Method configure

src/gpu/cl/kernels/ClMulKernel.cpp:135–291

}

void ClMulKernel::configure(const CLCompileContext    &compile_context,
                            ITensorInfo               *src1,
                            ITensorInfo               *src2,
                            ITensorInfo               *dst,
                            float                      scale,
                            ConvertPolicy              overflow_policy,
                            RoundingPolicy             rounding_policy,
                            const ActivationLayerInfo &act_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst);
    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src1, src2, dst, scale, overflow_policy, rounding_policy, act_info));

    auto padding_info = get_padding_info({src1, src2, dst});

    const TensorShape &out_shape = TensorShape::broadcast_shape(src1->tensor_shape(), src2->tensor_shape());
    auto_init_if_empty(*dst, src1->clone()->set_tensor_shape(out_shape));

    int scale_int = -1;
    // Extract sign, exponent and mantissa
    int   exponent            = 0;
    float normalized_mantissa = std::frexp(scale, &exponent);
    // Use int scaling if factor is equal to 1/2^n for 0 <= n <= 15
    // frexp returns 0.5 as mantissa which means that the exponent will be in the range of -1 <= e <= 14
    // Moreover, it will be negative as we deal with 1/2^n
    if ((normalized_mantissa == 0.5f) && (-14 <= exponent) && (exponent <= 1))
    {
        // Store the positive exponent. We know that we compute 1/2^n
        // Additionally we need to subtract 1 to compensate that frexp used a mantissa of 0.5
        scale_int = std::abs(exponent - 1);
    }

    std::string acc_type;
    // Check if it has float src and dst
    if (is_data_type_float(src1->data_type()) || is_data_type_float(src2->data_type()))
    {
        scale_int = -1;
        acc_type  = (src1->data_type() == DataType::F32 || src2->data_type() == DataType::F32) ? "float" : "half";
    }
    else
    {
        if (src1->element_size() == 4 || src2->element_size() == 4)
        {
            // use 64 bit accumulator for 32-bit input
            acc_type = "long";
        }
        else if (src1->element_size() == 2 || src2->element_size() == 2)
        {
            // Use 32-bit accumulator for 16-bit input
            acc_type = "int";
        }
        else
        {
            // Use 16-bit accumulator for 8-bit input
            acc_type = "ushort";
        }
    }

    const bool         is_quantized      = is_data_type_quantized(src1->data_type());