--- Quote Start ---
The address alignment problem ("WARNING: NOT using DMA for transfer of size 1024 because of lack of alignment"). When we malloc memory for the host, please align the address first before use in the host program.
--- Quote End ---
Exactly. It's explained in the optimization guide (
http://www.altera.com/literature/hb/opencl-sdk/aocl_optimization_guide.pdf), page 1-7. It says the data should be 64-byte aligned.
In my case, without proper alignement, one of my memory transfers and kernel execution takes ~300secs. With alignment, it gets down to ~3sec. So you better align it. :D
Honored Contributor
