Warning
I've stopped maintaining this project !
Parallel Eigen value(s)/ vector(s) Computation on Accelerator
This is an implementation of maximum eigen value( with respective vector ) computation using Similarity Transform method for positive matrices.
Read paper for better understanding.
I keep both sequential & parallel implementation.
- You should have Intel oneAPI basekit installed, which you can find here.
- Or you may want to build it from source, check here.
- I'm working on a GNU/ Linux machine, with
make,clang-formatinstalled.
$ lsb_release -d Description: Ubuntu 20.04.3 LTS- My
dpcppcompiler version is
$ dpcpp --version Intel(R) oneAPI DPC++/C++ Compiler 2021.4.0 (2021.4.0.20210924) Target: x86_64-unknown-linux-gnu Thread model: posix InstalledDir: /opt/intel/oneapi/compiler/2021.4.0/linux/bin- Assuming you've all requirements installed on your machine, compile program
make # or make aot_{cpu,gpu}- Run produced binary
./run # or ./a.outClean up generated object files using
make clean
After editing source, you can reformat those using
make format
I ran parallel implementation of similarity transform algorithm on multiple hardwares, with hilbert matrix of various dimensions, while setting maximum iteration count to 1000 and dynamic work group size.
Except following ones, I keep detailed benchmark results for all kernels involved in similarity transform method here.
$ make aot_cpu && ./a.out running on Intel(R) Core(TM) i9-10920X CPU @ 3.50GHz Parallel Algorithm using Similarity Transform for finding max eigen value (with vector) 128 x 128 7 ms 9 round(s) 256 x 256 1 ms 10 round(s) 512 x 512 3 ms 12 round(s) 1024 x 1024 5 ms 13 round(s) 2048 x 2048 18 ms 14 round(s) 4096 x 4096 101 ms 15 round(s) 8192 x 8192 510 ms 17 round(s)$ make aot_cpu && ./a.out running on Intel(R) Xeon(R) Platinum 8358 CPU @ 2.60GHz Parallel Algorithm using Similarity Transform for finding max eigen value (with vector) 128 x 128 22 ms 9 round(s) 256 x 256 6 ms 10 round(s) 512 x 512 7 ms 12 round(s) 1024 x 1024 8 ms 13 round(s) 2048 x 2048 19 ms 14 round(s) 4096 x 4096 40 ms 15 round(s) 8192 x 8192 126 ms 17 round(s)$ make aot_cpu && ./a.out running on Intel(R) Xeon(R) CPU E5-2686 v4 @ 2.30GHz Parallel Algorithm using Similarity Transform for finding max eigen value (with vector) 128 x 128 11 ms 9 round(s) 256 x 256 6 ms 10 round(s) 512 x 512 16 ms 12 round(s) 1024 x 1024 51 ms 13 round(s) 2048 x 2048 200 ms 14 round(s) 4096 x 4096 847 ms 15 round(s) 8192 x 8192 3759 ms 17 round(s)$ make aot_cpu && ./a.out running on Intel(R) Xeon(R) Gold 6128 CPU @ 3.40GHz Parallel Algorithm using Similarity Transform for finding max eigen value (with vector) 128 x 128 18 ms 9 round(s) 256 x 256 4 ms 10 round(s) 512 x 512 6 ms 12 round(s) 1024 x 1024 8 ms 13 round(s) 2048 x 2048 27 ms 14 round(s) 4096 x 4096 82 ms 15 round(s) 8192 x 8192 339 ms 17 round(s)$ make aot_gpu && ./a.out running on Intel(R) Iris(R) Xe MAX Graphics [0x4905] Parallel Algorithm using Similarity Transform for finding max eigen value (with vector) 128 x 128 10 ms 9 round(s) 256 x 256 8 ms 10 round(s) 512 x 512 14 ms 12 round(s) 1024 x 1024 37 ms 13 round(s) 2048 x 2048 138 ms 14 round(s) 4096 x 4096 564 ms 15 round(s) 8192 x 8192 2509 ms 17 round(s)$ make && ./run # JIT compiled kernels, check time column of first row of below table [ way more than second row, because kernel being JIT-ed ] running on Intel(R) UHD Graphics P630 [0x3e96] Parallel Algorithm using Similarity Transform for finding max eigen value (with vector) 128 x 128 122 ms 9 round(s) 256 x 256 9 ms 10 round(s) 512 x 512 28 ms 12 round(s) 1024 x 1024 104 ms 13 round(s) 2048 x 2048 446 ms 14 round(s) 4096 x 4096 1858 ms 15 round(s) 8192 x 8192 8259 ms 17 round(s)A set of minimal test cases are present, which you may want to run using
make testRun sequential implementation using 👇, while ensuring you've numpy installed. It also compares computed maximum eigen value with numpy computed result.
python3 -m pip install -U numpy # you may need it python3 main.pySequential Similarity Transform, for finding maximum eigen value ( with vector ) 32 x 32 1.27 ms 5 round(s) 64 x 64 2.44 ms 5 round(s) 128 x 128 9.67 ms 5 round(s) 256 x 256 20.58 ms 4 round(s) 512 x 512 74.24 ms 4 round(s) 1024 x 1024 335.77 ms 4 round(s) I provide you with one build recipe which can be used for compiling Parallel Similarity Transform's implementation into dynamically linked shared object.
# AOT compiled for x86_64 CPU; check Makefile make lib # now check file wrapper/libsimilarity_transform.soOnce shared object is ready, you can now interact with C++ implementation of Maximum EigenValue Finder function from Python wrapper function.
pushd wrapper/python python3import similarity_transform as st # must be positive square matrix of `float32` # # this is input data matrix m = st.np.random.random((16, 16)).astype('f') ev = st.EigenValue() λ, v, ts, itr = ev.similarity_transform(m) # λ = maximum eigen value # v = eigen vector # ts = execution time ( in ms ) # itr = iterations required before convergenceYou may want to take a look at test case written using Python wrapper.
There's also one script for running tests on randomly generated positive square matrices.
python3 test.py popd