Skip to content

Latest commit

 

History

History
73 lines (58 loc) · 2.31 KB

readme.md

File metadata and controls

73 lines (58 loc) · 2.31 KB

StreamBlocks Platform Profilers

In this directory you can find two basic profilers for software (FIFO bandwidth) and hardware (OpenCL r/w over PCIe).

Software bandwidth

Go to software directory and run the following commands:

> cd software
> mkdir -p build
> cd build
> cmake .. -DCMAKE_BUILD_TYPE=Release
> cmake --build .
> cd ../bin
> chmod +x bandwidth-test
> ./bandwidth-test

You can configure the profiler in terms of number of repetitions, minimum and maximum buffer size in bytes (should be power of 2). By either modifying the CMakeLists.txt file in software or passing extra flags to cmake, e.g.,:

> cmake .. -DCMAKE_BUILD_TYPE=Release -DNUM_LOOPS_VALUE=100 -DBUFFER_SIZE_MIN=8 -DBUFFER_SIZE_MAX=1024

Once the test is done, a bandwidth.xml file is generated. You can use that in the partitioning tool (i.e., provide its path in the json file).

Hardware-software bandwidth

Performing hardware-software profile is slightly more complicated. You need to first implement a loopback example in CAL for your platform of choice (you need the usual environment variables for Vivado and XRT):

> cd hardware
> mkdir -p build
> cd build
> cmake .. -DFPGA_NAME=... -DPLATFORM_NAME=... -DUSE_VITIS=on
> cmake --build . -t Loopback_1_10_xclbin

This will generate an FPGA bitstream using 10ns HSL clock target (though the actual kernel may run at a different clock). This kernel has a single input and a single output, you can change the number of inputs and outputs up to to what is defined in Loopback.cal (e.g, -t Loopback_4_10_xclbin would have 4 connections).

After you have the FPGA bitstream (environment variable XILINX_XRT should be defined):

cd ../bandwidth
mkdir -p build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
cmake --build .
cd ../
./bandwidth --repeats 1000 --min-buffer 8 --max-buffer 1048576  ---randomize 0 --width 1

You can change the argument values if you are profiling for different configurations. The result of this run would be dumped in the run directory prefixed with stat_. The profiling information is originally in a detailed json format per buffer size configuration. You can use bandwidth.py to obtain an xml summary that is used by the partitioning tool.

python3 bandwidth.py stats_ --use-mean --output opencl.xml