<< Chapter < Page Chapter >> Page >
This module contains the full compilation of testing results for the filter bank project.

Testing parameters

All data was collected on a laptop equipped with an AMD A6-3400M “Llano” quad-core processor which supports a clock rate of up to 2.3 GHz.

Test Parameters
Test Filter 4-pole Butterworth bandpass filter
Input channels 256
Time samples 600,000
Data filter cycles 100
Compiler GCC

The following results tables show the individual parameters of experiments, averaged run times of the program and an indicator of the real-time processing speed of the program (the formula to generate this figure is shown in the equation below).

F s N t = F s t N

where F s = sampling rate of incoming data (samples/sec), N = number of samples processed by filter bank (samples), and t = time to process all samples (sec).

Comparison of optimizations using compiler flags and intrinsics

Unique Filter Coefficients
Optimization Time (sec) secs/sec (25 KS/s)
None 619.940 0.25831
O3 Compiler 241.131 0.10047
O3 and SSE3 75.589 0.03150
O3 and our SSE3 69.258 0.02886
Constant Filter Coefficients
O3 and SSE3 55.835 0.02326
O3 and our SSE3 49.271 0.02053

Implementations of posix threads

4 tests under various PThread control conditions demonstrate similar patterns according to thenumber of PThreads used during code execution. Compiler Intrinsics : Implementation with naive data structures relying purely on compiler optimizations of SSE3 intrinsics Our Intrinsics : Implementation with naive data structures using our custom SSE3 intrinsics code Partial Reordering : Implementation assigning an output vector to each thread utilizing compiler intrinsics Full Reordering : Intermediate variables are aligned by channel and separated by thread
POSIX Threads
Optimization Time (sec) secs/sec (25 KS/s)
0 Threads and our SSE3 48.333 0.02014
1 Thread and our SSE3 50.109 0.02088
2 Threads and our SSE3 88.632 0.03693
4 Threads and our SSE3 138.090 0.05754
8 Threads and our SSE3 62.481 0.02603
16 Threads and our SSE3 103.901 0.04329
32 Threads and our SSE3 78.219 0.03259
0 Threads and SSE3 48.075 0.02003
1 Thread and SSE3 64.302 0.02679
2 Threads and SSE3 96.755 0.04031
4 Threads and SSE3 123.931 0.05164
8 Threads and SSE3 67.629 0.02818
16 Threads and SSE3 141.329 0.05889
32 Threads and SSE3 121.134 0.05047
Reordered Output Data
Optimization Time (sec) secs/sec (25 KS/s)
0 Threads and SSE3 50.106 0.02088
1 Thread and SSE3 49.302 0.02054
2 Threads and SSE3 78.888 0.02054
4 Threads and SSE3 89.939 0.03747
8 Threads and SSE3 35.283 0.01470
16 Threads and SSE3 71.337 0.02972
32 Threads and SSE3 109.112 0.04546
Reordered Intermediate Variables
0 Threads and SSE3 71.166 0.02965
1 Thread and SSE3 57.156 0.02382
2 Threads and SSE3 52.639 0.02193
4 Threads and SSE3 48.939 0.02039
8 Threads and SSE3 33.589 0.01400
16 Threads and SSE3 51.543 0.02148
32 Threads and SSE3 110.716 0.04613

Comparison of initial optimizations to posix thread implementation

Both tests show code execution times based on different methods of compiler optimization. Test 1 allowed for unique filter coefficients and test 2 had constant coefficients. Both tests used inefficient data arrangement.

Our optimal filter design incorporated a combination of several of the methods we used to optimize our filter bank implementation. It made use of compiler-level optimization, SSE instructions, and POSIX threads. It processed 60 million samples for each of the 256 channels in 33.589 seconds. Assuming a 25 KHz sampling rate, each second of of data is processed in 14 milliseconds (25 KS/s) and thus acceptable for real-time processing.

Get Jobilize Job Search Mobile App in your pocket Now!

Get it on Google Play Download on the App Store Now




Source:  OpenStax, Efficient real-time filter design for recording multichannel neural activity. OpenStax CNX. Dec 11, 2012 Download for free at http://cnx.org/content/col11461/1.1
Google Play and the Google Play logo are trademarks of Google Inc.

Notification Switch

Would you like to follow the 'Efficient real-time filter design for recording multichannel neural activity' conversation and receive update notifications?

Ask