I tried autotvm resnet-18 by setting like: target = tvm.target.Target(target=“opencl”, host=“llvm”) .
The tune tasks run sucessfully, however failed in tune_graph.
> Traceback (most recent call last):
> File "./tune_relay_local_opencl.py", line 245, in <module>
> tune_and_evaluate(tuning_option)
> File "./tune_relay_local_opencl.py", line 214, in tune_and_evaluate
> tune_graph(mod["main"], data_shape, log_file, graph_opt_sch_file)
> File "./tune_relay_local_opencl.py", line 193, in tune_graph
> executor = Tuner(graph, {input_name: dshape}, records, target_op, target)
> File "/Users/banma-1396/proj/tvm/tvm/python/tvm/autotvm/graph_tuner/dynamic_programming_tuner.py", line 44, in __init__
> super(DPTuner, self).__init__(*args, **kwargs)
> File "/Users/banma-1396/proj/tvm/tvm/python/tvm/autotvm/graph_tuner/base_graph_tuner.py", line 202, in __init__
> self._fetch_cfg()
> File "/Users/banma-1396/proj/tvm/tvm/python/tvm/autotvm/graph_tuner/base_graph_tuner.py", line 286, in _fetch_cfg
> for record in cfg_dict[workload]:
> KeyError: ('conv2d_NCHWc.x86', ('TENSOR', (1, 3, 224, 224), 'float32'), ('TENSOR', (64, 3, 7, 7), 'float32'), (2, 2), (3, 3, 3, 3), (1, 1), 'NCHW', 'NCHW', 'float32')
After add more detail print, I found the op of “conv2d_NCHWc.x86” generated from tune_graph is not right. From the autotvm tasks log, the scheduling primitives maybe something like this: conv2d_nchw.cuda.
I tried the target LLVM, the python script is OK. Does graph tuner work on openCL ?