inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx512_core_x8s8s32x_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 #include "jit_uni_eltwise.hpp"
  33
  34 namespace mkldnn {
  35 namespace impl {
  36 namespace cpu {
  37
  38 typedef enum {
  39     no_last_block = 0x1U,
  40     last_ic_block = 0x2U,
  41     last_sp_block = 0x4U,
  42 } ker_block_t;
  43
  44 struct jit_avx512_core_x8s8s32x_deconv_fwd_kernel : public jit_generator {
  45     DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx512_core_x8s8s32x_deconv_fwd_ker_t);
  46
  47     jit_avx512_core_x8s8s32x_deconv_fwd_kernel(
  48             jit_conv_conf_t ajcp, const primitive_attr_t &attr)
  49         : jcp(ajcp), attr_(attr), eltwise_injector_(nullptr) {
  50         if (jcp.with_eltwise)
  51             eltwise_injector_ = new jit_uni_eltwise_injector_f32<avx512_common>(
  52                     this, jcp.eltwise);
  53         generate();
  54         jit_ker = (void (*)(jit_deconv_call_s *))getCode();
  55     }
  56
  57     ~jit_avx512_core_x8s8s32x_deconv_fwd_kernel() {
  58             delete eltwise_injector_;
  59     }
  60
  61     static bool post_ops_ok(jit_conv_conf_t &jcp,
  62             const primitive_attr_t &attr);
  63
  64     static status_t init_conf(jit_conv_conf_t &jcp,
  65             const deconvolution_desc_t &cd,
  66             cpu_memory_t::pd_t &src_pd,
  67             cpu_memory_t::pd_t &weights_pd,
  68             cpu_memory_t::pd_t &dst_pd,
  69             const bool with_bias,
  70             cpu_memory_t::pd_t &bias_pd,
  71             const primitive_attr_t &attr);
  72
  73     static void init_scratchpad(memory_tracking::registrar_t &scratchpad,
  74             const jit_conv_conf_t &jcp, const primitive_attr_t &attr);
  75
  76     jit_conv_conf_t jcp;
  77     const primitive_attr_t &attr_;
  78     void (*jit_ker)(jit_deconv_call_s *);
  79 private:
  80     jit_uni_eltwise_injector_f32<avx512_common> *eltwise_injector_;
  81     using reg64_t = const Xbyak::Reg64;
  82     using zmm_t = const Xbyak::Zmm;
  83     using xmm_t = const Xbyak::Xmm;
  84
  85     reg64_t reg_src = r8;
  86     reg64_t reg_filt = r9;
  87     reg64_t reg_dst = r10;
  88     reg64_t param1 = abi_param1;
  89     reg64_t reg_kh = abi_not_param1;
  90     reg64_t reg_nur_w = rbx;
  91     reg64_t reg_bias = rdx;
  92     reg64_t reg_icb = reg_bias;
  93     reg64_t reg_ptr_scales = rax;
  94     reg64_t reg_oc_blocks = rsi;
  95
  96     reg64_t aux_reg_src = r11;
  97     reg64_t aux_reg_filt = r12;
  98
  99     reg64_t reg_compensation = r14;
 100     reg64_t reg_scratch = r14;
 101     reg64_t reg_ptr_sum_scale = r11;
 102     reg64_t reg_bias_alpha = abi_not_param1;
 103     reg64_t reg_overflow = rax;
 104     reg64_t reg_comp_strides = reg_overflow;
 105
 106     Xbyak::Opmask ktail_mask = Xbyak::Opmask(2);
 107     zmm_t zmm_tmp = zmm_t(28);
 108     zmm_t zmm_one = zmm_t(29);
 109     /* used during write-out section of store_output */
 110     zmm_t zmm_zero = zmm_t(31);
 111     zmm_t zmm_wei = zmm_t(31);
 112
 113     /* signed input */
 114     zmm_t zmm_shift = zmm_t(30);
 115     zmm_t zmm_comp = zmm_t(30);
 116     zmm_t zmm_bias = zmm_t(31);
 117     zmm_t zmm_prev_dst = zmm_t(31);
 118
 119     zmm_t zmm_out(int i_ur, int i_oc) {
 120         int idx = i_ur * jcp.nb_oc_blocking + i_oc;
 121         assert(idx < 31);
 122         return zmm_t(idx);
 123     }
 124     zmm_t zmm_inp(int i_ic, int nb_x_blocking) {
 125         int idx = i_ic + nb_x_blocking * jcp.ur_w;
 126         assert(idx < 31);
 127         return zmm_t(idx);
 128     }
 129     zmm_t zmm_bias_alpha() {
 130         return zmm_t(jcp.nb_oc_blocking * jcp.ur_w);
 131     }
 132     xmm_t xmm_bias_alpha() {
 133         return xmm_t(jcp.nb_oc_blocking * jcp.ur_w);
 134     }
 135
 136     int get_ow_start(int ki, int l_overflow) {
 137         int res = (jcp.ow - 1 + jcp.r_pad) % jcp.stride_w
 138                 + l_overflow * jcp.stride_w
 139                 - (jcp.kw - 1 - ki) * (jcp.dilate_w + 1);
 140         while (res < 0)
 141             res += jcp.stride_w;
 142         return res;
 143     }
 144
 145     int get_ow_end(int ur_w, int ki, int r_overflow) {
 146         if (utils::one_of(ur_w, jcp.ow, jcp.ur_w_tail))
 147                 ur_w += nstl::min(0, jcp.r_pad); // remove negative padding
 148         int res = (ur_w - 1 + jcp.l_pad) % jcp.stride_w
 149             + r_overflow * jcp.stride_w - ki * (jcp.dilate_w + 1);
 150         while (res < 0)
 151             res += jcp.stride_w;
 152         return ur_w - res;
 153     }
 154     bool maybe_eltwise(int position);
 155     void compute_eltwise(int ur_w);
 156     void prepare_output(int ur_w);
 157     void store_output(int ur_w, bool last_oc_block);
 158     void compute_ker(int ur_w, int l_overflow, int r_overflow,
 159              ker_block_t last_ic_block_flag, bool h_padded = false);
 160     void kh_loop(int ur_w, int pad_l, int pad_r, ker_block_t last_ker_block);
 161     void icb_loop(int ur_w, int pad_l, int pad_r, bool last_block);
 162     void generate();
 163     void cvt2ps(data_type_t type_in, zmm_t zmm_in, const Xbyak::Operand &op,
 164         bool mask_flag);
 165 };
 166
 167 template <impl::data_type_t src_type, impl::data_type_t dst_type>
 168 struct _jit_avx512_core_x8s8s32x_deconvolution_fwd_t : public cpu_primitive_t {
 169     struct pd_t : public cpu_deconvolution_fwd_pd_t {
 170         pd_t(engine_t *engine,
 171                 const deconvolution_desc_t *adesc,
 172                 const primitive_attr_t *attr,
 173                 const deconvolution_fwd_pd_t *hint_fwd_pd)
 174             : cpu_deconvolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}
 175
 176         DECLARE_COMMON_PD_T(JIT_IMPL_NAME_HELPER("jit_deconvolution:", avx512_core, ""),
 177                 _jit_avx512_core_x8s8s32x_deconvolution_fwd_t<src_type, dst_type>);
 178
 179         virtual status_t init() override {
 180             assert(this->engine()->kind() == engine_kind::cpu);
 181
 182             bool ok = true
 183                 && utils::one_of(this->desc()->prop_kind, prop_kind::forward_training,
 184                             prop_kind::forward_inference)
 185                 && this->desc()->alg_kind & alg_kind::deconvolution_direct
 186                 && this->desc()->src_desc.data_type == src_type
 187                 && this->desc()->dst_desc.data_type == dst_type
 188                 && IMPLICATION(this->with_bias(), utils::one_of(
 189                             this->desc()->bias_desc.data_type, data_type::f32,
 190                             data_type::s32, data_type::s8, data_type::u8))
 191                 && this->desc()->accum_data_type == data_type::s32;
 192             if (!ok) return status::unimplemented;
 193
 194             status_t status = jit_avx512_core_x8s8s32x_deconv_fwd_kernel::init_conf(
 195                     jcp_, *this->desc(), this->src_pd_,
 196                     this->weights_pd_, this->dst_pd_,
 197                     this->with_bias(), this->bias_pd_,
 198                     *this->attr());
 199
 200             if (status != status::success) return status;
 201
 202             auto scratchpad = scratchpad_registry().registrar();
 203             jit_avx512_core_x8s8s32x_deconv_fwd_kernel::init_scratchpad(scratchpad,
 204                     jcp_, *this->attr());
 205
 206             return status::success;
 207         }
 208         jit_conv_conf_t jcp_;
 209     };
 210
 211     _jit_avx512_core_x8s8s32x_deconvolution_fwd_t(const pd_t *apd,
 212            const input_vector &inputs, const output_vector &outputs)
 213        : cpu_primitive_t(apd, inputs, outputs) {
 214            kernel_ = new jit_avx512_core_x8s8s32x_deconv_fwd_kernel(pd()->jcp_,
 215                    *pd()->attr());
 216        }
 217
 218     ~_jit_avx512_core_x8s8s32x_deconvolution_fwd_t() {
 219         delete kernel_;
 220     }
 221
 222     typedef typename prec_traits<src_type>::type src_data_t;
 223     typedef typename prec_traits<data_type::s8>::type wei_data_t;
 224     typedef typename prec_traits<dst_type>::type dst_data_t;
 225
 226     virtual void execute(event_t *e) const
 227     {
 228         execute_forward();
 229         e->set_state(event_t::ready);
 230     }
 231
 232 private:
 233     void execute_forward() const;
 234     const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
 235     jit_avx512_core_x8s8s32x_deconv_fwd_kernel *kernel_;
 236 };
 237
 238
 239 }
 240 }
 241 }
 242
 243 #endif
 244
 245 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s