inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx512_core_x8s8s32x_convolution.hpp

   1 /*******************************************************************************
   2 * Copyright 2016-2018 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_CORE_X8S8S32X_CONVOLUTION_HPP
  18 #define CPU_JIT_AVX512_CORE_X8S8S32X_CONVOLUTION_HPP
  19
  20 #include "c_types_map.hpp"
  21 #include "memory_tracking.hpp"
  22 #include "mkldnn_thread.hpp"
  23 #include "utils.hpp"
  24
  25 #include "cpu_convolution_pd.hpp"
  26
  27 #include "jit_avx512_core_x8s8s32x_conv_kernel.hpp"
  28
  29 namespace mkldnn {
  30 namespace impl {
  31 namespace cpu {
  32
  33 template <impl::data_type_t src_type, impl::data_type_t dst_type>
  34 struct jit_avx512_core_x8s8s32x_convolution_fwd_t : public cpu_primitive_t {
  35     struct pd_t : public cpu_convolution_fwd_pd_t {
  36         pd_t(engine_t *engine, const convolution_desc_t *adesc,
  37                 const primitive_attr_t *attr,
  38                 const typename pd_t::base_class *hint_fwd_pd)
  39             : cpu_convolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd)
  40             , jcp_()
  41         {}
  42
  43         DECLARE_COMMON_PD_T(
  44                 JIT_IMPL_NAME_HELPER("jit_int8:", avx512_core, ""),
  45                 jit_avx512_core_x8s8s32x_convolution_fwd_t<src_type, dst_type>);
  46
  47         virtual status_t init() override {
  48             using namespace prop_kind;
  49             assert(this->engine()->kind() == engine_kind::cpu);
  50
  51             bool ok = true
  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                     && this->desc()->src_desc.data_type == src_type
  59                     && this->desc()->dst_desc.data_type == dst_type
  60                     && IMPLICATION(this->with_bias(), utils::one_of(
  61                             this->desc()->bias_desc.data_type, data_type::f32,
  62                             data_type::s32, data_type::s8, data_type::u8))
  63                     && this->desc()->accum_data_type == data_type::s32;
  64             if (!ok) return status::unimplemented;
  65
  66             status_t status = jit_avx512_core_x8s8s32x_fwd_kernel::init_conf(
  67                     jcp_, *this->desc(), this->src_pd_, this->weights_pd_,
  68                     this->dst_pd_,this->bias_pd_, *this->attr(),
  69                     mkldnn_get_max_threads());
  70             if (status != status::success) return status;
  71
  72             auto scratchpad = scratchpad_registry().registrar();
  73             jit_avx512_core_x8s8s32x_fwd_kernel::init_scratchpad(scratchpad,
  74                     jcp_, *this->attr());
  75
  76             if (status == status::success
  77                     && this->desc()->alg_kind == alg_kind::convolution_auto)
  78                 CHECK(this->set_alg_kind(alg_kind::convolution_direct));
  79             return status;
  80         }
  81
  82         jit_conv_conf_t jcp_;
  83     };
  84
  85     jit_avx512_core_x8s8s32x_convolution_fwd_t(const pd_t *apd,
  86             const input_vector &inputs, const output_vector &outputs)
  87         : cpu_primitive_t(apd, inputs, outputs)
  88     {
  89         kernel_ = new jit_avx512_core_x8s8s32x_fwd_kernel(pd()->jcp_,
  90                     *pd()->attr());
  91     }
  92
  93     ~jit_avx512_core_x8s8s32x_convolution_fwd_t() { delete kernel_; }
  94
  95     typedef typename prec_traits<src_type>::type src_data_t;
  96     typedef typename prec_traits<data_type::s8>::type wei_data_t;
  97     typedef typename prec_traits<dst_type>::type dst_data_t;
  98
  99     virtual void execute(event_t *e) const
 100     {
 101         const auto &jcp = pd()->jcp_;
 102         if (pd()->ndims() == 3)
 103             execute_forward_1d();
 104         else if (jcp.is_depthwise)
 105             execute_forward_2d_dw();
 106         else
 107             execute_forward_2d();
 108         e->set_state(event_t::ready);
 109     }
 110
 111 private:
 112     void execute_forward_1d() const;
 113     void execute_forward_2d() const;
 114     void execute_forward_2d_dw() const;
 115     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 116
 117     jit_avx512_core_x8s8s32x_fwd_kernel *kernel_;
 118 };
 119
 120 }
 121 }
 122 }
 123
 124 #endif
 125
 126 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s