1 | /******************************************************************************* |
2 | * Copyright 2017-2022 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 | #ifndef CPU_X64_JIT_UNI_BATCH_NORMALIZATION_HPP |
18 | #define CPU_X64_JIT_UNI_BATCH_NORMALIZATION_HPP |
19 | |
20 | #include <assert.h> |
21 | |
22 | #include "common/c_types_map.hpp" |
23 | #include "common/primitive.hpp" |
24 | #include "common/type_helpers.hpp" |
25 | #include "common/utils.hpp" |
26 | |
27 | #include "cpu/cpu_batch_normalization_pd.hpp" |
28 | #include "cpu/x64/cpu_isa_traits.hpp" |
29 | #include "cpu/x64/jit_avx512_core_bf16cvt.hpp" |
30 | |
31 | namespace dnnl { |
32 | namespace impl { |
33 | namespace cpu { |
34 | namespace x64 { |
35 | |
36 | namespace bnorm_impl { |
37 | template <cpu_isa_t isa> |
38 | struct driver_t; |
39 | } |
40 | |
41 | template <cpu_isa_t isa> |
42 | struct jit_uni_batch_normalization_fwd_t : public primitive_t { |
43 | struct pd_t : public cpu_batch_normalization_fwd_pd_t { |
44 | pd_t(const batch_normalization_desc_t *adesc, |
45 | const primitive_attr_t *attr, |
46 | const batch_normalization_fwd_pd_t *hint_fwd_pd) |
47 | : cpu_batch_normalization_fwd_pd_t(adesc, attr, hint_fwd_pd) {} |
48 | |
49 | DECLARE_COMMON_PD_T( |
50 | JIT_IMPL_NAME_HELPER("bnorm_jit:" , |
51 | (src_md()->data_type == data_type::bf16) |
52 | ? (mayiuse(avx512_core_bf16) |
53 | ? avx512_core_bf16 |
54 | : mayiuse(avx512_core) |
55 | ? bf16_emulation_t:: |
56 | get_isa() |
57 | : avx2_vnni_2) |
58 | : (src_md()->data_type == data_type::f16) |
59 | ? (mayiuse(avx512_core_fp16) |
60 | ? avx512_core_fp16 |
61 | : avx2_vnni_2) |
62 | : isa, |
63 | "" ), |
64 | jit_uni_batch_normalization_fwd_t); |
65 | |
66 | status_t init(engine_t *engine); |
67 | int nthr_; // To not exceed the limit in execute used for set up. |
68 | }; |
69 | |
70 | jit_uni_batch_normalization_fwd_t(const pd_t *apd); |
71 | ~jit_uni_batch_normalization_fwd_t(); |
72 | |
73 | status_t init(engine_t *engine) override; |
74 | |
75 | status_t execute(const exec_ctx_t &ctx) const override; |
76 | |
77 | private: |
78 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
79 | |
80 | bnorm_impl::driver_t<isa> *bnorm_driver_; |
81 | }; |
82 | |
83 | template <cpu_isa_t isa> |
84 | struct jit_uni_batch_normalization_bwd_t : public primitive_t { |
85 | struct pd_t : public cpu_batch_normalization_bwd_pd_t { |
86 | pd_t(const batch_normalization_desc_t *adesc, |
87 | const primitive_attr_t *attr, |
88 | const batch_normalization_fwd_pd_t *hint_fwd_pd) |
89 | : cpu_batch_normalization_bwd_pd_t(adesc, attr, hint_fwd_pd) {} |
90 | |
91 | DECLARE_COMMON_PD_T( |
92 | JIT_IMPL_NAME_HELPER("bnorm_jit:" , |
93 | (src_md()->data_type == data_type::bf16) |
94 | ? (mayiuse(avx512_core_bf16) |
95 | ? avx512_core_bf16 |
96 | : bf16_emulation_t::get_isa()) |
97 | : (src_md()->data_type == data_type::f16) |
98 | ? avx512_core_fp16 |
99 | : isa, |
100 | "" ), |
101 | jit_uni_batch_normalization_bwd_t); |
102 | |
103 | status_t init(engine_t *engine); |
104 | int nthr_; // To not exceed the limit in execute used for set up. |
105 | }; |
106 | |
107 | jit_uni_batch_normalization_bwd_t(const pd_t *apd); |
108 | ~jit_uni_batch_normalization_bwd_t(); |
109 | |
110 | status_t init(engine_t *engine) override; |
111 | |
112 | status_t execute(const exec_ctx_t &ctx) const override; |
113 | |
114 | private: |
115 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
116 | |
117 | bnorm_impl::driver_t<isa> *bnorm_driver_; |
118 | }; |
119 | |
120 | } // namespace x64 |
121 | } // namespace cpu |
122 | } // namespace impl |
123 | } // namespace dnnl |
124 | |
125 | #endif |
126 | |
127 | // vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s |
128 | |