simple_rnn_int8.cpp 31.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709
/*******************************************************************************
* Copyright 2018 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/

#include <cstring>
#include <iostream>
#include <math.h>
#include <numeric>
#include <string>

#include "mkldnn.hpp"

// MSVC doesn't support collapse clause in omp parallel
#if defined(_MSC_VER) && !defined(__clang__) && !defined(__INTEL_COMPILER)
#define collapse(x)
#endif

using namespace mkldnn;

const int batch = 64;
const int src_seq_length_max = 25;
const int tgt_seq_length_max = 27;

const int feature_size = 1024;

const int enc_bidir_n_layers = 1;
const int enc_unidir_n_layers = 7;
const int dec_n_layers = 8;

const int lstm_n_gates = 4;
const int lstm_n_states = 2;
std::vector<int32_t> weighted_src_layer(batch *feature_size, 1);
std::vector<float> alignment_model(
        src_seq_length_max *batch *feature_size, 1.0f);
std::vector<float> alignments(src_seq_length_max *batch, 1.0f);
std::vector<float> exp_sums(batch, 1.0f);

const float onef = 1.0, zerof = 0.0;
const int onei = 1;

void compute_weighted_annotations(float *weighted_annotations,
        int src_seq_length_max, int batch, int feature_size,
        float *weights_annot, float *annotations) {
    // annotations(aka enc_dst_layer) is (t, n, 2c)
    // weights_annot is (2c, c)

    int num_weighted_annotations = src_seq_length_max * batch;
    // annotation[i] = GEMM(weights_annot, enc_dst_layer[i]);
    mkldnn_sgemm("N", "N", &feature_size, &num_weighted_annotations,
            &feature_size, &onef, weights_annot, &feature_size, annotations,
            &feature_size, &zerof, weighted_annotations, &feature_size);
}

void compute_sum_of_rows(int8_t *a, int rows, int cols, int32_t *a_reduced) {
#ifdef _OPENMP
#pragma omp parallel for
#endif
    for (int i = 0; i < cols; i++) {
        a_reduced[i] = 0;
        for (int j = 0; j < rows; j++) {
            a_reduced[i] += (int32_t)a[i * rows + j];
        }
    }
}

void compute_attention(float *context_vectors, int src_seq_length_max,
        int batch, int feature_size, int8_t *weights_src_layer,
        float weights_src_layer_scale, int32_t *compensation,
        uint8_t *dec_src_layer, float dec_src_layer_scale,
        float dec_src_layer_shift, uint8_t *annotations,
        float *weighted_annotations, float *weights_alignments) {
    // dst_iter : (n, c) matrix
    // src_layer: (n, c) matrix
    // weighted_annotations (t, n, c)

    // weights_yi is (c, c)
    // weights_ai is (c, 1)
    // tmp[i] is (n, c)
    // a[i] is (n, 1)
    // p is (n, 1)

    // first we precompute the weighted_dec_src_layer
    int8_t ao = 0;
    int8_t bo = 0;
    int32_t co = 0;
    mkldnn_gemm_s8u8s32("N", "N", "F", &feature_size, &batch, &feature_size,
            &onef, weights_src_layer, &feature_size, &ao, dec_src_layer,
            &feature_size, &bo, &zerof, weighted_src_layer.data(),
            &feature_size, &co);

    // then we compute the alignment model
    float *alignment_model_ptr = alignment_model.data();
#ifdef _OPENMP
#pragma omp parallel for collapse(2)
#endif
    for (int i = 0; i < src_seq_length_max; i++) {
        for (int j = 0; j < batch; j++) {
            for (int k = 0; k < feature_size; k++) {
                size_t tnc_offset
                        = i * batch * feature_size + j * feature_size + k;
                alignment_model_ptr[tnc_offset] = tanhf(
                        (float)(weighted_src_layer.data()[j * feature_size + k]
                                - dec_src_layer_shift * compensation[k])
                                / (dec_src_layer_scale
                                          * weights_src_layer_scale)
                        + weighted_annotations[tnc_offset]);
            }
        }
    }

    // gemv with alignments weights. the resulting alignments are in alignments
    int num_weighted_annotations = src_seq_length_max * batch;
    mkldnn_sgemm("N", "N", &onei, &num_weighted_annotations, &feature_size,
            &onef, weights_alignments, &onei, alignment_model_ptr,
            &feature_size, &zerof, alignments.data(), &onei);

// softmax on alignments. the resulting context weights are in alignments
#ifdef _OPENMP
#pragma omp parallel for
#endif
    for (int i = 0; i < batch; i++)
        exp_sums[i] = 0.0f;
#ifdef _OPENMP
#pragma omp parallel for collapse(2)
#endif
    for (int i = 0; i < src_seq_length_max; i++) {
        for (int j = 0; j < batch; j++) {
            alignments[i * batch + j] = expf(alignments[i * batch + j]);
            exp_sums[j] += alignments[i * batch + j];
        }
    }

#ifdef _OPENMP
#pragma omp parallel for collapse(2)
#endif
    for (int i = 0; i < src_seq_length_max; i++)
        for (int j = 0; j < batch; j++)
            alignments[i * batch + j] /= exp_sums[j];

// then we compute the context vectors
#ifdef _OPENMP
#pragma omp parallel for collapse(2)
#endif
    for (int i = 0; i < batch; i++)
        for (int j = 0; j < feature_size; j++)
            context_vectors[i * (feature_size + feature_size) + feature_size
                    + j]
                    = 0.0f;

#ifdef _OPENMP
#pragma omp parallel for collapse(3)
#endif
    for (int i = 0; i < batch; i++)
        for (int k = 0; k < src_seq_length_max; k++)
            for (int j = 0; j < feature_size; j++)
                context_vectors[i * (feature_size + feature_size) + feature_size
                        + j]
                        += alignments[k * batch + i]
                        * (((float)annotations[j
                                   + feature_size * (i + batch * k)]
                                   - dec_src_layer_shift)
                        / dec_src_layer_scale);
}

void copy_context(float *src_iter, int n_layers, int n_states, int batch,
        int feature_size) {
// we copy the context from the first layer to all other layers
#ifdef _OPENMP
#pragma omp parallel for collapse(3)
#endif
    for (int k = 1; k < n_layers; k++)
        for (int j = 0; j < batch; j++)
            for (int i = 0; i < feature_size; i++)
                src_iter[(k * n_states * batch + j)
                                * (feature_size + feature_size)
                        + i]
                        = src_iter[j * (feature_size + feature_size) + i];
}

void simple_net() {
    auto cpu_engine = engine(engine::cpu, 0);
    auto null_memory_ = null_memory(cpu_engine);

    /*
      GNMT low precicion example.
      Note, we do not implement connection yet.
      For the encoder we use:
      - one primitive for the bidirectional layer of the encoder
      - one primitive for all remaining unidirectional layers in the encoder
      For the decoder we use:
      - one primitive for the first iteration
      - one primitive for all subsequent iterations in the decoder. Note that
        in this example, this primitive computes the states in place.
      - the attention mechanism is implemented separately as there is no support
        for the context vectors in MKL-DNN yet
     */

    std::vector<primitive> weights_reorders;
    std::vector<primitive> encoder_net;
    std::vector<primitive> decoder_net;

    std::vector<float> net_src(batch * src_seq_length_max * feature_size, 0.1f);
    std::vector<float> net_dst(batch * tgt_seq_length_max * feature_size, 0.1f);

    /* Quantization factors for fp32 data */

    const float data_shift = 64.;
    const float data_scale = 63.;
    const int weights_scale_mask = 3; // 11 for last two dimensions of ldigo
    std::vector<float> weights_scales(lstm_n_gates * feature_size);
    /* assign halves of vector with arbitrary values */
    const int scales_half = lstm_n_gates * feature_size / 2;
    std::fill(
            weights_scales.begin(), weights_scales.begin() + scales_half, 30.f);
    std::fill(weights_scales.begin() + scales_half + 1, weights_scales.end(),
            65.5f);

    /* Encoder */

    memory::dims enc_bidir_src_layer_tz
            = { src_seq_length_max, batch, feature_size };
    memory::dims enc_bidir_weights_layer_tz = { enc_bidir_n_layers, 2,
        feature_size, lstm_n_gates, feature_size };
    memory::dims enc_bidir_weights_iter_tz = { enc_bidir_n_layers, 2,
        feature_size, lstm_n_gates, feature_size };
    memory::dims enc_bidir_bias_tz
            = { enc_bidir_n_layers, 2, lstm_n_gates, feature_size };
    memory::dims enc_bidir_dst_layer_tz
            = { src_seq_length_max, batch, 2 * feature_size };

    /* GNMT encoder: 1 bidirectional layer and 7 unidirectional layers */

    std::vector<float> user_enc_bidir_wei_layer(
            enc_bidir_n_layers * 2 * feature_size * lstm_n_gates * feature_size,
            0.3f);
    std::vector<float> user_enc_bidir_wei_iter(
            enc_bidir_n_layers * 2 * feature_size * lstm_n_gates * feature_size,
            0.2f);
    std::vector<float> user_enc_bidir_bias(
            enc_bidir_n_layers * 2 * lstm_n_gates * feature_size, 1.0f);

    /* Create the memory for user data */
    auto user_enc_bidir_src_layer_md = memory::desc({ enc_bidir_src_layer_tz },
            memory::data_type::f32, memory::format::tnc);

    auto user_enc_bidir_wei_layer_md
            = memory::desc({ enc_bidir_weights_layer_tz },
                    memory::data_type::f32, memory::format::ldigo);

    auto user_enc_bidir_wei_iter_md
            = memory::desc({ enc_bidir_weights_iter_tz },
                    memory::data_type::f32, memory::format::ldigo);

    auto user_enc_bidir_bias_md = memory::desc({ enc_bidir_bias_tz },
            memory::data_type::f32, memory::format::ldgo);

    auto user_enc_bidir_src_layer_memory = memory(
            { user_enc_bidir_src_layer_md, cpu_engine }, net_src.data());
    auto user_enc_bidir_wei_layer_memory
            = memory({ user_enc_bidir_wei_layer_md, cpu_engine },
                    user_enc_bidir_wei_layer.data());
    auto user_enc_bidir_wei_iter_memory
            = memory({ user_enc_bidir_wei_iter_md, cpu_engine },
                    user_enc_bidir_wei_iter.data());
    auto user_enc_bidir_bias_memory = memory(
            { user_enc_bidir_bias_md, cpu_engine }, user_enc_bidir_bias.data());

    /* Create memory descriptors for RNN data w/o specified layout */
    auto enc_bidir_src_layer_md = memory::desc({ enc_bidir_src_layer_tz },
            memory::data_type::u8, memory::format::any);

    auto enc_bidir_wei_layer_md = memory::desc({ enc_bidir_weights_layer_tz },
            memory::data_type::s8, memory::format::any);

    auto enc_bidir_wei_iter_md = memory::desc({ enc_bidir_weights_iter_tz },
            memory::data_type::s8, memory::format::any);

    auto enc_bidir_dst_layer_md = memory::desc({ enc_bidir_dst_layer_tz },
            memory::data_type::u8, memory::format::any);

    /* Create bidirectional RNN */
    rnn_cell::desc bi_cell(algorithm::vanilla_lstm);

    /* Check if int8 RNN is supported */
    try {
        rnn_forward::desc bi_layer_desc(prop_kind::forward_inference, bi_cell,
                rnn_direction::bidirectional_concat, enc_bidir_src_layer_md,
                zero_md(), enc_bidir_wei_layer_md, enc_bidir_wei_iter_md,
                user_enc_bidir_bias_md, enc_bidir_dst_layer_md, zero_md());
    } catch (error &e) {
        if (e.status == mkldnn_unimplemented) {
            std::cerr
                    << "Dependency on Intel(R) MKL version 2019u2 or newer is "
                       "required for int8 RNN"
                    << std::endl;
        }
        throw;
    }

    rnn_forward::desc bi_layer_desc(prop_kind::forward_inference, bi_cell,
            rnn_direction::bidirectional_concat, enc_bidir_src_layer_md,
            zero_md(), enc_bidir_wei_layer_md, enc_bidir_wei_iter_md,
            user_enc_bidir_bias_md, enc_bidir_dst_layer_md, zero_md());

    /* Define RNN attributes that store quantization parameters */
    primitive_attr attr;
    attr.set_int_output_round_mode(round_mode::round_nearest);
    attr.set_rnn_data_qparams(data_scale, data_shift);
    attr.set_rnn_weights_qparams(weights_scale_mask, weights_scales);

    auto enc_bidir_prim_desc
            = rnn_forward::primitive_desc(bi_layer_desc, attr, cpu_engine);

    /* Create memory primitives for input data and use reorders to quantize
     * values to int8
     * NOTE: same attributes are used when creating RNN primitive and reorders
     */
    auto enc_bidir_src_layer_memory
            = memory(enc_bidir_prim_desc.src_layer_primitive_desc());
    auto enc_bidir_src_layer_reorder_pd = reorder::primitive_desc(
            user_enc_bidir_src_layer_memory.get_primitive_desc(),
            enc_bidir_src_layer_memory.get_primitive_desc(), attr);
    encoder_net.push_back(reorder(enc_bidir_src_layer_reorder_pd,
            user_enc_bidir_src_layer_memory, enc_bidir_src_layer_memory));

    auto enc_bidir_wei_layer_memory
            = memory(enc_bidir_prim_desc.weights_layer_primitive_desc());
    auto enc_bidir_wei_layer_reorder_pd = reorder::primitive_desc(
            user_enc_bidir_wei_layer_memory.get_primitive_desc(),
            enc_bidir_wei_layer_memory.get_primitive_desc(), attr);
    weights_reorders.push_back(reorder(enc_bidir_wei_layer_reorder_pd,
            user_enc_bidir_wei_layer_memory, enc_bidir_wei_layer_memory));

    auto enc_bidir_wei_iter_memory
            = memory(enc_bidir_prim_desc.weights_iter_primitive_desc());
    auto enc_bidir_wei_iter_reorder_pd = reorder::primitive_desc(
            user_enc_bidir_wei_iter_memory.get_primitive_desc(),
            enc_bidir_wei_iter_memory.get_primitive_desc(), attr);
    weights_reorders.push_back(reorder(enc_bidir_wei_iter_reorder_pd,
            user_enc_bidir_wei_iter_memory, enc_bidir_wei_iter_memory));

    auto enc_bidir_dst_layer_memory
            = memory(enc_bidir_prim_desc.dst_layer_primitive_desc());

    encoder_net.push_back(
            rnn_forward(enc_bidir_prim_desc, enc_bidir_src_layer_memory,
                    null_memory_, enc_bidir_wei_layer_memory,
                    enc_bidir_wei_iter_memory, user_enc_bidir_bias_memory,
                    enc_bidir_dst_layer_memory, null_memory_, null_memory_));

    /* GNMT encoder: unidirectional layers */
    // First unidirectinal layer scales 2 * feature_size output of bidirectional
    // layer to feature_size output
    std::vector<float> user_enc_uni_first_wei_layer(
            1 * 1 * 2 * feature_size * lstm_n_gates * feature_size, 0.3f);
    std::vector<float> user_enc_uni_first_wei_iter(
            1 * 1 * feature_size * lstm_n_gates * feature_size, 0.2f);
    std::vector<float> user_enc_uni_first_bias(
            1 * 1 * lstm_n_gates * feature_size, 1.0f);

    memory::dims user_enc_uni_first_wei_layer_dims
            = { 1, 1, 2 * feature_size, lstm_n_gates, feature_size };
    memory::dims user_enc_uni_first_wei_iter_dims
            = { 1, 1, feature_size, lstm_n_gates, feature_size };
    memory::dims user_enc_uni_first_bias_dims
            = { 1, 1, lstm_n_gates, feature_size };
    memory::dims enc_uni_first_dst_layer_dims
            = { src_seq_length_max, batch, feature_size };

    auto user_enc_uni_first_wei_layer_md
            = memory::desc({ user_enc_uni_first_wei_layer_dims },
                    memory::data_type::f32, memory::format::ldigo);
    auto user_enc_uni_first_wei_iter_md
            = memory::desc({ user_enc_uni_first_wei_iter_dims },
                    memory::data_type::f32, memory::format::ldigo);
    auto user_enc_uni_first_bias_md
            = memory::desc({ user_enc_uni_first_bias_dims },
                    memory::data_type::f32, memory::format::ldgo);
    auto user_enc_uni_first_wei_layer_memory
            = memory({ user_enc_uni_first_wei_layer_md, cpu_engine },
                    user_enc_uni_first_wei_layer.data());
    auto user_enc_uni_first_wei_iter_memory
            = memory({ user_enc_uni_first_wei_iter_md, cpu_engine },
                    user_enc_uni_first_wei_iter.data());
    auto user_enc_uni_first_bias_memory
            = memory({ user_enc_uni_first_bias_md, cpu_engine },
                    user_enc_uni_first_bias.data());

    auto enc_uni_first_wei_layer_md
            = memory::desc({ user_enc_uni_first_wei_layer_dims },
                    memory::data_type::s8, memory::format::any);
    auto enc_uni_first_wei_iter_md
            = memory::desc({ user_enc_uni_first_wei_iter_dims },
                    memory::data_type::s8, memory::format::any);
    auto enc_uni_first_dst_layer_md
            = memory::desc({ enc_uni_first_dst_layer_dims },
                    memory::data_type::u8, memory::format::any);

    rnn_cell::desc enc_uni_first_cell(algorithm::vanilla_lstm);
    rnn_forward::desc enc_uni_first_layer_desc(prop_kind::forward_inference,
            enc_uni_first_cell, rnn_direction::unidirectional_left2right,
            enc_bidir_dst_layer_md, zero_md(), enc_uni_first_wei_layer_md,
            enc_uni_first_wei_iter_md, user_enc_uni_first_bias_md,
            enc_uni_first_dst_layer_md, zero_md());

    auto enc_uni_first_prim_desc = rnn_forward::primitive_desc(
            enc_uni_first_layer_desc, attr, cpu_engine);

    auto enc_uni_first_wei_layer_memory
            = memory(enc_uni_first_prim_desc.weights_layer_primitive_desc());
    auto enc_uni_first_wei_layer_reorder_pd = reorder::primitive_desc(
            user_enc_uni_first_wei_layer_memory.get_primitive_desc(),
            enc_uni_first_wei_layer_memory.get_primitive_desc(), attr);
    weights_reorders.push_back(reorder(enc_uni_first_wei_layer_reorder_pd,
            user_enc_uni_first_wei_layer_memory,
            enc_uni_first_wei_layer_memory));

    auto enc_uni_first_wei_iter_memory
            = memory(enc_uni_first_prim_desc.weights_iter_primitive_desc());
    auto enc_uni_first_wei_iter_reorder_pd = reorder::primitive_desc(
            user_enc_uni_first_wei_iter_memory.get_primitive_desc(),
            enc_uni_first_wei_iter_memory.get_primitive_desc(), attr);
    weights_reorders.push_back(reorder(enc_uni_first_wei_iter_reorder_pd,
            user_enc_uni_first_wei_iter_memory, enc_uni_first_wei_iter_memory));

    auto enc_uni_first_dst_layer_memory
            = memory(enc_uni_first_prim_desc.dst_layer_primitive_desc());

    encoder_net.push_back(rnn_forward(enc_uni_first_prim_desc,
            enc_bidir_dst_layer_memory, null_memory_,
            enc_uni_first_wei_layer_memory, enc_uni_first_wei_iter_memory,
            user_enc_uni_first_bias_memory, enc_uni_first_dst_layer_memory,
            null_memory_, null_memory_));

    /* Remainging unidirectional layers */
    std::vector<float> user_enc_uni_wei_layer((enc_unidir_n_layers - 1) * 1
                    * feature_size * lstm_n_gates * feature_size,
            0.3f);
    std::vector<float> user_enc_uni_wei_iter((enc_unidir_n_layers - 1) * 1
                    * feature_size * lstm_n_gates * feature_size,
            0.2f);
    std::vector<float> user_enc_uni_bias(
            (enc_unidir_n_layers - 1) * 1 * lstm_n_gates * feature_size, 1.0f);

    memory::dims user_enc_uni_wei_layer_dims = { (enc_unidir_n_layers - 1), 1,
        feature_size, lstm_n_gates, feature_size };
    memory::dims user_enc_uni_wei_iter_dims = { (enc_unidir_n_layers - 1), 1,
        feature_size, lstm_n_gates, feature_size };
    memory::dims user_enc_uni_bias_dims
            = { (enc_unidir_n_layers - 1), 1, lstm_n_gates, feature_size };
    memory::dims enc_dst_layer_dims
            = { src_seq_length_max, batch, feature_size };

    auto user_enc_uni_wei_layer_md
            = memory::desc({ user_enc_uni_wei_layer_dims },
                    memory::data_type::f32, memory::format::ldigo);
    auto user_enc_uni_wei_iter_md = memory::desc({ user_enc_uni_wei_iter_dims },
            memory::data_type::f32, memory::format::ldigo);
    auto user_enc_uni_bias_md = memory::desc({ user_enc_uni_bias_dims },
            memory::data_type::f32, memory::format::ldgo);

    auto user_enc_uni_wei_layer_memory
            = memory({ user_enc_uni_wei_layer_md, cpu_engine },
                    user_enc_uni_wei_layer.data());
    auto user_enc_uni_wei_iter_memory
            = memory({ user_enc_uni_wei_iter_md, cpu_engine },
                    user_enc_uni_wei_iter.data());
    auto user_enc_uni_bias_memory = memory(
            { user_enc_uni_bias_md, cpu_engine }, user_enc_uni_bias.data());

    auto enc_uni_wei_layer_md = memory::desc({ user_enc_uni_wei_layer_dims },
            memory::data_type::s8, memory::format::any);
    auto enc_uni_wei_iter_md = memory::desc({ user_enc_uni_wei_iter_dims },
            memory::data_type::s8, memory::format::any);
    auto enc_dst_layer_md = memory::desc({ enc_dst_layer_dims },
            memory::data_type::f32, memory::format::any);

    rnn_cell::desc enc_uni_cell(algorithm::vanilla_lstm);
    rnn_forward::desc enc_uni_layer_desc(prop_kind::forward_inference,
            enc_uni_cell, rnn_direction::unidirectional_left2right,
            enc_uni_first_dst_layer_md, zero_md(), enc_uni_wei_layer_md,
            enc_uni_wei_iter_md, user_enc_uni_bias_md, enc_dst_layer_md,
            zero_md());
    auto enc_uni_prim_desc
            = rnn_forward::primitive_desc(enc_uni_layer_desc, attr, cpu_engine);

    auto enc_uni_wei_layer_memory
            = memory(enc_uni_prim_desc.weights_layer_primitive_desc());
    auto enc_uni_wei_layer_reorder_pd = reorder::primitive_desc(
            user_enc_uni_wei_layer_memory.get_primitive_desc(),
            enc_uni_wei_layer_memory.get_primitive_desc(), attr);
    weights_reorders.push_back(reorder(enc_uni_wei_layer_reorder_pd,
            user_enc_uni_wei_layer_memory, enc_uni_wei_layer_memory));

    auto enc_uni_wei_iter_memory
            = memory(enc_uni_prim_desc.weights_iter_primitive_desc());
    auto enc_uni_wei_iter_reorder_pd = reorder::primitive_desc(
            user_enc_uni_wei_iter_memory.get_primitive_desc(),
            enc_uni_wei_iter_memory.get_primitive_desc(), attr);
    weights_reorders.push_back(reorder(enc_uni_wei_iter_reorder_pd,
            user_enc_uni_wei_iter_memory, enc_uni_wei_iter_memory));

    auto enc_dst_layer_memory
            = memory(enc_uni_prim_desc.dst_layer_primitive_desc());

    encoder_net.push_back(
            rnn_forward(enc_uni_prim_desc, enc_uni_first_dst_layer_memory,
                    null_memory_, enc_uni_wei_layer_memory,
                    enc_uni_wei_iter_memory, user_enc_uni_bias_memory,
                    enc_dst_layer_memory, null_memory_, null_memory_));

    /* Decoder with attention mechanism */
    std::vector<float> user_dec_wei_layer(
            dec_n_layers * 1 * feature_size * lstm_n_gates * feature_size,
            0.2f);
    std::vector<float> user_dec_wei_iter(dec_n_layers * 1
                    * (feature_size + feature_size) * lstm_n_gates
                    * feature_size,
            0.3f);
    std::vector<float> user_dec_bias(
            dec_n_layers * 1 * lstm_n_gates * feature_size, 1.0f);
    std::vector<int8_t> user_weights_attention_src_layer(
            feature_size * feature_size, 1);
    float weights_attention_scale = 127.;
    std::vector<float> user_weights_annotation(
            feature_size * feature_size, 1.0f);
    std::vector<float> user_weights_alignments(feature_size, 1.0f);
    // Buffer to store decoder output for all iterations
    std::vector<uint8_t> dec_dst(tgt_seq_length_max * batch * feature_size, 0);

    memory::dims user_dec_wei_layer_dims
            = { dec_n_layers, 1, feature_size, lstm_n_gates, feature_size };
    memory::dims user_dec_wei_iter_dims = { dec_n_layers, 1,
        feature_size + feature_size, lstm_n_gates, feature_size };
    memory::dims user_dec_bias_dims
            = { dec_n_layers, 1, lstm_n_gates, feature_size };
    memory::dims dec_src_layer_dims = { 1, batch, feature_size };
    memory::dims dec_dst_layer_dims = { 1, batch, feature_size };

    // We will use the same memory for dec_src_iter and dec_dst_iter
    // However, dec_src_iter has a context vector but not
    // dec_dst_iter.
    // To resolve this we will create one memory that holds the
    // context vector as well as the both the hidden and cell states.
    // For the dst_iter, we will use a view on this memory.
    // Note that the cell state will be padded by
    // feature_size values. However, we do not compute or
    // access those.
    memory::dims dec_dst_iter_dims = { dec_n_layers, 1, lstm_n_states, batch,
        feature_size + feature_size };
    memory::dims dec_dst_iter_noctx_dims
            = { dec_n_layers, 1, lstm_n_states, batch, feature_size };

    auto user_dec_wei_layer_md = memory::desc({ user_dec_wei_layer_dims },
            memory::data_type::f32, memory::format::ldigo);
    auto user_dec_wei_iter_md = memory::desc({ user_dec_wei_iter_dims },
            memory::data_type::f32, memory::format::ldigo);
    auto user_dec_bias_md = memory::desc({ user_dec_bias_dims },
            memory::data_type::f32, memory::format::ldgo);
    auto dec_src_layer_md = memory::desc(
            { dec_src_layer_dims }, memory::data_type::u8, memory::format::tnc);
    auto dec_dst_layer_md = memory::desc(
            { dec_dst_layer_dims }, memory::data_type::u8, memory::format::tnc);
    auto dec_dst_iter_md = memory::desc({ dec_dst_iter_dims },
            memory::data_type::f32, memory::format::ldsnc);

    auto user_dec_wei_layer_memory = memory(
            { user_dec_wei_layer_md, cpu_engine }, user_dec_wei_layer.data());
    auto user_dec_wei_iter_memory = memory(
            { user_dec_wei_iter_md, cpu_engine }, user_dec_wei_iter.data());
    auto user_dec_bias_memory
            = memory({ user_dec_bias_md, cpu_engine }, user_dec_bias.data());
    auto dec_src_layer_memory = memory({ dec_src_layer_md, cpu_engine });
    auto dec_dst_layer_memory
            = memory({ dec_dst_layer_md, cpu_engine }, dec_dst.data());

    /* Create memory descriptors for RNN data w/o specified layout */
    auto dec_wei_layer_md = memory::desc({ user_dec_wei_layer_dims },
            memory::data_type::s8, memory::format::any);
    auto dec_wei_iter_md = memory::desc({ user_dec_wei_iter_dims },
            memory::data_type::s8, memory::format::any);

    /* As mentioned above, we create a view without context out of the
     memory with context. */
    auto dec_dst_iter_memory = memory({ dec_dst_iter_md, cpu_engine });
    auto dec_dst_iter_noctx_md
            = view::primitive_desc(dec_dst_iter_memory.get_primitive_desc(),
                      dec_dst_iter_noctx_dims, { 0, 0, 0, 0, 0 })
                      .dst_primitive_desc()
                      .desc();

    rnn_cell::desc dec_cell(algorithm::vanilla_lstm);
    rnn_forward::desc dec_ctx_desc(prop_kind::forward_inference, dec_cell,
            rnn_direction::unidirectional_left2right, dec_src_layer_md,
            dec_dst_iter_md, dec_wei_layer_md, dec_wei_iter_md,
            user_dec_bias_md, dec_dst_layer_md, dec_dst_iter_noctx_md);
    auto dec_ctx_prim_desc
            = rnn_forward::primitive_desc(dec_ctx_desc, attr, cpu_engine);

    /* Create memory primitives for input data and use reorders to quantize
     * values to int8 */
    auto dec_wei_layer_memory
            = memory(dec_ctx_prim_desc.weights_layer_primitive_desc());
    auto dec_wei_layer_reorder_pd = reorder::primitive_desc(
            user_dec_wei_layer_memory.get_primitive_desc(),
            dec_wei_layer_memory.get_primitive_desc(), attr);
    weights_reorders.push_back(reorder(dec_wei_layer_reorder_pd,
            user_dec_wei_layer_memory, dec_wei_layer_memory));

    auto dec_wei_iter_memory
            = memory(dec_ctx_prim_desc.weights_iter_primitive_desc());
    auto dec_wei_iter_reorder_pd = reorder::primitive_desc(
            user_dec_wei_iter_memory.get_primitive_desc(),
            dec_wei_iter_memory.get_primitive_desc(), attr);
    weights_reorders.push_back(reorder(dec_wei_iter_reorder_pd,
            user_dec_wei_iter_memory, dec_wei_iter_memory));

    decoder_net.push_back(rnn_forward(dec_ctx_prim_desc, dec_src_layer_memory,
            dec_dst_iter_memory, dec_wei_layer_memory, dec_wei_iter_memory,
            user_dec_bias_memory, dec_dst_layer_memory, dec_dst_iter_memory,
            null_memory_));

    /* Allocating temporary buffers for attention mechanism */
    std::vector<float> weighted_annotations(
            src_seq_length_max * batch * feature_size, 1.0f);
    std::vector<int32_t> weights_attention_sum_rows(feature_size, 1);

    /*
       Execution
     */
    auto execute = [&]() {
        // reorder weights to MKLDNN internal representation
        stream(stream::kind::eager).submit(weights_reorders).wait();

        // run encoder (1 stream)
        stream(stream::kind::eager).submit(encoder_net).wait();

        // compute the weighted annotations once before the decoder
        compute_weighted_annotations(weighted_annotations.data(),
                src_seq_length_max, batch, feature_size,
                user_weights_annotation.data(),
                (float *)enc_dst_layer_memory.get_data_handle());
        // precompute compensation for s8u8s32 gemm in compute attention
        compute_sum_of_rows(user_weights_attention_src_layer.data(),
                feature_size, feature_size, weights_attention_sum_rows.data());

        // We initialise src_layer to the embedding of </s>, which
        // are assumed to be 0 here
        memset(dec_src_layer_memory.get_data_handle(), 0,
                dec_src_layer_memory.get_primitive_desc().get_size());
        // From now on, src points to the output of the last iteration

        for (int i = 0; i < tgt_seq_length_max; i++) {
            uint8_t *src_att_layer_handle
                    = (uint8_t *)dec_src_layer_memory.get_data_handle();
            float *src_att_iter_handle
                    = (float *)dec_dst_iter_memory.get_data_handle();

            // Compute attention context vector into the first layer src_iter
            compute_attention(src_att_iter_handle, src_seq_length_max, batch,
                    feature_size, user_weights_attention_src_layer.data(),
                    weights_attention_scale, weights_attention_sum_rows.data(),
                    src_att_layer_handle, data_scale, data_shift,
                    (uint8_t *)enc_bidir_dst_layer_memory.get_data_handle(),
                    weighted_annotations.data(),
                    user_weights_alignments.data());

            // copy the context vectors to all layers of src_iter
            copy_context(src_att_iter_handle, dec_n_layers, lstm_n_states,
                    batch, feature_size);

            // run the decoder iteration
            stream(stream::kind::eager).submit(decoder_net).wait();

            // Move the handle on the src/dst layer to the next iteration
            auto dst_layer_handle
                    = (uint8_t *)dec_dst_layer_memory.get_data_handle();
            dec_src_layer_memory.set_data_handle(dst_layer_handle);
            dec_dst_layer_memory.set_data_handle(
                    dst_layer_handle + batch * feature_size);
        }

    };

    execute();
}

int main(int argc, char **argv) {
    try {
        simple_net();
        std::cout << "ok\n";
    } catch (error &e) {
        std::cerr << "status: " << e.status << std::endl;
        std::cerr << "message: " << e.message << std::endl;
    }
    return 0;
}