1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
/*******************************************************************************
* Copyright 2022 Intel Corporation
*
* 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.
*******************************************************************************/
#include "graph/backend/dnnl/kernels/layer_norm.hpp"
#include "graph/backend/dnnl/patterns/fusions.hpp"
#include "graph/backend/dnnl/patterns/pattern_matcher_pass.hpp"
#include "graph/backend/dnnl/patterns/utils.hpp"
namespace dnnl {
namespace impl {
namespace graph {
namespace dnnl_impl {
namespace pattern {
namespace pm = graph::utils::pm;
using in_edges_t = pm::in_edges_t;
using pb_graph_t = pm::pb_graph_t;
using FCreatePattern = graph::pass::FCreatePattern;
// LayerNorm
// |
// [TypeCast]*
// |
// [unary/binary]*[0,MAX_REPETITION)
// |
// [Quantize]*
DNNL_BACKEND_REGISTER_PATTERN_DEF_BEGIN(layernorm_fusion)
#if DNNL_CPU_RUNTIME != DNNL_RUNTIME_NONE
DNNL_BACKEND_REGISTER_PATTERN_MATCHER_PASS(dnnl, layernorm_post_ops_fusion_cpu)
.set_priority(8.2f)
.set_kind(graph::partition_kind_t::misc_post_ops)
.set_engine_kind(engine_kind::cpu)
.set_attr<FCreatePattern>("FCreatePattern",
[](const std::shared_ptr<pb_graph_t> &pgraph) -> void {
pm::pb_op_t *layernorm_base
= pgraph->append_op(graph::op_kind::LayerNorm);
layernorm_base->append_decision_function(
check_input_dtype_from_offset<impl::data_type::f32,
1>);
layernorm_base->append_decision_function(
check_begin_norm_axis_attr);
// primitive only support 2-5D data tensor for layernorm
layernorm_base->append_decision_function(
check_input_ndim_from_offset<0, 2, 5>);
// optional typecast
auto tc_graph = std::make_shared<pb_graph_t>();
pm::pb_op_t *ptypecast
= tc_graph->append_op(graph::op_kind::TypeCast);
tc_graph->create_input_port(0, ptypecast, 0);
tc_graph->create_output_port(0, ptypecast, 0);
auto pre_tc = pgraph->append_optional(tc_graph,
in_edges_t {in_edge(0, layernorm_base, 0)});
// repetition(alternation(unary | binary))
auto alt_unary_binary = std::make_shared<pb_graph_t>();
auto palt = alt_unary_binary->append_alternation(
get_unary_binary_ops());
palt->allow_internal_inputs();
alt_unary_binary->create_input_port(0, palt, 0);
alt_unary_binary->create_output_port(0, palt, 0);
auto prep = pgraph->append_repetition(alt_unary_binary,
{0, 0}, 0, MAX_REPETITION,
in_edges_t {in_edge(0, pre_tc, 0)});
// optional quantize
auto q_graph = std::make_shared<pb_graph_t>();
pm::pb_op_t *pquantize
= q_graph->append_op(graph::op_kind::Quantize);
q_graph->create_input_port(0, pquantize, 0);
q_graph->create_output_port(0, pquantize, 0);
pgraph->append_optional(
q_graph, in_edges_t {in_edge(0, prep, 0)});
})
.set_attr<FCreateKernel>("FCreateKernel", []() -> kernel_ptr {
return std::make_shared<layer_norm_fwd_t>();
});
#endif
DNNL_BACKEND_REGISTER_PATTERN_DEF_END
} // namespace pattern
} // namespace dnnl_impl
} // namespace graph
} // namespace impl
} // namespace dnnl