add function ids to the spec #310
Conversation
|
This is a bit janky, so I'm open to other ideas. In SYCL they have an autogenerated enum. Right now in #300 I'm just using a loop counter inside of the template. Both of those implicitly depend on the order of functions in the spec. |
XPTI function tracing requires "a stable ABI function id" that is used to programmatically identify a traced entry point. This patch adds a requirement for functions to have an unique id, populates that field for existing functions and creates an enum with each function id. It also adds a script that automatically adds those ids if they are missing.
| infunc = False | ||
| hasid = False | ||
|
|
||
| lines = f.readlines() |
There was a problem hiding this comment.
Storing the whole file is unnecessary, try having the following loop this way:
for line in f
And later when checking for a newline at EOF:
f.seek(-1,2)
if f.read() != '\n'
There was a problem hiding this comment.
That won't work because the file is being overwritten. I'd have to buffer the output lines, so it would ultimately come down to the same thing perf- or memory-wise (not that it matters whether this executes in 0.05s or 0.1s... :P).
| UR_FUNCTION_ENQUEUE_MEM_IMAGE_WRITE = 28, | ||
| UR_FUNCTION_ENQUEUE_MEM_IMAGE_COPY = 29, | ||
| UR_FUNCTION_ENQUEUE_MEM_BUFFER_MAP = 31, | ||
| UR_FUNCTION_ENQUEUE_MEM_UNMAP = 33, |
There was a problem hiding this comment.
Missing 30 and 32, isn't this enum autogenerated?
There was a problem hiding this comment.
I had to rebase this many times now, which is very fiddly... In general, ids don't have to be monotonically increasing.
There was a problem hiding this comment.
Turns out it works fine, entry points with ids 30 and 32 are disabled, see #50.
In Khronos specifications they have the concept of a registry, this is used for hardware vendors to reserve ranges of enumeration values for vendor extensions among other things. I wonder if it would make sense to have a |
I like that idea. The script would be simpler, and such a registry would keep the autogenerated id separate from the actual spec. On the other hand, accessing the id variable in templates will be a little bit more involved, but nothing a helper function can't solve. I'll sleep on it. |
|
It needs to be stable across versions as if it's part of the ABI. XPTI-based tools will target this and shouldn't break when, e.g., a new function is added in the middle of some yml file. That's why this version adds the ids to the spec, and the generate script respects existing ids. I started reworking this in line with your suggestion for ids to be in a separate file. I'll probably close this PR and open a new one. |
|
Closing in favor of a different approach in #343 |
XPTI function tracing requires "a stable ABI function id" that
is used to programmatically identify a traced entry point.
This patch adds a requirement for functions to have a unique id,
populates that field for existing functions, and creates an enum
with each function id. It also adds a script that automatically
adds those ids if they are missing.