BLAS for FPGA based on the DaCe framework

This repository demonstrates fast linear algebra on FPGA using the DaCe parallel programming framework. The following program serves as a synthetic benchmark:

M1 = np.outer(x, y1) + A  // GER
M2 = np.outer(x, y2) + A  // GER
M3 = np.outer(x, y3) + A  // GER
M4 = np.outer(x, y4) + A  // GER
M5 = np.outer(x, y5) + A  // GER
M6 = np.outer(x, y6) + A  // GER
M7 = np.outer(x, y7) + A  // GER
M8 = np.outer(x, y8) + A  // GER

vector1 = M1 @ x  // GEMV
vector2 = M2 @ x  // GEMV
vector3 = M3 @ x  // GEMV
vector4 = M4 @ x  // GEMV
vector5 = M5 @ x  // GEMV
vector6 = M6 @ x  // GEMV
vector7 = M7 @ x  // GEMV
vector8 = M8 @ x  // GEMV

vector1 = vector1 + vector2  // AXPY
vector3 = vector3 + vector4  // AXPY
vector5 = vector5 + vector6  // AXPY
vector7 = vector7 + vector8  // AXPY

vector1 = vector1 + vector3  // AXPY
vector5 = vector5 + vector7  // AXPY

buf = vector1 + vector5  // AXPY
result = np.sum(buf)  // ASUM

Precompiled bitstreams for the FPGA (Xilinx U250 XDMA) are in /bitstreams. The code requires a recent version of GCC to be available in the environment and a recent version of CMake. To run and compile for the FPGA a Xilinx SDK has to be available in the environment as well.

Compile

To preprare for compilation run:

source setSubmodules.sh

to install and setup the DaCe framework and

source setupPython.sh (you must have the virtualenv package installed)

to setup the Python environment. Run the below commands to compile, after compiling once one can ran with the -c flag, which prevents recompilation and directly runs the previously compiled code.

FPGA

Run: performance_synthetic1.py -m hardware 8192

CPU

Make sure OpenBLAS is loaded in your environment Run: performance_synthetic1_cpu.py 8192

GPU

Make sure the CUDA toolkit is available in your environment Run: performance_synthetic1_gpu.py 8192

Run Benchmark

To run the benchmarks which will run the above scripts mutliple times and collect the data and report median timings run the following commands.

FPGA

Run run_experiments_platforms_fpga.py -v xilinx

CPU

Run run_experiments_platforms_cpu.py

GPU

Run run_experiments_platforms_gpu.py