From 1e9796608a05b0c824ad03a6845832b9ece1edf1 Mon Sep 17 00:00:00 2001 From: David Phillips Date: Wed, 22 Feb 2017 17:19:10 +1300 Subject: Add quick platform fix to demo image script --- README.md | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) (limited to 'README.md') diff --git a/README.md b/README.md index 7c1d41f..1e7601e 100644 --- a/README.md +++ b/README.md @@ -2,14 +2,15 @@ This is a work-in-progress port of my simple fractal-gen software to OpenCL. That software was an experiment of mine to generate mandelbrot (and some of -its cousin) fractals on CPU. This is my attempt at porting that software to -OpenCL so it can be used on a multitude of computation devices, including GPUs. +its cousin) fractals on the CPU. This is my attempt at porting that software to +OpenCL so it can be used on a multitude of computation devices, chiefly GPUs. I had started to port it to CUDA in October 2016, but changed to CL because of its portability and open nature. -Software is still in early days and needs more CL kernels for such fractals as -tricorn, burning ship, and julia sets to name a few. +The software is working with basic mandelbrot fractals, but needs more CL +kernels for such fractals as tricorn, burning ship, and julia sets to name a +few. Below is a simple demo image produced with the software. It is a simple mandelbrot fractal using only 75 iterations, at 768 square pixels. The @@ -21,9 +22,9 @@ Such a small, low-detail image will not provide a case for using GPU rather than CPU, but once you start upping the image size and detail, a modern GPU will provide endless benefit over a modern CPU. -For example, using the CPU-based fractal-gen, a 10240x10240 pixel image at +For example, using the CPU-based fractal-gen, a 10240x10240 pixel image at a 10000 iteration cutout per pixel, the image will complete in about 2200 seconds when running on all 32 threads of a dual-Xeon E5-2670 setup. Compare this to the runtime of this software on a (much cheaper) NVIDIA GTX 1070; 1 second. - -This is using single-precision floats on the GPU. +This is using single-precision floats on the GPU, but changing to `double` data +types only slows the 1070 to about 8 seconds. -- cgit v1.1