[Apache TVM Discuss] [Questions] Layout of custom layer

[fush via Apache TVM Discuss]

I have a custom layer "nucfpga.l2norm", and follow the doc [Convert Layout Pass 
— tvm 0.8.dev0 documentation 
(apache.org)](https://tvm.apache.org/docs/dev/convert_layout.html?highlight=finfercorrectlayout)
 ,I have set the  `FInferCorrectLayout`.

    RELAY_REGISTER_OP("nucfpga.l2norm")
        .set_attrs_type<NucfpgaL2normAttrs>()
        .set_num_inputs(2)
        .add_argument("data", "Tensor", "The input tensor.")
        .add_argument("scale", "1D Tensor", "scale.")
        .set_support_level(3)
        .add_type_rel("nucfpgaL2norm", NucfpgaL2normRel)
        .set_attr<FInferCorrectLayout>("FInferCorrectLayout", 
NucfpgaL2normInferCorrectLayout);

    Array<Array<Layout>> NucfpgaL2normInferCorrectLayout(const Attrs& attrs,
                              const Array<Layout>& new_in_layouts,
                              const Array<Layout>& old_in_layouts,
                              const Array<tvm::relay::Type>& old_in_types) {

      ICHECK_EQ(old_in_layouts.size(), 2U);
      ICHECK_EQ(old_in_types.size(), 2U);
      Layout data_layout = old_in_layouts[0];
      if (new_in_layouts.defined()) {
        ICHECK_EQ(new_in_layouts.size(), 2U);
      }
      return Array<Array<Layout>>{{data_layout, Layout("C")}, {data_layout}};
    }




In the host ,

            desired_layouts = {'nn.conv2d': ['NHWC', 'OHWI16o16i']} 
            # Convert the layout to NCHW
            # RemoveUnunsedFunctions is used to clean up the graph.
            seq = 
tvm.transform.Sequential([relay.transform.RemoveUnusedFunctions(),
                                            
relay.transform.ConvertLayout(desired_layouts)])
            with tvm.transform.PassContext(opt_level=3):
              mod = seq(mod)
            print(mod)


and get output

      %73 = nn.relu(%72) /* ty=Tensor[(1, 38, 38, 512), int8] */;
      %74 = layout_transform(%73, src_layout="NHWC", dst_layout="NCHW") /* 
ty=Tensor[(1, 512, 38, 38), int8] */;
      %75 = nucfpga.l2norm(%74, %L2Norm.weight, ifrac_input="", 
ifrac_kernel="", ifrac_output="") /* ty=Tensor[(1, 512, 38, 38), int8] */;
      %76 = layout_transform(%75, src_layout="NCHW", dst_layout="NHWC") /* 
ty=Tensor[(1, 38, 38, 512), int8] */;


My l2norm layer input layout transfrom NHWC to NCHW,and output layout transfrom 
NCHW to NHWC，**but my op does not need do this because the compute is already 
NHWC** ，I do not need the `layout_transform()` .

 What should I do?

Thanks.





---
[Visit Topic](https://discuss.tvm.apache.org/t/layout-of-custom-layer/10413/1) 
to respond.

You are receiving this because you enabled mailing list mode.

To unsubscribe from these emails, [click 
here](https://discuss.tvm.apache.org/email/unsubscribe/72d1c66931706341d6ddcb8cdc8d79a217e8482a6d3c430fa737939014c86c33).