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

[Igor Paunovic]

Hi Jiaxing,

Here is the RK3588 column, all 27 operators, ROCKET_SEED=7, scored the
way perch.py scores: a channel is good when its maxdiff (md below)
against max(cpu, output zero point) is at most 1.

Setup: Orange Pi 5 Plus (RK3588), kernel 7.2.0-rc6, the rocket driver
from this kernel's tree rebuilt with my clocks-by-name and devfreq
patches on top, Mesa at bf70ab68a21, teflon delegate, model
mobilenet_v1_1_224_quant.tflite from the Mesa test suite
(md5 4f348b87dca3315d2b3646cf5a3b31cf), per-operator models generated
with the four byte output patch you described. The "correct hw" column
is your chainmodel.py against the same model file. One difference to
flag up front: against this model file chainmodel prints 36/256 for
operator 8 where your table has 34/256, so our model files are not
byte-identical, and the columns below should be read against each
other rather than against your FINDINGS numbers.

  op  kind        correct hw       RK3588
  0   conv         32/32   md 1     32/32   md 1
  1   depthwise    28/32   md 3     28/32   md 3
  2   1x1          22/64   md 6     22/64   md 6
  3   depthwise    21/64   md 13    21/64   md 13
  4   1x1          18/128  md 14     8/128  md 15
  5   depthwise     9/128  md 13     3/128  md 15
  6   1x1           4/128  md 23     1/128  md 24
  7   depthwise     7/128  md 10     3/128  md 12
  8   1x1          36/256  md 7     20/256  md 17
  9   depthwise    32/256  md 11    22/256  md 13
  10  1x1          29/256  md 9     28/256  md 10
  11  depthwise    53/256  md 8     48/256  md 11
  12  1x1         166/512  md 11   121/512  md 13
  13  depthwise   142/512  md 9    137/512  md 12
  14  1x1          82/512  md 7     72/512  md 10
  15  depthwise   154/512  md 10   115/512  md 15
  16  1x1          82/512  md 7     71/512  md 10
  17  depthwise   156/512  md 14   141/512  md 21
  18  1x1          92/512  md 7     77/512  md 10
  19  depthwise   170/512  md 7    152/512  md 12
  20  1x1         102/512  md 8     84/512  md 10
  21  depthwise   166/512  md 13   150/512  md 12
  22  1x1         174/512  md 6    161/512  md 7
  23  depthwise   293/512  md 5    261/512  md 7
  24  1x1         671/1024 md 6    636/1024 md 5
  25  depthwise   718/1024 md 9    684/1024 md 8
  26  1x1         572/1024 md 25   574/1024 md 23

Three things stand out from here.

Operators 0 through 3 score identically to your chain simulation --
same good-channel counts, same maxdiff -- including operator 3, the
stride 2 depthwise with the asymmetric padding you suspect for the
first RK3576 divergence. They are not byte-identical to the simulated
hardware: diffing the raw tensors against the requant_hw chain shows a
few hundred elements per surface already off by 1-4 at operators 0-3.
That looks like the same small extra rounding difference that pushes
the scores below your column from operator 4 on; through operator 3 it
just stays under the maxdiff <= 1 scoring threshold.

There is no md 255 anywhere. From operator 4 on, RK3588 sits somewhat
below the simulation (8 vs 18 at op 4, 1 vs 4 at op 6), but the
maxdiff never exceeds 24 across all 27 operators and the deep layers
track the simulation closely (574 vs 572 at op 26).

A control run with ROCKET_SEED=11 keeps the same character: operator 0
still 32/32, no saturated maxdiff anywhere, worst case md 35 at
operator 6.

So from the RK3588 side your read looks right: the deep-layer
compounding is the reference artifact, and the RK3576 collapse from
operator 4 with maxdiff 255 has no counterpart here.

Igor