Hardware Reference
In-Depth Information
Stream #1
Stream #2
Start
FE-GA #0
CPU #0
CPU #1
FE-GA #1
DTU0 : Off-
CSM-URAM
DTU1 : Off-
CSM-URAM
DTU0 : URAM
to CRAM
DTU1 : URAM
to CRAM
Frame read
Frame read
Filter bank
& M/S Stereo
Filter bank
& M/S Stereo
DTU0 :CRAM
to URAM
DTU1 : CRAM
to URAM
DTU0 : URAM
to CRAM
DTU1 : URAM
to CRAM
Target bit rate calc.
Target bit rate calc.
DTU0 : CRAM
to URAM
DTU1 : CRAM
to URAM
Quantization
Quantization
Huffman coding
Huffman coding
Bit rate adjustment
Bit rate adjustment
N
N
Less than speci-
fied bit rate?
Less than speci-
fied bit rate?
DTU1 : URAM
to CRAM
DTU0 : URAM
to CRAM
Y
Y
Stream generation
Stream generation
DTU0 : URAM
to Off-CSM
DTU1 : URAM
to Off-CSM
Processing flow (data transfer)
Fig. 6.7
Heterogeneous parallelization of AAC encoding
1.8
60
x54.1
Encoding speed
Power consumption
1.5
50
1.46
1.17
1.36
1.2
40
1.22
0.9
30
x27.1
20
0.6
10
x8.0
0.3
x4.0
0
0.0
CPUx1
[1 stream]
CPUx2
[2 streams]
CPUx1+FEx1
[1 stream]
CPUx2+FEx2
[2 streams]
3.4
xEnc/W
5.8
xEnc/W
22.2
xEnc/W
37.1
xEnc/W
Encoding
speed / W
Fig. 6.8
Performance and power results with various con fi gurations
input stream was encoded twice on one CPU and one FE-GA, and the two streams
of the same input music were encoded simultaneously on two CPUs and two
FE-GAs. Input PCM and output AAC stream data were placed in the off-chip main
memory. The DTU transferred data by using transfer lists.
Figure 6.8 plots the evaluation results. The speedup was 4.0 and the average
power consumption was 1.17 W with encoding on a single CPU. The encoding
 
Search WWH ::




Custom Search