When using ROCm and the device has largeBar enabled, we can read scalars directly from GPU memory. That way, there's no need for a malloc and memcpy.

## Benchmark
Simple synthetic benchmark:
```python
import torch
import time

device = torch.device("cuda")

N = 1000000
tensor = torch.arange(N)
tensor = tensor.to(device)

start = time.time()
for i in range(0, N):
    x = tensor[i].item()
    assert(x == i)
end = time.time()

print('retrieving {} scalars took {} s'.format(N, end - start))
```

Result before the change on MI-355X:
```
# python test-noprof.py
retrieving 1000000 scalars took 15.402856349945068 s
```

After the change:
```
# python test-noprof.py
retrieving 1000000 scalars took 2.8192553520202637 s
```

Pull Request resolved: #177023
Approved by: https://github.com/jeffdaily

Co-authored-by: Jeff Daily <jeff.daily@amd.com>

Co-authored-by: Xia-Weiwen <12522207+Xia-Weiwen@users.noreply.github.com>

When using ROCm and the device has largeBar enabled, we can read scalars directly from GPU memory. That way, there's no need for a malloc and memcpy.

## Benchmark
Simple synthetic benchmark:
```python
import torch
import time

device = torch.device("cuda")

N = 1000000
tensor = torch.arange(N)
tensor = tensor.to(device)

start = time.time()
for i in range(0, N):
    x = tensor[i].item()
    assert(x == i)
end = time.time()

print('retrieving {} scalars took {} s'.format(N, end - start))
```

Result before the change on MI-355X:
```
# python test-noprof.py
retrieving 1000000 scalars took 15.402856349945068 s
```

After the change:
```
# python test-noprof.py
retrieving 1000000 scalars took 2.8192553520202637 s
```

Pull Request resolved: pytorch#177023
Approved by: https://github.com/jeffdaily

Co-authored-by: Jeff Daily <jeff.daily@amd.com>

When using ROCm and the device has largeBar enabled, we can read scalars directly from GPU memory. That way, there's no need for a malloc and memcpy.

## Benchmark
Simple synthetic benchmark:
```python
import torch
import time

device = torch.device("cuda")

N = 1000000
tensor = torch.arange(N)
tensor = tensor.to(device)

start = time.time()
for i in range(0, N):
    x = tensor[i].item()
    assert(x == i)
end = time.time()

print('retrieving {} scalars took {} s'.format(N, end - start))
```

Result before the change on MI-355X:
```
# python test-noprof.py
retrieving 1000000 scalars took 15.402856349945068 s
```

After the change:
```
# python test-noprof.py
retrieving 1000000 scalars took 2.8192553520202637 s
```

Pull Request resolved: #177023
Approved by: https://github.com/jeffdaily

Co-authored-by: Jeff Daily <jeff.daily@amd.com>

When using ROCm and the device has largeBar enabled, we can read scalars directly from GPU memory. That way, there's no need for a malloc and memcpy.

## Benchmark
Simple synthetic benchmark:
```python
import torch
import time

device = torch.device("cuda")

N = 1000000
tensor = torch.arange(N)
tensor = tensor.to(device)

start = time.time()
for i in range(0, N):
    x = tensor[i].item()
    assert(x == i)
end = time.time()

print('retrieving {} scalars took {} s'.format(N, end - start))
```

Result before the change on MI-355X:
```
# python test-noprof.py
retrieving 1000000 scalars took 15.402856349945068 s
```

After the change:
```
# python test-noprof.py
retrieving 1000000 scalars took 2.8192553520202637 s
```

Pull Request resolved: pytorch#177023
Approved by: https://github.com/jeffdaily

Co-authored-by: Jeff Daily <jeff.daily@amd.com>

When using ROCm and the device has largeBar enabled, we can read scalars directly from GPU memory. That way, there's no need for a malloc and memcpy.

## Benchmark
Simple synthetic benchmark:
```python
import torch
import time

device = torch.device("cuda")

N = 1000000
tensor = torch.arange(N)
tensor = tensor.to(device)

start = time.time()
for i in range(0, N):
    x = tensor[i].item()
    assert(x == i)
end = time.time()

print('retrieving {} scalars took {} s'.format(N, end - start))
```

Result before the change on MI-355X:
```
# python test-noprof.py
retrieving 1000000 scalars took 15.402856349945068 s
```

After the change:
```
# python test-noprof.py
retrieving 1000000 scalars took 2.8192553520202637 s
```

Pull Request resolved: pytorch#177023
Approved by: https://github.com/jeffdaily

Co-authored-by: Jeff Daily <jeff.daily@amd.com>