inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx512_core_u8s8s32x_deconvolution.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_DECONVOLUTION_HPP
  18 #define CPU_JIT_AVX512_CORE_U8S8S32X_DECONVOLUTION_HPP
  19
  20
  21 #include "c_types_map.hpp"
  22 #include "cpu_engine.hpp"
  23 #include "cpu_memory.hpp"
  24 #include "mkldnn_thread.hpp"
  25 #include "type_helpers.hpp"
  26 #include "utils.hpp"
  27 #include "nstl.hpp"
  28
  29 #include "cpu_deconvolution_pd.hpp"
  30 #include "jit_generator.hpp"
  31 #include "jit_primitive_conf.hpp"
  32
  33 namespace mkldnn {
  34 namespace impl {
  35 namespace cpu {
  36
  37 typedef enum {
  38     no_last_block = 0x1U,
  39     last_ic_block = 0x2U,
  40     last_sp_block = 0x4U,
  41     last_ic
  42 } ker_block_t;
  43
  44 struct jit_avx512_core_u8s8s32x_deconv_fwd_kernel : public jit_generator {
  45     DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx512_core_u8s8s32x_deconv_fwd_ker_t);
  46
  47     jit_avx512_core_u8s8s32x_deconv_fwd_kernel(jit_conv_conf_t ajcp,
  48             const primitive_attr_t &attr) : jcp(ajcp), attr_(attr) {
  49         generate();
  50         jit_ker = (void (*)(jit_deconv_call_s *))getCode();
  51     }
  52
  53     static status_t init_conf(jit_conv_conf_t &jcp,
  54             const deconvolution_desc_t &cd,
  55             cpu_memory_t::pd_t &src_pd,
  56             cpu_memory_t::pd_t &weights_pd,
  57             cpu_memory_t::pd_t &dst_pd,
  58             const bool with_bias,
  59             cpu_memory_t::pd_t &bias_pd,
  60             const primitive_attr_t &attr);
  61
  62     jit_conv_conf_t jcp;
  63     const primitive_attr_t &attr_;
  64     void (*jit_ker)(jit_deconv_call_s *);
  65 private:
  66     using reg64_t = const Xbyak::Reg64;
  67     using zmm_t = const Xbyak::Zmm;
  68     using xmm_t = const Xbyak::Xmm;
  69
  70     reg64_t reg_src = r8;
  71     reg64_t reg_filt = r9;
  72     reg64_t reg_dst = r10;
  73     reg64_t param1 = abi_param1;
  74     reg64_t reg_kh = abi_not_param1;
  75     reg64_t reg_nur_w = rbx;
  76     reg64_t reg_bias = rdx;
  77     reg64_t reg_icb = reg_bias;
  78     reg64_t reg_ptr_scales = rax;
  79     reg64_t reg_oc_blocks = rsi;
  80
  81     reg64_t reg_scratch = r14;
  82     reg64_t aux_reg_src = r11;
  83     reg64_t aux_reg_filt = r12;
  84     reg64_t reg_kj = rax;
  85
  86     Xbyak::Opmask ktail_mask = Xbyak::Opmask(2);
  87     zmm_t zmm_tmp = zmm_t(29);
  88     zmm_t zmm_one = zmm_t(30);
  89     zmm_t zmm_zero = zmm_t(31);
  90     zmm_t zmm_wei = zmm_t(31);
  91
  92     zmm_t zmm_out(int i_ur, int i_oc) {
  93         int idx = i_ur * jcp.nb_oc_blocking + i_oc;
  94         assert(idx < 31);
  95         return zmm_t(idx);
  96     }
  97     zmm_t zmm_inp(int i_ic, int nb_x_blocking) {
  98         int idx = i_ic + nb_x_blocking * jcp.ur_w;
  99         assert(idx < 31);
 100         return zmm_t(idx);
 101     }
 102
 103     int get_ow_start(int ki, int l_overflow) {
 104         int res = (jcp.ow - 1 + jcp.r_pad) % jcp.stride_w
 105                 + l_overflow * jcp.stride_w
 106                 - (jcp.kw - 1 - ki) * (jcp.dilate_w + 1);
 107         while (res < 0)
 108             res += jcp.stride_w;
 109         return res;
 110     }
 111
 112     int get_ow_end(int ur_w, int ki, int r_overflow) {
 113         if (utils::one_of(ur_w, jcp.ow, jcp.ur_w_tail))
 114                 ur_w += nstl::min(0, jcp.r_pad);
 115         int res = (ur_w - 1 + jcp.l_pad) % jcp.stride_w
 116             + r_overflow * jcp.stride_w - ki * (jcp.dilate_w + 1);
 117         while (res < 0)
 118             res += jcp.stride_w;
 119         return ur_w - res;
 120     }
 121
 122     void prepare_output(int ur_w);
 123     void store_output(int ur_w, bool last_oc_block);
 124     void compute_ker(int ur_w, int pad_l, int pad_r, ker_block_t last_ker_block);
 125     void compute_loop(int ur_w, int pad_l, int pad_r, bool last_block);
 126     void generate();
 127     void cvt2ps(data_type_t type_in, zmm_t zmm_in, const Xbyak::Operand &op,
 128         bool mask_flag);
 129 };
 130
 131 template <impl::data_type_t dst_type>
 132 struct _jit_avx512_core_u8s8s32x_deconvolution_fwd_t : public cpu_primitive_t {
 133     struct pd_t : public cpu_deconvolution_fwd_pd_t {
 134         pd_t(engine_t *engine,
 135                 const deconvolution_desc_t *adesc,
 136                 const primitive_attr_t *attr,
 137                 const deconvolution_fwd_pd_t *hint_fwd_pd)
 138             : cpu_deconvolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}
 139
 140         DECLARE_COMMON_PD_T(JIT_IMPL_NAME_HELPER("jit_deconvolution:", avx512_core, ""),
 141                 _jit_avx512_core_u8s8s32x_deconvolution_fwd_t<dst_type>);
 142
 143         virtual status_t init() override {
 144             assert(this->engine()->kind() == engine_kind::cpu);
 145
 146             bool ok = true
 147                 && utils::one_of(this->desc()->prop_kind, prop_kind::forward_training,
 148                             prop_kind::forward_inference)
 149                 && this->desc()->alg_kind & alg_kind::deconvolution_direct
 150                 && this->desc()->dst_desc.data_type == dst_type
 151                 && IMPLICATION(this->with_bias(), utils::one_of(
 152                             this->desc()->bias_desc.data_type, data_type::f32,
 153                             data_type::s32, data_type::s8, data_type::u8))
 154                 && this->desc()->accum_data_type == data_type::s32;
 155             if (!ok) return status::unimplemented;
 156
 157             /*TODO: support signed input and postops */
 158             return jit_avx512_core_u8s8s32x_deconv_fwd_kernel::init_conf(
 159                     jcp_, *this->desc(), this->src_pd_,
 160                     this->weights_pd_, this->dst_pd_,
 161                     this->with_bias(), this->bias_pd_,
 162                     *this->attr());
 163         }
 164         jit_conv_conf_t jcp_;
 165     };
 166
 167     _jit_avx512_core_u8s8s32x_deconvolution_fwd_t(const pd_t *pd,
 168            const input_vector &inputs, const output_vector &outputs)
 169        : cpu_primitive_t(&conf_, inputs, outputs), conf_(*pd) {
 170            kernel_ = new jit_avx512_core_u8s8s32x_deconv_fwd_kernel(conf_.jcp_,
 171                    *conf_.attr());
 172        }
 173
 174     ~_jit_avx512_core_u8s8s32x_deconvolution_fwd_t() {
 175         delete kernel_;
 176     }
 177
 178     typedef typename prec_traits<data_type::u8>::type src_data_t;
 179     typedef typename prec_traits<data_type::s8>::type wei_data_t;
 180     typedef typename prec_traits<dst_type>::type dst_data_t;
 181
 182     virtual void execute(event_t *e)
 183     {
 184         execute_forward();
 185         e->set_state(event_t::ready);
 186     }
 187
 188 private:
 189     void execute_forward();
 190     pd_t conf_;
 191     jit_avx512_core_u8s8s32x_deconv_fwd_kernel *kernel_;
 192 };
 193
 194
 195 }
 196 }
 197 }
 198
 199 #endif
 200
 201 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s