inference-engine/thirdparty/mkl-dnn/src/cpu/gemm_x8s8s32x_convolution.hpp

   1 /*******************************************************************************
   2 * Copyright 2017-2019 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 GEMM_X8S8S32X_CONVOLUTION_HPP
  18 #define GEMM_X8S8S32X_CONVOLUTION_HPP
  19
  20 #include "c_types_map.hpp"
  21 #include "memory_tracking.hpp"
  22
  23 #include "cpu_convolution_pd.hpp"
  24 #include "cpu_engine.hpp"
  25 #include "jit_primitive_conf.hpp"
  26 #include "jit_generator.hpp"
  27 #include "gemm_convolution_utils.hpp"
  28
  29 #include "gemm/gemm.hpp"
  30
  31 namespace mkldnn {
  32 namespace impl {
  33 namespace cpu {
  34
  35 template <data_type_t src_type, data_type_t dst_type>
  36 struct _gemm_x8s8s32x_convolution_fwd_t: public cpu_primitive_t {
  37     struct pd_t: public cpu_convolution_fwd_pd_t {
  38         pd_t(engine_t *engine, const convolution_desc_t *adesc,
  39                 const primitive_attr_t *attr,
  40                 const typename pd_t::base_class *hint_fwd_pd)
  41             : cpu_convolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd)
  42             , jcp_() {}
  43
  44         DECLARE_COMMON_PD_T(IGEMM_S8U8S32_IMPL_STR,
  45                 _gemm_x8s8s32x_convolution_fwd_t<src_type, dst_type>);
  46
  47         virtual status_t init() override {
  48             using namespace data_type;
  49             using namespace memory_format;
  50
  51             assert(this->engine()->kind() == engine_kind::cpu);
  52
  53             bool ok = true
  54                 && this->set_default_params() == status::success
  55                 && utils::one_of(this->desc()->prop_kind,
  56                         prop_kind::forward_training,
  57                         prop_kind::forward_inference)
  58                 && utils::one_of(this->desc()->alg_kind,
  59                         alg_kind::convolution_auto,
  60                         alg_kind::convolution_direct)
  61                 && !this->has_zero_dim_memory()
  62                 && this->desc()->src_desc.data_type == src_type
  63                 && this->desc()->dst_desc.data_type == dst_type
  64                 && this->desc()->weights_desc.data_type == s8
  65                 && IMPLICATION(this->with_bias(), utils::one_of(
  66                             this->desc()->bias_desc.data_type, f32, s32, s8,
  67                             u8))
  68                 && this->desc()->accum_data_type == data_type::s32
  69                 && utils::one_of(this->src_pd_.desc()->format, nhwc, ndhwc)
  70                 && this->src_pd_.desc()->format == this->dst_pd_.desc()->format
  71                 && IMPLICATION(this->src_pd_.desc()->format == nhwc,
  72                         this->weights_pd_.desc()->format == (this->with_groups()
  73                                 ? ((src_type == data_type::s8) ? hwigo_s8s8 : hwigo)
  74                                 : ((src_type == data_type::s8) ? hwio_s8s8 : hwio)))
  75                 && IMPLICATION(this->src_pd_.desc()->format == ndhwc,
  76                         this->weights_pd_.desc()->format == (this->with_groups()
  77                                 ? ((src_type == data_type::s8) ? dhwigo_s8s8 : dhwigo)
  78                                 : ((src_type == data_type::s8) ? dhwio_s8s8 : dhwio)))
  79                 && this->is_gemm_conv_format();
  80             if (!ok) return status::unimplemented;
  81
  82             auto scratchpad = scratchpad_registry().registrar();
  83             return jit_gemm_convolution_utils::init_conf(jcp_, scratchpad,
  84                     *this->desc(), this->src_pd(), this->weights_pd(0),
  85                     this->dst_pd(), mkldnn_get_max_threads());
  86         }
  87
  88         jit_gemm_conv_conf_t jcp_;
  89
  90     protected:
  91         memory_format_t src_format() const {
  92             using namespace memory_format;
  93             const size_t ndims_sp = this->desc()->src_desc.ndims - 4;
  94             return (utils::pick(ndims_sp, nhwc, ndhwc));
  95         }
  96
  97         memory_format_t wei_format() const {
  98             using namespace memory_format;
  99             const size_t ndims_sp = this->desc()->src_desc.ndims - 4;
 100             return this->with_groups()
 101                 ? (src_type == data_type::s8) ? utils::pick(ndims_sp, hwigo_s8s8, dhwigo_s8s8)
 102                                               : utils::pick(ndims_sp, hwigo, dhwigo)
 103                 : (src_type == data_type::s8) ? utils::pick(ndims_sp, hwio_s8s8, dhwio_s8s8)
 104                                               : utils::pick(ndims_sp, hwio, dhwio);
 105         }
 106
 107         virtual status_t set_default_params() override {
 108             using namespace memory_format;
 109             if (this->src_pd_.desc()->format == any)
 110                 CHECK(this->src_pd_.set_format(src_format()));
 111             if (this->dst_pd_.desc()->format == any)
 112                 CHECK(this->dst_pd_.set_format(src_format()));
 113             if (this->weights_pd_.desc()->format == any)
 114                 CHECK(this->weights_pd_.set_format(wei_format()));
 115             if (this->bias_pd_.desc()->format == any)
 116                 CHECK(this->bias_pd_.set_format(x));
 117             if (this->desc()->alg_kind == alg_kind::convolution_auto)
 118                 CHECK(this->set_alg_kind(alg_kind::convolution_direct));
 119             return status::success;
 120         }
 121
 122         virtual bool is_gemm_conv_format() const {
 123             using namespace mkldnn::impl::primitive_kind;
 124             auto const &po = this->attr()->post_ops_;
 125             auto is_relu = [&](int idx) {
 126                 return po.entry_[idx].is_relu(true, false); };
 127
 128             switch (po.len_) {
 129             case 0: return true;
 130             case 1: return is_relu(0) || po.contain(sum, 0);
 131             case 2: return po.contain(sum, 0) && is_relu(1);
 132             default: return false;
 133             }
 134         }
 135     };
 136
 137     _gemm_x8s8s32x_convolution_fwd_t(const pd_t *apd, const input_vector &inputs,
 138            const output_vector &outputs)
 139         : cpu_primitive_t(apd, inputs, outputs, true) {
 140         pp_ker_ = new pp_ker_t(this->pd());
 141     }
 142     ~_gemm_x8s8s32x_convolution_fwd_t() {
 143         delete pp_ker_;
 144     }
 145
 146     typedef typename prec_traits<src_type>::type src_data_t;
 147     typedef typename prec_traits<data_type::s8>::type wei_data_t;
 148     typedef typename prec_traits<dst_type>::type dst_data_t;
 149     typedef typename prec_traits<data_type::s32>::type acc_data_t;
 150
 151     virtual void execute(event_t *e) const {
 152         execute_forward();
 153         e->set_state(event_t::ready);
 154     }
 155
 156 private:
 157     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 158     void execute_forward() const;
 159     // XXX: this is throwaway code that will become unnecessary when we have a
 160     // sufficiently advanced igemm jit generator that supports quantization,
 161     // relu, and whatnot
 162     class pp_ker_t : jit_generator {
 163     public:
 164         DECLARE_CPU_JIT_AUX_FUNCTIONS(
 165         _gemm_x8s8s32x_convolution_fwd_t::pp_kernel);
 166         pp_ker_t(const pd_t *pd);
 167
 168         void operator()(dst_data_t *dst, const acc_data_t *acc,
 169             const char *bias, const float *scales,
 170             float nslope, float sum_scale, float signed_scale,
 171             int g, size_t start, size_t end);
 172
 173         size_t dst_os_stride_;
 174
 175     private:
 176         void generate();
 177
 178         struct ker_args {
 179             dst_data_t *dst;
 180             const acc_data_t *acc;
 181             const char *bias;
 182             const float *scales;
 183             float nslope;
 184             float sum_scale;
 185             float signed_scale;
 186             size_t len;
 187             size_t oc_offset;
 188         };
 189         void(*ker_)(const ker_args *args);
 190
 191         const jit_gemm_conv_conf_t jcp_;
 192         size_t OC_;
 193         size_t OS_;
 194         data_type_t bias_data_type_;
 195         size_t bias_data_type_size_;
 196         size_t scale_idx_mult_;
 197         round_mode_t rmode_;
 198         bool do_bias_;
 199         bool do_relu_;
 200         bool do_sum_;
 201         bool do_signed_scaling_;
 202         size_t vlen_;
 203     };
 204
 205
 206     void execute_forward_thr(const int ithr, const int nthr,
 207             const src_data_t *src_base, const wei_data_t *wei_base,
 208             const char *bia_base, dst_data_t *dst_base,
 209             const memory_tracking::grantor_t &scratchpad) const;
 210
 211     int nthr_;
 212     pp_ker_t *pp_ker_;
 213
 214 };
 215
 216 template <data_type_t dst_type>
 217 struct _gemm_u8s8s32x_convolution_bwd_data_t: public cpu_primitive_t {
 218     struct pd_t: public cpu_convolution_bwd_data_pd_t{
 219         pd_t(engine_t *engine,
 220                 const convolution_desc_t *adesc, const primitive_attr_t *attr,
 221                 const convolution_fwd_pd_t *hint_fwd_pd)
 222             : cpu_convolution_bwd_data_pd_t(engine, adesc, attr, hint_fwd_pd)
 223             , jcp_() {}
 224
 225         DECLARE_COMMON_PD_T(IGEMM_S8U8S32_IMPL_STR,
 226                 _gemm_u8s8s32x_convolution_bwd_data_t<dst_type>);
 227
 228         virtual status_t init() override {
 229             using namespace data_type;
 230             using namespace memory_format;
 231
 232             assert(this->engine()->kind() == engine_kind::cpu);
 233
 234             bool ok = true
 235                 && this->set_default_params() == status::success
 236                 && this->desc()->prop_kind == prop_kind::backward_data
 237                 && utils::one_of(this->desc()->alg_kind, alg_kind::convolution_auto,
 238                            alg_kind::convolution_direct)
 239                 && !this->has_zero_dim_memory()
 240                 && this->desc()->diff_src_desc.data_type == dst_type
 241                 && this->desc()->diff_dst_desc.data_type == u8
 242                 && this->desc()->weights_desc.data_type == s8
 243                 && IMPLICATION(this->with_bias(), utils::one_of(
 244                             this->desc()->bias_desc.data_type, f32, s32, s8,
 245                             u8))
 246                 && this->desc()->accum_data_type == data_type::s32
 247                 && utils::everyone_is(nhwc, this->diff_src_pd_.desc()->format,
 248                         this->diff_dst_pd_.desc()->format)
 249                 && this->weights_pd_.desc()->format == (this->with_groups()
 250                         ? hwigo : hwio)
 251                 && attr()->post_ops_.has_default_values();
 252             if (!ok) return status::unimplemented;
 253
 254             auto scratchpad = scratchpad_registry().registrar();
 255             return jit_gemm_convolution_utils::init_conf(jcp_, scratchpad,
 256                     *this->desc(), this->diff_src_pd(), this->weights_pd(0),
 257                     this->diff_dst_pd(), mkldnn_get_max_threads());
 258         }
 259
 260         virtual bool support_bias() const override { return true; }
 261
 262         jit_gemm_conv_conf_t jcp_;
 263
 264     protected:
 265         virtual status_t set_default_params() override {
 266             using namespace memory_format;
 267
 268             if (this->diff_src_pd_.desc()->format == any)
 269                 CHECK(this->diff_src_pd_.set_format(nhwc));
 270             if (this->diff_dst_pd_.desc()->format == any)
 271                 CHECK(this->diff_dst_pd_.set_format(nhwc));
 272             if (this->weights_pd_.desc()->format == any)
 273                 CHECK(this->weights_pd_.set_format(
 274                             this->with_groups() ? hwigo : hwio));
 275             if (bias_pd_.desc()->format == any)
 276                 CHECK(bias_pd_.set_format(x));
 277             if (this->desc()->alg_kind == alg_kind::convolution_auto)
 278                 CHECK(this->set_alg_kind(alg_kind::convolution_direct));
 279              return status::success;
 280         }
 281     };
 282
 283     _gemm_u8s8s32x_convolution_bwd_data_t(const pd_t *apd, const input_vector &inputs,
 284            const output_vector &outputs)
 285         : cpu_primitive_t(apd, inputs, outputs, true) {}
 286     ~_gemm_u8s8s32x_convolution_bwd_data_t() {}
 287
 288     typedef typename prec_traits<data_type::u8>::type diff_dst_data_t;
 289     typedef typename prec_traits<data_type::s8>::type wei_data_t;
 290     typedef typename prec_traits<dst_type>::type diff_src_data_t;
 291     typedef typename prec_traits<data_type::s32>::type acc_data_t;
 292
 293     virtual void execute(event_t *e) const {
 294         execute_backward_data();
 295         e->set_state(event_t::ready);
 296     }
 297
 298 private:
 299     void execute_backward_data() const;
 300     void execute_backward_data_thr(const int ithr, const int nthr,
 301             const diff_dst_data_t *diff_dst_base, const wei_data_t *wei_base,
 302             const char *bia_base, diff_src_data_t *diff_src_base,
 303             const memory_tracking::grantor_t &scratchpad) const;
 304     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 305 };
 306
 307 }
 308 }
 309 }
 310
 311 #endif