CUDA cuFFT vs FFTW

My initial thoughts were that I was going to perform FFT on the GPU as it intuitively sounds fast to use the multi-core system that the GPU provides. However, depending on the size of the FFT, GPU based algorithms may actually be outperformed by running FFT on the CPU.

Transferring memory-buffers to the GPU is a slow operation, and the overhead of transferring data to the GPU does affect the overall performance of running FFT particularly. When using CUDA cuFFT compared to serial FFTW, it has been shown that for small N, approximately N <= 4096, the gain from running cuFFT is lost due to slow memory transfer rates.


From lecture: "Fast Fourier Transforms (FFTs) and Graphical Processing Units (GPUs)" - Kate Despain, University of Maryland Institute for Advanced Computer Studies. Originally from University of Waterloo (2007). Comparing FFT and cuFFT including and excluding memory transfer rates. Y axis labeling flops, X axis labeling size of FFT work set. 

In my project, I'm running 2D FFT since I'm simulating a water surface, so 4096 elements would equal to a grid of size 64x64, which is a rather small surface. In the original research paper by Jerry Tessendorf, Simulating Ocean Water, it is stated that: "The values of N and M can be between 16 and 2048, in powers of two. For many situations, values in the range 128 to 512 are sufficient.", and it is also concluded that for interactive performance, a CPU of +1 GHz could should nearly suffice for a 512x512 sized grid. 

With this in mind, I think I will be able to still use cuFFT and achieve reasonable performance, as 512x512 >> 4096. However, if time allows, it would be interesting to compare the performance of serial FFTW and cuFFT for different sizes of grids. However, my initial expectation is still that cuFFT is the way to go.

Kommentarer

Skicka en kommentar

Populära inlägg i den här bloggen

Dear ImGui - Interactive settings for ocean water

Bild
The last crucial piece of the project implementation relates to GUI. It is tedious having to restart the application for testing different settings, and I wanted to be able to change the settings of the ocean in real time. Using a GUI I wanted the following settings to be modifiable in real time. For ocean: Wave length Amplitude Wind speed Wind direction Normal displacement amount  Vertex count Tile count For application: Camera movement speed Camera rotation speed Enable/disable wireframe mode Enable/disable skybox Enable/disable skybox lighting Selecting skybox cubemap A very convenient and handy tool for debugging graphics applications is Dear ImGui , or simply ImGui. Integrating ImGui is simple and it has an implementation agnostic interface for whichever window environment and graphics library you are using. When the ImGui library is linked properly to the application, dialogs and GUI utils are easily drawn to the screen by initializing an ImGui context. When a context is runn...
Läs mer

Gradients and Choppy waves.

Bild
As mentioned in the last post, the focus of the last days was to fix proper normal calculations using the gradient of the height-field. Gradient, together with creating a displacement in x and z-direction, only required setting up a four more cuda-buffers and performing an inverse FFT-pass on each of them. Calculating the gradient and the x-z displacement is easy, and comes from equation 20 and 29 in Tessendorf's paper, which you should be familiar with now. The input for calculating the x-gradient and z-gradient simply meant multiplying the input for the displacement map with the imaginary number ik, for each k where k is the propagation direction of the wave at each point. Formula for gradient calculation for realistic surface normals (Tessendorf 2001) Left: z-gradient normalized to range 0-1. Right:  z-gradient normalized to range 0-1 One problem that occurred when calculating the surface normals using the gradient values was that the result of the FFT is naturally rather small....
Läs mer