Draft: [luci/service] Support Conv2D ops dynamic shape inference

[pcs1265]

This supports dynamic shape inference for Conv2D and DepthwiseConv2D.

If the input feature map (ifm) has unknown dimensions, it forwards those unknown dimensions directly to the output shape.

related to : #13697

I am planning to make some PRs for this changes. (Splitting into smaller changes..)

1. Migrate existing sinf rules (Conv2D, DepthwiseConv2D) to sinf::Algorithm
   • Separate the infer_conv2d_type function used in Conv2D and DepthwiseConv2D into the HelperConv2Ds.h header file.
2. Implement dynamic shape inference support for Conv2D
   • Add known states for height and width to the OutputSize struct.
   • Make infer_conv2d_type function can aware unknown dimensions.
3. Implement dynamic shape inference support for DepthwiseConv2D
   • Use known states that added in right before PR step.
4. (Optional) Refactor to make the infer_conv2d_type function returns a loco::TensorShape object
   • Return the object with only the inferred height and width set (in NHWC format).
   • Other dimensions will be returned as unknown.
   • This can remove redundant casting to get shape
5. (Optional too) Refactor to remove template from infer_conv2d_type function.
   • But this makes the infer_conv2d_type function has 5(!!) arguments.
6. Add some tests related to Conv2D and DepthwiseConv2D
   • I am thinking hard about how to add NEG test to dynamic shape inference.

ONE-DCO-1.0-Signed-off-by: Chansu Park [email protected]