Suffix implementation name to GPU kernel names #20
Conversation
| @@ -114,7 +114,8 @@ def template_args(self): | |||
| dry_runs=self.dry_runs, | |||
| timers=self.timers, | |||
| strides=self.strides, | |||
| index_type=self.index_type) | |||
| index_type=self.index_type, | |||
| implementation_name=self.template_file().partition('.')[0]) | |||
There was a problem hiding this comment.
I guess the implementation_name is used for easier analysis with NCU? But if we for example run using two different domain sizes, template_file is the same, so this does not really help in general, or does it?
Wouldn’t it be better to either just add the hash of all parameters (which is completely unreadable, but enforces a unique name for unique parameters) or create a human-readable implementation_name that includes all relevant parameters?
There was a problem hiding this comment.
But if we for example run using two different domain sizes, template_file is the same, so this does not really help in general, or does it?
It's true that it only partly helps. IIRC one can infer the domain size by the grid size reported by NCU though?
When I was looking into this, distinguishing the kernel names was helpful enough since I could figure out which size was each report then. For different sizes one can also launch the benchmarks separately and create separate ncu-rep files that encode the domain sizes as well.
Tbh I don't think that a hash adds much value towards the desired purpose.
There was a problem hiding this comment.
Running just a single kernel at a time was what I always did. But then we don’t need any change from now because you either know exactly what you have run or otherwise you have a problem because of missing information, right?
So I am for either completely fixing the problem (if it is one) or completely ignoring it.
There was a problem hiding this comment.
If we run each kernel with separate commands then we can avoid those issues and simplify things. I was just using the sbench-<arch>-collection to run the benchmarks and collect results in batches.
It would still be useful to have a way to inspect and distinguish the generated code for each implementation by either encoding the implementation name in the function name or the filename saved to disk.
--generate-line-infocompilation option fornvccto improve source code investigation inNSight Compute. (To be used with--import-source yesinncu)