Method run_op

src/gpu/cl/kernels/ClMatMulNativeMMULKernel.cpp:208–240

}

void ClMatMulNativeMMULKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
{
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);

    const ICLTensor *lhs =
        utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
    const ICLTensor *rhs =
        utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
    const ICLTensor *bias = utils::cast::polymorphic_downcast<const ICLTensor *>(
        tensors.get_const_tensor(TensorType::ACL_SRC_2)); // nullptr if bias is not present
    ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
    ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst);
    ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst);
    unsigned int idx = 0;

    add_3d_tensor_nhw_argument(idx, lhs);
    add_3d_tensor_nhw_argument(idx, rhs);
    if (bias != nullptr)
    {
        add_3d_tensor_nhw_argument(idx, bias);
    }
    add_3d_tensor_nhw_argument(idx, dst);

    // Pass m and n at runtime as signed ints, to ensure results of any subtractions they could be operand in, would still be signed.
    _kernel.setArg<cl_int>(idx++, _m);
    _kernel.setArg<cl_int>(idx++, _n);
    _kernel.setArg<cl_int>(idx++, _k);

    // LWS_x should be multiple of 16 at least. (32, 2) has been chosen to have more work-items on a single core
    // LWS also enforces the order of execution of the work items which improves cache utilization
    enqueue(queue, *this, window, cl::NDRange(32, 2), false);
}

} // namespace kernels
} // namespace opencl