Method run

src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp:647–811

}

void CpuGemmLowpMatrixMultiplyCore::run(ITensorPack &tensors)
{
    ARM_COMPUTE_TRACE_EVENT(ARM_COMPUTE_PROF_CAT_CPU, ARM_COMPUTE_PROF_LVL_CPU, "CpuGemmLowpMatrixMultiplyCore::run");
    prepare(tensors);

    auto a        = tensors.get_const_tensor(TensorType::ACL_SRC_0);
    auto b        = tensors.get_const_tensor(TensorType::ACL_SRC_1);
    auto c        = tensors.get_const_tensor(TensorType::ACL_SRC_2);
    auto dst      = tensors.get_tensor(TensorType::ACL_DST);
    auto a_to_use = a;
    auto matrix_a = a;
    auto matrix_b = b;

    CpuAuxTensorHandler vector_sum_col(offset_int_vec(VectorSumCol), _vector_sum_col, tensors, false);
    CpuAuxTensorHandler vector_sum_row(offset_int_vec(VectorSumRow), _vector_sum_row, tensors, false);
    CpuAuxTensorHandler tmp_a(offset_int_vec(TmpA), _tmp_a, tensors, false);
    CpuAuxTensorHandler tmp_b(offset_int_vec(TmpB), _tmp_b, tensors, true);
    CpuAuxTensorHandler mm_result_s32(offset_int_vec(MMResultS32), _mm_result_s32, tensors, false);
    CpuAuxTensorHandler signed_a(offset_int_vec(SignedA), _signed_a, tensors, false);
    CpuAuxTensorHandler signed_output(offset_int_vec(SignedOutput), _signed_output, tensors, false);

    const QuantizationInfo a_qinfo = a->info()->quantization_info();
    const QuantizationInfo b_qinfo = b->info()->quantization_info();

    if (a_qinfo.is_dynamic())
        _a_offset = a_qinfo.uniform().offset;
    if (b_qinfo.is_dynamic())
        _b_offset = b_qinfo.uniform().offset;

    // Convert QASYMM8->QASYMM8_SIGNED
    if (_flip_signedness)
    {
        ITensorPack pack = {{TensorType::ACL_SRC, a}, {TensorType::ACL_DST, signed_a.get()}};
        NEScheduler::get().schedule_op(_convert_to_signed_asymm.get(), Window::DimY, _convert_to_signed_asymm->window(),
                                       pack);
        a_to_use = signed_a.get();
        matrix_a = signed_a.get();
    }

    // Run GEMM
    if (_asm_glue->is_configured())
    {
        ITensorPack asm_glue_tensors = tensors;
        if (is_data_type_quantized_asymmetric(a_to_use->info()->data_type()) &&
            _gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
        {
            asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_0, a_to_use);
            asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_1, b);
            asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_2, c);
            asm_glue_tensors.add_tensor(TensorType::ACL_DST, dst);
        }
        else
        {
            auto output_to_use = (_fuse_output_stage ? mm_result_s32.get() : dst);
            asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_0, a_to_use);
            asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_1, b);
            asm_glue_tensors.add_tensor(TensorType::ACL_DST, output_to_use);
        }