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

[Jiaxing Hu <gahing@gahingwoo.com>]

Hi Igor,

A correction to what this cover said about your differential run, and to
what I said to you about it on 27 August.

The cover said a job had signalled completion with its output buffer
never written, all 48 channels 0x80, and that across 102 induced resets
in nine runs it appeared only on the arm without 1/13 and 2/13. None of
that is what you reported.

Your table is 12 + 8 + 12 + 13 = 45 resets, two passes per kernel, and
you wrote that the race did not manifest in the 45 on either kernel,
with the oracle 48/48 on both arms. The 0x80 is ours: on RK3576 a fresh
shmem BO is zeroed and teflon's readback adds 0x80, so an output buffer
that was never written reads back as a uniform 128. I carried our own
signature onto your RK3588 run, which never reported it.

The patch text is the part that was right, and it is the only form of
your result I will use: 45 induced resets across three cores, the domain
dropping every time with no MMU message on either kernel, so this is not
rocket-wide. You were careful to say your protocol could not settle
whether a genuinely hung block behaves differently. I turned a bounding
result into a reproduction.

Jiaxing

_______________________________________________
Linux-rockchip mailing list
Linux-rockchip@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-rockchip