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 template <data_type_t data_type>
  34 struct ncsp_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("ncsp_bnorm:any", ncsp_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
  49             bool ok = true
  50                 && is_fwd()
  51                 && !has_zero_dim_memory()
  52                 && utils::one_of(desc()->prop_kind, forward_training,
  53                         forward_inference)
  54                 && desc()->data_desc.data_type == data_type
  55                 && IMPLICATION(use_scaleshift(),
  56                         desc()->data_scaleshift_desc.data_type == f32)
  57                 && utils::everyone_is(f32,
  58                         desc()->mean_desc.data_type,
  59                         desc()->variance_desc.data_type)
  60                 && utils::one_of(data_pd_.desc()->format, memory_format::nchw,
  61                         memory_format::ncdhw, memory_format::nc)
  62                 && IMPLICATION(data_type == bf16, mayiuse(avx512_core))
  63                 && (attr()->has_default_values() || this->with_relu_post_op());
  64             if (!ok) return status::unimplemented;
  65
  66             if (is_training() && fuse_bn_relu())
  67                 bn_init_default_ws(this, this->workspace_pd_, 8);
  68
  69             if (stats_is_src() || is_training()) {
  70                 memory_desc_t stats_d;
  71                 dims_t stats_dims = { C() };
  72                 mkldnn_memory_desc_init(
  73                         &stats_d, 1, stats_dims, f32, memory_format::x);
  74                 mean_pd_ = cpu_memory_t::pd_t(engine_, &stats_d);
  75                 variance_pd_ = cpu_memory_t::pd_t(engine_, &stats_d);
  76             }
  77
  78             init_scratchpad();
  79
  80             return success;
  81         }
  82
  83     private:
  84         void init_scratchpad() {
  85             using namespace memory_tracking::names;
  86             auto scratchpad = scratchpad_registry().registrar();
  87             if (!stats_is_src()) {
  88                 scratchpad.book(key_bnorm_reduction,
  89                         sizeof(acc_data_t) * C() * mkldnn_get_max_threads());
  90
  91                 if (!is_training()) {
  92                     scratchpad.book(key_bnorm_tmp_mean, sizeof(acc_data_t) * C());
  93                     scratchpad.book(key_bnorm_tmp_var, sizeof(acc_data_t) * C());
  94                 }
  95             }
  96
  97             if (data_type == data_type::bf16) {
  98                 const int simd_w = 16;
  99                 const bool has_spatial = utils::one_of(ndims(), 4, 5);
 100                 const int SP = has_spatial ? D() * H() * W() : 1;
 101                 const int nbufs = 2;
 102                 const size_t bf16cvt_buf_sz = sizeof(acc_data_t) * nbufs
 103                     * mkldnn_get_max_threads() * utils::rnd_up(SP, simd_w);
 104                 scratchpad.book(key_bnorm_bf16cvt, bf16cvt_buf_sz);
 105             }
 106         }
 107     };
 108
 109     typedef typename prec_traits<data_type>::type data_t;
 110     typedef float acc_data_t;
 111
 112     ncsp_batch_normalization_fwd_t(const pd_t *apd, const input_vector &inputs,
 113             const output_vector &outputs)
 114         : cpu_primitive_t(apd, inputs, outputs) {}
 115
 116     ~ncsp_batch_normalization_fwd_t() {}
 117
 118     virtual void execute(event_t *e) const {
 119         execute_forward();
 120         e->set_state(event_t::ready);
 121     }
 122
 123 private:
 124     void execute_forward() const;
 125     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 126 };
 127
 128 template <data_type_t data_type>
 129 struct ncsp_batch_normalization_bwd_t : public cpu_primitive_t {
 130     struct pd_t : public cpu_batch_normalization_bwd_pd_t {
 131         pd_t(engine_t *engine, const batch_normalization_desc_t *adesc,
 132                 const primitive_attr_t *attr,
 133                 const batch_normalization_fwd_pd_t *hint_fwd_pd)
 134             : cpu_batch_normalization_bwd_pd_t(
 135                     engine, adesc, attr, hint_fwd_pd) {}
 136
 137         DECLARE_COMMON_PD_T("ncsp_bnorm:any", ncsp_batch_normalization_bwd_t);
 138
 139         virtual status_t init() override {
 140             using namespace data_type;
 141             using namespace prop_kind;
 142             assert(engine()->kind() == engine_kind::cpu);
 143
 144             bool ok = true
 145                 && is_bwd()
 146                 && !has_zero_dim_memory()
 147                 && utils::one_of(desc()->prop_kind, backward, backward_data)
 148                 && utils::everyone_is(data_type, desc()->data_desc.data_type,
 149                         desc()->diff_data_desc.data_type)
 150                 && utils::everyone_is(f32, desc()->mean_desc.data_type,
 151                         desc()->variance_desc.data_type)
 152                 && IMPLICATION(use_scaleshift(),
 153                         desc()->diff_data_scaleshift_desc.data_type == f32
 154                         && desc()->data_scaleshift_desc.data_type == f32)
 155                 && IMPLICATION(data_type == bf16, mayiuse(avx512_core))
 156                 && utils::one_of(data_pd_.desc()->format, memory_format::nchw,
 157                         memory_format::ncdhw, memory_format::nc)
 158                 && attr()->has_default_values()
 159                 && hint_fwd_pd_ != nullptr;
 160             if (!ok)
 161                 return status::unimplemented;
 162
 163             if (fuse_bn_relu()) {
 164                 bn_init_default_ws(this, this->workspace_pd_, 8);
 165                 const size_t this_ws_sz
 166                     = memory_desc_wrapper(this->workspace_pd()).size();
 167
 168                 bool ws_ok = true
 169                     && hint_fwd_pd_->workspace_pd()
 170                     && memory_desc_wrapper(hint_fwd_pd_->workspace_pd()).size()
 171                     == this_ws_sz;
 172                 if (!ws_ok) return status::unimplemented;
 173             }
 174
 175             init_scratchpad();
 176
 177             return status::success;
 178         }
 179
 180     private:
 181         void init_scratchpad() {
 182             using namespace memory_tracking::names;
 183             auto scratchpad = scratchpad_registry().registrar();
 184             scratchpad.book(key_bnorm_reduction,
 185                     sizeof(acc_data_t) * 2 * C() * mkldnn_get_max_threads());
 186             if (!(use_scaleshift() && desc()->prop_kind == prop_kind::backward))
 187                 scratchpad.book(key_bnorm_tmp_diff_ss,
 188                         sizeof(acc_data_t) * 2 * C());
 189
 190             if (data_type == data_type::bf16) {
 191                 const int simd_w = 16;
 192                 const bool has_spatial = utils::one_of(ndims(), 4, 5);
 193                 const int SP = has_spatial ? D() * H() * W() : 1;
 194                 const int nbufs = 2 + !use_global_stats();
 195                 const size_t bf16cvt_buf_sz = sizeof(acc_data_t) * nbufs
 196                     * mkldnn_get_max_threads() * utils::rnd_up(SP, simd_w);
 197                 scratchpad.book(key_bnorm_bf16cvt, bf16cvt_buf_sz);
 198             }
 199         }
 200     };
 201
 202     typedef typename prec_traits<data_type>::type data_t;
 203     typedef float acc_data_t;
 204
 205     ncsp_batch_normalization_bwd_t(const pd_t *apd, const input_vector &inputs,
 206             const output_vector &outputs)
 207         : cpu_primitive_t(apd, inputs, outputs) {}
 208
 209     ~ncsp_batch_normalization_bwd_t() {}
 210
 211     virtual void execute(event_t *e) const {
 212         execute_backward();
 213         e->set_state(event_t::ready);
 214     }
 215
 216 private:
 217     void execute_backward() const;
 218     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 219 };
 220
 221 }
 222 }
 223 }
 224
 225 #endif
 226
 227 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s