vinx13 commented on code in PR #70:
URL: https://github.com/apache/tvm-rfcs/pull/70#discussion_r893900504
##########
rfcs/0070-introducing-decl-buffer.md:
##########
@@ -136,11 +138,26 @@ def elemwise(A: T.Buffer[(16, 16), "float32"], C:
T.Buffer[(16, 16), "float32"])
C_flattened[i * 16 + j] = A[i * 16 + j]
```
+Specifically, the updated flow of buffer flattening using `DeclBuffer` will be:
+1. Before `FlattenBuffer/StorageFlatten`: Buffers are declared in the
`buffer_map`, and are not flattened. Buffer access is done using N-d
unflattened indices.
+2. After `FlattenBuffer/StorageFlatten`, but before `MakePackedAPI`: Buffers
are declared in the `buffer_map`, and are not flattened. Buffer access is done
through a buffer alias explicitly created via `DeclBuffer`, where the alias
shares the same data pointer, but has a flattened shape and is accessed with
flattened indices.
Review Comment:
It is not changed. The buffer in `buffer_map` is always unflattened. After
`FlattenBuffer/StorageFlatten`, we will create a flattened view of the buffer
using buffer alias, and in IR all accesses to the buffer will be done via the
flattened view. This is because from the perspective of the calling convention
of `PrimFunc`, we always expect the input to have unflattened shape (e.g. the
shape of `DLTensor`). But internally in the IR, we need to flatten to physical
shape and indices which is how it is done in codegen and runtime
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
