inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx512_core_u8s8s32x_1x1_convolution.hpp

   1 /*******************************************************************************
   2 * Copyright 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_U8S8S32X_1X1_CONVOLUTION_HPP
  18 #define CPU_JIT_AVX512_CORE_U8S8S32X_1X1_CONVOLUTION_HPP
  19
  20 #include "c_types_map.hpp"
  21 #include "cpu_convolution_pd.hpp"
  22 #include "cpu_engine.hpp"
  23 #include "cpu_reducer.hpp"
  24 #include "mkldnn_thread.hpp"
  25 #include "utils.hpp"
  26
  27 #include "jit_uni_1x1_conv_utils.hpp"
  28 #include "jit_avx512_core_u8s8s32x_1x1_conv_kernel.hpp"
  29
  30 namespace mkldnn {
  31 namespace impl {
  32 namespace cpu {
  33
  34 template <bool with_relu, impl::data_type_t dst_type>
  35 struct _jit_avx512_core_u8s8s32x_1x1_convolution_fwd_t : public cpu_primitive_t {
  36     struct pd_t: public _cpu_convolution_fwd_pd_t<with_relu> {
  37         pd_t(engine_t *engine,
  38                 const typename pd_t::base_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<with_relu>(engine, adesc, attr,
  42                     hint_fwd_pd)
  43             , jcp_(), rtus_() {}
  44
  45         DECLARE_COMMON_PD_T(
  46                 JIT_IMPL_NAME_HELPER("jit_1x1:", avx512_core, ""),
  47                 _jit_avx512_core_u8s8s32x_1x1_convolution_fwd_t<with_relu,
  48                 dst_type>);
  49
  50         virtual status_t init() override {
  51             using namespace prop_kind;
  52             using namespace utils;
  53             assert(this->engine()->kind() == engine_kind::cpu);
  54             bool ok = true
  55                 && this->set_default_params() == status::success
  56                 && utils::one_of(this->cdesc_().prop_kind, forward_training,
  57                         forward_inference)
  58                 && this->cdesc_().alg_kind == alg_kind::convolution_direct
  59                 && this->cdesc_().src_desc.data_type == data_type::u8
  60                 && this->cdesc_().dst_desc.data_type == dst_type
  61                 && this->cdesc_().weights_desc.data_type == data_type::s8
  62                 && utils::implication(this->with_bias(), utils::one_of(
  63                             this->cdesc_().bias_desc.data_type, data_type::f32,
  64                             data_type::s32, data_type::s8, data_type::u8))
  65                 && this->cdesc_().accum_data_type == data_type::s32;
  66
  67             if (!ok) return status::unimplemented;
  68
  69             const convolution_desc_t *conv_d = &this->cdesc_();
  70             const memory_desc_t *src_d = this->src_pd_.desc();
  71             rtus_prepare(this, conv_d, src_d, this->dst_pd_.desc());
  72             return jit_avx512_core_u8s8s32x_1x1_conv_kernel::init_conf(jcp_,
  73                     *conv_d, *src_d, *this->weights_pd_.desc(),
  74                     *this->dst_pd_.desc(), *this->bias_pd_.desc(), *this->attr(),
  75                     with_relu, this->negative_slope(),
  76                     omp_get_max_threads(), rtus_.reduce_src_);
  77         }
  78
  79         jit_1x1_conv_conf_t jcp_;
  80         struct reduce_to_unit_stride_t {
  81             convolution_desc_t conv_d_;
  82             bool reduce_src_;
  83         } rtus_;
  84
  85       protected:
  86         virtual status_t set_default_params() override {
  87             using namespace memory_format;
  88             if (this->src_pd_.desc()->format == any)
  89                 CHECK(this->src_pd_.set_format(nhwc));
  90             if (this->dst_pd_.desc()->format == any)
  91                 CHECK(this->dst_pd_.set_format(nhwc));
  92             if (this->weights_pd_.desc()->format == any)
  93                 CHECK(this->weights_pd_.set_format(this->with_groups()
  94                                         ? gOIhw4i16o4i : OIhw4i16o4i));
  95             if (this->bias_pd_.desc()->format == any)
  96                 CHECK(this->bias_pd_.set_format(x));
  97             return status::success;
  98         }
  99     };
 100
 101     template <cpu_isa_t isa, typename conv_t>
 102     friend void init_rtus_driver(conv_t *self);
 103     _jit_avx512_core_u8s8s32x_1x1_convolution_fwd_t(const pd_t *pd,
 104                                           const input_vector &inputs,
 105                                           const output_vector &outputs)
 106         : cpu_primitive_t(&conf_, inputs, outputs), conf_(*pd)
 107         , kernel_(nullptr), rtus_driver_(nullptr), ws_per_thread_(0)
 108         , scratch_(nullptr)
 109     {
 110         kernel_ = new jit_avx512_core_u8s8s32x_1x1_conv_kernel(conf_.jcp_,
 111                     *conf_.attr());
 112
 113         ws_size_ = conf_.jcp_.mb * conf_.jcp_.oc * conf_.jcp_.ow * conf_.jcp_.oh;
 114         ws_ = (acc_data_t *)malloc(ws_size_ * sizeof(acc_data_t), 64);
 115         init_rtus_driver<avx512_common>(this);
 116     }
 117     ~_jit_avx512_core_u8s8s32x_1x1_convolution_fwd_t() {
 118         delete kernel_;
 119         delete rtus_driver_;
 120         free(ws_);
 121         free(scratch_);
 122     }
 123
 124     typedef typename prec_traits<data_type::u8>::type src_data_t;
 125     typedef typename prec_traits<data_type::s8>::type wei_data_t;
 126     typedef typename prec_traits<dst_type>::type dst_data_t;
 127     typedef typename prec_traits<data_type::s32>::type acc_data_t;
 128
 129     virtual void execute(event_t *e) {
 130         execute_forward();
 131         e->set_state(event_t::ready);
 132     }
 133
 134   private:
 135     void execute_forward();
 136     pd_t conf_;
 137     jit_avx512_core_u8s8s32x_1x1_conv_kernel *kernel_;
 138
 139     rtus_driver_t<avx512_common> *rtus_driver_;
 140     size_t ws_per_thread_;
 141     src_data_t *scratch_;
 142     acc_data_t *ws_;
 143     size_t ws_size_;
 144 };
 145
 146 template <impl::data_type_t dst_type>
 147 using jit_avx512_core_u8s8s32x_1x1_convolution_fwd_t =
 148     _jit_avx512_core_u8s8s32x_1x1_convolution_fwd_t<false, dst_type>;
 149
 150 template <impl::data_type_t dst_type>
 151 using jit_avx512_core_u8s8s32x_1x1_convolution_relu_t =
 152     _jit_avx512_core_u8s8s32x_1x1_convolution_fwd_t<true, dst_type>;
 153
 154 }
 155 }
 156 }
 157
 158 #endif