Method configure_mm

src/cpu/operators/CpuFullyConnected.cpp:157–199

CpuFullyConnected::~CpuFullyConnected() = default;

void CpuFullyConnected::configure_mm(const ITensorInfo         *src,
                                     const ITensorInfo         *weights,
                                     const ITensorInfo         *biases,
                                     ITensorInfo               *dst,
                                     const ActivationLayerInfo &act)
{
    if (_is_quantized_asymmetric)
    {
        // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
        // Extract and negate src and weights offset
        const QuantizationInfo src_quantization_info(src->quantization_info().uniform().scale,
                                                     -src->quantization_info().uniform().offset);
        const QuantizationInfo weights_quantization_info(weights->quantization_info().uniform().scale,
                                                         -weights->quantization_info().uniform().offset);

        TensorInfo src_info     = src->clone()->set_quantization_info(src_quantization_info);
        TensorInfo weights_info = weights->clone()->set_quantization_info(weights_quantization_info);

        // Configure gemmlowp function and output stage for asymmetric quantized types
        GEMMLowpOutputStageInfo gemmlowp_output_stage_info;
        const Status            status =
            get_gemmlowp_output_stage_info(&src_info, &weights_info, dst, act, gemmlowp_output_stage_info);
        ARM_COMPUTE_ERROR_ON(status.error_code() != ErrorCode::OK);

        GEMMInfo gemm_info;
        gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info);
        gemm_info.set_activation_info(act);
        gemm_info.set_fast_math(_enable_fast_math);
        _mm_gemmlowp = std::make_unique<CpuGemmLowpMatrixMultiplyCore>();
        _mm_gemmlowp->configure(&src_info, &weights_info, biases, dst, gemm_info);
    }
    else
    {
        // Configure matrix multiply kernel
        GEMMInfo gemm_info;
        gemm_info.set_activation_info(act);
        gemm_info.set_fast_math(_enable_fast_math);
        gemm_info.set_fixed_format(_fixed_format);
        gemm_info.set_weight_format(_weight_format);
        _mm_gemm = std::make_unique<CpuGemm>();
        _mm_gemm->configure(src, weights, biases, dst, 1.f, 1.0f, gemm_info);
    }
}

void CpuFullyConnected::configure_conv_fc(const ITensorInfo         *src,
                                          const ITensorInfo         *weights,