An example using cuBLAS library in alpaka

[mehmetyusufoglu]

This example uses cuBLAS library for matrix multiplication by using allocated alpaka buffers and alpaka queue. Another example is using rocBLAS library.

Cmake file is still to be changed for CI to not fail for other backends.

    auto alpakaStream = alpaka::getNativeHandle(queue);

    // cuBLAS setup from alpaka stream
    cublasHandle_t cublasHandle;
    cublasCreate(&cublasHandle);
    cublasSetStream(cublasHandle, alpakaStream);

    // Perform matrix multiplication: C = A * B
    float alpha = 1.0f, beta = 0.0f; // Set beta to 0.0f to overwrite C
    cublasSgemm(
        cublasHandle,
        CUBLAS_OP_N,
        CUBLAS_OP_N, // No transpose for A and B
        M,
        N,
        K, // Dimensions: C = A * B
        &alpha,
        alpaka::getPtrNative(bufDevA), M ...
    );
    alpaka::wait(queue); // Wait for multiplication to complete```

[psychocoderHPC]

better use std::data() instead of alpaka::getPtrNative()

[mehmetyusufoglu]

ok, done, thanks.

[psychocoderHPC]

Could you please use the CUDA tag and derive the ACC from the tag? THis will reduce the work as soon as we refactor the accelerators.

[mehmetyusufoglu]

I used the standard tags like other examples, but prevented the configuration of this example at cmake if ACC_CUDA_ONLY cmake variable is not set.

[mehmetyusufoglu]

Anyway i used cuda tag as you suggested. This example could have a direct main rather than using ExampleTags since only will run with single backend.

using Acc = alpaka::TagToAcc<alpaka::TagGpuCudaRt, Dim1D, Idx>;

[SimeonEhrig]

@psychocoderHPC I agree with @mehmetyusufoglu . It does not make sense to use the same template, like in the other examples. The code can be only used with the CUDA backend. Therefore we need no complicated iteration over the enabled tags.

[mehmetyusufoglu]

This PR is closed because the changes is added to #2433 since 2 PR's will share the same directory in examples directory. Rather than being 2 separate directories in examples.