scatter reduce lowering with include_self=False

Author: apbose

py/torch_tensorrt/dynamo/lowering/_decompositions.py

@@ -303,16 +303,30 @@ def __new__(cls, description, func):
         obj.func = func
         return obj
 
-    def reduce_operation_with_scatter(
-        self, operation_lhs, initial_tensor, dim, index_tensor, src_tensor
+    def reduce_operation_with_scatter_include_self(
+        self,
+        operation_lhs,
+        initial_tensor,
+        dim,
+        index_tensor,
+        src_tensor,
+        min_ele=float("-inf"),
+        max_ele=float("inf"),
+        include_self=True,
     ):
         scatter_tensor = None
         if self == ReduceOperation.SUM or self == ReduceOperation.MEAN:
             scatter_tensor = torch.zeros_like(initial_tensor)
         elif self == ReduceOperation.PROD:
             scatter_tensor = torch.ones_like(initial_tensor)
-        elif self == ReduceOperation.AMIN or self == ReduceOperation.AMAX:
+        elif self == ReduceOperation.AMAX:
             scatter_tensor = initial_tensor
+            if not (include_self):
+                scatter_tensor = torch.full_like(initial_tensor, min_ele)
+        elif self == ReduceOperation.AMIN:
+            scatter_tensor = initial_tensor
+            if not (include_self):
+                scatter_tensor = torch.full_like(initial_tensor, max_ele)
         else:
             # This case would not be encountered from torch itself
             print("Invalid Operation for Reduce op!!")
@@ -336,13 +350,39 @@ def scatter_reduce_decomposition(
     include_self: bool = True,
 ) -> torch.Tensor:
     scatter_loop_tensor = input_tensor
+    MAX_ELE = 0
+    MIN_ELE = 0
+    if src_tensor.dtype == torch.int32 or input_tensor.dtype == torch.int32:
+        MAX_ELE = 2147483647
+        MIN_ELE = -2147483648
+    else:
+        MAX_ELE = float("inf")
+        MIN_ELE = float("-inf")
+    if not (include_self):
+        if reduce == "sum" or reduce == "mean":
+            scatter_loop_tensor = torch.scatter(
+                scatter_loop_tensor, dim, index, torch.zeros_like(src_tensor)
+            )
+        if reduce == "prod":
+            scatter_loop_tensor = torch.scatter(
+                scatter_loop_tensor, dim, index, torch.ones_like(src_tensor)
+            )
+        if reduce == "amax":
+            src_red_tensor = torch.full_like(src_tensor, MIN_ELE)
+            scatter_loop_tensor = torch.scatter(
+                scatter_loop_tensor, dim, index, src_red_tensor
+            )
+        if reduce == "amin":
+            src_red_tensor = torch.full_like(src_tensor, MAX_ELE)
+            scatter_loop_tensor = torch.scatter(
+                scatter_loop_tensor, dim, index, src_red_tensor
+            )
+
     device_input_tensor = input_tensor.device
     # required for mean reduce operation
     scatter_count_tensor = torch.zeros_like(input_tensor)
     src_shape = list(src_tensor.shape)
     src_dim = src_shape[dim]
-    if include_self == False:
-        raise AssertionError("include_self False for scatter reduce not yet supported")
     for i in range(0, src_dim):
         src_slice = torch.select(src_tensor, dim, i)
         index_slice = torch.select(index, dim, i)
@@ -366,20 +406,53 @@ def scatter_reduce_decomposition(
                 dim,
                 index_slice,
                 torch.ones_like(src_slice),
+                MIN_ELE,
+                MAX_ELE,
+                include_self,
             )
         elif reduce == "amax":
             reduceOp = ReduceOperation.AMAX
         elif reduce == "amin":
             reduceOp = ReduceOperation.AMIN
-        scatter_loop_tensor = reduceOp.reduce_operation_with_scatter(
-            scatter_loop_tensor, input_tensor, dim, index_slice, src_slice
+        scatter_loop_tensor = reduceOp.reduce_operation_with_scatter_include_self(
+            scatter_loop_tensor,
+            input_tensor,
+            dim,
+            index_slice,
+            src_slice,
+            MIN_ELE,
+            MAX_ELE,
+            include_self,
         )
     if reduce == "mean":
         scatter_loop_tensor = torch.div(
             scatter_loop_tensor,
-            torch.add(scatter_count_tensor, torch.ones_like(scatter_count_tensor)),
+            (
+                torch.add(scatter_count_tensor, torch.ones_like(scatter_count_tensor))
+                if include_self
+                else scatter_count_tensor
+            ),
             rounding_mode="trunc",
         )
+    # for include_self cases for amax and amin additional processing is required
+    # except for the max elements in amax, rest are -inf or INT_MIN
+    # except for the min elements in amin, rest are +inf or INT_MAX
+    if reduce == "amax" and not (include_self):
+        # the relevant should be min, rest original
+        return torch.max(
+            scatter_loop_tensor,
+            torch.scatter(
+                input_tensor, dim, index, torch.full_like(src_tensor, MIN_ELE)
+            ),
+        )
+    if reduce == "amin" and not (include_self):
+        # the relevant should be min, rest original
+        return torch.min(
+            scatter_loop_tensor,
+            torch.scatter(
+                input_tensor, dim, index, torch.full_like(src_tensor, MAX_ELE)
+            ),
+        )
     return scatter_loop_tensor