Skip to content

UCLA-VAST/fpga-runtime

 
 

Repository files navigation

FPGA Runtime

This project provides a convenient runtime for PCIe-based FPGAs programmed under the OpenCL host-kernel model. Both Intel and Xilinx platforms are supported.

Prerequisites

  • Ubuntu 20.04+

Install from Binary

./install.sh

Usage

Invoking

template <typename... Args>
fpga::Instance Invoke(const std::string& bitstream, Args&&... args);

This invokes the kernel contained in file bitstream. bitstream should be a file that can be read via ifstream and can be a pipe with proper EOF. args are the arguments to the kernel. If an argument is not a scalar, it needs to be wrapped in one of the following wrappers:

ReadOnly(T* ptr, size_t n);
WriteOnly(T* ptr, size_t n);
ReadWrite(T* ptr, size_t n);

This will tell the runtime the data exchange direction and how many elements are allocated. The directions are with respect to the host, not the device (because this is host code). Passing a host pointer directly will not work (doesn't even compile).

Device Selection

By default, FRT selects devices using metadata from the bitstream. This may not always work as expected, often due to the following reasons:

  1. Xilinx 2RP shell platforms must be flashed by admin (root) before running any user logic.
  2. FRT may not know how to match the device name in the bitstream and the runtime device name. If you encounter this issue, please feel free to file a bug.

Selecting Xilinx Device by PCIe BDF

For Xilinx devices, it is possible to select the device by its PCIe BDF.

To do this, make sure you parsed gflags in your main function:

#include <gflags/gflags.h>
...
int main(int argc, char* argv[]) {
  gflags::ParseCommandLineFlags(&argc, &argv, /*remove_flags=*/true);
  ...
}

When running the host program, add --xocl_bdf=<bdf>, e.g.,

./host --xocl_bdf=0000:d8:00.1 ...

Profiling

Invoke returns an fpga::Instance object that contains profiling information.

double Instance::LoadTimeSeconds();
double Instance::ComputeTimeSeconds();
double Instance::StoreTimeSeconds();
double Instance::LoadThroughputGbps();
double Instance::StoreThroughputGbps();

Streaming

Streaming is supported (on legacy Xilinx platforms).

class fpga::ReadStream;
class fpga::WriteStream;

The streams need to be created and passed to fpga::Invoke as a parameter. If the arguments to fpga::Invoke contains a stream, it will not wait for the kernel to finish; instead, it will return an fpga::Instance object immediately. The host program can read from fpga::ReadStream and/or write to fpga::WriteStream. When all stream I/O are done, instance.Finish() should be invoked to wait until the kernel finishes.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • C++ 60.8%
  • CMake 26.8%
  • Shell 7.5%
  • C 2.7%
  • Python 1.6%
  • Roff 0.6%