[ERROR] Error while compiling mobilenet_v1_1.0_224.tflite on opencl target

Troubleshooting
#1

Hi,

I was trying execute following code:

tflite_model_file = os.path.join("mobilenet_v1_1.0_224.tflite")
tflite_model_buf = open(tflite_model_file, "rb").read()

try:
      import tflite
      tflite_model = tflite.Model.GetRootAsModel(tflite_model_buf, 0)
  except AttributeError:
      print('AttributeError')
      import tflite.Model
      tflite_model = tflite.Model.Model.GetRootAsModel(tflite_model_buf, 0)

shape_dict = {'input': (1, 244, 244, 3)}

sym, params = tvm.relay.frontend.from_tflite(tflite_model, shape_dict=shape_dict, dtype_dict={'input': 'float32'})

graph, lib, params = tvm.relay.build(sym, target='opencl', params=params)

and I get following error.

VMError: Check failed: match: iter_var(threadIdx.x, , threadIdx.x) domain already inferred, cannot prove their extents are the same 256 vs 128

How to troubleshoot this issue?

Thank you for help in advance.

Best Regards

Marcin Sielski

#2

Your problem is similar to mine. I think that my solution can be applied to the problem.

I compile a TFLite model according to the following TVM tutorial: Deploy a Framework-prequantized Model with TVM with the TVM commit da27e6d9a466263a9a0025aba92086a8bf837edb on 2020/8/7

The tutorial run successfully with “llvm” target; however, it fails with “opencl” target, and I get an error message similar to yours. Note that, to run the tutorial with “opencl” target, the modification is as follows:

target = 'opencl'
ctx = tvm.cl(0)
rt_mod = graph_runtime.GraphModule(lib['default'](tvm.cl(0)))

To solve the problem, at first I convert the layout into NCHW after relay.frontend.from_tflite:

desired_layouts = {'nn.conv2d': ['NCHW', 'default'], 'qnn.conv2d': ['NCHW', 'default']}
seq = tvm.transform.Sequential([relay.transform.RemoveUnusedFunctions(),
                                relay.transform.ConvertLayout(desired_layouts)])
with tvm.transform.PassContext(opt_level=3):
    mod = seq(mod)

Then, I remove topi.cuda.conv2d_nchw_int8 form python/tvm/relay/op/strategy/cuda.py because TVM uses cuda’s schedules for scheduling opencl, and I find that conv2d_nchw.cuda and depthwise_conv2d_nchw.cuda are stable enough to run the MobileNet model.

Finally, we can run the modified tutorial successfully and get the following result:

TVM Top-5 labels: [387 102 386 349 341]
TFLite Top-5 labels: [387 102 386 341 880]
Elapsed average ms: 2.1388191100000005