Re: [PATCH v10 00/13] accel/rocket: RK3576 NPU (RKNN) enablement

From: Igor Paunovic <royalnet026@gmail.com>
To: Jiaxing Hu <gahing@gahingwoo.com>
Cc: Tomeu Vizoso <tomeu@tomeuvizoso.net>,
	linux-rockchip@lists.infradead.org,
	dri-devel@lists.freedesktop.org,
	Igor Paunovic <royalnet026@gmail.com>
Subject: Re: [PATCH v10 00/13] accel/rocket: RK3576 NPU (RKNN) enablement
Date: Fri, 11 Sep 2026 21:28:32 +0200	[thread overview]
Message-ID: <20260911192833.105634-1-royalnet026@gmail.com> (raw)
In-Reply-To: <20260911122832.1364839-1-gahing@gahingwoo.com>

Hi Jiaxing,

Thanks for going back over it, but the correction needs one more
turn: both sets of numbers are mine, from two different days.

The 45 (12 + 8 + 12 + 13) is the 19 August A/B square on v8 1-2/12
[1]: no manifestation on either arm, oracle 48/48 throughout. That
is what the patch text describes, and it is right.

The 102 is the 25 August re-run on v9 as posted [2]: v9 arm 12 + 11
resets, all 48/48; differential arm (same base, 1/13 and 2/13
removed) five runs, 8/10/12/8/15; plus 13 + 13 on a third arm with
3/13 added [3]. In one of the five differential runs the inference
issued after the forced autosuspend/resume "succeeded" and returned
all 48 output channels at 0x80 while the CPU reference varied
normally, with nothing in dmesg, lockdep or on the serial console.
That run was on RK3588, on my board, and the v10 cover quoted it as
I wrote it.

So the signature is shared, not carried over. Your explanation of
it (fresh shmem BO zeroed, teflon readback adding 0x80) is the same
mechanism I inferred from the scorer, which flagged 128 as not this
model's zero point, i.e. an output buffer that was never written.
What holds is the caveat I put on it then: one event in 53
differential resets, zero in 49 with the patches, timing-dependent,
and my protocol cannot tell a genuinely hung block from a lost
completion. It bounds; it does not prove. The artifacts (scorer
output, runtime and genpd state, journal, serial log) are preserved
if you or Tomeu want them.

If you respin, both statements can stand next to each other: 45
resets with no manifestation on 19 August, and one silent completion
on the unpatched arm out of 102 resets on 25 August.

[1] https://lore.kernel.org/all/20260819073530.6087-1-royalnet026@gmail.com/
[2] https://lore.kernel.org/all/CAEWPSH5mxTbUkNouxm6yecMZYvDowquhvYvhaXQ8HoMtHD5U1g@mail.gmail.com/
[3] https://lore.kernel.org/all/CAEWPSH5MVzoQMaFOEQ93+1pOPmcXCDNfuXsv5j-X8Lb1fy_hrg@mail.gmail.com/

Regards,
Igor