[BUG] HW barrier needed after vector load (32 bit ew)

[callme-sam]

Description

In some cases, it is necessary to insert a hardware barrier (snrt_cluster_hw_barrier()) after a vle32.v loads to get correct results. This does not happen consistently, making the issue intermittent.

I was able to reproduce the problem by implementing a vectorized backward_solve function for upper-triangular linear systems. For this reason, I suspect the issue may be related to the vector length (vl) not being a multiple of 2, although this still needs verification.

How to reproduce

• Extract the attached zip archive into sw/spatzBenchmarks
• Add the following lines to sw/spatzBenchmarks/CMakeLists.txt

add_library(backward-solve backward-solve/kernel/backward-solve.c)
add_spatz_test_oneParam(backward-solve backward-solve/main.c 64)

• Build and run the test

Expected Result

The test should fail computing the correct solution of the linera system.

Workaround

• Uncomment the hw barrier after the two loads in kernel/backward-solve.c

asm volatile ("vle32.v v8, (%0)" :: "r"(p_dst));
asm volatile ("vle32.v v0, (%0)" :: "r"(p_mat));
snrt_cluster_hw_barrier(); 

• re-build and run test
• now test computes the correct solution of the linear system

backward-solve.zip