About Loop Unrolling

[Altera_Forum]

Hi everyone, 

 I am wondering if setting a unroll factor that is not multiple of loop count is OK in Altera OpenCL?  For example:# pragma unroll 10 for( int i = 0; i < n; i++){  ... }  Here n is either a number not divisible by 10 (ex: 128) or not known at compilation time (ex: a variable passed in from host). I know this can be done on CPU or GPU but I am wondering if this will generate incorrect result or cause performance issues on FPGA.  Thanks!

[Altera_Forum]

No, this will work. Depending on whether the iterations are independent or not, you may get performance increase because essentially FPGA hardware of the loop will be replicated 10 times.

[Altera_Forum]

Thank you for the reply! 

 So basically during the last unrolled iteration, only the reminder (in my example 128%10 = 8) number of the FPGA datapath will be used and the rest (again in my example 10 - 8 = 2) of the datapath will be bypassed right?    --- Quote Start ---  No, this will work. Depending on whether the iterations are independent or not, you may get performance increase because essentially FPGA hardware of the loop will be replicated 10 times.  --- Quote End ---  

[Altera_Forum]

--- Quote Start ---  Thank you for the reply!  So basically during the last unrolled iteration, only the reminder (in my example 128%10 = 8) number of the FPGA datapath will be used and the rest (again in my example 10 - 8 = 2) of the datapath will be bypassed right?  --- Quote End ---    Yes, this is correct.

[Altera_Forum]

--- Quote Start ---  Yes, this is correct.  --- Quote End ---    Thank you!