Issue with Smaug & gem5-Aladdin Simulation Output After Hardware Modification

[Henry]

Hi all,

I am using Smaug with gem5-Aladdin to simulate an SoC integrating a systolic array accelerator.

I know that adding the -p option to the simulation command allows Smaug to print the ML model inference results at the end of the stdout file.

This works correctly when running gem5-Aladdin using the following command:

bash

CopyEdit

${gem5_dir}/build/X86/gem5.opt \ --debug-flags=Aladdin,HybridDatapath,SystolicToplevel \ --outdir=${cfg_home}/outputs \ --stats-db-file=stats.db \ ${gem5_dir}/configs/aladdin/aladdin_se.py \ --num-cpus=1 \ --mem-size=4GB \ --mem-type=LPDDR4_3200_2x16 \ --sys-clock=1.25GHz \ --cpu-clock=2.5GHz \ --cpu-type=TimingSimpleCPU \ --ruby \ --l2_size=2097152 \ --l2_assoc=16 \ --cacheline_size=32 \ --access-backing-store \ --accel_cfg_file=gem5.cfg \ --fast-forward=10000000000 \ -c ${bmk_dir}/smaug \ -o "${topo_file} ${params_file} --sample-level=no --gem5 --debug-level=0 --num-accels=1 --use-systolic-array -p" \ > stdout 2> stderr

However, after modifying the PE.cc hardware design—which should have affected the inference result—the output remains unchanged when running the simulation with the same command.

I am sure that I have recompiled the code. I am simulating lenet5 model which has the convulotion layer calling systolic array accelerator. Does anyone know why this is happening? Is there a possibility that the modified PE design is not being properly compiled, integrated, or used in the simulation? Is it possibile that the inference result from the gem5 simulation is not runtime calculated by the gem5 modeled hardware?

Any insights or suggestions would be greatly appreciated!

Many thanks,
Henry