Simplify UnpackInitializerData API

[guoyu-wang]

Description: Simplify UnpackInitializerData API

Motivation and Context

• There are crashes reported due to unaligned memory access on arm, such as Fix arm crash due to unaligned raw_data buffer of initializer tensors onnx/onnx#3626, and Add timeseries imputer transformer featurizer kernel #2813, there are many places in our mobile EP code still access the initializer data directly, which may cause issue later
• Plan is to move all the EP code accessing initializer data directly to use UnpackInitializerData
• This change is to simplify the API of UnpackInitializerData, to use std:vector instead of a pair of std::unique_ptr for data and a size_t for size, also add one extra UnpackInitializerData to work on internal initializer only (return error when it is used on tensor with external data)
  • The only disadvantage of using vector is that the memory need to be initialized, which may have slight perf hit while creating session
• There will be on extra change to update all the EP code, one typical update is at onnxruntime/core/providers/shared/utils/utils.cc

[edgchen1]

is it possible to use std::byte?

[guoyu-wang]

Unfortunately, seems CUDA does not have std::byte support yet, revert back to use uint8+t

[edgchen1]

is this vector<uint8_t>::data guaranteed to be suitably aligned for floats?

[guoyu-wang]

In theory the data in the vector will be aligned to std::max_align_t, which is usually 16 or maybe 8 on some 32bit system, this makes it enough for all the scalar types we support for now.
To make it more robust, (so far I don't think we have 16bit data type yet, but just in case it will be added in the future), we can change the unpacked_tensor.resize(tensor_byte_size); to something like
unpacked_tensor = std::vector<std::byte>(tensor_byte_size, custom_alligned_allocator);, can look into this later

[edgchen1]

thanks - sounds like it should be fine for now then

[yuslepukhin]

unpacked_tensor[0]

Need to check that the data is not empty. Perhaps, this was the reason for crashes? #Resolved

[guoyu-wang]

Yes, we should check the length of the buffer, but this is not the reason for the crashes

[yuslepukhin]

ORT_RETURN_IF_ERROR(onnxruntime::Env::Default().ReadFileIntoBuffer(

Should we check for zero len? #Resolved

[guoyu-wang]

Seems zero len is fine here, same in posix version

onnxruntime/onnxruntime/core/platform/windows/env.cc

Lines 255 to 256 in 2243804

	if (length == 0)
	return Status::OK();

[yuslepukhin]