initial version

Author: yf225

helion/_compiler/ast_extension.py

@@ -8,6 +8,8 @@
 from typing import TYPE_CHECKING
 from typing import TypeVar
 
+import torch
+
 from .. import exc
 from .source_location import SourceLocation
 from .source_location import current_location
@@ -82,10 +84,31 @@ def __repr__(self) -> str:
 
     def update_type_info(self, type_info: TypeInfo) -> TypeInfo:
         if self._type_info is not None and type_info != self._type_info:
+            prev_rank = self._tensor_rank(self._type_info)
+            new_rank = self._tensor_rank(type_info)
+            if (
+                prev_rank is not None
+                and new_rank is not None
+                and prev_rank != new_rank
+            ):
+                self._type_info = type_info
+                return self._type_info
             type_info = self._type_info.merge(type_info)
         self._type_info = type_info
         return self._type_info
 
+    @staticmethod
+    def _tensor_rank(type_info: "TypeInfo") -> int | None:
+        fake_value = getattr(type_info, "fake_value", None)
+        if isinstance(fake_value, torch.Tensor):
+            return fake_value.dim()
+        tensor = getattr(type_info, "tensor", None)
+        if tensor is not None:
+            fake_value = getattr(tensor, "fake_value", None)
+            if isinstance(fake_value, torch.Tensor):
+                return fake_value.dim()
+        return None
+
     def debug_annotations(self) -> list[str]:
         result = []
         if self._type_info:

helion/_compiler/compile_environment.py

@@ -142,6 +142,30 @@ def allocate_reduction_dimension(self, size: torch.SymInt | int) -> BlockSizeInf
             if rdim.reduction and rdim.size == size:
                 return rdim
 
+        # Check if size matches any tile dimension for symbolic equality.
+        # When building expressions that mix sizes derived from tiles
+        # (e.g., via slicing) with sizes coming directly from tile block vars, we
+        # want them to share the same SymInt variable whenever they are equal by
+        # construction. This preserves equality in the shape environment and avoids
+        # spurious "size mismatch" issues during fake-tensor broadcasting and
+        # arithmetic in type propagation.
+        if isinstance(size, torch.SymInt):
+            size_str = str(size)
+            for block_info in self.block_sizes:
+                if not block_info.reduction and str(block_info.var) == size_str:
+                    # Create reduction dimension with the same var to preserve
+                    # symbolic equality and ensure all later users see identical
+                    # symbols (rather than equal-but-distinct SymInts).
+                    rdim_idx = self.allocate_block_size(
+                        size,
+                        reduction=True,
+                        source=ReductionLoopBlockSizeSource(
+                            reduction_loop=len([b for b in self.block_sizes if b.reduction])
+                        ),
+                    )
+                    self.block_sizes[rdim_idx].var = block_info.var
+                    return self.block_sizes[rdim_idx]
+
         # Allocate a new reduction dimension
         rdim_idx = self.allocate_block_size(
             size,
@@ -203,6 +227,91 @@ def cached_create_unbacked_symint(
             self._symint_cache[key] = result
         return result
 
+
+    def register_tile_index_tensor_block_id(self, tensor: torch.Tensor, block_id: int) -> None:
+        """Annotate ``tensor`` as originating from ``tile.index`` with ``block_id`` provenance."""
+        tensor._tile_index_block_id = block_id  # type: ignore[attr-defined]
+
+    def get_tile_index_tensor_block_id(self, tensor: torch.Tensor) -> int | None:
+        """Return the originating ``tile.index`` block id if present."""
+        return getattr(tensor, "_tile_index_block_id", None)
+
+    def get_indexer_output_dims(
+        self,
+        indexer_tensor: torch.Tensor,
+        base_dim_size: int | torch.SymInt | None,
+    ) -> list[int | torch.SymInt]:
+        """Map a tensor indexer's shape to the output dimensions for advanced indexing."""
+
+        dims = list(indexer_tensor.size())
+        non_broadcast_dims = [d for d in dims if self.size_hint(d) != 1]
+
+        # Multi-dimensional indexer - return full shape
+        if len(non_broadcast_dims) > 1:
+            return dims
+
+        block_id = self.get_tile_index_tensor_block_id(indexer_tensor)
+        if block_id is None and base_dim_size is not None:
+            block_id = self.get_block_id(base_dim_size)
+        if block_id is None and non_broadcast_dims:
+            block_id = self.get_block_id(non_broadcast_dims[0])
+
+        if block_id is not None:
+            return [self.block_sizes[block_id].var]
+        if non_broadcast_dims:
+            return [non_broadcast_dims[0]]
+        return [1]
+
+    def tensor_indexer_broadcast_shape(
+        self, tensors: typing.Sequence[torch.Tensor]
+    ) -> list[int | torch.SymInt] | None:
+        """Compute a shared broadcast shape for tensor indexers when needed."""
+
+        tensor_list = [t for t in tensors if isinstance(t, torch.Tensor)]
+        if not tensor_list:
+            return None
+
+        if all(self.get_tile_index_tensor_block_id(t) is not None for t in tensor_list):
+            return None
+
+        shapes = [list(t.size()) for t in tensor_list]
+        return compute_broadcast_shape_for_tensor_indexers(shapes, self)
+
+    def resolve_tile_index_shape(
+        self, input_tensor: torch.Tensor, output_shape: typing.Sequence[int | torch.SymInt]
+    ) -> tuple[list[int | torch.SymInt], int | None]:
+        """Resolve the symbolic shape for tensors derived from ``tile.index``.
+
+        Returns a copy of ``output_shape`` where the single non-broadcast
+        dimension is replaced with the canonical block-symbol and the associated
+        block_id to register on the new tensor. If the tensor is not a tile
+        indexer or it introduces more than one non-broadcast dimension, the
+        original shape and ``None`` are returned.
+        """
+
+        block_id = self.get_tile_index_tensor_block_id(input_tensor)
+        if block_id is None:
+            return list(output_shape), None
+
+        resolved = list(output_shape)
+        non_broadcast = [i for i, s in enumerate(resolved) if self.size_hint(s) != 1]
+        if len(non_broadcast) <= 1:
+            if non_broadcast:
+                resolved[non_broadcast[0]] = self.block_sizes[block_id].var
+            return resolved, block_id
+        return resolved, None
+
+    def new_index_result(
+        self, tensor: torch.Tensor, output_shape: typing.Sequence[int | torch.SymInt]
+    ) -> torch.Tensor:
+        """Create a new tensor for indexing/view ops while preserving tile index provenance."""
+
+        resolved_shape, block_id = self.resolve_tile_index_shape(tensor, output_shape)
+        result = tensor.new_empty(resolved_shape)
+        if block_id is not None:
+            self.register_tile_index_tensor_block_id(result, block_id)
+        return result
+
     def to_fake(self, obj: object, origin: Origin) -> object:
         if isinstance(obj, torch.Tensor):
             return self._to_fake_tensor(obj, origin.to_source())
@@ -283,6 +392,10 @@ def _to_fake_tensor(self, tensor: torch.Tensor, source: Source) -> torch.Tensor:
                 self.fake_mode, tensor, shape_env=self.shape_env, source=source
             )
         self.input_sources[result] = source
+        if hasattr(tensor, "_tile_index_block_id"):
+            self.register_tile_index_tensor_block_id(
+                result, typing.cast(int, getattr(tensor, "_tile_index_block_id"))
+            )
         if isinstance(source, LocalSource):
             for i, s in enumerate(result.size()):
                 if isinstance(s, torch.SymInt) and isinstance(
@@ -357,9 +470,9 @@ def current() -> CompileEnvironment:
     @staticmethod
     def has_current() -> bool:
         try:
-            CompileEnvironment.current()
-            return True
-        except NoCurrentEnvironment:
+            CompileEnvironment.current()	
+            return True	
+        except NoCurrentEnvironment:	
             return False
 
     def get_block_id(self, size: int | torch.SymInt | sympy.Expr) -> int | None:
@@ -535,3 +648,35 @@ def _to_sympy(x: int | torch.SymInt) -> sympy.Expr:
 
 def _has_unbacked(expr: sympy.Expr) -> bool:
     return any(n.name.startswith("u") for n in expr.free_symbols)  # pyright: ignore[reportAttributeAccessIssue]
+
+
+def compute_broadcast_shape_for_tensor_indexers(
+    shapes: list[list[int | torch.SymInt]], 
+    env: "CompileEnvironment"
+) -> list[int | torch.SymInt]:
+    """
+    Compute broadcast shape for multiple tensor indexers using right-aligned broadcasting.
+    
+    Args:
+        shapes: List of shapes from each tensor indexer
+        env: CompileEnvironment for size_hint and known_equal checks
+        
+    Returns:
+        Broadcast shape as list of dimensions
+    """
+    if not shapes:
+        return []
+    
+    max_ndim = max(len(s) for s in shapes)
+    padded = [([1] * (max_ndim - len(s)) + s) for s in shapes]
+    broadcast_shape: list[int | torch.SymInt] = []
+    
+    for dims_at_pos in zip(*padded, strict=True):
+        chosen: int | torch.SymInt | None = None
+        for d in dims_at_pos:
+            if env.size_hint(d) != 1:
+                if chosen is None or env.known_equal(chosen, d):
+                    chosen = d
+        broadcast_shape.append(chosen if chosen is not None else 1)
+    
+    return broadcast_shape