simulate BNN and INT8 quantized models

requires

• CPU compatible with ARMv7-A ISA
• or ARMv7-M ISA (for simulate.ino)

how to compile

gcc -mfpu=neon -O3 [bitcount.c | int8.c]

test computes

ARMv7-A

• bitcount.c : NUM_RANDOM * 32 * NUM_REPEAT MACs
• int8.c : NUM_RANDOM * 4 * NUM_REPEAT MACs

ARMv7-M (simulate.ino)

• BNN : NUM_RANDOM * 32 * NUM_REPEAT MACs
• INT8 : NUM_RANDOM * NUM_REPEAT MACs

test result

On Raspberry Pi 2, both test spends about 6 seconds

• BNN : ~715 MMACS
• INT8 : ~89.5 MMACS

On Arduino Nano 33 BLE,

• BNN : ~67.1 MMACS
• INT8 : ~6.87 MMACS