1 /*******************************************************************************
2 * Copyright 2018 Intel Corporation
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
8 * http://www.apache.org/licenses/LICENSE-2.0
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 *******************************************************************************/
17 #ifndef CPU_JIT_AVX512_CORE_X8S8S32X_1X1_CONVOLUTION_HPP
18 #define CPU_JIT_AVX512_CORE_X8S8S32X_1X1_CONVOLUTION_HPP
20 #include "c_types_map.hpp"
21 #include "memory_tracking.hpp"
22 #include "mkldnn_thread.hpp"
25 #include "cpu_convolution_pd.hpp"
26 #include "cpu_engine.hpp"
28 #include "jit_avx512_core_x8s8s32x_1x1_conv_kernel.hpp"
29 #include "jit_uni_1x1_conv_utils.hpp"
35 template<impl::data_type_t src_type, impl::data_type_t dst_type>
36 struct jit_avx512_core_x8s8s32x_1x1_convolution_fwd_t : public cpu_primitive_t {
37 struct pd_t: public cpu_convolution_fwd_pd_t {
38 pd_t(engine_t *engine, const convolution_desc_t *adesc,
39 const primitive_attr_t *attr,
40 const typename pd_t::base_class *hint_fwd_pd)
41 : cpu_convolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd)
45 JIT_IMPL_NAME_HELPER("jit_int8_1x1:", avx512_core, ""),
46 jit_avx512_core_x8s8s32x_1x1_convolution_fwd_t<
49 virtual status_t init() override {
50 using namespace prop_kind;
51 using namespace utils;
52 assert(this->engine()->kind() == engine_kind::cpu);
54 && this->set_default_params() == status::success
55 && utils::one_of(this->desc()->prop_kind, forward_training,
57 && utils::one_of(this->desc()->alg_kind,
58 alg_kind::convolution_auto,
59 alg_kind::convolution_direct)
60 && !this->has_zero_dim_memory()
61 && this->desc()->src_desc.data_type == src_type
62 && this->desc()->dst_desc.data_type == dst_type
63 && this->desc()->weights_desc.data_type == data_type::s8
64 && IMPLICATION(this->with_bias(), utils::one_of(
65 this->desc()->bias_desc.data_type, data_type::f32,
66 data_type::s32, data_type::s8, data_type::u8))
67 && this->desc()->accum_data_type == data_type::s32;
68 if (!ok) return status::unimplemented;
70 const convolution_desc_t *conv_d = this->desc();
71 const memory_desc_t *src_d = this->src_pd_.desc();
72 rtus_prepare(this, conv_d, src_d, this->dst_pd_.desc());
75 jit_avx512_core_x8s8s32x_1x1_conv_kernel::init_conf(jcp_,
76 *conv_d, *src_d, *this->weights_pd_.desc(),
77 *this->dst_pd_.desc(), *this->bias_pd_.desc(),
78 *this->attr(), mkldnn_get_max_threads(),
80 if (status != status::success) return status;
82 auto scratchpad = scratchpad_registry().registrar();
83 jit_avx512_core_x8s8s32x_1x1_conv_kernel::init_scratchpad(
84 scratchpad, jcp_, *this->attr());
86 rtus_prepare_space_info(this, scratchpad);
88 return status::success;
91 jit_1x1_conv_conf_t jcp_;
92 reduce_to_unit_stride_t rtus_;
95 virtual status_t set_default_params() override {
96 using namespace memory_format;
98 this->desc()->src_desc.data_type == data_type::s8;
100 if (this->src_pd_.desc()->format == any)
101 CHECK(this->src_pd_.set_format(nhwc));
102 if (this->dst_pd_.desc()->format == any)
103 CHECK(this->dst_pd_.set_format(nhwc));
104 if (this->weights_pd_.desc()->format == any)
105 CHECK(this->weights_pd_.set_format(this->with_groups()
106 ? (is_sign_input ? gOIhw4i16o4i_s8s8 : gOIhw4i16o4i)
107 : (is_sign_input ? OIhw4i16o4i_s8s8 : OIhw4i16o4i)));
108 if (this->bias_pd_.desc()->format == any)
109 CHECK(this->bias_pd_.set_format(x));
110 if (this->desc()->alg_kind == alg_kind::convolution_auto)
111 CHECK(this->set_alg_kind(alg_kind::convolution_direct));
113 return status::success;
117 template <cpu_isa_t isa, typename conv_t>
118 friend void init_rtus_driver(conv_t *self);
120 jit_avx512_core_x8s8s32x_1x1_convolution_fwd_t(const pd_t *apd,
121 const input_vector &inputs, const output_vector &outputs)
122 : cpu_primitive_t(apd, inputs, outputs)
123 , kernel_(nullptr), rtus_driver_(nullptr)
125 kernel_ = new jit_avx512_core_x8s8s32x_1x1_conv_kernel(pd()->jcp_,
127 init_rtus_driver<avx512_common>(this);
130 ~jit_avx512_core_x8s8s32x_1x1_convolution_fwd_t() {
135 typedef typename prec_traits<src_type>::type src_data_t;
136 typedef typename prec_traits<data_type::s8>::type wei_data_t;
137 typedef typename prec_traits<dst_type>::type dst_data_t;
138 typedef typename prec_traits<data_type::s32>::type acc_data_t;
140 virtual void execute(event_t *e) const {
142 e->set_state(event_t::ready);
146 void execute_forward() const;
147 void execute_forward_thr(const int ithr, const int nthr,
148 const src_data_t *src, const wei_data_t *weights,
149 const char *bias, dst_data_t *dst,
150 const memory_tracking::grantor_t &scratchpad) const;
151 const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
153 jit_avx512_core_x8s8s32x_1x1_conv_kernel *kernel_;
154 rtus_driver_t<avx512_common> *rtus_driver_;