To implement machine learning applications in real-time safety-critical systems, we previously introduced a predictable framework named ACETONE. This framework compiles the detailed description of an off-line trained feed-forward deep neural network into an equivalent C code. In this paper, we improve the performance of the generated C code by including gemm-based convolutions in ACETONE. The code incorporating the gemm routines maintains the ACETONE properties of semantics preservation and timing predictability. We compare the proposed method with ACETONE ’s initial version, Keras2c and uTVM on a realistic set of machine learning benchmarks and show that the introduced convolution algorithms allow a trade-off between performance and memory footprint.