This repo was created to host my optimized implemetation of a Buddhabrot generator using PyCuda.
- CUDA 9.0
- PyCuda built for CUDA 9.0
numpyscipymatplotlib(only for generating graphs during benchmarking)
Running the code is as simple as doing python3 buddha_kernel.py.
The grainy image was generated in 20ms, the fine image was generated in 1840ms.
The grainy image was generated in 10ms, the fine image was generated in 2490ms.
This is some example output from my laptop (GTX 1050) and my desktop (GTX 1080Ti). In both tests, settings are adjusted to make the execution time about 30 seconds for the actual buddha calculation. However, the important distinction between the two tests is the total number of iterations, or the number of iterations per second. For the GTX 1050, a speed of about 750 Msamples/second is achieved, but for the GTX 1080Ti, a speed of 10 Gsamples/second is achieved.
Mask generated in 24.94 seconds
Formatting...
Total iterations: 2.18473e+10
Iterations per pixel: 5926.45
Maximum frequency: 68115
Minimum frequency: 0
Total time: 29.19s
Iterations per second: 7.48e+08
Mask generated in 14.47 seconds
Formatting...
Total iterations: 3.49555e+11
Iterations per pixel: 94822.95
Maximum frequency: 1086572
Minimum frequency: 0
Total time: 32.98s
Iterations per second: 1.06e+10
There are multiple ways of increasing performance, but a lot of them come with a tradeoff that execution time will be longer. In order to maximize the number of samples per second, the discretization constant should be at least 8192. The thread size should be 128 or 256 (can also depend on your GPU), and the block size should be as large as possible. After this, the repeat variable can be set to linearly scale the execution time while slightly improving performance. With all of this configured, it is possible to squeeze around 40% more performance out of the script. For the benchmark below, the repeat variable was set to 4.
Total iterations: 1.78973e+14
Iterations per pixel: 48549443.94
Maximum frequency: 556520308
Minimum frequency: 0
Total time: 12279.47s
Iterations per second: 1.46e+10
Another factor which impacts performance is the size of the output image. The number of iterations per second drops as the image size increases. For all of these benchmarks, the image size has been fixed at 1440x2560 pixels in order to ensure measurability of performance. The sampling area is also adjusted for this aspect ratio, but could be automatically calculated in a future version.