inference-engine/thirdparty/mkl-dnn/src/cpu/ref_batch_normalization.hpp

   1 /*******************************************************************************
   2 * Copyright 2016-2018 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_REF_BATCH_NORMALIZATION_FWD_HPP
  18 #define CPU_REF_BATCH_NORMALIZATION_FWD_HPP
  19
  20 #include <assert.h>
  21
  22 #include "c_types_map.hpp"
  23 #include "cpu_batch_normalization_pd.hpp"
  24 #include "cpu_engine.hpp"
  25 #include "cpu_isa_traits.hpp"
  26 #include "type_helpers.hpp"
  27 #include "utils.hpp"
  28
  29 namespace mkldnn {
  30 namespace impl {
  31 namespace cpu {
  32
  33 template <data_type_t data_type>
  34 struct ref_batch_normalization_fwd_t : public cpu_primitive_t {
  35     struct pd_t: public cpu_batch_normalization_fwd_pd_t {
  36         pd_t(engine_t *engine, const batch_normalization_desc_t *adesc,
  37                 const primitive_attr_t *attr,
  38                 const batch_normalization_fwd_pd_t *hint_fwd_pd)
  39             : cpu_batch_normalization_fwd_pd_t(engine, adesc, attr,
  40                     hint_fwd_pd) {}
  41
  42         DECLARE_COMMON_PD_T("ref:any", ref_batch_normalization_fwd_t);
  43
  44         virtual status_t init() override {
  45             using namespace data_type;
  46             using namespace prop_kind;
  47             assert(engine()->kind() == engine_kind::cpu);
  48             bool ok = true
  49                 && is_fwd()
  50                 && !has_zero_dim_memory()
  51                 && utils::one_of(desc()->prop_kind, forward_training,
  52                         forward_inference)
  53                 && desc()->data_desc.data_type == data_type
  54                 && IMPLICATION(use_scaleshift(),
  55                         desc()->data_scaleshift_desc.data_type == f32)
  56                 && utils::everyone_is(f32,
  57                         desc()->mean_desc.data_type,
  58                         desc()->variance_desc.data_type)
  59                 && IMPLICATION(data_type == bf16, mayiuse(avx512_core))
  60                 && (attr()->has_default_values() || this->with_relu_post_op());
  61             if (!ok) return status::unimplemented;
  62
  63             if (desc()->data_desc.data_type == data_type::s8 && !stats_is_src())
  64                 return status::unimplemented;
  65
  66             if (stats_is_src() || is_training()) {
  67                 memory_desc_t stats_d;
  68                 dims_t stats_dims = { C() };
  69                 mkldnn_memory_desc_init(
  70                         &stats_d, 1, stats_dims, f32, memory_format::x);
  71                 mean_pd_ = cpu_memory_t::pd_t(engine_, &stats_d);
  72                 variance_pd_ = cpu_memory_t::pd_t(engine_, &stats_d);
  73             }
  74
  75             if (is_training() && fuse_bn_relu())
  76                 bn_init_default_ws(this, this->workspace_pd_, 8);
  77
  78             return status::success;
  79         }
  80     };
  81
  82     ref_batch_normalization_fwd_t(const pd_t *apd, const input_vector &inputs,
  83             const output_vector &outputs)
  84         : cpu_primitive_t(apd, inputs, outputs) {}
  85
  86     typedef typename prec_traits<data_type>::type data_t;
  87
  88     virtual void execute(event_t *e) const {
  89         execute_forward();
  90         e->set_state(event_t::ready);
  91     }
  92
  93 private:
  94     void execute_forward() const;
  95     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
  96 };
  97
  98 template <data_type_t data_type>
  99 struct ref_batch_normalization_bwd_t : public cpu_primitive_t {
 100     struct pd_t: public cpu_batch_normalization_bwd_pd_t {
 101         pd_t(engine_t *engine, const batch_normalization_desc_t *adesc,
 102                 const primitive_attr_t *attr,
 103                 const batch_normalization_fwd_pd_t *hint_fwd_pd)
 104             : cpu_batch_normalization_bwd_pd_t(engine, adesc, attr,
 105                     hint_fwd_pd) {}
 106
 107         DECLARE_COMMON_PD_T("ref:any", ref_batch_normalization_bwd_t);
 108
 109         virtual status_t init() override {
 110             using namespace data_type;
 111             using namespace prop_kind;
 112             assert(engine()->kind() == engine_kind::cpu);
 113             bool ok = true
 114                 && is_bwd()
 115                 && !has_zero_dim_memory()
 116                 && utils::one_of(desc()->prop_kind, backward, backward_data)
 117                 && utils::everyone_is(data_type, desc()->data_desc.data_type,
 118                         desc()->diff_data_desc.data_type)
 119                 && utils::everyone_is(f32,
 120                         desc()->mean_desc.data_type,
 121                         desc()->variance_desc.data_type)
 122                 && IMPLICATION(use_scaleshift(),
 123                         desc()->diff_data_scaleshift_desc.data_type == f32
 124                         && desc()->data_scaleshift_desc.data_type == f32)
 125                 && IMPLICATION(data_type == bf16, mayiuse(avx512_core))
 126                 && attr()->has_default_values()
 127                 && hint_fwd_pd_ != nullptr;
 128             if (!ok) return status::unimplemented;
 129
 130             if (fuse_bn_relu()) {
 131                 bn_init_default_ws(this, this->workspace_pd_, 8);
 132                 const size_t this_ws_sz
 133                     = memory_desc_wrapper(this->workspace_pd()).size();
 134
 135                 bool ws_ok = true
 136                     && hint_fwd_pd_->workspace_pd()
 137                     && memory_desc_wrapper(hint_fwd_pd_->workspace_pd()).size()
 138                             == this_ws_sz;
 139                 if (!ws_ok)
 140                     return status::unimplemented;
 141             }
 142
 143             bool stats_ok = true
 144                 && hint_fwd_pd_->mean_pd()->desc()->ndims == 1
 145                 && hint_fwd_pd_->mean_pd()->desc()->format == memory_format::x
 146                 && hint_fwd_pd_->mean_pd()->desc()->data_type == f32
 147                 && hint_fwd_pd_->variance_pd()->desc()->ndims == 1
 148                 && hint_fwd_pd_->variance_pd()->desc()->format == memory_format::x
 149                 && hint_fwd_pd_->variance_pd()->desc()->data_type == f32;
 150             if (!stats_ok) return status::unimplemented;
 151
 152             return status::success;
 153         }
 154     };
 155
 156     ref_batch_normalization_bwd_t(const pd_t *apd, const input_vector &inputs,
 157             const output_vector &outputs)
 158         : cpu_primitive_t(apd, inputs, outputs) {}
 159     typedef typename prec_traits<data_type>::type data_t;
 160
 161     virtual void execute(event_t *e) const {
 162         execute_backward();
 163         e->set_state(event_t::ready);
 164     }
 165
 166 private:
 167     void execute_backward() const;
 168     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 169 };
 170
 171 }
 172 }
 173 }
 174
 175 #endif
 176
 177 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s