inference-engine/thirdparty/mkl-dnn/src/cpu/jit_sse42_1x1_convolution.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_SSE42_1x1_CONVOLUTION_HPP
  18 #define CPU_JIT_SSE42_1x1_CONVOLUTION_HPP
  19
  20 #include "c_types_map.hpp"
  21 #include "cpu_convolution_pd.hpp"
  22 #include "cpu_engine.hpp"
  23 #include "jit_sse42_1x1_conv_kernel_f32.hpp"
  24 #include "mkldnn_thread.hpp"
  25 #include "utils.hpp"
  26 #include "jit_uni_depthwise.hpp"
  27
  28 namespace mkldnn {
  29 namespace impl {
  30 namespace cpu {
  31
  32 struct jit_sse42_1x1_convolution_fwd_t: public cpu_primitive_t {
  33     // TODO: (Roma) Code duplication duplication! Remove with templates
  34     //              (maybe...)!
  35     struct pd_t: public cpu_convolution_fwd_pd_t {
  36         pd_t(engine_t *engine,
  37                 const convolution_desc_t *adesc,
  38                 const primitive_attr_t *attr,
  39                 const typename pd_t::base_class *hint_fwd_pd)
  40             : cpu_convolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd)
  41             , jcp_(), jcp_dw_() {}
  42
  43         DECLARE_COMMON_PD_T(
  44                 JIT_IMPL_NAME_HELPER("jit_1x1:", sse42, ""),
  45                 jit_sse42_1x1_convolution_fwd_t);
  46
  47         virtual status_t init() override {
  48             using namespace prop_kind;
  49             assert(this->engine()->kind() == engine_kind::cpu);
  50             bool ok = true
  51                 && this->set_default_params() == status::success
  52                 && utils::one_of(this->desc()->prop_kind, forward_training,
  53                         forward_inference)
  54                 && utils::one_of(this->desc()->alg_kind,
  55                            alg_kind::convolution_auto,
  56                            alg_kind::convolution_direct)
  57                 && !this->has_zero_dim_memory()
  58                 && utils::everyone_is(data_type::f32,
  59                         this->desc()->src_desc.data_type,
  60                         this->desc()->weights_desc.data_type,
  61                         this->desc()->dst_desc.data_type)
  62                 && IMPLICATION(this->with_bias(),
  63                         data_type::f32 == this->desc()->bias_desc.data_type);
  64             if (!ok) return status::unimplemented;
  65
  66             status_t sts_1x1 = jit_sse42_1x1_conv_kernel_f32::init_conf(jcp_,
  67                     *this->desc(),
  68                     *this->src_pd_.desc(), *this->weights_pd_.desc(),
  69                     *this->dst_pd_.desc(), *this->attr());
  70             if (sts_1x1 != status::success) return sts_1x1;
  71
  72             if (jcp_.with_dw_conv) {
  73                 status_t sts_dw = jit_uni_dw_conv_row_f32<sse42>::init_conf(jcp_, jcp_dw_, *this->attr());
  74                 if (sts_dw != status::success) return sts_dw;
  75             }
  76
  77             auto scratchpad = scratchpad_registry().registrar();
  78             jit_sse42_1x1_conv_kernel_f32::init_scratchpad(scratchpad, jcp_, jcp_dw_);
  79
  80             return status::success;
  81         }
  82
  83         jit_1x1_conv_conf_t jcp_;
  84         jit_conv_conf_t jcp_dw_;
  85
  86     protected:
  87         virtual status_t set_default_params() override {
  88             using namespace memory_format;
  89             if (this->src_pd_.desc()->format == any)
  90                 CHECK(this->src_pd_.set_format(utils::pick(this->ndims() - 3,
  91                     nCw8c, nChw8c)));
  92             if (this->dst_pd_.desc()->format == any)
  93                 CHECK(this->dst_pd_.set_format(utils::pick(this->ndims() - 3,
  94                     nCw8c, nChw8c)));
  95             if (this->weights_pd_.desc()->format == any)
  96                 CHECK(this->weights_pd_.set_format(this->with_groups()
  97                     ? utils::pick(this->ndims() - 3, gOIw8i8o, gOIhw8i8o)
  98                     : utils::pick(this->ndims() - 3, OIw8i8o, OIhw8i8o)));
  99             if (this->bias_pd_.desc()->format == any)
 100                 CHECK(this->bias_pd_.set_format(x));
 101             if (this->desc()->alg_kind == alg_kind::convolution_auto)
 102                 CHECK(this->set_alg_kind(alg_kind::convolution_direct));
 103             return status::success;
 104         }
 105     };
 106
 107     jit_sse42_1x1_convolution_fwd_t(const pd_t *apd,
 108             const input_vector &inputs, const output_vector &outputs)
 109         : cpu_primitive_t(apd, inputs, outputs)
 110     {
 111         kernel_ = new jit_sse42_1x1_conv_kernel_f32(pd()->jcp_, pd()->jcp_dw_, *pd()->attr());
 112
 113         if (pd()->jcp_.with_dw_conv) {
 114             kernel_dw_ = new jit_uni_dw_conv_row_f32<sse42>(pd()->jcp_dw_, *pd()->attr(), pd()->jcp_dw_.ch_block);
 115         }
 116     }
 117
 118     ~jit_sse42_1x1_convolution_fwd_t() {
 119         delete kernel_;
 120
 121         if (pd()->jcp_.with_dw_conv) {
 122             delete kernel_dw_;
 123         }
 124     };
 125
 126     typedef typename prec_traits<data_type::f32>::type data_t;
 127
 128     virtual void execute(event_t *e) const {
 129         if (pd()->jcp_.with_dw_conv)
 130             execute_forward_with_dw_conv();
 131         else
 132             execute_forward();
 133
 134         e->set_state(event_t::ready);
 135     }
 136
 137 private:
 138     void execute_forward() const;
 139     void execute_forward_with_dw_conv() const;
 140     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 141
 142     jit_sse42_1x1_conv_kernel_f32 *kernel_;
 143     jit_uni_dw_conv_row_f32<sse42> *kernel_dw_;
 144 };
 145
 146 }
 147 }
 148 }
 149
 150 #endif