inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx512_common_convolution_winograd.hpp

   1 /*******************************************************************************
   2 * Copyright 2017-2018 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_JIT_AVX512_COMMON_CONVOLUTION_WINOGRAD_HPP
  18 #define CPU_JIT_AVX512_COMMON_CONVOLUTION_WINOGRAD_HPP
  19
  20 #include "c_types_map.hpp"
  21 #include "memory_tracking.hpp"
  22 #include "cpu_convolution_pd.hpp"
  23 #include "cpu_engine.hpp"
  24 #include "mkldnn_thread.hpp"
  25
  26 #include "jit_avx512_common_conv_winograd_kernel_f32.hpp"
  27
  28 namespace mkldnn {
  29 namespace impl {
  30 namespace cpu {
  31
  32 namespace winograd_avx512_common {
  33 inline void init_scratchpad(memory_tracking::registrar_t &scratchpad,
  34         const jit_conv_winograd_conf_t &jcp) {
  35     using namespace memory_tracking::names;
  36
  37     size_t U_sz = (size_t)alpha * alpha * jcp.ic * jcp.oc;
  38     size_t V_sz = (size_t)alpha * alpha * jcp.mb * jcp.ic
  39         * (jcp.itiles * jcp.jtiles + jcp.tile_4fma_padding);
  40     size_t M_sz = (size_t)alpha * alpha * jcp.mb * jcp.oc
  41         * (jcp.itiles * jcp.jtiles + jcp.tile_4fma_padding);
  42
  43     scratchpad.book(key_wino_U, sizeof(float) * U_sz, PAGE_2M);
  44     scratchpad.book(key_wino_V, sizeof(float) * V_sz, PAGE_2M);
  45     scratchpad.book(key_wino_M, sizeof(float) * M_sz, PAGE_2M);
  46
  47     if (jcp.sched_policy == WSCHED_WEI_S_D_G_W) {
  48         const int nthr = mkldnn_get_max_threads();
  49
  50         size_t tr_src_sz = jcp.ver != ver_4fma ? 0 : (size_t)nthr
  51             * alpha * alpha * jcp.tile_4fma * jcp.ic_simd_block;
  52         scratchpad.book(key_conv_tr_src, sizeof(float) * tr_src_sz, PAGE_2M);
  53
  54         size_t br_sz = jcp.with_bias ? nthr * jcp.oc : 0;
  55         scratchpad.book(key_conv_bia_reduction, sizeof(float) * br_sz, PAGE_2M);
  56
  57         size_t padded_bias_sz =
  58             jcp.with_bias && jcp.oc_without_padding != jcp.oc ? jcp.oc : 0;
  59         scratchpad.book(key_conv_padded_bias, sizeof(float) * padded_bias_sz);
  60     }
  61 }
  62 }
  63
  64 template <bool is_fwd>
  65 struct _jit_avx512_common_convolution_winograd_t {
  66
  67     _jit_avx512_common_convolution_winograd_t(
  68             const jit_conv_winograd_conf_t &jcp, const primitive_attr_t *attr)
  69         : kernel_(nullptr), attr_(attr) {
  70         kernel_ = new _jit_avx512_common_conv_winograd_data_kernel_f32(jcp);
  71         }
  72
  73     ~_jit_avx512_common_convolution_winograd_t() { delete kernel_; }
  74
  75     protected:
  76         void _execute_data_W_S_G_D(const int MB, float *inp_ptr, float *out_ptr,
  77                 float *wei_ptr, float *bias_ptr,
  78                 const memory_tracking::grantor_t &scratchpad) const;
  79         _jit_avx512_common_conv_winograd_data_kernel_f32 *kernel_;
  80         const primitive_attr_t *attr_;
  81 };
  82
  83 struct jit_avx512_common_convolution_winograd_fwd_t
  84      : _jit_avx512_common_convolution_winograd_t<true>
  85      , public cpu_primitive_t
  86     {
  87     struct pd_t : public cpu_convolution_fwd_pd_t {
  88         pd_t(engine_t *engine, const convolution_desc_t *adesc,
  89                 const primitive_attr_t *attr,
  90                 const typename pd_t::base_class *hint_fwd_pd)
  91             : cpu_convolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd)
  92             , jcp_() {}
  93
  94         DECLARE_COMMON_PD_T(
  95                 JIT_IMPL_NAME_HELPER("jit_wino:", avx512_common, ""),
  96                 jit_avx512_common_convolution_winograd_fwd_t);
  97
  98         virtual status_t init() override
  99         {
 100             using namespace prop_kind;
 101             assert(this->engine()->kind() == engine_kind::cpu);
 102             bool ok = true && this->set_default_params() == status::success
 103                     && utils::one_of(this->desc()->prop_kind, forward_training,
 104                                forward_inference)
 105                     && utils::one_of(this->desc()->alg_kind,
 106                                alg_kind::convolution_auto,
 107                                alg_kind::convolution_winograd)
 108                     && !this->has_zero_dim_memory()
 109                     && utils::everyone_is(data_type::f32,
 110                                this->desc()->src_desc.data_type,
 111                                this->desc()->weights_desc.data_type,
 112                                this->desc()->dst_desc.data_type)
 113                     && IMPLICATION(this->with_bias(), data_type::f32
 114                                        == this->desc()->bias_desc.data_type)
 115                     && mkldnn_thr_syncable();
 116
 117             if (!ok)
 118                 return status::unimplemented;
 119
 120             status_t status =
 121                 jit_avx512_common_conv_winograd_fwd_kernel_f32::init_conf(
 122                         jcp_, *this->desc(), *this->src_pd_.desc(),
 123                         *this->weights_pd_.desc(), *this->dst_pd_.desc(),
 124                         *this->attr());
 125             if (status != status::success) return status;
 126
 127             auto scratchpad = this->scratchpad_registry().registrar();
 128             winograd_avx512_common::init_scratchpad(scratchpad, jcp_);
 129
 130             if (status == status::success
 131                     && this->desc()->alg_kind == alg_kind::convolution_auto)
 132                 CHECK(this->set_alg_kind(alg_kind::convolution_winograd));
 133
 134             return status;
 135         }
 136
 137         jit_conv_winograd_conf_t jcp_;
 138
 139     protected:
 140         virtual status_t set_default_params() override
 141         {
 142             using namespace memory_format;
 143             if (this->src_pd_.desc()->format == any)
 144                 CHECK(this->src_pd_.set_format(nChw16c));
 145             if (this->dst_pd_.desc()->format == any)
 146                 CHECK(this->dst_pd_.set_format(nChw16c));
 147             if (this->weights_pd_.desc()->format == any)
 148                 CHECK(this->weights_pd_.set_format(
 149                         this->with_groups() ? gOIhw16i16o : OIhw16i16o));
 150             if (this->bias_pd_.desc()->format == any)
 151                 CHECK(this->bias_pd_.set_format(x));
 152             return status::success;
 153         }
 154     };
 155
 156     jit_avx512_common_convolution_winograd_fwd_t(const pd_t *apd,
 157             const input_vector &inputs, const output_vector &outputs)
 158         : _jit_avx512_common_convolution_winograd_t<true>(apd->jcp_, apd->attr())
 159         , cpu_primitive_t(apd, inputs, outputs, true) {}
 160
 161     ~jit_avx512_common_convolution_winograd_fwd_t(){};
 162
 163     typedef typename prec_traits<data_type::f32>::type data_t;
 164
 165     virtual void execute(event_t *e) const
 166     {
 167         float *src = (float *)this->input_memory(0);
 168         float *dst = (float *)this->memory();
 169         float *weights = (float *)this->input_memory(1);
 170         float *bias = (float *)this->input_memory(2);
 171
 172         this->_execute_data_W_S_G_D(pd()->MB(), src, dst, weights, bias,
 173                 this->scratchpad());
 174
 175         e->set_state(event_t::ready);
 176     }
 177
 178 private:
 179     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 180 };
 181
 182 struct jit_avx512_common_convolution_winograd_bwd_data_t
 183         : _jit_avx512_common_convolution_winograd_t<false>,
 184         public cpu_primitive_t {
 185     struct pd_t : public cpu_convolution_bwd_data_pd_t {
 186         pd_t(engine_t *engine, const convolution_desc_t *adesc,
 187                 const primitive_attr_t *attr,
 188                 const convolution_fwd_pd_t *hint_fwd_pd)
 189             : cpu_convolution_bwd_data_pd_t(engine, adesc, attr, hint_fwd_pd)
 190             , jcp_() {}
 191
 192         DECLARE_COMMON_PD_T(
 193                 JIT_IMPL_NAME_HELPER("jit_wino:", avx512_common, ""),
 194                 jit_avx512_common_convolution_winograd_bwd_data_t);
 195
 196         virtual status_t init() override
 197         {
 198             using namespace prop_kind;
 199             assert(this->engine()->kind() == engine_kind::cpu);
 200             bool ok = true && this->set_default_params() == status::success
 201                     && utils::one_of(this->desc()->prop_kind, backward_data)
 202                     && utils::one_of(this->desc()->alg_kind,
 203                                alg_kind::convolution_auto,
 204                                alg_kind::convolution_winograd)
 205                     && !this->has_zero_dim_memory()
 206                     && utils::everyone_is(data_type::f32,
 207                                this->desc()->diff_src_desc.data_type,
 208                                this->desc()->weights_desc.data_type,
 209                                this->desc()->diff_dst_desc.data_type)
 210                     && mkldnn_thr_syncable();
 211
 212             if (!ok)
 213                 return status::unimplemented;
 214
 215             status_t status =
 216                 jit_avx512_common_conv_winograd_bwd_data_kernel_f32::init_conf(
 217                         jcp_, *this->desc(), *this->diff_src_pd_.desc(),
 218                         *this->weights_pd_.desc(), *this->diff_dst_pd_.desc());
 219             if (status != status::success) return status;
 220
 221             auto scratchpad = this->scratchpad_registry().registrar();
 222             winograd_avx512_common::init_scratchpad(scratchpad, jcp_);
 223
 224             if (status == status::success
 225                     && this->desc()->alg_kind == alg_kind::convolution_auto)
 226                 CHECK(this->set_alg_kind(alg_kind::convolution_winograd));
 227
 228             return status;
 229         }
 230
 231         jit_conv_winograd_conf_t jcp_;
 232
 233     protected:
 234         virtual status_t set_default_params() override
 235         {
 236             using namespace memory_format;
 237
 238             if (this->diff_src_pd_.desc()->format == any)
 239                 CHECK(this->diff_src_pd_.set_format(nChw16c));
 240             if (this->diff_dst_pd_.desc()->format == any)
 241                 CHECK(this->diff_dst_pd_.set_format(nChw16c));
 242             if (this->weights_pd_.desc()->format == any)
 243                 CHECK(this->weights_pd_.set_format(
 244                         this->with_groups() ? gOIhw16i16o : OIhw16i16o));
 245             return status::success;
 246         }
 247     };
 248
 249     jit_avx512_common_convolution_winograd_bwd_data_t(const pd_t *apd,
 250             const input_vector &inputs, const output_vector &outputs)
 251         : _jit_avx512_common_convolution_winograd_t<false>(apd->jcp_, apd->attr())
 252         , cpu_primitive_t(apd, inputs, outputs, true) {}
 253
 254     ~jit_avx512_common_convolution_winograd_bwd_data_t(){};
 255
 256     typedef typename prec_traits<data_type::f32>::type data_t;
 257
 258     virtual void execute(event_t *e) const
 259     {
 260         assert(pd()->desc()->prop_kind == prop_kind::backward_data
 261                 && "invalid prop_kind");
 262
 263         float *diff_dst = (float *)this->input_memory(0);
 264         float *diff_src = (float *)this->memory();
 265         float *weights = (float *)this->input_memory(1);
 266
 267         this->_execute_data_W_S_G_D(pd()->MB(), diff_dst, diff_src, weights, nullptr,
 268                 this->scratchpad());
 269
 270         e->set_state(event_t::ready);
 271     }
 272
 273 private:
 274     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 275 };
 276
 277 struct jit_avx512_common_convolution_winograd_bwd_weights_t
 278         : public cpu_primitive_t {
 279     struct pd_t : public cpu_convolution_bwd_weights_pd_t {
 280         pd_t(engine_t *engine, const convolution_desc_t *adesc,
 281                 const primitive_attr_t *attr,
 282                 const convolution_fwd_pd_t *hint_fwd_pd)
 283             : cpu_convolution_bwd_weights_pd_t(engine, adesc, attr,
 284                     hint_fwd_pd)
 285             , jcp_() {}
 286
 287         DECLARE_COMMON_PD_T(
 288                 JIT_IMPL_NAME_HELPER("jit_wino:", avx512_common, ""),
 289                 jit_avx512_common_convolution_winograd_bwd_weights_t);
 290
 291         virtual status_t init() override
 292         {
 293             using namespace prop_kind;
 294             assert(this->engine()->kind() == engine_kind::cpu);
 295             bool ok = true && this->set_default_params() == status::success
 296                     && utils::one_of(this->desc()->prop_kind, backward_weights)
 297                     && utils::one_of(this->desc()->alg_kind,
 298                                alg_kind::convolution_auto,
 299                                alg_kind::convolution_winograd)
 300                     && !this->has_zero_dim_memory()
 301                     && utils::everyone_is(data_type::f32,
 302                                this->desc()->src_desc.data_type,
 303                                this->desc()->diff_dst_desc.data_type,
 304                                this->desc()->diff_weights_desc.data_type)
 305                     && mkldnn_thr_syncable();
 306             if (!ok)
 307                 return status::unimplemented;
 308
 309             status_t status =
 310                 jit_avx512_common_conv_winograd_bwd_weights_kernel_f32::
 311                 init_conf(jcp_, *this->desc(), *this->src_pd_.desc(),
 312                         *this->diff_dst_pd_.desc(),
 313                         *this->diff_weights_pd_.desc());
 314             if (status != status::success) return status;
 315
 316             auto scratchpad = this->scratchpad_registry().registrar();
 317             winograd_avx512_common::init_scratchpad(scratchpad, jcp_);
 318
 319             if (status == status::success
 320                     && this->desc()->alg_kind == alg_kind::convolution_auto)
 321                 CHECK(this->set_alg_kind(alg_kind::convolution_winograd));
 322
 323             return status;
 324         }
 325
 326         jit_conv_winograd_conf_t jcp_;
 327
 328     protected:
 329         virtual status_t set_default_params() override
 330         {
 331             using namespace memory_format;
 332
 333             if (this->src_pd_.desc()->format == any)
 334                 CHECK(this->src_pd_.set_format(nChw16c));
 335             if (this->diff_dst_pd_.desc()->format == any)
 336                 CHECK(this->diff_dst_pd_.set_format(nChw16c));
 337             if (this->diff_weights_pd_.desc()->format == any)
 338                 CHECK(this->diff_weights_pd_.set_format(
 339                         this->with_groups() ? gOIhw16i16o : OIhw16i16o));
 340             if (diff_bias_pd_.desc()->format == any)
 341                 CHECK(diff_bias_pd_.set_format(x));
 342             return status::success;
 343         }
 344     };
 345
 346     jit_avx512_common_convolution_winograd_bwd_weights_t(const pd_t *apd,
 347             const input_vector &inputs, const output_vector &outputs)
 348         : cpu_primitive_t(apd, inputs, outputs, true), kernel_(nullptr)
 349     {
 350         kernel_ = new jit_avx512_common_conv_winograd_bwd_weights_kernel_f32(
 351                 pd()->jcp_);
 352     }
 353
 354     ~jit_avx512_common_convolution_winograd_bwd_weights_t()
 355     { delete kernel_; }
 356
 357     typedef typename prec_traits<data_type::f32>::type data_t;
 358
 359     virtual void execute(event_t *e) const
 360     {
 361         assert(pd()->desc()->prop_kind == prop_kind::backward_weights
 362                 && "invalid prop_kind");
 363         _execute_backward_weights_S_D_G_W(scratchpad());
 364         e->set_state(event_t::ready);
 365     }
 366
 367 private:
 368     void _execute_backward_weights_S_D_G_W(
 369             const memory_tracking::grantor_t &scratchpad) const;
 370     void _maybe_execute_diff_bias_copy(
 371             const memory_tracking::grantor_t &scratchpad) const;
 372
 373     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 374     jit_avx512_common_conv_winograd_bwd_weights_kernel_f32 *kernel_;
 375 };
 376
 377 void trans_W_4x4_3x3(float Fw_[6][6][16][16], float F[3][3][16][16]);
 378 void trans_O_4x4_3x3(float Mw[6][6][16], float O[4][4][16]);
 379 void trans_W_3x3_4x4(float Fw[6][6][16], float F[4][6][16]);
 380 void trans_O_3x3_4x4(float Mw[6][6][16][16], float M[3][3][16][16]);
 381 void trans_I_4x4_3x3(float Iw[6][6][16], float I[6][6][16]);
 382 void trans_W_3x3_4x4_wu(float Fw[6][6][16], float F[4][6][16]);
 383 void trans_O_3x3_4x4_wu(float Mw[6][6][16][16], float M[3][3][16][16]);
 384
 385 }
 386 }
 387 }
 388
 389 #endif
 390
 391 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s