[CUDA][HIP] Use device to get native context

[hdelan]

Since oneapi-src/unified-runtime#999 it is no longer valid to get the native context from the SYCL context on a multi GPU system. The get native func for contexts has been deprecated for this reason. See intel/llvm#10975

Similar ticket: uxlfoundation/oneDNN#1765

[jinz2014]

reading your changes, I have a question.

For example,

auto cudaDevice = sycl::get_nativesycl::backend::ext_oneapi_cuda(queue.get_device());

Is the type of cudaDevice "CUdevice" ?

[hdelan]

Hi @jinz2014 yes you are correct!

[FMarno]

cufft_run.txt
All the DFT changes look good to me and I've run the DFT tests successfully.
I'd like to see test logs for the other backends before I approve.

[hdelan]

AMD tests for lapack and blas all passing:
test_amd.txt

8 lapack nvidia test failing on GTX1050 but these tests are also failing on develop branch:
test_cuda_lapack.txt

Nvidia blas tests passing
test_cuda_blas.txt

[muhammad-tanvir-1211]

I see all the buffer tests failing for the rocblas backend with PI_ERROR_INVALID_OPERATION.

Logs:
PR_425.txt

The failures are not because of the changes in this PR, but rather a recent change in the compiler. All these tests are expected to pass once oneapi-src/unified-runtime#1226 and intel/llvm#12297 are merged.

[Rbiessy]

Would you be able to attach test logs again? Also does this change compile with the 2024.0 icpx release?

[Rbiessy]

Should we expect a performance cost from this change? From what I understand cuCtxSetCurrent was expected to be called only once before, assuming the context active was not changed outside of oneMKL.
This constructor is called once before each calls to blas functions so I am wary that the cost may add up.

[hdelan]

The cost of cuDevicePrimaryContextRetain is minimal once the primary context is not being initialized for the first time, which it should not be here. A simple benchmark like this:

for (int i = 0; i < NUM_ITERATIONS; i++) {
    CHECK(cuDevicePrimaryCtxRetain(&context, device));
    CHECK(cuDevicePrimaryCtxRelease(device));
}

Gives 32ns per loop, so calls to these funcs are almost free.

Using setup:

$ nvidia-smi
Mon Mar 25 16:27:32 2024       
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 535.104.12             Driver Version: 535.104.12   CUDA Version: 12.2     |
|-----------------------------------------+----------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
|                                         |                      |               MIG M. |
|=========================================+======================+======================|
|   0  NVIDIA GeForce GTX 1050 Ti     On  | 00000000:01:00.0 Off |                  N/A |
| 31%   24C    P8              N/A /  75W |     14MiB /  4096MiB |      0%      Default |
|                                         |                      |                  N/A |
+-----------------------------------------+----------------------+----------------------+
                                                                                         

[hjabird]

Generally LGTM. I've tested the rocFFT and cuFFT backends with DPC++ 2024.0's icpx.
It would be good to see logs again after the rebase as Rbiessy suggests.

[hdelan]

Some test results:

CUDA

gtx1050.txt
Some failures due to precision also present on develop branch.

HIP

gfx90a_oneMKL_test.txt
Test failures in HIP are also present on the develop branch:
gfx90a_oneMKL_test_develop_branch.txt

I am not sure how to build/run the FFT tests. Are there some build/test instructions that I can follow?

[hdelan]

In terms of building with icpx 2024.0.0 for CUDA. I am getting a segfault at linking with develop branch.

Fixed. LD_LIBRARY_PATH problems -_-

I can successfully build this branch with icpx 2024.0.2 for CUDA

[Rbiessy]

Thanks a lot @hdelan ! The instructions are here but need to be improved.

The short answer is that you should just need to add -DENABLE_CUFFT_BACKEND=True -DENABLE_ROCFFT_BACKEND=True to also test the DFT domain with the native CUDA and HIP backends.
If you are explicitly setting -DTARGET_DOMAINS in your CMake command you will also need to append dft to the list, otherwise it will be enabled by default.
If you don't want to build and test the other domains again you can use -DTARGET_DOMAINS=dft.

[hdelan]

Thanks @Rbiessy !

Building rocFFT is broken for me but this PR does not touch that code. Building with cuFFT is OK. Here is updated tests for all oneMKL for CUDA including cuBLAS, cuFFT, cuRAND, cuSOLVER:

gtx1050.txt

[ericlars]

Thanks! LGTM

[Rbiessy]

Thanks for the review. Let me know @lhuot or @mmeterel if you need more time, otherwise I will go ahead and merge this on Monday.