1 | /******************************************************************************* |
2 | * Copyright 2018-2022 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_AVX512_CORE_X8S8S32X_1X1_DECONVOLUTION_HPP |
18 | #define CPU_X64_JIT_AVX512_CORE_X8S8S32X_1X1_DECONVOLUTION_HPP |
19 | |
20 | #include "common/c_types_map.hpp" |
21 | #include "common/dnnl_thread.hpp" |
22 | #include "common/primitive.hpp" |
23 | #include "common/primitive_desc_iterator.hpp" |
24 | #include "common/type_helpers.hpp" |
25 | #include "common/utils.hpp" |
26 | |
27 | #include "cpu/cpu_convolution_pd.hpp" |
28 | #include "cpu/cpu_deconvolution_pd.hpp" |
29 | #include "cpu/zero_point_utils.hpp" |
30 | |
31 | #include "cpu/x64/jit_avx512_core_x8s8s32x_1x1_convolution.hpp" |
32 | #include "cpu/x64/jit_uni_1x1_conv_utils.hpp" |
33 | |
34 | namespace dnnl { |
35 | namespace impl { |
36 | namespace cpu { |
37 | namespace x64 { |
38 | |
39 | struct jit_avx512_core_x8s8s32x_1x1_deconvolution_fwd_t : public primitive_t { |
40 | struct pd_t : public cpu_deconvolution_fwd_pd_t { |
41 | pd_t(const deconvolution_desc_t *adesc, const primitive_attr_t *attr, |
42 | const deconvolution_fwd_pd_t *hint_fwd_pd) |
43 | : cpu_deconvolution_fwd_pd_t(adesc, attr, hint_fwd_pd) {} |
44 | |
45 | pd_t(const pd_t &other) |
46 | : cpu_deconvolution_fwd_pd_t(other) |
47 | , conv_pd_(other.conv_pd_->clone()) {} |
48 | |
49 | ~pd_t() = default; |
50 | |
51 | DECLARE_COMMON_PD_T(conv_pd_->name(), |
52 | jit_avx512_core_x8s8s32x_1x1_deconvolution_fwd_t); |
53 | |
54 | status_t init_convolution(engine_t *engine) { |
55 | convolution_desc_t cd; |
56 | |
57 | auto dd = desc(); |
58 | CHECK(conv_desc_init(&cd, prop_kind::forward_training, |
59 | alg_kind::convolution_direct, &(dd->src_desc), |
60 | &(dd->weights_desc), &(dd->bias_desc), &(dd->dst_desc), |
61 | dd->strides, dd->dilates, dd->padding[0], dd->padding[1])); |
62 | |
63 | primitive_attr_t conv_attr(*attr()); |
64 | if (!conv_attr.is_initialized()) return status::out_of_memory; |
65 | primitive_desc_iterator_t it( |
66 | engine, (op_desc_t *)&cd, &conv_attr, nullptr); |
67 | if (!it.is_initialized()) return status::out_of_memory; |
68 | |
69 | while (++it != it.end()) { |
70 | conv_pd_ = *it; |
71 | // XXX: find another way to create required implementation. |
72 | if (dynamic_cast<conv_pd_t *>(conv_pd_.get())) |
73 | return set_default_params(); |
74 | } |
75 | |
76 | return status::unimplemented; |
77 | }; |
78 | |
79 | status_t init(engine_t *engine) { |
80 | using namespace data_type; |
81 | using skip_mask_t = primitive_attr_t::skip_mask_t; |
82 | bool ok = is_fwd() |
83 | && desc()->alg_kind == alg_kind::deconvolution_direct |
84 | && !has_zero_dim_memory() |
85 | && utils::one_of(src_md(0)->data_type, s8, u8) |
86 | && weights_md(0)->data_type == s8 |
87 | && IMPLICATION(with_bias(), |
88 | utils::one_of( |
89 | weights_md(1)->data_type, f32, s32, s8, u8)) |
90 | && utils::one_of(dst_md(0)->data_type, f32, s32, s8, u8) |
91 | && desc()->accum_data_type == s32 |
92 | && attr()->has_default_values(skip_mask_t::scales_runtime |
93 | | skip_mask_t::post_ops |
94 | | skip_mask_t::zero_points_runtime) |
95 | && zero_points_valid( |
96 | attr(), true /*per_oc_bcast_accepted*/); |
97 | |
98 | if (!ok) return status::unimplemented; |
99 | |
100 | CHECK(init_convolution(engine)); |
101 | CHECK(attr_.set_default_formats(dst_md(0))); |
102 | init_scratchpad(); |
103 | |
104 | return status::success; |
105 | } |
106 | |
107 | protected: |
108 | status_t set_default_params() { |
109 | auto conv_1x1_pd_ = static_cast<conv_pd_t *>(conv_pd_.get()); |
110 | src_md_ = *conv_1x1_pd_->src_md(); |
111 | dst_md_ = *conv_1x1_pd_->dst_md(); |
112 | weights_md_ = *conv_1x1_pd_->weights_md(); |
113 | if (with_bias()) bias_md_ = *conv_1x1_pd_->weights_md(1); |
114 | return status::success; |
115 | } |
116 | |
117 | using conv_pd_t = |
118 | typename jit_avx512_core_x8s8s32x_1x1_convolution_fwd_t::pd_t; |
119 | friend jit_avx512_core_x8s8s32x_1x1_deconvolution_fwd_t; |
120 | |
121 | std::shared_ptr<primitive_desc_t> conv_pd_; |
122 | |
123 | private: |
124 | void init_scratchpad() { |
125 | auto scratchpad = scratchpad_registry().registrar(); |
126 | scratchpad.book(memory_tracking::names::key_nested, |
127 | conv_pd_->scratchpad_registry()); |
128 | } |
129 | }; |
130 | |
131 | jit_avx512_core_x8s8s32x_1x1_deconvolution_fwd_t(const pd_t *apd) |
132 | : primitive_t(apd) {} |
133 | |
134 | status_t init(engine_t *engine) override { |
135 | pd()->conv_pd_->create_primitive(conv_p_, engine); |
136 | return status::success; |
137 | } |
138 | |
139 | status_t execute(const exec_ctx_t &ctx) const override { |
140 | nested_scratchpad_t ns( |
141 | ctx, memory_tracking::names::key_nested, conv_p_); |
142 | // XXX: create a new ctx for convolution? |
143 | auto &tmp_ctx = const_cast<exec_ctx_t &>(ctx); |
144 | tmp_ctx.set_scratchpad_grantor(ns.grantor()); |
145 | return conv_p_->execute(tmp_ctx); |
146 | } |
147 | |
148 | private: |
149 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
150 | std::shared_ptr<primitive_t> conv_p_; |
151 | }; |
152 | |
153 | } // namespace x64 |
154 | } // namespace cpu |
155 | } // namespace impl |
156 | } // namespace dnnl |
157 | |
158 | #endif /* CPU_X64_JIT_AVX512_CORE_X8S8S32X_1X1_DECONVOLUTION_HPP */ |
159 | |