Matrix arithmetic program results do not match software and homegrown IP core

Asked 2 years ago, Updated 2 years ago, 107 views

"Based on the magazine ""Interface"" (January 2019 issue), published by CQ publisher, we are thinking of accelerating YOLOv3 with Ultra96.

"

Currently, we have built a matrix multiplication logic circuit for acceleration and compiled the program that was on the book with gcc to create an executable matrix.When I ran it, it looked like this:

root@xilinx-ultra96-reva-2018_2:/uio_irq_sample#./matrix 
NG: hw 0.000000, sw55551.996094
NG: hw 0.000000, sw 55634.667969
NG: hw 0.000000, sw 55717.335938
               omission
NG: hw 0.000000, sw 2830176.000000
NG: hw 0.000000, sw 2835549.500000
NG: hw 0.000000, sw 2840922.750000
NG: hw 0.000000, sw 2846296.000000
OK: 0, NG: 1024, Total: 1024.

Results did not match like this.
If anyone has done it, please give me some advice.

fpga zynq vivado

2022-09-30 21:44

1 Answers

If the output is all zero, I think it means it's completely wrong.


2022-09-30 21:44

If you have any answers or tips


© 2024 OneMinuteCode. All rights reserved.