- CPU compatible with
ARMv7-AISA - or
ARMv7-MISA (forsimulate.ino)
gcc -mfpu=neon -O3 [bitcount.c | int8.c]
bitcount.c : NUM_RANDOM * 32 * NUM_REPEAT MACsint8.c : NUM_RANDOM * 4 * NUM_REPEAT MACs
BNN : NUM_RANDOM * 32 * NUM_REPEAT MACsINT8 : NUM_RANDOM * NUM_REPEAT MACs
On Raspberry Pi 2, both test spends about 6 seconds
BNN : ~715 MMACSINT8 : ~89.5 MMACS
On Arduino Nano 33 BLE,
BNN : ~67.1 MMACSINT8 : ~6.87 MMACS