inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx512_common_1x1_conv_kernel.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 JIT_AVX512_COMMON_1x1_CONV_KERNEL_HPP
  18 #define JIT_AVX512_COMMON_1x1_CONV_KERNEL_HPP
  19
  20 #include "c_types_map.hpp"
  21 #include "jit_generator.hpp"
  22 #include "jit_primitive_conf.hpp"
  23 #include "jit_uni_eltwise.hpp"
  24
  25 namespace mkldnn {
  26 namespace impl {
  27 namespace cpu {
  28
  29 struct jit_avx512_common_1x1_conv_kernel : public jit_generator {
  30     jit_avx512_common_1x1_conv_kernel(jit_1x1_conv_conf_t ajcp,
  31             const primitive_attr_t &attr) : jcp(ajcp), attr_(attr)
  32     {
  33         this->generate();
  34         jit_ker = (void (*)(jit_1x1_conv_call_s *)) this->getCode();
  35     }
  36
  37     DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx512_common_1x1_conv_kernel)
  38
  39     static bool post_ops_ok(jit_1x1_conv_conf_t &jcp,
  40                                 const primitive_attr_t &attr);
  41
  42     static status_t init_conf(jit_1x1_conv_conf_t &jcp,
  43                                 const convolution_desc_t &cd,
  44                                 const memory_desc_wrapper &src_d,
  45                                 const memory_desc_wrapper &weights_d,
  46                                 const memory_desc_wrapper &dst_d,
  47                                 const primitive_attr_t &attr,
  48                                 bool with_relu, float relu_negative_slope,
  49                                 int nthreads, bool reduce_src);
  50
  51     static status_t init_conf(jit_1x1_conv_conf_t &jcp,
  52                               const convolution_desc_t &cd,
  53                               const memory_desc_wrapper &src_d,
  54                               const memory_desc_wrapper &weights_d,
  55                               const memory_desc_wrapper &dst_d,
  56                               const primitive_attr_t &attr,
  57                               int nthreads, bool reduce_src)
  58     {
  59         return init_conf(jcp, cd, src_d, weights_d, dst_d, attr, false, 0.0,
  60         nthreads, reduce_src);
  61     }
  62
  63     jit_1x1_conv_conf_t jcp;
  64     const primitive_attr_t &attr_;
  65     void (*jit_ker)(jit_1x1_conv_call_s *);
  66
  67   private:
  68     using reg64_t = const Xbyak::Reg64;
  69     using zmm_t = const Xbyak::Zmm;
  70     using mask_t = const Xbyak::Opmask;
  71
  72     reg64_t reg_bcast_data = r8;
  73     reg64_t reg_load_data = r10;
  74     reg64_t reg_output_data = r9;
  75     reg64_t aux_reg_bcast_data = r14;
  76     reg64_t aux1_reg_bcast_data = rbx;
  77     reg64_t aux_reg_load_data = r15;
  78     reg64_t imm_addr64 = aux_reg_load_data;
  79     reg64_t aux_reg_output_data = abi_not_param1;
  80     reg64_t reg_load_loop_work = rsi;
  81     reg64_t reg_reduce_loop_work = r11;
  82     reg64_t bcast_loop_iter = rdx;
  83     reg64_t reduce_loop_iter = abi_param1;
  84     reg64_t reg_reduce_pos_flag = rax;
  85     reg64_t reg_output_stride = r13;
  86     reg64_t reg_bias_data = r12;
  87     reg64_t reg_relu_ns = r13;
  88     reg64_t reg_bcast_loop_work = aux1_reg_bcast_data;
  89     mask_t vmask = k7;
  90
  91     Xbyak::Zmm vreg_bcast = Xbyak::Zmm(31);
  92
  93     int bcast_loop_work_offt = 0;
  94     int stack_space_needed = 16;
  95
  96     void bcast_loop(int load_loop_blk);
  97     void reduce_loop(int load_loop_blk, int ur, int substep, bool wraparound);
  98
  99     void generate();
 100     static void balance(jit_1x1_conv_conf_t &jcp, int nthreads);
 101
 102     Xbyak::Label l_table;
 103     jit_uni_eltwise_vector_f32<avx512_common> eltwise_generator;
 104 };
 105 }
 106 }
 107 }
 108
 109 #endif