http://www.k-wave.org/forum/topic/cuda-replacement-libraries-for-fftw-and-blas Support for the k-Wave MATLAB toolbox en-US Tue, 12 May 2026 23:28:20 +0000 http://bbpress.org/?v=1.0.2 q http://www.k-wave.org/forum/search.php http://www.k-wave.org/forum/topic/cuda-replacement-libraries-for-fftw-and-blas#post-4581 Wed, 18 Jun 2014 21:08:29 +0000 marcelr 4581@http://www.k-wave.org/forum/ <p>Hi Jiri,</p> <p>Thanks for the update. I have the exact same configuration CPU/GPU you tested on. No big speed-ups to be reckoned with, then. I would love to get the patch file anyway, might learn a thing or two.</p> <p>Good to hear that a GPU-only version of k-Wave is in the making. Looking forward to it. </p> <p>Marcel </p> http://www.k-wave.org/forum/topic/cuda-replacement-libraries-for-fftw-and-blas#post-4580 Tue, 17 Jun 2014 20:26:42 +0000 Jiri Jaros 4580@http://www.k-wave.org/forum/ <p>Hi marcelr,</p> <p>I've performed two simple benchmarks and the results are not surprising :-)<br /> <pre><code>size iCore7 920 (4c/8T) GTX 580 2xIntel E5-2670 (16T) 128^3 0.09918s/step 0.0967s/step 0.025s/step 256^3 1.10s/step 0.7347s/step 0.42s/step</code></pre> <p>Offloading the FFT on the GPU is not going to speed up k-Wave substantially because the data has to be moved back and forth 14 times per time step. This virtually kills all the benefit gained from the GPU. The only way to get a reasonable speed-up is to run k-Wave entirely on the GPU (this version is being tested).</p> <p>The maximum size of the simulation domain I was able to run with a GTX580/1.5GB VRAM was 256^3 (cufft is incredibly memory voracious).</p> <p>However, if you have a slow CPU and a powerful GPU, the speed-up is measurable. If interested, I can send you a patch file to be applied before compiling.</p> <p>Jiri </p> http://www.k-wave.org/forum/topic/cuda-replacement-libraries-for-fftw-and-blas#post-4576 Mon, 16 Jun 2014 08:46:49 +0000 marcelr 4576@http://www.k-wave.org/forum/ <p>In their recent cuda tookits, NVidia have published plugin replacement libs for FFTW and BLAS. I want to have a look at these libs, see if they are any good in speeding up things related to k-Wave. I wonder if forum members have experience with those, in particular in combination with k-Wave?</p> <p>best,</p> <p>Marcel </p>