You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The ACCL XRT test suite fails for large counts on the bcast with root 1 tests.
ACCL was synthesized with TCP stack and UDP stack from dev at 36eebbb. All other tests passed.
The error appears in both versions.
Setup:
2x Alveo U280, directly connected
XRT:2.12
Platform: xilinx_u280_xdma_201920_3
Vitis: 2021.2
Tests using streams are commented out to prevent deadlocks
Interesting. I've seen this before when testing the segmentation in hardware, but I assumed this was because of a problem in the segmentation implementation.
Actually looking further into it, I'm pretty sure this is caused by the segmentation implementation. In the current dev branch the maximum segment size is not set to the rxbufsize, which makes the segmenting fall back to the default 1 KiB. On the dev branch only bcast implements segmentation. This has been 'fixed' (the segmentation of broadcast is still broken) in #127, where the segment size is set correctly.
TristanLaan
changed the title
ACCL XRT Test suite: bcast with root 1 fails for large counts
Broadcast segmentation fails when root != 0
Nov 28, 2022
The ACCL XRT test suite fails for large counts on the bcast with root 1 tests.
ACCL was synthesized with TCP stack and UDP stack from
dev
at 36eebbb. All other tests passed.The error appears in both versions.
Setup:
Execution:
TCP version:
mpirun -n 2 bin/test -f -x the.xclbin -t -b 2 -s 512
UDP version:
mpirun -n 2 bin/test -f -x the.xclbin -u -b 2 -s 512
Output for the two failing tests:
The text was updated successfully, but these errors were encountered: