[SYCL] Investigate arbitrary attribute bounds for FPGA attributes

[elizabethandrews]

Are these bounds required? Can we just test for unsigned int values?

See comment thread #2877 (comment) for details

[MrSidims]

@mohammadfawaz @jessicadavies-intel @GarveyJoe to address this question.

From what I remember now - we definitly need such limitation for max_global_work_dim kernel attribute (which has a natural limitation of '3').

[mohammadfawaz]

Many attributes have sane limitations such as force_pow2_depth (which can be either 0 or 1). Other memory attributes, such as bankwidth, need to have a lower bound of 1. However, an upper bound of 1024*1024 seems quite arbitrary. I tried searching for where this number comes from, but I couldn't find anything. I am okay with removing the restriction, as long as:

1. We make sure that Clang doesn't care if we do so. I know the backend will be okay.
2. We keep all other bounds that actually make sense.

[elizabethandrews]

Thanks for the inputs!

I took a look at the attributes and identified these with what looks like an arbitrary upper bound

0 - 1024*1024
SYCLIntelSchedulerTargetFmaxMhz 
SYCLIntelLoopFuse
IntelFPGAPrivateCopies
IntelFPGABankBits

1 - 1024*1024
IntelFPGABankWidth
IntelFPGANumBanks
IntelFPGAMaxReplicates

I can keep the lower bounds but remove the upper bounds for these.

I know the backend will be okay.

@mohammadfawaz are you confirming the backend doesn't care? Or is this something I need to verify with someone?

[mohammadfawaz]

Thanks for the inputs!

I took a look at the attributes and identified these with what looks like an arbitrary upper bound

0 - 1024*1024
SYCLIntelSchedulerTargetFmaxMhz 
SYCLIntelLoopFuse
IntelFPGAPrivateCopies
IntelFPGABankBits

1 - 1024*1024
IntelFPGABankWidth
IntelFPGANumBanks
IntelFPGAMaxReplicates

I can keep the lower bounds but remove the upper bounds for these.

I know the backend will be okay.

@mohammadfawaz are you confirming the backend doesn't care? Or is this something I need to verify with someone?

Already verified that the backend doesn't care. Thanks!