[GitHub] [tvm] giuseros commented on a change in pull request #7925: Add a 'rolling_buffer' scheduling primitive

[GitBox <gi...@apache.org>]

giuseros commented on a change in pull request #7925:
URL: https://github.com/apache/tvm/pull/7925#discussion_r631173425



##########
File path: python/tvm/tir/transform/inject_rolling_buffer.py
##########
@@ -0,0 +1,214 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""Inject rolling buffers through a TIR transformation."""
+# pylint: disable=invalid-name,unused-argument,inconsistent-return-statements
+from collections import defaultdict
+
+import tvm
+from tvm import arith
+
+
+def InjectRollingBuffer():
+    """Inject rolling buffer statements.
+
+    Returns
+    -------
+    fpass : tvm.transform.Pass
+        The pass
+    """
+    buffer_to_attrs = defaultdict(list)
+    rolling_buffers = set()
+    rolling_buffer_to_info = dict()
+    iter_vars = list()
+    hoist_buffer_to_for = defaultdict(list)
+
+    class RollingBufferInfo:
+        def __init__(self, rolling_axis, rolling_extent, axis_overlaps, axis_iter_vars):
+            self.rolling_axis = rolling_axis
+            self.rolling_extent = rolling_extent
+            self.axis_overlaps = axis_overlaps
+            self.axis_iter_vars = axis_iter_vars
+
+    def _pre_visit(stmt):
+        if isinstance(stmt, tvm.tir.For):
+            # Manage the stack of iter_vars
+            iter_vars.append(stmt)
+
+        elif isinstance(stmt, tvm.tir.AttrStmt):
+            if isinstance(stmt.node, tvm.tir.Buffer):
+                if stmt.attr_key == "rolling_buffer" and stmt.value.value:
+                    # If the attribute is indicating that a buffer should be a rolling
+                    # buffer, then update the rolling_buffers set to include the bufffer
+                    rolling_buffers.add(stmt.node)
+                # Keep a dictionary associating attribute statements with the buffers
+                # they reference. We'll need this if the buffer gets hoisted and we
+                # need to hoist all of its attributes at the same time.
+                buffer_to_attrs[stmt.node].append(stmt)
+
+        elif isinstance(stmt, tvm.tir.BufferRealize):
+            if stmt.buffer in rolling_buffers:
+                # If a BufferRealize has been identified as needing to be made into
+                # a rolling buffer, begin the analysis...
+                bound_iter_vars = []
+                bound_strides = []
+                bound_overlaps = []
+                # We use the bound information of the BufferRealize to calculate
+                # how we can legally roll
+                for bound in stmt.bounds:
+                    # If the bound is an int, we can't roll over it
+                    if isinstance(bound.min, tvm.tir.IntImm):
+                        iter_var = None
+                        stride = 0
+                    # If the bound is just a Var, that implies the stride is 1
+                    elif isinstance(bound.min, tvm.tir.Var):
+                        iter_var = bound.min
+                        stride = 1
+                    # Otherwise, it's the iter var multiplied by the stride
+                    # If not we're in unknown behaviour, so assert
+                    else:
+                        assert isinstance(bound.min, tvm.tir.Mul)

Review comment:
       Should we simply not unrolling instead of asserting?

##########
File path: src/te/schedule/schedule_ops.cc
##########
@@ -54,6 +54,9 @@ Stmt MakePipeline(const Stage& s, const std::unordered_map<IterVar, Range>& dom_
   pipeline = s->op->BuildRealize(s, dom_map, pipeline);
   // use attribute to mark scope of the operation.
   pipeline = AttrStmt(s->op, tir::attr::realize_scope, StringImm(s->scope), pipeline);
+  if (s->rolling_buffer) {
+    pipeline = AttrStmt(s->op, tir::attr::rolling_buffer, Bool(true), pipeline);

Review comment:
       What about `rolling_buffer_scope`?




-- 
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.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org