mindspore 2.4.0__cp311-cp311-macosx_10_15_x86_64.whl

mindspore/_extends/graph_kernel/model/graph_split.py

@@ -0,0 +1,1348 @@
+# Copyright 2020-2022 Huawei Technologies Co., Ltd
+#
+# Licensed 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.
+# ===========================================================================
+"""Cost model splitter"""
+from functools import reduce as prod_reduce
+from functools import partial
+from .model import PrimLib, Graph, Tensor, Operator
+
+
+def tensor_size(tensor):
+    """get tensor size"""
+    size = 1
+    for i in tensor.shape:
+        size *= i
+    return size
+
+
+def reduce_nums(ops):
+    """get reduce nums"""
+    count = 0
+    for op in ops:
+        if op.prim.startswith('Reduce'):
+            count += 1
+    return count
+
+
+def may_stitch(dom, a, r, stitch_axis_size, stitch_buffer_size):
+    """check if can stitch"""
+
+    def _same_stitch_axis(stitch_tensors, final_outs, stitch_axis_size):
+        """does a and b have same stitch axis"""
+
+        def _stitch_axis(shape, stitch_axis_size):
+            """get stitch axis"""
+            stitchaxis = []
+            size = 1
+            for i in shape:
+                size = size * i
+                stitchaxis.append(i)
+                if size >= stitch_axis_size:
+                    return stitchaxis
+            return []
+
+        x = []
+        x.extend(stitch_tensors)
+        x.extend(final_outs)
+        stitch_axis_0 = _stitch_axis(x[0].shape, stitch_axis_size)
+        for item in x:
+            i_stitch_axis = _stitch_axis(item.shape, stitch_axis_size)
+            if not i_stitch_axis or i_stitch_axis != stitch_axis_0:
+                return False
+        return True
+
+    if a.pattern <= PrimLib.REDUCE and r <= PrimLib.BROADCAST and dom.check_acyclic(a):
+        if reduce_nums(a.ops) >= 2:
+            return False
+        dom_outs = set(op.output for op in dom.ops)
+        a_ins = set(op_input for op in a.ops for op_input in op.inputs)
+        a_outs = set(op.output for op in a.ops)
+        a_final_outs = list(tensor for tensor in a_outs if tensor not in a_ins)
+        stitch_tensors = list(tensor for tensor in dom_outs if tensor in a_ins)
+        if not _same_stitch_axis(stitch_tensors, a_final_outs, stitch_axis_size):
+            return False
+        return any((tensor_size(tensor) >= stitch_buffer_size for tensor in stitch_tensors))
+    return False
+
+
+class CommonPattern:
+    """common fuse strategies across various devices"""
+
+    @staticmethod
+    def reshape(dom):
+        """fuse strategy for reshape dom"""
+        if dom.pattern != PrimLib.RESHAPE:
+            return [], False
+        min_area, forward_fuse = None, False
+        for a, _ in dom.out_relations.items():
+            if a.pattern <= PrimLib.BROADCAST and dom.check_acyclic(a):
+                if min_area is None or a.pattern < min_area.pattern:
+                    min_area = a
+        for a, _ in dom.in_relations.items():
+            if a.pattern <= PrimLib.BROADCAST and a.check_acyclic(dom):
+                if min_area is None or a.pattern < min_area.pattern:
+                    min_area, forward_fuse = a, True
+        return ([min_area], forward_fuse) if min_area else ([], False)
+
+    @staticmethod
+    def isolate_reshape(dom):
+        """fuse strategy for isolate reshape dom"""
+        if dom.pattern != PrimLib.RESHAPE or len(dom.ops) != 1:
+            return [], False
+        for a, _ in dom.out_relations.items():
+            if a.mode == GraphSplitByPattern.Area.MODE_COMPOSITE and dom.check_acyclic(a):
+                return [a], False
+        for a, _ in dom.in_relations.items():
+            if a.mode == GraphSplitByPattern.Area.MODE_COMPOSITE and a.pattern <= PrimLib.BROADCAST and \
+                    a.check_acyclic(dom):
+                return [a], True
+        return [], False
+
+    @staticmethod
+    def elemwise_depth(dom):
+        """fuse strategy in depth for elemwise dom"""
+        if dom.pattern != PrimLib.ELEMWISE or len(dom.in_relations) != 1:
+            return [], False
+        a, r = list(dom.in_relations.items())[0]
+        if a.pattern > PrimLib.ELEMWISE or len(a.out_relations) != 1 or r > PrimLib.ELEMWISE:
+            return [], False
+        if tensor_size(a.dom_op().output) != tensor_size(dom.dom_op().output):
+            return [], False
+        return [a], True
+
+    @staticmethod
+    def elemwise_width(dom):
+        """fuse strategy in width for elemwise dom"""
+        if dom.pattern != PrimLib.ELEMWISE:
+            return [], False
+        fused = []
+        for a, r in dom.in_relations.items():
+            if a.pattern <= PrimLib.ELEMWISE and r <= PrimLib.ELEMWISE and a.check_acyclic(dom):
+                if tensor_size(a.dom_op().output) == tensor_size(dom.dom_op().output):
+                    fused.append(a)
+        return fused, True
+
+    @staticmethod
+    def broadcast_depth(dom):
+        """fuse strategy in depth for broadcast dom"""
+        if dom.pattern not in (PrimLib.ELEMWISE, PrimLib.BROADCAST) or len(dom.in_relations) != 1:
+            return [], False
+        a, r = list(dom.in_relations.items())[0]
+        if a.pattern > PrimLib.BROADCAST or len(a.out_relations) != 1 or r > PrimLib.ELEMWISE:
+            return [], False
+        if tensor_size(a.dom_op().output) != tensor_size(dom.dom_op().output):
+            return [], False
+        return [a], True
+
+    @staticmethod
+    def broadcast_width(dom):
+        """fuse strategy in width for broadcast dom"""
+        if dom.pattern not in (PrimLib.ELEMWISE, PrimLib.BROADCAST):
+            return [], False
+        fused = []
+        for a, r in dom.in_relations.items():
+            if a.pattern <= PrimLib.BROADCAST and r <= PrimLib.ELEMWISE and a.check_acyclic(dom):
+                if tensor_size(a.dom_op().output) == tensor_size(dom.dom_op().output):
+                    fused.append(a)
+        return fused, True
+
+    @staticmethod
+    def assign(dom):
+        """fuse strategy for assign dom"""
+        if len(dom.ops) != 1 or dom.dom_op().prim != "Assign":
+            return [], False
+        fused = []
+        for a, _ in dom.in_relations.items():
+            fused.append(a)
+        return fused, True
+
+
+class ReshapeElimChecker:
+    """ check reshape elim """
+
+    def __init__(self, reshape):
+        def _get_remap_axis(in_shape, out_shape):
+            rin, rout = [], []
+            in_prod, out_prod, out_idx = 1, out_shape[-1], -1
+            in_ext, out_ext = -len(in_shape) - 1, -len(out_shape) - 1
+            for in_idx in range(-1, in_ext, -1):
+                in_prod = in_prod * in_shape[in_idx]
+                while out_prod < in_prod:
+                    rout.append(out_idx)
+                    out_idx = out_idx - 1
+                    out_prod = out_prod * out_shape[out_idx]
+                if out_prod == in_prod and out_idx > out_ext and out_shape[out_idx] == in_shape[in_idx]:
+                    out_idx = out_idx - 1
+                    if out_idx > out_ext:
+                        out_prod = out_prod * out_shape[out_idx]
+                else:
+                    rin.append(in_idx)
+            if out_idx > out_ext:
+                rout.extend([i for i in range(out_idx, out_ext, -1)])
+            return rin, rout
+
+        remap_in, remap_out = _get_remap_axis(reshape.inputs[0].shape, reshape.output.shape)
+        self.exc_fwd = self._collect_exc_ops(reshape, remap_in, True)
+        self.exc_bwd = self._collect_exc_ops(reshape, remap_out, False)
+
+    @staticmethod
+    def _collect_exc_ops(reshape, remap_axis, is_fwd):
+        """collect exclude ops of reshape"""
+
+        def _propagate(remap, src, des):
+            out_remap = []
+            src_prod, des_prod, des_idx = 1, 1, 0
+            for src_idx in range(-1, -len(src) - 1, -1):
+                src_prod = src_prod * src[src_idx]
+                if src_idx in remap:
+                    while des_prod < src_prod:
+                        des_idx = des_idx - 1
+                        des_prod = des_prod * des[des_idx]
+                        out_remap.append(des_idx)
+                else:
+                    while des_prod < src_prod:
+                        prod = des_prod * des[des_idx - 1]
+                        if prod > src_prod:
+                            break
+                        des_idx, des_prod = des_idx - 1, prod
+            return out_remap
+
+        def _remap_check(op, remap, iter_type):
+            if iter_type not in (PrimLib.ELEMWISE, PrimLib.BROADCAST):
+                return False
+            for t in op.inputs:
+                for i in remap:
+                    if -i <= len(t.shape) and t.shape[i] != op.output.shape[i]:
+                        return False
+            return True
+
+        def push_stack(op, remap):
+            stack.append((op, remap))
+            visited.add(op)
+
+        def _visit_fwd(op, remap):
+            for t in op.inputs:
+                if t.op is None:
+                    _visit_bwd(t, remap)
+                elif tensor_size(t) > 1 and t.op not in visited:  # all broadcast
+                    iter_type = PrimLib.iter_type(t.op)
+                    if iter_type == PrimLib.RESHAPE:
+                        new_remap = _propagate(remap, t.shape, t.op.inputs[0].shape)
+                        push_stack(t.op, new_remap)
+                    elif _remap_check(t.op, remap, iter_type):
+                        push_stack(t.op, remap)
+                    else:
+                        exc_ops.add(t.op)
+
+        def _visit_bwd(t, remap):
+            for op in t.to_ops:
+                if op not in visited:
+                    iter_type = PrimLib.iter_type(op)
+                    if iter_type == PrimLib.REDUCE and tensor_size(op.output) == 1:  # all reduce
+                        continue
+                    if iter_type == PrimLib.RESHAPE:
+                        new_remap = _propagate(remap, t.shape, op.output.shape)
+                        push_stack(op, new_remap)
+                    elif _remap_check(op, remap, iter_type):
+                        push_stack(op, remap)
+                    else:
+                        exc_ops.add(op)
+
+        exc_ops, stack, visited = set(), [], {reshape}
+        if is_fwd:
+            _visit_fwd(reshape, remap_axis)
+        else:
+            _visit_bwd(reshape.output, remap_axis)
+        while stack:
+            top, remap = stack.pop()
+            _visit_bwd(top.output, remap)
+            _visit_fwd(top, remap)
+        return exc_ops
+
+    def check(self, ops, is_fwd):
+        """ fuse check """
+        if is_fwd:
+            fwd_res = all([op not in self.exc_fwd for op in ops]) if self.exc_fwd is not None else False
+            bwd_res = self.exc_bwd is not None
+        else:
+            fwd_res = self.exc_fwd is not None
+            bwd_res = all([op not in self.exc_bwd for op in ops]) if self.exc_bwd is not None else False
+        return [fwd_res, bwd_res] if fwd_res or bwd_res else False
+
+    def commit(self, res):
+        """ commit fuse result """
+        if not res[0] and self.exc_fwd is not None:
+            self.exc_fwd = None
+        if not res[1] and self.exc_bwd is not None:
+            self.exc_bwd = None
+
+
+class ReduceOutFuseChecker:
+    """Reduce output fuse checker """
+
+    def __init__(self, red_op):
+        self.output_excluded = set()
+        recursion_stack = [red_op]
+        while recursion_stack:
+            op = recursion_stack.pop()
+            for to in op.output.to_ops:
+                idx = to.inputs.index(op.output)
+                if PrimLib.iter_type(to) > PrimLib.ELEMWISE or \
+                        tensor_size(to.inputs[idx]) != tensor_size(to.output):
+                    self.output_excluded.add(to)
+                else:
+                    recursion_stack.append(to)
+
+    def check(self, ops, is_fwd):
+        """ fuse check """
+        if not is_fwd and self.output_excluded:
+            for op in self.output_excluded:
+                if op in ops:
+                    return False
+        return True
+
+    def commit(self, res):
+        """ commit fuse result """
+        del res
+        return self.output_excluded  # I'm not static
+
+
+class GraphSplitByPattern:
+    """Graph splitter"""
+
+    class ReachTable:
+        """Reachable table"""
+
+        def __init__(self, size):
+            self.map = []
+            self.alive = set(range(size))
+            for i in range(0, size):
+                self.map.append([False] * size)
+                self.map[i][i] = True
+
+        def reachable(self, x, y):
+            """reachable from x to y"""
+            return self.map[x][y]
+
+        def sync(self, x, y):
+            """sync from y to x"""
+            for i in self.alive:
+                self._link(self.map[y][i], x, i)
+
+        def _link(self, cond, f, t):
+            """link from `f` to `t`"""
+            if cond:
+                self.map[f][t] = True
+
+        def fuse(self, x, y):
+            """fuse y to x"""
+            for i in self.alive:
+                # i is the succeeding node of y, links the x's previous nodes to i
+                if self.map[y][i] and not self.map[x][i]:
+                    for pre in self.alive:
+                        self._link(self.map[pre][x], pre, i)
+                # i is the previous node of y, link i to x's succeeding nodes
+                if self.map[i][y] and not self.map[i][x]:
+                    for suc in self.alive:
+                        self._link(self.map[x][suc], i, suc)
+            self.alive.remove(y)
+
+    class Area:
+        """Area"""
+        MODE_BASIC = 1
+        MODE_COMPOSITE = 2
+
+        class StitchInfo:
+            """StitchInfo"""
+
+            def __init__(self):
+                self.stitch_ops = set()
+                self.stitch_atomic_ops = set()
+
+            def has_stitch_op(self):
+                """check stitch_op exists"""
+                return self.stitch_ops or self.stitch_atomic_ops
+
+        def __init__(self, init_op, is_output, unique_id, reach_tab):
+            self.pattern = PrimLib.iter_type(init_op) if init_op is not None else PrimLib.UNKNOWN
+            self.ops = [] if init_op is None else [init_op]
+            self.in_relations = dict()  # {area1: relation1, area2: relation2, ...}
+            self.out_relations = dict()  # {area1: relation1, area2: relation2, ...}
+            self.mode = None
+            self.stitch_info = self.StitchInfo()
+            self.recompute_ops = []
+            self.is_output = is_output
+            self.output_excluded = set()
+            self.unique_id = unique_id
+            self.reach_tab = reach_tab
+            self.checkers = []
+            if self.pattern == PrimLib.RESHAPE and init_op.inputs:  # reshape's input may be empty (const value)
+                self.checkers.append(ReshapeElimChecker(init_op))
+            elif self.pattern == PrimLib.REDUCE:
+                self.checkers.append(ReduceOutFuseChecker(init_op))
+
+        def __str__(self):
+            return '<' + '-'.join((op.output.name for op in self.ops)) + '>'
+
+        def __repr__(self):
+            return str(self)
+
+        @staticmethod
+        def get_relation(op, i):
+            """Get op relation"""
+            relation = PrimLib.UNKNOWN
+            _, elem_relation = PrimLib.input_relation(op, i)
+            for r in elem_relation:
+                if r is None:
+                    relation = max(relation, PrimLib.BROADCAST)
+                elif r > relation:
+                    relation = r
+            return relation
+
+        def link_input(self, area_map):
+            """Link inputs"""
+            for i, t in enumerate(self.ops[0].inputs):
+                if t.op is not None:
+                    area, relation = area_map[t.op], self.get_relation(self.ops[0], i)
+                    self.in_relations[area] = relation
+
+        def link_output(self):
+            """Link outputs"""
+            for input_area, r in self.in_relations.items():
+                input_area.out_relations[self] = r
+            for out, _ in self.out_relations.items():
+                self.reach_tab.sync(self.unique_id, out.unique_id)
+
+        def update_stitch_info(self, stitch_info):
+            """Update stitch info"""
+            if stitch_info.stitch_ops:
+                self.stitch_info.stitch_ops.update(stitch_info.stitch_ops)
+            if stitch_info.stitch_atomic_ops:
+                self.stitch_info.stitch_atomic_ops.update(stitch_info.stitch_atomic_ops)
+
+        def fuse_confirm(self, area):
+            """confirm if area can be fused"""
+
+            def _check(a, b, res, fwd):
+                for checker in a.checkers:
+                    r = checker.check(b.ops, fwd)
+                    if not r:
+                        return False
+                    res.append(r)
+                return True
+
+            def _commit(a, res):
+                for i, checker in enumerate(a.checkers):
+                    checker.commit(res[i])
+
+            res1, res2 = [], []
+            if not _check(self, area, res1, True) or not _check(area, self, res2, False):
+                return False
+            _commit(self, res1)
+            _commit(area, res2)
+            return True
+
+        def fuse_prepare(self, dom):
+            """do some prepare before fused to dom"""
+            del dom
+            return self.unique_id  # I'm not static method
+
+        def fuse_done(self, dom):
+            """do some thing after fused to dom"""
+            dom.reach_tab.fuse(dom.unique_id, self.unique_id)
+
+        def fuse(self, area):
+            """Fuse `area` to `self`"""
+
+            def _update_relation(relations, a, r):
+                relations[a] = max(r, relations[a]) if a in relations else r
+
+            def _update_pattern():
+                if area.pattern > self.pattern:
+                    self.pattern = area.pattern
+                if area in self.in_relations and self.in_relations.get(area) > self.pattern:
+                    self.pattern = self.in_relations.get(area)
+
+            def _fuse_relation(self_relations, new_relations):
+                for a, r in new_relations.items():
+                    if a != self:
+                        _update_relation(self_relations, a, r)
+                if area in self_relations:
+                    self_relations.pop(area)
+
+            def _redirect_relation(rels):
+                """Replace `area` with `self` in relations"""
+                if area in rels:
+                    r = rels.pop(area)
+                    _update_relation(rels, self, r)
+
+            area.fuse_prepare(self)
+            if self.pattern >= area.pattern:
+                self.ops.extend(area.ops)
+            else:
+                self.ops = area.ops + self.ops
+            _update_pattern()
+            _fuse_relation(self.in_relations, area.in_relations)
+            _fuse_relation(self.out_relations, area.out_relations)
+            for a, _ in area.in_relations.items():
+                _redirect_relation(a.out_relations)
+            for a, _ in area.out_relations.items():
+                _redirect_relation(a.in_relations)
+            if self.pattern > PrimLib.RESHAPE:
+                self.mode = self.MODE_COMPOSITE
+            if area.is_output and not self.is_output:
+                self.is_output = True
+            self.update_stitch_info(area.stitch_info)
+            self.recompute_ops.extend(area.recompute_ops)
+            self.checkers.extend(area.checkers)
+            area.fuse_done(self)
+
+        def check_acyclic(self, to):
+            """Check circle. It returns false if circle exists"""
+            for out, _ in self.out_relations.items():
+                if out != to and self.reach_tab.reachable(out.unique_id, to.unique_id):
+                    return False
+            return True
+
+        def dom_op(self):
+            """Get dom op"""
+            return self.ops[0]
+
+    class RecomputeArea(Area):
+        """RecomputeArea"""
+
+        def __init__(self, unique_id, reach_tab):
+            super().__init__(None, False, unique_id, reach_tab)
+            self.recom_pre = None
+            self.recom_user = None
+            self.recom_dom = None
+            self.dom_user_r = PrimLib.UNKNOWN
+            self.ori_op_map = {}
+            self.recom_map = {}
+            self.fuse_success = False
+
+        def fuse_prepare(self, dom):
+            """copy recompute_ops in area to ops, self is area's user"""
+            tail_tensor = self.recompute_ops[-1].output
+            # copy tensors, all copied are Tensor.PARA_NONE
+            tensor_map = {}
+            if self.recompute_ops[0].inputs:
+                tensor_map[self.recompute_ops[0].inputs[0]] = self.recompute_ops[0].inputs[0]
+            for op in self.recompute_ops:
+                orig_tensor = op.output
+                cp_tensor = Tensor(orig_tensor.name, orig_tensor.shape, orig_tensor.dtype, orig_tensor.data_format)
+                tensor_map[orig_tensor] = cp_tensor
+            # copy ops
+            cp_ops = []
+            for op in self.recompute_ops:
+                inputs = [tensor_map.get(op.inputs[0])] if op.inputs else []
+                cp_op = Operator(op.prim, inputs, tensor_map.get(op.output), op.attrs)
+                cp_op.all_inputs = cp_op.inputs
+                cp_ops.append(cp_op)
+                self.ori_op_map[cp_op] = op
+            # connect copied ops
+            for op in dom.ops:
+                if tail_tensor in op.inputs:
+                    op.inputs.remove(tail_tensor)
+                    op.inputs.append(tensor_map.get(tail_tensor))
+                    tail_tensor.to_ops.remove(op)
+                    tensor_map.get(tail_tensor).to_ops.append(op)
+            # fill cp_ops in self.recompute_area
+            cp_dom_op = None
+            for cp, ori in self.ori_op_map.items():
+                if ori == self.dom_op():
+                    cp_dom_op = cp
+            self.ops.clear()
+            self.ops.append(cp_dom_op)
+            self.ops.extend((op for op in cp_ops if op != cp_dom_op))
+
+        def fuse_done(self, dom):
+            """do some thing after fused to dom"""
+            del dom
+            self.fuse_success = True
+
+        def reset(self, dom_area, ops, user_area, pre_area):
+            """set the recompute area and connect with other areas"""
+            self.recompute_ops.extend(ops)
+            # recom_area: set dom_op and correct ops length
+            patterns = list(PrimLib.iter_type(op) for op in ops)
+            self.pattern = max(patterns)
+            for i, pat in enumerate(patterns):
+                if pat == self.pattern:
+                    self.ops = [ops[i]] * len(ops)
+                    break
+            # disconnect dom_area and user_area
+            self.dom_user_r = dom_area.out_relations[user_area]
+            dom_area.out_relations.pop(user_area)
+            user_area.in_relations.pop(dom_area)
+            # connect recom_area and user_area
+            user_area.in_relations[self] = self.dom_user_r
+            self.out_relations[user_area] = self.dom_user_r
+            # connect recom_pre and recom_area
+            self.recom_pre = pre_area
+            if self.recom_pre is not None:
+                self.in_relations[self.recom_pre] = dom_area.in_relations[self.recom_pre]
+                self.recom_pre.out_relations[self] = dom_area.in_relations[self.recom_pre]
+            # set related areas
+            self.recom_user = user_area
+            self.recom_dom = dom_area
+            self.fuse_success = False
+
+        def clear(self):
+            """disconnect recom_area from other areas, and clear recom_area"""
+            self.out_relations.clear()
+            self.in_relations.clear()
+            if not self.fuse_success:
+                self.recom_user.in_relations.pop(self)
+                self.recom_user.in_relations[self.recom_dom] = self.dom_user_r
+                self.recom_dom.out_relations[self.recom_user] = self.dom_user_r
+                if self.recom_pre:
+                    self.recom_pre.out_relations.pop(self)
+            self.ops.clear()
+            self.recompute_ops.clear()
+            self.recom_map.update(self.ori_op_map)
+            self.ori_op_map.clear()
+
+    def __init__(self, graph, flags):
+        self.graph = graph
+        self.areas = []
+        self.flags = flags
+        self.enable_recompute = self.flags.get("enable_recompute_fusion", False)
+        self.enable_stitch_fusion = self.flags.get("enable_stitch_fusion", False)
+        self.enable_horizontal_fusion = self.flags.get("enable_horizontal_fusion", False)
+        self.reduce_fuse_depth = self.flags.get("reduce_fuse_depth", -1)
+        self.reach_tab = self.ReachTable(len(graph.ops) + 1 if self.enable_recompute else len(graph.ops))
+        self.area_map = {}
+        _, outputs = graph.deduce_parameters()
+        idx = 0
+        for op in graph.ops:
+            is_output = op.output in outputs
+            a = self.Area(op, is_output, idx, self.reach_tab)
+            idx += 1
+            self.set_default_mode(a)
+            self.areas.append(a)
+            self.set_area_map([op], a)
+        for a in self.areas:
+            a.link_input(self.area_map)
+        for i in range(len(self.areas) - 1, -1, -1):
+            self.areas[i].link_output()
+        if self.enable_recompute:
+            self.recom_area = self.RecomputeArea(idx, self.reach_tab)
+
+    def set_area_map(self, ops, area):
+        """update area_map after op fused to area"""
+        for op in ops:
+            self.area_map[op] = area
+
+    def set_default_mode(self, area):
+        """Set default mode"""
+        area.mode = self.get_default_mode(area.ops[0])
+
+    @staticmethod
+    def limit_area_size(dominant, fuse_areas, limit_size):
+        """Remove some areas if the size is too large"""
+        area_sizes = map(lambda area: len(area.ops), fuse_areas)
+        dom_size = len(dominant.ops)
+        if dom_size + prod_reduce(lambda x, y: x + y, area_sizes) <= limit_size:
+            return fuse_areas
+        # fuse the smaller area in priority
+        fuse_areas.sort(key=lambda area: len(area.ops))
+        new_fuse_areas = []
+        for area in fuse_areas:
+            if dom_size + len(area.ops) > limit_size:
+                break
+            dom_size += len(area.ops)
+            new_fuse_areas.append(area)
+        return new_fuse_areas
+
+    def fuse(self, selector, is_stitch=False):
+        """Fuse areas"""
+
+        def _fuse_area():
+            for dominant in self.areas:
+                result = selector(dominant)
+                if not result or not result[0]:
+                    continue
+                fuse_areas, is_forward = result
+                if not is_stitch:
+                    fuse_areas = self.limit_area_size(dominant, fuse_areas, self.flags['composite_op_limit_size'])
+                    if not fuse_areas:
+                        continue
+                changed = False
+                if is_forward:
+                    for area in fuse_areas:
+                        if is_stitch or dominant.fuse_confirm(area):
+                            dominant.fuse(area)
+                            self.set_area_map(area.ops, dominant)
+                            self.areas.remove(area)
+                            changed = True
+                else:
+                    forward_area = dominant
+                    for area in fuse_areas:
+                        if is_stitch or area.fuse_confirm(forward_area):
+                            area.fuse(forward_area)
+                            self.set_area_map(forward_area.ops, area)
+                            self.areas.remove(forward_area)
+                            forward_area = area
+                            changed = True
+                if changed:
+                    return True
+            return False
+
+        changed, do_again = False, True
+        while do_again:
+            do_again = _fuse_area()
+            changed = changed or do_again
+        return changed
+
+    def hfuse(self, selector):
+        """Fuse horizontal areas with same input tensor"""
+
+        def _do_fuse(areas):
+            for i in range(len(areas) - 1):
+                dom = areas[i]
+                for a in areas[i + 1:]:
+                    can_fuse = dom.check_acyclic(a) and a.check_acyclic(dom) and selector(dom, a) \
+                        and self.limit_area_size(dom, [a], 64) and dom.fuse_confirm(a)
+                    if can_fuse:
+                        dom.fuse(a)
+                        self.set_area_map(a.ops, dom)
+                        self.areas.remove(a)
+                        return True
+            return False
+
+        def _update_areas(areas, from_op):
+            for op in from_op.to_ops:
+                a = self.area_map.get(op)
+                if a in self.areas and a not in areas:
+                    areas.append(a)
+
+        changed = False
+        while True:
+            for dom in self.areas:
+                if len(dom.out_relations) > 1 and _do_fuse(list(dom.out_relations.keys())):
+                    changed = True
+                    break
+            else:
+                break
+        inputs, _ = self.graph.deduce_parameters()
+        while True:
+            for t in inputs:
+                areas = []
+                _update_areas(areas, t)
+                if len(areas) > 1 and _do_fuse(areas):
+                    changed = True
+                    break
+            else:
+                break
+        return changed
+
+    def fuse_recom(self, selector):
+        """Fuse recompute area to its user"""
+        user = self.recom_area.recom_user
+        for dominant in [self.recom_area, user]:
+            result = selector(dominant)
+            if result and result[0]:
+                fuse_areas, _ = result
+                fuse_areas = self.limit_area_size(dominant, fuse_areas, self.flags['composite_op_limit_size'])
+                if not fuse_areas:
+                    continue
+                if fuse_areas[0] in [self.recom_area, user] and user.fuse_confirm(self.recom_area):
+                    user.fuse(self.recom_area)
+                    self.set_area_map(self.recom_area.ops, user)
+                    return True
+        return False
+
+    def index_op(self):
+        """index op by order, the copied op share id with original op, for topo-sort"""
+        ids = {}
+        for i, op in enumerate(self.graph.ops):
+            ids[op] = i
+        if self.enable_recompute:
+            for k, v in self.recom_area.recom_map.items():
+                ids[k] = ids.get(v)
+        return ids
+
+    def to_subgraphs(self):
+        """Transform op groups to subgraphs"""
+        ids = self.index_op()
+        subgraphs = []
+        graphmodes = []
+        for i, area in enumerate(self.areas):
+            area.ops.sort(key=ids.get)
+            subgraphs.append(Graph('{}_{}'.format(self.graph.name, i), area.ops, area.stitch_info, area.recompute_ops))
+            graphmodes.append("basic" if area.mode == self.Area.MODE_BASIC else "composite")
+        return subgraphs, graphmodes
+
+    def pattern_fuse(self, fuse_func=None):
+        """fuse Areas by pattern repeatedly"""
+        del fuse_func
+        raise Exception("pattern_fuse() is not implemented in {}".format(self.__class__.__name__))
+
+    def split(self):
+        """Split graph by pattern"""
+        self.pattern_fuse()
+        if self.enable_recompute:
+            self.recompute_fuse()
+        # The reshape should not be output node
+        # Note: after this function, the input output relation is not maintained.
+        self.split_output_reshapes()
+        subgraphs, graphmodes = self.to_subgraphs()
+        return subgraphs, graphmodes
+
+    def split_output_reshapes(self):
+        """Force split the output Reshapes into other new area"""
+
+        def _remove_output_reshape(reshape_ops, other_ops):
+            def _run():
+                for op in reshape_ops:
+                    if any((to_op in other_ops for to_op in op.output.to_ops)):
+                        reshape_ops.remove(op)
+                        other_ops.append(op)
+                        return True
+                return False
+
+            while _run():
+                pass
+
+        new_areas = []
+        for area in self.areas:
+            reshape_ops = list(op for op in area.ops if PrimLib.iter_type(op) == PrimLib.RESHAPE)
+            other_ops = list(op for op in area.ops if op not in reshape_ops)
+            if not other_ops or not reshape_ops:
+                continue
+            # remove the output reshape from "reshape_ops" and add it into "other_ops"
+            _remove_output_reshape(reshape_ops, other_ops)
+            if not reshape_ops:
+                continue
+            for op in reshape_ops:
+                a = self.Area(op, False, 0, self.reach_tab)
+                self.set_default_mode(a)
+                new_areas.append(a)
+            area.ops = other_ops
+            if len(other_ops) == 1:
+                self.set_default_mode(area)
+        if new_areas:
+            self.areas += new_areas
+
+    def recompute_fuse(self):
+        """find recompute regions and copy them out to new Areas"""
+
+        def _get_prods(area, border):
+            """get producer region of border op"""
+            max_weight = 10
+            stack = [border]
+            ops, inputs = [], []
+            while stack:
+                op = stack.pop()
+                if len(op.inputs) > 1 or PrimLib.iter_type(op) > PrimLib.BROADCAST or len(ops) > max_weight:
+                    return [], []
+                ops.append(op)
+                for t in op.inputs:
+                    if t.op in area.ops:
+                        stack.append(t.op)
+                    else:
+                        inputs.append(t)
+            return ops, inputs
+
+        def _get_border_info(area):
+            """get border information"""
+            prods, users = {}, {}
+            for op in area.ops:
+                if len(op.output.to_ops) <= 1 and op.output.para_type != Tensor.PARA_OUTPUT:
+                    continue
+                for to in op.output.to_ops:
+                    if to in area.ops:
+                        continue
+                    user = self.area_map.get(to)
+                    if user.pattern > PrimLib.RESHAPE:
+                        if user in users:
+                            users.get(user).append(op)
+                        else:
+                            users[user] = [op]
+                        if op not in prods:
+                            prods[op] = _get_prods(area, op)
+            return prods, users
+
+        def _get_cheap_region(prods, borders):
+            """get cheap region of border ops"""
+            if len(borders) > 1:
+                return []
+            result = []
+            for op in borders:
+                prod_ops, inputs = prods[op]
+                if prod_ops:
+                    if sum([t.get_size() for t in inputs]) <= op.output.get_size():
+                        pred = self.area_map.get(inputs[0].op) if inputs and inputs[0].op else None
+                        result.append([pred, prod_ops[::-1]])
+            return result
+
+        def _do_recompute(area):
+            """split the unfusing pattern by add recompute area"""
+            prods, users = _get_border_info(area)
+            for user, borders in users.items():
+                result = _get_cheap_region(prods, borders)
+                for pred, region in result:
+                    self.recom_area.reset(area, region, user, pred)
+                    self.pattern_fuse(self.fuse_recom)
+                    self.recom_area.clear()
+                    if self.recom_area.fuse_success:
+                        return True
+            return False
+
+        changed = True
+        while changed:
+            changed = False
+            orig_areas = []
+            orig_areas.extend(self.areas)
+            for area in orig_areas:
+                if area in self.areas and area.out_relations:
+                    changed = _do_recompute(area) or changed
+            if changed:
+                self.pattern_fuse()
+
+
+class GraphSplitGpu(GraphSplitByPattern):
+    """Graph splitter"""
+    BROADCAST_FUSE_DEPTH = 20
+    TRANSPOSE_FUSE_DEPTH = 6
+
+    def __init__(self, graph, flags):
+        super().__init__(graph, flags)
+        self.reduce_fuse_depth = 20 if self.reduce_fuse_depth < 0 else self.reduce_fuse_depth
+
+    def get_default_mode(self, op):
+        """Get default mode in GPU"""
+        if op.prim == "MatMul":
+            return self.Area.MODE_COMPOSITE if op.inputs[0].dtype == "float16" and op.attrs['Akg'] else \
+                self.Area.MODE_BASIC
+        if op.prim == "Assign":
+            return self.Area.MODE_BASIC
+        pattern = PrimLib.iter_type(op)
+        return self.Area.MODE_BASIC if pattern == PrimLib.RESHAPE else self.Area.MODE_COMPOSITE
+
+    def pattern_fuse(self, fuse_func=None):
+        """fuse Areas by pattern"""
+
+        def _broadcast_pat_exclude(dom, a, r):
+            if a.pattern == PrimLib.REDUCE:
+                return dom.pattern > PrimLib.ELEMWISE or r > PrimLib.ELEMWISE
+            return a.pattern > PrimLib.REDUCE or r > PrimLib.BROADCAST
+
+        def _broadcast_bwd_depth(dom):
+            if dom.pattern not in (PrimLib.ELEMWISE, PrimLib.BROADCAST) or len(dom.out_relations) != 1:
+                return [], False
+            if dom.is_output or len(dom.ops) > self.BROADCAST_FUSE_DEPTH:
+                return [], False
+            a, r = list(dom.out_relations.items())[0]
+            if _broadcast_pat_exclude(dom, a, r) or len(a.in_relations) != 1:
+                return [], False
+            return [a], False
+
+        def _broadcast_bwd_width(dom):
+            if dom.pattern not in (PrimLib.ELEMWISE, PrimLib.BROADCAST) or \
+                    dom.is_output or len(dom.ops) > self.BROADCAST_FUSE_DEPTH:
+                return [], False
+            fused = []
+            for a, r in dom.out_relations.items():
+                if _broadcast_pat_exclude(dom, a, r) or not dom.check_acyclic(a):
+                    return [], False
+                if fused and tensor_size(fused[0].dom_op().output) != tensor_size(a.dom_op().output):
+                    return [], False
+                fused.append(a)
+            return fused, False
+
+        def _reduce_pat_exclude(_, a, r):
+            if len(a.ops) > self.reduce_fuse_depth:
+                return True
+            return a.pattern > PrimLib.ELEMWISE or r > PrimLib.REDUCE or r == PrimLib.BROADCAST
+
+        def _reduce_depth(dom):
+            if dom.pattern != PrimLib.REDUCE or len(dom.in_relations) != 1:
+                return [], False
+            a, r = list(dom.in_relations.items())[0]
+            if dom.ops[0].inputs[0].dtype == "float16" and a.is_output:
+                if len(a.ops) >= 10 and _is_atomic_add_available(dom):
+                    # to evade the precision problem.
+                    return [], False
+            if _reduce_pat_exclude(dom, a, r) or len(a.out_relations) != 1:
+                return []
+            return [a], True
+
+        def _reduce_width(dom):
+            if dom.pattern != PrimLib.REDUCE:
+                return [], False
+            fused = []
+            for a, r in dom.in_relations.items():
+                if dom.ops[0].inputs[0].dtype == "float16" and a.is_output:
+                    if len(a.ops) >= 10 and _is_atomic_add_available(dom):
+                        # to evade the precision problem.
+                        continue
+                if not _reduce_pat_exclude(dom, a, r) and a.check_acyclic(dom):
+                    fused.append(a)
+            return fused, True
+
+        def _is_atomic_add_available(dom):
+            if any(("Reduce" in x.prim for x in dom.ops[1:])):
+                return False
+            op = dom.ops[0]
+            if "reduce_axis" in op.attrs:
+                reduce_axis = op.attrs["reduce_axis"]
+            elif "axis" in op.attrs:
+                reduce_axis = [op.attrs["axis"]]
+            else:
+                raise Exception("For '{}', can not find the attr 'reduce_axis' or 'axis'".format(op.prim))
+            if op.inputs and len(op.inputs[0].shape) - 1 in reduce_axis:
+                reduce_size = prod_reduce(lambda x, y: x * y, (op.inputs[0].shape[i] for i in reduce_axis))
+                return reduce_size >= 1024
+            return True
+
+        def _may_multi_filter(dom_ops):
+            count = 1
+            stack = [dom_ops[0]]
+            while stack:
+                op = stack.pop()
+                for t in op.inputs:
+                    if t.op and t.op in dom_ops:
+                        count = count + 1
+                        stack.append(t.op)
+            return count < len(dom_ops)
+
+        def _reduce_output(dom):
+            if dom.pattern != PrimLib.REDUCE:
+                return [], False
+            if _may_multi_filter(dom.ops):
+                return [], False
+            if _is_atomic_add_available(dom):
+                return [], False
+            is_all_reduce = tensor_size(dom.ops[0].output) == 1
+            # excluded large size all reduce
+            if is_all_reduce and dom.ops[0].inputs and tensor_size(dom.ops[0].inputs[0]) > 1024 * 12:
+                return [], False
+
+            fused = []
+            for a, r in dom.out_relations.items():
+                if a.pattern <= PrimLib.BROADCAST and r <= PrimLib.BROADCAST and dom.check_acyclic(a):
+                    fused.append(a)
+            return fused, False
+
+        def _reduce_stitch(dom):
+            if dom.pattern != PrimLib.REDUCE:
+                return [], False
+            if tensor_size(dom.ops[0].output) == 1:
+                return [], False
+            if tensor_size(dom.ops[0].inputs[0]) < 1024 * 12:
+                return [], False
+
+            fused = []
+            for a, r in dom.out_relations.items():
+                if not may_stitch(dom, a, r, 1024 * 8, 1024 * 1024):
+                    continue
+                if a.pattern == PrimLib.REDUCE:
+                    if a.ops[0].attrs['reduce_axis'] == dom.ops[0].attrs['reduce_axis']:
+                        dom.stitch_info.stitch_ops.add(dom.ops[0].output.name)
+                        fused.append(a)
+                elif a.pattern == PrimLib.BROADCAST:
+                    dom.stitch_info.stitch_ops.add(dom.ops[0].output.name)
+                    fused.append(a)
+            return fused, False
+
+        def _transpose(dom):
+            if len(dom.ops) != 1 or dom.ops[0].prim != "Transpose":
+                return [], False
+            fused = []
+            for a, _ in dom.in_relations.items():
+                if a.pattern <= PrimLib.BROADCAST and a.check_acyclic(dom) and len(a.ops) <= self.TRANSPOSE_FUSE_DEPTH:
+                    fused.append(a)
+            return fused, True
+
+        def _strided_slice(dom):
+            if dom.dom_op().prim != "StridedSlice":
+                return [], False
+            fused = []
+            for a, _ in dom.in_relations.items():
+                if a.pattern <= PrimLib.BROADCAST and a.check_acyclic(dom) and \
+                        len(a.out_relations) == 1 and not a.is_output:
+                    fused.append(a)
+            return fused, True
+
+        def _gather_output(dom, reduce_fusion=False):
+            gather_prims = ("Gather", "GatherNd", "CSRGather")
+            if not dom.dom_op().prim in gather_prims:
+                return [], False
+
+            def _reduce_exclude(op, axis_list):
+                """ Whether this operator should be excluded.
+                Excluding condition:
+                1. There are at least one same axis between reduce axes and axis_list.
+
+                Args:
+                    op (Operator): Target reduce operator.
+                    axis_list (list): List to check whether it is intersected by reduce axis.
+                Returns:
+                    Boolean. Whether this operator should be excluded.
+                """
+                axis = op.attrs["reduce_axis"]
+                if isinstance(axis, int):
+                    axis = [axis]
+                in_shape_len = len(op.inputs[0].shape)
+                for i, dim in enumerate(axis):
+                    axis[i] = in_shape_len + dim if dim < 0 else dim
+                fix_axis = []
+                for ax in axis:
+                    if op.inputs[0].shape[ax] == 1:
+                        continue
+                    fix_axis.append(ax)
+                return bool(set(fix_axis) & set(axis_list))
+
+            def _bfs_visit(start_op, start_prims, total_ops, end_ops, gather_axis):
+                consisten_shape = start_op.output.shape
+                visited = []
+                op_queue = [start_op]
+
+                def _early_stop(cur_op):
+                    if cur_op in end_ops:
+                        # If reduce the gather axis, stop early for not fusion.
+                        if cur_op.prim == "ReduceSum" and _reduce_exclude(cur_op, gather_axis):
+                            return True
+                    else:
+                        if (cur_op.prim in start_prims and cur_op != start_op) or \
+                                consisten_shape != cur_op.output.shape:
+                            return True
+                    return False
+
+                while op_queue:
+                    tmp_queue = []
+                    for op in op_queue:
+                        if op in visited or op not in total_ops:
+                            continue
+                        if _early_stop(op):
+                            return False
+                        if op in end_ops:
+                            continue
+                        for to_op in op.output.to_ops:
+                            tmp_queue.append(to_op)
+                        visited.append(op)
+                    op_queue = tmp_queue
+                return True
+
+            def _shape_consistent(start_prims, end_prims, source, target):
+                """
+                Check whether it is always shape consistent from source nodes to target nodes.
+                Excluding condition:
+                    When fusing ReduceSum, first check if TensorScatterAdd and/or UnsortedSegmentSum
+                    has already been fused, if so, stop ReduceSum fusion.
+                """
+                total_ops = source.ops + target.ops
+                op_prims_set = {op.prim for op in total_ops}
+                if reduce_fusion and (len({"TensorScatterAdd", "UnsortedSegmentSum"} & op_prims_set) >= 1):
+                    return False
+                start_ops = []
+                for op in source.ops:
+                    if op.prim in start_prims:
+                        start_ops.append(op)
+                end_ops = []
+                for op in total_ops:
+                    if op.prim in end_prims and not any((to_op in total_ops for to_op in op.output.to_ops)):
+                        end_ops.append(op)
+
+                for start_op in start_ops:
+                    gather_axis = start_op.attrs.get("axis", None)
+                    if gather_axis is None:
+                        # For GatherNd
+                        gather_axis = list(range(len(start_op.inputs[1].shape)))
+                    elif isinstance(gather_axis, int):
+                        gather_axis = [gather_axis]
+
+                    is_consistent = _bfs_visit(start_op, start_prims, total_ops, end_ops, gather_axis)
+                    if not is_consistent:
+                        return False
+                return True
+
+            if reduce_fusion:
+                appected_areas = {"ReduceSum", "CSRReduceSum"}
+            else:
+                appected_areas = {"TensorScatterAdd", "UnsortedSegmentSum"}
+
+            for a, _ in dom.out_relations.items():
+                if _shape_consistent(gather_prims, appected_areas, dom, a) and dom.check_acyclic(a):
+                    return [a], False
+            return [], False
+
+        def _broadcast_tot(dom):
+            """Fuse rule for TensorScatterAdd and UnsortedSegmentSum."""
+
+            def _same_input(op1, op2):
+                return bool(set(op1.inputs) & set(op2.inputs))
+
+            if len(dom.ops) != 1:
+                return [], False
+
+            # Only fuse the first input for `TensorScatterAdd`` and the first and second input for `UnsortedSegmentSum`.
+            fuse_arg = {"TensorScatterAdd": slice(1, None), "UnsortedSegmentSum": slice(0, 2)}
+            arg_idx = fuse_arg.get(dom.dom_op().prim, -1)
+            if arg_idx == -1:
+                return [], False
+            fuse_tensor = dom.dom_op().inputs[arg_idx]
+
+            for a, _ in dom.in_relations.items():
+                if not a.check_acyclic(dom):
+                    continue
+                # Rule 1: Same type with at lease one same input.
+                if a.dom_op().prim == dom.dom_op().prim and _same_input(dom.dom_op(), a.dom_op()):
+                    return [a], True
+                # Rule 2: Fuse op(reshape/elementwise/broadcast) in specified position inputs.
+                if a.pattern <= PrimLib.BROADCAST and any((op.output in fuse_tensor for op in a.ops)):
+                    return [a], True
+            return [], False
+
+        def _broadcast_onehot(dom, fwd=True):
+            """Fuse rule for OneHot."""
+            if dom.dom_op().prim != "OneHot":
+                return [], False
+
+            fused = []
+            neighbours = dom.in_relations.items() if fwd else dom.out_relations.items()
+            for a, _ in neighbours:
+                if a.pattern <= PrimLib.BROADCAST:
+                    if fwd:
+                        if a.check_acyclic(dom) and len(a.out_relations) == 1 and not a.is_output:
+                            fused.append(a)
+                    else:
+                        if dom.check_acyclic(a):
+                            fused.append(a)
+
+            return fused, fwd
+
+        def _elemwise_elemany(dom):
+            """Fuse rule for elemany."""
+            if dom.dom_op().prim != "ElemAny":
+                return [], False
+
+            fused = []
+            for a, r in dom.in_relations.items():
+                if a.pattern < PrimLib.BROADCAST and r <= PrimLib.ELEMWISE and a.check_acyclic(dom):
+                    fused.append(a)
+
+            return fused, True
+
+        def _injective_output(dom):
+            """Fuse rule for injective """
+            injective_ops = {"Transpose", "StridedSlice"}
+            if dom.dom_op().prim not in injective_ops:
+                return [], False
+            to_ops = dom.dom_op().output.to_ops
+            if dom.is_output or len(to_ops) != 1 or len(dom.out_relations) != 1:
+                return [], False
+            to_area = list(dom.out_relations.keys())[0]
+            if (to_area.pattern >= PrimLib.REDUCE and to_area.dom_op().prim not in injective_ops) or \
+                    to_ops[0] not in to_area.ops:
+                return [], False
+            if len(to_area.ops) > self.TRANSPOSE_FUSE_DEPTH:
+                return [], False
+            return [to_area], False
+
+        def _h_broadcast(dom, a):
+            if dom.pattern > PrimLib.BROADCAST:
+                return [], False
+            return a.pattern <= PrimLib.BROADCAST and dom.ops[0].output.shape == a.ops[0].output.shape
+
+        def _h_reduce(dom, a):
+            if dom.pattern != PrimLib.REDUCE or dom.stitch_info.stitch_ops:
+                return []
+            dom_op = dom.ops[0]
+            if not PrimLib.is_reduce(dom_op) or _is_atomic_add_available(dom):
+                return []
+            op = a.ops[0]
+            return a.pattern == PrimLib.REDUCE and not a.stitch_info.stitch_ops and \
+                PrimLib.is_reduce(op) and dom_op.inputs[0].shape == op.inputs[0].shape and \
+                dom_op.attrs.get("reduce_axis") == op.attrs.get("reduce_axis")
+
+        def _h_opaque(dom, a):
+            if dom.ops[0].prim not in {"StridedSlice"}:
+                return []
+            return a.ops[0].prim == dom.ops[0].prim and dom.ops[0].output.shape == a.ops[0].output.shape and \
+                dom.ops[0].inputs[0].shape == a.ops[0].inputs[0].shape
+
+        def _link_csr(dom):
+            def _same_input(op1, op2):
+                return bool(set(op1.inputs.copy()) & set(op2.inputs.copy()))
+
+            fuse_arg = {"CSRReduceSum": slice(1, 3), "CSRGather": slice(2, 3)}
+            arg_idx = fuse_arg.get(dom.dom_op().prim, -1)
+            if arg_idx == -1:
+                return [], False
+            fuse_tensor = dom.dom_op().inputs[arg_idx]
+            for a, _ in dom.in_relations.items():
+                if (a.dom_op().prim == "CSRGather" and a.dom_op().prim == dom.dom_op().prim and
+                        _same_input(dom.dom_op(), a.dom_op())):
+                    return [a], True
+                if a.pattern <= PrimLib.BROADCAST and dom.check_acyclic(a) and \
+                        any([op.output in fuse_tensor for op in a.ops]):
+                    return [a], True
+            return [], False
+
+        def _fuse_loop():
+            self.fuse(CommonPattern.reshape)
+            self.fuse(CommonPattern.assign)
+            self.fuse(CommonPattern.elemwise_depth)
+            self.fuse(CommonPattern.elemwise_width)
+            self.fuse(_broadcast_tot)
+            self.fuse(_link_csr)
+            self.fuse(CommonPattern.broadcast_depth)
+            self.fuse(CommonPattern.broadcast_width)
+            self.fuse(_reduce_depth)
+            self.fuse(_reduce_width)
+            self.fuse(_broadcast_bwd_depth)
+            self.fuse(_broadcast_bwd_width)
+            self.fuse(_strided_slice)
+            self.fuse(partial(_broadcast_onehot, fwd=True))
+            self.fuse(partial(_broadcast_onehot, fwd=False))
+            self.fuse(partial(_gather_output, reduce_fusion=False))
+            self.fuse(partial(_gather_output, reduce_fusion=True))
+            self.fuse(_reduce_output)
+            if self.enable_stitch_fusion:
+                self.fuse(_reduce_stitch, True)
+            self.fuse(_transpose)
+            self.fuse(_injective_output)
+            self.fuse(CommonPattern.isolate_reshape)
+            if self.enable_horizontal_fusion:
+                self.hfuse(_h_broadcast)
+                self.hfuse(_h_reduce)
+                self.hfuse(_h_opaque)
+            self.fuse(_elemwise_elemany)
+
+        def _fuse_once(fuse_func):
+            if fuse_func(CommonPattern.reshape) or \
+                    fuse_func(CommonPattern.elemwise_depth) or fuse_func(CommonPattern.elemwise_width) or \
+                    fuse_func(CommonPattern.broadcast_depth) or fuse_func(CommonPattern.broadcast_width) or \
+                    fuse_func(_reduce_depth) or fuse_func(_reduce_width) or \
+                    fuse_func(_broadcast_bwd_depth) or fuse_func(_broadcast_bwd_width):
+                return
+            if fuse_func(_reduce_output):
+                return
+            fuse_func(_transpose)
+
+        if fuse_func is None:
+            _fuse_loop()
+        else:
+            _fuse_once(fuse_func)
+
+
+def split(graph, target, flags):
+    """Split graph"""
+    result = None
+    if target == "cuda":
+        result = GraphSplitGpu(graph, flags).split()
+    return result