NI LabVIEW GPU Demo - Multi-channel FFT.AVI скачать в хорошем качестве

NI LabVIEW GPU Demo - Multi-channel FFT.AVI

=Overview= This demo acquires a live audio source, computes the spectral information (Complex FFT) for multiple audio channels simultaneously on an NVIDIA GPU and displays spectrum results on the host computer. =Details= The application is acquiring one audio channel at 50KHz -- 1K in size -- and stores the channel in a history buffer containing 1024 signals. The channel is acquired into a double-precision array. Before being stored in the history buffer, the signal is converted to complex single precision where the double-precision values are truncated and stored in the single-precision real component of the complex data. This means the original signal size and the converted download signal are the same in bytes. At every process iteration, the entire history buffer is transferred to the GPU. A multi-channel FFT implementation based on the CUFFT library (v2.2) is invoked which performs a complex FFT (single-precision components) on all (1024) 1K signals. The results from all channels are transferred back to the host for display. On the front panel, we display the most current acquired channel and it's spectrum tothe right. The intensity graph shows the results for all spectrums in the history buffer. To improve performance, we actually use a chart to avoid redraw overhead of information that does not change. The application maintained a loop time around 20ms which showed we were able to keep up with acquisition. Because of this, limited benchmarking was done beyond verifying the accuracy of the results. It appears as if the GPU FFT computations were responsible for less that 5% of the overall time -- perhaps less. =Purpose= The make-up of this demo may seem odd at first. Why only acquire a single channel then recompute spectrums for other channels? The idea was to provide a demo that could satisfy two goals. 1. Minimize the hardware requirements for the demo to be used away from NI. 2. Show the computing power of the GPU in live acquisition. Hence, I used USB DAQ for acquisition while still throwing the kitchen sink at the GPU from a computational standpoint. Because we've had requests from customers regarding high-performance DSP computations on GPUs from within a LabVIEW, the demo is capable of plugging in new GPU functions for common operations such as Windowing or THD (peak/harmonic detection). I hope to add these types of functions to the demo once I can map them to a real customer application. =Components= Here's some specs on development system. • LabVIEW 2009 • NILabs LabVIEW GPU Computing Module (v1.0) • NI USB-9162 with NI cRIO-9233 • After-market computer (home system) Intel Core2Duo E8400 (3.0GHz, 6MB Cache, 1333MHz FSB) NVIDIA 8800GTS-512M (display) NVIDIA GTX280 (all CUDA computations) 2G RAM (1066MHz) • Windows XP SP3 (32-bit) • MS Visual Studio 2005 (VC8) • Multi-channel FFT CUDA interface for LabVIEW (unreleased) The demo video was captured on the same system using a Canon PowerShot TX1. The video source was from an iPod Touch playing an MP4 converted copy of the first Harry Potter film. A cable from the iPod was connected from the headphone jack to AI0 on the cRIO. NOTE: The audio in the playback from the video was acquired from computer speakers and it matches the current signal displayed on the front panel. =Developers= • Darren Schmidt • Michael Chen • Strong Yao • Michael Cerna