TaPaSCo HPX Wrapper

Overview

The TaPaSCo HPX Wrapper allows to use TaPaSCo FPGA and AI Engine tasks transparently in HPX applications. It adopts the light-weight threading model of HPX in contrast to multi-threading on OS level when using TaPaSCo directly.

Usage

In your CMakeLists.txt, include HPX, TaPaSCo and TaPaSCoHPX and link against all required libraries:

find_package(HPX REQUIRED)
find_package(Tapasco REQUIRED)
find_package(TapascoHPX REQUIRED)

target_link_libraries(your_executable HPX::hpx HPX::wrap_main tapasco TapascoHPX)

In your application use #include <tapasco-hpx.hpp> to include the HPX wrapper, which will also include the tapasco.hpp header file. At the beginning of your application, instantiate the Tapascoobject:

tapasco::Tapasco tapasco;

If you have multiple FPGAs in your host computer, you need to pass the ID of your device when creating the object:

tapasco::Tapasco tapasco(tapasco::tlkm_access::TlkmAccessExclusive, dev_id);

You can then launch the TaPaSCo task by calling the provided wrapper function. Pass the TaPaSCo object, ID of the requested PE-type and all arguments to the wrapper function:

tapasco_hpx_wrapper(tapasco, pe_type, arg1, arg2, ...);

Accepted argument types are the same as for the reqular tapasco.launch() call. The wrapper function will handle and pass the arguments to the PE, start the PE and wait for the interrupt, while suspending the HPX light-weight thread in-between if necessary.

Please, consult the provided examples and TaPaSCo C++ examples for more details.

Data transfers

If you need to transfer larger data buffers, we recommend to do allocations and data transfers in separate tasks using the io_pool_executor of HPX. The TaPaSCo API provides the alloc() and free() calls for memory management, and copy_to() and copy_from() for data transfers. Have a look in the arraysum-pipeline example.

Examples

In order to build and run the provided examples, first clone and build TaPaSCo as described in the Readme. The examples require a bitstream containing at least one arraysum kernel, the arraysum-pipeline example also arrayupdate and arrayinit kernels. All three kernels are already provided by TaPaSCo. You can build the required bitstream (e.g. for the Alveo U280) using:

tapasco compose [arrayinit x 5, arrayupdate x 5, arraysum x 5] @ 250 MHz -p AU280

To build the example software, export the following variables so that CMake finds all required modules:

export HPX_DIR=/path/to/HPX
export TapascoHPX_DIR=/path/to/tapasco-hpx-wrapper
source /path/to/tapasco-workspace/tapasco-setup.sh

Then switch to the example directory, create a build directory and build with CMake:

cd /path/to/example/directory
mkdir build && cd build
cmake .. && make

Load the bitstream with

tapasco-load-bitstream your_bitstream.bit --reload-driver

and run the example application.

Publication

If you want to cite this work, please use the following information:

[Kalkhof2024] Torben Kalkhof, Carsten Heinz, and Andreas Koch. 2024. Enabling FPGA and AI Engine Tasks in the HPX Programming Framework for Heterogeneous High-Performance Computing. In Applied Reconfigurable Computing. Architectures, Tools, and Applications (ARC). DOI: 10.1007/978-3-031-55673-9_6

Important Note

Tested with TaPaSCo release 2024.1