[MetaSchedule] Schedule Rule: Add RFactor (apache#9975)

* add rfactor

* format

* fix ci

include/tvm/meta_schedule/schedule_rule.h

@@ -152,6 +152,16 @@ class ScheduleRule : public runtime::ObjectRef {
                                                Optional<Integer> vector_load_max_len,        //
                                                Optional<Map<String, ObjectRef>> reuse_read,  //
                                                Optional<Map<String, ObjectRef>> reuse_write);
+  /*!
+   * \brief Create a rule: add-rfactor to some blocks if needed
+   * \param max_jobs_per_core The maximum number of jobs to be launched per CPU core. It sets the
+   * uplimit of CPU parallelism, i.e. `num_cores * max_jobs_per_core`. Use -1 to disable
+   * parallelism.
+   * \param max_innermost_factor The maximum size of the innermost factor. NullOpt means no limit
+   * \return The schedule rule created
+   */
+  TVM_DLL static ScheduleRule AddRFactor(int max_jobs_per_core,  //
+                                         Optional<Integer> max_innermost_factor);
   /*!
    * \brief A rule that randomly select a compute-at location for a free block
    * \return The rule created

python/tvm/meta_schedule/schedule_rule/__init__.py

@@ -16,6 +16,7 @@
 Meta Schedule schedule rules are used for modification of
 blocks in a schedule. See also PostOrderApply.
 """
+from .add_rfactor import AddRFactor
 from .auto_inline import AutoInline
 from .schedule_rule import PyScheduleRule, ScheduleRule
 from .random_compute_location import RandomComputeLocation

python/tvm/meta_schedule/schedule_rule/add_rfactor.py

@@ -0,0 +1,49 @@
+# 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.
+"""Add-rfactor Rule that add-rfactor to some blocks if needed"""
+from typing import Optional
+
+from tvm._ffi import register_object
+
+from .. import _ffi_api
+from .schedule_rule import ScheduleRule
+
+
+@register_object("meta_schedule.AddRFactor")
+class AddRFactor(ScheduleRule):
+    """Rules for add-rfactor to some blocks if needed.
+
+    Parameters
+    ----------
+    max_jobs_per_core: int
+        The maximum number of jobs to be launched per CPU core. It sets the uplimit of CPU
+        parallelism, i.e. `num_cores * max_jobs_per_core`.
+        Use -1 to disable parallelism.
+    max_innermost_factor: Optional[int] = None
+        The maximum size of the innermost factor. None means no limit.
+    """
+
+    def __init__(
+        self,
+        max_jobs_per_core: int = 16,
+        max_innermost_factor: Optional[int] = None,
+    ) -> None:
+        self.__init_handle_by_constructor__(
+            _ffi_api.ScheduleRuleAddRFactor,  # type: ignore # pylint: disable=no-member
+            max_jobs_per_core,
+            max_innermost_factor,
+        )

python/tvm/meta_schedule/testing/schedule_rule.py

@@ -16,6 +16,7 @@
 # under the License.
 """Default schedule rules"""
 from tvm.meta_schedule.schedule_rule import (
+    AddRFactor,
     AutoInline,
     ScheduleRule,
 )
@@ -45,3 +46,10 @@ def auto_inline(target: Target) -> ScheduleRule:
             disallow_op=None,
         )
     raise NotImplementedError(f"{target.kind.name} is not supported")
+
+
+def add_rfactor(target: Target) -> ScheduleRule:
+    """Default schedule rules for with add_rfactor"""
+    if target.kind.name == "llvm":
+        return AddRFactor(max_jobs_per_core=16, max_innermost_factor=64)
+    raise NotImplementedError(f"{target.kind.name} is not supported")