pooling_layer.cpp 27.1 KB
Newer Older
wester committed
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
/*M///////////////////////////////////////////////////////////////////////////////////////
//
//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
//  By downloading, copying, installing or using the software you agree to this license.
//  If you do not agree to this license, do not download, install,
//  copy or use the software.
//
//
//                           License Agreement
//                For Open Source Computer Vision Library
//
// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
// Copyright (C) 2017, Intel Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
//   * Redistribution's of source code must retain the above copyright notice,
//     this list of conditions and the following disclaimer.
//
//   * Redistribution's in binary form must reproduce the above copyright notice,
//     this list of conditions and the following disclaimer in the documentation
//     and/or other materials provided with the distribution.
//
//   * The name of the copyright holders may not be used to endorse or promote products
//     derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/

#include "../precomp.hpp"
#include "layers_common.hpp"
#include "opencv2/core/hal/intrin.hpp"
#include "op_halide.hpp"
#include <float.h>
#include <algorithm>
using std::max;
using std::min;

namespace cv
{
namespace dnn
{

//TODO: add ceil_mode param
class PoolingLayerImpl : public PoolingLayer
{
public:
    PoolingLayerImpl(const LayerParams& params)
    {
        type = PoolingLayer::MAX;
        computeMaxIdx = true;

        if (params.has("pool"))
        {
            String pool = params.get<String>("pool").toLowerCase();
            if (pool == "max")
                type = PoolingLayer::MAX;
            else if (pool == "ave")
                type = PoolingLayer::AVE;
            else if (pool == "stochastic")
                type = PoolingLayer::STOCHASTIC;
            else
                CV_Error(Error::StsBadArg, "Unknown pooling type \"" + pool + "\"");
        }

        getPoolingKernelParams(params, kernel.height, kernel.width, globalPooling,
                               pad.height, pad.width, stride.height, stride.width, padMode);
        setParamsFrom(params);
    }

    void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs)
    {
        CV_Assert(inputs.size() == 1);

        cv::Size inp(inputs[0]->size[3], inputs[0]->size[2]),
                out(outputs[0].size[3], outputs[0].size[2]);

        if(globalPooling)
        {
            kernel = inp;
        }

        getConvPoolPaddings(inp, out, kernel, stride, padMode, pad);
    }

    virtual bool supportBackend(int backendId)
    {
        return backendId == DNN_BACKEND_DEFAULT ||
               backendId == DNN_BACKEND_HALIDE && haveHalide() &&
               (type == PoolingLayer::MAX ||
                type == PoolingLayer::AVE && !pad.width && !pad.height);
    }

    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
    {
        CV_TRACE_FUNCTION();
        CV_TRACE_ARG_VALUE(name, "name", name.c_str());

        for (size_t ii = 0; ii < inputs.size(); ii++)
        {
            switch (type)
            {
                case MAX:
                    maxPooling(*inputs[ii], outputs[2 * ii], outputs[2 * ii + 1]);
                    break;
                case AVE:
                    avePooling(*inputs[ii], outputs[ii]);
                    break;
                default:
                    CV_Error(Error::StsNotImplemented, "Not implemented");
                    break;
            }
        }
    }

    virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
    {
        if (type == PoolingLayer::MAX)
            return initMaxPoolingHalide(inputs);
        else if (type == PoolingLayer::AVE)
            return initAvePoolingHalide(inputs);
        else
            return Ptr<BackendNode>();
    }

    class PoolingInvoker : public ParallelLoopBody
    {
    public:
        const Mat* src;
        Mat *dst, *mask;
        Size kernel, stride, pad;
        int nstripes;
        bool computeMaxIdx;
        std::vector<int> ofsbuf;
        int poolingType;

        PoolingInvoker() : src(0), dst(0), mask(0), nstripes(0), computeMaxIdx(0), poolingType(PoolingLayer::MAX) {}

        static void run(const Mat& src, Mat& dst, Mat& mask, Size kernel,
                        Size stride, Size pad, int poolingType,
                        bool computeMaxIdx, int nstripes)
        {
            CV_Assert(src.isContinuous() && dst.isContinuous() &&
                      src.type() == CV_32F && src.type() == dst.type() &&
                      src.dims == 4 && dst.dims == 4 &&
                      src.size[0] == dst.size[0] && src.size[1] == dst.size[1] &&
                      (mask.empty() || (mask.type() == src.type() && mask.size == dst.size)));

            PoolingInvoker p;

            p.src = &src;
            p.dst = &dst;
            p.mask = &mask;
            p.kernel = kernel;
            p.stride = stride;
            p.pad = pad;
            p.nstripes = nstripes;
            p.computeMaxIdx = computeMaxIdx;
            p.poolingType = poolingType;

            if( !computeMaxIdx )
            {
                p.ofsbuf.resize(kernel.width*kernel.height);
                for( int i = 0; i < kernel.height; i++ )
                    for( int j = 0; j < kernel.width; j++ )
                        p.ofsbuf[i*kernel.width + j] = src.size[3]*i + j;
            }

            parallel_for_(Range(0, nstripes), p, nstripes);
        }

        void operator()(const Range& r) const
        {
            int channels = dst->size[1], width = dst->size[3], height = dst->size[2];
            int inp_width = src->size[3], inp_height = src->size[2];
            size_t total = dst->total();
            size_t stripeSize = (total + nstripes - 1)/nstripes;
            size_t stripeStart = r.start*stripeSize;
            size_t stripeEnd = std::min(r.end*stripeSize, total);
            int kernel_w = kernel.width, kernel_h = kernel.height;
            int pad_w = pad.width, pad_h = pad.height;
            int stride_w = stride.width, stride_h = stride.height;
            bool compMaxIdx = computeMaxIdx;

#if CV_SIMD128
            const int* ofsptr = &ofsbuf[0];
            v_float32x4 idx00(0.f, (float)stride_w, (float)(stride_w*2), (float)(stride_w*3));
            v_float32x4 ones = v_setall_f32(1.f);
            v_float32x4 idx_delta = v_setall_f32((float)(inp_width - kernel_w));
#endif

            for( size_t ofs0 = stripeStart; ofs0 < stripeEnd; )
            {
                size_t ofs = ofs0;
                int x0 = (int)(ofs % width);
                ofs /= width;
                int y0 = (int)(ofs % height);
                ofs /= height;
                int c = (int)(ofs % channels);
                int n = (int)(ofs / channels);
                int ystart = y0 * stride_h - pad_h;
                int yend = min(ystart + kernel_h, inp_height + pad_h);
                int ydelta = yend - ystart;
                ystart = max(ystart, 0);
                yend = min(yend, inp_height);
                const float *srcData = src->ptr<float>(n, c);
                float *dstData = dst->ptr<float>(n, c, y0);
                float *dstMaskData = mask->data ? mask->ptr<float>(n, c, y0) : 0;

                int delta = std::min((int)(stripeEnd - ofs0), width - x0);
                ofs0 += delta;
                int x1 = x0 + delta;

                if( poolingType == PoolingLayer::MAX )
                    for( ; x0 < x1; x0++ )
                    {
                        int xstart = x0 * stride_w - pad_w;
                        int xend = min(xstart + kernel_w, inp_width);
                        xstart = max(xstart, 0);

#if CV_SIMD128
                        if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
                        {
                            if( compMaxIdx )
                            {
                                v_float32x4 max_val0 = v_setall_f32(-FLT_MAX);
                                v_float32x4 max_val1 = max_val0;
                                v_float32x4 max_idx0 = v_setall_f32(-1.f);
                                v_float32x4 max_idx1 = max_idx0;
                                int index0 = ystart * inp_width + xstart;
                                v_float32x4 idx0 = idx00 + v_setall_f32((float)index0);
                                v_float32x4 idx1 = idx0 + v_setall_f32((float)(stride_w*4));

                                for (int y = ystart; y < yend; ++y)
                                {
                                    for (int x = xstart; x < xend; ++x, idx0 += ones, idx1 += ones)
                                    {
                                        const int index = y * inp_width + x;
                                        v_float32x4 v0(srcData[index], srcData[index + stride_w],
                                                       srcData[index + stride_w*2], srcData[index + stride_w*3]);
                                        v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
                                                       srcData[index + stride_w*6], srcData[index + stride_w*7]);
                                        max_idx0 = v_select(v0 > max_val0, idx0, max_idx0);
                                        max_idx1 = v_select(v1 > max_val1, idx1, max_idx1);
                                        max_val0 = v_max(max_val0, v0);
                                        max_val1 = v_max(max_val1, v1);
                                    }
                                    idx0 += idx_delta;
                                    idx1 += idx_delta;
                                }
                                v_store(dstData + x0, max_val0);
                                v_store(dstData + x0 + 4, max_val1);
                                if (dstMaskData)
                                {
                                    v_store(dstMaskData + x0, max_idx0);
                                    v_store(dstMaskData + x0 + 4, max_idx1);
                                }
                                x0 += 7;
                            }
                            else
                            {
                                v_float32x4 max_val0 = v_setall_f32(-FLT_MAX);
                                v_float32x4 max_val1 = max_val0;

                                if( yend - ystart == kernel_h )
                                {
                                    const float* srcData1 = srcData + ystart*inp_width + xstart;
                                    if( stride_w == 1 )
                                        for (int k = 0; k < kernel_w*kernel_h; k++)
                                        {
                                            int index = ofsptr[k];
                                            v_float32x4 v0 = v_load(srcData1 + index);
                                            v_float32x4 v1 = v_load(srcData1 + index + 4);
                                            max_val0 = v_max(max_val0, v0);
                                            max_val1 = v_max(max_val1, v1);
                                        }
#if CV_SSE2
                                    else if( stride_w == 2 )
                                        for (int k = 0; k < kernel_w*kernel_h; k++)
                                        {
                                            int index = ofsptr[k];
                                            v_float32x4 v00 = v_load(srcData1 + index), v01 = v_load(srcData1 + index + 4);
                                            v_float32x4 v0(_mm_shuffle_ps(v00.val, v01.val, _MM_SHUFFLE(2, 0, 2, 0)));
                                            v_float32x4 v10 = v_load(srcData1 + index + 8), v11 = v_load(srcData1 + index + 12);
                                            v_float32x4 v1(_mm_shuffle_ps(v10.val, v11.val, _MM_SHUFFLE(2, 0, 2, 0)));
                                            max_val0 = v_max(max_val0, v0);
                                            max_val1 = v_max(max_val1, v1);
                                        }
#endif
                                    else
                                        for (int k = 0; k < kernel_w*kernel_h; k++)
                                        {
                                            int index = ofsptr[k];
                                            v_float32x4 v0(srcData1[index], srcData1[index + stride_w],
                                                           srcData1[index + stride_w*2], srcData1[index + stride_w*3]);
                                            v_float32x4 v1(srcData1[index + stride_w*4], srcData1[index + stride_w*5],
                                                           srcData1[index + stride_w*6], srcData1[index + stride_w*7]);
                                            max_val0 = v_max(max_val0, v0);
                                            max_val1 = v_max(max_val1, v1);
                                        }
                                }
                                else
                                {
                                    for (int y = ystart; y < yend; ++y)
                                    {
                                        for (int x = xstart; x < xend; ++x)
                                        {
                                            const int index = y * inp_width + x;
                                            v_float32x4 v0(srcData[index], srcData[index + stride_w],
                                                           srcData[index + stride_w*2], srcData[index + stride_w*3]);
                                            v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
                                                           srcData[index + stride_w*6], srcData[index + stride_w*7]);
                                            max_val0 = v_max(max_val0, v0);
                                            max_val1 = v_max(max_val1, v1);
                                        }
                                    }
                                }
                                v_store(dstData + x0, max_val0);
                                v_store(dstData + x0 + 4, max_val1);
                                x0 += 7;
                            }
                        }
                        else
#endif
                        {
                            float max_val = -FLT_MAX;
                            if( compMaxIdx )
                            {
                                int max_index = -1;
                                for (int y = ystart; y < yend; ++y)
                                    for (int x = xstart; x < xend; ++x)
                                    {
                                        const int index = y * inp_width + x;
                                        float val = srcData[index];
                                        if (val > max_val)
                                        {
                                            max_val = val;
                                            max_index = index;
                                        }
                                    }

                                dstData[x0] = max_val;
                                if (dstMaskData)
                                    dstMaskData[x0] = max_index;
                            }
                            else
                            {
                                for (int y = ystart; y < yend; ++y)
                                    for (int x = xstart; x < xend; ++x)
                                    {
                                        const int index = y * inp_width + x;
                                        float val = srcData[index];
                                        max_val = std::max(max_val, val);
                                    }

                                dstData[x0] = max_val;
                            }
                        }
                    }
                else
                {
                    for( ; x0 < x1; x0++ )
                    {
                        int xstart = x0 * stride_w - pad_w;
                        int xend = min(xstart + kernel_w, inp_width + pad_w);
                        int xdelta = xend - xstart;
                        xstart = max(xstart, 0);
                        xend = min(xend, inp_width);
                        float inv_kernel_area = 1.f/(ydelta*xdelta);

#if CV_SIMD128
                        if( xstart > 0 && x0 + 7 < x1 && (x0 + 7) * stride_w - pad_w + kernel_w < inp_width )
                        {
                            v_float32x4 sum_val0 = v_setzero_f32(), sum_val1 = v_setzero_f32();
                            v_float32x4 ikarea = v_setall_f32(inv_kernel_area);

                            for (int y = ystart; y < yend; ++y)
                            {
                                for (int x = xstart; x < xend; ++x)
                                {
                                    const int index = y * inp_width + x;
                                    v_float32x4 v0(srcData[index], srcData[index + stride_w],
                                                   srcData[index + stride_w*2], srcData[index + stride_w*3]);
                                    v_float32x4 v1(srcData[index + stride_w*4], srcData[index + stride_w*5],
                                                   srcData[index + stride_w*6], srcData[index + stride_w*7]);
                                    sum_val0 += v0;
                                    sum_val1 += v1;
                                }
                            }
                            v_store(dstData + x0, sum_val0*ikarea);
                            v_store(dstData + x0 + 4, sum_val1*ikarea);
                            x0 += 7;
                        }
                        else
#endif
                        {
                            float sum_val = 0.f;
                            for (int y = ystart; y < yend; ++y)
                                for (int x = xstart; x < xend; ++x)
                                {
                                    const int index = y * inp_width + x;
                                    float val = srcData[index];
                                    sum_val += val;
                                }

                            dstData[x0] = sum_val*inv_kernel_area;
                        }
                    }
                }
            }
        }
    };

    void maxPooling(Mat &src, Mat &dst, Mat &mask)
    {
        const int nstripes = getNumThreads();
        PoolingInvoker::run(src, dst, mask, kernel, stride, pad, type, computeMaxIdx, nstripes);
    }

    void avePooling(Mat &src, Mat &dst)
    {
        const int nstripes = getNumThreads();
        Mat mask;
        PoolingInvoker::run(src, dst, mask, kernel, stride, pad, type, computeMaxIdx, nstripes);
    }

    virtual Ptr<BackendNode> initMaxPoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
    {
#ifdef HAVE_HALIDE
        Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);
        const int inWidth = inputBuffer.width();
        const int inHeight = inputBuffer.height();

        Halide::Var x("x"), y("y"), c("c"), n("n");
        Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
        Halide::RDom r(0, kernel.width, 0, kernel.height);
        Halide::Expr kx, ky;
        if (pad.width || pad.height)
        {
            kx = clamp(x * stride.width + r.x - pad.width, 0, inWidth - 1);
            ky = clamp(y * stride.height + r.y - pad.height, 0, inHeight - 1);
        }
        else
        {
            kx = min(x * stride.width + r.x, inWidth - 1);
            ky = min(y * stride.height + r.y, inHeight - 1);
        }

        // Halide::argmax returns tuple (r.x, r.y, max).
        Halide::Tuple res = argmax(inputBuffer(kx, ky, c, n));

        // Compute offset from argmax in range [0, kernel_size).
        Halide::Expr max_index;
        if (pad.width || pad.height)
        {
            max_index = clamp(y * stride.height + res[1] - pad.height,
                              0, inHeight - 1) * inWidth +
                        clamp(x * stride.width + res[0] - pad.width,
                              0, inWidth - 1);
        }
        else
        {
            max_index = min(y * stride.height + res[1], inHeight - 1) * inWidth +
                        min(x * stride.width + res[0], inWidth - 1);
        }
        top(x, y, c, n) = { res[2], Halide::cast<float>(max_index) };
        return Ptr<BackendNode>(new HalideBackendNode(top));
#endif  // HAVE_HALIDE
        return Ptr<BackendNode>();
    }

    virtual Ptr<BackendNode> initAvePoolingHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
    {
#ifdef HAVE_HALIDE
        Halide::Buffer<float> inputBuffer = halideBuffer(inputs[0]);

        const int inW = inputBuffer.width(), inH = inputBuffer.height();
        if ((inW - kernel.width) % stride.width || (inH - kernel.height) % stride.height)
        {
            CV_Error(cv::Error::StsNotImplemented,
                     "Halide backend for average pooling with partial "
                     "kernels is not implemented");
        }

        const float norm = 1.0f / (kernel.width * kernel.height);

        Halide::Var x("x"), y("y"), c("c"), n("n");
        Halide::Func top = (name.empty() ? Halide::Func() : Halide::Func(name));
        Halide::RDom r(0, kernel.width, 0, kernel.height);
        top(x, y, c, n) = sum(
            inputBuffer(x * stride.width + r.x,
                        y * stride.height + r.y, c, n)) * norm;
        return Ptr<BackendNode>(new HalideBackendNode(top));
#endif  // HAVE_HALIDE
        return Ptr<BackendNode>();
    }

    virtual void applyHalideScheduler(Ptr<BackendNode>& node,
                                      const std::vector<Mat*> &inputs,
                                      const std::vector<Mat> &outputs,
                                      int targetId) const
    {
#ifdef  HAVE_HALIDE
        if (targetId != DNN_TARGET_CPU)
        {
            Layer::applyHalideScheduler(node, inputs, outputs, targetId);
            return;
        }
        Halide::Var x("x"), y("y"), c("c"), n("n"), tile("tile"),
                    xi("xi"), yi("yi"), ci("ci"), xo("xo"), yo("yo"), co("co");
        Halide::Func& top = node.dynamicCast<HalideBackendNode>()->funcs.back();

        int outW, outH, outC, outN;
        getCanonicalSize(outputs[0].size, &outW, &outH, &outC, &outN);

        if (outW < 8 || outH < 8)
        {
            if (outC > 8)
                top.split(c, co, ci, 8)
                   .fuse(x, y, tile).fuse(co, tile, tile).fuse(n, tile, tile)
                   .parallel(tile)
                   .vectorize(ci);
            else
            {
                top.fuse(y, c, tile).fuse(n, tile, tile)
                   .parallel(tile);
                if (outW > 1)
                    top.vectorize(x);
            }
        }
        else
        {
            if (outC > 8)
                top.split(x, xo, xi, 8).split(y, yo, yi, 8).split(c, co, ci, 8)
                   .fuse(xo, yo, tile).fuse(co, tile, tile).fuse(n, tile, tile)
                   .parallel(tile)
                   .vectorize(xi);
            else
                top.split(x, xo, xi, 8).split(y, yo, yi, 8)
                   .fuse(xo, yo, tile).fuse(c, tile, tile).fuse(n, tile, tile)
                   .parallel(tile)
                   .vectorize(xi);
        }
#endif  // HAVE_HALIDE
    }

    bool getMemoryShapes(const std::vector<MatShape> &inputs,
                         const int requiredOutputs,
                         std::vector<MatShape> &outputs,
                         std::vector<MatShape> &internals) const
    {
        CV_Assert(inputs.size() != 0);
        Size in(inputs[0][3], inputs[0][2]), out;

        if (globalPooling)
        {
            out.height = 1;
            out.width = 1;
        }
        else if (padMode.empty())
        {
            //Yeah, something strange Caffe scheme-)
            out.height = static_cast<int>(ceil(static_cast<float>(in.height + 2 * pad.height -
                                                                  kernel.height) / stride.height)) + 1;
            out.width = static_cast<int>(ceil(static_cast<float>(in.width + 2 * pad.width -
                                                                 kernel.width) / stride.width)) + 1;

            if (pad.height || pad.width)
            {
                // If we have padding, ensure that the last pooling starts strictly
                // inside the image (instead of at the padding); otherwise clip the last.
                if ((out.height - 1) * stride.height >= in.height + pad.height)
                    --out.height;
                if ((out.width - 1) * stride.width >= in.width + pad.width)
                    --out.width;
                CV_Assert((out.height - 1) * stride.height < in.height + pad.height);
                CV_Assert((out.width - 1) * stride.width < in.width + pad.width);
            }
        }
        else
        {
            getConvPoolOutParams(in, kernel, stride,
                                 padMode, out);
        }

        outputs.resize(type == MAX ? 2 * inputs.size() : inputs.size());
        for (size_t i = 0; i < inputs.size(); i++)
        {
            size_t index = type == MAX ? 2*i : i;
            int dims[] = {inputs[i][0], inputs[i][1], out.height, out.width};
            outputs[index] = shape(dims);

            if (type == MAX)
                outputs[index + 1] = shape(dims);
        }

        return false;
    }

    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
                           const std::vector<MatShape> &outputs) const
    {
        (void)inputs; // suppress unused variable warning
        long flops = 0;

        for(int i = 0; i < outputs.size(); i++)
        {
            if (type == MAX)
            {
                if (i%2 == 0)
                    flops += total(outputs[i])*kernel.area();
            }
            else
            {
                flops += total(outputs[i])*(kernel.area() + 1);
            }
        }
        return flops;
    }
};

Ptr<PoolingLayer> PoolingLayer::create(const LayerParams& params)
{
    return Ptr<PoolingLayer>(new PoolingLayerImpl(params));
}

}
}