Method configure

src/cpu/operators/CpuMatMul.cpp:180–286

}

void CpuMatMul::configure(ITensorInfo               *lhs,
                          ITensorInfo               *rhs,
                          ITensorInfo               *dst,
                          const MatMulInfo          &info,
                          const CpuMatMulSettings   &settings,
                          const ActivationLayerInfo &act_info)
{
    ARM_COMPUTE_TRACE_EVENT(ARM_COMPUTE_PROF_CAT_CPU, ARM_COMPUTE_PROF_LVL_CPU, "CpuMatMul::configure");
    ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst);
    ARM_COMPUTE_LOG_PARAMS(lhs, rhs, dst, info, settings);
    ARM_COMPUTE_ERROR_THROW_ON(CpuMatMul::validate(lhs, rhs, dst, info, settings));

    _adj_lhs   = info.adj_lhs();
    _adj_rhs   = info.adj_rhs();
    _fast_math = settings.fast_math();

    // 1. Create and reshape tensors
    // ------------------------------------------------------
    // a. Clone TensorInfo to prevent changing original tensor values during setup
    // b. Change shape of lhs/dst to [x, y, 1, collapsed(z)] to match assembly kernel configuration
    // c. For rhs collapse all dimensions larger than 3 to z dimension
    TensorInfo lhs_to_use = *lhs->clone();
    TensorInfo dst_to_use = *dst->clone();
    TensorInfo rhs_to_use = *rhs->clone();

    // Save starting shape of tensors
    _original_lhs_shape = lhs_to_use.tensor_shape();
    _original_dst_shape = dst_to_use.tensor_shape();
    _original_rhs_shape = rhs_to_use.tensor_shape();

    // Reshape lhs for use with assembly kernels.
    lhs_to_use.set_tensor_shape(
        TensorShape(_original_lhs_shape.x(), _original_lhs_shape.y(), 1, _original_lhs_shape.collapsed_from(2).z()));
    dst_to_use.set_tensor_shape(
        TensorShape(_original_dst_shape.x(), _original_dst_shape.y(), 1, _original_dst_shape.collapsed_from(2).z()));
    rhs_to_use.set_tensor_shape(_original_rhs_shape.collapsed_from(2));

    // 2.  Configuration for transpose of lhs/rhs
    // ------------------------------------------------------
    // Initialise transposed TensorInfo class for aux tensors (intermediary tensors)
    if (_adj_lhs)
    {
        // Setup transpose LHS
        _transpose_kernel_lhs = std::make_unique<cpu::kernels::CpuTransposeKernel>();
        _transpose_kernel_lhs->configure(&lhs_to_use, &_lhs_transposed);

        _aux_mem[TransposeLHS] = MemoryInfo(offset_int_vec(TransposeLHS), MemoryLifetime::Temporary, lhs->total_size());
    }

    if (_adj_rhs)
    {
        // Setup transpose RHS
        _transpose_kernel_rhs = std::make_unique<cpu::kernels::CpuTransposeKernel>();
        _transpose_kernel_rhs->configure(&rhs_to_use, &_rhs_transposed);

        _aux_mem[TransposeRHS] = MemoryInfo(offset_int_vec(TransposeRHS), MemoryLifetime::Temporary, rhs->total_size());
    }