Add separate function to apply shared memory indexing #129
Conversation
| write.index = remove_global_indexing( | ||
| read.index, tilingConstraints | ||
| ) | ||
| write.index = deepcopy(read.index) |
There was a problem hiding this comment.
Why do we need deepcopy?
There was a problem hiding this comment.
So if we just assign read.index to write.index, then when we modify write.index in the corrections pass it will also affect read.index. But did you have a better way to handle this?
There was a problem hiding this comment.
But remove_global_indexing is already creating a copy, it shouldn't affect original index object.
There was a problem hiding this comment.
Yes, that's right. remove_global_indexing is creating a copy, but that copy will be assigned to write.index and read.index which is not what we want. If you remove the deepcopy, two of the lit tests fail.
There was a problem hiding this comment.
okay I was able to figure this out. I have removed the deep copy. @Hardcode84 - can you take a look when you get a chance? thanks!
This PR introduces a new function to apply shared memory indexing corrections. Previously, this was being done as part of post expansion. With the new PR, we look at the entire graph holistically and apply indexing changes to reads from shared memory. Signed-off-by: Harsh Menon <harsh@nod-labs.com>
Signed-off-by: Harsh Menon <harsh@nod-labs.com>
593cc79 to
64204c5
Compare
This PR modifies the way the index is being updated so that we allocate a new index whenever the index is being set. Also, rather than overload the write index to also return the register index, we add a separate register_index property for Write nodes. Signed-off-by: Harsh Menon <harsh@nod-labs.com>
64204c5 to
9274b0a
Compare
This PR introduces a new function to apply shared
memory indexing corrections. Previously, this was being done as part of post expansion. With the new PR, we look at the entire graph holistically and apply indexing changes to reads from shared memory.