TheRock/rocm-libraries/dnn-providers/integration-tests/reference_data_scripts/utilities/tensor.py

## Copyright © Advanced Micro Devices, Inc., or its affiliates.
## SPDX-License-Identifier:  MIT

import torch
from common import DTypeConverter, tensor_uid
import os


class TensorAttributes:
    def __init__(
        self,
        tensor: torch.Tensor,
        uid: int = None,
        name: str = "",
        virtual: bool = False,
    ):
        self.uid = uid if uid != None else tensor_uid.get()
        self.tensor = tensor
        self.name = name
        self.virtual = virtual

    def __eq__(self, other):
        return self.uid == other.uid

    def __hash__(self):
        return self.uid

    def as_dict(self):
        return {
            "name": self.name,
            "uid": self.uid,
            "strides": list(self.tensor.stride()),
            "dims": list(self.tensor.size()),
            "data_type": DTypeConverter.to_string(self.tensor.dtype),
            "virtual": self.virtual,
        }

    def to_gpu(self):
        self.tensor = self.tensor.cuda()

    def to_cpu(self):
        self.tensor = self.tensor.cpu()

    @staticmethod
    def empty(
        dtype: torch.dtype = torch.float,
        dims: list[int] = [],
        uid: int = None,
        name: str = "",
        virtual: bool = False,
    ):
        return TensorAttributes(torch.empty(dims, dtype=dtype), uid, name, virtual)

    @staticmethod
    def random(
        min_val,
        max_val,
        dtype: torch.dtype,
        dims: list[int],
        generator: torch.Generator = None,
    ):
        if (
            dtype == torch.int
            or dtype == torch.int8
            or dtype == torch.int16
            or dtype == torch.int32
            or dtype == torch.int64
        ):
            return TensorAttributes(
                torch.randint(
                    low=int(min_val),
                    high=int(max_val),
                    dtype=dtype,
                    size=dims,
                    generator=generator,
                )
            )

        tensor = torch.rand(dtype=dtype, size=dims, generator=generator)
        return TensorAttributes(tensor * (max_val - min_val) + min_val)

    @staticmethod
    def from_dict(d: dict):
        dtype = DTypeConverter.from_string(d["data_type"])
        return TensorAttributes.empty(
            dtype, d["dims"], d["uid"], d["name"], d["virtual"]
        )


def dump_data_as_binary(filename: str, tensor_attr: TensorAttributes):
    tensor = tensor_attr.tensor
    with open(filename, "wb") as file:
        bytes = bytearray(tensor.untyped_storage())
        file.write(bytes)


def load_data_from_binary(filename: str, tensor_attr: TensorAttributes):
    tensor = tensor_attr.tensor
    num_bytes = os.path.getsize(filename)
    storage = torch.UntypedStorage.from_file(filename, nbytes=num_bytes)
    tensor.set_(
        storage, storage_offset=0, size=tuple(tensor.size()), stride=tensor.stride()
    )