[Bug] Hang after load with stride (32 bit ew)

[callme-sam]

Description

I created a test case to reproduce a hang that occurs after a strided vector load (vlse32.v) when no hardware barrier is inserted.

The test case tests a linear algebra kernel (matrix multiplication with transposed B).
The main.c computes the result and compares it against a golden model.
Input data and expected output are provided in data.h and were generated via a Python script.

How to reproduce

• Extract the attached zip archive into sw/spatzBenchmarks

• Add the following lines to sw/spatzBenchmarks/CMakeLists.txt:

add_library(mat-mul-trans-B mat-mul-trans-B/kernel/mat-mul-trans-B.c)
add_spatz_test_threeParam(mat-mul-trans-B mat-mul-trans-B/main.c 64 32 64)

• Build and run

Expected Behaviour

The code hangs after executing a strided vector load (vlse32.v).
Inserting a hardware barrier immediately after the load avoids the hang.

Workaround

Uncommenting the hardware barrier after the strided load fixes the issue:

asm volatile ("vlse32.v v16, (%0), %1" :: "r"(col_b1), "r"(stride));
asm volatile ("vlse32.v v24, (%0), %1" :: "r"(col_b1_next), "r"(stride));
snrt_cluster_hw_barrier();

Recompiling with the barrier enabled allows the test to run correctly.

# Loading entry point: 80000000
#
##################################### MATRIX_MUL_TRANS_B TEST ####################################
#
# INFO | Running 'matrix_mul_trans_B' test on Spatz Cluster
# INFO | Test SUCCESS
#
##########################################################################################
#
# ** Info: [SUCCESS] Program finished successfully

mat-mul-trans-B.zip