The result is right in emulator mode, but is wrong in FPGA

--- Quote Start ---

This could be anything from a race condition that is not properly emulated by the emulator to a bug in the compiler. Pretty much the only way to debug code on the FPGA is to use printf from inside the kernel, and that is only if the kernel does not deadlock.

--- Quote End ---

Thanks for your reply!

My kernels are all task kernel. That is to say mutiple kernels read/write one global variable ?

--- Quote Start ---

My kernels are all task kernel. That is to say mutiple kernels read/write one global variable ?

--- Quote End ---

Yes, if you have multiple kernels running in parallel which read from or write to the same global memory buffer, and read/write order will affect your output, you will likely get incorrect results. Trying to synchronize such kernels using channels will NOT work.

--- Quote Start ---

Yes, if you have multiple kernels running in parallel which read from or write to the same global memory buffer, and read/write order will affect your output, you will likely get incorrect results. Trying to synchronize such kernels using channels will NOT work.

--- Quote End ---

But my code have no global variable that multiple kernels write to it. Is the problem of channel? In my code, many channel read/write have conditions. Or the problem of local memory access?? My local buffer is doubble-buffer, in a loop, there is read and write to it, but their address is different.

HRZ,

In response to your comment:

--- Quote Start ---

Yes, if you have multiple kernels running in parallel which read from or write to the same global memory buffer, and read/write order will affect your output, you will likely get incorrect results. Trying to synchronize such kernels using channels will NOT work.

--- Quote End ---

I have not explored this design pattern yet; but is this due to a lack of cache coherence (in the LSUs that are automatically inferred)? Did you ever bring this up with Altera or get a response?

--- Quote Start ---

I have not explored this design pattern yet; but is this due to a lack of cache coherence (in the LSUs that are automatically inferred)? Did you ever bring this up with Altera or get a response?

--- Quote End ---

This is not a bug, this is part of the OpenCL specification. Global memory consistency is only guaranteed at the end of kernel execution and hence, one should not expect consistency when kernels running in parallel are reading from/writing to the same global address. Channels are meant to be used for passing data between kernels, not synchronizing global memory load/stores.

--- Quote Start ---

But my code have no global variable that multiple kernels write to it. Is the problem of channel? In my code, many channel read/write have conditions. Or the problem of local memory access?? My local buffer is doubble-buffer, in a loop, there is read and write to it, but their address is different.

--- Quote End ---

Channel ordering could be the problem, try to print it out. My case was the emulator had the wrong ordering but FPGA was correct.