README.md

This Repository is archived

The code of the STREAM Benchmark for FPGA was integrated into the HPCC FPGA Benchmark suite. As a consequence, this repository is no longer maintained and the development will continue in the new repository as part of HPCC FPGA.

STREAM for FPGA

This repository contains the STREAM benchmark for FPGA and its OpenCL kernels. Currently only the Intel® FPGA SDK for OpenCL™ utility is supported.

It is a modified implementation of the STREAM benchmark by John D. McCaplin, Ph.D. available at https://www.cs.virginia.edu/stream/.

Build

Depending on the use case you may want to change certain lines in the Makefile:

1. Check the location of the used compilers (A C++ compiler and the aoc/aocl)
2. Update the board name in the Makefile or give the new board name as an argument to make.
3. Set the used data type using the STREAM_TYPE variable and the number the for loops in the kernels should be unrolled with UNROLL_COUNT.
4. Set the OpenCL specific flags PLATFORM_ID and DEVICE_ID to the use the desired FPGA device. For more information look into the stream_fpga.cpp
5. Build the host program or the kernels by using the available build targets.

For more detailed information about the available build targets call:

make

without specifying a target. This will print a list of the targets together with a short description.

Similar to the original STREAM benchmark it is also possible to modify the size of the used buffers as well as the number of iterations with the variables STREAM_ARRAY_SIZE and NTIMES.

To make it easier to generate different versions of the kernels, it is possible to specify a variable BUILD_SUFFIX when executing make. This suffix will be added to the kernel name after generation.

Example:

make host BUILD_SUFFIX=18.1.1

Will build the host and name the binary after the given build suffix. So the host would be named 'stream_fpga_18.1.1'. The file will be placed in a folder called 'bin' in the root of this project.

Execution

The created host uses the kernel with the same build suffix by default. It tries to load it from the same directory it is executed in.

./stream_fpga_18.1.1

will try to load the kernel file with the name stream_kernels_18.1.1.aocx by default. Additionally, the executable will interpret the first argument given as the path to the kernel that should be used. For example to use the kernel 'other.aocx':

./stream_fpga_18.1.1 other.aocx

Also, relative and absolute paths to the kernel can be given.

Result interpretation

The output of the host application is similar to the original STREAM benchmark:

Function    Best Rate MB/s  Avg time     Min time     Max time
Copy:           30875.9     0.025914     0.025910     0.025919
Scale:          30885.6     0.025905     0.025902     0.025911
Add:            46289.2     0.025928     0.025924     0.025935
Triad:          45613.4     0.026310     0.026308     0.026312
PCI Write:       6324.0     0.189800     0.189753     0.189862
PCI Read:        5587.3     0.214869     0.214773     0.214943

In addition it also measures the bandwidth of the connection between host and device. It is distinguished between writing to and reading from the devices memory. The buffers are written to the device before every iteration and read back after each iteration.

Different Kernel Source Files

The repository contains two OpenCL files with implementations of the STREAM kernels. stream_kernels.cl is using only the unrolling pragma without specific code optimizations. stream_kernels_vec.cl is making use of the OpenCL vector types. Moreover, it is optimized for boards with 4 or more banks by manually unrolling the scale and copy kernel. To synthesize the kernels from other files, the kernel source file can be given with KERNEL_SRCS For synthesizing the vector type version run:

make kernel KERNEL_SRCS=stream_kernels_vec.cl

Exemplary Results

The benchmark was executed on Bittware 520N cards for different Intel® Quartus® Prime versions. The detailed results of the runs are given in results.txt and as CSV files in the subfolder csv_result_export.

Single Precision

Double Precision

Usage of the Global Ring

It is possible to force a ring interconnect for the global memory with the compiler command -global-ring. To test the impact of this type of interconnect, the benchmark was also synthesized with the additional parameters -global-ring -duplicate-ring for all SDK versions supporting this option.

The raw data of these runs can be found in the folder csv_result_export.

Single Precision

Double Precision