Non-negative Matrix Factorization for DPC++

Non-negative Matrix Factorization algorithm implemented in DPC++, CUDA and OpenMP

Requirements

To compile and launch the code you need to install oneAPI base Toolkit. To run on Nvidia GPUs you have to install the intel open compiler as well as build the oneMKL toolchain.

How to compile it?

To compile the code you can simply do it with make command. However, the Makefile offers the following compilation options:

• OPTIMIZE=[yes, no] --> Enables compiler -O3 option. Activated by default.
• DEBUG=[yes, no] --> Sets the -g compiler option. Disabled by default.
• DEVICE=[cpu, nvidia, igpu] --> States where to run the SYCL kernels. It is set by default to cpu.
• REAL=[simple, double] --> Determines whether to use simple or double float point precision. It is set by default to simple.

How to run it?

Once you copile it, you will get the nmf executable. This binary takes the following parameters:

• matrix binary sample used to initialize the V matrix.
• N --> An unsigned integer which defines the N dimension of the matrix.
• M --> An unsigned integer which defines the M dimension of the matrix.
• K --> It is an unsigned integer, and states the factorization parameter of the matrix.
• Tests --> The number of tests to perform.
• Stop threshold --> The NMF stop threshold.

An example to run it could be:

./nmf V.bin 1000 500 4 1 10

Running on a hyper-threding CPU could reduce performance, to disable it you can use the following variable set to the number of phisical cores on the CPU.

export DPCPP_CPU_NUM_CUS=12

To sufficient splitting to balance load, use:

export DPCPP_CPU_SCHEDULE=dynamic

For changing the backend where the device executes, you can use "SYCL_DEVICE_FILTER" variables, for example:

export SYCL_DEVICE_FILTER=gpu,level_zero

Publications

• Youssef Faqir-Rhazoui, Carlos García and Francisco Tirado (2023). "Performance Portability Assessment: Non-negative Matrix Factorization as a Case Study". Euro-Par 2022: Parallel Processing Workshops. Euro-Par 2022. Lecture Notes in Computer Science, vol 13835. Springer, Cham.
  • Free available here.

Acknowledgements

This work has been supported by the EU (FEDER) and the Spanish MINECO and CM under grants S2018/TCS-4423 and RTI2018-093684-B-I00.