Intel Opencl kernel compiling time

[ADua0]

I am trying to compile one of my kernel, but it is stuck at routing for more than 10 hours, but I have not even utilized even 50 % of the FPGA I have added my resource utilization report image and here is the message I got after 10 hours. Can anyone tell me what could be the possible reason for that ?

[Kenny_Tan]

This usually cause by high utilization of the design for the particular are in the chip planner. Design will face this if your frequencies requirement is high. Can you try to follow the optimization base on the document https://www.intel.com/content/dam/www/programmable/us/en/pdfs/literature/hb/opencl-sdk/archives/aocl-best-practices-guide-15.0.pdf

[ADua0]

Hi thanks for the reply. Can you suggest like which page you are referring to in the above document

[Kenny_Tan]

I would suggest you read all. Also, try to run the design longer to see if it fit.

[HRZ]

As it seems from the report, you have significant routing congestion in your design which is lengthening the routing process and there is no guarantee routing will ever succeed, either. Make sure you are using the latest BSP and compiler that supports your board and try to simplify your design. Specifically, try to avoid barrel shifters and large register-based buffers. Also, this is the latest version of the Best Practices Guide:

https://www.intel.com/content/dam/www/programmable/us/en/pdfs/literature/hb/opencl-sdk/aocl-best-practices-guide.pdf

P.S. Remember that the area estimation is not necessarily accurate and actual utilization could be a lot higher (or lower) than the estimation.

[ADua0]

I am using 18.0 Bsp version. I have lot of register , also I have some register structure such as "float __attribute__((register)) sum[4][2][16] ( ", although I am not able to understand how this 3d register is possible , do you know what that means ? could that also be a reason for long compile time

[HRZ]

That buffer is not too large but if you have on-chip buffers that are by default implemented using Block RAMs by the compiler due to their size and you force implementation using registers by adding the "__attribute__((register))" attribute, then indeed that could cause routing congestion problems. Having large register-based buffers with a lot of reads/writes will result in high fan-in/fan-out and routing congestion in the area the buffer is implemented.

[Kenny_Tan]

Still working on this?