OpenVX Kernel Benchmark

Mainly includes 44 standard kernels of OpenVX 1.2, run each kernel independently to verify and understand how each kernel works.

It also includes some kernels with tiling feature, which are included in the directory of tiling.

Prerequisite

To run the openvx kernel benchmark programs, you need to set up openvx environment first. There is an easy way to build openvx environment by following this.

Usage

After the envoriment is ready, download the files from this repo. We can use Makefile to create executable files and run each of them.

$ cd openvx-kernels

$ make

$ cd ./bin_kernels

$ ./gaussian (or other executable files)

Note: Since we are trying to run only one kernel node once, some of them need to use the output of other kernel running results as input. So there are some running order for certain kernels:

1. Before run 'absdiff', run 'lut' first.
2. Before run 'add, sub, multiply, bitwise_and, bitwise_or, bitwise_xor', run 'absdiff' first.
3. Before run 'erode, dilate', run 'thresh' first.

Others

4. 'canny_1, canny_2 ' with different thresholds, 'VX_NORM_L1' for 'canny_1', and 'VX_NORM_L2' for 'canny_2 '. Mainly to observe the difference canny edge detection results.
5. 'multiply, remap' have different policies for setting parameters, we just choose one of them to run a single node. Differet options have been implemented in 'multiply.c, remap.c' and need to be commented out if needed.
6. 'absdiff_1.c, multiply_1.c, channel_combine_1.c' are used to verify the results of 'absdiff, multiply, channel_combine'. They have been listed in Makefile but been commented, so if you want to run them, comment out them in Makefile to generate excutable files.

Tiling

The above are basic kernels without using tiling. In the dir of tiling, we have implemented kernels with tiling. In order to run the kernels:

$ cd tiling

$ make

$ cd ./bin_tiling

$ ./tiling_gaussian (or other executable files)

To get perf stat when run the kernels:

$ ./run_kernels.sh ./tiling_gaussian 4x4 (./run_kernel.sh + executable kernel files + tile block size)

After you get all the tiled and non-tiled perf output (.txt) files, you can extract the perf data:

$ python extract_perf_data.py (use python3 if needed)

Note:

1. Before run 'tiling_add', run 'tiling_alpha' first.
2. Before run 'non_tiling_add', run 'non_tiling_alpha' first.
3. We put part of non_tiling kernels in tiling dir, mainly for comparison.
4. If you want to change tile blocksize, you need to change the origin vx_tiling_ext.c file in 'openvx_sample'(what we built at the beginning) and then rebuild by using (in /openvx_sample/) python Build.py --os=Linux --tiling --openmp --conf=Debug --rebuild=True --c=gcc --cpp=g++
5. If you want to print out the block size, what we do now is add some printf lines in original tiling definition. You can use files in this repo: vx_tiling_add.c, vx_tiling_alpha.c, vx_tiling_box.c, and vx_tiling_gaussian.c (in /openvx-kernels/tiling/examples_tiling) to replace these files in /openvx_sample/examples to rebuild the library of tiling. Same way to rebuild the openvx(python Build.py --os=Linux --tiling --openmp --conf=Debug --rebuild=True --c=gcc --cpp=g++ ). You can do these steps(4&5) together and then rebuild together or rebuild separately. Btw, you can also back up files before being replaced if needed.
6. When use ./run_kernels.sh, for non-tiled kernel, you can use '0x0' for parameter of block size or just skip it. The parameter of execution file, add './' in front of the filename.

Credits: OpenVX (https://www.khronos.org/openvx/)