inference-engine/thirdparty/mkl-dnn/src/cpu/gemm_x8s8s32x_inner_product.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 GEMM_X8S8S32X_INNER_PRODUCT_HPP
  18 #define GEMM_X8S8S32X_INNER_PRODUCT_HPP
  19
  20 #include <assert.h>
  21
  22 #include "c_types_map.hpp"
  23 #include "memory_tracking.hpp"
  24 #include "type_helpers.hpp"
  25 #include "utils.hpp"
  26
  27 #include "gemm/gemm.hpp"
  28 #include "jit_generator.hpp"
  29 #include "gemm_inner_product_utils.hpp"
  30
  31 #include "cpu_inner_product_pd.hpp"
  32
  33 namespace mkldnn {
  34 namespace impl {
  35 namespace cpu {
  36
  37 template <impl::data_type_t src_type, impl::data_type_t dst_type>
  38 struct gemm_x8s8s32x_inner_product_fwd_t: public cpu_primitive_t {
  39     struct pd_t: public cpu_inner_product_fwd_pd_t {
  40         pd_t(engine_t *engine, const inner_product_desc_t *adesc,
  41                 const primitive_attr_t *attr,
  42                 const inner_product_fwd_pd_t *hint_fwd_pd)
  43             : cpu_inner_product_fwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}
  44
  45         DECLARE_COMMON_PD_T(src_type == data_type::u8
  46                 ? IGEMM_S8U8S32_IMPL_STR
  47                 : IGEMM_S8S8S32_IMPL_STR,
  48                 gemm_x8s8s32x_inner_product_fwd_t);
  49
  50         virtual status_t init() override {
  51             using namespace utils;
  52             using namespace data_type;
  53
  54             assert(engine()->kind() == engine_kind::cpu);
  55
  56             bool ok = true
  57                 && this->set_default_params() == status::success
  58                 && one_of(desc()->prop_kind, prop_kind::forward_training,
  59                         prop_kind::forward_inference)
  60                 && !has_zero_dim_memory()
  61                 && this->desc()->src_desc.data_type == src_type
  62                 && this->desc()->dst_desc.data_type == dst_type
  63                 && this->desc()->weights_desc.data_type == s8
  64                 && IMPLICATION(this->with_bias(), utils::one_of(
  65                             this->desc()->bias_desc.data_type, f32, s32, s8,
  66                             u8))
  67                 && attr()->post_ops_.len_ <= 1
  68                 && IMPLICATION(attr()->post_ops_.len_,
  69                         attr()->post_ops_.entry_[0].is_eltwise())
  70                 && dense_gemm_consitency_check(src_pd(), weights_pd(),
  71                         dst_pd());
  72             if (!ok) return status::unimplemented;
  73
  74             dst_is_acc_ = one_of(dst_type, s32, f32);
  75
  76             init_scratchpad();
  77
  78             return status::success;
  79         }
  80
  81         bool dst_is_acc_;
  82
  83     protected:
  84         virtual status_t set_default_params() override {
  85             using namespace memory_format;
  86
  87             if (this->src_pd_.desc()->format == any) {
  88                 CHECK(this->src_pd_.set_format(
  89                         utils::pick(ndims() - 2, nc, nwc, nhwc, ndhwc)));
  90             }
  91             if (this->dst_pd_.desc()->format == any)
  92                 CHECK(this->dst_pd_.set_format(nc));
  93             if (this->weights_pd_.desc()->format == any) {
  94                 if (MB() > 1) {
  95                     CHECK(this->weights_pd_.set_format(
  96                         utils::pick(ndims() - 2, io, wio, hwio, dhwio)));
  97                 } else {
  98                     CHECK(this->weights_pd_.set_format(
  99                         utils::pick(ndims() - 2, oi, owi, ohwi, odhwi)));
 100                 }
 101             }
 102             if (this->bias_pd_.desc()->format == any)
 103                 CHECK(this->bias_pd_.set_format(x));
 104
 105             return status::success;
 106         }
 107
 108     private:
 109         void init_scratchpad() {
 110             if (!dst_is_acc_) {
 111                 auto scratchpad = scratchpad_registry().registrar();
 112                 scratchpad.book(
 113                         memory_tracking::names::key_iprod_int_dat_in_acc_dt,
 114                         sizeof(acc_data_t) * MB() * OC());
 115             }
 116         }
 117     };
 118
 119     gemm_x8s8s32x_inner_product_fwd_t(const pd_t *apd,
 120             const input_vector &inputs, const output_vector &outputs)
 121         : cpu_primitive_t(apd, inputs, outputs, true) {
 122         pp_kernel_ = new inner_product_utils::pp_kernel_t<data_type::s32,
 123                 dst_type>(apd);
 124     }
 125     ~gemm_x8s8s32x_inner_product_fwd_t() { delete pp_kernel_; }
 126
 127     typedef typename prec_traits<dst_type>::type data_t;
 128
 129     typedef typename prec_traits<src_type>::type src_data_t;
 130     typedef typename prec_traits<data_type::s8>::type wei_data_t;
 131     typedef typename prec_traits<dst_type>::type dst_data_t;
 132     typedef typename prec_traits<data_type::s32>::type acc_data_t;
 133
 134     virtual void execute(event_t *e) const {
 135         execute_forward();
 136         e->set_state(event_t::ready);
 137     }
 138
 139 private:
 140     void execute_forward() const;
 141     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 142
 143     inner_product_utils::pp_kernel_t<data_type::s32, dst_type> *pp_kernel_;
 144 };
 145 }
 146 }
 147 }
 148
 149 #endif
 150
 151 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s