1 | /******************************************************************************* |
2 | * Copyright 2019-2022 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_X8S8S32X_CONVOLUTION_HPP |
18 | #define CPU_X64_JIT_UNI_X8S8S32X_CONVOLUTION_HPP |
19 | |
20 | #include "common/c_types_map.hpp" |
21 | #include "common/dnnl_thread.hpp" |
22 | #include "common/memory_tracking.hpp" |
23 | #include "common/primitive.hpp" |
24 | #include "common/utils.hpp" |
25 | |
26 | #include "cpu/cpu_convolution_pd.hpp" |
27 | |
28 | #include "cpu/x64/jit_uni_x8s8s32x_conv_kernel.hpp" |
29 | |
30 | namespace dnnl { |
31 | namespace impl { |
32 | namespace cpu { |
33 | namespace x64 { |
34 | |
35 | template <cpu_isa_t isa> |
36 | struct jit_uni_x8s8s32x_convolution_fwd_t : public primitive_t { |
37 | struct pd_t : public cpu_convolution_fwd_pd_t { |
38 | pd_t(const convolution_desc_t *adesc, const primitive_attr_t *attr, |
39 | const typename pd_t::base_class *hint_fwd_pd) |
40 | : cpu_convolution_fwd_pd_t(adesc, attr, hint_fwd_pd), jcp_() {} |
41 | |
42 | DECLARE_COMMON_PD_T( |
43 | JIT_IMPL_NAME_HELPER("jit_uni_int8:" , |
44 | isa == avx2 && jcp_.has_vnni ? avx2_vnni : isa, "" ), |
45 | jit_uni_x8s8s32x_convolution_fwd_t); |
46 | |
47 | status_t init(engine_t *engine) { |
48 | using namespace data_type; |
49 | using smask_t = primitive_attr_t::skip_mask_t; |
50 | const bool args_ok = is_fwd() |
51 | && set_default_alg_kind(alg_kind::convolution_direct) |
52 | && utils::one_of(src_md(0)->data_type, s8, u8) |
53 | && weights_md(0)->data_type == s8 |
54 | && IMPLICATION(with_bias(), |
55 | utils::one_of( |
56 | weights_md(1)->data_type, f32, s32, s8, u8)) |
57 | && utils::one_of(dst_md(0)->data_type, f32, s32, s8, u8) |
58 | && desc()->accum_data_type == s32 |
59 | && attr()->has_default_values(smask_t::scales_runtime |
60 | | smask_t::zero_points_runtime |
61 | | smask_t::post_ops | smask_t::sum_dt, |
62 | dst_md(0)->data_type) |
63 | && attr()->post_ops_.check_sum_consistent_dt( |
64 | dst_md(0)->data_type) |
65 | && !has_zero_dim_memory() && zero_points_ok(); |
66 | if (!args_ok) return status::unimplemented; |
67 | |
68 | CHECK(jit_uni_x8s8s32x_fwd_kernel<isa>::init_conf(jcp_, *desc(), |
69 | src_md_, weights_md_, dst_md_, bias_md_, attr_, |
70 | dnnl_get_max_threads())); |
71 | |
72 | auto scratchpad = scratchpad_registry().registrar(); |
73 | jit_uni_x8s8s32x_fwd_kernel<isa>::init_scratchpad( |
74 | scratchpad, jcp_, *attr()); |
75 | |
76 | return attr_.set_default_formats(dst_md(0)); |
77 | } |
78 | |
79 | jit_conv_conf_t jcp_; |
80 | |
81 | protected: |
82 | bool zero_points_ok() const { |
83 | // Only common zero points are supported -> mask should only be 0 |
84 | int mask_src = 0, mask_dst = 0; |
85 | attr()->zero_points_.get(DNNL_ARG_SRC, &mask_src); |
86 | attr()->zero_points_.get(DNNL_ARG_DST, &mask_dst); |
87 | return attr()->zero_points_.has_default_values(DNNL_ARG_WEIGHTS) |
88 | && mask_src == 0 && mask_dst == 0; |
89 | } |
90 | }; |
91 | |
92 | jit_uni_x8s8s32x_convolution_fwd_t(const pd_t *apd) : primitive_t(apd) {} |
93 | |
94 | status_t init(engine_t *engine) override { |
95 | CHECK(safe_ptr_assign(kernel_, |
96 | new jit_uni_x8s8s32x_fwd_kernel<isa>( |
97 | pd()->jcp_, *pd()->attr(), *pd()->dst_md()))); |
98 | return kernel_->create_kernel(); |
99 | } |
100 | |
101 | status_t execute(const exec_ctx_t &ctx) const override { |
102 | const auto &_pd = pd(); |
103 | const int ndims = _pd->ndims(); |
104 | const bool is_dw = _pd->jcp_.is_depthwise; |
105 | |
106 | switch (ndims) { |
107 | case 3: return execute_forward_1d(ctx); |
108 | case 4: |
109 | if (is_dw) return execute_forward_2d_dw(ctx); |
110 | return execute_forward_2d(ctx); |
111 | case 5: return execute_forward_3d(ctx); |
112 | } |
113 | return status::unimplemented; |
114 | } |
115 | |
116 | private: |
117 | status_t execute_forward_1d(const exec_ctx_t &ctx) const; |
118 | status_t execute_forward_2d(const exec_ctx_t &ctx) const; |
119 | status_t execute_forward_3d(const exec_ctx_t &ctx) const; |
120 | status_t execute_forward_2d_dw(const exec_ctx_t &ctx) const; |
121 | const pd_t *pd() const { |
122 | return static_cast<const pd_t *>(primitive_t::pd().get()); |
123 | } |
124 | const float *adjust_oscales(const memory_tracking::grantor_t &scratchpad, |
125 | const float *src_scales, const float *wei_scales) const; |
126 | |
127 | std::unique_ptr<jit_uni_x8s8s32x_fwd_kernel<isa>> kernel_; |
128 | }; |
129 | |
130 | } // namespace x64 |
131 | } // namespace cpu |
132 | } // namespace impl |
133 | } // namespace dnnl |
134 | |
135 | #endif |
136 | |