inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx512_core_bf16_conv_kernel.hpp

   1 /*******************************************************************************
   2 * Copyright 2019 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_BF16_FWD_KERNEL_HPP
  18 #define JIT_AVX512_BF16_FWD_KERNEL_HPP
  19
  20 #include "c_types_map.hpp"
  21 #include "memory_tracking.hpp"
  22
  23 #include "cpu_memory.hpp"
  24 #include "jit_generator.hpp"
  25 #include "jit_primitive_conf.hpp"
  26 #include "jit_uni_eltwise.hpp"
  27 #include "jit_avx512_core_bf16cvt.hpp"
  28
  29 //#define BF16_CONV_BWD_W_JIT_KER_USES_PERMW_TRANSPOSITION
  30 //#define BF16_CONV_BWD_W_DOES_NOT_USE_BARRIERS
  31
  32 namespace mkldnn {
  33 namespace impl {
  34 namespace cpu {
  35
  36 struct jit_avx512_core_bf16_fwd_kernel : public jit_generator {
  37
  38     jit_avx512_core_bf16_fwd_kernel(jit_conv_conf_t ajcp,
  39             const primitive_attr_t &attr) :
  40         jit_generator(nullptr, ker_code_size),
  41         jcp(ajcp),
  42         attr_(attr),
  43         eltwise_injector_(nullptr),
  44         bf16_emu_(nullptr)
  45     {
  46         if (jcp.with_eltwise)
  47             eltwise_injector_ = new jit_uni_eltwise_injector_f32<avx512_common>(
  48                     this, jcp.eltwise);
  49         if (!mayiuse(avx512_core_bf16))
  50             bf16_emu_ = new bf16_emulation_t(this,
  51                     bf16_emu_reserv_1, bf16_emu_reserv_2,
  52                     bf16_emu_reserv_3, bf16_emu_reserv_4,
  53                     bf16_emu_reserv_5, bf16_emu_reserv_6);
  54
  55         generate();
  56         jit_ker = (void (*)(jit_conv_call_s *))getCode();
  57     }
  58
  59     ~jit_avx512_core_bf16_fwd_kernel() {
  60         delete bf16_emu_;
  61         delete eltwise_injector_;
  62     }
  63
  64     DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx512_core_bf16_fwd_kernel)
  65
  66     static bool post_ops_ok(jit_conv_conf_t &jcp,
  67             const primitive_attr_t &attr);
  68     static status_t init_conf(jit_conv_conf_t &jcp,
  69             const convolution_desc_t &cd,
  70             cpu_memory_t::pd_t &src_pd,
  71             cpu_memory_t::pd_t &weights_pd,
  72             cpu_memory_t::pd_t &dst_pd,
  73             cpu_memory_t::pd_t &bias_pd,
  74             const primitive_attr_t &attr,
  75             int nthreads);
  76
  77     jit_conv_conf_t jcp;
  78     const primitive_attr_t &attr_;
  79     void (*jit_ker)(jit_conv_call_s *);
  80
  81 private:
  82     using reg64_t = const Xbyak::Reg64;
  83     enum {
  84         ker_reg_base_idx = 28,
  85         ker_code_size = 1024 * 1024,
  86     };
  87
  88     reg64_t param = abi_param1; //L: RDI, W: RCX
  89
  90     reg64_t reg_inp = r8;
  91     reg64_t reg_ker = r9;
  92     reg64_t reg_out = r10;
  93     reg64_t reg_owb = r11;
  94
  95     reg64_t aux_reg_inp = r12;
  96     reg64_t aux_reg_ker = r13;
  97
  98     reg64_t aux_reg_ker_d = r14;
  99     reg64_t aux_reg_inp_d = r15;
 100
 101     reg64_t reg_icb = rax;
 102     reg64_t reg_bias = rbx;
 103
 104     reg64_t reg_kj = abi_not_param1;
 105     reg64_t reg_oi = rdx;
 106     reg64_t reg_kh = rsi;
 107
 108
 109     reg64_t reg_out_long_offt = r14;
 110
 111     Xbyak::Zmm zmm_out(int i_ur, int i_oc) {
 112         int idx = i_ur + i_oc * jcp.ur_w;
 113         assert(idx < ker_reg_base_idx);
 114         return Xbyak::Zmm(idx);
 115     }
 116
 117     Xbyak::Ymm ymm_out(int i_ur, int i_oc) {
 118         int idx = i_ur + i_oc * jcp.ur_w;
 119         assert(idx < ker_reg_base_idx);
 120         return Xbyak::Ymm(idx);
 121     }
 122
 123     Xbyak::Zmm zmm_inp(int i_ic, int nb_x_blocking) {
 124         int idx = i_ic + nb_x_blocking * jcp.ur_w;
 125         assert(idx < 31);
 126         return Xbyak::Zmm(idx);
 127     }
 128
 129     Xbyak::Ymm ymm_inp(int i_ic, int nb_x_blocking) {
 130         int idx = i_ic + nb_x_blocking * jcp.ur_w;
 131         assert(idx < 31);
 132         return Xbyak::Ymm(idx);
 133     }
 134
 135     Xbyak::Reg64 imm_addr64 = r15;
 136     Xbyak::Xmm xmm_relu_ns = Xbyak::Xmm(30);
 137     Xbyak::Zmm zmm_relu_ns = Xbyak::Zmm(30);
 138     Xbyak::Zmm zmm_zero = Xbyak::Zmm(31);
 139     Xbyak::Zmm zmm_wei = Xbyak::Zmm(31);
 140     Xbyak::Zmm zmm_prev_dst = Xbyak::Zmm(31);
 141
 142     Xbyak::Zmm bf16_emu_reserv_1 = Xbyak::Zmm(26);
 143     Xbyak::Zmm bf16_emu_reserv_2 = Xbyak::Zmm(27);
 144     Xbyak::Zmm bf16_emu_reserv_3 = Xbyak::Zmm(28);
 145     reg64_t bf16_emu_reserv_4 = reg_icb;
 146     Xbyak::Zmm bf16_emu_reserv_5 = Xbyak::Zmm(29);
 147     Xbyak::Zmm bf16_emu_reserv_6 = Xbyak::Zmm(30);
 148
 149     jit_uni_eltwise_injector_f32<avx512_common> *eltwise_injector_;
 150     bf16_emulation_t *bf16_emu_;
 151
 152     inline void prepare_output(int ur_w);
 153     inline void store_output(int ur_w);
 154     inline void compute_loop(int ur_w, int pad_l, int pad_r);
 155
 156     void generate();
 157
 158     size_t get_output_offset(int oi, int n_oc_block) {
 159         return (size_t)jcp.typesize_out * ((size_t)n_oc_block * jcp.oh
 160             * jcp.ow * jcp.od + oi) * jcp.oc_block;
 161     }
 162
 163     size_t get_input_offset(int ki, int ic, int oi, int pad_l) {
 164         /*TODO: remove 4vnni */
 165         size_t scale = (jcp.ver == ver_4vnni || jcp.ver == ver_vnni) ? 2 : 1;
 166         size_t iw_str = jcp.ic_block;
 167         size_t ic_str = 1;
 168         return (size_t)jcp.typesize_in
 169                 * ((size_t)(ki * (jcp.dilate_w + 1) + oi * jcp.stride_w - pad_l)
 170                           * iw_str
 171                           + scale * ic * ic_str);
 172     }
 173
 174     size_t get_kernel_offset(int ki, int ic,
 175                                     int n_oc_block, int ker_number) {
 176         int scale = (jcp.ver == ver_4vnni || jcp.ver == ver_vnni) ? 2 : 1;
 177         size_t oc_block_stride = (size_t)jcp.nb_ic
 178                                * jcp.ic_block * jcp.kh * jcp.kw * jcp.kd;
 179         return jcp.typesize_in * jcp.oc_block
 180             * (n_oc_block * oc_block_stride
 181                     + (ic + ker_number) * scale + ki * jcp.ic_block);
 182     }
 183
 184     int get_ow_start(int ki, int pad_l) {
 185         return nstl::max(0,
 186                 utils::div_up(pad_l - ki * (jcp.dilate_w + 1), jcp.stride_w));
 187     }
 188
 189     int get_ow_end(int ur_w, int ki, int pad_r) {
 190         return ur_w - nstl::max(0, utils::div_up(pad_r
 191                                                    - (jcp.kw - 1 - ki)
 192                                                            * (jcp.dilate_w + 1),
 193                                            jcp.stride_w));
 194     }
 195 };
 196
 197 struct jit_avx512_core_bf16_bwd_data_kernel: public jit_generator {
 198
 199     jit_avx512_core_bf16_bwd_data_kernel(jit_conv_conf_t ajcp):
 200         jit_generator(nullptr, ker_code_size),
 201         jcp(ajcp), bf16_emu_(nullptr)
 202     {
 203         if (!mayiuse(avx512_core_bf16))
 204             bf16_emu_ = new bf16_emulation_t(this,
 205                     bf16_emu_reserv_1, bf16_emu_reserv_2,
 206                     bf16_emu_reserv_3, bf16_emu_reserv_4,
 207                     bf16_emu_reserv_5, bf16_emu_reserv_6);
 208         generate();
 209         jit_ker = (void (*)(jit_conv_call_s *))getCode();
 210     }
 211
 212     ~jit_avx512_core_bf16_bwd_data_kernel() {
 213         delete bf16_emu_;
 214     }
 215
 216     DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx512_core_bf16_bwd_data_kernel_f32)
 217
 218     static status_t init_conf(jit_conv_conf_t &jcp,
 219             const convolution_desc_t &cd,
 220             const memory_desc_wrapper &diff_src_d,
 221             const memory_desc_wrapper &weights_d,
 222             const memory_desc_wrapper &diff_dst_d);
 223
 224     jit_conv_conf_t jcp;
 225     void (*jit_ker)(jit_conv_call_s *);
 226
 227 private:
 228     using reg64_t = const Xbyak::Reg64;
 229     enum {
 230         ker_reg_base_idx = 31,
 231         ker_code_size = 1024 * 1024,
 232     };
 233
 234     reg64_t param = abi_param1;
 235     reg64_t reg_dst = r8;
 236     reg64_t reg_ker = r9;
 237     reg64_t reg_src = r10;
 238
 239     reg64_t aux_reg_dst = r14;
 240     reg64_t aux_reg_ker = r15;
 241
 242
 243     reg64_t reg_kj = rax;
 244     reg64_t reg_oi = rbx;
 245     reg64_t reg_kh = abi_not_param1;
 246
 247     reg64_t reg_ocb = r11;
 248
 249     Xbyak::Zmm zmm_inp(int i_ic) {
 250         int idx = i_ic + jcp.nb_ic_blocking * jcp.ur_w;
 251         assert(idx < ker_reg_base_idx);
 252         return Xbyak::Zmm(idx);
 253     }
 254
 255     Xbyak::Ymm ymm_inp(int i_ic) {
 256         int idx = i_ic + jcp.nb_ic_blocking * jcp.ur_w;
 257         assert(idx < ker_reg_base_idx);
 258         return Xbyak::Ymm(idx);
 259     }
 260
 261     Xbyak::Zmm zmm_out(int i_ur, int i_oc) {
 262         int idx = i_ur + i_oc * jcp.ur_w;
 263         assert(idx < ker_reg_base_idx);
 264         return Xbyak::Zmm(idx);
 265     }
 266
 267     Xbyak::Zmm bf16_emu_reserv_1 = Xbyak::Zmm(26);
 268     Xbyak::Zmm bf16_emu_reserv_2 = Xbyak::Zmm(27);
 269     Xbyak::Zmm bf16_emu_reserv_3 = Xbyak::Zmm(28);
 270     reg64_t bf16_emu_reserv_4 = reg_kj;
 271     Xbyak::Zmm bf16_emu_reserv_5 = Xbyak::Zmm(29);
 272     Xbyak::Zmm bf16_emu_reserv_6 = Xbyak::Zmm(30);
 273
 274     Xbyak::Zmm zmm_wei = Xbyak::Zmm(31);
 275     bf16_emulation_t *bf16_emu_;
 276
 277     inline void prepare_output(int ur_w);
 278     inline void store_output(int ur_w);
 279     inline void compute_loop(int ur_w, int l_overflow, int r_overflow);
 280     void generate();
 281
 282     int get_iw_start(int ki, int l_overflow)
 283     {
 284         int res = (jcp.iw - 1 + jcp.r_pad) % jcp.stride_w
 285                 + l_overflow * jcp.stride_w
 286                 - (jcp.kw - 1 - ki) * (jcp.dilate_w + 1);
 287         while (res < 0)
 288             res += jcp.stride_w;
 289
 290         return res;
 291     }
 292
 293     int get_iw_end(int ur_w, int ki, int r_overflow)
 294     {
 295         if (utils::one_of(ur_w, jcp.iw, jcp.ur_w_tail))
 296             ur_w += nstl::min(0, jcp.r_pad); // remove negative padding
 297         int res = (ur_w - 1 + jcp.l_pad) % jcp.stride_w
 298                 + r_overflow * jcp.stride_w - ki * (jcp.dilate_w + 1);
 299         while (res < 0)
 300             res += jcp.stride_w;
 301
 302         return ur_w - res;
 303     }
 304 };
 305
 306 struct jit_avx512_core_bf16_conv_bwd_weights_kernel_f32 : public jit_generator {
 307
 308     jit_avx512_core_bf16_conv_bwd_weights_kernel_f32(jit_conv_conf_t ajcp) :
 309         jit_generator(nullptr, ker_code_size),
 310         jcp(ajcp), bf16_emu_(nullptr)
 311     {
 312         if (!jcp.is_cpx) {
 313             bf16_emu_ = new bf16_emulation_t(this, one, even, selector, scratch,
 314                                         tmp0, tmp1);
 315         }
 316         generate();
 317         jit_ker = (void (*)(jit_conv_call_s *))getCode();
 318     }
 319
 320     ~jit_avx512_core_bf16_conv_bwd_weights_kernel_f32() {
 321         delete bf16_emu_;
 322     }
 323
 324     DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_avx512_core_bf16_conv_bwd_weights_kernel_f32)
 325
 326     static status_t init_conf(jit_conv_conf_t &jcp,
 327             const convolution_desc_t &cd, cpu_memory_t::pd_t &src_pd,
 328             cpu_memory_t::pd_t &diff_weights_pd,
 329             cpu_memory_t::pd_t &diff_bias_pd, cpu_memory_t::pd_t &diff_dst_pd);
 330     static void init_scratchpad(memory_tracking::registrar_t &scratchpad,
 331             const jit_conv_conf_t &jcp);
 332
 333     jit_conv_conf_t jcp;
 334     void (*jit_ker)(jit_conv_call_s *);
 335
 336 private:
 337     Xbyak::Label dst_prm_table;
 338
 339     using reg64_t = const Xbyak::Reg64;
 340     enum {
 341         typesize = sizeof(float),
 342         ker_code_size = 1024 * 1024,
 343     };
 344     static const int max_ur_w;
 345
 346     reg64_t param = abi_param1;
 347     reg64_t reg_input = rax;
 348     reg64_t reg_kernel = rdx;
 349     reg64_t reg_output = rsi;
 350     reg64_t b_ic = abi_not_param1;
 351     reg64_t kj = r8;
 352     reg64_t reg_kh = r9;
 353     reg64_t reg_ur_w_trips = r10;
 354     reg64_t reg_oj = r15;
 355     reg64_t reg_ih_count = rbx;
 356     reg64_t reg_tmp = r14;
 357     reg64_t reg_long_offt = r14;
 358
 359     reg64_t ki = r11;
 360     reg64_t reg_kd_count = r12;
 361     reg64_t reg_oi = r12;
 362     reg64_t reg_d_index = r13;
 363     reg64_t reg_input_d = r15;
 364     reg64_t reg_output_d = rbx;
 365     reg64_t aux_reg_input = r12;
 366     reg64_t aux_reg_kernel = r13;
 367     reg64_t reg_bias = rbx;
 368
 369     Xbyak::Zmm one = Xbyak::Zmm(27);
 370     Xbyak::Zmm even = Xbyak::Zmm(28);
 371     Xbyak::Zmm selector = Xbyak::Zmm(29);
 372     Xbyak::Zmm tmp0 = Xbyak::Zmm(30);
 373     Xbyak::Zmm tmp1 = Xbyak::Zmm(31);
 374     reg64_t scratch = r11;
 375
 376     inline void maybe_zero_kernel();
 377     inline void compute_oh_step_unroll_ow_icblock(int ic_block_step,
 378             int max_ur_w);
 379     inline void od_step_comeback_pointers();
 380     inline void oh_step_comeback_pointers();
 381     inline void compute_oh_step_unroll_ow(int ic_block_step, int max_ur_w);
 382     inline void compute_ic_block_step(int ur_w,
 383             int pad_l, int pad_r, int ic_block_step,
 384             int input_offset, int kernel_offset, int output_offset,
 385             bool is_tail = false);
 386     inline void compute_oh_step_common(int ic_block_step, int max_ur_w);
 387     inline void compute_oh_step_disp();
 388     inline void compute_loop();
 389
 390     void generate();
 391
 392     static void balance(const jit_conv_conf_t &j, int &nthr, int &nthr_mb,
 393             int &nthr_g, int &nthr_oc_b, int &nthr_ic_b);
 394
 395     bf16_emulation_t *bf16_emu_;
 396 #ifdef BF16_CONV_BWD_W_JIT_KER_USES_PERMW_TRANSPOSITION
 397     int stack_space_needed = 64;
 398 #endif
 399 };
 400
 401 }
 402 }
 403 }
 404 #endif