ytn86 / hlslib

A collection of extensions for Vivado HLS and Intel FPGA OpenCL to improve developer quality of life.

Do you use hlslib? Please consider citing us, and let us know so we can feature your project in the list of examples.

Quick introduction

hlslib is a collection of C++ headers, CMake files, and examples, aimed at improving the quality of life of HLS developers. The current repertoire primarily supports Vivado HLS, but some Intel FPGA OpenCL support is being added. An extended abstract describing the project is available here.

This project is developed at the Scalable Parallel Computing Lab (SPCL) at ETH Zurich (see our github).

How do I install it?

There are a few ways:

• Grab the headers and/or CMake files you need and stick them in your project.
• Install hlslib using the standard CMake installation procedure to a location of your choice.
• Clone this repository into your project as a git submodule and integrate it, with or without CMake.

How do I use it?

Just #include the header(s) you are interested in. You can see an example here.

When a Xilinx hlslib header is included, compilation must allow C++11 features, and the macro HLSLIB_SYNTHESIS must be set whenever HLS is run. Set -cflags "-std=c++11 -DHLSLIB_SYNTHESIS" in your synthesis script, and --xp prop:kernel.<entry function>.kernel_flags="-std=c++11 -DHLSLIB_SYNTHESIS" when building SDAccel kernels. See the included xilinx_test/CMakeLists.txt for reference.

Feature overview

We have Doxygen now! Simply run make to generate the docs.

A brief overview of hlslib features is given below.

CMake integration

For integrating the Xilinx or Intel HLS tools in your project, the FindSDAccel.cmake and FindIntelFPGAOpenCL.cmake are provided in the cmake subdirectory. The scripts will set all necessary variables required to build both host and device code.

Example CMakeLists.txt:

set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} hlslib/cmake)
find_package(SDAccel REQUIRED)

add_executable(MyHostExecutable src/MyHostExecutable.cpp)
include_directories(${SDAccel_INCLUDE_DIRS})
target_link_libraries(MyHostExecutable ${SDAccel_LIBRARIES})

add_custom_target(compile_kernel ${SDAccel_XOCC} -t hw src/MyKernel.cpp
                  --kernel MyKernel --platform xilinx_vcu1525_dynamic_5_1 -o MyKernel.xclbin)

DataPack

The hlslib::DataPack class located in hlslib/xilinx/DataPack.h facilitates SIMD-style vectorization, and makes it easy to build wide data paths in your code.

Examples usage:

hlslib::DataPack<float, 4> Foo(hlslib::DataPack<float, 4> &a,
                               hlslib::DataPack<float, 4> &b) {
  #pragma HLS INLINE     
  auto add = a + b; // Vector addition
  add[0] = add[1];  // Indexing for both reads and writes
  return 0.5 * add; // Element-wise multiplication by a scalar
}

Simulation

For kernels with multiple processing elements (PEs) executing in parallel, the hlslib/xilinx/Simulation.h adds some convenient macros to simulate this behavior, by wrapping each PE in a thread executed in parallel, all of which are joined when the program terminates.

Example usage:

HLSLIB_DATAFLOW_INIT();
hlslib::Stream<Data_t> pipes[kStages + 1];
HLSLIB_DATAFLOW_FUNCTION(MemoryToStream, memory_in, pipes[0]);
for (int i = 0; i < kStages; ++i) {
  #pragma HLS UNROLL
  HLSLIB_DATAFLOW_FUNCTION(PE, pipes[i], pipes[i + 1]); // Launches new C++ thread
}
HLSLIB_DATAFLOW_FUNCTION(StreamToMemory, pipes[kStages], memory_out);
HLSLIB_DATAFLOW_FINALIZE(); // In simulation mode, joins threads created as dataflow functions.

When building programs using the simulation features, you must link against a thread library (e.g., pthreads).

Stream

While Vivado HLS provides the hls::stream class, it is somewhat lacking in features, in particular when simulating multiple processing elements. The hlslib::Stream class in hlslib/xilinx/Stream.h compiles to Vivado HLS streams, but provides a richer interface. hlslib streams are:

• thread-safe during simulation, allowing producer and consumer to be executed in parallel;
• bounded, simulating the finite capacity of hardware FIFOs, allowing easier detection of deadlocks in software; and
• self-contained, allowing the stream depth and implementation (e.g., using LUTRAM or BRAM) to be specified directly in the object with excess pragmas.

Example usage:

void Bar(hlslib::Stream<int> &a, hlslib::Stream<int> &b, int N) {
  for (int i = 0; i < N; ++i) {
    #pragma HLS PIPELINE II=1
    auto read = a.Pop(); // Queue-like interface
    b.Push(read + 1);
  }
}

void Foo(hlslib::Stream<int> &in_stream, // Specifying stream depth is optional
         hlslib::Stream<int> &out_stream, int N) {
  #pragma HLS DATAFLOW
  
  hlslib::Stream<int, 4> foo_pipe; // Implements a FIFO of depth 4
  
  // Dataflow functions running in parallel
  HLSLIB_DATAFLOW_INIT();
  HLSLIB_DATAFLOW_FUNCTION(Bar, in_stream, foo_pipe, N);
  HLSLIB_DATAFLOW_FUNCTION(Bar, foo_pipe, out_stream, N);
  HLSLIB_DATAFLOW_FINALIZE();
}

OpenCL host code

To greatly reduce the amount of boilerplate code required to create and launch OpenCL kernels, and to handle FPGA-specific configuration required by the vendors, hlslib provides a C++14 convenience interface in hlslib/xilinx/SDAccel.h and hlslib/intel/OpenCL.h for Vivado HLS and Intel FPGA OpenCL, respectively.

Example usage:

using hlslib::ocl;
Context context;
std::vector<float> input_host(N, 5);   
std::vector<float> output_host(N, 5);
auto input_device = context.MakeBuffer<float, Access::read>(
    MemoryBank::bank0, input_host.cbegin(), input_end.cend());
auto output_device = context.MakeBuffer<float, Access::write>(MemoryBank::bank1, N);
auto program = context.MakeProgram("MyKernel.xclbin");
auto kernel = program.MakeKernel("MyKernel", input_device, output_device, N);
kernel.ExecuteTask();
output_device.CopyToHost(output_host.begin());

Other features

Various other features are provided, including:

• Classes to flatten loop nests and keep track of indices (include/hlslib/xilinx/Flatten.h), both for bounds known at runtime (hlslib::Flatten) and bounds known at compile-time (hlslib::ConstFlatten). Example usage can be found in xilinx_test/kernels/Flatten.cpp.
• Various compile-time functions commonly used when designing hardware, such as log2, in include/hlslib/xilinx/Utility.h Some of these headers are interdependent, while others can be included standalone. Refer to the source code for details.
• A template tcl-file that can be used with CMake or another templating engine to produce a high-level synthesis script
• xilinx_test/CMakeLists.txt that builds a number of tests to verify hlslib functionality, doubling as a reference for how to integrate HLS projects with CMake using the provided module files
• An example of how to use the Simulation and Stream headers, at xilinx_test/kernels/MultiStageAdd.cpp, both as a host-only simulation (xilinx_test/test/TestMultiStageAdd.cpp), and as a hardware kernel (xilinx_test/host/RunMultiStageAdd.cpp)
• include/hlslib/xilinx/Accumulate.h, which includes a streaming implementation of accumulation, including for type/operator combinations with a non-zero latency (such as floating point addition). Example kernels of usage for both integer and floating point types are included as xilinx_test/kernel/AccumulateInt.cpp and xilinx_test/kernel/AccumulateFloat.cpp, respectively.
• include/hlslib/xilinx/Operators.h, which includes some commonly used operators as functors to be plugged into templated functions such as TreeReduce and Accumulate.
• include/hlslib/xilinx/Axi.h, which implements the AXI Stream interface and the bus interfaces required by the DataMover IP, enabling the use of a command stream-based memory interface for HLS kernels if packaged as an RTL kernel where the DataMover IP is connected to the AXI interfaces.

Bundled libc++ with SDAccel

When linking against the Xilinx OpenCL libraries, the linker must add the runtime library folder to the library search path. This folder contains an ancient version of libc++.so, which can break compilation when using a newer compiler. To avoid this, backup or delete the library file located at:

<SDx or SDAccel folder>/runtime/lib/<architecture>/libstdc++.so

Ubuntu packages

On Ubuntu, the following package might need to be installed to run hardware emulation:

sudo apt install libc6-dev-i386

Projects using hlslib

• Matrix multiplication code: uses a wide range of hlslib features, including simulation, streams, vectors, CMake integration, and OpenCL wrapper code.
• SMI [1]: streaming message passing library for inter-FPGA communication in OpenCL. Uses hlslib for OpenCL host code.
• HelmGemm [2]: uses the simulation features of hlslib, and incorporates the matrix multiplication code above.
• DaCe [3]: a data-centric parallel programming framework targeting a wide range of architectures, including both Intel and Xilinx FPGAs. Uses hlslib for CMake integration, OpenCL host code, vectors, streams, and simulation.

If you use hlslib in your project, please let us know, so we can add you to the list.

Bugs and feature requests

Please use the issue tracker.

Publication

If your project uses hlslib, please consider citing us:

BibTeX:

@article{hlslib,
  title={{hlslib}: Software Engineering for Hardware Design},
  author={de~Fine~Licht, Johannes and Hoefler, Torsten},
  journal={arXiv preprint arXiv:1910.04436},
  year={2019}
}

Plain text:

de Fine Licht, Johannes, and Hoefler, Torsten. "hlslib: Software Engineering for Hardware Design" arXiv preprint arXiv:1910.04436 (2019).

References

• [1] De Matteis, Tiziano, de Fine Licht, Johannes, Beránek, Jakub, and Hoefler, Torsten. "Streaming Message Interface: High-Performance Distributed Memory Programming on Reconfigurable Hardware." arXiv preprint arXiv:1909.03231 (2019).
• [2] Diamantopoulos, Dionysios, and Hagleitner, Christoph. "HelmGemm: Managing GPUs and FPGAs for transprecision GEMM workloads in containerized environments." 2019 IEEE 30th International Conference on Application-specific Systems, Architectures and Processors (ASAP). Vol. 2160. IEEE, 2019.
• [3] Ben-Nun, Tal, de Fine Licht, Johannes, Ziogas, Alexandros Nikolaos, Schneider, Timo, and Hoefler, Torsten. "Stateful Dataflow Multigraphs: A Data-Centric Model for High-Performance Parallel Programs." arXiv preprint arXiv:1902.10345 (2019).