[SYCL][FPGA] Add support for two new function attributes #3441
Conversation
This patch implements the support for two new FPGA function attributes:
1. disable_loop_pipelining
2. initiation_interval
Both attributes alreday exist for loops as statement attributes.
Function attribute "initiation_interval" takes a single unsigned integer argument.
Syntax:
[[intel::initiation_interval(n)]] void foo() {}
The argument can be constexpr. The parameter n is a template parameter.
The LLVM IR representation will be function metadata:
!initiation_interval !0
!0 = !{!i32 n}
Function attribute "disable_loop_pipelining" takes no arguments.
Syntax:
[[intel::disable_loop_pipelining]] void foo() {}
The LLVM IR representation will be function metadata:
!disable_loop_pipelining !0
!0 = !{!i32 1}
Both attributes are only valid in device code and do not get propagated to the caller.
The function attributes apply to a lambda function, or function definition.
Signed-off-by: Soumi Manna <[email protected]>
Signed-off-by: Soumi Manna <[email protected]>
Signed-off-by: Soumi Manna <[email protected]>
Signed-off-by: Soumi Manna <[email protected]>
Signed-off-by: Soumi Manna <[email protected]>
| let Args = [ExprArgument<"IntervalExpr", /*opt*/1>]; | ||
| let LangOpts = [SYCLIsDevice, SilentlyIgnoreSYCLIsHost]; | ||
| let HasCustomTypeTransform = 1; | ||
| let Documentation = [SYCLIntelFPGAInitiationIntervalAttrDocs]; | ||
| let SupportsNonconformingLambdaSyntax = 1; |
There was a problem hiding this comment.
i checked with feature request author, the attribute needs to apply in this position.
There was a problem hiding this comment.
Thanks for checking with the authors on these! Just to check -- when you ask their opinion, are you letting them know that a "yes" answer means their attribute isn't valid C++ code? (I have a suspicion that not everyone is thinking about the language design of these attributes so much as what immediate problem they think they're solving. Being asked if they want to add something that's not valid C++ may help them think a bit deeper about the problem.)
Signed-off-by: Soumi Manna <[email protected]>
| let Args = [ExprArgument<"IntervalExpr", /*opt*/1>]; | ||
| let LangOpts = [SYCLIsDevice, SilentlyIgnoreSYCLIsHost]; | ||
| let HasCustomTypeTransform = 1; | ||
| let Documentation = [SYCLIntelFPGAInitiationIntervalAttrDocs]; | ||
| let SupportsNonconformingLambdaSyntax = 1; |
There was a problem hiding this comment.
Thanks for checking with the authors on these! Just to check -- when you ask their opinion, are you letting them know that a "yes" answer means their attribute isn't valid C++ code? (I have a suspicion that not everyone is thinking about the language design of these attributes so much as what immediate problem they think they're solving. Being asked if they want to add something that's not valid C++ may help them think a bit deeper about the problem.)
Signed-off-by: Soumi Manna <[email protected]>
|
This generally LGTM but it's still marked as a draft -- are you planning more changes? |
smanna12
requested review from
elizabethandrews and
premanandrao
as code owners
No, the feature is complete. |
Signed-off-by: Soumi Manna <[email protected]>
Signed-off-by: Soumi Manna <[email protected]>
|
Thanks for the reviews. |
…disable_loop_pipelining attribute The support for disable_loop_pipelining was implemented on intel#3441 This patch 1. removes FIXME 2. Updtaes document 3. adds tests 4. fixes typos Signed-off-by: Soumi Manna <[email protected]>
This patch implements the support for two new FPGA function attributes:
Both attributes alreday exist for loops as statement attributes.
Function attribute "initiation_interval" takes a single unsigned integer argument.
Syntax:
[[intel::initiation_interval(n)]] void foo() {}
The argument can be constexpr. The parameter n is a template parameter.
The LLVM IR representation will be function metadata:
!initiation_interval !0
!0 = !{!i32 n}
Function attribute "disable_loop_pipelining" takes no arguments.
Syntax:
[[intel::disable_loop_pipelining]] void foo() {}
The LLVM IR representation will be function metadata:
!disable_loop_pipelining !0
!0 = !{!i32 1}
Both attributes are only valid in device code and do not get propagated to the caller.
The function attributes apply to a lambda function, or function definition.
Signed-off-by: Soumi Manna [email protected]