[llvm] Use GEP for array access instead of ptrtoint/inttoptr #4276
Conversation
|
✔️ Deploy Preview for docsite-preview ready! 🔨 Explore the source changes: ece9d3a 🔍 Inspect the deploy log: https://app.netlify.com/sites/docsite-preview/deploys/620b172ff79809000887c413 😎 Browse the preview: https://deploy-preview-4276--docsite-preview.netlify.app |
|
/benchmark |
| @@ -69,7 +69,7 @@ def __init__(self, n=1, m=1, dt=None, suppress_warning=False): | |||
| self.local_tensor_proxy = impl.expr_init_local_tensor( | |||
| [len(n)], dt, | |||
| expr.make_expr_group([expr.Expr(x) for x in n])) | |||
| self.dynamic_index_stride = ti_core.data_type_size(dt) | |||
| self.dynamic_index_stride = 1 | |||
There was a problem hiding this comment.
Has the meaning of dynamic_index_stride been changed (from bytes to index)?
There was a problem hiding this comment.
After this PR, handling dynamic indexing of matrix fields requires manipulating pointers directly, while handling dynamic indexing of local matrices only requires array indices. However, currently they share the same expression (TensorElementExpression) and statement (PtrOffsetStmt), so the stride actually stands for two different things (bytes & indices) in these two cases. A future step must be to split them up. As it will add even one more type of pointer statement, I suggest doing this while we refactor those pointer statements.
Related issue = #2590
Using
ptrtoint/inttoptrfor array access prevents LLVM from demoting array alloca into individual allocas, which further prevents mem2reg optimizations. After this PR, withdynamic_index=True, we can have the same runtime performance as withdynamic_index=Falsefor old code which can only access matrices with constant indices.