1 | /******************************************************************************* |
---|---|
2 | * Copyright 2022 Intel Corporation |
3 | * |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
5 | * you may not use this file except in compliance with the License. |
6 | * You may obtain a copy of the License at |
7 | * |
8 | * http://www.apache.org/licenses/LICENSE-2.0 |
9 | * |
10 | * Unless required by applicable law or agreed to in writing, software |
11 | * distributed under the License is distributed on an "AS IS" BASIS, |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | * See the License for the specific language governing permissions and |
14 | * limitations under the License. |
15 | *******************************************************************************/ |
16 | |
17 | #include "gpu/jit/pass/barrier.hpp" |
18 | |
19 | #include "gpu/jit/ir/message.hpp" |
20 | #include "gpu/jit/utils/trace.hpp" |
21 | |
22 | namespace dnnl { |
23 | namespace impl { |
24 | namespace gpu { |
25 | namespace jit { |
26 | |
27 | class barrier_optimizer_t : public ir_mutator_t { |
28 | public: |
29 | object_t _mutate(const for_t &obj) override { |
30 | loop_level_++; |
31 | auto new_obj = ir_mutator_t::_mutate(obj); |
32 | loop_level_--; |
33 | return new_obj; |
34 | } |
35 | |
36 | object_t _mutate(const func_call_t &obj) override { |
37 | if (is_func_call<send_t>(obj)) { |
38 | auto &send = obj.func.as<send_t>(); |
39 | if (send.is_slm()) can_remove_barrier_ = false; |
40 | } else if (obj.func.is_same(funcs::barrier_func())) { |
41 | bool can_remove = can_remove_barrier_; |
42 | can_remove_barrier_ = false; |
43 | |
44 | // If not in a loop and this is the first barrier -> can be removed. |
45 | if (loop_level_ == 0 && can_remove) return stmt_t(); |
46 | return obj; |
47 | } |
48 | |
49 | return obj; |
50 | } |
51 | |
52 | // Store doesn't contain nested statements, return as is. |
53 | object_t _mutate(const store_t &obj) override { return obj; } |
54 | |
55 | private: |
56 | int loop_level_ = 0; |
57 | bool can_remove_barrier_ = true; |
58 | }; |
59 | |
60 | stmt_t optimize_barrier(const stmt_t &s, ir_context_t &ir_ctx) { |
61 | trace_start(); |
62 | auto ret = barrier_optimizer_t().mutate(s); |
63 | trace_pass("optimize_barrier", ret, ir_ctx); |
64 | return ret; |
65 | } |
66 | |
67 | } // namespace jit |
68 | } // namespace gpu |
69 | } // namespace impl |
70 | } // namespace dnnl |
71 |