Sobel Benchmark ...
I am interested in finding out what is the exact test you are using for the Sobel filter benchmark? Is it just a horizontal or vertical filter or does it compute the magnitude and angle as well in fixed point? Can you specify the dimensions of the test image used?
Keyboard shortcuts
Generic
| ? | Show this help |
|---|---|
| ESC | Blurs the current field |
Comment Form
| r | Focus the comment reply box |
|---|---|
| ^ + ↩ | Submit the comment |
You can use Command ⌘ instead of Control ^ on Mac
1 Posted by Anthony on 05 Mar, 2014 02:37 PM
Hi Kalyan,
Thank you for your question. The Sobel workload operates on a 2583×3874 3-channel RGB image for desktop benchmarks and a 1801x2712 3-channel RGB image for mobile benchmarks. The first thing it does is convert the image to greyscale using a weighted combination of the R, G, and B values. This conversion time is included in the execution time of the workload.
For the Sobel operator we compute the components of the gradient using 3x3 spatial filters (one for the x derivative and one for the y derivative). We then approximate the magnitude using |grad| = |dx| + |dy|. We do not compute the direction of the gradient vector and we do not simplify the computation using the fact that the Sobel filter is separable.
In case it is useful in the future, there is a document available that has high-level descriptions of each workload: http://support.primatelabs.com/kb/geekbench/geekbench-3-benchmarks.
Anthony