-echoes masking) 18-point IMDCT When long blocks are used

HYBRID SYNTHESIS Fast IMDCT algorithm (Szu-Wei Lee ) Based on simmetric properties of cosine function It needs a rearranging stage to restore values to their original positions Drastically reduces number of operations if compared to direct implementation × (short/long)+ (short/long) Direct Implementation216 / / 612 Fast IMDCT (Szu-Wei Lee)33 / 4369 / 115 Improvement84.7 % / 93.3 %61.7 % / 81.2 %

HYBRID SYNTHESIS Windowing Once transformed, subbands are windowed according to value of block_type (subbands with short blocks are separately transformed for each window and then overlapped) Overlap-adding First half of transformed blocks is overlapped with second half of the corresponding blocks in the previous granule

FREQUENCY INVERSION Every second sample in every second subband has to be multiplied by -1.

SYNTHESIS POLYPHASE FILTERBANK This process produces 32 PCM audio samples. 576 / granule 1152 / frame (equal to 26 ms of 44,1 kHz) Composed of several steps, it turns out to be the most time-consuming stage of the overall decoding process

SYNTHESIS POLYPHASE FILTERBANK Polyphase Matrixing It is a cosine-like transform (non standard ) The direct computation involves a 64×32 matrix and requires almost ¼ of decoding time Needs optimization to perform real-time decoding K. Konstantinides’ algorithm 32-point Fast DCT (B.G.Lee)

SYNTHESIS POLYPHASE FILTERBANK Konstantinides’ Algorithm

SYNTHESIS POLYPHASE FILTERBANK FCT Algorithm (Byeong-Gi Lee ) Using trigonometrical properties a 2 M DCT can be performed by 2 M-1 2-point DCTs Direct computation × = N ² + = N· ( N-1 ) FCT × = N/2 · log 2 ( N ) + < 3· N/2 · log 2 ( N )

WAVE STANDARD Individuated by a 44-byte header, holds information about: sampling frequency number of channels... Uncompressed PCM audio samples (normally with 16 bits/sample resolution) stored in following way: Istante di campionamento Canale 0 1 (Left) 2 (Right) 1 1 (Left) 2 (Right) 2 1 (Left) 2 (Right)

PERFORMANCE ANALYSIS PC Performances The decoder, without optimization, works in real time on the following CPUs: The decoder, with optimization, reaches 17,5× on Pentium IV CPU

PERFORMANCE ANALYSIS C6711 DSK Performances The decoder, without optimization, doesn’t work in real time on the board Parallel port is used for data transfer and it’s very very slow Most algorithms need optimization (only Huffman Decoding is optimized) Code needs some ASM optimization to use the full-potential of the board architecture Whole decoding process (except data transfer TO external hard disk) takes about 10 times more than needed to work in real time. With optimization it is an easy goal to reach.

PERFORMANCE ANALYSIS Time-occupation of optimized processes on C6711 DSK:

PERFORMANCE ANALYSIS Time-occupation of optimized processes on C6711 DSK:

PERFORMANCE ANALYSIS Time-occupation of other processes on C6711 DSK: