### **Pull Request: Optimized Non-Contiguous Tensor Copy for CPU to GPU in PyTorch**
#### **Summary**
This PR addresses the performance issue identified in [#111570](https://github.com/pytorch/pytorch/issues/111570), where non-contiguous tensors took significantly longer to transfer from CPU to GPU. Through detailed tracing of the call flow, we identified that PyTorch was creating temporary contiguous buffers for non-contiguous tensor transfers, which introduced unnecessary overhead.
#### **Tracing the Issue**
To pinpoint the cause of the slowdown, we followed the call flow from Python’s `tensor.cuda()` method through PyTorch’s backend, ultimately identifying `copy_kernel_cuda` as the key function responsible for CPU-to-GPU tensor transfers. Here’s a summary of the tracing process:
1. **Python Call: `tensor.cuda()`**
- Starting from Python, the `cuda()` method initiates the tensor transfer to the GPU.
2. **`TensorBody.h: cuda()`**
- The `cuda()` method calls `to()`, specifying the target device as CUDA.
3. **`Tensor.cpp: TensorBase::to()`**
- The `to()` function prepares device and data type options before invoking `_ops::to_dtype_layout::call()`.
4. **Operator Call: `_ops::to_dtype_layout::call()`**
- This operator dispatches the request to the backend-specific function responsible for managing the transfer.
5. **`Copy.cpp: copy_()`**
- The `copy_()` function performs preliminary checks (e.g., zero-tensor immutability) and proceeds to call `copy_impl()`.
6. **`Copy.cpp: copy_impl()`**
- This function sets up a tensor iterator and dispatches the copy operation to the appropriate backend through `copy_stub`.
7. **Dispatch to CUDA: `copy_stub`**
- The dispatch mechanism routes the call to the CUDA-specific function, `copy_kernel_cuda`.
8. **`Copy.cu: copy_kernel_cuda()`**
- Here, we identified that PyTorch was creating temporary contiguous buffers for 1D and 2D non-contiguous tensors, which slowed down the copy process. This behavior is managed by the `copy_requires_temporaries()` function.
#### **Solution**
To address this, we modified `copy_kernel_cuda` to handle non-contiguous 1D and 2D tensors directly by using `cudaMemcpy2DAsync`, which allows efficient, stride-aware memory transfers without temporary buffers. Here’s why this approach improves performance:
- **Efficiency of `cudaMemcpy2DAsync`**: This CUDA function is optimized for pitched (stride-based) memory transfers, allowing it to handle non-contiguous data layouts effectively by specifying memory strides for source and destination tensors.
- **Reduction of Overhead**: By directly copying non-contiguous tensors without intermediate buffers, we eliminate extra memory allocation and achieve faster CPU-to-GPU transfers.
- **Asynchronous Execution**: `cudaMemcpy2DAsync` enables asynchronous transfer on the CUDA stream, further improving performance by taking advantage of CUDA's optimized memory handling for non-contiguous layouts.
#### **Performance Results**
In my testing, I created tensors of size `327680 x 2000` and used slices for transfer performance measurements. The tests show that the average time for transferring a non-contiguous slice (e.g., rows 10,000 to 50,000) from CPU to GPU now closely matches the contiguous case. This improvement indicates that the updated implementation effectively addresses the performance discrepancy. Below are the measured times and validation checks:
```plaintext
Average time for contiguous slice (rows 10,000-50,000): 66 ms
Average time for non-contiguous slice (rows 10,000-50,000): 66 ms
Validation of contiguous and non-contiguous tensor copies:
✅ PASS: Tensor shapes match.
✅ PASS: Tensor contiguity matches.
✅ PASS: Tensor contents match.
✅ PASS: Tensor data types match.
✅ Success: Both contiguous and non-contiguous tensors were copied correctly to the GPU.
```
#### **Conclusion**
This PR resolves the identified performance issue by eliminating the need for temporary buffers in non-contiguous 1D and 2D tensor transfers, ensuring faster and more efficient copies from CPU to GPU. Future optimizations could further enhance performance for higher-dimensional non-contiguous tensors.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138964
Approved by: https://github.com/jeffdaily
Co-authored-by: Natalia Gimelshein <ngimel@gmail.com>
Co-authored-by: Jeff Daily <jeff.daily@amd.com>
…s_pinned if device is not initialized
Do not land
RFC
potential fix for #144687
Now `.is_pinned(device="cuda")` does not initialize device and thus doesn't poison the fork (but it complains about `device` arg being deprecated). To not need `device=` arg we'd need to fix get_accelerator to not initialize device.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145752
Approved by: https://github.com/albanD
Co-authored-by: albanD <albandes@fb.com>
# Motivation
for https://github.com/pytorch/pytorch/issues/143914
On Windows, there are two separate SYCL platforms for iGPU and dGPU. To simplify the logic, we will exclude iGPUs when a dGPU is present. This ensures that all XPU devices enumerated by PyTorch share the same SYCL context.
Now I generalize the logic as below:
1. We find the first L0 platform containing at least one dGPU and enumerate all dGPUs of that platform.
2. If no dGPU is found, we find the first L0 platform containing iGPU and enumerate all iGPUs of that platform.
3. No GPU is found (neither iGPU nor dGPU).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144378
Approved by: https://github.com/EikanWang, https://github.com/gujinghui
Using Philox4 as PRNG
Test plan (other that CI)
Run
```python
mport torch
from torch._inductor.utils import run_and_get_code
from contextlib import nullcontext
def foo(x):
return x * torch.randn_like(x)
foo_c = torch.compile(foo)
x = torch.ones(100, 100, device="mps")
y = foo_c(x)
print(y.mean().item(), y.std().item())
for i in range(25):
print(y[i].mean(), y[i].std())
```
And observe that printed values are close to 0 and 1
TODO: Better `randint` algorithm for large ranges
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145705
Approved by: https://github.com/dcci, https://github.com/jansel
#136627 has almost fixed the issue that test binaries' runpath has not been set correctly, with few cases left.
This PR fixes the rest.
The binaries are found by `auditwheel repair` a wheel built with `BUILD_TEST=1`.
@malfet
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144305
Approved by: https://github.com/malfet
May be to be later reused from eager op as well
Also, didn't know that Metal already have type_traits
And use `metal::isunorderder(a, b)` instead of `metal::isnan(a + b)` is it is defined as function that is equivalent `a != a || b != b`, but I suspect it might have a best native implementation for the specific architecture
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145157
Approved by: https://github.com/dcci
Useful for code reuse for Metal shader build both for eager mode and MPSInductor, but it requires one to implement `_cpp_embed_headers` tool that, as name suggests, would preprocess and embeds the for shader to be used in dynamic compilation.
Test using:
- `TestMetalLibrary.test_metal_include`
- Moving `i0`/`i1` implementation to `c10/util/metal_special_math.h` and call it from `SpecialOps.metal` shader, which now looks much more compact:
```metal
template <typename T, typename Tout = T>
void kernel
i0(constant T* input,
device Tout* output,
uint index [[thread_position_in_grid]]) {
output[index] = c10::i0(static_cast<Tout>(input[index]));
}
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145087
Approved by: https://github.com/dcci
ghstack dependencies: #145023
I added this to support code sharing with ExecuTorch, but the operator<< overrides are load-bearing for builds -- we have other code that attempts to pretty-print Half/BFloat16, and implicit conversions can't be used to make that work because there are *multiple* implicit conversions from Half/BFloat16 to primitive types, so which one to select is ambiguous. Also, we don't actually seem to need it now in ExecuTorch core because we have `include <ostream>` in there at the moment anyway.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144808
Approved by: https://github.com/janeyx99, https://github.com/malfet
PyTorch now support many private1 backend names like `AutogradPrivateUse1` or `QuantizedPrivateUse1`, not mentioned the original `PrivateUse1` backend.
However, users that implement `PrivateUse1` funtionalities would modified the backend name by calling `torch.utils.rename_privateuse1_backend("my_backend")`, in that case, all `PrivateUse1` backend string would not be found when we call other functions related to it. For example, we utilize `torch.library` to register some customize functions to our new backend, we would use "my_backend" as the backend name instead of "PrivateUse1", in which the error will be throw:
```
could not parse dispatch key 'my_backend'
```
So, this PR changed the function `c10::DispatchKey parseDispatchKey(const std::string& k)`, it would double check if the `PrivateUse1` has been modified, and if so, we would change `k` to adapt new backend name then find it again.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144325
Approved by: https://github.com/albanD
For correct import and export of functions when the dynamic linkage is used for HIP libraries on windows, the appropriate export/import macros need to be put in place. This Pull Request utilizes existing CUDA import/export macros by converting them to corresponding HIP macros during the hipification process.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144098
Approved by: https://github.com/jeffdaily
# Motivation
Fix https://github.com/pytorch/pytorch/issues/143543
# Solution
We should raise python exception instead of aborting...
# Additional Context
without this PR:
```python
>>> import torch
>>> torch.accelerator.current_stream(torch.accelerator.device_count())
terminate called after throwing an instance of 'c10::Error'
what(): device is out of range, device is 2, total number of device is 2.
Exception raised from check_device_index at /home/dvrogozh/git/pytorch/pytorch/c10/xpu/XPUFunctions.h:36 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >) + 0xac (0x7f30707eb95c in /home/dvrogozh/git/pytorch/pytorch/torch/lib/libc10.so)
frame #1: c10::detail::torchCheckFail(char const*, char const*, unsigned int, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&) + 0xf3 (0x7f307078fc57 in /home/dvrogozh/git/pytorch/pytorch/torch/lib/libc10.so)
frame #2: <unknown function> + 0x19a3e (0x7f3070c2ba3e in /home/dvrogozh/git/pytorch/pytorch/torch/lib/libc10_xpu.so)
frame #3: c10::xpu::getCurrentXPUStream(signed char) + 0x2f (0x7f3070c2c83f in /home/dvrogozh/git/pytorch/pytorch/torch/lib/libc10_xpu.so)
frame #4: <unknown function> + 0x1ca35 (0x7f3070c2ea35 in /home/dvrogozh/git/pytorch/pytorch/torch/lib/libc10_xpu.so)
frame #5: <unknown function> + 0x653f15 (0x7f3083391f15 in /home/dvrogozh/git/pytorch/pytorch/torch/lib/libtorch_python.so)
frame #6: <unknown function> + 0x39e5f2 (0x7f30830dc5f2 in /home/dvrogozh/git/pytorch/pytorch/torch/lib/libtorch_python.so)
<omitting python frames>
frame #20: <unknown function> + 0x29d90 (0x7f308b19bd90 in /lib/x86_64-linux-gnu/libc.so.6)
frame #21: __libc_start_main + 0x80 (0x7f308b19be40 in /lib/x86_64-linux-gnu/libc.so.6)
Aborted (core dumped)
```
with this PR:
```python
>>> import torch
>>> torch.accelerator.current_stream(torch.accelerator.device_count())
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/pt-gpu/4T-4652/guangyey/stock-pytorch/torch/accelerator/__init__.py", line 123, in current_stream
return torch._C._accelerator_getStream(device_index)
RuntimeError: The device index is out of range. It must be in [0, 2), but got 2.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143550
Approved by: https://github.com/EikanWang, https://github.com/dvrogozh, https://github.com/albanD
Certain `cpp_wrapper`-enabled tests were OOM-ing in the CI pipeline, with error messages suggesting that sufficient memory was accessible. This ultimately resulted from an internal memory limitation that was not queryable in the API. This PR adds querying for that limit.
Additionally, the failing tests had incorrect memory availability checks, and are updated with measured memory requirements.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/140620
Approved by: https://github.com/malfet, https://github.com/eqy
ghstack dependencies: #141367
the `__ANDROID__` macro was used as a proxy to check whether compilation is targeting a 32 or 64 bit system, causing build failure on non-android 32 bit linux targets like arm v7.
This modification adjusts the check to fail if and only if int64_t and long and not the same on 64-bit systems, on systems where `sizeof(void*) == 8`
Like I said in the issue #141043 , I'm not sure whether a different `Scalar` constructor should be defined in the 32 bit case. My code does not break but I'm not sure other people's code won't.
Fixes#141043
Pull Request resolved: https://github.com/pytorch/pytorch/pull/141244
Approved by: https://github.com/malfet
Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>
