Advanced Technologies Division
Channelization and Resampling Using
a Graphics Processing Unit
SAIC
September 23rd, 2008
8/1/2008
1
Overview
Advanced Technologies Division
• Channelization and resampling of digital signals
are common elements in SoftwareDefined/Cognitive Radio Applications.
• However, channelization and resampling
consume a significant portion of the overall
system computation budget
• This can limit the data rates at which the system
can operate.
• SAIC’s polyphase GPU implementation was
typically on the order of 4-10 times faster than
the associated CPU implementation.
8/1/2008
2
Polyphase GPU Implementation
Advanced Technologies Division
Interleave by Q,
Segment into N
blocks
Data
manipulation of
N*Q*(F-L+1)
elements
1,055,872
elements
(3.6% time)
Example:
P=1,Q=64,
L=31,N=73,
F=256
(~0.0276 sec
on 8800GTX
incl file I/O)
8/1/2008
FFT
Multiply with
each 'P-block'
of H(w)
IFFT
(N*Q)
FFTs of
length F
P*N*Q*F
complex
multiplies
(P*N*Q)
IFFTs of
length F
4672
256-point
FFTs
(1.1% time)
1,196, 032
complex
multiplies
(3.4% time)
4672
256-point
FFTs
(1.7% time)
P: Upsample Rate
Q: Downsample Rate
L: (Filter Length / (P * Q))
N: Number of input blocks
to process in parallel
F: FFT length
Rearrange
into
Q channels
Move
P*N*Q*(F-L+1)
elements
1,055,872
elements
(5.3% time)
FFT
(P*N*(F-L+1))
FFTs of
length Q
16,498
64-point
FFTs
(1.6% time)
•FFT length determined by filter length
•This is related to max(P, Q) / min(P, Q)
•Number of parallel input blocks N is
constrained by memory and computing resources
•~83% of time spent in data conversion and file I/O for
this example
3
Implementation Details
Advanced Technologies Division
Host (CPU) implementation:
•Uses fftw floating-point libraries (3.1.2, not multi-threaded)
•Uses parallel dft plans
GPU implementation:
•Uses NVIDIA's CUFFT library for parallel FFTs
•Custom kernels for data manipulations, block multiplies, and floating-point
conversion
•Tests run on 8800 GTX card under RHEL5 (32-bit)
•8800 GTX has 16 multiprocessors
•Maximum active threads = (16*768) = 12,288, each running at 1.35Mhz
General Notes:
•All timing results include reading data from the file
•GPU timing results include transferring data to and from the GPU
•All computations use single-precision IEEE-754 arithmetic
8/1/2008
4
Results
Advanced Technologies Division
GPU implementation was typically on the order of 4-10 times faster than
the associated CPU polyphase implementation across a wide range of
channelization levels.
Channelization Time for 1GB 16-bit Complex
Signed Data File
8/1/2008
Resample Time for 1GB 16-bit Complex,
Signed Data File
5