joaomiguelvieira / knn-stuff Goto Github PK

I attempted to use the Sift 1M dataset on your module, taking the first hundred thousand vectors of the Sift dataset as the training set. I noticed some errors in the output data (as shown in the attachment). The results were correct initially, but discrepancies started at index 101563. However, the results became correct again at index 101627.

Do you have any suggestions for me? Thanks.

Also, I would like to know how to solve the dma transfer issue. Dma have a limitation with 0x3FFFFFF data size. And in your paper, you say: the maximum stream length supported by the Xilinx DMA Intellectual Property (IP) limits the training set to a maximum size of N×M≤226×8 bytes (is mitigated by re-engineering the accelerator to receive the training set in multiple streams).

How to do re-engineering the accelerator to receive the training set in multiple streams?

Thanks for the help.

Hello joaomiguelvieira,

I don't have any access to the Zybo or Zedboard FPGA board right now. Is it possible to simulate kNN-STUFF VHDL files using Xilinx Isim or any other HDL simulator?

Thanks