algos.cpp 23.4 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
/**
 * \file dnn/src/arm_common/conv_bias/int8x8x16/algos.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 */

#include "src/arm_common/conv_bias/int8x8x16/algos.h"
#include "src/arm_common/conv_bias/int8x8x16/conv_direct.h"
#include "src/arm_common/conv_bias/int8x8x16/conv_stride2.h"

#include "midout.h"
#include "src/common/opr_delegate.h"
MIDOUT_DECL(megdnn_arm_common_conv_bias_int8816_kimpl)

#include <atomic>
#include <cstring>
#include <mutex>

using namespace megdnn;
using namespace arm_common;

namespace {
bool need_dst_copy_str1(
        const megdnn::fallback::ConvolutionImpl::NCBKernSizeParam& param) {
    if (param.osz[0] % 1 != 0 || param.osz[1] % 8 != 0)
        return true;
    return false;
}
bool need_src_copy_str1(
        const megdnn::fallback::ConvBiasImpl::NCBKernSizeParam& param) {
    auto&& fm = param.filter_meta;

    if (fm.padding[0] != 0 || fm.padding[1] != 0)
        return true;

    return need_dst_copy_str1(param);
}
void get_rectified_size_str1(size_t IH, size_t IW, size_t OH, size_t OW,
                             size_t PH, size_t PW, size_t& IH2, size_t& IW2,
                             size_t& OH2, size_t& OW2) {
    OH2 = OH;
    OW2 = (OW + 7) & ~7;
    IH2 = OH2 + (IH - OH) + 2 * PH;
    IW2 = OW2 + (IW - OW) + 2 * PW;
}
bool need_dst_copy_str2(
        const megdnn::fallback::ConvBiasImpl::NCBKernSizeParam& param) {
    // If the size of output is not multiples of 8, we need to copy it.
    if (param.osz[0] % 8 != 0 || param.osz[1] % 8 != 0)
        return true;
    return false;
}
bool need_src_copy_str2(
        const megdnn::fallback::ConvBiasImpl::NCBKernSizeParam& param) {
    auto&& fm = param.filter_meta;
    // If padding is not zero, we need to copy to eliminate padding effect.
    if (fm.padding[0] != 0 || fm.padding[1] != 0)
        return true;

    return need_dst_copy_str2(param);
}
void get_rectified_size_str2(size_t IH, size_t IW, size_t OH, size_t OW,
                             size_t FH, size_t FW, size_t PH, size_t PW,
                             size_t& IH2, size_t& IW2, size_t& OH2,
                             size_t& OW2) {
    MEGDNN_MARK_USED_VAR(PH);
    MEGDNN_MARK_USED_VAR(PW);
    OH2 = (OH + 7) & ~7;
    OW2 = (OW + 7) & ~7;
    IH2 = 2 * OH2 + FH - 2;
    IW2 = 2 * OW2 + FW - 2;
    // Because stride is 2, sometimes IH/W == IH/W2 + 1
    // Do a max update to handle this case.
    IH2 = std::max(IH2, IH);
    IW2 = std::max(IW2, IW);
}
}  // namespace

/* ===================== direct algo ===================== */
bool ConvBiasImpl::AlgoI8x8x16Direct::usable(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
        AlgoSelectionStrategy algo_selection_strategy) const {
    MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 1, 0) {
        auto&& fm = param.filter_meta;
        auto FH = fm.spatial[0];
        bool aviliable =
                param.bias_mode == BiasMode::NO_BIAS &&
                param.nonlineMode == NonlineMode::IDENTITY &&
                fm.format == param::ConvBias::Format::NCHW && !fm.should_flip &&
                param.src_type.enumv() == DTypeEnum::Int8 &&
                param.filter_type.enumv() == DTypeEnum::Int8 &&
                param.dst_type.enumv() == DTypeEnum::Int16 &&
                fm.spatial_ndim == 2 && fm.dilation[0] == 1 &&
                fm.dilation[1] == 1 && fm.stride[0] == 1 && fm.stride[1] == 1 &&
                FH == fm.spatial[1] && (FH == 2 || FH == 3 || FH == 5);
        if (algo_selection_strategy ==
            ConvBiasImpl::AlgoSelectionStrategy::HEURISTIC) {
            bool large_group = param.filter_meta.group >= param.nr_threads;
            aviliable &= (large_group == m_large_group);
        }
        return aviliable;
    }
    MIDOUT_END();
    return false;
}
WorkspaceBundle ConvBiasImpl::AlgoI8x8x16Direct::get_bundle(
        const NCBKernSizeParam& param) const {
    auto&& fm = param.filter_meta;
    size_t nr_threads = param.nr_threads;
    size_t group = fm.group, batch = param.n;
    auto IC = fm.icpg, IH = param.isz[0], IW = param.isz[1];
    auto OH = param.osz[0], OW = param.osz[1];
    auto PH = fm.padding[0], PW = fm.padding[1];
    size_t OH2, OW2, IH2, IW2;
    get_rectified_size_str1(IH, IW, OH, OW, PH, PW, IH2, IW2, OH2, OW2);
    size_t part0 = 0u, part1 = 0u;
    if (need_src_copy_str1(param)) {
        part0 = m_large_group ? IC * IH2 * IW2 * sizeof(int8_t) * nr_threads
                              : IC * IH2 * IW2 * sizeof(int8_t) * group * batch;
    }
    if (need_dst_copy_str1(param)) {
        part1 = OH2 * OW2 * sizeof(int16_t) * nr_threads + 16;
    }
    return {nullptr, {part0, part1}};
}
size_t ConvBiasImpl::AlgoI8x8x16Direct::get_workspace(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
    MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 1, 1) {
        auto bundle = get_bundle(param);
        return bundle.total_size_in_bytes();
    }
    MIDOUT_END();
    return 0;
}
//! Process one input channel copy padding
void ConvBiasImpl::AlgoI8x8x16Direct::copy_padding_kern(
        WorkspaceBundle bundle, const ConvBiasImpl::NCBKernParam& kern_param,
        const ConvBiasImpl::NCBKernIndex& ncb_index,
        const CpuNDRange& workspace_ids) {
    size_t IH = kern_param.isz[0];
    size_t IW = kern_param.isz[1];
    size_t IC = kern_param.filter_meta.icpg;
    size_t OH = kern_param.osz[0];
    size_t OW = kern_param.osz[1];
    size_t PH = kern_param.filter_meta.padding[0];
    size_t PW = kern_param.filter_meta.padding[1];
    size_t GROUP = kern_param.filter_meta.group;
    size_t OH2, OW2, IH2, IW2;
    get_rectified_size_str1(IH, IW, OH, OW, PH, PW, IH2, IW2, OH2, OW2);
    bool need_src_copy_var = need_src_copy_str1(kern_param);
    size_t padding_group_size = IH2 * IW2 * IC;
    bundle.set(kern_param.workspace_ptr);

    //! Used for get the workspace offset
    size_t workspace_group_id = workspace_ids[0],
           workspace_batch_id = workspace_ids[1],
           channel_id = workspace_ids[2];
    size_t group_id = ncb_index.ndrange_id[0],
           batch_id = ncb_index.ndrange_id[1];
    const int8_t* sptr = kern_param.src<int8_t>(batch_id, group_id, channel_id);
    if (need_src_copy_var) {
        //! copy to sptr_base to eliminate padding effect
        int8_t* sptr_base = static_cast<int8_t*>(bundle.get(0)) +
                            workspace_group_id * padding_group_size +
                            workspace_batch_id * GROUP * padding_group_size +
                            channel_id * IH2 * IW2;
        std::memset(sptr_base, 0, sizeof(int8_t) * IH2 * IW2);
        rep(ih, IH) {
            std::memcpy(sptr_base + (ih + PH) * IW2 + PW, sptr + ih * IW,
                        sizeof(int8_t) * IW);
        }
    }
};
//! compute one output channel
void ConvBiasImpl::AlgoI8x8x16Direct::do_conv_kern(
        WorkspaceBundle bundle, const NCBKernParam& kern_param,
        const NCBKernIndex& ncb_index, const CpuNDRange& workspace_ids) {
    size_t OH = kern_param.osz[0];
    size_t OW = kern_param.osz[1];
    size_t IH = kern_param.isz[0];
    size_t IW = kern_param.isz[1];
    size_t FH = kern_param.filter_meta.spatial[0];
    size_t FW = kern_param.filter_meta.spatial[1];
    size_t IC = kern_param.filter_meta.icpg;
    size_t PH = kern_param.filter_meta.padding[0];
    size_t PW = kern_param.filter_meta.padding[1];
    size_t GROUP = kern_param.filter_meta.group;
    size_t OH2, OW2, IH2, IW2;
    get_rectified_size_str1(IH, IW, OH, OW, PH, PW, IH2, IW2, OH2, OW2);
    bool need_src_copy_var = need_src_copy_str1(kern_param);
    bool need_dst_copy_var = need_dst_copy_str1(kern_param);
    size_t padding_group_size = IH2 * IW2 * IC;
    //! Choose the compute kernel
    using Func =
            std::function<void(const int8_t*, const int8_t*, int16_t*, size_t,
                               size_t, size_t, size_t, size_t, size_t)>;
    Func fun_not_add_to_dst = nullptr, fun_add_to_dst = nullptr;
    if (FH == 2) {
        fun_not_add_to_dst =
                conv_bias::conv_direct_2x2_sc_int8_int8_int16<false>;
        fun_add_to_dst = conv_bias::conv_direct_2x2_sc_int8_int8_int16<true>;
    } else if (FH == 3) {
        fun_not_add_to_dst =
                conv_bias::conv_direct_3x3_sc_int8_int8_int16<false>;
        fun_add_to_dst = conv_bias::conv_direct_3x3_sc_int8_int8_int16<true>;
    } else if (FH == 5) {
        fun_not_add_to_dst =
                conv_bias::conv_direct_5x5_sc_int8_int8_int16<false>;
        fun_add_to_dst = conv_bias::conv_direct_5x5_sc_int8_int8_int16<true>;
    }

    bundle.set(kern_param.workspace_ptr);
    //! Used for get the workspace offset
    size_t workspace_group_id = workspace_ids[0],
           workspace_batch_id = workspace_ids[1], oc = workspace_ids[2];

    size_t group_id = ncb_index.ndrange_id[0],
           batch_id = ncb_index.ndrange_id[1];

    const int8_t* sptr = kern_param.src<dt_int8>(batch_id, group_id);
    const int8_t* filter =
            kern_param.filter<dt_int8>(group_id) + oc * FH * FW * IC;
    int16_t* dst = kern_param.dst<dt_int16>(batch_id, group_id, oc);
    if (need_src_copy_var) {
        sptr = static_cast<int8_t*>(bundle.get(0)) +
               workspace_group_id * padding_group_size +
               workspace_batch_id * GROUP * padding_group_size;
    }
    int16_t* dptr = nullptr;
    if (need_dst_copy_var) {
        dptr = static_cast<int16_t*>(bundle.get(1)) +
               ncb_index.thread_id * OH2 * OW2;
    } else {
        dptr = dst;
    }
    fun_not_add_to_dst(sptr, filter, dptr, IH2, IW2, OH2, OW2, 0, 0);
    for (size_t ic = 1; ic < IC; ++ic) {
        fun_add_to_dst(sptr + ic * IH2 * IW2, filter + ic * FH * FW, dptr, IH2,
                       IW2, OH2, OW2, 0, 0);
    }
    if (need_dst_copy_var) {
        rep(oh, OH) {
            std::memcpy(dst + oh * OW, dptr + oh * OW2, sizeof(int16_t) * OW);
        }
    }
}
SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoI8x8x16Direct::get_kimpls(
        const NCBKernSizeParam& param) const {
    auto fm = param.filter_meta;
    size_t N = param.n;
    size_t IC = param.filter_meta.icpg;
    size_t OC = param.filter_meta.ocpg;
    size_t group = fm.group;
    WorkspaceBundle wbundle = get_bundle(param);
    SmallVector<NCBKern> ret_kerns;
    if (m_large_group) {
        auto exec_one_group = [wbundle](const NCBKernParam& kern_param,
                                        const NCBKernIndex& ncb_index) {
            auto fm = kern_param.filter_meta;
            size_t IC = fm.icpg;
            size_t OC = fm.ocpg;
            WorkspaceBundle bundle = wbundle;
            for (size_t ic = 0; ic < IC; ic++) {
                copy_padding_kern(bundle, kern_param, ncb_index,
                                  {ncb_index.thread_id, 0, ic});
            }
            for (size_t oc = 0; oc < OC; oc++) {
                do_conv_kern(bundle, kern_param, ncb_index,
                             {ncb_index.thread_id, 0, oc});
            }
        };
        ret_kerns.push_back({exec_one_group, {group, N, 1_z}});
    } else {
        WorkspaceBundle bundle = wbundle;
        auto copy_padding = [bundle](const NCBKernParam& kern_param,
                                     const NCBKernIndex& ncb_index) {
            copy_padding_kern(bundle, kern_param, ncb_index,
                              ncb_index.ndrange_id);
        };
        ret_kerns.push_back({copy_padding, {group, N, IC}});
        auto do_conv = [bundle](const NCBKernParam& kern_param,
                                const NCBKernIndex& ncb_index) {
            do_conv_kern(bundle, kern_param, ncb_index, ncb_index.ndrange_id);
        };
        ret_kerns.push_back({do_conv, {group, N, OC}});
    }
    return ret_kerns;
}
SmallVector<ConvBiasImpl::NCBKern>
ConvBiasImpl::AlgoI8x8x16Direct::dispatch_kerns(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
    MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 1, 2) {
        return get_kimpls(param);
    }
    MIDOUT_END();
    return {};
}

/* ===================== stride-2 algo ===================== */
bool ConvBiasImpl::AlgoI8x8x16Stride2::usable(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
        AlgoSelectionStrategy algo_selection_strategy) const {
    MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 2, 0) {
        auto&& fm = param.filter_meta;
        auto FH = fm.spatial[0];
        bool aviliable = param.bias_mode == BiasMode::NO_BIAS &&
                         param.nonlineMode == NonlineMode::IDENTITY &&
                         fm.format == param::ConvBias::Format::NCHW &&
                         !fm.should_flip &&
                         param.src_type.enumv() == DTypeEnum::Int8 &&
                         param.filter_type.enumv() == DTypeEnum::Int8 &&
                         param.dst_type.enumv() == DTypeEnum::Int16 &&
                         fm.dilation[0] == 1 && fm.dilation[1] == 1 &&
                         fm.stride[0] == 2 && fm.stride[1] == 2 &&
                         FH == fm.spatial[1] && (FH == 2 || FH == 3 || FH == 5);
        if (algo_selection_strategy ==
            ConvBiasImpl::AlgoSelectionStrategy::HEURISTIC) {
            bool large_group = param.filter_meta.group >= param.nr_threads;
            aviliable &= (large_group == m_large_group);
        }
        return aviliable;
    }
    MIDOUT_END();
    return false;
}
WorkspaceBundle ConvBiasImpl::AlgoI8x8x16Stride2::get_bundle(
        const NCBKernSizeParam& param) const {
    auto&& fm = param.filter_meta;
    size_t nr_threads = param.nr_threads;
    size_t group = fm.group, batch = param.n;
    auto IC = fm.icpg, IH = param.isz[0], IW = param.isz[1];
    auto OH = param.osz[0], OW = param.osz[1];
    auto PH = fm.padding[0], PW = fm.padding[1];
    auto FH = fm.spatial[0], FW = fm.spatial[1];
    size_t OH2, OW2, IH2, IW2;
    get_rectified_size_str2(IH, IW, OH, OW, FH, FW, PH, PW, IH2, IW2, OH2, OW2);
    size_t part0 = 0u, part1 = 0u;
    if (need_src_copy_str2(param)) {
        part0 = m_large_group ? IC * IH2 * IW2 * sizeof(int8_t) * nr_threads
                              : IC * IH2 * IW2 * sizeof(int8_t) * group * batch;
    }
    if (need_dst_copy_str2(param)) {
        part1 = OH2 * OW2 * sizeof(int16_t) * nr_threads + 16;
    }
    return {nullptr, {part0, part1}};
}
size_t ConvBiasImpl::AlgoI8x8x16Stride2::get_workspace(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
    MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 2, 1) {
        auto bundle = get_bundle(param);
        return bundle.total_size_in_bytes();
    }
    MIDOUT_END();
    return 0;
}
//! Process one input channel copy padding
void ConvBiasImpl::AlgoI8x8x16Stride2::copy_padding_kern(
        WorkspaceBundle bundle, const ConvBiasImpl::NCBKernParam& kern_param,
        const ConvBiasImpl::NCBKernIndex& ncb_index,
        const CpuNDRange& workspace_ids) {
    size_t IH = kern_param.isz[0];
    size_t IW = kern_param.isz[1];
    size_t IC = kern_param.filter_meta.icpg;
    size_t OH = kern_param.osz[0];
    size_t OW = kern_param.osz[1];
    size_t PH = kern_param.filter_meta.padding[0];
    size_t PW = kern_param.filter_meta.padding[1];
    auto FH = kern_param.filter_meta.spatial[0],
         FW = kern_param.filter_meta.spatial[1];
    size_t GROUP = kern_param.filter_meta.group;
    size_t IH2, IW2, OH2, OW2;
    get_rectified_size_str2(IH, IW, OH, OW, FH, FW, PH, PW, IH2, IW2, OH2, OW2);
    bool need_src_copy_var = need_src_copy_str2(kern_param);
    size_t padding_group_size = IH2 * IW2 * IC;

    bundle.set(kern_param.workspace_ptr);
    size_t workspace_group_id = workspace_ids[0],
           workspace_batch_id = workspace_ids[1],
           channel_id = workspace_ids[2];
    size_t group_id = ncb_index.ndrange_id[0],
           batch_id = ncb_index.ndrange_id[1];
    const int8_t* sptr = kern_param.src<int8_t>(batch_id, group_id, channel_id);
    if (need_src_copy_var) {
        //! copy to sptr_base to eliminate padding effect
        int8_t* sptr_base = static_cast<int8_t*>(bundle.get(0)) +
                            workspace_group_id * padding_group_size +
                            workspace_batch_id * GROUP * padding_group_size +
                            channel_id * IH2 * IW2;
        std::memset(sptr_base, 0, sizeof(int8_t) * IH2 * IW2);
        rep(ih, IH) {
            std::memcpy(sptr_base + (ih + PH) * IW2 + PW, sptr + ih * IW,
                        sizeof(int8_t) * IW);
        }
    }
};
//! compute one output channel
void ConvBiasImpl::AlgoI8x8x16Stride2::do_conv_kern(
        WorkspaceBundle bundle, const NCBKernParam& kern_param,
        const NCBKernIndex& ncb_index, const CpuNDRange& workspace_ids) {
    size_t OH = kern_param.osz[0];
    size_t OW = kern_param.osz[1];
    size_t IH = kern_param.isz[0];
    size_t IW = kern_param.isz[1];
    size_t FH = kern_param.filter_meta.spatial[0];
    size_t FW = kern_param.filter_meta.spatial[1];
    size_t IC = kern_param.filter_meta.icpg;
    size_t PH = kern_param.filter_meta.padding[0];
    size_t PW = kern_param.filter_meta.padding[1];
    size_t GROUP = kern_param.filter_meta.group;
    size_t IH2, IW2, OH2, OW2;
    get_rectified_size_str2(IH, IW, OH, OW, FH, FW, PH, PW, IH2, IW2, OH2, OW2);
    bool need_src_copy_var = need_src_copy_str2(kern_param);
    bool need_dst_copy_var = need_dst_copy_str2(kern_param);
    size_t padding_group_size = IH2 * IW2 * IC;
    //! Choose the compute kernel
    using Func =
            std::function<void(const int8_t*, const int8_t*, int16_t*, size_t,
                               size_t, size_t, size_t, size_t, size_t)>;
    Func fun_not_add_to_dst = nullptr, fun_add_to_dst = nullptr;
    if (FH == 2) {
        fun_not_add_to_dst =
                conv_bias::conv_stride2_2x2_sc_int8_int8_int16<false>;
        fun_add_to_dst = conv_bias::conv_stride2_2x2_sc_int8_int8_int16<true>;
    } else if (FH == 3) {
        fun_not_add_to_dst =
                conv_bias::conv_stride2_3x3_sc_int8_int8_int16<false>;
        fun_add_to_dst = conv_bias::conv_stride2_3x3_sc_int8_int8_int16<true>;
    } else if (FH == 5) {
        fun_not_add_to_dst =
                conv_bias::conv_stride2_5x5_sc_int8_int8_int16<false>;
        fun_add_to_dst = conv_bias::conv_stride2_5x5_sc_int8_int8_int16<true>;
    }

    bundle.set(kern_param.workspace_ptr);
    //! Used for get the workspace offset
    size_t workspace_group_id = workspace_ids[0],
           workspace_batch_id = workspace_ids[1], oc = workspace_ids[2];
    size_t group_id = ncb_index.ndrange_id[0],
           batch_id = ncb_index.ndrange_id[1];
    const int8_t* sptr = kern_param.src<dt_int8>(batch_id, group_id);
    const int8_t* filter =
            kern_param.filter<dt_int8>(group_id) + oc * FH * FW * IC;
    int16_t* dst = kern_param.dst<dt_int16>(batch_id, group_id, oc);
    if (need_src_copy_var) {
        sptr = static_cast<int8_t*>(bundle.get(0)) +
               workspace_group_id * padding_group_size +
               workspace_batch_id * GROUP * padding_group_size;
    }
    int16_t* dptr = nullptr;
    if (need_dst_copy_var) {
        dptr = static_cast<int16_t*>(bundle.get(1)) +
               ncb_index.thread_id * OH2 * OW2;
    } else {
        dptr = dst;
    }
    fun_not_add_to_dst(sptr, filter, dptr, IH2, IW2, OH2, OW2, 0, 0);
    for (size_t ic = 1; ic < IC; ++ic) {
        fun_add_to_dst(sptr + ic * IH2 * IW2, filter + ic * FH * FW, dptr, IH2,
                       IW2, OH2, OW2, 0, 0);
    }
    if (need_dst_copy_var) {
        rep(oh, OH) {
            std::memcpy(dst + oh * OW, dptr + oh * OW2, sizeof(int16_t) * OW);
        }
    }
}
SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoI8x8x16Stride2::get_kimpls(
        const NCBKernSizeParam& param) const {
    auto fm = param.filter_meta;
    size_t N = param.n;
    size_t IC = param.filter_meta.icpg;
    size_t OC = param.filter_meta.ocpg;
    size_t group = fm.group;
    WorkspaceBundle wbundle = get_bundle(param);
    SmallVector<NCBKern> ret_kerns;
    if (m_large_group) {
        auto exec_one_group = [wbundle](const NCBKernParam& kern_param,
                                        const NCBKernIndex& ncb_index) {
            auto fm = kern_param.filter_meta;
            size_t IC = fm.icpg;
            size_t OC = fm.ocpg;
            WorkspaceBundle bundle = wbundle;
            for (size_t ic = 0; ic < IC; ic++) {
                copy_padding_kern(bundle, kern_param, ncb_index,
                                  {ncb_index.thread_id, 0, ic});
            }
            for (size_t oc = 0; oc < OC; oc++) {
                do_conv_kern(bundle, kern_param, ncb_index,
                             {ncb_index.thread_id, 0, oc});
            }
        };
        ret_kerns.push_back({exec_one_group, {group, N, 1_z}});
    } else {
        WorkspaceBundle bundle = wbundle;
        auto copy_padding = [bundle](const NCBKernParam& kern_param,
                                     const NCBKernIndex& ncb_index) {
            copy_padding_kern(bundle, kern_param, ncb_index,
                              ncb_index.ndrange_id);
        };
        ret_kerns.push_back({copy_padding, {group, N, IC}});
        auto do_conv = [bundle](const NCBKernParam& kern_param,
                                const NCBKernIndex& ncb_index) {
            do_conv_kern(bundle, kern_param, ncb_index, ncb_index.ndrange_id);
        };
        ret_kerns.push_back({do_conv, {group, N, OC}});
    }
    return ret_kerns;
}
SmallVector<ConvBiasImpl::NCBKern>
ConvBiasImpl::AlgoI8x8x16Stride2::dispatch_kerns(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
    MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 2, 2) {
        return get_kimpls(param);
    }
    MIDOUT_END();
    return {};
}
bool ConvBiasImpl::AlgoI8x8x16Stride2Filter2::usable(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
        AlgoSelectionStrategy /*algo_selection_strategy*/) const {
    MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 3, 0) {
        return param.bias_mode == BiasMode::NO_BIAS &&
               param.nonlineMode == NonlineMode::IDENTITY &&
               param.nr_threads == 1_z &&
               conv_bias::can_conv_int8x8x16_stride2_flt2(param);
    }
    MIDOUT_END();
    return false;
}

size_t ConvBiasImpl::AlgoI8x8x16Stride2Filter2::get_workspace(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
    MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 3, 1) {
        return conv_bias::get_workspace_in_bytes_conv_int8x8x16_stride2_flt2(
                param);
    }
    MIDOUT_END();
    return 0;
}

SmallVector<ConvBiasImpl::NCBKern>
ConvBiasImpl::AlgoI8x8x16Stride2Filter2::dispatch_kerns(
        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
    // return {conv_bias::conv_int8x8x16_stride2_flt2,true};
    auto kern = [](const NCBKernParam& param, const NCBKernIndex& ncb_index) {
        MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 3, 2) {
            auto ncb_param = param;
            ncb_param.src_ptr = param.src<void>(0, ncb_index.ndrange_id[0]);
            ncb_param.dst_ptr = param.dst<void>(0, ncb_index.ndrange_id[0]);
            ncb_param.filter_ptr = param.filter<void>(ncb_index.ndrange_id[0]);
            ncb_param.bias_ptr = param.bias<void>(0, ncb_index.ndrange_id[0]);
            conv_bias::conv_int8x8x16_stride2_flt2(ncb_param);
        }
        MIDOUT_END();
    };
    size_t group = param.filter_meta.group;
    return {{kern, {group, 1_z, 1_z}}};
}

// vim: syntax=cpp.doxygen