Vitis System Compilation mode can also be used with standard Vitis kernels (.xo) and the application acceleration design flow. Therefore, you can use VSC with Vitis HLS C++ kernels, as described in Vitis High-Level Synthesis User Guide (UG1399), and RTL kernels, as described in RTL Kernel Development Flow.
Given a VSC accelerator code, sc_acc.cpp, and a Vitis kernel, kernel.xo, the following would be the command lines required to build the project:
v++ -c sc_acc.cpp -o sc_acc.o
v++ -l sc_acc.o kernel.xo -o hw.xclbin # also produces hw.o
g++ -c host.cpp -o host.o
g++ -l -o host.exe host.o sc_acc.o hw.o -lvpp_acc -lxrt_hw -lpthread
In the scenario as it has been described, the kernel.xo is already compiled. So, the VSC accelerator code (sc_acc.cpp) needs to be compiled as shown in the first
v++ command above. The output of this is the object
file sc_acc.o. The v++
command recognizes the VSC accelerator code, and compiles it properly into the object
file.
The v++ --link command is then run to link the
.xo and .o
files into the .xclbin. This produces a hw.o file in addition to the hw.xclbin file. The hw.o file is
required for compilation of the host application by the g++ command as is shown in the next steps.
The g++ command compiles the host code,
and then links the various object files (host.o,
sc_acc.o, and hw.o) to produce the host.exe
executable. Notice the addition of the vpp_acc and the
xrt_hw libraries to the g++ command line as they are required for the linking process.
In a standard Vitis flow, using
native-XRT or OpenCL code, the user host code has to
explicitly load the .xclbin. In a mixed VSC-XO flow
the user host code can still load the .xclbin
explicitly, but is not required. Instead, the device-ID can be obtained by calling API
vpp::sc::get_xrt_device().
Buffers objects can be easily passed between the native XRT and VSC APIs. The two possible buffer transfer scenarios are described in the following sections.