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
 116             if (!ok)
 117                 return status::unimplemented;
 118
 119             status_t status =
 120                 jit_avx512_common_conv_winograd_fwd_kernel_f32::init_conf(
 121                         jcp_, *this->desc(), *this->src_pd_.desc(),
 122                         *this->weights_pd_.desc(), *this->dst_pd_.desc(),
 123                         *this->attr());
 124             if (status != status::success) return status;
 125
 126             auto scratchpad = this->scratchpad_registry().registrar();
 127             winograd_avx512_common::init_scratchpad(scratchpad, jcp_);
 128
 129             if (status == status::success
 130                     && this->desc()->alg_kind == alg_kind::convolution_auto)
 131                 CHECK(this->set_alg_kind(alg_kind::convolution_winograd));
 132
 133             return status;
 134         }
 135
 136         jit_conv_winograd_conf_t jcp_;
 137
 138     protected:
 139         virtual status_t set_default_params() override
 140         {
 141             using namespace memory_format;
 142             if (this->src_pd_.desc()->format == any)
 143                 CHECK(this->src_pd_.set_format(nChw16c));
 144             if (this->dst_pd_.desc()->format == any)
 145                 CHECK(this->dst_pd_.set_format(nChw16c));
 146             if (this->weights_pd_.desc()->format == any)
 147                 CHECK(this->weights_pd_.set_format(
 148                         this->with_groups() ? gOIhw16i16o : OIhw16i16o));
 149             if (this->bias_pd_.desc()->format == any)
 150                 CHECK(this->bias_pd_.set_format(x));
 151             return status::success;
 152         }
 153     };
 154
 155     jit_avx512_common_convolution_winograd_fwd_t(const pd_t *apd,
 156             const input_vector &inputs, const output_vector &outputs)
 157         : _jit_avx512_common_convolution_winograd_t<true>(apd->jcp_, apd->attr())
 158         , cpu_primitive_t(apd, inputs, outputs, true) {}
 159
 160     ~jit_avx512_common_convolution_winograd_fwd_t(){};
 161
 162     typedef typename prec_traits<data_type::f32>::type data_t;
 163
 164     virtual void execute(event_t *e) const
 165     {
 166         float *src = (float *)this->input_memory(0);
 167         float *dst = (float *)this->memory();
 168         float *weights = (float *)this->input_memory(1);
 169         float *bias = (float *)this->input_memory(2);
 170
 171         this->_execute_data_W_S_G_D(pd()->MB(), src, dst, weights, bias,
 172                 this->scratchpad());
 173
 174         e->set_state(event_t::ready);
 175     }
 176
 177 private:
 178     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 179 };
 180
 181 struct jit_avx512_common_convolution_winograd_bwd_data_t
 182         : _jit_avx512_common_convolution_winograd_t<false>,
 183         public cpu_primitive_t {
 184     struct pd_t : public cpu_convolution_bwd_data_pd_t {
 185         pd_t(engine_t *engine, const convolution_desc_t *adesc,
 186                 const primitive_attr_t *attr,
 187                 const convolution_fwd_pd_t *hint_fwd_pd)
 188             : cpu_convolution_bwd_data_pd_t(engine, adesc, attr, hint_fwd_pd)
 189             , jcp_() {}
 190
 191         DECLARE_COMMON_PD_T(
 192                 JIT_IMPL_NAME_HELPER("jit_wino:", avx512_common, ""),
 193                 jit_avx512_common_convolution_winograd_bwd_data_t);
 194
 195         virtual status_t init() override
 196         {
 197             using namespace prop_kind;
 198             assert(this->engine()->kind() == engine_kind::cpu);
 199             bool ok = true && this->set_default_params() == status::success
 200                     && utils::one_of(this->desc()->prop_kind, backward_data)
 201                     && utils::one_of(this->desc()->alg_kind,
 202                                alg_kind::convolution_auto,
 203                                alg_kind::convolution_winograd)
 204                     && !this->has_zero_dim_memory()
 205                     && utils::everyone_is(data_type::f32,
 206                                this->desc()->diff_src_desc.data_type,
 207                                this->desc()->weights_desc.data_type,
 208                                this->desc()->diff_dst_desc.data_type)
 209                     && mkldnn_thr_syncable();
 210
 211             if (!ok)
 212                 return status::unimplemented;
 213
 214             status_t status =
 215                 jit_avx512_common_conv_winograd_bwd_data_kernel_f32::init_conf(
 216                         jcp_, *this->desc(), *this->diff_src_pd_.desc(),
 217                         *this->weights_pd_.desc(), *this->diff_dst_pd_.desc());
 218             if (status != status::success) return status;
 219
 220             auto scratchpad = this->scratchpad_registry().registrar();
 221             winograd_avx512_common::init_scratchpad(scratchpad, jcp_);
 222
 223             if (status == status::success
 224                     && this->desc()->alg_kind == alg_kind::convolution_auto)
 225                 CHECK(this->set_alg_kind(alg_kind::convolution_winograd));
 226
 227             return status;
 228         }
 229
 230         jit_conv_winograd_conf_t jcp_;
 231
 232     protected:
 233         virtual status_t set_default_params() override
 234         {
 235             using namespace memory_format;
 236
 237             if (this->diff_src_pd_.desc()->format == any)
 238                 CHECK(this->diff_src_pd_.set_format(nChw16c));
 239             if (this->diff_dst_pd_.desc()->format == any)
 240                 CHECK(this->diff_dst_pd_.set_format(nChw16c));
 241             if (this->weights_pd_.desc()->format == any)
 242                 CHECK(this->weights_pd_.set_format(
 243                         this->with_groups() ? gOIhw16i16o : OIhw16i16o));
 244             return status::success;
 245         }
 246     };
 247
 248     jit_avx512_common_convolution_winograd_bwd_data_t(const pd_t *apd,
 249             const input_vector &inputs, const output_vector &outputs)
 250         : _jit_avx512_common_convolution_winograd_t<false>(apd->jcp_, apd->attr())
 251         , cpu_primitive_t(apd, inputs, outputs, true) {}
 252
 253     ~jit_avx512_common_convolution_winograd_bwd_data_t(){};
 254
 255     typedef typename prec_traits<data_type::f32>::type data_t;
 256
 257     virtual void execute(event_t *e) const
 258     {
 259         assert(pd()->desc()->prop_kind == prop_kind::backward_data
 260                 && "invalid prop_kind");
 261
 262         float *diff_dst = (float *)this->input_memory(0);
 263         float *diff_src = (float *)this->memory();
 264         float *weights = (float *)this->input_memory(1);
 265
 266         this->_execute_data_W_S_G_D(pd()->MB(), diff_dst, diff_src, weights, nullptr,
 267                 this->scratchpad());
 268
 269         e->set_state(event_t::ready);
 270     }
 271
 272 private:
 273     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 274 };
 275
 276 struct jit_avx512_common_convolution_winograd_bwd_weights_t
 277         : public cpu_primitive_t {
 278     struct pd_t : public cpu_convolution_bwd_weights_pd_t {
 279         pd_t(engine_t *engine, const convolution_desc_t *adesc,
 280                 const primitive_attr_t *attr,
 281                 const convolution_fwd_pd_t *hint_fwd_pd)
 282             : cpu_convolution_bwd_weights_pd_t(engine, adesc, attr,
 283                     hint_fwd_pd)
 284             , jcp_() {}
 285
 286         DECLARE_COMMON_PD_T(
 287                 JIT_IMPL_NAME_HELPER("jit_wino:", avx512_common, ""),
 288                 jit_avx512_common_convolution_winograd_bwd_weights_t);
 289
 290         virtual status_t init() override
 291         {
 292             using namespace prop_kind;
 293             assert(this->engine()->kind() == engine_kind::cpu);
 294             bool ok = true && this->set_default_params() == status::success
 295                     && utils::one_of(this->desc()->prop_kind, backward_weights)
 296                     && utils::one_of(this->desc()->alg_kind,
 297                                alg_kind::convolution_auto,
 298                                alg_kind::convolution_winograd)
 299                     && !this->has_zero_dim_memory()
 300                     && utils::everyone_is(data_type::f32,
 301                                this->desc()->src_desc.data_type,
 302                                this->desc()->diff_dst_desc.data_type,
 303                                this->desc()->diff_weights_desc.data_type)
 304                     && mkldnn_thr_syncable();
 305             if (!ok)
 306                 return status::unimplemented;
 307
 308             status_t status =
 309                 jit_avx512_common_conv_winograd_bwd_weights_kernel_f32::
 310                 init_conf(jcp_, *this->desc(), *this->src_pd_.desc(),
 311                         *this->diff_dst_pd_.desc(),
 312                         *this->diff_weights_pd_.desc());
 313             if (status != status::success) return status;
 314
 315             auto scratchpad = this->scratchpad_registry().registrar();
 316             winograd_avx512_common::init_scratchpad(scratchpad, jcp_);
 317
 318             if (status == status::success
 319                     && this->desc()->alg_kind == alg_kind::convolution_auto)
 320                 CHECK(this->set_alg_kind(alg_kind::convolution_winograd));
 321
 322             return status;
 323         }
 324
 325         jit_conv_winograd_conf_t jcp_;
 326
 327     protected:
 328         virtual status_t set_default_params() override
 329         {
 330             using namespace memory_format;
 331
 332             if (this->src_pd_.desc()->format == any)
 333                 CHECK(this->src_pd_.set_format(nChw16c));
 334             if (this->diff_dst_pd_.desc()->format == any)
 335                 CHECK(this->diff_dst_pd_.set_format(nChw16c));
 336             if (this->diff_weights_pd_.desc()->format == any)
 337                 CHECK(this->diff_weights_pd_.set_format(
 338                         this->with_groups() ? gOIhw16i16o : OIhw16i16o));
 339             if (diff_bias_pd_.desc()->format == any)
 340                 CHECK(diff_bias_pd_.set_format(x));
 341             return status::success;
 342         }
 343     };
 344
 345     jit_avx512_common_convolution_winograd_bwd_weights_t(const pd_t *apd,
 346             const input_vector &inputs, const output_vector &outputs)
 347         : cpu_primitive_t(apd, inputs, outputs, true), kernel_(nullptr)
 348     {
 349         kernel_ = new jit_avx512_common_conv_winograd_bwd_weights_kernel_f32(
 350                 pd()->jcp_);
 351     }
 352
 353     ~jit_avx512_common_convolution_winograd_bwd_weights_t()
 354     { delete kernel_; }
 355
 356     typedef typename prec_traits<data_type::f32>::type data_t;
 357
 358     virtual void execute(event_t *e) const
 359     {
 360         assert(pd()->desc()->prop_kind == prop_kind::backward_weights
 361                 && "invalid prop_kind");
 362         _execute_backward_weights_S_D_G_W(scratchpad());
 363         e->set_state(event_t::ready);
 364     }
 365
 366 private:
 367     void _execute_backward_weights_S_D_G_W(
 368             const memory_tracking::grantor_t &scratchpad) const;
 369     void _maybe_execute_diff_bias_copy(
 370             const memory_tracking::grantor_t &scratchpad) const;
 371
 372     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 373     jit_avx512_common_conv_winograd_bwd_weights_kernel_f32 *kernel_;
 374 };
 375
 376 void trans_W_4x4_3x3(float Fw_[6][6][16][16], float F[3][3][16][16]);
 377 void trans_O_4x4_3x3(float Mw[6][6][16], float O[4][4][16]);
 378 void trans_W_3x3_4x4(float Fw[6][6][16], float F[4][6][16]);
 379 void trans_O_3x3_4x4(float Mw[6][6][16][16], float M[3][3][16][16]);
 380 void trans_I_4x4_3x3(float Iw[6][6][16], float I[6][6][16]);
 381 void trans_W_3x3_4x4_wu(float Fw[6][6][16], float F[4][6][16]);
 382 void trans_O_3x3_4x4_wu(float Mw[6][6][16][16], float M[3][3][16][16]);
 383
 384 }
 385 }
 386 }
 387
 388 #endif
 389
 390 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s