[PATCH v7 08/10] accel/rocket: add RK3576 NPU (RKNN) support

[Jiaxing Hu]

Hi Igor,

> I had the window right and the location wrong.

That is the useful half. v8 carries synchronize_irq(core->irq) before the
guard in rocket_reset(), with the comment above it saying what is
actually guaranteed rather than "Remaining interrupts have been handled".
Driving the reset path deliberately rather than waiting for a timeout is
worth more than the rest of the run put together, since that is the only
path either change is for.

> Say the word and I will point it at the CPU reference the way you did.

Please do, and thank you. It is the one comparison I cannot produce, and
it separates two things that look identical from here: a defect specific
to this SoC, and the reference's own rounding compounding through a
chain.

Two things before you spend time on it.

The number moved. When I wrote 995 of 1001 there was still one fault
left in Mesa, an output channel count that is not a multiple of two,
which the CNA reads in pairs. With that fixed it is 1000 of 1001, so it
is one output rather than six, and whether one output is even worth
chasing is a fair question. The comparison is still worth having for the
LAYERS rather than the final vector.

And the oracle matters more than the run. A per output comparison
against the CPU is not enough on its own past the first layer or two,
because tflite's requant and the hardware's disagree by design and that
disagreement compounds: at operator 6 a flawless accelerator scores 4 of
128 channels against the CPU. vendor-capture/chainmodel.py in

  https://github.com/gahingwoo/linux-rk3576-npu

runs the graph twice from the model file, once with tflite's
SaturatingRoundingDoublingHighMul and RoundingDivideByPOT and once with
the hardware's single half up shift, and prints what a perfect
accelerator would score at every operator. Read your numbers against
that column rather than against 128 of 128, or every deep layer will
look broken on both SoCs.

If it is easier, mn_L00 through mn_L26 in that repository are MobileNet
with its graph output moved to each operator's output, which is a four
byte patch of the flatbuffer and needs no converter. Those are what the
per layer table came from.

Jiaxing