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
  30 #include "cpu_inner_product_pd.hpp"
  31
  32 namespace mkldnn {
  33 namespace impl {
  34 namespace cpu {
  35
  36 template <impl::data_type_t src_type, impl::data_type_t dst_type>
  37 struct gemm_x8s8s32x_inner_product_fwd_t: public cpu_primitive_t {
  38     struct pd_t: public cpu_inner_product_fwd_pd_t {
  39         pd_t(engine_t *engine, const inner_product_desc_t *adesc,
  40                 const primitive_attr_t *attr,
  41                 const inner_product_fwd_pd_t *hint_fwd_pd)
  42             : cpu_inner_product_fwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}
  43
  44         DECLARE_COMMON_PD_T(src_type == data_type::u8
  45                 ? IGEMM_S8U8S32_IMPL_STR
  46                 : IGEMM_S8S8S32_IMPL_STR,
  47                 gemm_x8s8s32x_inner_product_fwd_t);
  48
  49         virtual status_t init() override {
  50             using namespace utils;
  51             using namespace data_type;
  52
  53             assert(engine()->kind() == engine_kind::cpu);
  54
  55             bool ok = true
  56                 && this->set_default_params() == status::success
  57                 && one_of(desc()->prop_kind, prop_kind::forward_training,
  58                         prop_kind::forward_inference)
  59                 && !has_zero_dim_memory()
  60                 && this->desc()->src_desc.data_type == src_type
  61                 && this->desc()->dst_desc.data_type == dst_type
  62                 && this->desc()->weights_desc.data_type == s8
  63                 && IMPLICATION(this->with_bias(), utils::one_of(
  64                             this->desc()->bias_desc.data_type, f32, s32, s8,
  65                             u8))
  66                 && attr()->post_ops_.len_ <= 1
  67                 && IMPLICATION(attr()->post_ops_.len_,
  68                         attr()->post_ops_.entry_[0].is_relu(true, false))
  69                 && dense_gemm_consitency_check(src_pd(), weights_pd(),
  70                         dst_pd());
  71             if (!ok) return status::unimplemented;
  72
  73             dst_is_acc_ = one_of(dst_type, s32, f32);
  74
  75             init_scratchpad();
  76
  77             return status::success;
  78         }
  79
  80         bool dst_is_acc_;
  81
  82     protected:
  83         virtual status_t set_default_params() override {
  84             using namespace memory_format;
  85
  86             if (this->src_pd_.desc()->format == any) {
  87                 if (ndims() == 4) CHECK(this->src_pd_.set_format(nhwc));
  88                 else if (ndims() == 5) CHECK(this->src_pd_.set_format(ndhwc));
  89                 else CHECK(this->src_pd_.set_format(nc));
  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 (ndims() == 4) CHECK(this->weights_pd_.set_format(hwio));
  95                 else if (ndims() == 5) CHECK(this->weights_pd_.set_format(dhwio));
  96                 else CHECK(this->weights_pd_.set_format(io));
  97             }
  98             if (this->bias_pd_.desc()->format == any)
  99                 CHECK(this->bias_pd_.set_format(x));
 100
 101             return status::success;
 102         }
 103
 104     private:
 105         void init_scratchpad() {
 106             if (!dst_is_acc_) {
 107                 auto scratchpad = scratchpad_registry().registrar();
 108                 scratchpad.book(
 109                         memory_tracking::names::key_iprod_int_dat_in_acc_dt,
 110                         sizeof(acc_data_t) * MB() * OC());
 111             }
 112         }
 113     };
 114
 115     gemm_x8s8s32x_inner_product_fwd_t(const pd_t *apd, const input_vector &inputs,
 116             const output_vector &outputs)
 117         : cpu_primitive_t(apd, inputs, outputs, true)
 118     { pp_kernel_ = new pp_kernel_t(apd, pd()->dst_is_acc_); }
 119     ~gemm_x8s8s32x_inner_product_fwd_t() { delete pp_kernel_; }
 120
 121     typedef typename prec_traits<dst_type>::type data_t;
 122
 123     typedef typename prec_traits<src_type>::type src_data_t;
 124     typedef typename prec_traits<data_type::s8>::type wei_data_t;
 125     typedef typename prec_traits<dst_type>::type dst_data_t;
 126     typedef typename prec_traits<data_type::s32>::type acc_data_t;
 127
 128     virtual void execute(event_t *e) const {
 129         execute_forward();
 130         e->set_state(event_t::ready);
 131     }
 132
 133 private:
 134     // XXX: this is throwaway code that will become unnecessary when we have a
 135     // sufficiently advanced igemm jit generator that supports quantization,
 136     // relu, and whatnot
 137     class pp_kernel_t: jit_generator {
 138     public:
 139         DECLARE_CPU_JIT_AUX_FUNCTIONS(
 140                 gemm_x8s8s32x_inner_product_fwd_t::pp_kernel);
 141         pp_kernel_t(const pd_t *pd, bool dst_is_acc);
 142
 143         void operator()(dst_data_t *dst, const acc_data_t *acc,
 144                 const char *bias, const float *scales, float nslope,
 145                 size_t start, size_t end);
 146     private:
 147         void generate();
 148
 149         struct ker_args {
 150             dst_data_t *dst;
 151             const acc_data_t *acc;
 152             const char *bias;
 153             const float *scales;
 154             float nslope;
 155             size_t len;
 156             size_t oc_offset;
 157         };
 158         void (*ker_)(const ker_args *args);
 159
 160         size_t OC_;
 161         data_type_t bias_data_type_;
 162         size_t bias_data_type_size_;
 163         size_t scale_idx_mult_;
 164         round_mode_t rmode_;
 165         bool do_bias_;
 166         bool do_relu_;
 167     };
 168
 169     void execute_forward() const;
 170     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 171
 172     pp_kernel_t *pp_kernel_;
 173 };
 174 }
 175 }
 176 }
 177
 178 #endif
 179
 180 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s