brute_force_match.cl

/*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) 2010-2012, Multicoreware, Inc., all rights reserved.
// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// @Authors
//    Nathan, [e-mail]
//    Peng Xiao, [e-mail]
//    Baichuan Su, [e-mail]
//
// 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*/

#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics:enable
#define MAX_FLOAT 3.40282e+038f

#ifndef T
#define T float
#endif

#ifndef BLOCK_SIZE
#define BLOCK_SIZE 16
#endif
#ifndef MAX_DESC_LEN
#define MAX_DESC_LEN 64
#endif

#define BLOCK_SIZE_ODD          (BLOCK_SIZE + 1)
#ifndef SHARED_MEM_SZ
#  if (BLOCK_SIZE < MAX_DESC_LEN)
#    define SHARED_MEM_SZ      (kercn * (BLOCK_SIZE * MAX_DESC_LEN + BLOCK_SIZE * BLOCK_SIZE))
#  else
#    define SHARED_MEM_SZ      (kercn * 2 * BLOCK_SIZE_ODD * BLOCK_SIZE)
#  endif
#endif

#ifndef DIST_TYPE
#define DIST_TYPE 2
#endif

// dirty fix for non-template support
#if (DIST_TYPE == 2) // L1Dist
#   ifdef T_FLOAT
        typedef float result_type;
#       if (8 == kercn)
            typedef float8 value_type;
#           define DIST(x, y) {value_type d = fabs((x) - (y)); result += d.s0 + d.s1 + d.s2 + d.s3 + d.s4 + d.s5 + d.s6 + d.s7;}
#       elif (4 == kercn)
            typedef float4 value_type;
#           define DIST(x, y) {value_type d = fabs((x) - (y)); result += d.s0 + d.s1 + d.s2 + d.s3;}
#       else
            typedef float value_type;
#           define DIST(x, y) result += fabs((x) - (y))
#       endif
#   else
        typedef int result_type;
#       if (8 == kercn)
            typedef int8 value_type;
#           define DIST(x, y) {value_type d = abs((x) - (y)); result += d.s0 + d.s1 + d.s2 + d.s3 + d.s4 + d.s5 + d.s6 + d.s7;}
#       elif (4 == kercn)
            typedef int4 value_type;
#           define DIST(x, y) {value_type d = abs((x) - (y)); result += d.s0 + d.s1 + d.s2 + d.s3;}
#       else
            typedef int  value_type;
#           define DIST(x, y) result += abs((x) - (y))
#       endif
#   endif
#   define DIST_RES(x) (x)
#elif (DIST_TYPE == 4) // L2Dist
    typedef float result_type;
#   if (8 == kercn)
        typedef float8 value_type;
#       define DIST(x, y)   {value_type d = ((x) - (y)); result += dot(d.s0123, d.s0123) + dot(d.s4567, d.s4567);}
#   elif (4 == kercn)
        typedef float4      value_type;
#       define DIST(x, y)   {value_type d = ((x) - (y)); result += dot(d, d);}
#   else
        typedef float       value_type;
#       define DIST(x, y)   {value_type d = ((x) - (y)); result = mad(d, d, result);}
#   endif
#   define DIST_RES(x) sqrt(x)
#elif (DIST_TYPE == 6) // Hamming
#   if (8 == kercn)
        typedef int8 value_type;
#   elif (4 == kercn)
        typedef int4 value_type;
#   else
        typedef int value_type;
#   endif
    typedef int result_type;
#   define DIST(x, y) result += popcount( (x) ^ (y) )
#   define DIST_RES(x) (x)
#endif

inline result_type reduce_block(
    __local value_type *s_query,
    __local value_type *s_train,
    int lidx,
    int lidy
    )
{
    result_type result = 0;
    #pragma unroll
    for (int j = 0 ; j < BLOCK_SIZE ; j++)
    {
        DIST(s_query[lidy * BLOCK_SIZE_ODD + j], s_train[j * BLOCK_SIZE_ODD + lidx]);
    }
    return DIST_RES(result);
}

inline result_type reduce_block_match(
    __local value_type *s_query,
    __local value_type *s_train,
    int lidx,
    int lidy
    )
{
    result_type result = 0;
    #pragma unroll
    for (int j = 0 ; j < BLOCK_SIZE ; j++)
    {
        DIST(s_query[lidy * BLOCK_SIZE_ODD + j], s_train[j * BLOCK_SIZE_ODD + lidx]);
    }
    return result;
}

inline result_type reduce_multi_block(
    __local value_type *s_query,
    __local value_type *s_train,
    int block_index,
    int lidx,
    int lidy
    )
{
    result_type result = 0;
    #pragma unroll
    for (int j = 0 ; j < BLOCK_SIZE ; j++)
    {
        DIST(s_query[lidy * MAX_DESC_LEN + block_index * BLOCK_SIZE + j], s_train[j * BLOCK_SIZE + lidx]);
    }
    return result;
}

__kernel void BruteForceMatch_Match(
    __global T *query,
    __global T *train,
    __global int *bestTrainIdx,
    __global float *bestDistance,
    int query_rows,
    int query_cols,
    int train_rows,
    int train_cols,
    int step
)
{
    const int lidx = get_local_id(0);
    const int lidy = get_local_id(1);
    const int groupidx = get_group_id(0);

    const int queryIdx = mad24(BLOCK_SIZE, groupidx, lidy);
    const int queryOffset = min(queryIdx, query_rows - 1) * step;
    __global TN *query_vec = (__global TN *)(query + queryOffset);
    query_cols /= kercn;

    __local float sharebuffer[SHARED_MEM_SZ];
    __local value_type *s_query = (__local value_type *)sharebuffer;

#if 0 < MAX_DESC_LEN
    __local value_type *s_train = (__local value_type *)sharebuffer + BLOCK_SIZE * MAX_DESC_LEN;
    // load the query into local memory.
    #pragma unroll
    for (int i = 0; i < MAX_DESC_LEN / BLOCK_SIZE; i++)
    {
        const int loadx = mad24(BLOCK_SIZE, i, lidx);
        s_query[mad24(MAX_DESC_LEN, lidy, loadx)] = loadx < query_cols ? query_vec[loadx] : 0;
    }
#else
    __local value_type *s_train = (__local value_type *)sharebuffer + BLOCK_SIZE_ODD * BLOCK_SIZE;
    const int s_query_i = mad24(BLOCK_SIZE_ODD, lidy, lidx);
    const int s_train_i = mad24(BLOCK_SIZE_ODD, lidx, lidy);
#endif

    float myBestDistance = MAX_FLOAT;
    int myBestTrainIdx = -1;

    // loopUnrolledCached to find the best trainIdx and best distance.
    for (int t = 0, endt = (train_rows + BLOCK_SIZE - 1) / BLOCK_SIZE; t < endt; t++)
    {
        result_type result = 0;

        const int trainOffset = min(mad24(BLOCK_SIZE, t, lidy), train_rows - 1) * step;
        __global TN *train_vec = (__global TN *)(train + trainOffset);
#if 0 < MAX_DESC_LEN
        #pragma unroll
        for (int i = 0; i < MAX_DESC_LEN / BLOCK_SIZE; i++)
        {
            //load a BLOCK_SIZE * BLOCK_SIZE block into local train.
            const int loadx = mad24(BLOCK_SIZE, i, lidx);
            s_train[mad24(BLOCK_SIZE, lidx, lidy)] = loadx < train_cols ? train_vec[loadx] : 0;

            //synchronize to make sure each elem for reduceIteration in share memory is written already.
            barrier(CLK_LOCAL_MEM_FENCE);

            result += reduce_multi_block(s_query, s_train, i, lidx, lidy);

            barrier(CLK_LOCAL_MEM_FENCE);
        }
#else
        for (int i = 0, endq = (query_cols + BLOCK_SIZE - 1) / BLOCK_SIZE; i < endq; i++)
        {
            const int loadx = mad24(i, BLOCK_SIZE, lidx);
            //load query and train into local memory
            if (loadx < query_cols)
            {
                s_query[s_query_i] = query_vec[loadx];
                s_train[s_train_i] = train_vec[loadx];
            }
            else
            {
                s_query[s_query_i] = 0;
                s_train[s_train_i] = 0;
            }

            barrier(CLK_LOCAL_MEM_FENCE);

            result += reduce_block_match(s_query, s_train, lidx, lidy);

            barrier(CLK_LOCAL_MEM_FENCE);
        }
#endif
        result = DIST_RES(result);

        const int trainIdx = mad24(BLOCK_SIZE, t, lidx);

        if (queryIdx < query_rows && trainIdx < train_rows && result < myBestDistance /*&& mask(queryIdx, trainIdx)*/)
        {
            myBestDistance = result;
            myBestTrainIdx = trainIdx;
        }
    }

    barrier(CLK_LOCAL_MEM_FENCE);

    __local float *s_distance = (__local float *)sharebuffer;
    __local int *s_trainIdx = (__local int *)(sharebuffer + BLOCK_SIZE_ODD * BLOCK_SIZE);

    //findBestMatch
    s_distance += lidy * BLOCK_SIZE_ODD;
    s_trainIdx += lidy * BLOCK_SIZE_ODD;
    s_distance[lidx] = myBestDistance;
    s_trainIdx[lidx] = myBestTrainIdx;

    barrier(CLK_LOCAL_MEM_FENCE);

    //reduce -- now all reduce implement in each threads.
    #pragma unroll
    for (int k = 0 ; k < BLOCK_SIZE; k++)
    {
        if (myBestDistance > s_distance[k])
        {
            myBestDistance = s_distance[k];
            myBestTrainIdx = s_trainIdx[k];
        }
    }

    if (queryIdx < query_rows && lidx == 0)
    {
        bestTrainIdx[queryIdx] = myBestTrainIdx;
        bestDistance[queryIdx] = myBestDistance;
    }
}

//radius_match
__kernel void BruteForceMatch_RadiusMatch(
    __global T *query,
    __global T *train,
    float maxDistance,
    __global int *bestTrainIdx,
    __global float *bestDistance,
    __global int *nMatches,
    int query_rows,
    int query_cols,
    int train_rows,
    int train_cols,
    int bestTrainIdx_cols,
    int step,
    int ostep
)
{
    const int lidx = get_local_id(0);
    const int lidy = get_local_id(1);
    const int groupidx = get_group_id(0);
    const int groupidy = get_group_id(1);

    const int queryIdx = mad24(BLOCK_SIZE, groupidy, lidy);
    const int queryOffset = min(queryIdx, query_rows - 1) * step;
    __global TN *query_vec = (__global TN *)(query + queryOffset);

    const int trainIdx = mad24(BLOCK_SIZE, groupidx, lidx);
    const int trainOffset = min(mad24(BLOCK_SIZE, groupidx, lidy), train_rows - 1) * step;
    __global TN *train_vec = (__global TN *)(train + trainOffset);

    query_cols /= kercn;

    __local float sharebuffer[SHARED_MEM_SZ];
    __local value_type *s_query = (__local value_type *)sharebuffer;
    __local value_type *s_train = (__local value_type *)sharebuffer + BLOCK_SIZE_ODD * BLOCK_SIZE;

    result_type result = 0;
    const int s_query_i = mad24(BLOCK_SIZE_ODD, lidy, lidx);
    const int s_train_i = mad24(BLOCK_SIZE_ODD, lidx, lidy);
    for (int i = 0 ; i < (query_cols + BLOCK_SIZE - 1) / BLOCK_SIZE ; ++i)
    {
        //load a BLOCK_SIZE * BLOCK_SIZE block into local train.
        const int loadx = mad24(BLOCK_SIZE, i, lidx);

        if (loadx < query_cols)
        {
            s_query[s_query_i] = query_vec[loadx];
            s_train[s_train_i] = train_vec[loadx];
        }
        else
        {
            s_query[s_query_i] = 0;
            s_train[s_train_i] = 0;
        }

        //synchronize to make sure each elem for reduceIteration in share memory is written already.
        barrier(CLK_LOCAL_MEM_FENCE);

        result += reduce_block(s_query, s_train, lidx, lidy);

        barrier(CLK_LOCAL_MEM_FENCE);
    }
    if (queryIdx < query_rows && trainIdx < train_rows && convert_float(result) < maxDistance)
    {
        int ind = atom_inc(nMatches + queryIdx);

        if(ind < bestTrainIdx_cols)
        {
            bestTrainIdx[mad24(queryIdx, ostep, ind)] = trainIdx;
            bestDistance[mad24(queryIdx, ostep, ind)] = result;
        }
    }
}

__kernel void BruteForceMatch_knnMatch(
    __global T *query,
    __global T *train,
    __global int2 *bestTrainIdx,
    __global float2 *bestDistance,
    int query_rows,
    int query_cols,
    int train_rows,
    int train_cols,
    int step
)
{
    const int lidx = get_local_id(0);
    const int lidy = get_local_id(1);
    const int groupidx = get_group_id(0);

    const int queryIdx = mad24(BLOCK_SIZE, groupidx, lidy);
    const int queryOffset = min(queryIdx, query_rows - 1) * step;
    __global TN *query_vec = (__global TN *)(query + queryOffset);
    query_cols /= kercn;

    __local float sharebuffer[SHARED_MEM_SZ];
    __local value_type *s_query = (__local value_type *)sharebuffer;

#if 0 < MAX_DESC_LEN
    __local value_type *s_train = (__local value_type *)sharebuffer + BLOCK_SIZE * MAX_DESC_LEN;
    // load the query into local memory.
    #pragma unroll
    for (int i = 0 ;  i <  MAX_DESC_LEN / BLOCK_SIZE; i ++)
    {
        int loadx = mad24(BLOCK_SIZE, i, lidx);
        s_query[mad24(MAX_DESC_LEN, lidy, loadx)] = loadx < query_cols ? query_vec[loadx] : 0;
    }
#else
    __local value_type *s_train = (__local value_type *)sharebuffer + BLOCK_SIZE_ODD * BLOCK_SIZE;
    const int s_query_i = mad24(BLOCK_SIZE_ODD, lidy, lidx);
    const int s_train_i = mad24(BLOCK_SIZE_ODD, lidx, lidy);
#endif

    float myBestDistance1 = MAX_FLOAT;
    float myBestDistance2 = MAX_FLOAT;
    int myBestTrainIdx1 = -1;
    int myBestTrainIdx2 = -1;

    for (int t = 0, endt = (train_rows + BLOCK_SIZE - 1) / BLOCK_SIZE; t < endt ; t++)
    {
        result_type result = 0;

        int trainOffset = min(mad24(BLOCK_SIZE, t, lidy), train_rows - 1) * step;
        __global TN *train_vec = (__global TN *)(train + trainOffset);
#if 0 < MAX_DESC_LEN
        #pragma unroll
        for (int i = 0 ; i < MAX_DESC_LEN / BLOCK_SIZE ; i++)
        {
            //load a BLOCK_SIZE * BLOCK_SIZE block into local train.
            const int loadx = mad24(BLOCK_SIZE, i, lidx);
            s_train[mad24(BLOCK_SIZE, lidx, lidy)] = loadx < train_cols ? train_vec[loadx] : 0;

            //synchronize to make sure each elem for reduceIteration in share memory is written already.
            barrier(CLK_LOCAL_MEM_FENCE);

            result += reduce_multi_block(s_query, s_train, i, lidx, lidy);

            barrier(CLK_LOCAL_MEM_FENCE);
        }
#else
        for (int i = 0, endq = (query_cols + BLOCK_SIZE -1) / BLOCK_SIZE; i < endq ; i++)
        {
            const int loadx = mad24(BLOCK_SIZE, i, lidx);
            //load query and train into local memory
            if (loadx < query_cols)
            {
                s_query[s_query_i] = query_vec[loadx];
                s_train[s_train_i] = train_vec[loadx];
            }
            else
            {
                s_query[s_query_i] = 0;
                s_train[s_train_i] = 0;
            }

            barrier(CLK_LOCAL_MEM_FENCE);

            result += reduce_block_match(s_query, s_train, lidx, lidy);

            barrier(CLK_LOCAL_MEM_FENCE);
        }
#endif
        result = DIST_RES(result);

        const int trainIdx = mad24(BLOCK_SIZE, t, lidx);

        if (queryIdx < query_rows && trainIdx < train_rows)
        {
            if (result < myBestDistance1)
            {
                myBestDistance2 = myBestDistance1;
                myBestTrainIdx2 = myBestTrainIdx1;
                myBestDistance1 = result;
                myBestTrainIdx1 = trainIdx;
            }
            else if (result < myBestDistance2)
            {
                myBestDistance2 = result;
                myBestTrainIdx2 = trainIdx;
            }
        }
    }

    barrier(CLK_LOCAL_MEM_FENCE);

    __local float *s_distance = (__local float *)sharebuffer;
    __local int *s_trainIdx = (__local int *)(sharebuffer + BLOCK_SIZE_ODD * BLOCK_SIZE);

    // find BestMatch
    s_distance += lidy * BLOCK_SIZE_ODD;
    s_trainIdx += lidy * BLOCK_SIZE_ODD;
    s_distance[lidx] = myBestDistance1;
    s_trainIdx[lidx] = myBestTrainIdx1;

    float bestDistance1 = MAX_FLOAT;
    float bestDistance2 = MAX_FLOAT;
    int bestTrainIdx1 = -1;
    int bestTrainIdx2 = -1;
    barrier(CLK_LOCAL_MEM_FENCE);

    if (lidx == 0)
    {
        for (int i = 0 ; i < BLOCK_SIZE ; i++)
        {
            float val = s_distance[i];
            if (val < bestDistance1)
            {
                bestDistance2 = bestDistance1;
                bestTrainIdx2 = bestTrainIdx1;

                bestDistance1 = val;
                bestTrainIdx1 = s_trainIdx[i];
            }
            else if (val < bestDistance2)
            {
                bestDistance2 = val;
                bestTrainIdx2 = s_trainIdx[i];
            }
        }
    }

    barrier(CLK_LOCAL_MEM_FENCE);

    s_distance[lidx] = myBestDistance2;
    s_trainIdx[lidx] = myBestTrainIdx2;

    barrier(CLK_LOCAL_MEM_FENCE);

    if (lidx == 0)
    {
        for (int i = 0 ; i < BLOCK_SIZE ; i++)
        {
            float val = s_distance[i];

            if (val < bestDistance2)
            {
                bestDistance2 = val;
                bestTrainIdx2 = s_trainIdx[i];
            }
        }
    }

    myBestDistance1 = bestDistance1;
    myBestDistance2 = bestDistance2;

    myBestTrainIdx1 = bestTrainIdx1;
    myBestTrainIdx2 = bestTrainIdx2;

    if (queryIdx < query_rows && lidx == 0)
    {
        bestTrainIdx[queryIdx] = (int2)(myBestTrainIdx1, myBestTrainIdx2);
        bestDistance[queryIdx] = (float2)(myBestDistance1, myBestDistance2);
    }
}