Open Zamion101 opened 3 days ago
According to your output, the application seems to crash on this line
xrt::device device = xrt::device(device_id);
or when retrieving the device info.
Is XRT properly set up on the compute node (maybe try to execute xbutil examine
)? Does creating a xrt::device
object work for you with other applications?
Please also note, that our project is currently only tested with the U280.
When running on Noctua 2, you should run "make xclbin" as a compute job, as it is a long running synthesis. There is a script for it (https://github.com/pc2/Aurora-HLS/blob/main/scripts/synth.sh):
The example is running on an U280 card. The n2fpga17 node has a custom configuration, so you need to set the correct PART and PLATFORM when running the synthesis. It could be u280 or u55. The following is just a guess for u55, better check what are the correct ones.
sbatch ./scripts/synth.sh make xclbin PART=xcu55c-fsvh2892-2L-e PLATFORM=xilinx_u55c_gen3x16_xdma_3_202210_1
And also the correct xrt version needs to be set in the env.sh file, if it is fixed one.
For running it is also best to use the provided scripts (adjusted to the configuration): https://github.com/pc2/Aurora-HLS/blob/main/scripts/run_N1.sh
Okay after talking with Riebler, found out that the custom configuration on n2fpga17
does not includes XRT
and that's the reason why the Aurora-HLS does not work on it. Also n2fpga17 does include 2 Alveo U280 cards and not U55C. I'm trying on another node at the moment.
I also would like to ask is there a way to import or port the Aurora-HLS for XDMA as well?
The focus of Aurora-HLS is the support for Vitis/XRT toolchains. But Maybe you can just build the object files by running
make aurora_hls_0.xo aurora_hls_1.xo
0 for QSFP0 and 1 for QSFP1.
Actually, the object files should already be there from your previous synthesis.
Both kernels should have two AXI streams (512bit by default) rx_axis
and tx_axis
. Then you need to link the object files into your design using your tool flow and connect the AXI streams accordingly. That should do the trick for the streaming configuration.
I'm trying to setup Aurora-HLS on Noctua2 on Node
n2fpga17
with 2xcu280_u55c_0
connected together in the following configuration. (https://pc2.github.io/fpgalink-gui/index.html?import=%20--fpgalink%3Dn00%3Aacl0%3Ach0-n00%3Aacl1%3Ach0%20--fpgalink%3Dn00%3Aacl0%3Ach1-n00%3Aacl1%3Ach1)I followed the README and compiled the project with
make aurora
than build the example withafter the 2 hours of wating for
xclbin
to be compiled, I tried using./host_aurora_hls_test
and got[n2fpga17:3185664] MCW rank 0 is not bound (or bound to all available processors)
error/warning as well as[n2fpga17:3185664:0:3185664] Caught signal 11 (Segmentation fault: address not mapped to object at address (nil))
.I added following debug prints to
host/host_aurora_hls_test.cpp
and tried againwhich gave me following output and the place where the problem occurs.
I know the cards are present and working as shown in the screenshot below: