opr_footprint.cpp 28.9 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
/**
 * \file src/plugin/impl/opr_footprint.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 "megbrain/plugin/opr_footprint.h"
#include "megbrain/opr/basic_arith.h"
#include "megbrain/opr/blas.h"
#include "megbrain/opr/dnn/convolution.h"
#include "megbrain/opr/dnn/images2neibs.h"
#include "megbrain/opr/dnn/local.h"
#include "megbrain/opr/dnn/lrn.h"
#include "megbrain/opr/dnn/pooling.h"
#include "megbrain/opr/imgproc.h"
#include "megbrain/opr/io.h"
#include "megbrain/opr/tensor_manip.h"
#if MGB_ENABLE_JSON
#include "megdnn/opr_param_json.h"
#endif

using namespace mgb;

namespace {

template <class T>
uint64_t opr_footprint_func(cg::OperatorNodeBase* opr);

// Elemwise
template <>
uint64_t opr_footprint_func<opr::Elemwise>(cg::OperatorNodeBase* opr) {
    return opr->output()[0]->shape().total_nr_elems() *
           (std::max<size_t>(opr->input().size(), 2) - 1);
}

// AddUpdate
template <>
uint64_t opr_footprint_func<opr::AddUpdate>(cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 2,
               "AddUpdate opr should have two inputs");
    auto&& out_shape = opr->output()[0]->shape();
    return out_shape.total_nr_elems() * 3;
}

template <class Conv>
uint64_t eval_conv_computation(const TensorShape& src_shape,
                               const TensorShape& filter_shape,
                               const TensorShape& dst_shape,
                               cg::OperatorNodeBase* opr) {
    using Param = opr::ConvolutionForward::Param;
    auto&& param = opr->cast_final_safe<Conv>().param();

    if (param.format == Param::Format::NHWCD4) {
        size_t fh, fw;
        size_t group = 1;
        if (param.sparse == Param::Sparse::DENSE) {
            fh = filter_shape[1];
            fw = filter_shape[2];
            group = 1;
        } else {
            // chanwise conv
            mgb_assert(param.sparse == Param::Sparse::GROUP);
            fh = filter_shape[2];
            fw = filter_shape[3];
            group = filter_shape[0];

            if (filter_shape.ndim == 5) {
                group *= 4;
            }
        }
        return dst_shape.total_nr_elems() * fh * fw *
            src_shape[2] * 4 / group * 2;
    }
    auto eval_conv_computation_nchwx = [&param, &src_shape, &filter_shape,
                                        &dst_shape]() -> uint64_t {
        size_t fh, fw;
        bool hybird_nchwx = false;
        size_t group = 1;
        if (param.sparse == Param::Sparse::DENSE) {
            //! if nchwxx mode src is nchw output is nchwxx
            if (dst_shape.ndim == 5 && src_shape.ndim == 4) {
                fh = filter_shape[1];
                fw = filter_shape[2];
                hybird_nchwx = true;
            } else {
                fh = filter_shape[2];
                fw = filter_shape[3];
            }
            group = 1;
        } else {
            mgb_assert(param.sparse == Param::Sparse::GROUP);
            fh = filter_shape[3];
            fw = filter_shape[4];
            group = filter_shape[0];
        }
        if (param.format == Param::Format::NCHW88) {
102
            //! if channel wise weight layout is {group/8, FH, FW, 1, 1, 8}
103 104 105 106 107 108 109
            if (filter_shape[1] == 1 && filter_shape[2] == 1) {
                group *= 8;
            }
            size_t computation = dst_shape.total_nr_elems() * fh * fw *
                                 src_shape[1] / group * 2;
            return hybird_nchwx ? computation : computation * 8;
        }
110 111 112 113 114 115 116 117 118
        if (param.format == Param::Format::NCHW44) {
            //! if channel wise weight layout is {group/4, FH, FW, 1, 1, 4}
            if (filter_shape[1] == 1 && filter_shape[2] == 1) {
                group *= 4;
            }
            size_t computation = dst_shape.total_nr_elems() * fh * fw *
                                 src_shape[1] / group * 2;
            return hybird_nchwx ? computation : computation * 4;
        }
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
        if (param.format == Param::Format::NCHW32) {
            return dst_shape.total_nr_elems() * fh * fw * src_shape[1] * 32 /
                   group * 2;
        }
        mgb_assert(param.format == Param::Format::NCHW4,
                   "format should be NCHW4/NCHW32");
        return dst_shape.total_nr_elems() * fh * fw * src_shape[1] * 4 / group *
               2;
    };
    auto eval_conv_computation_chwn4 = [&param, &src_shape, &filter_shape,
                                        &dst_shape]() -> uint64_t {
        size_t fh, fw;
        size_t group = 1;
        if (param.sparse == Param::Sparse::DENSE) {
            fh = filter_shape[1];
            fw = filter_shape[2];
            group = 1;
        } else {
            mgb_assert(param.sparse == Param::Sparse::GROUP);
            fh = filter_shape[2];
            fw = filter_shape[3];
            group = filter_shape[0];
        }
        return dst_shape.total_nr_elems() * fh * fw * src_shape[0] * 4 / group *
               2;
    };
    if (param.format == Param::Format::NCHW4 ||
        param.format == Param::Format::NCHW88 ||
147
        param.format == Param::Format::NCHW44 ||
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
        param.format == Param::Format::NCHW32) {
        return eval_conv_computation_nchwx();
    }
    if (param.format == Param::Format::CHWN4) {
        return eval_conv_computation_chwn4();
    }
    size_t cpos;
    size_t spatial_start;
    size_t group = 1;
    switch (param.format) {
        case Param::Format::NCHW:
            cpos = 1;
            spatial_start = 2;
            break;
        case Param::Format::NCHW_WINOGRAD:
        case Param::Format::NCHW88_WINOGRAD:
            cpos = 1;
            spatial_start = 0;
            break;
        case Param::Format::NHWC:
            cpos = 3;
            spatial_start = 1;
            break;
        default:
            mgb_assert(false, "Unknown CONV Param::Format type");
    }
    switch (param.sparse) {
        case Param::Sparse::DENSE:
            mgb_assert(filter_shape.ndim == 4 || filter_shape.ndim == 6,
                       "DENSE conv filter shape dimension should be "
                       "4/6(winograd mk4)");
            break;
        case Param::Sparse::GROUP:
            mgb_assert(filter_shape.ndim == 5 || filter_shape.ndim == 7,
                       "GROUP conv filter shape dimension should be "
                       "5/7(winograd mk4)");
            spatial_start++;
            group = filter_shape[0];
            break;
        default:
            mgb_assert(false, "Unkown CONV Param::Sparse type");
    }

    uint64_t fh = static_cast<uint64_t>(filter_shape[spatial_start]);
    uint64_t fw = static_cast<uint64_t>(filter_shape[spatial_start + 1]);
    if (param.format == Param::Format::NCHW_WINOGRAD ||
        param.format == Param::Format::NCHW88_WINOGRAD) {
        mgb_assert(opr->same_type<opr::ConvBias>(),
                   "Only conv bias support NCHW_WINOGRAD");
        auto&& conv_bias_opr = opr->cast_final_safe<opr::ConvBias>();
        uint32_t output_block_size = conv_bias_opr.param().output_block_size;
        mgb_assert(fh == fw,
                   "NCHW_WINOGRAD, NCHW88_WINOGRAD need fw==fh, got fw: %u fh "
                   "%u\n",
                   static_cast<uint32_t>(fh), static_cast<uint32_t>(fw));
        fh = fh + 1 - output_block_size;
        fw = fw + 1 - output_block_size;
    }
    // mul and add are counted as 2 operations
    if(param.format == Param::Format::NCHW88_WINOGRAD){
        return dst_shape.total_nr_elems() * fh * fw *
               static_cast<uint64_t>(src_shape[cpos] * 8) / group * 2;
    }
    return dst_shape.total_nr_elems() * fh * fw *
           static_cast<uint64_t>(src_shape[cpos]) / group * 2;
}

// ConvolutionForward
template <>
uint64_t opr_footprint_func<opr::ConvolutionForward>(
        cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 2,
               "ConvolutionFwd opr should have two inputs");
    auto&& out_shape = opr->output()[0]->shape();
    auto&& src_shape = opr->input()[0]->shape();
    auto&& filter_shape = opr->input()[1]->shape();
    return eval_conv_computation<opr::ConvolutionForward>(
            src_shape, filter_shape, out_shape, opr);
}
template <>
uint64_t opr_footprint_func<opr::ConvBiasForward>(
        cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 2 || opr->input().size() == 3 ||
                       opr->input().size() == 4,
               "ConvBiasForward opr should have two/three/four inputs");
    auto&& out_shape = opr->output()[0]->shape();
    auto&& src_shape = opr->input()[0]->shape();
    auto&& filter_shape = opr->input()[1]->shape();
    uint64_t res = eval_conv_computation<opr::ConvBiasForward>(
            src_shape, filter_shape, out_shape, opr);
    if (opr->input().size() == 3) {
        res += out_shape.total_nr_elems();
    }
    return res;
}

// ConvolutionBackwardData
template <>
uint64_t opr_footprint_func<opr::ConvolutionBackwardData>(
        cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 2 || opr->input().size() == 3,
               "ConvolutionBackwardData opr should have two or three inputs");
    auto&& filter_shape = opr->input()[0]->shape();
    auto&& diff_shape = opr->input()[1]->shape();
    auto&& grad_shape = opr->output()[0]->shape();
    return eval_conv_computation<opr::ConvolutionBackwardData>(
            grad_shape, filter_shape, diff_shape, opr);
}

// ConvolutionBackwardFilter
template <>
uint64_t opr_footprint_func<opr::ConvolutionBackwardFilter>(
        cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 3,
               "ConvolutionBackwardData opr should have three inputs");
    auto&& filter_shape = opr->input()[2]->shape();
    auto&& diff_shape = opr->input()[1]->shape();
    auto&& src_shape = opr->input()[0]->shape();
    return eval_conv_computation<opr::ConvolutionBackwardFilter>(
            src_shape, filter_shape, diff_shape, opr);
}

// MatrixMul
template <>
uint64_t opr_footprint_func<opr::MatrixMul>(cg::OperatorNodeBase* opr) {
    auto&& mopr = opr->cast_final_safe<opr::MatrixMul>();
    auto &&i0 = opr->input(0)->shape(), &&i1 = opr->input(1)->shape();
    mgb_assert(i0.ndim == 2 && i1.ndim == 2);
    auto m = i0[0], k0 = i0[1], k1 = i1[0], n = i1[1];
    if (mopr.param().transposeA) {
        std::swap(m, k0);
    }
    if (mopr.param().transposeB) {
        std::swap(k1, n);
    }
    mgb_assert(k0 == k1);
    // mul and add are counted as 2 operations
    return m * k0 * n * 2;
}

template <>
uint64_t opr_footprint_func<opr::LocalShareForward>(cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 2,
               "LocalShare opr should have two inputs");
    auto&& out_shape = opr->output()[0]->shape();
    auto&& src_shape = opr->input()[0]->shape();
    auto&& filter_shape = opr->input()[1]->shape();
    using Param = opr::LocalShareForward::Param;
    auto&& param = opr->cast_final_safe<opr::LocalShareForward>().param();
    mgb_assert(param.format == Param::Format::NCHW);
    size_t groups = 1;
    size_t kern_spatial_pos = 3;
    if (param.sparse == Param::Sparse::GROUP) {
        groups = filter_shape[0];
        kern_spatial_pos = 4;
    }
    size_t fh = filter_shape[kern_spatial_pos],
           fw = filter_shape[kern_spatial_pos + 1];
    return out_shape.total_nr_elems() * fh * fw * src_shape[1] * 2 / groups;
}

template <>
uint64_t opr_footprint_func<opr::LocalShareBackwardData>(cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 3,
               "LocalShareBackwardData opr should have three inputs");
    auto&& filter_shape = opr->input()[0]->shape();
    auto&& diff_shape = opr->input()[1]->shape();
    auto&& grad_shape = opr->output()[0]->shape();
    using Param = opr::LocalShareForward::Param;
    auto&& param = opr->cast_final_safe<opr::LocalShareBackwardData>().param();
    mgb_assert(param.format == Param::Format::NCHW);
    size_t groups = 1;
    size_t kern_spatial_pos = 3;
    if (param.sparse == Param::Sparse::GROUP) {
        groups = filter_shape[0];
        kern_spatial_pos = 4;
    }
    size_t fh = filter_shape[kern_spatial_pos],
           fw = filter_shape[kern_spatial_pos + 1];
    return diff_shape.total_nr_elems() * fh * fw * grad_shape[1] * 2 / groups;
}

template <>
uint64_t opr_footprint_func<opr::LocalShareBackwardFilter>(cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 3,
               "LocalShareBackwardFilter opr should have three inputs");
    auto&& src_shape = opr->input()[0]->shape();
    auto&& diff_shape = opr->input()[1]->shape();
    auto&& grad_shape = opr->output()[0]->shape();
    using Param = opr::LocalShareForward::Param;
    auto&& param = opr->cast_final_safe<opr::LocalShareBackwardFilter>().param();
    mgb_assert(param.format == Param::Format::NCHW);
    size_t groups = 1;
    size_t kern_spatial_pos = 3;
    if (param.sparse == Param::Sparse::GROUP) {
        groups = grad_shape[0];
        kern_spatial_pos = 4;
    }
    size_t fh = grad_shape[kern_spatial_pos],
           fw = grad_shape[kern_spatial_pos + 1];
    return diff_shape.total_nr_elems() * fh * fw * src_shape[1] * 2 / groups;
}

template <>
uint64_t opr_footprint_func<opr::DeformableConvForward>(
        cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 4,
               "DeformableConvForward opr should have four inputs");
    auto&& out_shape = opr->output()[0]->shape();
    auto&& filter_shape = opr->input()[1]->shape();
    using Param = opr::DeformableConvForward::Param;
    auto&& param = opr->cast_final_safe<opr::Convolution>().param();
    size_t fh, fw, icpg;
    mgb_assert(param.format == Param::Format::NCHW);
    if (param.sparse == Param::Sparse::GROUP) {
        icpg = filter_shape[2];
        fh = filter_shape[3], fw = filter_shape[4];
    } else {
        icpg = filter_shape[1];
        fh = filter_shape[2], fw = filter_shape[3];
    }
    //! conv(1 mul), mask(1, mul), accumulate(1 add)
    return out_shape.total_nr_elems() * fh * fw * icpg * 3;
}

template <>
uint64_t opr_footprint_func<opr::DeformableConvBackwardFilter>(
        cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 5,
               "DeformableConvBackwardFilter opr should have four inputs");
    auto&& out_shape = opr->output()[0]->shape();
    auto&& filter_shape = opr->input()[1]->shape();
    using Param = opr::DeformableConvBackwardFilter::Param;
    auto&& param = opr->cast_final_safe<opr::Convolution>().param();
    size_t fh, fw, icpg;
    mgb_assert(param.format == Param::Format::NCHW);
    if (param.sparse == Param::Sparse::GROUP) {
        icpg = filter_shape[2];
        fh = filter_shape[3], fw = filter_shape[4];
    } else {
        icpg = filter_shape[1];
        fh = filter_shape[2], fw = filter_shape[3];
    }
    //! deconv(1 mul), mask(1 mul), accumulate(1 add), bilinear(4 add, 4mul,
    //! skip)
    return out_shape.total_nr_elems() * fh * fw * icpg * 3;
}

template <>
uint64_t opr_footprint_func<opr::DeformableConvBackwardData>(
        cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 5,
               "DeformableConvBackwardData opr should have four inputs");
    auto&& out_shape = opr->output()[0]->shape();
    auto&& filter_shape = opr->input()[1]->shape();
    using Param = opr::DeformableConvForward::Param;
    auto&& param = opr->cast_final_safe<opr::Convolution>().param();
    size_t fh, fw, icpg;
    mgb_assert(param.format == Param::Format::NCHW);
    if (param.sparse == Param::Sparse::GROUP) {
        icpg = filter_shape[2];
        fh = filter_shape[3], fw = filter_shape[4];
    } else {
        icpg = filter_shape[1];
        fh = filter_shape[2], fw = filter_shape[3];
    }
    //! deconv(1 mul), mask(1 mul), accumulate(1 add), grad_weight(1 mul, skip),
    //! grad_coord(4mul, 4 add)
    return out_shape.total_nr_elems() * fh * fw * icpg * 12;
}

template <>
uint64_t opr_footprint_func<opr::BatchConvBiasForward>(
        cg::OperatorNodeBase* opr) {
    mgb_assert(opr->input().size() == 2 || opr->input().size() == 3 ||
                       opr->input().size() == 4,
               "BatchConvBias opr should have two/three/four inputs");
    auto&& out_shape = opr->output()[0]->shape();
    auto&& src_shape = opr->input()[0]->shape();
    auto&& filter_shape = opr->input()[1]->shape();
    using Param = opr::BatchConvBiasForward::Param;
    auto&& param = opr->cast_final_safe<opr::BatchConvBiasForward>().param();
    mgb_assert(param.format == Param::Format::NCHW4);
    size_t packed_channels = 4;
    size_t kern_spatial_pos = 3;
    size_t fh = filter_shape[kern_spatial_pos],
           fw = filter_shape[kern_spatial_pos + 1];
    return out_shape.total_nr_elems() * fh * fw * src_shape[1] *
           packed_channels * 2;
}

// Pooling
template <>
uint64_t opr_footprint_func<opr::PoolingForward>(cg::OperatorNodeBase* opr) {
    auto&& param = opr->cast_final_safe<opr::PoolingForward>().param();
    auto area = param.window_h * param.window_w;
    return opr->output(0)->shape().total_nr_elems() * area;
}

// Concat
template <>
uint64_t opr_footprint_func<opr::Concat>(cg::OperatorNodeBase* opr) {
    auto&& out_shape = opr->output()[0]->shape();
    return out_shape.total_nr_elems();
}

// Dimshuffle
template <>
uint64_t opr_footprint_func<opr::Dimshuffle>(cg::OperatorNodeBase* opr) {
    auto&& out = opr->output()[0];
    return out->shape().total_nr_elems();
}

// Reduce
template <>
uint64_t opr_footprint_func<opr::Reduce>(cg::OperatorNodeBase* opr) {
    return opr->input()[0]->shape().total_nr_elems();
}

// Host2DeviceCopy
template <>
uint64_t opr_footprint_func<opr::Host2DeviceCopy>(cg::OperatorNodeBase* opr) {
    auto&& out_shape = opr->output()[0]->shape();
    return out_shape.total_nr_elems();
}

/******************* Registe Param Json Functions *************************/
#if MGB_ENABLE_JSON
template <class T>
std::shared_ptr<json::Value> opr_param_json_func(cg::OperatorNodeBase* opr);

#define REGISTE_PARAM_JSON_FUNC(cls)                            \
    template <>                                                 \
    std::shared_ptr<json::Value> opr_param_json_func<opr::cls>( \
            cg::OperatorNodeBase * opr) {                       \
        return opr::opr_param_to_json(                          \
                opr->cast_final_safe<opr::cls>().param());      \
    }

REGISTE_PARAM_JSON_FUNC(Elemwise)
REGISTE_PARAM_JSON_FUNC(ConvolutionForward)
REGISTE_PARAM_JSON_FUNC(Convolution3D)
REGISTE_PARAM_JSON_FUNC(ConvBiasForward)
REGISTE_PARAM_JSON_FUNC(ConvolutionBackwardData)
REGISTE_PARAM_JSON_FUNC(Convolution3DBackwardData)
REGISTE_PARAM_JSON_FUNC(ConvolutionBackwardFilter)
REGISTE_PARAM_JSON_FUNC(MatrixMul)
REGISTE_PARAM_JSON_FUNC(BatchedMatrixMul)
REGISTE_PARAM_JSON_FUNC(Dot)
REGISTE_PARAM_JSON_FUNC(MatrixInverse)
REGISTE_PARAM_JSON_FUNC(PoolingForward)
REGISTE_PARAM_JSON_FUNC(SVD)
REGISTE_PARAM_JSON_FUNC(MaskConvolution)
REGISTE_PARAM_JSON_FUNC(Images2Neibs)
REGISTE_PARAM_JSON_FUNC(Local)
REGISTE_PARAM_JSON_FUNC(GroupLocal)
REGISTE_PARAM_JSON_FUNC(LRN)
REGISTE_PARAM_JSON_FUNC(Concat)
REGISTE_PARAM_JSON_FUNC(Reduce)
REGISTE_PARAM_JSON_FUNC(LocalShareForward)
REGISTE_PARAM_JSON_FUNC(LocalShareBackwardData)
REGISTE_PARAM_JSON_FUNC(LocalShareBackwardFilter)
REGISTE_PARAM_JSON_FUNC(DeformableConvForward)
REGISTE_PARAM_JSON_FUNC(DeformableConvBackwardFilter)
REGISTE_PARAM_JSON_FUNC(DeformableConvBackwardData)
REGISTE_PARAM_JSON_FUNC(BatchConvBiasForward)

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
template <>
std::shared_ptr<json::Value> opr_param_json_func<opr::Dimshuffle>(
    cg::OperatorNodeBase * opr) {
        auto param = opr->cast_final_safe<opr::Dimshuffle>().param();

        auto pattern = json::Array::make();
        for (size_t i = 0; i < param.pattern_len; i++)
            pattern->add(json::NumberInt::make(param.pattern[i]));

        return json::Object::make({
            {"ndim", json::NumberInt::make(param.ndim)},
            {"pattern", pattern},
        });
    }

template <>
std::shared_ptr<json::Value> opr_param_json_func<opr::AxisAddRemove>(
    cg::OperatorNodeBase * opr) {
        auto param = opr->cast_final_safe<opr::AxisAddRemove>().param();

        auto desc = json::Array::make();
        for (size_t i = 0; i < param.nr_desc; i++) {
            auto axisdesc = param.desc[i];
            desc->add(
                json::Object::make({
                    {"method", json::NumberInt::make(
                        static_cast<int32_t>(axisdesc.method))},
                    {"axisnum", json::NumberInt::make(axisdesc.axis.get_raw())},
                }));
        }

        return json::Object::make({
            {"nr_desc", json::NumberInt::make(param.nr_desc)},
            {"desc", desc},
        });
    }

template <>
std::shared_ptr<json::Value> opr_param_json_func<opr::Subtensor>(
    cg::OperatorNodeBase * opr) {
        auto desc = json::Array::make();
        auto indices = opr->cast_final_safe<opr::Subtensor>().index_desc();
        for (auto &index : indices){
            desc->add(
                json::Object::make({
                    {"axis", json::NumberInt::make(index.axis.get_raw())},
                    {"begin", json::NumberInt::make(index.begin.node() != nullptr)},
                    {"end", json::NumberInt::make(index.end.node() != nullptr)},
                    {"step", json::NumberInt::make(index.step.node() != nullptr)},
                    {"idx", json::NumberInt::make(index.idx.node() != nullptr)},
                }));
        }

        return desc;
    }
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
#endif // MGB_ENABLE_JSON

}  // namespace

template <class OprType>
void OprFootprint::add_single_comp_footprint() {
    auto&& record = m_type2comp_footprint.emplace(OprType::typeinfo(),
                                                  opr_footprint_func<OprType>);
    mgb_assert(record.second, "duplicate opr typeinfo");
}

#if MGB_ENABLE_JSON
template <class OprType>
void OprFootprint::add_single_param_json() {
    auto&& record = m_type2param_json.emplace(OprType::typeinfo(),
                                              opr_param_json_func<OprType>);
    mgb_assert(record.second, "duplicate opr typeinfo");
}
#endif

void OprFootprint::init_all_footprints() {
    add_single_comp_footprint<opr::Elemwise>();
    add_single_comp_footprint<opr::AddUpdate>();
    add_single_comp_footprint<opr::ConvolutionForward>();
    add_single_comp_footprint<opr::ConvBiasForward>();
    add_single_comp_footprint<opr::ConvolutionBackwardData>();
    add_single_comp_footprint<opr::ConvolutionBackwardFilter>();
    add_single_comp_footprint<opr::MatrixMul>();
    add_single_comp_footprint<opr::PoolingForward>();
    add_single_comp_footprint<opr::Concat>();
    add_single_comp_footprint<opr::Dimshuffle>();
    add_single_comp_footprint<opr::Reduce>();
    add_single_comp_footprint<opr::Host2DeviceCopy>();
    add_single_comp_footprint<opr::LocalShareForward>();
    add_single_comp_footprint<opr::LocalShareBackwardData>();
    add_single_comp_footprint<opr::LocalShareBackwardFilter>();
    add_single_comp_footprint<opr::DeformableConvForward>();
    add_single_comp_footprint<opr::DeformableConvBackwardFilter>();
    add_single_comp_footprint<opr::DeformableConvBackwardData>();
    add_single_comp_footprint<opr::BatchConvBiasForward>();

#if MGB_ENABLE_JSON
    add_single_param_json<opr::Elemwise>();
    add_single_param_json<opr::ConvolutionForward>();
    add_single_param_json<opr::Convolution3D>();
    add_single_param_json<opr::ConvBiasForward>();
    add_single_param_json<opr::ConvolutionBackwardData>();
    add_single_param_json<opr::Convolution3DBackwardData>();
    add_single_param_json<opr::ConvolutionBackwardFilter>();
    add_single_param_json<opr::MatrixMul>();
    add_single_param_json<opr::BatchedMatrixMul>();
    add_single_param_json<opr::Dot>();
    add_single_param_json<opr::MatrixInverse>();
    add_single_param_json<opr::PoolingForward>();
    add_single_param_json<opr::SVD>();
    add_single_param_json<opr::MaskConvolution>();
    add_single_param_json<opr::Images2Neibs>();
    add_single_param_json<opr::Local>();
    add_single_param_json<opr::GroupLocal>();
    add_single_param_json<opr::LRN>();
    add_single_param_json<opr::Concat>();
631 632 633
    add_single_param_json<opr::Dimshuffle>();
    add_single_param_json<opr::AxisAddRemove>();
    add_single_param_json<opr::Subtensor>();
634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768
    add_single_param_json<opr::Reduce>();
    add_single_param_json<opr::LocalShareForward>();
    add_single_param_json<opr::LocalShareBackwardData>();
    add_single_param_json<opr::LocalShareBackwardFilter>();
    add_single_param_json<opr::DeformableConvForward>();
    add_single_param_json<opr::DeformableConvBackwardFilter>();
    add_single_param_json<opr::DeformableConvBackwardData>();
    add_single_param_json<opr::BatchConvBiasForward>();

#endif
}

OprFootprint::Result OprFootprint::calc_footprint(cg::OperatorNodeBase* opr) {
    Result rst;
    auto&& dep_map = opr->node_prop().dep_map();
    for (auto&& inp : opr->input()) {
        if (inp->mem_plan().valid())
            rst.inp_layout.push_back(inp->layout());
        else
            rst.inp_layout.push_back({inp->shape(), inp->dtype()});
        if (cg::OperatorNodeBase::NodeProp::is_device_value_dep(
                    dep_map.at(inp))) {
            rst.memory += inp->dtype().size(inp->shape().total_nr_elems());
        }
    }
    for (auto&& out : opr->output()) {
        if (out->contain_flag(VarNode::Flag::VOLATILE_CONTENT))
            continue;
        rst.out_shape.push_back(out->shape());
        rst.memory += out->dtype().size(out->shape().total_nr_elems());
    }
    rst.computation = get_computation(opr);
#if MGB_ENABLE_JSON
    rst.param = get_param_json(opr);
#endif
    rst.opr_type = opr->dyn_typeinfo();
    return rst;
}

uint64_t OprFootprint::get_computation(cg::OperatorNodeBase* opr) {
    auto comp_trait = m_type2comp_footprint.find(opr->dyn_typeinfo());
    if (comp_trait != m_type2comp_footprint.end()) {
        return (comp_trait->second)(opr);
    }
    return 0;
}

#if MGB_ENABLE_JSON
std::shared_ptr<json::Value> OprFootprint::get_param_json(
        cg::OperatorNodeBase* opr) {
    auto param_trait = m_type2param_json.find(opr->dyn_typeinfo());
    if (param_trait != m_type2param_json.end()) {
        return (param_trait->second)(opr);
    }
    return json::Object::make();
}

std::shared_ptr<json::Value> OprFootprint::Result::to_json() const {
    using namespace json;
    std::shared_ptr<Value> comp;
    if (computation) {
        comp = NumberInt::make(computation);
    } else {
        comp = Null::make();
    }
    auto format_shape_arr = [](const TensorShapeArray& arr) {
        auto ret = Array::make();
        for (auto&& shp : arr) {
            auto cur = Array::make();
            for (size_t i = 0; i < shp.ndim; ++i) {
                cur->add(NumberInt::make(shp[i]));
            }
            ret->add(std::move(cur));
        }
        return ret;
    };
    auto format_layout_arr =
            [](const TensorLayoutArray& arr) -> std::shared_ptr<Value> {
        auto ret = Array::make();
        bool have_non_contig = false;
        for (auto&& item : arr) {
            if (item.is_contiguous()) {
                ret->add(json::Null::make());
            } else {
                have_non_contig = true;
                auto cur = Array::make();
                for (size_t i = 0; i < item.ndim; ++i) {
                    cur->add(NumberInt::make(item.stride[i]));
                }
                ret->add(std::move(cur));
            }
        }
        if (!have_non_contig) {
            ret.reset();
        }
        return ret;
    };

    TensorShapeArray inp_shape;
    for (auto&& i : inp_layout)
        inp_shape.push_back(i);
    auto ret = Object::make({{"computation", std::move(comp)},
                             {"memory", NumberInt::make(memory)},
                             {"in_shapes", format_shape_arr(inp_shape)},
                             {"out_shapes", format_shape_arr(out_shape)},
                             {"param", param}});
    if (auto inp_layout_json = format_layout_arr(inp_layout)) {
        ret->operator[]("in_layouts") = std::move(inp_layout_json);
    }
    return ret;
}

std::shared_ptr<json::Value> OprFootprint::get_opr_fp_graph_exec(
        cg::ComputingGraph& graph, const SymbolVarArray& outputs) {
    OprFootprint m_opr_footprint;
    ComputingGraph::OutputSpec out_spec;
    for (auto i : outputs) {
        out_spec.emplace_back(i, nullptr);
    }
    graph.options().allocate_static_mem_after_graph_compile = true;
    auto async_exec = graph.compile(out_spec);
    std::vector<std::pair<json::String, std::shared_ptr<json::Value>>> rst_vals;
    auto on_opr = [&m_opr_footprint, &rst_vals](cg::OperatorNodeBase* opr) {
        Result trait(m_opr_footprint.calc_footprint(opr));
        rst_vals.emplace_back(json::String(opr->id_str()), trait.to_json());
        return true;
    };
    async_exec->iter_opr_seq(on_opr);
    auto opr_fp = json::Object::make(rst_vals);
    return json::Object::make(
            {{"opr_footprint", opr_fp}, {"graph_exec", async_exec->to_json()}});
}
#endif

// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}