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

   1 /*******************************************************************************
   2 * Copyright 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_NCSP_BATCH_NORMALIZATION_HPP
  18 #define CPU_NCSP_BATCH_NORMALIZATION_HPP
  19
  20 #include <assert.h>
  21
  22 #include "c_types_map.hpp"
  23 #include "memory_tracking.hpp"
  24 #include "type_helpers.hpp"
  25 #include "utils.hpp"
  26
  27 #include "cpu_batch_normalization_pd.hpp"
  28
  29 namespace mkldnn {
  30 namespace impl {
  31 namespace cpu {
  32
  33 struct ncsp_batch_normalization_fwd_t : public cpu_primitive_t {
  34     struct pd_t : public cpu_batch_normalization_fwd_pd_t {
  35         pd_t(engine_t *engine, const batch_normalization_desc_t *adesc,
  36                 const primitive_attr_t *attr,
  37                 const batch_normalization_fwd_pd_t *hint_fwd_pd)
  38             : cpu_batch_normalization_fwd_pd_t(
  39                       engine, adesc, attr, hint_fwd_pd) {}
  40
  41         DECLARE_COMMON_PD_T("ncsp_bnorm:any", ncsp_batch_normalization_fwd_t);
  42
  43         virtual status_t init() override {
  44             using namespace data_type;
  45             using namespace prop_kind;
  46
  47             assert(engine()->kind() == engine_kind::cpu);
  48
  49             bool ok = true
  50                 && is_fwd()
  51                 && !has_zero_dim_memory()
  52                 && desc()->data_desc.data_type == f32
  53                 && IMPLICATION(use_scaleshift(),
  54                         desc()->data_scaleshift_desc.data_type == f32)
  55                 && utils::one_of(data_pd_.desc()->format, memory_format::nchw,
  56                         memory_format::ncdhw, memory_format::nc)
  57                 && (attr()->has_default_values() || this->with_relu_post_op());
  58             if (!ok) return status::unimplemented;
  59
  60             if (is_training() && fuse_bn_relu())
  61                 bn_init_default_ws(this, this->workspace_pd_, 8);
  62
  63             if (stats_is_src() || is_training()) {
  64                 memory_desc_t stats_d;
  65                 dims_t stats_dims = { C() };
  66                 mkldnn_memory_desc_init(&stats_d, 1, stats_dims,
  67                         data_type::f32, memory_format::x);
  68                 mean_pd_ = cpu_memory_t::pd_t(engine_, &stats_d);
  69                 variance_pd_ = cpu_memory_t::pd_t(engine_, &stats_d);
  70             }
  71
  72             init_scratchpad();
  73
  74             return success;
  75         }
  76
  77     private:
  78         void init_scratchpad() {
  79             using namespace memory_tracking::names;
  80             auto scratchpad = scratchpad_registry().registrar();
  81             if (!stats_is_src()) {
  82                 scratchpad.book(key_bnorm_reduction,
  83                         sizeof(data_t) * C() * mkldnn_get_max_threads());
  84
  85                 if (!is_training()) {
  86                     scratchpad.book(key_bnorm_tmp_mean, sizeof(data_t) * C());
  87                     scratchpad.book(key_bnorm_tmp_var, sizeof(data_t) * C());
  88                 }
  89             }
  90         }
  91     };
  92
  93     typedef typename prec_traits<data_type::f32>::type data_t;
  94
  95     ncsp_batch_normalization_fwd_t(const pd_t *apd, const input_vector &inputs,
  96             const output_vector &outputs)
  97         : cpu_primitive_t(apd, inputs, outputs) {}
  98     ~ncsp_batch_normalization_fwd_t() {}
  99
 100     virtual void execute(event_t *e) const {
 101         execute_forward();
 102         e->set_state(event_t::ready);
 103     }
 104
 105 private:
 106     void execute_forward() const;
 107     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 108 };
 109
 110 struct ncsp_batch_normalization_bwd_t : public cpu_primitive_t {
 111     struct pd_t : public cpu_batch_normalization_bwd_pd_t {
 112         pd_t(engine_t *engine, const batch_normalization_desc_t *adesc,
 113                 const primitive_attr_t *attr,
 114                 const batch_normalization_fwd_pd_t *hint_fwd_pd)
 115             : cpu_batch_normalization_bwd_pd_t(
 116                     engine, adesc, attr, hint_fwd_pd) {}
 117
 118         DECLARE_COMMON_PD_T("ncsp_bnorm:any", ncsp_batch_normalization_bwd_t);
 119
 120         virtual status_t init() override {
 121             using namespace data_type;
 122             assert(engine()->kind() == engine_kind::cpu);
 123
 124             bool ok = true
 125                 && is_bwd()
 126                 && !has_zero_dim_memory()
 127                 && desc()->data_desc.data_type == f32
 128                 && IMPLICATION(use_scaleshift(),
 129                         desc()->data_scaleshift_desc.data_type == f32)
 130                 && utils::one_of(data_pd_.desc()->format, memory_format::nchw,
 131                         memory_format::ncdhw, memory_format::nc)
 132                 && attr()->has_default_values();
 133             if (!ok) return status::unimplemented;
 134
 135             if (fuse_bn_relu()) {
 136                 bn_init_default_ws(this, this->workspace_pd_, 8);
 137                 const size_t this_ws_sz
 138                     = memory_desc_wrapper(this->workspace_pd()).size();
 139
 140                 bool ws_ok = true
 141                     && hint_fwd_pd_->workspace_pd()
 142                     && memory_desc_wrapper(hint_fwd_pd_->workspace_pd()).size()
 143                     == this_ws_sz;
 144                 if (!ws_ok) return status::unimplemented;
 145             }
 146
 147             init_scratchpad();
 148
 149             return success;
 150         }
 151
 152     private:
 153         void init_scratchpad() {
 154             using namespace memory_tracking::names;
 155             auto scratchpad = scratchpad_registry().registrar();
 156             scratchpad.book(key_bnorm_reduction,
 157                     sizeof(data_t) * 2 * C() * mkldnn_get_max_threads());
 158             if (!(use_scaleshift() && desc()->prop_kind == prop_kind::backward))
 159                 scratchpad.book(key_bnorm_tmp_diff_ss,
 160                         sizeof(data_t) * 2 * C());
 161         }
 162     };
 163
 164     typedef typename prec_traits<data_type::f32>::type data_t;
 165
 166     ncsp_batch_normalization_bwd_t(const pd_t *apd, const input_vector &inputs,
 167             const output_vector &outputs)
 168         : cpu_primitive_t(apd, inputs, outputs) {}
 169     ~ncsp_batch_normalization_bwd_t() {}
 170
 171     virtual void execute(event_t *e) const {
 172         execute_backward();
 173         e->set_state(event_t::ready);
 174     }
 175
 176 private:
 177     void execute_backward() const;
 178     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 179 };
 180
 181 }
 182 }
 183 }
 184
 185 #endif
 186
 187 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s