This discussion has been locked.
You can no longer post new replies to this discussion. If you have a question you can start a new discussion

ARM_import_memory API is very slow

I want to use clImportMemoryARM API to achieve zero copy between CPU and GPU. 

However, the performance is not what I expected. For a FHD image, it takes 4.4 ms for importing, almost identical to uploading explicitly.

Is this slow performance expected? I am using Mali G72 GPU.

Thanks,

-Shouwen

Parents
  • Hi Shouwen,

    Thanks for the feedback. Are you using the dma_buf or host memory path? Are multiple kernels using the same imported buffer as part of a single flush? What DDK version are you using?

    Regards,

    Kévin

Reply
  • Hi Shouwen,

    Thanks for the feedback. Are you using the dma_buf or host memory path? Are multiple kernels using the same imported buffer as part of a single flush? What DDK version are you using?

    Regards,

    Kévin

Children