Vitis HLS synthesis front-end OOM crash for some MVAU configs

[fpjentzsch]

In some cases clang memory usage during the initial HLS phase explodes, resulting in the clang process being killed. The user receives ERROR: [HLS 200-1715] Encountered problem during source synthesis in the vitis_hls.log with no further explanation.

Example configs for the test_fpgadataflow_mvau_rtlsim unit test I have tested:

#@pytest.mark.parametrize("mem_mode", ["internal_decoupled"])
#@pytest.mark.parametrize("act", [DataType["UINT7"]])
#@pytest.mark.parametrize("wdt", [DataType["INT7"]])
#@pytest.mark.parametrize("idt", [DataType["INT8"]])
#@pytest.mark.parametrize("nf", [5])
#@pytest.mark.parametrize("sf", [720])
#@pytest.mark.parametrize("mw", [3600])
#@pytest.mark.parametrize("mh", [400])

Result: As soon as

remark: /home/felixj/WD/finn/deps/finn-hlslib/mvau.hpp:234:33: Applying array_partition to 'accu': Complete partitioning on dimension 1. Complete partitioning on dimension 2.

is reached in \sol1\.autopilot\db\a.g.ld.0.bc.clang.err.log, RAM usage explodes to > 40 GB, followed by

clang: error: unable to execute command: Killed
clang: error: clang frontend command failed due to signal (use -v to see invocation)
clang version 7.0.0
Target: fpga64-xilinx-none
Thread model: posix
InstalledDir: /tools/Xilinx/Vitis_HLS/2022.2/lnx64/tools/clang-3.9-csynth/bin
clang: note: diagnostic msg: PLEASE submit a bug report to http://llvm.org/bugs/ and include the crash backtrace, preprocessed source, and associated run script.
clang: note: diagnostic msg: Error generating preprocessed source(s) - no preprocessable inputs.

This suggests that the problem is encountered during or directly after array partitioning. Tested with 2022.2 and 2023.1.

Curiously, if I keep SIMD=5 & PE=80 but reduce MW & MH to get folding factors of 1:

@pytest.mark.parametrize("mem_mode", ["internal_decoupled"])
@pytest.mark.parametrize("act", [DataType["UINT7"]])
@pytest.mark.parametrize("wdt", [DataType["INT7"]])
@pytest.mark.parametrize("idt", [DataType["INT8"]])
@pytest.mark.parametrize("nf", [1])
@pytest.mark.parametrize("sf", [1])
@pytest.mark.parametrize("mw", [5])
@pytest.mark.parametrize("mh", [80])

Result: Test passes, RAM usage peaks at ~12 GB. The clang log contains only one additional line beyond the line it failed at previously:

remark: /home/felixj/WD/finn/deps/finn-hlslib/mvau.hpp:237:30: Applying array_partition to 'w': Complete partitioning on dimension 1.

See also this AMD forum thread.