inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx512_common_conv_winograd_kernel_f32.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_CONV_WINOGRAD_KERNEL_F32_HPP
  18 #define JIT_AVX512_COMMON_CONV_WINOGRAD_KERNEL_F32_HPP
  19
  20 #include "c_types_map.hpp"
  21 #include "cpu_memory.hpp"
  22
  23 #include "jit_generator.hpp"
  24 #include "jit_primitive_conf.hpp"
  25
  26 namespace mkldnn {
  27 namespace impl {
  28 namespace cpu {
  29
  30 //alpha determines the output tile_size
  31 constexpr int alpha = 6;
  32 constexpr int tile_size = 4;
  33 //simd length used for vectorization
  34 constexpr int simd_w = 16;
  35
  36 struct _jit_avx512_common_conv_winograd_data_kernel_f32 : public jit_generator {
  37     _jit_avx512_common_conv_winograd_data_kernel_f32(
  38             jit_conv_winograd_conf_t ajcp)
  39         : jcp(ajcp)
  40     {
  41         //******************* First iter kernel ********************//
  42         this->gemm_loop_generate(true);
  43         gemm_loop_ker_first_iter
  44                 = (decltype(gemm_loop_ker_first_iter)) this->getCode();
  45
  46         //************** Subsequent iterations kernel **************//
  47         if (jcp.dimK_nb_block > 1) {
  48             align();
  49             const Xbyak::uint8 *addr = getCurr();
  50             this->gemm_loop_generate(false);
  51             gemm_loop_ker = (decltype(gemm_loop_ker))addr;
  52         }
  53     }
  54
  55     DECLARE_CPU_JIT_AUX_FUNCTIONS(_jit_avx512_common_conv_winograd_data_kernel_f32)
  56
  57     static status_t init_conf_common(jit_conv_winograd_conf_t &jcp,
  58             const convolution_desc_t &cd, const memory_desc_wrapper &src_d,
  59             const memory_desc_wrapper &weights_d,
  60             const memory_desc_wrapper &dst_d);
  61
  62     static status_t init_conf_kernel(
  63             jit_conv_winograd_conf_t &jcp, int dimM, int dimN, int dimK);
  64
  65     jit_conv_winograd_conf_t jcp;
  66     void (*gemm_loop_ker)(float *, const float *, const float *);
  67     void (*gemm_loop_ker_first_iter)(float *, const float *, const float *);
  68
  69 protected:
  70     using reg64_t = const Xbyak::Reg64;
  71     enum { typesize = sizeof(float) };
  72
  73     void gemm_loop_generate(bool is_beta_zero);
  74
  75     /* registers used for GEMM */
  76     reg64_t reg_dstC = abi_param1;
  77     reg64_t reg_srcA = abi_param2;
  78     reg64_t reg_srcB = abi_param3;
  79
  80     reg64_t reg_dimM_block_loop_cnt = r10;
  81     reg64_t reg_dimK_block_loop_cnt = r11;
  82 };
  83
  84 struct jit_avx512_common_conv_winograd_fwd_kernel_f32
  85         : _jit_avx512_common_conv_winograd_data_kernel_f32 {
  86     using _jit_avx512_common_conv_winograd_data_kernel_f32::
  87             _jit_avx512_common_conv_winograd_data_kernel_f32;
  88
  89     static bool post_ops_ok(jit_conv_conf_t &jcp, const primitive_attr_t &attr);
  90
  91     static status_t init_conf(jit_conv_winograd_conf_t &jcp,
  92             const convolution_desc_t &cd, const memory_desc_wrapper &src_d,
  93             const memory_desc_wrapper &weights_d,
  94             const memory_desc_wrapper &dst_d, const primitive_attr_t &attr);
  95 };
  96
  97 struct jit_avx512_common_conv_winograd_bwd_data_kernel_f32
  98         : public _jit_avx512_common_conv_winograd_data_kernel_f32 {
  99     using _jit_avx512_common_conv_winograd_data_kernel_f32::
 100             _jit_avx512_common_conv_winograd_data_kernel_f32;
 101
 102     static status_t init_conf(jit_conv_winograd_conf_t &jcp,
 103             const convolution_desc_t &cd, const memory_desc_wrapper &diff_src_d,
 104             const memory_desc_wrapper &weights_d,
 105             const memory_desc_wrapper &diff_dst_d);
 106 };
 107
 108 struct jit_avx512_common_conv_winograd_bwd_weights_kernel_f32
 109         : public jit_generator {
 110     DECLARE_CPU_JIT_AUX_FUNCTIONS(_jit_avx512_common_conv_winograd_bwd_weights_kernel_f32)
 111
 112     jit_avx512_common_conv_winograd_bwd_weights_kernel_f32(
 113             jit_conv_winograd_conf_t ajcp)
 114         : jcp(ajcp)
 115     {
 116
 117         //******************* First iter kernel ********************//
 118         {
 119             align();
 120             const Xbyak::uint8 *addr = getCurr();
 121             this->gemm_loop_generate(true);
 122             gemm_loop_ker_first_iter = (decltype(gemm_loop_ker_first_iter))addr;
 123         }
 124
 125         if (jcp.tile_block > 1) {
 126             align();
 127             const Xbyak::uint8 *addr = getCurr();
 128             this->gemm_loop_generate(false);
 129             gemm_loop_ker = (decltype(gemm_loop_ker))addr;
 130         }
 131
 132         if (jcp.ver == ver_4fma) {
 133             align();
 134             const Xbyak::uint8 *addr = getCurr();
 135             this->transpose_ker_generate();
 136             transpose_4fma_ker = (decltype(transpose_4fma_ker))addr;
 137         }
 138     }
 139
 140     static status_t init_conf(jit_conv_winograd_conf_t &jcp,
 141             const convolution_desc_t &cd, const memory_desc_wrapper &src_d,
 142             const memory_desc_wrapper &diff_dst_d,
 143             const memory_desc_wrapper &diff_weights_d);
 144
 145     jit_conv_winograd_conf_t jcp;
 146     void (*gemm_loop_ker)(float *, const float *, const float *);
 147     void (*gemm_loop_ker_first_iter)(float *, const float *, const float *);
 148     void (*transpose_4fma_ker)(float *, float *);
 149
 150 private:
 151     using reg64_t = const Xbyak::Reg64;
 152     enum { typesize = sizeof(float) };
 153
 154     void gemm_loop_generate(bool is_first_tile);
 155     void transpose_ker_generate();
 156
 157     reg64_t reg_origB = abi_param2;
 158     reg64_t reg_transB = abi_param1;
 159
 160     reg64_t reg_dstC = abi_param1;
 161     reg64_t reg_srcA_const = abi_param2;
 162     reg64_t reg_srcB = abi_param3;
 163
 164     reg64_t reg_sp = rsp;
 165     reg64_t reg_srcA = r9;
 166     reg64_t reg_nb_ic = r10;
 167     reg64_t reg_loop_cpt = r11;
 168     reg64_t reg_transB_idx = r13;
 169
 170     /* Registers used by new kernel */
 171     reg64_t reg_dimM_block_loop_cnt = r10;
 172     reg64_t reg_dimK_block_loop_cnt = r12;
 173     reg64_t reg_dimN_block_loop_cnt = r11;
 174 };
 175 }
 176 }
 177 }
 178
 179 #endif