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

[Jiaxing Hu]

Hi Igor,

That is the measurement I could not make from here, and between your row and one
of mine the question is answered.

The control you propose is not one I can run in that form. Upstream Mesa has no
RK3576 path, so its encoder emits the RK3588 register layout, and this SoC does
not share that layout at the same offsets. I would not expect the result to be a
convolution at all, and I have not run it to find out.

What I can run is the vendor userspace, which is the same trade the other way
around, same silicon and an entirely different stack. Five models, output tensor
read as int8, the NPU interrupt count advancing by exactly one per run so
each of them reached the hardware.

  model    zero point   values below it            sitting exactly on it
  a_lin       17          2478 of 4096    60.50%      32
  a_lin2     -14          2420 of 4096    59.08%      41
  g_cal       -8        108564 of 204800  53.01%    2551
  pq_oc        0         62720 of 128576  48.78%    3136
  w_160        0        236287 of 409600  57.69%    2877

g_cal is conv2d-cal's geometry exactly, 16 input channels to 128 output over an
80x80 surface, 5x5 at stride 2. pq_oc and w_160 carry conv2d-cal's zero point
exactly, 0 in int8 being the 128 my tables report in uint8. So the geometry and
the zero point are each covered by a model that does not clamp. The counts
sitting on the zero point are 0.7 to 2.4 percent of the surface, which is close
to what you measured on the other side and close to what the distribution gives.

The grid now reads

  RK3576, vendor userspace     does not clamp
  RK3588, upstream Mesa        does not clamp, your run
  RK3576, my Mesa              clamps

The first and third rows are the same silicon. So the clamp is mine, and there
is no hardware behaviour left for me to appeal to. Your third control is the one
that makes your row carry weight, since fifteen channels off by one is something
a delegate falling back to the CPU could not produce.

I have not found it yet. The register stream is byte identical to the vendor's
at this geometry apart from addresses, the requantisation, the pad value and the
padding. A, B and C swapped one at a time from the vendor's records each leave
the floor where it is. The weight buffer is the last thing I have not compared.

The 88 and 120 sweep is queued for the next time the board is flashed, and I
will send what it says either way.

Jiaxing