A compiler error when doing autotune on Intel graphics


#1

the tensorflow inference network which has conv2d and depthwise dilation ops, with layout =“NCHW”, target = tvm.target.intel_graphics(), the 48 tasks could be optimized with ‘n_trial’: 24, ‘early_stopping’: 10 (for test, used small num). but after the tune_taks, there is a compile error when trying to do graph, lib, params = relay.build_module.build( net, target=target, params=params) with
relay.build_config(opt_level=3):

File “/home/dolphin/projects/tvm/topi/python/topi/intel_graphics/conv2d.py”, line 206, in _alter_conv2d_layout
ic_bn, oc_bn = cfg[“tile_ic”].size[-1], cfg[“tile_oc”].size[-1]
File “/home/dolphin/projects/tvm/python/tvm/autotvm/task/space.py”, line 805, in getitem
return self._entity_map[name]
KeyError: ‘tile_ic’

Thanks a lot!
@Laurawly @merrymercy

ps: when opts_level set = 2, in with relay.build_config(opt_level=2), the autotune on intel_graphics passed successfully.