Investigate whether raw data or external data should be used

[mingmingtasd]

Some initializers must be uploaded to raw data, for example:

1. Reshape op needs parameter shape as raw data to do shape inference.
2. Reduce op needs parameter axes as raw data.
3. Expand op needs parameter shape as raw data.
4. Slice op needs starts, ends and steps as raw data.

While for some other ops, initializers have not to be raw data:
min/max value of clamp can be either raw data or external data

Are there other ops requiring initializer as raw data?
Currently we upload all webnn constants into external data (for potential zero-copy) but what if some ops like Reshape can't take external data as initializers?

@huningxin @shiyi9801 @lisa0314 @miaobin

[shiyi9801]

/cc @Honry

[huningxin]

@miaobin mentioned that the raw data is required by shape inference code.

#47 (comment)

[huningxin]

While for some other ops, initializers have not to be raw data: min/max value of clamp can be either raw data or external data

WebNN clamp min/max values are just scalars. Raw data (copying) vs. external data (non-copying) seems not to be much different.

[huningxin]

The guideline recommends using raw data when

• the value is less than 128 bytes (because there's about 60 bytes or so of overhead to use the external memory structure for pre-allocated memory), the scalar should be in this category
• the value is used for shape inference

We may need to add a helper to calculate the size of an initializer and use raw data if its size is less than 128 bytes.

See more discussions at microsoft/onnxruntime#23223 (comment)

[huningxin]

@mingmingtasd , your instanceNormalization PR mentions

// ONNX requires scale and bias inputs. And they must be uploaded as raw data,
// otherwise there will be runtime error.

https://github.com/shiyi9801/chromium/blame/ort_backend/services/webnn/ort/graph_builder_ort.cc#L666

I understand scale and bias won't be used for shape inference. What's the runtime error if using pre-allocated data?

[mingmingtasd]

I understand scale and bias won't be used for shape inference. What's the runtime error if using pre-allocated data

Sorry, after verifying, the scale and bias of instance norm can be external data, I added wrong annotation here. I will fix ~

We may need to add a helper to calculate the size of an initializer and use raw data if its size is less than 128 bytes.

I will do this.