"SIMD must divide IFMChannels" error but SIMD does divide IFMChannels

[julienmas]

Hi everyone,

I want to deploy a custom CNN on my Ultra96V2 board. I am trying to create a custom build flow, inspired from the end2end example notebooks tfc and cnv. When I call the transformation model = model.transform(ZynqBuild(platform = test_pynq_board, period_ns = target_clk_ns)), I get the error AssertionError: SIMD must divide IFMChannels, even though I made sure that SIMD values divides the IFMChannels values.
Here is my .onnx dataflow model unet_dataflow_model.txt (I put a .txt instead of .onnx to be able to upload it here) and the values I use for folding :

fc_layers = unet.get_nodes_by_op_type("MatrixVectorActivation")
# each tuple is (PE, SIMD, in_fifo_depth) for a layer
folding = [
    (4, 27, 32),
    (4, 72, 32),
    (1, 144, 32),
    (4, 72, 32),
    (2, 72, 32),
    (4, 72, 32),
    (2, 72, 32),
    (4, 72, 32),
    (2, 72, 32),
    (4, 72, 32),
    (1, 1, 32)
]
for fcl, (pe, simd, ififodepth) in zip(fc_layers, folding):
    fcl_inst = getCustomOp(fcl)
    fcl_inst.set_nodeattr("PE", pe)
    fcl_inst.set_nodeattr("SIMD", simd)
    fcl_inst.set_nodeattr("inFIFODepth", ififodepth)

# use same SIMD values for the sliding window operators
swg_layers = unet.get_nodes_by_op_type("ConvolutionInputGenerator")
for i in range(len(swg_layers)):
    swg_inst = getCustomOp(swg_layers[i])
    simd = folding[i][1]
    swg_inst.set_nodeattr("SIMD", simd)

I tried to use DataflowBuildConfig() and build_dataflow_cfg() functions to build a deployment package instead of a custom flow, but I get a similar error :

AssertionError: DWC OutWidth is bigger than InWidth and is not divisible by it.
                Please adjust PE and SIMD values so that OutWidth % InWidth = 0
                or alternatively use impl_style = vivado

Do you know what is going wrong?

Julien

[julienmas]

up

[MFJI]

I also met this problem.

[shakeelakram00]

I am also having the similar problem, did you found any solution for that?

---

KeyError Traceback (most recent call last)
/opt/finn/deps/qonnx/src/qonnx/custom_op/base.py in set_nodeattr(self, name, value)
102 try:
--> 103 (dtype, req, def_val, allowed_values) = self.get_nodeattr_def(name)
104 if allowed_values is not None:

/opt/finn/deps/qonnx/src/qonnx/custom_op/base.py in get_nodeattr_def(self, name)
48 allowed_values = None
---> 49 attrdef = self.get_nodeattr_types()[name]
50 if len(attrdef) == 3:

KeyError: 'inFIFODepth'

During handling of the above exception, another exception occurred:

AttributeError Traceback (most recent call last)
in
17 fcl_inst.set_nodeattr("PE", pe)
18 fcl_inst.set_nodeattr("SIMD", simd)
---> 19 fcl_inst.set_nodeattr("inFIFODepth", ififodepth)
20
21 # use same SIMD values for the sliding window operators

/opt/finn/deps/qonnx/src/qonnx/custom_op/base.py in set_nodeattr(self, name, value)
137 self.onnx_node.attribute.append(attr_proto)
138 except KeyError:
--> 139 raise AttributeError("Op has no such attribute: " + name)
140
141 def make_const_shape_op(self, shape):

AttributeError: Op has no such attribute: inFIFODepth

[anguyen9630]

I know this is a bit late but I finally managed to get it working so I am going to share for anyone who still needed it:
Error 1: AssertionError: SIMD must divide IFMChannels: If I remember correctly, this is when the convolution input generators IFMChannels % SIMD != 0. You must open up your ONNX file using Netron and check that.

Error 2: AssertionError: DWC OutWidth is bigger than InWidth and is not divisible by it.: This one I do not know how to fix the issue it self, because some of the modules you cannot set PE and some you cannot set SIMD so technically it is never achievable. HOWEVER, the solution is to instead use the Vivado config as suggested. Use force_rtl_conv_inp_gen = True and rtlsim_use_vivado_comps = True in the build config and then change your ENSURE THAT YOUR SIMD COUNT IS <= 512