hipBLAS is a Basic Linear Algebra Subprograms (BLAS) marshalling library, with multiple supported backends. It sits between the application and a 'worker' BLAS library, marshalling inputs into the backend library and marshalling results back to the application. hipBLAS exports an interface that does not require the client to change, regardless of the chosen backend. Currently, hipBLAS supports rocBLAS and cuBLAS as backends.

For a detailed description of the hipBLAS library, its implemented routines, the installation process and user guide, see the hipBLAS Documentation.

hipBLAS requires either rocBLAS + rocSOLVER or cuBLAS APIs for BLAS implementation. For more information dependent roc* libraries see rocBLAS documentation, and rocSolver documentation.

To download the hipBLAS source code, use the below command to clone the repository

    git clone https://github.com/ROCmSoftwarePlatform/hipBLAS.git

hipBLAS requires specific version of rocBLAS & rocSOLVER to be installed on the system. The required rocBLAS and rocSOLVER versions to build hipBLAS is provided here.

Once the dependent libraries are installed, the following command will build hipBLAS and install to /opt/rocm/hipblas:

    cd hipblas
    ./install.sh -i

The hipBLAS interface is compatible with rocBLAS and cuBLAS-v2 APIs. Porting a CUDA application which originally calls the cuBLAS API to an application calling hipBLAS API should be relatively straightforward. For example, the hipBLAS SGEMV interface is

hipblasStatus_t
hipblasSgemv( hipblasHandle_t handle,
              hipblasOperation_t trans,
              int m, int n, const float *alpha,
              const float *A, int lda,
              const float *x, int incx, const float *beta,
              float *y, int incy );

hipBLAS GEMM can process matrices in batches with regular strides. There are several permutations of these API's, the following is an example that takes everything

hipblasStatus_t
hipblasSgemmStridedBatched( hipblasHandle_t handle,
              hipblasOperation_t transa, hipblasOperation_t transb,
              int m, int n, int k, const float *alpha,
              const float *A, int lda, long long bsa,
              const float *B, int ldb, long long bsb, const float *beta,
              float *C, int ldc, long long bsc,
              int batchCount);

hipBLAS assumes matrices A and vectors x, y are allocated in GPU memory space filled with data. Users are responsible for copying data from/to the host and device memory.

For a complete list of all supported functions, see the hipBLAS user guide and hipBLAS functions.