ilongshan / iob-yolo

iob-SoC-Yolo

Repository for the implementation of an object detection application based on the YOLOv3 network, using the iob-soc and iob-versat submodules.

Software Tests

Utility Software

• bootloader: firmware to use as BOOT - waits to receive another firmware via UART
• console: program that runs on the machine, sends the firmware received by the bootloader and handles UART communication

Test Software

• ddr_test: writes N positions of DDR and reads them. Goal: verify correct functioning of the DDR.
• eth_repeat: wait for a message from pc and send it back. Goal: verify Ethernet is working.
• eth_weights: receive an image and weights for the tiny-yolo CNN from pc, each message is sent back to the pc. Goal: verify Ethernet can receive and send back many frames without blocking.
• eth_ddr: same as eth_weights, with the addition that the received data is also saved to DDR. Goal: simultaneous verification of DDR and Ethernet working with lot of data.
• yolo_sw: reception of the input image and weights and full execution of the tiny-yolo CNN in RISCV. Goal: baseline for hardware acceleration of tiny-yolo.
• yolo_sw_full: Full execution of the YOLO pipeline (image resize, tiny-yolo CNN and detections). Goal: baseline for hardware acceleration of full YOLO pipeline.
• versat_test: performs one 3D convolution with 5 input FMs of 5x5 and kernels of 3x3. Goal: verify deep_versat pipeline is working.
• yolo_hw_full: Full execution of YOLO pipeline using versat. Goal: adapt yolo-sw code to be easily accelerated by versat.
• new_versat_test: Computing yolov3-tiny layer 1 output with nSTAGES of nYOLOvect each. Goal: verify new versat is working.
• dma_w_test: Transfering 16 words of 256 bits each from vwrite to DDR via AXI-DMA (only works in simulation). Goal: verify AXI-DMA only-write is working.
• dma_r_test: Transfering 7 words of 256 bits each from DDR to vread via AXI-DMA (only works in simulation). Goal: verify AXI-DMA only-read is working.
• yolo_hw: Execution of yolov3-tiny CNN in new_versat. Goal: accelerate yolo with new versat system.
• pre_cnn: Performs input image resize before CNN. Goal: accelerate pre_cnn with new versat.
• pos_cnn: Performs detection filtering and box/label display after CNN.
• tiny_yolov3: Execution of full yolov3-tiny pipeline (pre-CNN, CNN and pos-CNN) in new_versat. Goal: run all pipeline in 30 fps.

FPGA execution

Python Dependencies

• Pillow
• numpy

Input files

• ../tiny_yolov3_in_x4.bin
• ../tiny_yolov3_out.bin

Target launch order

• make fpga-load on remote machine (or ld-hw locally)
• make ld-sw locally (launches console)
• make ld-eth locally in parallel with make ld-sw (send data and receive results)