Method configure_mm

src/gpu/cl/operators/ClGemmConv2d.cpp:76–138

ClGemmConv2d::~ClGemmConv2d() = default;

void ClGemmConv2d::configure_mm(const ClCompileContext        &compile_context,
                                const ITensorInfo             *src,
                                ITensorInfo                   *weights,
                                ITensorInfo                   *biases,
                                ITensorInfo                   *dst,
                                const GEMMLowpOutputStageInfo &gemmlowp_output_stage,
                                int                            gemm_3d_depth,
                                const ActivationLayerInfo     &act_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights);
    ARM_COMPUTE_ERROR_THROW_ON(
        validate_mm(src, weights, biases, dst, gemmlowp_output_stage, gemm_3d_depth, _skip_im2col, act_info));

    const GEMMInfo &gemm_info = GEMMInfo(false,                 // is_a_reshaped
                                         false,                 // is_b_reshaped
                                         true,                  // reshape_b_only_on_first_run
                                         gemm_3d_depth,         // depth_output_gemm3d
                                         _skip_im2col,          // reinterpret_input_as_3d
                                         false,                 // retain_internal_weights
                                         gemmlowp_output_stage, // gemmlowp_output_stage
                                         false,                 // fast_math
                                         false,                 // fp_mixed_precision
                                         true,                  // broadcast_bias
                                         act_info               // activation_info
    );

    TensorInfo tmp_src{*src};
    if (_is_quantized)
    {
        // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
        // Extract and negate input and weights offset
        const QuantizationInfo input_quantization_info   = src->quantization_info();
        const QuantizationInfo weights_quantization_info = weights->quantization_info();

        tmp_src.set_quantization_info(
            QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset));
        weights->set_quantization_info(
            QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset));

        _mm_gemmlowp = std::make_unique<ClGemmLowpMatrixMultiplyCore>();
        _mm_gemmlowp->configure(compile_context, &tmp_src, weights, biases, dst, gemm_info);

        // Revert back QuantizatioInfo as weights could be used in other convolution layers
        weights->set_quantization_info(weights_quantization_info);

        auto mm_mem_req = _mm_gemmlowp->workspace();
        for (unsigned int cont = 0; cont < mm_mem_req.size(); ++cont)
        {
            _aux_mem[cont] = mm_mem_req[cont];
        }
    }
    else
    {
        // Configure matrix multiply function
        _mm_gemm = std::make_unique<ClGemm>();
        _mm_gemm->configure(compile_context, &tmp_src, weights, biases, dst, 1.0f, 1.0f, gemm_info);
        auto mm_mem_req = _mm_gemm->workspace();
        for (unsigned int cont = 0; cont < mm_mem_req.size(); ++cont)