Wiki » History » Version 9

Chris Cannam, 2013-10-21 05:21 PM

1 1 Chris Cannam
h1. Summary of results
2 6 Chris Cannam
3 6 Chris Cannam
{{>toc}}
4 1 Chris Cannam
5 9 Chris Cannam
h2. What we're looking at
6 9 Chris Cannam
7 9 Chris Cannam
Here we're only looking at causal methods, so no forward/backward filtering. The question in my head is really whether faster IIR filters are still so much faster as to be worth using in preference to linear-phase methods with better (?) theoretical quality. Of course that would always depend on the application, but it's interesting to compare. 
8 9 Chris Cannam
9 1 Chris Cannam
We compared
10 1 Chris Cannam
11 7 Chris Cannam
 * @decimate@: the "Decimator":http://code.soundsoftware.ac.uk/projects/qm-dsp/embedded/classDecimator.html implementation in the "qm-dsp":/projects/qm-dsp library, which uses an 8-tap IIR lowpass filter;
12 1 Chris Cannam
13 2 Chris Cannam
 * @resample_hq@, @resample_mq@, @resample_lq@: the "Resampler":http://code.soundsoftware.ac.uk/projects/qm-dsp/embedded/classResampler.html implementation in the "qm-dsp":/projects/qm-dsp library, which uses a lengthy Kaiser-windowed sinc filter, at three different quality settings;
14 1 Chris Cannam
   
15 1 Chris Cannam
 * @src@: the sndfile-resample program which uses "libsamplerate":http://mega-nerd.com/SRC/, a well trusted resampler also using a Kaiser-windowed sinc implementation, at its default quality setting;
16 1 Chris Cannam
17 1 Chris Cannam
 * @zoh@: the sndfile-resample zero-order hold resampler, which just takes every Nth sample without any filtering, serving as a baseline.
18 3 Chris Cannam
19 3 Chris Cannam
h2. Speed
20 3 Chris Cannam
21 4 Chris Cannam
For 5292000 input frames on a Core i3-3229Y low-voltage CPU. (Frames-per-second values are for input frames.)
22 3 Chris Cannam
23 8 Chris Cannam
All code is 64-bit. The qm-dsp implementations (resample_* and decimate) were compiled with -O3 -ffast-math while the libsamplerate implementations (src and zoh) were standard Ubuntu packages, so probably -O2. This is likely to make a very significant difference, so these results are more useful for comparison among the qm-dsp implementations than between qm-dsp and libsamplerate.
24 1 Chris Cannam
25 5 Chris Cannam
The decimate implementation supports factors up to 8 only, so 16x, 32x and 64x are handled in two passes.
26 5 Chris Cannam
27 4 Chris Cannam
All implementations use libsndfile for audio file I/O, so that should not be a factor in overall speed.
28 4 Chris Cannam
29 4 Chris Cannam
h3. Implementations by decimation factor
30 4 Chris Cannam
31 3 Chris Cannam
h4. Factor 2
32 3 Chris Cannam
33 3 Chris Cannam
|Frames per second|Clock time|Implementation|
34 3 Chris Cannam
|  14225806|0.372|decimate|
35 3 Chris Cannam
|   8939189|0.592|zoh|
36 3 Chris Cannam
|   3732016|1.418|resample_lq|
37 3 Chris Cannam
|   1856842|2.850|resample_mq|
38 1 Chris Cannam
|    989158|5.350|resample_hq|
39 3 Chris Cannam
|    516141|10.253|src|
40 4 Chris Cannam
41 1 Chris Cannam
h4. Factor 4
42 4 Chris Cannam
43 4 Chris Cannam
|Frames per second|Clock time|Implementation|
44 4 Chris Cannam
|17070967|0.310|zoh|
45 4 Chris Cannam
|14659279|0.361|decimate|
46 4 Chris Cannam
| 4285020|1.235|resample_lq|
47 4 Chris Cannam
| 2186776|2.420|resample_mq|
48 4 Chris Cannam
| 1056287|5.010|resample_hq|
49 4 Chris Cannam
|  610099|8.674|src|
50 4 Chris Cannam
51 1 Chris Cannam
h4. Factor 8
52 4 Chris Cannam
53 4 Chris Cannam
|Frames per second|Clock time|Implementation|
54 4 Chris Cannam
|26328358|0.201|zoh|
55 4 Chris Cannam
|13926315|0.380|decimate|
56 4 Chris Cannam
| 4895467|1.081|resample_lq|
57 4 Chris Cannam
| 2470588|2.142|resample_mq|
58 4 Chris Cannam
| 1166409|4.537|resample_hq|
59 4 Chris Cannam
|  614919|8.606|src|
60 4 Chris Cannam
61 1 Chris Cannam
h4. Factor 16
62 4 Chris Cannam
63 4 Chris Cannam
|Frames per second|Clock time|Implementation|
64 4 Chris Cannam
|33493670|0.158|zoh|
65 4 Chris Cannam
|12721153|0.416|decimate|
66 4 Chris Cannam
| 5093358|1.039|resample_lq|
67 4 Chris Cannam
| 2515209|2.104|resample_mq|
68 4 Chris Cannam
| 1182041|4.477|resample_hq|
69 4 Chris Cannam
|  668857|7.912|src|
70 4 Chris Cannam
71 1 Chris Cannam
h4. Factor 32
72 4 Chris Cannam
73 4 Chris Cannam
|Frames per second|Clock time|Implementation|
74 4 Chris Cannam
|41669291|0.127|zoh|
75 4 Chris Cannam
|14498630|0.365|decimate|
76 4 Chris Cannam
| 5318592|0.995|resample_lq|
77 4 Chris Cannam
| 2312937|2.288|resample_mq|
78 4 Chris Cannam
| 1148936|4.606|resample_hq|
79 4 Chris Cannam
|  670467|7.893|src|
80 4 Chris Cannam
81 1 Chris Cannam
h4. Factor 64
82 4 Chris Cannam
83 4 Chris Cannam
|Frames per second|Clock time|Implementation|
84 4 Chris Cannam
|42000000|0.126|zoh|
85 4 Chris Cannam
|13397468|0.395|decimate|
86 4 Chris Cannam
| 5040000|1.050|resample_lq|
87 4 Chris Cannam
| 2365668|2.237|resample_mq|
88 4 Chris Cannam
| 1232704|4.293|resample_hq|
89 1 Chris Cannam
|  636057|8.320|src|
90 5 Chris Cannam
91 5 Chris Cannam
h3. Decimation factors by implementation
92 4 Chris Cannam
93 4 Chris Cannam
h4. Implementation zoh
94 4 Chris Cannam
95 4 Chris Cannam
|Frames per second|Clock time|Factor|
96 4 Chris Cannam
|42000000|0.126|factor 64|
97 4 Chris Cannam
|41669291|0.127|factor 32|
98 4 Chris Cannam
|33493670|0.158|factor 16|
99 4 Chris Cannam
|26328358|0.201|factor 8|
100 4 Chris Cannam
|17070967|0.310|factor 4|
101 4 Chris Cannam
| 8939189|0.592|factor 2|
102 4 Chris Cannam
103 4 Chris Cannam
h4. Implementation decimate
104 4 Chris Cannam
105 4 Chris Cannam
|Frames per second|Clock time|Factor|
106 4 Chris Cannam
|14659279|0.361|factor 4|
107 4 Chris Cannam
|14498630|0.365|factor 32|
108 4 Chris Cannam
|14225806|0.372|factor 2|
109 4 Chris Cannam
|13926315|0.380|factor 8|
110 4 Chris Cannam
|13397468|0.395|factor 64|
111 4 Chris Cannam
|12721153|0.416|factor 16|
112 4 Chris Cannam
113 4 Chris Cannam
h4. Implementation resample_hq
114 4 Chris Cannam
115 4 Chris Cannam
|Frames per second|Clock time|Factor|
116 4 Chris Cannam
| 1232704|4.293|factor 64|
117 4 Chris Cannam
| 1182041|4.477|factor 16|
118 4 Chris Cannam
| 1166409|4.537|factor 8|
119 4 Chris Cannam
| 1148936|4.606|factor 32|
120 4 Chris Cannam
| 1056287|5.010|factor 4|
121 4 Chris Cannam
|  989158|5.350|factor 2|
122 4 Chris Cannam
123 4 Chris Cannam
h4. Implementation resample_mq
124 4 Chris Cannam
125 4 Chris Cannam
|Frames per second|Clock time|Factor|
126 4 Chris Cannam
| 2515209|2.104|factor 16|
127 4 Chris Cannam
| 2470588|2.142|factor 8|
128 4 Chris Cannam
| 2365668|2.237|factor 64|
129 4 Chris Cannam
| 2312937|2.288|factor 32|
130 4 Chris Cannam
| 2186776|2.420|factor 4|
131 4 Chris Cannam
| 1856842|2.850|factor 2|
132 4 Chris Cannam
133 4 Chris Cannam
h4. Implementation resample_lq
134 4 Chris Cannam
135 4 Chris Cannam
|Frames per second|Clock time|Factor|
136 4 Chris Cannam
| 5318592|0.995|factor 32|
137 4 Chris Cannam
| 5093358|1.039|factor 16|
138 4 Chris Cannam
| 5040000|1.050|factor 64|
139 4 Chris Cannam
| 4895467|1.081|factor 8|
140 4 Chris Cannam
| 4285020|1.235|factor 4|
141 4 Chris Cannam
| 3732016|1.418|factor 2|
142 4 Chris Cannam
143 4 Chris Cannam
h4. Implementation src
144 4 Chris Cannam
145 4 Chris Cannam
|Frames per second|Clock time|Factor|
146 4 Chris Cannam
|  670467|7.893|factor 32|
147 4 Chris Cannam
|  668857|7.912|factor 16|
148 4 Chris Cannam
|  636057|8.320|factor 64|
149 4 Chris Cannam
|  614919|8.606|factor 8|
150 4 Chris Cannam
|  610099|8.674|factor 4|
151 4 Chris Cannam
|  516141|10.253|factor 2|