Skip to content

O
opencv
Commit 64e9cf5d authored by yao's avatar yao
Browse files
Options

add SURF and HOG to ocl module

parent a2df4909
Hide whitespace changes
Inline Side-by-side
Showing with 2810 additions and 0 deletions
modules/ocl/include/opencv2/ocl/ocl.hpp
View file @ 64e9cf5d
......@@ -12,6 +12,7 @@
//
// Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
// Copyright (C) 2010-2012, Multicoreware, 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,
......@@ -924,6 +925,154 @@ namespace cv
const oclMat& src3, double beta, oclMat& dst, int flags = 0);
#endif
//////////////// HOG (Histogram-of-Oriented-Gradients) Descriptor and Object Detector //////////////
struct CV_EXPORTS HOGDescriptor
{
enum { DEFAULT_WIN_SIGMA = -1 };
enum { DEFAULT_NLEVELS = 64 };
enum { DESCR_FORMAT_ROW_BY_ROW, DESCR_FORMAT_COL_BY_COL };
HOGDescriptor(Size win_size=Size(64, 128), Size block_size=Size(16, 16),
Size block_stride=Size(8, 8), Size cell_size=Size(8, 8),
int nbins=9, double win_sigma=DEFAULT_WIN_SIGMA,
double threshold_L2hys=0.2, bool gamma_correction=true,
int nlevels=DEFAULT_NLEVELS);
size_t getDescriptorSize() const;
size_t getBlockHistogramSize() const;
void setSVMDetector(const vector<float>& detector);
static vector<float> getDefaultPeopleDetector();
static vector<float> getPeopleDetector48x96();
static vector<float> getPeopleDetector64x128();
void detect(const oclMat& img, vector<Point>& found_locations,
double hit_threshold=0, Size win_stride=Size(),
Size padding=Size());
void detectMultiScale(const oclMat& img, vector<Rect>& found_locations,
double hit_threshold=0, Size win_stride=Size(),
Size padding=Size(), double scale0=1.05,
int group_threshold=2);
void getDescriptors(const oclMat& img, Size win_stride,
oclMat& descriptors,
int descr_format=DESCR_FORMAT_COL_BY_COL);
Size win_size;
Size block_size;
Size block_stride;
Size cell_size;
int nbins;
double win_sigma;
double threshold_L2hys;
bool gamma_correction;
int nlevels;
protected:
void computeBlockHistograms(const oclMat& img);
void computeGradient(const oclMat& img, oclMat& grad, oclMat& qangle);
double getWinSigma() const;
bool checkDetectorSize() const;
static int numPartsWithin(int size, int part_size, int stride);
static Size numPartsWithin(Size size, Size part_size, Size stride);
// Coefficients of the separating plane
float free_coef;
oclMat detector;
// Results of the last classification step
oclMat labels;
Mat labels_host;
// Results of the last histogram evaluation step
oclMat block_hists;
// Gradients conputation results
oclMat grad, qangle;
std::vector<oclMat> image_scales;
};
//! Speeded up robust features, port from GPU module.
////////////////////////////////// SURF //////////////////////////////////////////
class CV_EXPORTS SURF_OCL
{
public:
enum KeypointLayout
{
X_ROW = 0,
Y_ROW,
LAPLACIAN_ROW,
OCTAVE_ROW,
SIZE_ROW,
ANGLE_ROW,
HESSIAN_ROW,
ROWS_COUNT
};
//! the default constructor
SURF_OCL();
//! the full constructor taking all the necessary parameters
explicit SURF_OCL(double _hessianThreshold, int _nOctaves=4,
int _nOctaveLayers=2, bool _extended=false, float _keypointsRatio=0.01f, bool _upright = false);
//! returns the descriptor size in float's (64 or 128)
int descriptorSize() const;
//! upload host keypoints to device memory
void uploadKeypoints(const vector<cv::KeyPoint>& keypoints, oclMat& keypointsocl);
//! download keypoints from device to host memory
void downloadKeypoints(const oclMat& keypointsocl, vector<KeyPoint>& keypoints);
//! download descriptors from device to host memory
void downloadDescriptors(const oclMat& descriptorsocl, vector<float>& descriptors);
//! finds the keypoints using fast hessian detector used in SURF
//! supports CV_8UC1 images
//! keypoints will have nFeature cols and 6 rows
//! keypoints.ptr<float>(X_ROW)[i] will contain x coordinate of i'th feature
//! keypoints.ptr<float>(Y_ROW)[i] will contain y coordinate of i'th feature
//! keypoints.ptr<float>(LAPLACIAN_ROW)[i] will contain laplacian sign of i'th feature
//! keypoints.ptr<float>(OCTAVE_ROW)[i] will contain octave of i'th feature
//! keypoints.ptr<float>(SIZE_ROW)[i] will contain size of i'th feature
//! keypoints.ptr<float>(ANGLE_ROW)[i] will contain orientation of i'th feature
//! keypoints.ptr<float>(HESSIAN_ROW)[i] will contain response of i'th feature
void operator()(const oclMat& img, const oclMat& mask, oclMat& keypoints);
//! finds the keypoints and computes their descriptors.
//! Optionally it can compute descriptors for the user-provided keypoints and recompute keypoints direction
void operator()(const oclMat& img, const oclMat& mask, oclMat& keypoints, oclMat& descriptors,
bool useProvidedKeypoints = false);
void operator()(const oclMat& img, const oclMat& mask, std::vector<KeyPoint>& keypoints);
void operator()(const oclMat& img, const oclMat& mask, std::vector<KeyPoint>& keypoints, oclMat& descriptors,
bool useProvidedKeypoints = false);
void operator()(const oclMat& img, const oclMat& mask, std::vector<KeyPoint>& keypoints, std::vector<float>& descriptors,
bool useProvidedKeypoints = false);
void releaseMemory();
// SURF parameters
float hessianThreshold;
int nOctaves;
int nOctaveLayers;
bool extended;
bool upright;
//! max keypoints = min(keypointsRatio * img.size().area(), 65535)
float keypointsRatio;
oclMat sum, mask1, maskSum, intBuffer;
oclMat det, trace;
oclMat maxPosBuffer;
};
}
}
#include "opencv2/ocl/matrix_operations.hpp"
......
modules/ocl/src/hog.cpp 0 → 100644
View file @ 64e9cf5d
This source diff could not be displayed because it is too large. You can view the blob instead.
modules/ocl/src/kernels/nonfree_surf.cl 0 → 100644
View file @ 64e9cf5d
/*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
// Peng Xiao, pengxiao@multicorewareinc.com
//
// 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 oclMaterials 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_amd_printf : enable
#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
// dynamically change the precision used for floating type
#if defined (__ATI__) || defined (__NVIDIA__)
#define F double
#else
#define F float
#endif
// Image read mode
__constant sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;
#define CV_PI_F 3.14159265f
// print greyscale image to show image layout
__kernel void printImage(image2d_t img)
{
printf("(%d, %d) - %3d \n",
get_global_id(0),
get_global_id(1),
read_imageui(img, (int2)(get_global_id(0), get_global_id(1))).x
);
}
// Use integral image to calculate haar wavelets.
// N = 2
// for simple haar paatern
float icvCalcHaarPatternSum_2(image2d_t sumTex, __constant float src[2][5], int oldSize, int newSize, int y, int x)
{
float ratio = (float)newSize / oldSize;
F d = 0;
#pragma unroll
for (int k = 0; k < 2; ++k)
{
int dx1 = convert_int_rte(ratio * src[k][0]);
int dy1 = convert_int_rte(ratio * src[k][1]);
int dx2 = convert_int_rte(ratio * src[k][2]);
int dy2 = convert_int_rte(ratio * src[k][3]);
F t = 0;
t += read_imageui(sumTex, sampler, (int2)(x + dx1, y + dy1)).x;
t -= read_imageui(sumTex, sampler, (int2)(x + dx1, y + dy2)).x;
t -= read_imageui(sumTex, sampler, (int2)(x + dx2, y + dy1)).x;
t += read_imageui(sumTex, sampler, (int2)(x + dx2, y + dy2)).x;
d += t * src[k][4] / ((dx2 - dx1) * (dy2 - dy1));
}
return (float)d;
}
// N = 3
float icvCalcHaarPatternSum_3(image2d_t sumTex, __constant float src[3][5], int oldSize, int newSize, int y, int x)
{
float ratio = (float)newSize / oldSize;
F d = 0;
#pragma unroll
for (int k = 0; k < 3; ++k)
{
int dx1 = convert_int_rte(ratio * src[k][0]);
int dy1 = convert_int_rte(ratio * src[k][1]);
int dx2 = convert_int_rte(ratio * src[k][2]);
int dy2 = convert_int_rte(ratio * src[k][3]);
F t = 0;
t += read_imageui(sumTex, sampler, (int2)(x + dx1, y + dy1)).x;
t -= read_imageui(sumTex, sampler, (int2)(x + dx1, y + dy2)).x;
t -= read_imageui(sumTex, sampler, (int2)(x + dx2, y + dy1)).x;
t += read_imageui(sumTex, sampler, (int2)(x + dx2, y + dy2)).x;
d += t * src[k][4] / ((dx2 - dx1) * (dy2 - dy1));
}
return (float)d;
}
// N = 4
float icvCalcHaarPatternSum_4(image2d_t sumTex, __constant float src[4][5], int oldSize, int newSize, int y, int x)
{
float ratio = (float)newSize / oldSize;
F d = 0;
#pragma unroll
for (int k = 0; k < 4; ++k)
{
int dx1 = convert_int_rte(ratio * src[k][0]);
int dy1 = convert_int_rte(ratio * src[k][1]);
int dx2 = convert_int_rte(ratio * src[k][2]);
int dy2 = convert_int_rte(ratio * src[k][3]);
F t = 0;
t += read_imageui(sumTex, sampler, (int2)(x + dx1, y + dy1)).x;
t -= read_imageui(sumTex, sampler, (int2)(x + dx1, y + dy2)).x;
t -= read_imageui(sumTex, sampler, (int2)(x + dx2, y + dy1)).x;
t += read_imageui(sumTex, sampler, (int2)(x + dx2, y + dy2)).x;
d += t * src[k][4] / ((dx2 - dx1) * (dy2 - dy1));
}
return (float)d;
}
////////////////////////////////////////////////////////////////////////
// Hessian
__constant float c_DX [3][5] = { {0, 2, 3, 7, 1}, {3, 2, 6, 7, -2}, {6, 2, 9, 7, 1} };
__constant float c_DY [3][5] = { {2, 0, 7, 3, 1}, {2, 3, 7, 6, -2}, {2, 6, 7, 9, 1} };
__constant float c_DXY[4][5] = { {1, 1, 4, 4, 1}, {5, 1, 8, 4, -1}, {1, 5, 4, 8, -1}, {5, 5, 8, 8, 1} };
__inline int calcSize(int octave, int layer)
{
/* Wavelet size at first layer of first octave. */
const int HAAR_SIZE0 = 9;
/* Wavelet size increment between layers. This should be an even number,
such that the wavelet sizes in an octave are either all even or all odd.
This ensures that when looking for the neighbours of a sample, the layers
above and below are aligned correctly. */
const int HAAR_SIZE_INC = 6;
return (HAAR_SIZE0 + HAAR_SIZE_INC * layer) << octave;
}
//calculate targeted layer per-pixel determinant and trace with an integral image
__kernel void icvCalcLayerDetAndTrace(
image2d_t sumTex, // input integral image
__global float * det, // output Determinant
__global float * trace, // output trace
int det_step, // the step of det in bytes
int trace_step, // the step of trace in bytes
int c_img_rows,
int c_img_cols,
int c_nOctaveLayers,
int c_octave,
int c_layer_rows
)
{
det_step /= sizeof(*det);
trace_step /= sizeof(*trace);
// Determine the indices
const int gridDim_y = get_num_groups(1) / (c_nOctaveLayers + 2);
const int blockIdx_y = get_group_id(1) % gridDim_y;
const int blockIdx_z = get_group_id(1) / gridDim_y;
const int j = get_local_id(0) + get_group_id(0) * get_local_size(0);
const int i = get_local_id(1) + blockIdx_y * get_local_size(1);
const int layer = blockIdx_z;
const int size = calcSize(c_octave, layer);
const int samples_i = 1 + ((c_img_rows - size) >> c_octave);
const int samples_j = 1 + ((c_img_cols - size) >> c_octave);
// Ignore pixels where some of the kernel is outside the image
const int margin = (size >> 1) >> c_octave;
if (size <= c_img_rows && size <= c_img_cols && i < samples_i && j < samples_j)
{
const float dx = icvCalcHaarPatternSum_3(sumTex, c_DX , 9, size, i << c_octave, j << c_octave);
const float dy = icvCalcHaarPatternSum_3(sumTex, c_DY , 9, size, i << c_octave, j << c_octave);
const float dxy = icvCalcHaarPatternSum_4(sumTex, c_DXY, 9, size, i << c_octave, j << c_octave);
det [j + margin + det_step * (layer * c_layer_rows + i + margin)] = dx * dy - 0.81f * dxy * dxy;
trace[j + margin + trace_step * (layer * c_layer_rows + i + margin)] = dx + dy;
}
}
////////////////////////////////////////////////////////////////////////
// NONMAX
__constant float c_DM[5] = {0, 0, 9, 9, 1};
bool within_check(image2d_t maskSumTex, int sum_i, int sum_j, int size)
{
float ratio = (float)size / 9.0f;
float d = 0;
int dx1 = convert_int_rte(ratio * c_DM[0]);
int dy1 = convert_int_rte(ratio * c_DM[1]);
int dx2 = convert_int_rte(ratio * c_DM[2]);
int dy2 = convert_int_rte(ratio * c_DM[3]);
float t = 0;
t += read_imageui(maskSumTex, sampler, (int2)(sum_j + dx1, sum_i + dy1)).x;
t -= read_imageui(maskSumTex, sampler, (int2)(sum_j + dx1, sum_i + dy2)).x;
t -= read_imageui(maskSumTex, sampler, (int2)(sum_j + dx2, sum_i + dy1)).x;
t += read_imageui(maskSumTex, sampler, (int2)(sum_j + dx2, sum_i + dy2)).x;
d += t * c_DM[4] / ((dx2 - dx1) * (dy2 - dy1));
return (d >= 0.5f);
}
// Non-maximal suppression to further filtering the candidates from previous step
__kernel
void icvFindMaximaInLayer_withmask(
__global const float * det,
__global const float * trace,
__global int4 * maxPosBuffer,
volatile __global unsigned int* maxCounter,
int counter_offset,
int det_step, // the step of det in bytes
int trace_step, // the step of trace in bytes
int c_img_rows,
int c_img_cols,
int c_nOctaveLayers,
int c_octave,
int c_layer_rows,
int c_layer_cols,
int c_max_candidates,
float c_hessianThreshold,
image2d_t maskSumTex
)
{
volatile __local float N9[768]; // threads.x * threads.y * 3
det_step /= sizeof(*det);
trace_step /= sizeof(*trace);
maxCounter += counter_offset;
// Determine the indices
const int gridDim_y = get_num_groups(1) / c_nOctaveLayers;
const int blockIdx_y = get_group_id(1) % gridDim_y;
const int blockIdx_z = get_group_id(1) / gridDim_y;
const int layer = blockIdx_z + 1;
const int size = calcSize(c_octave, layer);
// Ignore pixels without a 3x3x3 neighbourhood in the layer above
const int margin = ((calcSize(c_octave, layer + 1) >> 1) >> c_octave) + 1;
const int j = get_local_id(0) + get_group_id(0) * (get_local_size(0) - 2) + margin - 1;
const int i = get_local_id(1) + blockIdx_y * (get_local_size(1) - 2) + margin - 1;
// Is this thread within the hessian buffer?
const int zoff = get_local_size(0) * get_local_size(1);
const int localLin = get_local_id(0) + get_local_id(1) * get_local_size(0) + zoff;
N9[localLin - zoff] =
det[det_step *
(c_layer_rows * (layer - 1) + min(max(i, 0), c_img_rows - 1)) // y
+ min(max(j, 0), c_img_cols - 1)]; // x
N9[localLin ] =
det[det_step *
(c_layer_rows * (layer ) + min(max(i, 0), c_img_rows - 1)) // y
+ min(max(j, 0), c_img_cols - 1)]; // x
N9[localLin + zoff] =
det[det_step *
(c_layer_rows * (layer + 1) + min(max(i, 0), c_img_rows - 1)) // y
+ min(max(j, 0), c_img_cols - 1)]; // x
barrier(CLK_LOCAL_MEM_FENCE);
if (i < c_layer_rows - margin
&& j < c_layer_cols - margin
&& get_local_id(0) > 0
&& get_local_id(0) < get_local_size(0) - 1
&& get_local_id(1) > 0
&& get_local_id(1) < get_local_size(1) - 1 // these are unnecessary conditions ported from CUDA
)
{
float val0 = N9[localLin];
if (val0 > c_hessianThreshold)
{
// Coordinates for the start of the wavelet in the sum image. There
// is some integer division involved, so don't try to simplify this
// (cancel out sampleStep) without checking the result is the same
const int sum_i = (i - ((size >> 1) >> c_octave)) << c_octave;
const int sum_j = (j - ((size >> 1) >> c_octave)) << c_octave;
if (within_check(maskSumTex, sum_i, sum_j, size))
{
// Check to see if we have a max (in its 26 neighbours)
const bool condmax = val0 > N9[localLin - 1 - get_local_size(0) - zoff]
&& val0 > N9[localLin - get_local_size(0) - zoff]
&& val0 > N9[localLin + 1 - get_local_size(0) - zoff]
&& val0 > N9[localLin - 1 - zoff]
&& val0 > N9[localLin - zoff]
&& val0 > N9[localLin + 1 - zoff]
&& val0 > N9[localLin - 1 + get_local_size(0) - zoff]
&& val0 > N9[localLin + get_local_size(0) - zoff]
&& val0 > N9[localLin + 1 + get_local_size(0) - zoff]
&& val0 > N9[localLin - 1 - get_local_size(0)]
&& val0 > N9[localLin - get_local_size(0)]
&& val0 > N9[localLin + 1 - get_local_size(0)]
&& val0 > N9[localLin - 1 ]
&& val0 > N9[localLin + 1 ]
&& val0 > N9[localLin - 1 + get_local_size(0)]
&& val0 > N9[localLin + get_local_size(0)]
&& val0 > N9[localLin + 1 + get_local_size(0)]
&& val0 > N9[localLin - 1 - get_local_size(0) + zoff]
&& val0 > N9[localLin - get_local_size(0) + zoff]
&& val0 > N9[localLin + 1 - get_local_size(0) + zoff]
&& val0 > N9[localLin - 1 + zoff]
&& val0 > N9[localLin + zoff]
&& val0 > N9[localLin + 1 + zoff]
&& val0 > N9[localLin - 1 + get_local_size(0) + zoff]
&& val0 > N9[localLin + get_local_size(0) + zoff]
&& val0 > N9[localLin + 1 + get_local_size(0) + zoff]
;
if(condmax)
{
unsigned int ind = atomic_inc(maxCounter);
if (ind < c_max_candidates)
{
const int laplacian = (int) copysign(1.0f, trace[trace_step* (layer * c_layer_rows + i) + j]);
maxPosBuffer[ind] = (int4)(j, i, layer, laplacian);
}
}
}
}
}
}
__kernel
void icvFindMaximaInLayer(
__global float * det,
__global float * trace,
__global int4 * maxPosBuffer,
volatile __global unsigned int* maxCounter,
int counter_offset,
int det_step, // the step of det in bytes
int trace_step, // the step of trace in bytes
int c_img_rows,
int c_img_cols,
int c_nOctaveLayers,
int c_octave,
int c_layer_rows,
int c_layer_cols,
int c_max_candidates,
float c_hessianThreshold
)
{
volatile __local float N9[768]; // threads.x * threads.y * 3
det_step /= sizeof(float);
trace_step /= sizeof(float);
maxCounter += counter_offset;
// Determine the indices
const int gridDim_y = get_num_groups(1) / c_nOctaveLayers;
const int blockIdx_y = get_group_id(1) % gridDim_y;
const int blockIdx_z = get_group_id(1) / gridDim_y;
const int layer = blockIdx_z + 1;
const int size = calcSize(c_octave, layer);
// Ignore pixels without a 3x3x3 neighbourhood in the layer above
const int margin = ((calcSize(c_octave, layer + 1) >> 1) >> c_octave) + 1;
const int j = get_local_id(0) + get_group_id(0) * (get_local_size(0) - 2) + margin - 1;
const int i = get_local_id(1) + blockIdx_y * (get_local_size(1) - 2) + margin - 1;
// Is this thread within the hessian buffer?
const int zoff = get_local_size(0) * get_local_size(1);
const int localLin = get_local_id(0) + get_local_id(1) * get_local_size(0) + zoff;
int l_x = min(max(j, 0), c_img_cols - 1);
int l_y = c_layer_rows * layer + min(max(i, 0), c_img_rows - 1);
N9[localLin - zoff] =
det[det_step * (l_y - c_layer_rows) + l_x];
N9[localLin ] =
det[det_step * (l_y ) + l_x];
N9[localLin + zoff] =
det[det_step * (l_y + c_layer_rows) + l_x];
barrier(CLK_LOCAL_MEM_FENCE);
if (i < c_layer_rows - margin
&& j < c_layer_cols - margin
&& get_local_id(0) > 0
&& get_local_id(0) < get_local_size(0) - 1
&& get_local_id(1) > 0
&& get_local_id(1) < get_local_size(1) - 1 // these are unnecessary conditions ported from CUDA
)
{
float val0 = N9[localLin];
if (val0 > c_hessianThreshold)
{
//printf(\"(%3d, %3d) N9[%3d]=%7.1f val0=%7.1f\\n\", l_x, l_y, localLin - zoff, N9[localLin], val0);
// Coordinates for the start of the wavelet in the sum image. There
// is some integer division involved, so don't try to simplify this
// (cancel out sampleStep) without checking the result is the same
// Check to see if we have a max (in its 26 neighbours)
const bool condmax = val0 > N9[localLin - 1 - get_local_size(0) - zoff]
&& val0 > N9[localLin - get_local_size(0) - zoff]
&& val0 > N9[localLin + 1 - get_local_size(0) - zoff]
&& val0 > N9[localLin - 1 - zoff]
&& val0 > N9[localLin - zoff]
&& val0 > N9[localLin + 1 - zoff]
&& val0 > N9[localLin - 1 + get_local_size(0) - zoff]
&& val0 > N9[localLin + get_local_size(0) - zoff]
&& val0 > N9[localLin + 1 + get_local_size(0) - zoff]
&& val0 > N9[localLin - 1 - get_local_size(0)]
&& val0 > N9[localLin - get_local_size(0)]
&& val0 > N9[localLin + 1 - get_local_size(0)]
&& val0 > N9[localLin - 1 ]
&& val0 > N9[localLin + 1 ]
&& val0 > N9[localLin - 1 + get_local_size(0)]
&& val0 > N9[localLin + get_local_size(0)]
&& val0 > N9[localLin + 1 + get_local_size(0)]
&& val0 > N9[localLin - 1 - get_local_size(0) + zoff]
&& val0 > N9[localLin - get_local_size(0) + zoff]
&& val0 > N9[localLin + 1 - get_local_size(0) + zoff]
&& val0 > N9[localLin - 1 + zoff]
&& val0 > N9[localLin + zoff]
&& val0 > N9[localLin + 1 + zoff]
&& val0 > N9[localLin - 1 + get_local_size(0) + zoff]
&& val0 > N9[localLin + get_local_size(0) + zoff]
&& val0 > N9[localLin + 1 + get_local_size(0) + zoff]
;
if(condmax)
{
unsigned int ind = atomic_inc(maxCounter);
if (ind < c_max_candidates)
{
const int laplacian = (int) copysign(1.0f, trace[trace_step* (layer * c_layer_rows + i) + j]);
maxPosBuffer[ind] = (int4)(j, i, layer, laplacian);
}
}
}
}
}
// solve 3x3 linear system Ax=b for floating point input
inline bool solve3x3_float(volatile __local const float A[3][3], volatile __local const float b[3], volatile __local float x[3])
{
float det = A[0][0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1])
- A[0][1] * (A[1][0] * A[2][2] - A[1][2] * A[2][0])
+ A[0][2] * (A[1][0] * A[2][1] - A[1][1] * A[2][0]);
if (det != 0)
{
F invdet = 1.0 / det;
x[0] = invdet *
(b[0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1]) -
A[0][1] * (b[1] * A[2][2] - A[1][2] * b[2] ) +
A[0][2] * (b[1] * A[2][1] - A[1][1] * b[2] ));
x[1] = invdet *
(A[0][0] * (b[1] * A[2][2] - A[1][2] * b[2] ) -
b[0] * (A[1][0] * A[2][2] - A[1][2] * A[2][0]) +
A[0][2] * (A[1][0] * b[2] - b[1] * A[2][0]));
x[2] = invdet *
(A[0][0] * (A[1][1] * b[2] - b[1] * A[2][1]) -
A[0][1] * (A[1][0] * b[2] - b[1] * A[2][0]) +
b[0] * (A[1][0] * A[2][1] - A[1][1] * A[2][0]));
return true;
}
return false;
}
#define X_ROW 0
#define Y_ROW 1
#define LAPLACIAN_ROW 2
#define OCTAVE_ROW 3
#define SIZE_ROW 4
#define ANGLE_ROW 5
#define HESSIAN_ROW 6
#define ROWS_COUNT 7
////////////////////////////////////////////////////////////////////////
// INTERPOLATION
__kernel
void icvInterpolateKeypoint(
__global const float * det,
__global const int4 * maxPosBuffer,
__global float * keypoints,
volatile __global unsigned int * featureCounter,
int det_step,
int keypoints_step,
int c_img_rows,
int c_img_cols,
int c_octave,
int c_layer_rows,
int c_max_features
)
{
det_step /= sizeof(*det);
keypoints_step /= sizeof(*keypoints);
__global float * featureX = keypoints + X_ROW * keypoints_step;
__global float * featureY = keypoints + Y_ROW * keypoints_step;
__global int * featureLaplacian = (__global int *)keypoints + LAPLACIAN_ROW * keypoints_step;
__global int * featureOctave = (__global int *)keypoints + OCTAVE_ROW * keypoints_step;
__global float * featureSize = keypoints + SIZE_ROW * keypoints_step;
__global float * featureHessian = keypoints + HESSIAN_ROW * keypoints_step;
const int4 maxPos = maxPosBuffer[get_group_id(0)];
const int j = maxPos.x - 1 + get_local_id(0);
const int i = maxPos.y - 1 + get_local_id(1);
const int layer = maxPos.z - 1 + get_local_id(2);
volatile __local float N9[3][3][3];
N9[get_local_id(2)][get_local_id(1)][get_local_id(0)] =
det[det_step * (c_layer_rows * layer + i) + j];
barrier(CLK_LOCAL_MEM_FENCE);
if (get_local_id(0) == 0 && get_local_id(1) == 0 && get_local_id(2) == 0)
{
volatile __local float dD[3];
//dx
dD[0] = -0.5f * (N9[1][1][2] - N9[1][1][0]);
//dy
dD[1] = -0.5f * (N9[1][2][1] - N9[1][0][1]);
//ds
dD[2] = -0.5f * (N9[2][1][1] - N9[0][1][1]);
volatile __local float H[3][3];
//dxx
H[0][0] = N9[1][1][0] - 2.0f * N9[1][1][1] + N9[1][1][2];
//dxy
H[0][1]= 0.25f * (N9[1][2][2] - N9[1][2][0] - N9[1][0][2] + N9[1][0][0]);
//dxs
H[0][2]= 0.25f * (N9[2][1][2] - N9[2][1][0] - N9[0][1][2] + N9[0][1][0]);
//dyx = dxy
H[1][0] = H[0][1];
//dyy
H[1][1] = N9[1][0][1] - 2.0f * N9[1][1][1] + N9[1][2][1];
//dys
H[1][2]= 0.25f * (N9[2][2][1] - N9[2][0][1] - N9[0][2][1] + N9[0][0][1]);
//dsx = dxs
H[2][0] = H[0][2];
//dsy = dys
H[2][1] = H[1][2];
//dss
H[2][2] = N9[0][1][1] - 2.0f * N9[1][1][1] + N9[2][1][1];
volatile __local float x[3];
if (solve3x3_float(H, dD, x))
{
if (fabs(x[0]) <= 1.f && fabs(x[1]) <= 1.f && fabs(x[2]) <= 1.f)
{
// if the step is within the interpolation region, perform it
const int size = calcSize(c_octave, maxPos.z);
const int sum_i = (maxPos.y - ((size >> 1) >> c_octave)) << c_octave;
const int sum_j = (maxPos.x - ((size >> 1) >> c_octave)) << c_octave;
const float center_i = sum_i + (float)(size - 1) / 2;
const float center_j = sum_j + (float)(size - 1) / 2;
const float px = center_j + x[0] * (1 << c_octave);
const float py = center_i + x[1] * (1 << c_octave);
const int ds = size - calcSize(c_octave, maxPos.z - 1);
const float psize = round(size + x[2] * ds);
/* The sampling intervals and wavelet sized for selecting an orientation
and building the keypoint descriptor are defined relative to 's' */
const float s = psize * 1.2f / 9.0f;
/* To find the dominant orientation, the gradients in x and y are
sampled in a circle of radius 6s using wavelets of size 4s.
We ensure the gradient wavelet size is even to ensure the
wavelet pattern is balanced and symmetric around its center */
const int grad_wav_size = 2 * convert_int_rte(2.0f * s);
// check when grad_wav_size is too big
if ((c_img_rows + 1) >= grad_wav_size && (c_img_cols + 1) >= grad_wav_size)
{
// Get a new feature index.
unsigned int ind = atomic_inc(featureCounter);
if (ind < c_max_features)
{
featureX[ind] = px;
featureY[ind] = py;
featureLaplacian[ind] = maxPos.w;
featureOctave[ind] = c_octave;
featureSize[ind] = psize;
featureHessian[ind] = N9[1][1][1];
}
} // grad_wav_size check
} // If the subpixel interpolation worked
}
} // If this is thread 0.
}
////////////////////////////////////////////////////////////////////////
// Orientation
#define ORI_SEARCH_INC 5
#define ORI_WIN 60
#define ORI_SAMPLES 113
__constant float c_aptX[ORI_SAMPLES] = {-6, -5, -5, -5, -5, -5, -5, -5, -4, -4, -4, -4, -4, -4, -4, -4, -4, -3, -3, -3, -3, -3, -3, -3, -3, -3, -3, -3, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -2, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 6};
__constant float c_aptY[ORI_SAMPLES] = {0, -3, -2, -1, 0, 1, 2, 3, -4, -3, -2, -1, 0, 1, 2, 3, 4, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, -6, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5, -4, -3, -2, -1, 0, 1, 2, 3, 4, -3, -2, -1, 0, 1, 2, 3, 0};
__constant float c_aptW[ORI_SAMPLES] = {0.001455130288377404f, 0.001707611023448408f, 0.002547456417232752f, 0.003238451667129993f, 0.0035081731621176f,
0.003238451667129993f, 0.002547456417232752f, 0.001707611023448408f, 0.002003900473937392f, 0.0035081731621176f, 0.005233579315245152f,
0.00665318313986063f, 0.00720730796456337f, 0.00665318313986063f, 0.005233579315245152f, 0.0035081731621176f,
0.002003900473937392f, 0.001707611023448408f, 0.0035081731621176f, 0.006141661666333675f, 0.009162282571196556f,
0.01164754293859005f, 0.01261763460934162f, 0.01164754293859005f, 0.009162282571196556f, 0.006141661666333675f,
0.0035081731621176f, 0.001707611023448408f, 0.002547456417232752f, 0.005233579315245152f, 0.009162282571196556f,
0.01366852037608624f, 0.01737609319388866f, 0.0188232995569706f, 0.01737609319388866f, 0.01366852037608624f,
0.009162282571196556f, 0.005233579315245152f, 0.002547456417232752f, 0.003238451667129993f, 0.00665318313986063f,
0.01164754293859005f, 0.01737609319388866f, 0.02208934165537357f, 0.02392910048365593f, 0.02208934165537357f,
0.01737609319388866f, 0.01164754293859005f, 0.00665318313986063f, 0.003238451667129993f, 0.001455130288377404f,
0.0035081731621176f, 0.00720730796456337f, 0.01261763460934162f, 0.0188232995569706f, 0.02392910048365593f,
0.02592208795249462f, 0.02392910048365593f, 0.0188232995569706f, 0.01261763460934162f, 0.00720730796456337f,
0.0035081731621176f, 0.001455130288377404f, 0.003238451667129993f, 0.00665318313986063f, 0.01164754293859005f,
0.01737609319388866f, 0.02208934165537357f, 0.02392910048365593f, 0.02208934165537357f, 0.01737609319388866f,
0.01164754293859005f, 0.00665318313986063f, 0.003238451667129993f, 0.002547456417232752f, 0.005233579315245152f,
0.009162282571196556f, 0.01366852037608624f, 0.01737609319388866f, 0.0188232995569706f, 0.01737609319388866f,
0.01366852037608624f, 0.009162282571196556f, 0.005233579315245152f, 0.002547456417232752f, 0.001707611023448408f,
0.0035081731621176f, 0.006141661666333675f, 0.009162282571196556f, 0.01164754293859005f, 0.01261763460934162f,
0.01164754293859005f, 0.009162282571196556f, 0.006141661666333675f, 0.0035081731621176f, 0.001707611023448408f,
0.002003900473937392f, 0.0035081731621176f, 0.005233579315245152f, 0.00665318313986063f, 0.00720730796456337f,
0.00665318313986063f, 0.005233579315245152f, 0.0035081731621176f, 0.002003900473937392f, 0.001707611023448408f,
0.002547456417232752f, 0.003238451667129993f, 0.0035081731621176f, 0.003238451667129993f, 0.002547456417232752f,
0.001707611023448408f, 0.001455130288377404f};
__constant float c_NX[2][5] = {{0, 0, 2, 4, -1}, {2, 0, 4, 4, 1}};
__constant float c_NY[2][5] = {{0, 0, 4, 2, 1}, {0, 2, 4, 4, -1}};
void reduce_32_sum(volatile __local float * data, float partial_reduction, int tid)
{
#define op(A, B) (A)+(B)
data[tid] = partial_reduction;
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 16)
{
data[tid] = partial_reduction = op(partial_reduction, data[tid + 16]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 8 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 4 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 2 ]);
data[tid] = partial_reduction = op(partial_reduction, data[tid + 1 ]);
}
#undef op
}
__kernel
void icvCalcOrientation(
image2d_t sumTex,
__global float * keypoints,
int keypoints_step,
int c_img_rows,
int c_img_cols
)
{
keypoints_step /= sizeof(*keypoints);
__global float* featureX = keypoints + X_ROW * keypoints_step;
__global float* featureY = keypoints + Y_ROW * keypoints_step;
__global float* featureSize = keypoints + SIZE_ROW * keypoints_step;
__global float* featureDir = keypoints + ANGLE_ROW * keypoints_step;
volatile __local float s_X[128];
volatile __local float s_Y[128];
volatile __local float s_angle[128];
volatile __local float s_sumx[32 * 4];
volatile __local float s_sumy[32 * 4];
/* The sampling intervals and wavelet sized for selecting an orientation
and building the keypoint descriptor are defined relative to 's' */
const float s = featureSize[get_group_id(0)] * 1.2f / 9.0f;
/* To find the dominant orientation, the gradients in x and y are
sampled in a circle of radius 6s using wavelets of size 4s.
We ensure the gradient wavelet size is even to ensure the
wavelet pattern is balanced and symmetric around its center */
const int grad_wav_size = 2 * convert_int_rte(2.0f * s);
// check when grad_wav_size is too big
if ((c_img_rows + 1) < grad_wav_size || (c_img_cols + 1) < grad_wav_size)
return;
// Calc X, Y, angle and store it to shared memory
const int tid = get_local_id(1) * get_local_size(0) + get_local_id(0);
float X = 0.0f, Y = 0.0f, angle = 0.0f;
if (tid < ORI_SAMPLES)
{
const float margin = (float)(grad_wav_size - 1) / 2.0f;
const int x = convert_int_rte(featureX[get_group_id(0)] + c_aptX[tid] * s - margin);
const int y = convert_int_rte(featureY[get_group_id(0)] + c_aptY[tid] * s - margin);
if (y >= 0 && y < (c_img_rows + 1) - grad_wav_size &&
x >= 0 && x < (c_img_cols + 1) - grad_wav_size)
{
X = c_aptW[tid] * icvCalcHaarPatternSum_2(sumTex, c_NX, 4, grad_wav_size, y, x);
Y = c_aptW[tid] * icvCalcHaarPatternSum_2(sumTex, c_NY, 4, grad_wav_size, y, x);
angle = atan2(Y, X);
if (angle < 0)
angle += 2.0f * CV_PI_F;
angle *= 180.0f / CV_PI_F;
}
}
s_X[tid] = X;
s_Y[tid] = Y;
s_angle[tid] = angle;
barrier(CLK_LOCAL_MEM_FENCE);
float bestx = 0, besty = 0, best_mod = 0;
#pragma unroll
for (int i = 0; i < 18; ++i)
{
const int dir = (i * 4 + get_local_id(1)) * ORI_SEARCH_INC;
float sumx = 0.0f, sumy = 0.0f;
int d = abs(convert_int_rte(s_angle[get_local_id(0)]) - dir);
if (d < ORI_WIN / 2 || d > 360 - ORI_WIN / 2)
{
sumx = s_X[get_local_id(0)];
sumy = s_Y[get_local_id(0)];
}
d = abs(convert_int_rte(s_angle[get_local_id(0) + 32]) - dir);
if (d < ORI_WIN / 2 || d > 360 - ORI_WIN / 2)
{
sumx += s_X[get_local_id(0) + 32];
sumy += s_Y[get_local_id(0) + 32];
}
d = abs(convert_int_rte(s_angle[get_local_id(0) + 64]) - dir);
if (d < ORI_WIN / 2 || d > 360 - ORI_WIN / 2)
{
sumx += s_X[get_local_id(0) + 64];
sumy += s_Y[get_local_id(0) + 64];
}
d = abs(convert_int_rte(s_angle[get_local_id(0) + 96]) - dir);
if (d < ORI_WIN / 2 || d > 360 - ORI_WIN / 2)
{
sumx += s_X[get_local_id(0) + 96];
sumy += s_Y[get_local_id(0) + 96];
}
reduce_32_sum(s_sumx + get_local_id(1) * 32, sumx, get_local_id(0));
reduce_32_sum(s_sumy + get_local_id(1) * 32, sumy, get_local_id(0));
const float temp_mod = sumx * sumx + sumy * sumy;
if (temp_mod > best_mod)
{
best_mod = temp_mod;
bestx = sumx;
besty = sumy;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (get_local_id(0) == 0)
{
s_X[get_local_id(1)] = bestx;
s_Y[get_local_id(1)] = besty;
s_angle[get_local_id(1)] = best_mod;
}
barrier(CLK_LOCAL_MEM_FENCE);
if (get_local_id(1) == 0 && get_local_id(0) == 0)
{
int bestIdx = 0;
if (s_angle[1] > s_angle[bestIdx])
bestIdx = 1;
if (s_angle[2] > s_angle[bestIdx])
bestIdx = 2;
if (s_angle[3] > s_angle[bestIdx])
bestIdx = 3;
float kp_dir = atan2(s_Y[bestIdx], s_X[bestIdx]);
if (kp_dir < 0)
kp_dir += 2.0f * CV_PI_F;
kp_dir *= 180.0f / CV_PI_F;
featureDir[get_group_id(0)] = kp_dir;
}
}
#undef ORI_SEARCH_INC
#undef ORI_WIN
#undef ORI_SAMPLES
////////////////////////////////////////////////////////////////////////
// Descriptors
#define PATCH_SZ 20
__constant float c_DW[PATCH_SZ * PATCH_SZ] =
{
3.695352233989979e-006f, 8.444558261544444e-006f, 1.760426494001877e-005f, 3.34794785885606e-005f, 5.808438800158911e-005f, 9.193058212986216e-005f, 0.0001327334757661447f, 0.0001748319627949968f, 0.0002100782439811155f, 0.0002302826324012131f, 0.0002302826324012131f, 0.0002100782439811155f, 0.0001748319627949968f, 0.0001327334757661447f, 9.193058212986216e-005f, 5.808438800158911e-005f, 3.34794785885606e-005f, 1.760426494001877e-005f, 8.444558261544444e-006f, 3.695352233989979e-006f,
8.444558261544444e-006f, 1.929736572492402e-005f, 4.022897701361217e-005f, 7.650675252079964e-005f, 0.0001327334903180599f, 0.0002100782585330308f, 0.0003033203829545528f, 0.0003995231236331165f, 0.0004800673632416874f, 0.0005262381164357066f, 0.0005262381164357066f, 0.0004800673632416874f, 0.0003995231236331165f, 0.0003033203829545528f, 0.0002100782585330308f, 0.0001327334903180599f, 7.650675252079964e-005f, 4.022897701361217e-005f, 1.929736572492402e-005f, 8.444558261544444e-006f,
1.760426494001877e-005f, 4.022897701361217e-005f, 8.386484114453197e-005f, 0.0001594926579855382f, 0.0002767078403849155f, 0.0004379475140012801f, 0.0006323281559161842f, 0.0008328808471560478f, 0.001000790391117334f, 0.001097041997127235f, 0.001097041997127235f, 0.001000790391117334f, 0.0008328808471560478f, 0.0006323281559161842f, 0.0004379475140012801f, 0.0002767078403849155f, 0.0001594926579855382f, 8.386484114453197e-005f, 4.022897701361217e-005f, 1.760426494001877e-005f,
3.34794785885606e-005f, 7.650675252079964e-005f, 0.0001594926579855382f, 0.0003033203247468919f, 0.0005262380582280457f, 0.0008328807889483869f, 0.001202550483867526f, 0.001583957928232849f, 0.001903285388834775f, 0.002086334861814976f, 0.002086334861814976f, 0.001903285388834775f, 0.001583957928232849f, 0.001202550483867526f, 0.0008328807889483869f, 0.0005262380582280457f, 0.0003033203247468919f, 0.0001594926579855382f, 7.650675252079964e-005f, 3.34794785885606e-005f,
5.808438800158911e-005f, 0.0001327334903180599f, 0.0002767078403849155f, 0.0005262380582280457f, 0.0009129836107604206f, 0.001444985857233405f, 0.002086335094645619f, 0.002748048631474376f, 0.00330205773934722f, 0.003619635012000799f, 0.003619635012000799f, 0.00330205773934722f, 0.002748048631474376f, 0.002086335094645619f, 0.001444985857233405f, 0.0009129836107604206f, 0.0005262380582280457f, 0.0002767078403849155f, 0.0001327334903180599f, 5.808438800158911e-005f,
9.193058212986216e-005f, 0.0002100782585330308f, 0.0004379475140012801f, 0.0008328807889483869f, 0.001444985857233405f, 0.002286989474669099f, 0.00330205773934722f, 0.004349356517195702f, 0.00522619066759944f, 0.005728822201490402f, 0.005728822201490402f, 0.00522619066759944f, 0.004349356517195702f, 0.00330205773934722f, 0.002286989474669099f, 0.001444985857233405f, 0.0008328807889483869f, 0.0004379475140012801f, 0.0002100782585330308f, 9.193058212986216e-005f,
0.0001327334757661447f, 0.0003033203829545528f, 0.0006323281559161842f, 0.001202550483867526f, 0.002086335094645619f, 0.00330205773934722f, 0.004767658654600382f, 0.006279794964939356f, 0.007545807864516974f, 0.008271530270576477f, 0.008271530270576477f, 0.007545807864516974f, 0.006279794964939356f, 0.004767658654600382f, 0.00330205773934722f, 0.002086335094645619f, 0.001202550483867526f, 0.0006323281559161842f, 0.0003033203829545528f, 0.0001327334757661447f,
0.0001748319627949968f, 0.0003995231236331165f, 0.0008328808471560478f, 0.001583957928232849f, 0.002748048631474376f, 0.004349356517195702f, 0.006279794964939356f, 0.008271529339253902f, 0.009939077310264111f, 0.01089497376233339f, 0.01089497376233339f, 0.009939077310264111f, 0.008271529339253902f, 0.006279794964939356f, 0.004349356517195702f, 0.002748048631474376f, 0.001583957928232849f, 0.0008328808471560478f, 0.0003995231236331165f, 0.0001748319627949968f,
0.0002100782439811155f, 0.0004800673632416874f, 0.001000790391117334f, 0.001903285388834775f, 0.00330205773934722f, 0.00522619066759944f, 0.007545807864516974f, 0.009939077310264111f, 0.01194280479103327f, 0.01309141051024199f, 0.01309141051024199f, 0.01194280479103327f, 0.009939077310264111f, 0.007545807864516974f, 0.00522619066759944f, 0.00330205773934722f, 0.001903285388834775f, 0.001000790391117334f, 0.0004800673632416874f, 0.0002100782439811155f,
0.0002302826324012131f, 0.0005262381164357066f, 0.001097041997127235f, 0.002086334861814976f, 0.003619635012000799f, 0.005728822201490402f, 0.008271530270576477f, 0.01089497376233339f, 0.01309141051024199f, 0.01435048412531614f, 0.01435048412531614f, 0.01309141051024199f, 0.01089497376233339f, 0.008271530270576477f, 0.005728822201490402f, 0.003619635012000799f, 0.002086334861814976f, 0.001097041997127235f, 0.0005262381164357066f, 0.0002302826324012131f,
0.0002302826324012131f, 0.0005262381164357066f, 0.001097041997127235f, 0.002086334861814976f, 0.003619635012000799f, 0.005728822201490402f, 0.008271530270576477f, 0.01089497376233339f, 0.01309141051024199f, 0.01435048412531614f, 0.01435048412531614f, 0.01309141051024199f, 0.01089497376233339f, 0.008271530270576477f, 0.005728822201490402f, 0.003619635012000799f, 0.002086334861814976f, 0.001097041997127235f, 0.0005262381164357066f, 0.0002302826324012131f,
0.0002100782439811155f, 0.0004800673632416874f, 0.001000790391117334f, 0.001903285388834775f, 0.00330205773934722f, 0.00522619066759944f, 0.007545807864516974f, 0.009939077310264111f, 0.01194280479103327f, 0.01309141051024199f, 0.01309141051024199f, 0.01194280479103327f, 0.009939077310264111f, 0.007545807864516974f, 0.00522619066759944f, 0.00330205773934722f, 0.001903285388834775f, 0.001000790391117334f, 0.0004800673632416874f, 0.0002100782439811155f,
0.0001748319627949968f, 0.0003995231236331165f, 0.0008328808471560478f, 0.001583957928232849f, 0.002748048631474376f, 0.004349356517195702f, 0.006279794964939356f, 0.008271529339253902f, 0.009939077310264111f, 0.01089497376233339f, 0.01089497376233339f, 0.009939077310264111f, 0.008271529339253902f, 0.006279794964939356f, 0.004349356517195702f, 0.002748048631474376f, 0.001583957928232849f, 0.0008328808471560478f, 0.0003995231236331165f, 0.0001748319627949968f,
0.0001327334757661447f, 0.0003033203829545528f, 0.0006323281559161842f, 0.001202550483867526f, 0.002086335094645619f, 0.00330205773934722f, 0.004767658654600382f, 0.006279794964939356f, 0.007545807864516974f, 0.008271530270576477f, 0.008271530270576477f, 0.007545807864516974f, 0.006279794964939356f, 0.004767658654600382f, 0.00330205773934722f, 0.002086335094645619f, 0.001202550483867526f, 0.0006323281559161842f, 0.0003033203829545528f, 0.0001327334757661447f,
9.193058212986216e-005f, 0.0002100782585330308f, 0.0004379475140012801f, 0.0008328807889483869f, 0.001444985857233405f, 0.002286989474669099f, 0.00330205773934722f, 0.004349356517195702f, 0.00522619066759944f, 0.005728822201490402f, 0.005728822201490402f, 0.00522619066759944f, 0.004349356517195702f, 0.00330205773934722f, 0.002286989474669099f, 0.001444985857233405f, 0.0008328807889483869f, 0.0004379475140012801f, 0.0002100782585330308f, 9.193058212986216e-005f,
5.808438800158911e-005f, 0.0001327334903180599f, 0.0002767078403849155f, 0.0005262380582280457f, 0.0009129836107604206f, 0.001444985857233405f, 0.002086335094645619f, 0.002748048631474376f, 0.00330205773934722f, 0.003619635012000799f, 0.003619635012000799f, 0.00330205773934722f, 0.002748048631474376f, 0.002086335094645619f, 0.001444985857233405f, 0.0009129836107604206f, 0.0005262380582280457f, 0.0002767078403849155f, 0.0001327334903180599f, 5.808438800158911e-005f,
3.34794785885606e-005f, 7.650675252079964e-005f, 0.0001594926579855382f, 0.0003033203247468919f, 0.0005262380582280457f, 0.0008328807889483869f, 0.001202550483867526f, 0.001583957928232849f, 0.001903285388834775f, 0.002086334861814976f, 0.002086334861814976f, 0.001903285388834775f, 0.001583957928232849f, 0.001202550483867526f, 0.0008328807889483869f, 0.0005262380582280457f, 0.0003033203247468919f, 0.0001594926579855382f, 7.650675252079964e-005f, 3.34794785885606e-005f,
1.760426494001877e-005f, 4.022897701361217e-005f, 8.386484114453197e-005f, 0.0001594926579855382f, 0.0002767078403849155f, 0.0004379475140012801f, 0.0006323281559161842f, 0.0008328808471560478f, 0.001000790391117334f, 0.001097041997127235f, 0.001097041997127235f, 0.001000790391117334f, 0.0008328808471560478f, 0.0006323281559161842f, 0.0004379475140012801f, 0.0002767078403849155f, 0.0001594926579855382f, 8.386484114453197e-005f, 4.022897701361217e-005f, 1.760426494001877e-005f,
8.444558261544444e-006f, 1.929736572492402e-005f, 4.022897701361217e-005f, 7.650675252079964e-005f, 0.0001327334903180599f, 0.0002100782585330308f, 0.0003033203829545528f, 0.0003995231236331165f, 0.0004800673632416874f, 0.0005262381164357066f, 0.0005262381164357066f, 0.0004800673632416874f, 0.0003995231236331165f, 0.0003033203829545528f, 0.0002100782585330308f, 0.0001327334903180599f, 7.650675252079964e-005f, 4.022897701361217e-005f, 1.929736572492402e-005f, 8.444558261544444e-006f,
3.695352233989979e-006f, 8.444558261544444e-006f, 1.760426494001877e-005f, 3.34794785885606e-005f, 5.808438800158911e-005f, 9.193058212986216e-005f, 0.0001327334757661447f, 0.0001748319627949968f, 0.0002100782439811155f, 0.0002302826324012131f, 0.0002302826324012131f, 0.0002100782439811155f, 0.0001748319627949968f, 0.0001327334757661447f, 9.193058212986216e-005f, 5.808438800158911e-005f, 3.34794785885606e-005f, 1.760426494001877e-005f, 8.444558261544444e-006f, 3.695352233989979e-006f
};
// utility for linear filter
inline uchar readerGet(
image2d_t src,
const float centerX, const float centerY, const float win_offset, const float cos_dir, const float sin_dir,
int i, int j
)
{
float pixel_x = centerX + (win_offset + j) * cos_dir + (win_offset + i) * sin_dir;
float pixel_y = centerY - (win_offset + j) * sin_dir + (win_offset + i) * cos_dir;
return (uchar)read_imageui(src, sampler, (float2)(pixel_x, pixel_y)).x;
}
inline float linearFilter(
image2d_t src,
const float centerX, const float centerY, const float win_offset, const float cos_dir, const float sin_dir,
float y, float x
)
{
x -= 0.5f;
y -= 0.5f;
float out = 0.0f;
const int x1 = convert_int_rtn(x);
const int y1 = convert_int_rtn(y);
const int x2 = x1 + 1;
const int y2 = y1 + 1;
uchar src_reg = readerGet(src, centerX, centerY, win_offset, cos_dir, sin_dir, y1, x1);
out = out + src_reg * ((x2 - x) * (y2 - y));
src_reg = readerGet(src, centerX, centerY, win_offset, cos_dir, sin_dir, y1, x2);
out = out + src_reg * ((x - x1) * (y2 - y));
src_reg = readerGet(src, centerX, centerY, win_offset, cos_dir, sin_dir, y2, x1);
out = out + src_reg * ((x2 - x) * (y - y1));
src_reg = readerGet(src, centerX, centerY, win_offset, cos_dir, sin_dir, y2, x2);
out = out + src_reg * ((x - x1) * (y - y1));
return out;
}
void calc_dx_dy(
image2d_t imgTex,
volatile __local float s_dx_bin[25],
volatile __local float s_dy_bin[25],
volatile __local float s_PATCH[6][6],
__global const float* featureX,
__global const float* featureY,
__global const float* featureSize,
__global const float* featureDir
)
{
const float centerX = featureX[get_group_id(0)];
const float centerY = featureY[get_group_id(0)];
const float size = featureSize[get_group_id(0)];
const float descriptor_dir = featureDir[get_group_id(0)] * (float)(CV_PI_F / 180.0f);
/* The sampling intervals and wavelet sized for selecting an orientation
and building the keypoint descriptor are defined relative to 's' */
const float s = size * 1.2f / 9.0f;
/* Extract a window of pixels around the keypoint of size 20s */
const int win_size = (int)((PATCH_SZ + 1) * s);
float sin_dir;
float cos_dir;
sin_dir = sincos(descriptor_dir, &cos_dir);
/* Nearest neighbour version (faster) */
const float win_offset = -(float)(win_size - 1) / 2;
// Compute sampling points
// since grids are 2D, need to compute xBlock and yBlock indices
const int xBlock = (get_group_id(1) & 3); // get_group_id(1) % 4
const int yBlock = (get_group_id(1) >> 2); // floor(get_group_id(1)/4)
const int xIndex = xBlock * 5 + get_local_id(0);
const int yIndex = yBlock * 5 + get_local_id(1);
const float icoo = ((float)yIndex / (PATCH_SZ + 1)) * win_size;
const float jcoo = ((float)xIndex / (PATCH_SZ + 1)) * win_size;
s_PATCH[get_local_id(1)][get_local_id(0)] = linearFilter(imgTex, centerX, centerY, win_offset, cos_dir, sin_dir, icoo, jcoo);
barrier(CLK_LOCAL_MEM_FENCE);
if (get_local_id(0) < 5 && get_local_id(1) < 5)
{
const int tid = get_local_id(1) * 5 + get_local_id(0);
const float dw = c_DW[yIndex * PATCH_SZ + xIndex];
const float vx = (
s_PATCH[get_local_id(1) ][get_local_id(0) + 1] -
s_PATCH[get_local_id(1) ][get_local_id(0) ] +
s_PATCH[get_local_id(1) + 1][get_local_id(0) + 1] -
s_PATCH[get_local_id(1) + 1][get_local_id(0) ])
* dw;
const float vy = (
s_PATCH[get_local_id(1) + 1][get_local_id(0) ] -
s_PATCH[get_local_id(1) ][get_local_id(0) ] +
s_PATCH[get_local_id(1) + 1][get_local_id(0) + 1] -
s_PATCH[get_local_id(1) ][get_local_id(0) + 1])
* dw;
s_dx_bin[tid] = vx;
s_dy_bin[tid] = vy;
}
}
void reduce_sum25(
volatile __local float* sdata1,
volatile __local float* sdata2,
volatile __local float* sdata3,
volatile __local float* sdata4,
int tid
)
{
// first step is to reduce from 25 to 16
if (tid < 9) // use 9 threads
{
sdata1[tid] += sdata1[tid + 16];
sdata2[tid] += sdata2[tid + 16];
sdata3[tid] += sdata3[tid + 16];
sdata4[tid] += sdata4[tid + 16];
}
// sum (reduce) from 16 to 1 (unrolled - aligned to a half-warp)
if (tid < 8)
{
sdata1[tid] += sdata1[tid + 8];
sdata1[tid] += sdata1[tid + 4];
sdata1[tid] += sdata1[tid + 2];
sdata1[tid] += sdata1[tid + 1];
sdata2[tid] += sdata2[tid + 8];
sdata2[tid] += sdata2[tid + 4];
sdata2[tid] += sdata2[tid + 2];
sdata2[tid] += sdata2[tid + 1];
sdata3[tid] += sdata3[tid + 8];
sdata3[tid] += sdata3[tid + 4];
sdata3[tid] += sdata3[tid + 2];
sdata3[tid] += sdata3[tid + 1];
sdata4[tid] += sdata4[tid + 8];
sdata4[tid] += sdata4[tid + 4];
sdata4[tid] += sdata4[tid + 2];
sdata4[tid] += sdata4[tid + 1];
}
}
__kernel
void compute_descriptors64(
image2d_t imgTex,
volatile __global float * descriptors,
__global const float * keypoints,
int descriptors_step,
int keypoints_step
)
{
descriptors_step /= sizeof(float);
keypoints_step /= sizeof(float);
__global const float * featureX = keypoints + X_ROW * keypoints_step;
__global const float * featureY = keypoints + Y_ROW * keypoints_step;
__global const float * featureSize = keypoints + SIZE_ROW * keypoints_step;
__global const float * featureDir = keypoints + ANGLE_ROW * keypoints_step;
// 2 floats (dx,dy) for each thread (5x5 sample points in each sub-region)
volatile __local float sdx[25];
volatile __local float sdy[25];
volatile __local float sdxabs[25];
volatile __local float sdyabs[25];
volatile __local float s_PATCH[6][6];
calc_dx_dy(imgTex, sdx, sdy, s_PATCH, featureX, featureY, featureSize, featureDir);
barrier(CLK_LOCAL_MEM_FENCE);
const int tid = get_local_id(1) * get_local_size(0) + get_local_id(0);
if (tid < 25)
{
sdxabs[tid] = fabs(sdx[tid]); // |dx| array
sdyabs[tid] = fabs(sdy[tid]); // |dy| array
barrier(CLK_LOCAL_MEM_FENCE);
reduce_sum25(sdx, sdy, sdxabs, sdyabs, tid);
barrier(CLK_LOCAL_MEM_FENCE);
volatile __global float* descriptors_block = descriptors + descriptors_step * get_group_id(0) + (get_group_id(1) << 2);
// write dx, dy, |dx|, |dy|
if (tid == 0)
{
descriptors_block[0] = sdx[0];
descriptors_block[1] = sdy[0];
descriptors_block[2] = sdxabs[0];
descriptors_block[3] = sdyabs[0];
}
}
}
__kernel
void compute_descriptors128(
image2d_t imgTex,
__global volatile float * descriptors,
__global float * keypoints,
int descriptors_step,
int keypoints_step
)
{
descriptors_step /= sizeof(*descriptors);
keypoints_step /= sizeof(*keypoints);
__global float * featureX = keypoints + X_ROW * keypoints_step;
__global float * featureY = keypoints + Y_ROW * keypoints_step;
__global float* featureSize = keypoints + SIZE_ROW * keypoints_step;
__global float* featureDir = keypoints + ANGLE_ROW * keypoints_step;
// 2 floats (dx,dy) for each thread (5x5 sample points in each sub-region)
volatile __local float sdx[25];
volatile __local float sdy[25];
// sum (reduce) 5x5 area response
volatile __local float sd1[25];
volatile __local float sd2[25];
volatile __local float sdabs1[25];
volatile __local float sdabs2[25];
volatile __local float s_PATCH[6][6];
calc_dx_dy(imgTex, sdx, sdy, s_PATCH, featureX, featureY, featureSize, featureDir);
barrier(CLK_LOCAL_MEM_FENCE);
const int tid = get_local_id(1) * get_local_size(0) + get_local_id(0);
if (tid < 25)
{
if (sdy[tid] >= 0)
{
sd1[tid] = sdx[tid];
sdabs1[tid] = fabs(sdx[tid]);
sd2[tid] = 0;
sdabs2[tid] = 0;
}
else
{
sd1[tid] = 0;
sdabs1[tid] = 0;
sd2[tid] = sdx[tid];
sdabs2[tid] = fabs(sdx[tid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
reduce_sum25(sd1, sd2, sdabs1, sdabs2, tid);
barrier(CLK_LOCAL_MEM_FENCE);
volatile __global float* descriptors_block = descriptors + descriptors_step * get_group_id(0) + (get_group_id(1) << 3);
// write dx (dy >= 0), |dx| (dy >= 0), dx (dy < 0), |dx| (dy < 0)
if (tid == 0)
{
descriptors_block[0] = sd1[0];
descriptors_block[1] = sdabs1[0];
descriptors_block[2] = sd2[0];
descriptors_block[3] = sdabs2[0];
}
if (sdx[tid] >= 0)
{
sd1[tid] = sdy[tid];
sdabs1[tid] = fabs(sdy[tid]);
sd2[tid] = 0;
sdabs2[tid] = 0;
}
else
{
sd1[tid] = 0;
sdabs1[tid] = 0;
sd2[tid] = sdy[tid];
sdabs2[tid] = fabs(sdy[tid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
reduce_sum25(sd1, sd2, sdabs1, sdabs2, tid);
barrier(CLK_LOCAL_MEM_FENCE);
// write dy (dx >= 0), |dy| (dx >= 0), dy (dx < 0), |dy| (dx < 0)
if (tid == 0)
{
descriptors_block[4] = sd1[0];
descriptors_block[5] = sdabs1[0];
descriptors_block[6] = sd2[0];
descriptors_block[7] = sdabs2[0];
}
}
}
__kernel
void normalize_descriptors128(__global float * descriptors, int descriptors_step)
{
descriptors_step /= sizeof(*descriptors);
// no need for thread ID
__global float* descriptor_base = descriptors + descriptors_step * get_group_id(0);
// read in the unnormalized descriptor values (squared)
volatile __local float sqDesc[128];
const float lookup = descriptor_base[get_local_id(0)];
sqDesc[get_local_id(0)] = lookup * lookup;
barrier(CLK_LOCAL_MEM_FENCE);
if (get_local_id(0) < 64)
sqDesc[get_local_id(0)] += sqDesc[get_local_id(0) + 64];
barrier(CLK_LOCAL_MEM_FENCE);
// reduction to get total
if (get_local_id(0) < 32)
{
volatile __local float* smem = sqDesc;
smem[get_local_id(0)] += smem[get_local_id(0) + 32];
smem[get_local_id(0)] += smem[get_local_id(0) + 16];
smem[get_local_id(0)] += smem[get_local_id(0) + 8];
smem[get_local_id(0)] += smem[get_local_id(0) + 4];
smem[get_local_id(0)] += smem[get_local_id(0) + 2];
smem[get_local_id(0)] += smem[get_local_id(0) + 1];
}
// compute length (square root)
volatile __local float len;
if (get_local_id(0) == 0)
{
len = sqrt(sqDesc[0]);
}
barrier(CLK_LOCAL_MEM_FENCE);
// normalize and store in output
descriptor_base[get_local_id(0)] = lookup / len;
}
__kernel
void normalize_descriptors64(__global float * descriptors, int descriptors_step)
{
descriptors_step /= sizeof(*descriptors);
// no need for thread ID
__global float* descriptor_base = descriptors + descriptors_step * get_group_id(0);
// read in the unnormalized descriptor values (squared)
volatile __local float sqDesc[64];
const float lookup = descriptor_base[get_local_id(0)];
sqDesc[get_local_id(0)] = lookup * lookup;
barrier(CLK_LOCAL_MEM_FENCE);
// reduction to get total
if (get_local_id(0) < 32)
{
volatile __local float* smem = sqDesc;
smem[get_local_id(0)] += smem[get_local_id(0) + 32];
smem[get_local_id(0)] += smem[get_local_id(0) + 16];
smem[get_local_id(0)] += smem[get_local_id(0) + 8];
smem[get_local_id(0)] += smem[get_local_id(0) + 4];
smem[get_local_id(0)] += smem[get_local_id(0) + 2];
smem[get_local_id(0)] += smem[get_local_id(0) + 1];
}
// compute length (square root)
volatile __local float len;
if (get_local_id(0) == 0)
{
len = sqrt(sqDesc[0]);
}
barrier(CLK_LOCAL_MEM_FENCE);
// normalize and store in output
descriptor_base[get_local_id(0)] = lookup / len;
}
modules/ocl/src/kernels/objdetect_hog.cl 0 → 100644
View file @ 64e9cf5d
/*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
// Wenju He, wenju@multicorewareinc.com
//
// 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*/
#define CELL_WIDTH 8
#define CELL_HEIGHT 8
#define CELLS_PER_BLOCK_X 2
#define CELLS_PER_BLOCK_Y 2
#define NTHREADS 256
#define CV_PI_F 3.1415926535897932384626433832795f
//----------------------------------------------------------------------------
// Histogram computation
__kernel void compute_hists_kernel(const int width, const int cblock_stride_x, const int cblock_stride_y,
const int cnbins, const int cblock_hist_size, const int img_block_width,
const int grad_quadstep, const int qangle_step,
__global const float* grad, __global const uchar* qangle,
const float scale, __global float* block_hists, __local float* smem)
{
const int lidX = get_local_id(0);
const int lidY = get_local_id(1);
const int gidX = get_group_id(0);
const int gidY = get_group_id(1);
const int cell_x = lidX / 16;
const int cell_y = lidY;
const int cell_thread_x = lidX & 0xF;
__local float* hists = smem;
__local float* final_hist = smem + cnbins * 48;
const int offset_x = gidX * cblock_stride_x + (cell_x << 2) + cell_thread_x;
const int offset_y = gidY * cblock_stride_y + (cell_y << 2);
__global const float* grad_ptr = grad + offset_y * grad_quadstep + (offset_x << 1);
__global const uchar* qangle_ptr = qangle + offset_y * qangle_step + (offset_x << 1);
// 12 means that 12 pixels affect on block's cell (in one row)
if (cell_thread_x < 12)
{
__local float* hist = hists + 12 * (cell_y * CELLS_PER_BLOCK_Y + cell_x) + cell_thread_x;
for (int bin_id = 0; bin_id < cnbins; ++bin_id)
hist[bin_id * 48] = 0.f;
const int dist_x = -4 + cell_thread_x - 4 * cell_x;
const int dist_y_begin = -4 - 4 * lidY;
for (int dist_y = dist_y_begin; dist_y < dist_y_begin + 12; ++dist_y)
{
float2 vote = (float2) (grad_ptr[0], grad_ptr[1]);
uchar2 bin = (uchar2) (qangle_ptr[0], qangle_ptr[1]);
grad_ptr += grad_quadstep;
qangle_ptr += qangle_step;
int dist_center_y = dist_y - 4 * (1 - 2 * cell_y);
int dist_center_x = dist_x - 4 * (1 - 2 * cell_x);
float gaussian = exp(-(dist_center_y * dist_center_y + dist_center_x * dist_center_x) * scale);
float interp_weight = (8.f - fabs(dist_y + 0.5f)) * (8.f - fabs(dist_x + 0.5f)) / 64.f;
hist[bin.x * 48] += gaussian * interp_weight * vote.x;
hist[bin.y * 48] += gaussian * interp_weight * vote.y;
}
volatile __local float* hist_ = hist;
for (int bin_id = 0; bin_id < cnbins; ++bin_id, hist_ += 48)
{
if (cell_thread_x < 6) hist_[0] += hist_[6];
if (cell_thread_x < 3) hist_[0] += hist_[3];
if (cell_thread_x == 0)
final_hist[(cell_x * 2 + cell_y) * cnbins + bin_id] = hist_[0] + hist_[1] + hist_[2];
}
}
barrier(CLK_LOCAL_MEM_FENCE);
__global float* block_hist = block_hists + (gidY * img_block_width + gidX) * cblock_hist_size;
int tid = (cell_y * CELLS_PER_BLOCK_Y + cell_x) * 16 + cell_thread_x;
if (tid < cblock_hist_size)
block_hist[tid] = final_hist[tid];
}
//-------------------------------------------------------------
// Normalization of histograms via L2Hys_norm
//
float reduce_smem(volatile __local float* smem, int size)
{
unsigned int tid = get_local_id(0);
float sum = smem[tid];
if (size >= 512) { if (tid < 256) smem[tid] = sum = sum + smem[tid + 256]; barrier(CLK_LOCAL_MEM_FENCE); }
if (size >= 256) { if (tid < 128) smem[tid] = sum = sum + smem[tid + 128]; barrier(CLK_LOCAL_MEM_FENCE); }
if (size >= 128) { if (tid < 64) smem[tid] = sum = sum + smem[tid + 64]; barrier(CLK_LOCAL_MEM_FENCE); }
if (tid < 32)
{
if (size >= 64) smem[tid] = sum = sum + smem[tid + 32];
if (size >= 32) smem[tid] = sum = sum + smem[tid + 16];
if (size >= 16) smem[tid] = sum = sum + smem[tid + 8];
if (size >= 8) smem[tid] = sum = sum + smem[tid + 4];
if (size >= 4) smem[tid] = sum = sum + smem[tid + 2];
if (size >= 2) smem[tid] = sum = sum + smem[tid + 1];
}
barrier(CLK_LOCAL_MEM_FENCE);
sum = smem[0];
return sum;
}
__kernel void normalize_hists_kernel(const int nthreads, const int block_hist_size, const int img_block_width,
__global float* block_hists, const float threshold, __local float *squares)
{
const int tid = get_local_id(0);
const int gidX = get_group_id(0);
const int gidY = get_group_id(1);
__global float* hist = block_hists + (gidY * img_block_width + gidX) * block_hist_size + tid;
float elem = 0.f;
if (tid < block_hist_size)
elem = hist[0];
squares[tid] = elem * elem;
barrier(CLK_LOCAL_MEM_FENCE);
float sum = reduce_smem(squares, nthreads);
float scale = 1.0f / (sqrt(sum) + 0.1f * block_hist_size);
elem = min(elem * scale, threshold);
barrier(CLK_LOCAL_MEM_FENCE);
squares[tid] = elem * elem;
barrier(CLK_LOCAL_MEM_FENCE);
sum = reduce_smem(squares, nthreads);
scale = 1.0f / (sqrt(sum) + 1e-3f);
if (tid < block_hist_size)
hist[0] = elem * scale;
}
//---------------------------------------------------------------------
// Linear SVM based classification
//
__kernel void classify_hists_kernel(const int cblock_hist_size, const int cdescr_size, const int cdescr_width,
const int img_win_width, const int img_block_width,
const int win_block_stride_x, const int win_block_stride_y,
__global const float * block_hists, __global const float* coefs,
float free_coef, float threshold, __global uchar* labels)
{
const int tid = get_local_id(0);
const int gidX = get_group_id(0);
const int gidY = get_group_id(1);
__global const float* hist = block_hists + (gidY * win_block_stride_y * img_block_width + gidX * win_block_stride_x) * cblock_hist_size;
float product = 0.f;
for (int i = tid; i < cdescr_size; i += NTHREADS)
{
int offset_y = i / cdescr_width;
int offset_x = i - offset_y * cdescr_width;
product += coefs[i] * hist[offset_y * img_block_width * cblock_hist_size + offset_x];
}
__local float products[NTHREADS];
products[tid] = product;
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 128) products[tid] = product = product + products[tid + 128];
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 64) products[tid] = product = product + products[tid + 64];
barrier(CLK_LOCAL_MEM_FENCE);
if (tid < 32)
{
volatile __local float* smem = products;
smem[tid] = product = product + smem[tid + 32];
smem[tid] = product = product + smem[tid + 16];
smem[tid] = product = product + smem[tid + 8];
smem[tid] = product = product + smem[tid + 4];
smem[tid] = product = product + smem[tid + 2];
smem[tid] = product = product + smem[tid + 1];
}
if (tid == 0)
labels[gidY * img_win_width + gidX] = (product + free_coef >= threshold);
}
//----------------------------------------------------------------------------
// Extract descriptors
__kernel void extract_descrs_by_rows_kernel(const int cblock_hist_size, const int descriptors_quadstep, const int cdescr_size, const int cdescr_width,
const int img_block_width, const int win_block_stride_x, const int win_block_stride_y,
__global const float* block_hists, __global float* descriptors)
{
int tid = get_local_id(0);
int gidX = get_group_id(0);
int gidY = get_group_id(1);
// Get left top corner of the window in src
__global const float* hist = block_hists + (gidY * win_block_stride_y * img_block_width + gidX * win_block_stride_x) * cblock_hist_size;
// Get left top corner of the window in dst
__global float* descriptor = descriptors + (gidY * get_num_groups(0) + gidX) * descriptors_quadstep;
// Copy elements from src to dst
for (int i = tid; i < cdescr_size; i += NTHREADS)
{
int offset_y = i / cdescr_width;
int offset_x = i - offset_y * cdescr_width;
descriptor[i] = hist[offset_y * img_block_width * cblock_hist_size + offset_x];
}
}
__kernel void extract_descrs_by_cols_kernel(const int cblock_hist_size, const int descriptors_quadstep, const int cdescr_size,
const int cnblocks_win_x, const int cnblocks_win_y, const int img_block_width, const int win_block_stride_x,
const int win_block_stride_y, __global const float* block_hists, __global float* descriptors)
{
int tid = get_local_id(0);
int gidX = get_group_id(0);
int gidY = get_group_id(1);
// Get left top corner of the window in src
__global const float* hist = block_hists + (gidY * win_block_stride_y * img_block_width + gidX * win_block_stride_x) * cblock_hist_size;
// Get left top corner of the window in dst
__global float* descriptor = descriptors + (gidY * get_num_groups(0) + gidX) * descriptors_quadstep;
// Copy elements from src to dst
for (int i = tid; i < cdescr_size; i += NTHREADS)
{
int block_idx = i / cblock_hist_size;
int idx_in_block = i - block_idx * cblock_hist_size;
int y = block_idx / cnblocks_win_x;
int x = block_idx - y * cnblocks_win_x;
descriptor[(x * cnblocks_win_y + y) * cblock_hist_size + idx_in_block] = hist[(y * img_block_width + x) * cblock_hist_size + idx_in_block];
}
}
//----------------------------------------------------------------------------
// Gradients computation
__kernel void compute_gradients_8UC4_kernel(const int height, const int width, const int img_step, const int grad_quadstep, const int qangle_step,
const __global uchar4 * img, __global float * grad, __global uchar * qangle,
const float angle_scale, const char correct_gamma, const int cnbins)
{
const int x = get_global_id(0);
const int tid = get_local_id(0);
const int gSizeX = get_local_size(0);
const int gidX = get_group_id(0);
const int gidY = get_group_id(1);
__global const uchar4* row = img + gidY * img_step;
__local float sh_row[(NTHREADS + 2) * 3];
uchar4 val;
if (x < width)
val = row[x];
else
val = row[width - 2];
sh_row[tid + 1] = val.x;
sh_row[tid + 1 + (NTHREADS + 2)] = val.y;
sh_row[tid + 1 + 2 * (NTHREADS + 2)] = val.z;
if (tid == 0)
{
val = row[max(x - 1, 1)];
sh_row[0] = val.x;
sh_row[(NTHREADS + 2)] = val.y;
sh_row[2 * (NTHREADS + 2)] = val.z;
}
if (tid == gSizeX - 1)
{
val = row[min(x + 1, width - 2)];
sh_row[gSizeX + 1] = val.x;
sh_row[gSizeX + 1 + (NTHREADS + 2)] = val.y;
sh_row[gSizeX + 1 + 2 * (NTHREADS + 2)] = val.z;
}
barrier(CLK_LOCAL_MEM_FENCE);
if (x < width)
{
float3 a = (float3) (sh_row[tid], sh_row[tid + (NTHREADS + 2)], sh_row[tid + 2 * (NTHREADS + 2)]);
float3 b = (float3) (sh_row[tid + 2], sh_row[tid + 2 + (NTHREADS + 2)], sh_row[tid + 2 + 2 * (NTHREADS + 2)]);
float3 dx;
if (correct_gamma == 1)
dx = sqrt(b) - sqrt(a);
else
dx = b - a;
float3 dy = (float3) 0.f;
if (gidY > 0 && gidY < height - 1)
{
a = convert_float3(img[(gidY - 1) * img_step + x].xyz);
b = convert_float3(img[(gidY + 1) * img_step + x].xyz);
if (correct_gamma == 1)
dy = sqrt(b) - sqrt(a);
else
dy = b - a;
}
float best_dx = dx.x;
float best_dy = dy.x;
float mag0 = dx.x * dx.x + dy.x * dy.x;
float mag1 = dx.y * dx.y + dy.y * dy.y;
if (mag0 < mag1)
{
best_dx = dx.y;
best_dy = dy.y;
mag0 = mag1;
}
mag1 = dx.z * dx.z + dy.z * dy.z;
if (mag0 < mag1)
{
best_dx = dx.z;
best_dy = dy.z;
mag0 = mag1;
}
mag0 = sqrt(mag0);
float ang = (atan2(best_dy, best_dx) + CV_PI_F) * angle_scale - 0.5f;
int hidx = (int)floor(ang);
ang -= hidx;
hidx = (hidx + cnbins) % cnbins;
qangle[(gidY * qangle_step + x) << 1] = hidx;
qangle[((gidY * qangle_step + x) << 1) + 1] = (hidx + 1) % cnbins;
grad[(gidY * grad_quadstep + x) << 1] = mag0 * (1.f - ang);
grad[((gidY * grad_quadstep + x) << 1) + 1] = mag0 * ang;
}
}
__kernel void compute_gradients_8UC1_kernel(const int height, const int width, const int img_step, const int grad_quadstep, const int qangle_step,
__global const uchar * img, __global float * grad, __global uchar * qangle,
const float angle_scale, const char correct_gamma, const int cnbins)
{
const int x = get_global_id(0);
const int tid = get_local_id(0);
const int gSizeX = get_local_size(0);
const int gidX = get_group_id(0);
const int gidY = get_group_id(1);
__global const uchar* row = img + gidY * img_step;
__local float sh_row[NTHREADS + 2];
if (x < width)
sh_row[tid + 1] = row[x];
else
sh_row[tid + 1] = row[width - 2];
if (tid == 0)
sh_row[0] = row[max(x - 1, 1)];
if (tid == gSizeX - 1)
sh_row[gSizeX + 1] = row[min(x + 1, width - 2)];
barrier(CLK_LOCAL_MEM_FENCE);
if (x < width)
{
float dx;
if (correct_gamma == 1)
dx = sqrt(sh_row[tid + 2]) - sqrt(sh_row[tid]);
else
dx = sh_row[tid + 2] - sh_row[tid];
float dy = 0.f;
if (gidY > 0 && gidY < height - 1)
{
float a = (float) img[ (gidY + 1) * img_step + x ];
float b = (float) img[ (gidY - 1) * img_step + x ];
if (correct_gamma == 1)
dy = sqrt(a) - sqrt(b);
else
dy = a - b;
}
float mag = sqrt(dx * dx + dy * dy);
float ang = (atan2(dy, dx) + CV_PI_F) * angle_scale - 0.5f;
int hidx = (int)floor(ang);
ang -= hidx;
hidx = (hidx + cnbins) % cnbins;
qangle[ (gidY * qangle_step + x) << 1 ] = hidx;
qangle[ ((gidY * qangle_step + x) << 1) + 1 ] = (hidx + 1) % cnbins;
grad[ (gidY * grad_quadstep + x) << 1 ] = mag * (1.f - ang);
grad[ ((gidY * grad_quadstep + x) << 1) + 1 ] = mag * ang;
}
}
modules/ocl/src/surf.cpp 0 → 100644
View file @ 64e9cf5d
/*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
// Peng Xiao, pengxiao@multicorewareinc.com
//
// 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 oclMaterials 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 <iomanip>
#include "precomp.hpp"
using namespace cv;
using namespace cv::ocl;
using namespace std;
#if !defined (HAVE_OPENCL)
cv::ocl::SURF_OCL::SURF_OCL() { throw_nogpu(); }
cv::ocl::SURF_OCL::SURF_OCL(double, int, int, bool, float, bool) { throw_nogpu(); }
int cv::ocl::SURF_OCL::descriptorSize() const { throw_nogpu(); return 0;}
void cv::ocl::SURF_OCL::uploadKeypoints(const vector<KeyPoint>&, oclMat&) { throw_nogpu(); }
void cv::ocl::SURF_OCL::downloadKeypoints(const oclMat&, vector<KeyPoint>&) { throw_nogpu(); }
void cv::ocl::SURF_OCL::downloadDescriptors(const oclMat&, vector<float>&) { throw_nogpu(); }
void cv::ocl::SURF_OCL::operator()(const oclMat&, const oclMat&, oclMat&) { throw_nogpu(); }
void cv::ocl::SURF_OCL::operator()(const oclMat&, const oclMat&, oclMat&, oclMat&, bool) { throw_nogpu(); }
void cv::ocl::SURF_OCL::operator()(const oclMat&, const oclMat&, vector<KeyPoint>&) { throw_nogpu(); }
void cv::ocl::SURF_OCL::operator()(const oclMat&, const oclMat&, vector<KeyPoint>&, oclMat&, bool) { throw_nogpu(); }
void cv::ocl::SURF_OCL::operator()(const oclMat&, const oclMat&, vector<KeyPoint>&, vector<float>&, bool) { throw_nogpu(); }
void cv::ocl::SURF_OCL::releaseMemory() { throw_nogpu(); }
#else /* !defined (HAVE_OPENCL) */
namespace cv { namespace ocl
{
///////////////////////////OpenCL kernel strings///////////////////////////
extern const char * nonfree_surf;
}}
namespace
{
static inline int divUp(int total, int grain)
{
return (total + grain - 1) / grain;
}
static inline int calcSize(int octave, int layer)
{
/* Wavelet size at first layer of first octave. */
const int HAAR_SIZE0 = 9;
/* Wavelet size increment between layers. This should be an even number,
such that the wavelet sizes in an octave are either all even or all odd.
This ensures that when looking for the neighbours of a sample, the layers
above and below are aligned correctly. */
const int HAAR_SIZE_INC = 6;
return (HAAR_SIZE0 + HAAR_SIZE_INC * layer) << octave;
}
class SURF_OCL_Invoker
{
public:
// facilities
void bindImgTex(const oclMat& img);
void bindSumTex(const oclMat& sum);
void bindMaskSumTex(const oclMat& maskSum);
//void loadGlobalConstants(int maxCandidates, int maxFeatures, int img_rows, int img_cols, int nOctaveLayers, float hessianThreshold);
//void loadOctaveConstants(int octave, int layer_rows, int layer_cols);
// kernel callers declearations
void icvCalcLayerDetAndTrace_gpu(oclMat& det, oclMat& trace, int octave, int nOctaveLayers, int layer_rows);
void icvFindMaximaInLayer_gpu(const oclMat& det, const oclMat& trace, oclMat& maxPosBuffer, oclMat& maxCounter, int counterOffset,
int octave, bool use_mask, int nLayers, int layer_rows, int layer_cols);
void icvInterpolateKeypoint_gpu(const oclMat& det, const oclMat& maxPosBuffer, unsigned int maxCounter,
oclMat& keypoints, oclMat& counters, int octave, int layer_rows, int maxFeatures);
void icvCalcOrientation_gpu(const oclMat& keypoints, int nFeatures);
void compute_descriptors_gpu(const oclMat& descriptors, const oclMat& keypoints, int nFeatures);
// end of kernel callers declearations
SURF_OCL_Invoker(SURF_OCL& surf, const oclMat& img, const oclMat& mask) :
surf_(surf),
img_cols(img.cols), img_rows(img.rows),
use_mask(!mask.empty())
{
CV_Assert(!img.empty() && img.type() == CV_8UC1);
CV_Assert(mask.empty() || (mask.size() == img.size() && mask.type() == CV_8UC1));
CV_Assert(surf_.nOctaves > 0 && surf_.nOctaveLayers > 0);
const int min_size = calcSize(surf_.nOctaves - 1, 0);
CV_Assert(img_rows - min_size >= 0);
CV_Assert(img_cols - min_size >= 0);
const int layer_rows = img_rows >> (surf_.nOctaves - 1);
const int layer_cols = img_cols >> (surf_.nOctaves - 1);
const int min_margin = ((calcSize((surf_.nOctaves - 1), 2) >> 1) >> (surf_.nOctaves - 1)) + 1;
CV_Assert(layer_rows - 2 * min_margin > 0);
CV_Assert(layer_cols - 2 * min_margin > 0);
maxFeatures = std::min(static_cast<int>(img.size().area() * surf.keypointsRatio), 65535);
maxCandidates = std::min(static_cast<int>(1.5 * maxFeatures), 65535);
CV_Assert(maxFeatures > 0);
counters.create(1, surf_.nOctaves + 1, CV_32SC1);
counters.setTo(Scalar::all(0));
//loadGlobalConstants(maxCandidates, maxFeatures, img_rows, img_cols, surf_.nOctaveLayers, static_cast<float>(surf_.hessianThreshold));
bindImgTex(img);
oclMat integral_sqsum;
integral(img, surf_.sum, integral_sqsum); // the two argumented integral version is incorrect
bindSumTex(surf_.sum);
maskSumTex = 0;
if (use_mask)
{
throw std::exception();
//!FIXME
// temp fix for missing min overload
oclMat temp(mask.size(), mask.type());
temp.setTo(Scalar::all(1.0));
//cv::ocl::min(mask, temp, surf_.mask1); ///////// disable this
integral(surf_.mask1, surf_.maskSum);
bindMaskSumTex(surf_.maskSum);
}
}
void detectKeypoints(oclMat& keypoints)
{
// create image pyramid buffers
// different layers have same sized buffers, but they are sampled from gaussin kernel.
surf_.det.create(img_rows * (surf_.nOctaveLayers + 2), img_cols, CV_32FC1);
surf_.trace.create(img_rows * (surf_.nOctaveLayers + 2), img_cols, CV_32FC1);
surf_.maxPosBuffer.create(1, maxCandidates, CV_32SC4);
keypoints.create(SURF_OCL::ROWS_COUNT, maxFeatures, CV_32FC1);
keypoints.setTo(Scalar::all(0));
for (int octave = 0; octave < surf_.nOctaves; ++octave)
{
const int layer_rows = img_rows >> octave;
const int layer_cols = img_cols >> octave;
//loadOctaveConstants(octave, layer_rows, layer_cols);
icvCalcLayerDetAndTrace_gpu(surf_.det, surf_.trace, octave, surf_.nOctaveLayers, layer_rows);
icvFindMaximaInLayer_gpu(surf_.det, surf_.trace, surf_.maxPosBuffer, counters, 1 + octave,
octave, use_mask, surf_.nOctaveLayers, layer_rows, layer_cols);
unsigned int maxCounter = Mat(counters).at<unsigned int>(1 + octave);
maxCounter = std::min(maxCounter, static_cast<unsigned int>(maxCandidates));
if (maxCounter > 0)
{
icvInterpolateKeypoint_gpu(surf_.det, surf_.maxPosBuffer, maxCounter,
keypoints, counters, octave, layer_rows, maxFeatures);
}
}
unsigned int featureCounter = Mat(counters).at<unsigned int>(0);
featureCounter = std::min(featureCounter, static_cast<unsigned int>(maxFeatures));
keypoints.cols = featureCounter;
if (surf_.upright)
keypoints.row(SURF_OCL::ANGLE_ROW).setTo(Scalar::all(90.0));
else
findOrientation(keypoints);
}
void findOrientation(oclMat& keypoints)
{
const int nFeatures = keypoints.cols;
if (nFeatures > 0)
{
icvCalcOrientation_gpu(keypoints, nFeatures);
}
}
void computeDescriptors(const oclMat& keypoints, oclMat& descriptors, int descriptorSize)
{
const int nFeatures = keypoints.cols;
if (nFeatures > 0)
{
descriptors.create(nFeatures, descriptorSize, CV_32F);
compute_descriptors_gpu(descriptors, keypoints, nFeatures);
}
}
~SURF_OCL_Invoker()
{
if(imgTex)
openCLFree(imgTex);
if(sumTex)
openCLFree(sumTex);
if(maskSumTex)
openCLFree(maskSumTex);
additioalParamBuffer.release();
}
private:
SURF_OCL& surf_;
int img_cols, img_rows;
bool use_mask;
int maxCandidates;
int maxFeatures;
oclMat counters;
// texture buffers
cl_mem imgTex;
cl_mem sumTex;
cl_mem maskSumTex;
oclMat additioalParamBuffer;
};
}
cv::ocl::SURF_OCL::SURF_OCL()
{
hessianThreshold = 100.0f;
extended = true;
nOctaves = 4;
nOctaveLayers = 2;
keypointsRatio = 0.01f;
upright = false;
}
cv::ocl::SURF_OCL::SURF_OCL(double _threshold, int _nOctaves, int _nOctaveLayers, bool _extended, float _keypointsRatio, bool _upright)
{
hessianThreshold = _threshold;
extended = _extended;
nOctaves = _nOctaves;
nOctaveLayers = _nOctaveLayers;
keypointsRatio = _keypointsRatio;
upright = _upright;
}
int cv::ocl::SURF_OCL::descriptorSize() const
{
return extended ? 128 : 64;
}
void cv::ocl::SURF_OCL::uploadKeypoints(const vector<KeyPoint>& keypoints, oclMat& keypointsGPU)
{
if (keypoints.empty())
keypointsGPU.release();
else
{
Mat keypointsCPU(SURF_OCL::ROWS_COUNT, static_cast<int>(keypoints.size()), CV_32FC1);
float* kp_x = keypointsCPU.ptr<float>(SURF_OCL::X_ROW);
float* kp_y = keypointsCPU.ptr<float>(SURF_OCL::Y_ROW);
int* kp_laplacian = keypointsCPU.ptr<int>(SURF_OCL::LAPLACIAN_ROW);
int* kp_octave = keypointsCPU.ptr<int>(SURF_OCL::OCTAVE_ROW);
float* kp_size = keypointsCPU.ptr<float>(SURF_OCL::SIZE_ROW);
float* kp_dir = keypointsCPU.ptr<float>(SURF_OCL::ANGLE_ROW);
float* kp_hessian = keypointsCPU.ptr<float>(SURF_OCL::HESSIAN_ROW);
for (size_t i = 0, size = keypoints.size(); i < size; ++i)
{
const KeyPoint& kp = keypoints[i];
kp_x[i] = kp.pt.x;
kp_y[i] = kp.pt.y;
kp_octave[i] = kp.octave;
kp_size[i] = kp.size;
kp_dir[i] = kp.angle;
kp_hessian[i] = kp.response;
kp_laplacian[i] = 1;
}
keypointsGPU.upload(keypointsCPU);
}
}
void cv::ocl::SURF_OCL::downloadKeypoints(const oclMat& keypointsGPU, vector<KeyPoint>& keypoints)
{
const int nFeatures = keypointsGPU.cols;
if (nFeatures == 0)
keypoints.clear();
else
{
CV_Assert(keypointsGPU.type() == CV_32FC1 && keypointsGPU.rows == ROWS_COUNT);
Mat keypointsCPU(keypointsGPU);
keypoints.resize(nFeatures);
float* kp_x = keypointsCPU.ptr<float>(SURF_OCL::X_ROW);
float* kp_y = keypointsCPU.ptr<float>(SURF_OCL::Y_ROW);
int* kp_laplacian = keypointsCPU.ptr<int>(SURF_OCL::LAPLACIAN_ROW);
int* kp_octave = keypointsCPU.ptr<int>(SURF_OCL::OCTAVE_ROW);
float* kp_size = keypointsCPU.ptr<float>(SURF_OCL::SIZE_ROW);
float* kp_dir = keypointsCPU.ptr<float>(SURF_OCL::ANGLE_ROW);
float* kp_hessian = keypointsCPU.ptr<float>(SURF_OCL::HESSIAN_ROW);
for (int i = 0; i < nFeatures; ++i)
{
KeyPoint& kp = keypoints[i];
kp.pt.x = kp_x[i];
kp.pt.y = kp_y[i];
kp.class_id = kp_laplacian[i];
kp.octave = kp_octave[i];
kp.size = kp_size[i];
kp.angle = kp_dir[i];
kp.response = kp_hessian[i];
}
}
}
void cv::ocl::SURF_OCL::downloadDescriptors(const oclMat& descriptorsGPU, vector<float>& descriptors)
{
if (descriptorsGPU.empty())
descriptors.clear();
else
{
CV_Assert(descriptorsGPU.type() == CV_32F);
descriptors.resize(descriptorsGPU.rows * descriptorsGPU.cols);
Mat descriptorsCPU(descriptorsGPU.size(), CV_32F, &descriptors[0]);
descriptorsGPU.download(descriptorsCPU);
}
}
void cv::ocl::SURF_OCL::operator()(const oclMat& img, const oclMat& mask, oclMat& keypoints)
{
if (!img.empty())
{
SURF_OCL_Invoker surf(*this, img, mask);
surf.detectKeypoints(keypoints);
}
}
void cv::ocl::SURF_OCL::operator()(const oclMat& img, const oclMat& mask, oclMat& keypoints, oclMat& descriptors,
bool useProvidedKeypoints)
{
if (!img.empty())
{
SURF_OCL_Invoker surf(*this, img, mask);
if (!useProvidedKeypoints)
surf.detectKeypoints(keypoints);
else if (!upright)
{
surf.findOrientation(keypoints);
}
surf.computeDescriptors(keypoints, descriptors, descriptorSize());
}
}
void cv::ocl::SURF_OCL::operator()(const oclMat& img, const oclMat& mask, vector<KeyPoint>& keypoints)
{
oclMat keypointsGPU;
(*this)(img, mask, keypointsGPU);
downloadKeypoints(keypointsGPU, keypoints);
}
void cv::ocl::SURF_OCL::operator()(const oclMat& img, const oclMat& mask, vector<KeyPoint>& keypoints,
oclMat& descriptors, bool useProvidedKeypoints)
{
oclMat keypointsGPU;
if (useProvidedKeypoints)
uploadKeypoints(keypoints, keypointsGPU);
(*this)(img, mask, keypointsGPU, descriptors, useProvidedKeypoints);
downloadKeypoints(keypointsGPU, keypoints);
}
void cv::ocl::SURF_OCL::operator()(const oclMat& img, const oclMat& mask, vector<KeyPoint>& keypoints,
vector<float>& descriptors, bool useProvidedKeypoints)
{
oclMat descriptorsGPU;
(*this)(img, mask, keypoints, descriptorsGPU, useProvidedKeypoints);
downloadDescriptors(descriptorsGPU, descriptors);
}
void cv::ocl::SURF_OCL::releaseMemory()
{
sum.release();
mask1.release();
maskSum.release();
intBuffer.release();
det.release();
trace.release();
maxPosBuffer.release();
}
// Facilities
//// load SURF constants into device memory
//void SURF_OCL_Invoker::loadGlobalConstants(int maxCandidates, int maxFeatures, int img_rows, int img_cols, int nOctaveLayers, float hessianThreshold)
//{
// Mat tmp(1, 9, CV_32FC1);
// float * tmp_data = tmp.ptr<float>();
// *tmp_data = maxCandidates;
// *(++tmp_data) = maxFeatures;
// *(++tmp_data) = img_rows;
// *(++tmp_data) = img_cols;
// *(++tmp_data) = nOctaveLayers;
// *(++tmp_data) = hessianThreshold;
// additioalParamBuffer = tmp;
//}
//void SURF_OCL_Invoker::loadOctaveConstants(int octave, int layer_rows, int layer_cols)
//{
// Mat tmp = additioalParamBuffer;
// float * tmp_data = tmp.ptr<float>();
// tmp_data += 6;
// *tmp_data = octave;
// *(++tmp_data) = layer_rows;
// *(++tmp_data) = layer_cols;
// additioalParamBuffer = tmp;
//}
// create and bind source buffer to image oject.
void SURF_OCL_Invoker::bindImgTex(const oclMat& img)
{
Mat cpu_img(img); // time consuming
cl_image_format format;
int err;
format.image_channel_data_type = CL_UNSIGNED_INT8;
format.image_channel_order = CL_R;
#if CL_VERSION_1_2
cl_image_desc desc;
desc.image_type = CL_MEM_OBJECT_IMAGE2D;
desc.image_width = cpu_img.cols;
desc.image_height = cpu_img.rows;
desc.image_depth = NULL;
desc.image_array_size = 1;
desc.image_row_pitch = cpu_img.step;
desc.image_slice_pitch= 0;
desc.buffer = NULL;
desc.num_mip_levels = 0;
desc.num_samples = 0;
imgTex = clCreateImage(img.clCxt->impl->clContext, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, &format, &desc, cpu_img.data, &err);
#else
imgTex = clCreateImage2D(
img.clCxt->impl->clContext,
CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
&format,
cpu_img.cols,
cpu_img.rows,
cpu_img.step,
cpu_img.data,
&err);
#endif
openCLSafeCall(err);
}
void SURF_OCL_Invoker::bindSumTex(const oclMat& sum)
{
Mat cpu_img(sum); // time consuming
cl_image_format format;
int err;
format.image_channel_data_type = CL_UNSIGNED_INT32;
format.image_channel_order = CL_R;
#if CL_VERSION_1_2
cl_image_desc desc;
desc.image_type = CL_MEM_OBJECT_IMAGE2D;
desc.image_width = cpu_img.cols;
desc.image_height = cpu_img.rows;
desc.image_depth = NULL;
desc.image_array_size = 1;
desc.image_row_pitch = cpu_img.step;
desc.image_slice_pitch= 0;
desc.buffer = NULL;
desc.num_mip_levels = 0;
desc.num_samples = 0;
sumTex = clCreateImage(sum.clCxt->impl->clContext, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, &format, &desc, cpu_img.data, &err);
#else
sumTex = clCreateImage2D(
sum.clCxt->impl->clContext,
CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
&format,
cpu_img.cols,
cpu_img.rows,
cpu_img.step,
cpu_img.data,
&err);
#endif
openCLSafeCall(err);
}
void SURF_OCL_Invoker::bindMaskSumTex(const oclMat& maskSum)
{
Mat cpu_img(maskSum); // time consuming
cl_image_format format;
int err;
format.image_channel_data_type = CL_UNSIGNED_INT32;
format.image_channel_order = CL_R;
#if CL_VERSION_1_2
cl_image_desc desc;
desc.image_type = CL_MEM_OBJECT_IMAGE2D;
desc.image_width = cpu_img.cols;
desc.image_height = cpu_img.rows;
desc.image_depth = NULL;
desc.image_array_size = 1;
desc.image_row_pitch = cpu_img.step;
desc.image_slice_pitch= 0;
desc.buffer = NULL;
desc.num_mip_levels = 0;
desc.num_samples = 0;
maskSumTex = clCreateImage(maskSum.clCxt->impl->clContext, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR, &format, &desc, cpu_img.data, &err);
#else
maskSumTex = clCreateImage2D(
maskSum.clCxt->impl->clContext,
CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
&format,
cpu_img.cols,
cpu_img.rows,
cpu_img.step,
cpu_img.data,
&err);
#endif
openCLSafeCall(err);
}
////////////////////////////
// kernel caller definitions
void SURF_OCL_Invoker::icvCalcLayerDetAndTrace_gpu(oclMat& det, oclMat& trace, int octave, int nOctaveLayers, int c_layer_rows)
{
const int min_size = calcSize(octave, 0);
const int max_samples_i = 1 + ((img_rows - min_size) >> octave);
const int max_samples_j = 1 + ((img_cols - min_size) >> octave);
Context *clCxt = det.clCxt;
string kernelName = "icvCalcLayerDetAndTrace";
vector< pair<size_t, const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&sumTex));
args.push_back( make_pair( sizeof(cl_mem), (void *)&det.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&trace.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&det.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&trace.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&img_rows));
args.push_back( make_pair( sizeof(cl_int), (void *)&img_cols));
args.push_back( make_pair( sizeof(cl_int), (void *)&nOctaveLayers));
args.push_back( make_pair( sizeof(cl_int), (void *)&octave));
args.push_back( make_pair( sizeof(cl_int), (void *)&c_layer_rows));
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3] = {
divUp(max_samples_j, localThreads[0]) * localThreads[0],
divUp(max_samples_i, localThreads[1]) * localThreads[1] * (nOctaveLayers + 2),
1};
openCLExecuteKernel(clCxt, &nonfree_surf, kernelName, globalThreads, localThreads, args, -1, -1);
}
void SURF_OCL_Invoker::icvFindMaximaInLayer_gpu(const oclMat& det, const oclMat& trace, oclMat& maxPosBuffer, oclMat& maxCounter, int counterOffset,
int octave, bool use_mask, int nLayers, int layer_rows, int layer_cols)
{
const int min_margin = ((calcSize(octave, 2) >> 1) >> octave) + 1;
Context *clCxt = det.clCxt;
string kernelName = use_mask ? "icvFindMaximaInLayer_withmask" : "icvFindMaximaInLayer";
vector< pair<size_t, const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&det.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&trace.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&maxPosBuffer.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&maxCounter.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&counterOffset));
args.push_back( make_pair( sizeof(cl_int), (void *)&det.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&trace.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&img_rows));
args.push_back( make_pair( sizeof(cl_int), (void *)&img_cols));
args.push_back( make_pair( sizeof(cl_int), (void *)&nLayers));
args.push_back( make_pair( sizeof(cl_int), (void *)&octave));
args.push_back( make_pair( sizeof(cl_int), (void *)&layer_rows));
args.push_back( make_pair( sizeof(cl_int), (void *)&layer_cols));
args.push_back( make_pair( sizeof(cl_int), (void *)&maxCandidates));
args.push_back( make_pair( sizeof(cl_float), (void *)&surf_.hessianThreshold));
if(use_mask)
{
args.push_back( make_pair( sizeof(cl_mem), (void *)&maskSumTex));
}
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3] = {divUp(layer_cols - 2 * min_margin, localThreads[0] - 2) * localThreads[0],
divUp(layer_rows - 2 * min_margin, localThreads[1] - 2) * nLayers * localThreads[1],
1};
openCLExecuteKernel(clCxt, &nonfree_surf, kernelName, globalThreads, localThreads, args, -1, -1);
}
void SURF_OCL_Invoker::icvInterpolateKeypoint_gpu(const oclMat& det, const oclMat& maxPosBuffer, unsigned int maxCounter,
oclMat& keypoints, oclMat& counters, int octave, int layer_rows, int maxFeatures)
{
Context *clCxt = det.clCxt;
string kernelName = "icvInterpolateKeypoint";
vector< pair<size_t, const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&det.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&maxPosBuffer.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&counters.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&det.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&keypoints.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&img_rows));
args.push_back( make_pair( sizeof(cl_int), (void *)&img_cols));
args.push_back( make_pair( sizeof(cl_int), (void *)&octave));
args.push_back( make_pair( sizeof(cl_int), (void *)&layer_rows));
args.push_back( make_pair( sizeof(cl_int), (void *)&maxFeatures));
size_t localThreads[3] = {3, 3, 3};
size_t globalThreads[3] = {maxCounter * localThreads[0], 1, 1};
openCLExecuteKernel(clCxt, &nonfree_surf, kernelName, globalThreads, localThreads, args, -1, -1);
}
void SURF_OCL_Invoker::icvCalcOrientation_gpu(const oclMat& keypoints, int nFeatures)
{
Context * clCxt = counters.clCxt;
string kernelName = "icvCalcOrientation";
vector< pair<size_t, const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&sumTex));
args.push_back( make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&keypoints.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&img_rows));
args.push_back( make_pair( sizeof(cl_int), (void *)&img_cols));
size_t localThreads[3] = {32, 4, 1};
size_t globalThreads[3] = {nFeatures * localThreads[0], localThreads[1], 1};
openCLExecuteKernel(clCxt, &nonfree_surf, kernelName, globalThreads, localThreads, args, -1, -1);
}
void SURF_OCL_Invoker::compute_descriptors_gpu(const oclMat& descriptors, const oclMat& keypoints, int nFeatures)
{
// compute unnormalized descriptors, then normalize them - odd indexing since grid must be 2D
Context *clCxt = descriptors.clCxt;
string kernelName = "";
vector< pair<size_t, const void *> > args;
size_t localThreads[3] = {1, 1, 1};
size_t globalThreads[3] = {1, 1, 1};
if(descriptors.cols == 64)
{
kernelName = "compute_descriptors64";
localThreads[0] = 6;
localThreads[1] = 6;
globalThreads[0] = nFeatures * localThreads[0];
globalThreads[1] = 16 * localThreads[1];
args.clear();
args.push_back( make_pair( sizeof(cl_mem), (void *)&imgTex));
args.push_back( make_pair( sizeof(cl_mem), (void *)&descriptors.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&descriptors.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&keypoints.step));
openCLExecuteKernel(clCxt, &nonfree_surf, kernelName, globalThreads, localThreads, args, -1, -1);
kernelName = "normalize_descriptors64";
localThreads[0] = 64;
localThreads[1] = 1;
globalThreads[0] = nFeatures * localThreads[0];
globalThreads[1] = localThreads[1];
args.clear();
args.push_back( make_pair( sizeof(cl_mem), (void *)&descriptors.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&descriptors.step));
openCLExecuteKernel(clCxt, &nonfree_surf, kernelName, globalThreads, localThreads, args, -1, -1);
}
else
{
kernelName = "compute_descriptors128";
localThreads[0] = 6;
localThreads[1] = 6;
globalThreads[0] = nFeatures * localThreads[0];
globalThreads[1] = 16 * localThreads[1];
args.clear();
args.push_back( make_pair( sizeof(cl_mem), (void *)&imgTex));
args.push_back( make_pair( sizeof(cl_mem), (void *)&descriptors.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&descriptors.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&keypoints.step));
openCLExecuteKernel(clCxt, &nonfree_surf, kernelName, globalThreads, localThreads, args, -1, -1);
kernelName = "normalize_descriptors128";
localThreads[0] = 128;
localThreads[1] = 1;
globalThreads[0] = nFeatures * localThreads[0];
globalThreads[1] = localThreads[1];
args.clear();
args.push_back( make_pair( sizeof(cl_mem), (void *)&descriptors.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&descriptors.step));
openCLExecuteKernel(clCxt, &nonfree_surf, kernelName, globalThreads, localThreads, args, -1, -1);
}
}
#endif // /* !defined (HAVE_OPENCL) */
modules/ocl/test/test_hog.cpp 0 → 100644
View file @ 64e9cf5d
/*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.
//
//
// Intel 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
// Wenju He, wenju@multicorewareinc.com
//
// 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 Intel Corporation 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 "opencv2/core/core.hpp"
using namespace std;
#ifdef HAVE_OPENCL
PARAM_TEST_CASE(HOG,cv::Size,int)
{
cv::Size winSize;
int type;
vector<cv::ocl::Info> info;
virtual void SetUp()
{
winSize = GET_PARAM(0);
type = GET_PARAM(1);
cv::ocl::getDevice(info);
}
};
TEST_P(HOG, GetDescriptors)
{
// Load image
cv::Mat img_rgb = readImage("../../../samples/gpu/road.png");
ASSERT_FALSE(img_rgb.empty());
// Convert image
cv::Mat img;
switch (type)
{
case CV_8UC1:
cv::cvtColor(img_rgb, img, CV_BGR2GRAY);
break;
case CV_8UC4:
default:
cv::cvtColor(img_rgb, img, CV_BGR2BGRA);
break;
}
cv::ocl::oclMat d_img(img);
// HOGs
cv::ocl::HOGDescriptor ocl_hog;
ocl_hog.gamma_correction = true;
cv::HOGDescriptor hog;
hog.gammaCorrection = true;
// Compute descriptor
cv::ocl::oclMat d_descriptors;
ocl_hog.getDescriptors(d_img, ocl_hog.win_size, d_descriptors, ocl_hog.DESCR_FORMAT_COL_BY_COL);
cv::Mat down_descriptors;
d_descriptors.download(down_descriptors);
down_descriptors = down_descriptors.reshape(0, down_descriptors.cols * down_descriptors.rows);
hog.setSVMDetector(hog.getDefaultPeopleDetector());
std::vector<float> descriptors;
switch (type)
{
case CV_8UC1:
hog.compute(img, descriptors, ocl_hog.win_size);
break;
case CV_8UC4:
default:
hog.compute(img_rgb, descriptors, ocl_hog.win_size);
break;
}
cv::Mat cpu_descriptors(descriptors);
EXPECT_MAT_SIMILAR(down_descriptors, cpu_descriptors, 1e-2);
}
TEST_P(HOG, Detect)
{
// Load image
cv::Mat img_rgb = readImage("../../../samples/gpu/road.png");
ASSERT_FALSE(img_rgb.empty());
// Convert image
cv::Mat img;
switch (type)
{
case CV_8UC1:
cv::cvtColor(img_rgb, img, CV_BGR2GRAY);
break;
case CV_8UC4:
default:
cv::cvtColor(img_rgb, img, CV_BGR2BGRA);
break;
}
cv::ocl::oclMat d_img(img);
// HOGs
if ((winSize != cv::Size(48, 96)) && (winSize != cv::Size(64, 128)))
winSize = cv::Size(64, 128);
cv::ocl::HOGDescriptor ocl_hog(winSize);
ocl_hog.gamma_correction = true;
cv::HOGDescriptor hog;
hog.winSize = winSize;
hog.gammaCorrection = true;
if (winSize.width == 48 && winSize.height == 96)
{
// daimler's base
ocl_hog.setSVMDetector(ocl_hog.getPeopleDetector48x96());
hog.setSVMDetector(hog.getDaimlerPeopleDetector());
}
else if (winSize.width == 64 && winSize.height == 128)
{
ocl_hog.setSVMDetector(ocl_hog.getPeopleDetector64x128());
hog.setSVMDetector(hog.getDefaultPeopleDetector());
}
else
{
ocl_hog.setSVMDetector(ocl_hog.getDefaultPeopleDetector());
hog.setSVMDetector(hog.getDefaultPeopleDetector());
}
// OpenCL detection
std::vector<cv::Point> d_v_locations;
ocl_hog.detect(d_img, d_v_locations, 0);
cv::Mat d_locations(d_v_locations);
// CPU detection
std::vector<cv::Point> v_locations;
switch (type)
{
case CV_8UC1:
hog.detect(img, v_locations, 0);
break;
case CV_8UC4:
default:
hog.detect(img_rgb, v_locations, 0);
break;
}
cv::Mat locations(v_locations);
char s[100]={0};
EXPECT_MAT_NEAR(d_locations, locations, 0, s);
}
INSTANTIATE_TEST_CASE_P(OCL_ObjDetect, HOG, testing::Combine(
testing::Values(cv::Size(64, 128), cv::Size(48, 96)),
testing::Values(MatType(CV_8UC1), MatType(CV_8UC4))));
#endif //HAVE_OPENCL
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment