gemm test regression after PMF integration

[AngryLoki]

sudo xrt-smi validate shows less TOPS in recent commits (up to current main branch). Bisecting (via <install> && sudo modprobe -r amdxdna && sudo modprobe amdxdna && sudo xrt-smi validate) shows that regression started at f0a2f48:

27516b2 (and every commit before, incl. release) - good, always at TOPS: 51.0:

Validate Device           : [0000:c5:00.1]
    Platform              : NPU Strix Halo
    Power Mode            : performance
-------------------------------------------------------------------------------
Test 1 [0000:c5:00.1]     : gemm 
    Details               : TOPS: 51.0                                          
    Test Status           : [PASSED]

f0a2f48 - bad, always at TOPS: 21.3

Test 1 [0000:c5:00.1]     : gemm 
    Details               : TOPS: 21.3                                          
    Test Status           : [PASSED]

Checked with kernel 7.0.6 (i. e. + HAVE_7_0_amd_pmf_get_npu_data: yes). Power measurement indeed works now, but it somehow affects gemm performance.

[maxzhen]

@NishadSaraf , please look into this issue.

[AngryLoki]

Hi, just a small note: I don't really see any regressions in the inference tasks. xrt-smi validate shows 2x less gemm TOPS, but in reality I can't measure any notable slowdown in token/sec.