Opis: This thesis explores the possibility of running neural networks on microcontrollers and how
to optimize their performance using instruction set extensions. Microcontrollers are seen as
too weak to run neural networks. We challenge this view and show that stripped-down neural
networks can run and be useful for some applications. We used an open-source microcontroller
called PULPino to run our neural network. The benefit of various instructions and optimizations
for minimizing energy consumption to run deep learning algorithms was evaluated. Hardware
loops, loop unrolling, and the dot-product unit were implemented and tested.
We developed an FPGA-based testing system to evaluate our hardware. We also developed a
deep learning library and a test neural network for our hardware. We wrote two versions of
the deep learning library. One version is the reference code, and the other is the optimized
code that uses the dot product unit. Using the testing system, we tested the performance of
the two versions. The synthesis was run to determine the power and energy consumption. We
also tried out various optimizations to see if the performance could be improved.
Using instruction set extensions and algorithmic optimizations we reduced the clock cycle count by 72% for the convolutional layers and by 78% for fully-connected layers. This reduced
power consumption by 73%. We compare our results with related research.Ključne besede: deep learning, embedded system, instruction set, RISC-VObjavljeno: 03.11.2020; Ogledov: 417; Prenosov: 83 Celotno besedilo (2,69 MB)