AlterOpLayout issue when using AutoTVM

taknevski · June 6, 2019, 5:38pm

I’m trying to follow the approach here:
https://github.com/dmlc/tvm/blob/master/tutorials/autotvm/tune_relay_x86.py
to tune a full resnet network.
However I’m trying to use my own tunable conv2d routine as a autotvm template.
Is there a way to do this? Currently the tuning proceeds without error but when build_module.build is called I get an error in AlterOpLayout saying that the configuration lacks value for “tile_ic” and “tile_oc”
the innermost dimension in NCHWc.

Thanks

eqy · June 6, 2019, 9:09pm

There has been a recent patch for an AlterOpLayout fix merged into master, have you updated your local copy of TVM?

merrymercy · June 7, 2019, 11:48pm

You should disable the alter_op_layout pass, because it assumes the configuration space has “tile_ic” and “tile_oc” fileds for x86 cpu.
This assumption is hard-coded in this function

github.com

dmlc/tvm/blob/4cee98ba6c9c20beb3b9744a758786a345695537/topi/python/topi/x86/conv2d.py#L391


new_data = tvm.placeholder(new_data_shape, data.dtype)
new_kernel = tvm.placeholder(new_kernel_shape, kernel.dtype)


C = _declaration_conv_NCHWc(cfg, new_data, new_kernel, strides, padding, dilation,
                            data_layout, out_layout, dtype)
s = _schedule_conv2d_NCHWc(cfg, [C])
return s, [new_data, new_kernel, C]




@conv2d_alter_layout.register("cpu")
def _alter_conv2d_layout(attrs, inputs, tinfo, F):


copy_inputs = [s for s in inputs]
new_attrs = {k : attrs[k] for k in attrs.keys()}


if F.__name__ == 'tvm.relay.op':
    # Derive channels for frontends (e.g ONNX) that miss "channel" field.
    new_attrs["channels"] = inputs[1].checked_type.shape[attrs['kernel_layout'].index('O')]


data, kernel = tinfo[0], tinfo[1]
batch_size, in_channel, height, width = get_const_tuple(data.shape)