[feature]ascend 支持dlpack · Pull Request !23870 · Ascend/pytorch

1. ascend 支持dlpack
---

**What type of PR is this?**
> Uncomment only one ` /kind <>` line, hit enter to put that in a new line, and remove leading whitespaces from that line:
>
> /kind bug
> /kind task
> /kind feature

**What does this PR do / why do we need it**:
torch_npu 支持dlpack，torch原dlpack相关功能移到torch_npu 使用monkey patch替换原torch接口

测试代码如下：

```
import torch
import torch_npu
import numpy as np
torch.set_grad_enabled(False)

def test_torch_slicing_to_torch():
    print(f'===================test_torch_slicing_to_torch=====================', flush = True)
    src_tensor = torch.zeros((3,4), device='npu', dtype=torch.float16)
    print("=======start  src_tensor dtype=", src_tensor.dtype, ", device=", src_tensor.device, ", shape=", src_tensor.shape, flush=True)
    print("=======start  src_tensor storage_size=", src_tensor.untyped_storage().size(), ", storage_offset=", src_tensor.storage_offset(), ", stride=", src_tensor.stride(), flush=True)
    print("=======start  src_tensor.data_ptr()=", src_tensor.data_ptr(), flush=True)
    # 单行切片
    row_slice = src_tensor[1, :]

print(f'row_slice is: {row_slice}', flush = True)

print("=======before  row_slice dtype=", row_slice.dtype, ", device=", row_slice.device, ", shape=", row_slice.shape, flush=True)
    print("=======before  row_slice storage_size=", row_slice.untyped_storage().size(), ", storage_offset=", row_slice.storage_offset(), ", stride=", row_slice.stride(), flush=True)
    print("====== before  row_slice.data_ptr()=", row_slice.data_ptr(), flush=True)
    pack = torch.utils.dlpack.to_dlpack(row_slice)
    print(f'torch pack: {pack}', flush = True)
    convert_tensor = torch.utils.dlpack.from_dlpack(pack)
    print("convert_tensor: ", convert_tensor, flush=True)
    print("=======after  convert_tensor dtype=", convert_tensor.dtype, ", device=", convert_tensor.device, ", shape=", convert_tensor.shape, flush=True)
    print("=======after  convert_tensor storage_size=", convert_tensor.untyped_storage().size(), ", storage_offset=", convert_tensor.storage_offset(), ", stride=", convert_tensor.stride(), flush=True)
    print("====== after  convert_tensor.data_ptr()=", convert_tensor.data_ptr(), flush=True)

def test_torch_to_torch():
    print(f'===================test_torch_to_torch=====================', flush = True)
    src_tensor = torch.zeros((3,4), device='npu', dtype=torch.float16)
    print("=======start  src_tensor dtype=", src_tensor.dtype, ", device=", src_tensor.device, ", shape=", src_tensor.shape, flush=True)
    print("=======start  src_tensor storage_size=", src_tensor.untyped_storage().size(), ", storage_offset=", src_tensor.storage_offset(), ", stride=", src_tensor.stride(), flush=True)
    print("=======start  src_tensor.data_ptr()=", src_tensor.data_ptr(), flush=True)

pack = torch.utils.dlpack.to_dlpack(src_tensor)
    print(f'torch pack: {pack}', flush = True)
    convert_tensor = torch.utils.dlpack.from_dlpack(pack)
    print("convert_tensor: ", convert_tensor, flush=True)
    print("=======after  convert_tensor dtype=", convert_tensor.dtype, ", device=", convert_tensor.device, ", shape=", convert_tensor.shape, flush=True)
    print("=======after  convert_tensor storage_size=", convert_tensor.untyped_storage().size(), ", storage_offset=", convert_tensor.storage_offset(), ", stride=", convert_tensor.stride(), flush=True)
    print("====== after  convert_tensor.data_ptr()=", convert_tensor.data_ptr(), flush=True)

def test_torch_slicing_to_torch_2():
    print(f'===================test_torch_slicing_to_torch_2=====================', flush = True)
    src_tensor = torch.ones((3,4), device='npu', dtype=torch.float16)
    print("=======start  src_tensor dtype=", src_tensor.dtype, ", device=", src_tensor.device, ", shape=", src_tensor.shape, flush=True)
    print("=======start  src_tensor storage_size=", src_tensor.untyped_storage().size(), ", storage_offset=", src_tensor.storage_offset(), ", stride=", src_tensor.stride(), flush=True)
    print("=======start  src_tensor.data_ptr()=", src_tensor.data_ptr(), flush=True)
    # 单行切片
    row_slice = src_tensor[1, :]

print(f'row_slice is: {row_slice}', flush = True)

print("=======before  row_slice dtype=", row_slice.dtype, ", device=", row_slice.device, ", shape=", row_slice.shape, flush=True)
    print("=======before  row_slice storage_size=", row_slice.untyped_storage().size(), ", storage_offset=", row_slice.storage_offset(), ", stride=", row_slice.stride(), flush=True)
    print("====== before  row_slice.data_ptr()=", row_slice.data_ptr(), flush=True)
    pack = torch.utils.dlpack.to_dlpack(row_slice)
    print(f'torch pack: {pack}', flush = True)
    convert_tensor = torch.utils.dlpack.from_dlpack(pack)
    print("convert_tensor: ", convert_tensor, flush=True)
    print("=======after  convert_tensor dtype=", convert_tensor.dtype, ", device=", convert_tensor.device, ", shape=", convert_tensor.shape, flush=True)
    print("=======after  convert_tensor storage_size=", convert_tensor.untyped_storage().size(), ", storage_offset=", convert_tensor.storage_offset(), ", stride=", convert_tensor.stride(), flush=True)
    print("====== after  convert_tensor.data_ptr()=", convert_tensor.data_ptr(), flush=True)

def test_torch_to_torch_2():
    print(f'===================test_torch_to_torch_2=====================', flush = True)
    src_tensor = torch.ones((3,4), device='npu', dtype=torch.float16)
    print("=======start  src_tensor dtype=", src_tensor.dtype, ", device=", src_tensor.device, ", shape=", src_tensor.shape, flush=True)
    print("=======start  src_tensor storage_size=", src_tensor.untyped_storage().size(), ", storage_offset=", src_tensor.storage_offset(), ", stride=", src_tensor.stride(), flush=True)
    print("=======start  src_tensor.data_ptr()=", src_tensor.data_ptr(), flush=True)

pack = torch.utils.dlpack.to_dlpack(src_tensor)
    print(f'torch pack: {pack}', flush = True)
    convert_tensor = torch.utils.dlpack.from_dlpack(pack)
    print("convert_tensor: ", convert_tensor, flush=True)
    print("=======after  convert_tensor dtype=", convert_tensor.dtype, ", device=", convert_tensor.device, ", shape=", convert_tensor.shape, flush=True)
    print("=======after  convert_tensor storage_size=", convert_tensor.untyped_storage().size(), ", storage_offset=", convert_tensor.storage_offset(), ", stride=", convert_tensor.stride(), flush=True)
    print("====== after  convert_tensor.data_ptr()=", convert_tensor.data_ptr(), flush=True)

test_torch_slicing_to_torch()
test_torch_to_torch()
test_torch_slicing_to_torch_2()
test_torch_to_torch_2()
```

**Which issue(s) this PR fixes**:

Fixes #

**Special notes for your reviewers**:

Ascend/pytorch .gitee-modal { width: 500px !important; }

搜索帮助

Ascend/pytorch