inference-engine/thirdparty/mkl-dnn/src/cpu/nspc_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_NSPC_BATCH_NORMALIZATION_HPP
  18 #define CPU_NSPC_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 template <data_type_t data_type>
  34 struct nspc_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(
  40                       engine, adesc, attr, hint_fwd_pd) {}
  41
  42         DECLARE_COMMON_PD_T("nspc_bnorm:any", nspc_batch_normalization_fwd_t);
  43
  44         virtual status_t init() override {
  45             using namespace data_type;
  46             using namespace prop_kind;
  47
  48             assert(engine()->kind() == engine_kind::cpu);
  49
  50             bool ok = true
  51                     /* the algorithm requires barriers while switching
  52                      * between parallelization over N and C dimensions */
  53                     && mkldnn_thr_syncable() && is_fwd()
  54                     && !has_zero_dim_memory()
  55                     && utils::one_of(desc()->prop_kind, forward_training,
  56                                forward_inference)
  57                     && desc()->data_desc.data_type == data_type
  58                     && IMPLICATION(use_scaleshift(),
  59                                desc()->data_scaleshift_desc.data_type
  60                                        == f32)
  61                     && utils::everyone_is(f32, desc()->mean_desc.data_type,
  62                                desc()->variance_desc.data_type)
  63                     // TODO: add ndhwc support?
  64                     && utils::one_of(
  65                                data_pd_.desc()->format, memory_format::nhwc)
  66                     && IMPLICATION(data_type == bf16, mayiuse(avx512_core))
  67                     && (attr()->has_default_values()
  68                                || this->with_relu_post_op());
  69             if (!ok)
  70                 return status::unimplemented;
  71
  72             if (is_training() && fuse_bn_relu())
  73                 bn_init_default_ws(this, this->workspace_pd_, 8);
  74
  75             if (stats_is_src() || is_training()) {
  76                 memory_desc_t stats_d;
  77                 dims_t stats_dims = { C() };
  78                 mkldnn_memory_desc_init(
  79                         &stats_d, 1, stats_dims, f32, memory_format::x);
  80                 mean_pd_ = cpu_memory_t::pd_t(engine_, &stats_d);
  81                 variance_pd_ = cpu_memory_t::pd_t(engine_, &stats_d);
  82             }
  83
  84             init_scratchpad();
  85
  86             return status::success;
  87         }
  88
  89     private:
  90         void init_scratchpad() {
  91             using namespace memory_tracking::names;
  92             auto scratchpad = scratchpad_registry().registrar();
  93             if (!stats_is_src()) {
  94                 const size_t stats_buf_sz = sizeof(acc_data_t)
  95                     * nstl::max(C(), 16) * mkldnn_get_max_threads();
  96                 scratchpad.book(key_bnorm_reduction, stats_buf_sz);
  97                 scratchpad.book(key_bnorm_tmp_mean, stats_buf_sz);
  98                 scratchpad.book(key_bnorm_tmp_var, stats_buf_sz);
  99             }
 100             if (data_type == data_type::bf16) {
 101                 const int simd_w = 16;
 102                 const int nbufs = 2;
 103                 const size_t bf16cvt_buf_sz = sizeof(acc_data_t) * nbufs
 104                     * mkldnn_get_max_threads() * utils::rnd_up(C(), simd_w);
 105                 scratchpad.book(key_bnorm_bf16cvt, bf16cvt_buf_sz);
 106             }
 107         }
 108     };
 109
 110     typedef typename prec_traits<data_type>::type data_t;
 111     typedef float acc_data_t;
 112
 113     nspc_batch_normalization_fwd_t(const pd_t *apd, const input_vector &inputs,
 114             const output_vector &outputs)
 115         : cpu_primitive_t(apd, inputs, outputs) {}
 116
 117     ~nspc_batch_normalization_fwd_t() {}
 118
 119     virtual void execute(event_t *e) const {
 120         execute_forward();
 121         e->set_state(event_t::ready);
 122     }
 123
 124 private:
 125     void execute_forward() const;
 126     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 127 };
 128
 129 template <data_type_t data_type>
 130 struct nspc_batch_normalization_bwd_t : public cpu_primitive_t {
 131     struct pd_t : public cpu_batch_normalization_bwd_pd_t {
 132         pd_t(engine_t *engine, const batch_normalization_desc_t *adesc,
 133                 const primitive_attr_t *attr,
 134                 const batch_normalization_fwd_pd_t *hint_fwd_pd)
 135             : cpu_batch_normalization_bwd_pd_t(
 136                       engine, adesc, attr, hint_fwd_pd) {}
 137
 138         DECLARE_COMMON_PD_T("nspc_bnorm:any", nspc_batch_normalization_bwd_t);
 139
 140         virtual status_t init() override {
 141             using namespace data_type;
 142             using namespace prop_kind;
 143
 144             assert(engine()->kind() == engine_kind::cpu);
 145
 146             bool ok = true
 147                     /* the algorithm requires barriers while switching
 148                      * between parallelization over N and C dimensions */
 149                     && mkldnn_thr_syncable() && is_bwd()
 150                     && !has_zero_dim_memory()
 151                     && utils::one_of(desc()->prop_kind, backward, backward_data)
 152                     && utils::everyone_is(data_type,
 153                                desc()->data_desc.data_type,
 154                                desc()->diff_data_desc.data_type)
 155                     && utils::everyone_is(f32, desc()->mean_desc.data_type,
 156                                desc()->variance_desc.data_type)
 157                     && IMPLICATION(use_scaleshift(),
 158                                desc()->diff_data_scaleshift_desc.data_type
 159                                                == f32
 160                                        && desc()->data_scaleshift_desc.data_type
 161                                                == f32)
 162                     && IMPLICATION(data_type == bf16, mayiuse(avx512_core))
 163                     // TODO: add ndhwc support?
 164                     && utils::one_of(
 165                                data_pd_.desc()->format, memory_format::nhwc)
 166                     && (attr()->has_default_values()
 167                                || this->with_relu_post_op())
 168                     && hint_fwd_pd_ != nullptr;
 169             if (!ok)
 170                 return status::unimplemented;
 171
 172             if (fuse_bn_relu()) {
 173                 bn_init_default_ws(this, this->workspace_pd_, 8);
 174                 const size_t this_ws_sz
 175                     = memory_desc_wrapper(this->workspace_pd()).size();
 176
 177                 bool ws_ok = true
 178                     && hint_fwd_pd_->workspace_pd()
 179                     && memory_desc_wrapper(hint_fwd_pd_->workspace_pd()).size()
 180                     == this_ws_sz;
 181                 if (!ws_ok) return status::unimplemented;
 182             }
 183
 184             init_scratchpad();
 185
 186             return status::success;
 187         }
 188
 189     private:
 190         void init_scratchpad() {
 191             using namespace memory_tracking::names;
 192             auto scratchpad = scratchpad_registry().registrar();
 193             scratchpad.book(key_bnorm_reduction,
 194                     sizeof(acc_data_t) * 2 * C() * mkldnn_get_max_threads());
 195             scratchpad.book(key_bnorm_tmp_diff_ss, sizeof(acc_data_t) * 2 * C()
 196                     * (mkldnn_get_max_threads() + 1));
 197             if (data_type == data_type::bf16) {
 198                 const int simd_w = 16;
 199                 const int nbufs = 2 + !use_global_stats();
 200                 const size_t bf16cvt_buf_sz = sizeof(acc_data_t) * nbufs
 201                     * mkldnn_get_max_threads() * utils::rnd_up(C(), simd_w);
 202                 scratchpad.book(key_bnorm_bf16cvt, bf16cvt_buf_sz);
 203             }
 204         }
 205     };
 206
 207     typedef typename prec_traits<data_type>::type data_t;
 208     typedef float acc_data_t;
 209
 210     nspc_batch_normalization_bwd_t(const pd_t *apd, const input_vector &inputs,
 211             const output_vector &outputs)
 212         : cpu_primitive_t(apd, inputs, outputs) {}
 213
 214     ~nspc_batch_normalization_bwd_t() {}
 215
 216     virtual void execute(event_t *e) const {
 217         execute_backward();
 218         e->set_state(event_t::ready);
 219     }
 220
 221 private:
 222     void execute_backward() const;
 223     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 224 };
 225
 226 }
 227 }
 228 }
 229
 230 #endif
 231
 232 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s