Defining an Advanced Compute Operator (Conv2D)

The following takes the Conv2D operator as an example to describe how to define an advanced operator.

Conv2D operator prototype definition is as follows.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

REG_OP(Conv2D)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT8, DT_BF16}))
    .INPUT(filter, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT8, DT_BF16}))
    .OPTIONAL_INPUT(bias, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT32}))
    .OPTIONAL_INPUT(offset_w, TensorType({DT_INT8}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT32, DT_BF16}))
    .REQUIRED_ATTR(strides, ListInt)
    .REQUIRED_ATTR(pads, ListInt)
    .ATTR(dilations, ListInt, {1, 1, 1, 1})
    .ATTR(groups, Int, 1)
    .ATTR(data_format, String, "NHWC")
    .ATTR(offset_x, Int, 0)
    .OP_END_FACTORY_REG(Conv2D)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17

auto conv2d = op::Conv2D("Conv2d")
    // quant, conv_weight, conv_bias are three input nodes.
    .set_input_x(quant)
    .set_input_filter(conv_weight)
    .set_input_bias(conv_bias)
    .set_attr_strides({ 1, 1, 1, 1 })
    .set_attr_pads({ 0, 0, 0, 0 })
    .set_attr_dilations({ 1, 1, 1, 1 });

TensorDesc conv2d_input_desc_x(ge::Shape(), FORMAT_NCHW, DT_INT8);
TensorDesc conv2d_input_desc_filter(ge::Shape(), FORMAT_NCHW, DT_INT8);
TensorDesc conv2d_input_desc_bias(ge::Shape(), FORMAT_NCHW, DT_INT32);
TensorDesc conv2d_output_desc_y(ge::Shape(), FORMAT_NCHW, DT_INT32);
conv2d.update_input_desc_x(conv2d_input_desc_x);
conv2d.update_input_desc_filter(conv2d_input_desc_filter);
conv2d.update_input_desc_bias(conv2d_input_desc_bias);
conv2d.update_output_desc_y(conv2d_output_desc_y);

The major steps are as follows:

1. Call an operator constructor, for example, call the Conv2D(const char* name) constructor to create an operator instance, and pass the operator name (for example, Conv2D) to this call.

   1	auto conv2d1 = op::Conv2D("Conv2d")

2. Call set_input_inputName to set operator inputs.

   1 2 3

   .set_input_x(data) .set_input_filter(conv_weight) .set_input_bias(conv_bias)

   data is the input node of the graph. It is constructed from the Data operator. For details, see Defining a Data Node (Data).

   conv_weight is a constant data constructed from the Const operator. For details, see Defining a Data Node (Const).

   conv_bias is a constant data constructed from the Const operator. For details, see Defining a Data Node (Const).

3. Call set_attr_attributeName to set operator attributes.

   1 2 3

   .set_attr_strides({1, 1, 1, 1}) // Set the strides attribute values. .set_attr_pads({0, 0, 0, 0}) // Set the pads attribute values. .set_attr_dilations({1, 1, 1, 1}); // Set the dilations attribute values.

4. For convolution operators such as Conv2D or operators that are sensitive to processing along the C axis, you are advised to set the format to NCHW or NHWC as required in the update_input_desc_inputName call.

   1 2 3 4 5 6 7 8

   TensorDesc conv2d_input_desc_x(ge::Shape(), FORMAT_NCHW, DT_INT8); TensorDesc conv2d_input_desc_filter(ge::Shape(), FORMAT_NCHW, DT_INT8); TensorDesc conv2d_input_desc_bias(ge::Shape(), FORMAT_NCHW, DT_INT32); TensorDesc conv2d_output_desc_y(ge::Shape(), FORMAT_NCHW, DT_INT32); conv2d.update_input_desc_x(conv2d_input_desc_x); conv2d.update_input_desc_filter(conv2d_input_desc_filter); conv2d.update_input_desc_bias(conv2d_input_desc_bias); conv2d.update_output_desc_y(conv2d_output_desc_y);

   IR graph building does not support the following formats.

   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

   FORMAT_NC1HWC0 FORMAT_FRACTAL_Z FORMAT_NC1C0HWPAD FORMAT_NHWC1C0 FORMAT_FRACTAL_DECONV FORMAT_C1HWNC0 FORMAT_FRACTAL_DECONV_TRANSPOSE FORMAT_FRACTAL_DECONV_SP_STRIDE_TRANS FORMAT_NC1HWC0_C04 FORMAT_FRACTAL_Z_C04 FORMAT_FRACTAL_DECONV_SP_STRIDE8_TRANS FORMAT_NC1KHKWHWC0 FORMAT_C1HWNCoC0 FORMAT_FRACTAL_ZZ FORMAT_FRACTAL_NZ FORMAT_NDC1HWC0 FORMAT_FRACTAL_Z_3D FORMAT_FRACTAL_Z_3D_TRANSPOSE FORMAT_FRACTAL_ZN_LSTM FORMAT_FRACTAL_Z_G FORMAT_ND_RNN_BIAS FORMAT_FRACTAL_ZN_RNN FORMAT_NYUV FORMAT_NYUV_A