bm_fast.cu 10.4 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
/*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) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., 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*/

#if !defined CUDA_DISABLER

#include "opencv2/core/cuda/common.hpp"
#include "opencv2/core/cuda/limits.hpp"
#include "opencv2/core/cuda/functional.hpp"
#include "opencv2/core/cuda/reduce.hpp"

using namespace cv::cuda;
using namespace cv::cuda::device;

namespace optflowbm_fast
{
    enum
    {
        CTA_SIZE = 128,

        TILE_COLS = 128,
        TILE_ROWS = 32,

        STRIDE = CTA_SIZE
    };

    template <typename T> __device__ __forceinline__ int calcDist(T a, T b)
    {
        return ::abs(a - b);
    }

    template <class T> struct FastOptFlowBM
    {

        int search_radius;
        int block_radius;

        int search_window;
        int block_window;

        PtrStepSz<T> I0;
        PtrStep<T> I1;

        mutable PtrStepi buffer;

        FastOptFlowBM(int search_window_, int block_window_,
                      PtrStepSz<T> I0_, PtrStepSz<T> I1_,
                      PtrStepi buffer_) :
            search_radius(search_window_ / 2), block_radius(block_window_ / 2),
            search_window(search_window_), block_window(block_window_),
            I0(I0_), I1(I1_),
            buffer(buffer_)
        {
        }

        __device__ __forceinline__ void initSums_BruteForce(int i, int j, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
        {
            for (int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
            {
                dist_sums[index] = 0;

                for (int tx = 0; tx < block_window; ++tx)
                    col_sums(tx, index) = 0;

                int y = index / search_window;
                int x = index - y * search_window;

                int ay = i;
                int ax = j;

                int by = i + y - search_radius;
                int bx = j + x - search_radius;

                for (int tx = -block_radius; tx <= block_radius; ++tx)
                {
                    int col_sum = 0;
                    for (int ty = -block_radius; ty <= block_radius; ++ty)
                    {
                        int dist = calcDist(I0(ay + ty, ax + tx), I1(by + ty, bx + tx));

                        dist_sums[index] += dist;
                        col_sum += dist;
                    }

                    col_sums(tx + block_radius, index) = col_sum;
                }

                up_col_sums(j, index) = col_sums(block_window - 1, index);
            }
        }

        __device__ __forceinline__ void shiftRight_FirstRow(int i, int j, int first, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
        {
            for (int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
            {
                int y = index / search_window;
                int x = index - y * search_window;

                int ay = i;
                int ax = j + block_radius;

                int by = i + y - search_radius;
                int bx = j + x - search_radius + block_radius;

                int col_sum = 0;

                for (int ty = -block_radius; ty <= block_radius; ++ty)
                    col_sum += calcDist(I0(ay + ty, ax), I1(by + ty, bx));

                dist_sums[index] += col_sum - col_sums(first, index);

                col_sums(first, index) = col_sum;
                up_col_sums(j, index) = col_sum;
            }
        }

        __device__ __forceinline__ void shiftRight_UpSums(int i, int j, int first, int* dist_sums, PtrStepi& col_sums, PtrStepi& up_col_sums) const
        {
            int ay = i;
            int ax = j + block_radius;

            T a_up   = I0(ay - block_radius - 1, ax);
            T a_down = I0(ay + block_radius, ax);

            for(int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
            {
                int y = index / search_window;
                int x = index - y * search_window;

                int by = i + y - search_radius;
                int bx = j + x - search_radius + block_radius;

                T b_up   = I1(by - block_radius - 1, bx);
                T b_down = I1(by + block_radius, bx);

                int col_sum = up_col_sums(j, index) + calcDist(a_down, b_down) - calcDist(a_up, b_up);

                dist_sums[index] += col_sum  - col_sums(first, index);
                col_sums(first, index) = col_sum;
                up_col_sums(j, index) = col_sum;
            }
        }

        __device__ __forceinline__ void convolve_window(int i, int j, const int* dist_sums, float& velx, float& vely) const
        {
            int bestDist = numeric_limits<int>::max();
            int bestInd = -1;

            for (int index = threadIdx.x; index < search_window * search_window; index += STRIDE)
            {
                int curDist = dist_sums[index];
                if (curDist < bestDist)
                {
                    bestDist = curDist;
                    bestInd = index;
                }
            }

            __shared__ int cta_dist_buffer[CTA_SIZE];
            __shared__ int cta_ind_buffer[CTA_SIZE];

            reduceKeyVal<CTA_SIZE>(cta_dist_buffer, bestDist, cta_ind_buffer, bestInd, threadIdx.x, less<int>());

            if (threadIdx.x == 0)
            {
                int y = bestInd / search_window;
                int x = bestInd - y * search_window;

                velx = x - search_radius;
                vely = y - search_radius;
            }
        }

        __device__ __forceinline__ void operator()(PtrStepf velx, PtrStepf vely) const
        {
            int tbx = blockIdx.x * TILE_COLS;
            int tby = blockIdx.y * TILE_ROWS;

            int tex = ::min(tbx + TILE_COLS, I0.cols);
            int tey = ::min(tby + TILE_ROWS, I0.rows);

            PtrStepi col_sums;
            col_sums.data = buffer.ptr(I0.cols + blockIdx.x * block_window) + blockIdx.y * search_window * search_window;
            col_sums.step = buffer.step;

            PtrStepi up_col_sums;
            up_col_sums.data = buffer.data + blockIdx.y * search_window * search_window;
            up_col_sums.step = buffer.step;

            extern __shared__ int dist_sums[]; //search_window * search_window

            int first = 0;

            for (int i = tby; i < tey; ++i)
            {
                for (int j = tbx; j < tex; ++j)
                {
                    __syncthreads();

                    if (j == tbx)
                    {
                        initSums_BruteForce(i, j, dist_sums, col_sums, up_col_sums);
                        first = 0;
                    }
                    else
                    {
                        if (i == tby)
                          shiftRight_FirstRow(i, j, first, dist_sums, col_sums, up_col_sums);
                        else
                          shiftRight_UpSums(i, j, first, dist_sums, col_sums, up_col_sums);

                        first = (first + 1) % block_window;
                    }

                    __syncthreads();

                    convolve_window(i, j, dist_sums, velx(i, j), vely(i, j));
                }
            }
        }

    };

    template<typename T> __global__ void optflowbm_fast_kernel(const FastOptFlowBM<T> fbm, PtrStepf velx, PtrStepf vely)
    {
        fbm(velx, vely);
    }

    void get_buffer_size(int src_cols, int src_rows, int search_window, int block_window, int& buffer_cols, int& buffer_rows)
    {
        dim3 grid(divUp(src_cols, TILE_COLS), divUp(src_rows, TILE_ROWS));

        buffer_cols = search_window * search_window * grid.y;
        buffer_rows = src_cols + block_window * grid.x;
    }

    template <typename T>
    void calc(PtrStepSzb I0, PtrStepSzb I1, PtrStepSzf velx, PtrStepSzf vely, PtrStepi buffer, int search_window, int block_window, cudaStream_t stream)
    {
        FastOptFlowBM<T> fbm(search_window, block_window, I0, I1, buffer);

        dim3 block(CTA_SIZE, 1);
        dim3 grid(divUp(I0.cols, TILE_COLS), divUp(I0.rows, TILE_ROWS));

        size_t smem = search_window * search_window * sizeof(int);

        optflowbm_fast_kernel<<<grid, block, smem, stream>>>(fbm, velx, vely);
        cudaSafeCall ( cudaGetLastError () );

        if (stream == 0)
            cudaSafeCall( cudaDeviceSynchronize() );
    }

    template void calc<uchar>(PtrStepSzb I0, PtrStepSzb I1, PtrStepSzf velx, PtrStepSzf vely, PtrStepi buffer, int search_window, int block_window, cudaStream_t stream);
}

#endif // !defined CUDA_DISABLER