Commit e6f3c9b0 authored by Andrey Pavlenko's avatar Andrey Pavlenko Committed by OpenCV Buildbot

Merge pull request #2281 from vpisarev:ocl_surf

parents 055f41c9 3e0c72a8
......@@ -246,105 +246,3 @@ The class ``SURF_CUDA`` uses some buffers and provides access to it. All buffers
.. note::
* An example for using the SURF keypoint matcher on GPU can be found at opencv_source_code/samples/gpu/surf_keypoint_matcher.cpp
ocl::SURF_OCL
-------------
.. ocv:class:: ocl::SURF_OCL
Class used for extracting Speeded Up Robust Features (SURF) from an image. ::
class 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<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);
void operator()(const oclMat& img, const oclMat& mask,
oclMat& keypoints);
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
double 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;
};
The class ``SURF_OCL`` implements Speeded Up Robust Features descriptor. There is a fast multi-scale Hessian keypoint detector that can be used to find the keypoints (which is the default option). But the descriptors can also be computed for the user-specified keypoints. Only 8-bit grayscale images are supported.
The class ``SURF_OCL`` can store results in the GPU and CPU memory. It provides functions to convert results between CPU and GPU version ( ``uploadKeypoints``, ``downloadKeypoints``, ``downloadDescriptors`` ). The format of CPU results is the same as ``SURF`` results. GPU results are stored in ``oclMat``. The ``keypoints`` matrix is :math:`\texttt{nFeatures} \times 7` matrix with the ``CV_32FC1`` type.
* ``keypoints.ptr<float>(X_ROW)[i]`` contains x coordinate of the i-th feature.
* ``keypoints.ptr<float>(Y_ROW)[i]`` contains y coordinate of the i-th feature.
* ``keypoints.ptr<float>(LAPLACIAN_ROW)[i]`` contains the laplacian sign of the i-th feature.
* ``keypoints.ptr<float>(OCTAVE_ROW)[i]`` contains the octave of the i-th feature.
* ``keypoints.ptr<float>(SIZE_ROW)[i]`` contains the size of the i-th feature.
* ``keypoints.ptr<float>(ANGLE_ROW)[i]`` contain orientation of the i-th feature.
* ``keypoints.ptr<float>(HESSIAN_ROW)[i]`` contains the response of the i-th feature.
The ``descriptors`` matrix is :math:`\texttt{nFeatures} \times \texttt{descriptorSize}` matrix with the ``CV_32FC1`` type.
The class ``SURF_OCL`` uses some buffers and provides access to it. All buffers can be safely released between function calls.
.. seealso:: :ocv:class:`SURF`
.. note::
* OCL : An example of the SURF detector can be found at opencv_source_code/samples/ocl/surf_matcher.cpp
......@@ -142,7 +142,6 @@ public:
CV_PROP_RW bool upright;
protected:
void detectImpl( InputArray image, std::vector<KeyPoint>& keypoints, InputArray mask = noArray() ) const;
void computeImpl( const Mat& image, std::vector<KeyPoint>& keypoints, Mat& descriptors ) const;
};
......
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#ifndef __OPENCV_NONFREE_OCL_HPP__
#define __OPENCV_NONFREE_OCL_HPP__
#include "opencv2/ocl.hpp"
namespace cv
{
namespace ocl
{
//! Speeded up robust features, port from CUDA 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;
//! returns the default norm type
int defaultNorm() const;
//! upload host keypoints to device memory
void uploadKeypoints(const std::vector<cv::KeyPoint> &keypoints, oclMat &keypointsocl);
//! download keypoints from device to host memory
void downloadKeypoints(const oclMat &keypointsocl, std::vector<KeyPoint> &keypoints);
//! download descriptors from device to host memory
void downloadDescriptors(const oclMat &descriptorsocl, std::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;
};
}
}
#endif //__OPENCV_NONFREE_OCL_HPP__
......@@ -45,36 +45,59 @@
//
//M*/
// The number of degrees between orientation samples in calcOrientation
#define ORI_SEARCH_INC 5
// The local size of the calcOrientation kernel
#define ORI_LOCAL_SIZE (360 / ORI_SEARCH_INC)
// specialized for non-image2d_t supported platform, intel HD4000, for example
#ifdef DISABLE_IMAGE2D
#define IMAGE_INT32 __global uint *
#define IMAGE_INT8 __global uchar *
#else
#define IMAGE_INT32 image2d_t
#define IMAGE_INT8 image2d_t
#endif
#ifndef HAVE_IMAGE2D
__inline uint read_sumTex_(__global uint* sumTex, int sum_step, int img_rows, int img_cols, int2 coord)
{
int x = clamp(coord.x, 0, img_cols);
int y = clamp(coord.y, 0, img_rows);
return sumTex[sum_step * y + x];
}
uint read_sumTex(IMAGE_INT32 img, sampler_t sam, int2 coord, int rows, int cols, int elemPerRow)
__inline uchar read_imgTex_(__global uchar* imgTex, int img_step, int img_rows, int img_cols, float2 coord)
{
#ifdef DISABLE_IMAGE2D
int x = clamp(coord.x, 0, cols);
int y = clamp(coord.y, 0, rows);
return img[elemPerRow * y + x];
int x = clamp(convert_int_rte(coord.x), 0, img_cols-1);
int y = clamp(convert_int_rte(coord.y), 0, img_rows-1);
return imgTex[img_step * y + x];
}
#define read_sumTex(coord) read_sumTex_(sumTex, sum_step, img_rows, img_cols, coord)
#define read_imgTex(coord) read_imgTex_(imgTex, img_step, img_rows, img_cols, coord)
#define __PARAM_sumTex__ __global uint* sumTex, int sum_step, int sum_offset
#define __PARAM_imgTex__ __global uchar* imgTex, int img_step, int img_offset
#define __PASS_sumTex__ sumTex, sum_step, sum_offset
#define __PASS_imgTex__ imgTex, img_step, img_offset
#else
return read_imageui(img, sam, coord).x;
#endif
__inline uint read_sumTex_(image2d_t sumTex, sampler_t sam, int2 coord)
{
return read_imageui(sumTex, sam, coord).x;
}
uchar read_imgTex(IMAGE_INT8 img, sampler_t sam, float2 coord, int rows, int cols, int elemPerRow)
__inline uchar read_imgTex_(image2d_t imgTex, sampler_t sam, float2 coord)
{
#ifdef DISABLE_IMAGE2D
int x = clamp(round(coord.x), 0, cols - 1);
int y = clamp(round(coord.y), 0, rows - 1);
return img[elemPerRow * y + x];
#else
return (uchar)read_imageui(img, sam, coord).x;
#endif
return (uchar)read_imageui(imgTex, sam, coord).x;
}
#define read_sumTex(coord) read_sumTex_(sumTex, sampler, coord)
#define read_imgTex(coord) read_imgTex_(imgTex, sampler, coord)
#define __PARAM_sumTex__ image2d_t sumTex
#define __PARAM_imgTex__ image2d_t imgTex
#define __PASS_sumTex__ sumTex
#define __PASS_imgTex__ imgTex
#endif
// dynamically change the precision used for floating type
#if defined (DOUBLE_SUPPORT)
......@@ -99,45 +122,6 @@ __constant sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAM
#define CV_PI_F 3.14159265f
#endif
// Use integral image to calculate haar wavelets.
// N = 2
// for simple haar paatern
float icvCalcHaarPatternSum_2(
IMAGE_INT32 sumTex,
__constant float2 *src,
int oldSize,
int newSize,
int y, int x,
int rows, int cols, int elemPerRow)
{
float ratio = (float)newSize / oldSize;
F d = 0;
int2 dx1 = convert_int2(round(ratio * src[0]));
int2 dy1 = convert_int2(round(ratio * src[1]));
int2 dx2 = convert_int2(round(ratio * src[2]));
int2 dy2 = convert_int2(round(ratio * src[3]));
F t = 0;
t += read_sumTex( sumTex, sampler, (int2)(x + dx1.x, y + dy1.x), rows, cols, elemPerRow );
t -= read_sumTex( sumTex, sampler, (int2)(x + dx1.x, y + dy2.x), rows, cols, elemPerRow );
t -= read_sumTex( sumTex, sampler, (int2)(x + dx2.x, y + dy1.x), rows, cols, elemPerRow );
t += read_sumTex( sumTex, sampler, (int2)(x + dx2.x, y + dy2.x), rows, cols, elemPerRow );
d += t * src[4].x / ((dx2.x - dx1.x) * (dy2.x - dy1.x));
t = 0;
t += read_sumTex( sumTex, sampler, (int2)(x + dx1.y, y + dy1.y), rows, cols, elemPerRow );
t -= read_sumTex( sumTex, sampler, (int2)(x + dx1.y, y + dy2.y), rows, cols, elemPerRow );
t -= read_sumTex( sumTex, sampler, (int2)(x + dx2.y, y + dy1.y), rows, cols, elemPerRow );
t += read_sumTex( sumTex, sampler, (int2)(x + dx2.y, y + dy2.y), rows, cols, elemPerRow );
d += t * src[4].y / ((dx2.y - dx1.y) * (dy2.y - dy1.y));
return (float)d;
}
////////////////////////////////////////////////////////////////////////
// Hessian
......@@ -175,23 +159,21 @@ F calcAxisAlignedDerivative(
}
//calculate targeted layer per-pixel determinant and trace with an integral image
__kernel void icvCalcLayerDetAndTrace(
IMAGE_INT32 sumTex, // input integral image
__global float * det, // output Determinant
__kernel void SURF_calcLayerDetAndTrace(
__PARAM_sumTex__, // input integral image
int img_rows, int img_cols,
int c_nOctaveLayers, int c_octave, int c_layer_rows,
__global float * det, // output determinant
int det_step, int det_offset,
__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,
int sumTex_step
)
int trace_step, int trace_offset)
{
det_step /= sizeof(*det);
trace_step /= sizeof(*trace);
sumTex_step/= sizeof(uint);
#ifndef HAVE_IMAGE2D
sum_step/= sizeof(uint);
#endif
// Determine the indices
const int gridDim_y = get_num_groups(1) / (c_nOctaveLayers + 2);
const int blockIdx_y = get_group_id(1) % gridDim_y;
......@@ -203,13 +185,13 @@ __kernel void icvCalcLayerDetAndTrace(
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);
const int samples_i = 1 + ((img_rows - size) >> c_octave);
const int samples_j = 1 + ((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)
if (size <= img_rows && size <= img_cols && i < samples_i && j < samples_j)
{
int x = j << c_octave;
int y = i << c_octave;
......@@ -233,14 +215,14 @@ __kernel void icvCalcLayerDetAndTrace(
{
// Some of the pixels needed to compute the derivative are
// repeated, so we only don't duplicate the fetch here.
int t02 = read_sumTex( sumTex, sampler, (int2)(x, y + r2), c_img_rows, c_img_cols, sumTex_step );
int t07 = read_sumTex( sumTex, sampler, (int2)(x, y + r7), c_img_rows, c_img_cols, sumTex_step );
int t32 = read_sumTex( sumTex, sampler, (int2)(x + r3, y + r2), c_img_rows, c_img_cols, sumTex_step );
int t37 = read_sumTex( sumTex, sampler, (int2)(x + r3, y + r7), c_img_rows, c_img_cols, sumTex_step );
int t62 = read_sumTex( sumTex, sampler, (int2)(x + r6, y + r2), c_img_rows, c_img_cols, sumTex_step );
int t67 = read_sumTex( sumTex, sampler, (int2)(x + r6, y + r7), c_img_rows, c_img_cols, sumTex_step );
int t92 = read_sumTex( sumTex, sampler, (int2)(x + r9, y + r2), c_img_rows, c_img_cols, sumTex_step );
int t97 = read_sumTex( sumTex, sampler, (int2)(x + r9, y + r7), c_img_rows, c_img_cols, sumTex_step );
int t02 = read_sumTex( (int2)(x, y + r2));
int t07 = read_sumTex( (int2)(x, y + r7));
int t32 = read_sumTex( (int2)(x + r3, y + r2));
int t37 = read_sumTex( (int2)(x + r3, y + r7));
int t62 = read_sumTex( (int2)(x + r6, y + r2));
int t67 = read_sumTex( (int2)(x + r6, y + r7));
int t92 = read_sumTex( (int2)(x + r9, y + r2));
int t97 = read_sumTex( (int2)(x + r9, y + r7));
d = calcAxisAlignedDerivative(t02, t07, t32, t37, (r3) * (r7 - r2),
t62, t67, t92, t97, (r9 - r6) * (r7 - r2),
......@@ -253,14 +235,14 @@ __kernel void icvCalcLayerDetAndTrace(
{
// Some of the pixels needed to compute the derivative are
// repeated, so we only don't duplicate the fetch here.
int t20 = read_sumTex( sumTex, sampler, (int2)(x + r2, y), c_img_rows, c_img_cols, sumTex_step );
int t23 = read_sumTex( sumTex, sampler, (int2)(x + r2, y + r3), c_img_rows, c_img_cols, sumTex_step );
int t70 = read_sumTex( sumTex, sampler, (int2)(x + r7, y), c_img_rows, c_img_cols, sumTex_step );
int t73 = read_sumTex( sumTex, sampler, (int2)(x + r7, y + r3), c_img_rows, c_img_cols, sumTex_step );
int t26 = read_sumTex( sumTex, sampler, (int2)(x + r2, y + r6), c_img_rows, c_img_cols, sumTex_step );
int t76 = read_sumTex( sumTex, sampler, (int2)(x + r7, y + r6), c_img_rows, c_img_cols, sumTex_step );
int t29 = read_sumTex( sumTex, sampler, (int2)(x + r2, y + r9), c_img_rows, c_img_cols, sumTex_step );
int t79 = read_sumTex( sumTex, sampler, (int2)(x + r7, y + r9), c_img_rows, c_img_cols, sumTex_step );
int t20 = read_sumTex( (int2)(x + r2, y) );
int t23 = read_sumTex( (int2)(x + r2, y + r3) );
int t70 = read_sumTex( (int2)(x + r7, y) );
int t73 = read_sumTex( (int2)(x + r7, y + r3) );
int t26 = read_sumTex( (int2)(x + r2, y + r6) );
int t76 = read_sumTex( (int2)(x + r7, y + r6) );
int t29 = read_sumTex( (int2)(x + r2, y + r9) );
int t79 = read_sumTex( (int2)(x + r7, y + r9) );
d = calcAxisAlignedDerivative(t20, t23, t70, t73, (r7 - r2) * (r3),
t26, t29, t76, t79, (r7 - r2) * (r9 - r6),
......@@ -274,31 +256,31 @@ __kernel void icvCalcLayerDetAndTrace(
// There's no saving us here, we just have to get all of the pixels in
// separate fetches
F t = 0;
t += read_sumTex( sumTex, sampler, (int2)(x + r1, y + r1), c_img_rows, c_img_cols, sumTex_step );
t -= read_sumTex( sumTex, sampler, (int2)(x + r1, y + r4), c_img_rows, c_img_cols, sumTex_step );
t -= read_sumTex( sumTex, sampler, (int2)(x + r4, y + r1), c_img_rows, c_img_cols, sumTex_step );
t += read_sumTex( sumTex, sampler, (int2)(x + r4, y + r4), c_img_rows, c_img_cols, sumTex_step );
t += read_sumTex( (int2)(x + r1, y + r1) );
t -= read_sumTex( (int2)(x + r1, y + r4) );
t -= read_sumTex( (int2)(x + r4, y + r1) );
t += read_sumTex( (int2)(x + r4, y + r4) );
d += t / ((r4 - r1) * (r4 - r1));
t = 0;
t += read_sumTex( sumTex, sampler, (int2)(x + r5, y + r1), c_img_rows, c_img_cols, sumTex_step );
t -= read_sumTex( sumTex, sampler, (int2)(x + r5, y + r4), c_img_rows, c_img_cols, sumTex_step );
t -= read_sumTex( sumTex, sampler, (int2)(x + r8, y + r1), c_img_rows, c_img_cols, sumTex_step );
t += read_sumTex( sumTex, sampler, (int2)(x + r8, y + r4), c_img_rows, c_img_cols, sumTex_step );
t += read_sumTex( (int2)(x + r5, y + r1) );
t -= read_sumTex( (int2)(x + r5, y + r4) );
t -= read_sumTex( (int2)(x + r8, y + r1) );
t += read_sumTex( (int2)(x + r8, y + r4) );
d -= t / ((r8 - r5) * (r4 - r1));
t = 0;
t += read_sumTex( sumTex, sampler, (int2)(x + r1, y + r5), c_img_rows, c_img_cols, sumTex_step );
t -= read_sumTex( sumTex, sampler, (int2)(x + r1, y + r8), c_img_rows, c_img_cols, sumTex_step );
t -= read_sumTex( sumTex, sampler, (int2)(x + r4, y + r5), c_img_rows, c_img_cols, sumTex_step );
t += read_sumTex( sumTex, sampler, (int2)(x + r4, y + r8), c_img_rows, c_img_cols, sumTex_step );
t += read_sumTex( (int2)(x + r1, y + r5) );
t -= read_sumTex( (int2)(x + r1, y + r8) );
t -= read_sumTex( (int2)(x + r4, y + r5) );
t += read_sumTex( (int2)(x + r4, y + r8) );
d -= t / ((r4 - r1) * (r8 - r5));
t = 0;
t += read_sumTex( sumTex, sampler, (int2)(x + r5, y + r5), c_img_rows, c_img_cols, sumTex_step );
t -= read_sumTex( sumTex, sampler, (int2)(x + r5, y + r8), c_img_rows, c_img_cols, sumTex_step );
t -= read_sumTex( sumTex, sampler, (int2)(x + r8, y + r5), c_img_rows, c_img_cols, sumTex_step );
t += read_sumTex( sumTex, sampler, (int2)(x + r8, y + r8), c_img_rows, c_img_cols, sumTex_step );
t += read_sumTex( (int2)(x + r5, y + r5) );
t -= read_sumTex( (int2)(x + r5, y + r8) );
t -= read_sumTex( (int2)(x + r8, y + r5) );
t += read_sumTex( (int2)(x + r8, y + r8) );
d += t / ((r8 - r5) * (r8 - r5));
}
const float dxy = (float)d;
......@@ -311,171 +293,17 @@ __kernel void icvCalcLayerDetAndTrace(
////////////////////////////////////////////////////////////////////////
// NONMAX
__constant float c_DM[5] = {0, 0, 9, 9, 1};
bool within_check(IMAGE_INT32 maskSumTex, int sum_i, int sum_j, int size, int rows, int cols, int step)
{
float ratio = (float)size / 9.0f;
float d = 0;
int dx1 = round(ratio * c_DM[0]);
int dy1 = round(ratio * c_DM[1]);
int dx2 = round(ratio * c_DM[2]);
int dy2 = round(ratio * c_DM[3]);
float t = 0;
t += read_sumTex(maskSumTex, sampler, (int2)(sum_j + dx1, sum_i + dy1), rows, cols, step);
t -= read_sumTex(maskSumTex, sampler, (int2)(sum_j + dx1, sum_i + dy2), rows, cols, step);
t -= read_sumTex(maskSumTex, sampler, (int2)(sum_j + dx2, sum_i + dy1), rows, cols, step);
t += read_sumTex(maskSumTex, sampler, (int2)(sum_j + dx2, sum_i + dy2), rows, cols, step);
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 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,
IMAGE_INT32 maskSumTex,
int mask_step
)
{
volatile __local float N9[768]; // threads.x * threads.y * 3
det_step /= sizeof(*det);
trace_step /= sizeof(*trace);
maxCounter += counter_offset;
mask_step /= sizeof(uint);
// 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, c_img_rows, c_img_cols, mask_step))
{
// 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)
{
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(
void SURF_findMaximaInLayer(
__global float * det,
int det_step, int det_offset,
__global float * trace,
int trace_step, int trace_offset,
__global int4 * maxPosBuffer,
volatile __global 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 img_rows,
int img_cols,
int c_nOctaveLayers,
int c_octave,
int c_layer_rows,
......@@ -509,8 +337,8 @@ void icvFindMaximaInLayer(
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);
int l_x = min(max(j, 0), img_cols - 1);
int l_y = c_layer_rows * layer + min(max(i, 0), img_rows - 1);
N9[localLin - zoff] =
det[det_step * (l_y - c_layer_rows) + l_x];
......@@ -590,7 +418,7 @@ inline bool solve3x3_float(const float4 *A, const float *b, float *x)
if (det != 0)
{
F invdet = 1.0 / det;
F invdet = 1.0f / det;
x[0] = invdet *
(b[0] * (A[1].y * A[2].z - A[1].z * A[2].y) -
......@@ -624,15 +452,15 @@ inline bool solve3x3_float(const float4 *A, const float *b, float *x)
////////////////////////////////////////////////////////////////////////
// INTERPOLATION
__kernel
void icvInterpolateKeypoint(
void SURF_interpolateKeypoint(
__global const float * det,
int det_step, int det_offset,
__global const int4 * maxPosBuffer,
__global float * keypoints,
volatile __global int * featureCounter,
int det_step,
int keypoints_step,
int c_img_rows,
int c_img_cols,
int keypoints_step, int keypoints_offset,
volatile __global int* featureCounter,
int img_rows,
int img_cols,
int c_octave,
int c_layer_rows,
int c_max_features
......@@ -724,7 +552,7 @@ void icvInterpolateKeypoint(
const int grad_wav_size = 2 * round(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)
if ((img_rows + 1) >= grad_wav_size && (img_cols + 1) >= grad_wav_size)
{
// Get a new feature index.
int ind = atomic_inc(featureCounter);
......@@ -829,23 +657,19 @@ void reduce_32_sum(volatile __local float * data, volatile float* partial_reduc
}
__kernel
void icvCalcOrientation(
IMAGE_INT32 sumTex,
__global float * keypoints,
int keypoints_step,
int c_img_rows,
int c_img_cols,
int sum_step
)
void SURF_calcOrientation(
__PARAM_sumTex__, int img_rows, int img_cols,
__global float * keypoints, int keypoints_step, int keypoints_offset )
{
keypoints_step /= sizeof(*keypoints);
#ifndef HAVE_IMAGE2D
sum_step /= sizeof(uint);
#endif
__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;
__local float s_X[ORI_SAMPLES];
__local float s_Y[ORI_SAMPLES];
__local float s_angle[ORI_SAMPLES];
......@@ -860,7 +684,6 @@ void icvCalcOrientation(
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
......@@ -868,7 +691,7 @@ void icvCalcOrientation(
const int grad_wav_size = 2 * round(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)
if ((img_rows + 1) < grad_wav_size || (img_cols + 1) < grad_wav_size)
return;
// Calc X, Y, angle and store it to shared memory
......@@ -880,8 +703,8 @@ void icvCalcOrientation(
float ratio = (float)grad_wav_size / 4;
int r2 = round(ratio * 2.0);
int r4 = round(ratio * 4.0);
int r2 = round(ratio * 2.0f);
int r4 = round(ratio * 4.0f);
for (int i = tid; i < ORI_SAMPLES; i += ORI_LOCAL_SIZE )
{
float X = 0.0f, Y = 0.0f, angle = 0.0f;
......@@ -889,21 +712,20 @@ void icvCalcOrientation(
const int x = round(featureX[get_group_id(0)] + c_aptX[i] * s - margin);
const int y = round(featureY[get_group_id(0)] + c_aptY[i] * s - margin);
if (y >= 0 && y < (c_img_rows + 1) - grad_wav_size &&
x >= 0 && x < (c_img_cols + 1) - grad_wav_size)
if (y >= 0 && y < (img_rows + 1) - grad_wav_size &&
x >= 0 && x < (img_cols + 1) - grad_wav_size)
{
float apt = c_aptW[i];
// Compute the haar sum without fetching duplicate pixels.
float t00 = read_sumTex( sumTex, sampler, (int2)(x, y), c_img_rows, c_img_cols, sum_step);
float t02 = read_sumTex( sumTex, sampler, (int2)(x, y + r2), c_img_rows, c_img_cols, sum_step);
float t04 = read_sumTex( sumTex, sampler, (int2)(x, y + r4), c_img_rows, c_img_cols, sum_step);
float t20 = read_sumTex( sumTex, sampler, (int2)(x + r2, y), c_img_rows, c_img_cols, sum_step);
float t24 = read_sumTex( sumTex, sampler, (int2)(x + r2, y + r4), c_img_rows, c_img_cols, sum_step);
float t40 = read_sumTex( sumTex, sampler, (int2)(x + r4, y), c_img_rows, c_img_cols, sum_step);
float t42 = read_sumTex( sumTex, sampler, (int2)(x + r4, y + r2), c_img_rows, c_img_cols, sum_step);
float t44 = read_sumTex( sumTex, sampler, (int2)(x + r4, y + r4), c_img_rows, c_img_cols, sum_step);
float t00 = read_sumTex( (int2)(x, y));
float t02 = read_sumTex( (int2)(x, y + r2));
float t04 = read_sumTex( (int2)(x, y + r4));
float t20 = read_sumTex( (int2)(x + r2, y));
float t24 = read_sumTex( (int2)(x + r2, y + r4));
float t40 = read_sumTex( (int2)(x + r4, y));
float t42 = read_sumTex( (int2)(x + r4, y + r2));
float t44 = read_sumTex( (int2)(x + r4, y + r4));
F t = t00 - t04 - t20 + t24;
X -= t / ((r2) * (r4));
......@@ -995,18 +817,17 @@ void icvCalcOrientation(
}
__kernel
void icvSetUpright(
void SURF_setUpRight(
__global float * keypoints,
int keypoints_step,
int nFeatures
)
int keypoints_step, int keypoints_offset,
int rows, int cols )
{
int i = get_global_id(0);
keypoints_step /= sizeof(*keypoints);
__global float* featureDir = keypoints + ANGLE_ROW * keypoints_step;
if(get_global_id(0) <= nFeatures)
if(i < cols)
{
featureDir[get_global_id(0)] = 270.0f;
keypoints[mad24(keypoints_step, ANGLE_ROW, i)] = 270.f;
}
}
......@@ -1045,22 +866,14 @@ __constant float c_DW[PATCH_SZ * PATCH_SZ] =
};
// utility for linear filter
inline uchar readerGet(
IMAGE_INT8 src,
const float centerX, const float centerY, const float win_offset, const float cos_dir, const float sin_dir,
int i, int j, int rows, int cols, int elemPerRow
)
{
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 read_imgTex(src, sampler, (float2)(pixel_x, pixel_y), rows, cols, elemPerRow);
}
#define readerGet(centerX, centerY, win_offset, cos_dir, sin_dir, i, j) \
read_imgTex((float2)(centerX + (win_offset + j) * cos_dir + (win_offset + i) * sin_dir, \
centerY - (win_offset + j) * sin_dir + (win_offset + i) * cos_dir))
inline float linearFilter(
IMAGE_INT8 src,
const float centerX, const float centerY, const float win_offset, const float cos_dir, const float sin_dir,
float y, float x, int rows, int cols, int elemPerRow
)
__PARAM_imgTex__, int img_rows, int img_cols,
float centerX, float centerY, float win_offset,
float cos_dir, float sin_dir, float y, float x )
{
x -= 0.5f;
y -= 0.5f;
......@@ -1072,34 +885,31 @@ inline float linearFilter(
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, rows, cols, elemPerRow);
uchar src_reg = readerGet(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, rows, cols, elemPerRow);
src_reg = readerGet(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, rows, cols, elemPerRow);
src_reg = readerGet(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, rows, cols, elemPerRow);
src_reg = readerGet(centerX, centerY, win_offset, cos_dir, sin_dir, y2, x2);
out = out + src_reg * ((x - x1) * (y - y1));
return out;
}
void calc_dx_dy(
IMAGE_INT8 imgTex,
__PARAM_imgTex__,
int img_rows, int img_cols,
volatile __local float *s_dx_bin,
volatile __local float *s_dy_bin,
volatile __local float *s_PATCH,
__global const float* featureX,
__global const float* featureY,
__global const float* featureSize,
__global const float* featureDir,
int rows,
int cols,
int elemPerRow
)
__global const float* featureDir )
{
const float centerX = featureX[get_group_id(0)];
const float centerY = featureY[get_group_id(0)];
......@@ -1136,7 +946,9 @@ void calc_dx_dy(
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) * 6 + get_local_id(0)] = linearFilter(imgTex, centerX, centerY, win_offset, cos_dir, sin_dir, icoo, jcoo, rows, cols, elemPerRow);
s_PATCH[get_local_id(1) * 6 + get_local_id(0)] =
linearFilter(__PASS_imgTex__, img_rows, img_cols, centerX, centerY,
win_offset, cos_dir, sin_dir, icoo, jcoo);
barrier(CLK_LOCAL_MEM_FENCE);
......@@ -1162,6 +974,7 @@ void calc_dx_dy(
s_dy_bin[tid] = vy;
}
}
void reduce_sum25(
volatile __local float* sdata1,
volatile __local float* sdata2,
......@@ -1225,16 +1038,13 @@ void reduce_sum25(
}
__kernel
void compute_descriptors64(
IMAGE_INT8 imgTex,
void SURF_computeDescriptors64(
__PARAM_imgTex__,
int img_rows, int img_cols,
__global const float* keypoints,
int keypoints_step, int keypoints_offset,
__global float * descriptors,
__global const float * keypoints,
int descriptors_step,
int keypoints_step,
int rows,
int cols,
int img_step
)
int descriptors_step, int descriptors_offset)
{
descriptors_step /= sizeof(float);
keypoints_step /= sizeof(float);
......@@ -1250,7 +1060,7 @@ void compute_descriptors64(
volatile __local float sdyabs[25];
volatile __local float s_PATCH[6*6];
calc_dx_dy(imgTex, sdx, sdy, s_PATCH, featureX, featureY, featureSize, featureDir, rows, cols, img_step);
calc_dx_dy(__PASS_imgTex__, img_rows, img_cols, 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);
......@@ -1279,17 +1089,15 @@ void compute_descriptors64(
}
}
}
__kernel
void compute_descriptors128(
IMAGE_INT8 imgTex,
__global float * descriptors,
__global float * keypoints,
int descriptors_step,
int keypoints_step,
int rows,
int cols,
int img_step
)
void SURF_computeDescriptors128(
__PARAM_imgTex__,
int img_rows, int img_cols,
__global const float* keypoints,
int keypoints_step, int keypoints_offset,
__global float* descriptors,
int descriptors_step, int descriptors_offset)
{
descriptors_step /= sizeof(*descriptors);
keypoints_step /= sizeof(*keypoints);
......@@ -1310,7 +1118,7 @@ void compute_descriptors128(
volatile __local float sdabs2[25];
volatile __local float s_PATCH[6*6];
calc_dx_dy(imgTex, sdx, sdy, s_PATCH, featureX, featureY, featureSize, featureDir, rows, cols, img_step);
calc_dx_dy(__PASS_imgTex__, img_rows, img_cols, 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);
......@@ -1483,7 +1291,7 @@ void reduce_sum64(volatile __local float* smem, int tid)
}
__kernel
void normalize_descriptors128(__global float * descriptors, int descriptors_step)
void SURF_normalizeDescriptors128(__global float * descriptors, int descriptors_step, int descriptors_offset)
{
descriptors_step /= sizeof(*descriptors);
// no need for thread ID
......@@ -1509,8 +1317,9 @@ void normalize_descriptors128(__global float * descriptors, int descriptors_step
// normalize and store in output
descriptor_base[get_local_id(0)] = lookup / len;
}
__kernel
void normalize_descriptors64(__global float * descriptors, int descriptors_step)
void SURF_normalizeDescriptors64(__global float * descriptors, int descriptors_step, int descriptors_offset)
{
descriptors_step /= sizeof(*descriptors);
// no need for thread ID
......
......@@ -60,11 +60,6 @@
# include "opencv2/cudaarithm.hpp"
#endif
#ifdef HAVE_OPENCV_OCL
# include "opencv2/nonfree/ocl.hpp"
# include "opencv2/ocl/private/util.hpp"
#endif
#include "opencv2/core/private.hpp"
#endif
......@@ -108,6 +108,7 @@ Modifications by Ian Mahon
*/
#include "precomp.hpp"
#include "surf.hpp"
namespace cv
{
......@@ -897,11 +898,42 @@ void SURF::operator()(InputArray _img, InputArray _mask,
OutputArray _descriptors,
bool useProvidedKeypoints) const
{
Mat img = _img.getMat(), mask = _mask.getMat(), mask1, sum, msum;
int imgtype = _img.type(), imgcn = CV_MAT_CN(imgtype);
bool doDescriptors = _descriptors.needed();
CV_Assert(!img.empty() && img.depth() == CV_8U);
if( img.channels() > 1 )
CV_Assert(!_img.empty() && CV_MAT_DEPTH(imgtype) == CV_8U && (imgcn == 1 || imgcn == 3 || imgcn == 4));
CV_Assert(_descriptors.needed() || !useProvidedKeypoints);
if( ocl::useOpenCL() )
{
SURF_OCL ocl_surf;
UMat gpu_kpt;
bool ok = ocl_surf.init(this);
if( ok )
{
if( !_descriptors.needed() )
{
ok = ocl_surf.detect(_img, _mask, gpu_kpt);
}
else
{
if(useProvidedKeypoints)
ocl_surf.uploadKeypoints(keypoints, gpu_kpt);
ok = ocl_surf.detectAndCompute(_img, _mask, gpu_kpt, _descriptors, useProvidedKeypoints);
}
}
if( ok )
{
if(!useProvidedKeypoints)
ocl_surf.downloadKeypoints(gpu_kpt, keypoints);
return;
}
}
Mat img = _img.getMat(), mask = _mask.getMat(), mask1, sum, msum;
if( imgcn > 1 )
cvtColor(img, img, COLOR_BGR2GRAY);
CV_Assert(mask.empty() || (mask.type() == CV_8U && mask.size() == img.size()));
......
///////////// see LICENSE.txt in the OpenCV root directory //////////////
#ifndef __OPENCV_NONFREE_SURF_HPP__
#define __OPENCV_NONFREE_SURF_HPP__
namespace cv
{
//! Speeded up robust features, port from CUDA module.
////////////////////////////////// SURF //////////////////////////////////////////
class SURF_OCL
{
public:
enum KeypointLayout
{
X_ROW = 0,
Y_ROW,
LAPLACIAN_ROW,
OCTAVE_ROW,
SIZE_ROW,
ANGLE_ROW,
HESSIAN_ROW,
ROWS_COUNT
};
//! the full constructor taking all the necessary parameters
SURF_OCL();
bool init(const SURF* params);
//! returns the descriptor size in float's (64 or 128)
int descriptorSize() const { return params->extended ? 128 : 64; }
void uploadKeypoints(const std::vector<KeyPoint> &keypoints, UMat &keypointsGPU);
void downloadKeypoints(const UMat &keypointsGPU, std::vector<KeyPoint> &keypoints);
//! 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
bool detect(InputArray img, InputArray mask, UMat& keypoints);
//! finds the keypoints and computes their descriptors.
//! Optionally it can compute descriptors for the user-provided keypoints and recompute keypoints direction
bool detectAndCompute(InputArray img, InputArray mask, UMat& keypoints,
OutputArray descriptors, bool useProvidedKeypoints = false);
protected:
bool setImage(InputArray img, InputArray mask);
// kernel callers declarations
bool calcLayerDetAndTrace(int octave, int layer_rows);
bool findMaximaInLayer(int counterOffset, int octave, int layer_rows, int layer_cols);
bool interpolateKeypoint(int maxCounter, UMat &keypoints, int octave, int layer_rows, int maxFeatures);
bool calcOrientation(UMat &keypoints);
bool setUpRight(UMat &keypoints);
bool computeDescriptors(const UMat &keypoints, OutputArray descriptors);
bool detectKeypoints(UMat &keypoints);
const SURF* params;
int refcount;
//! max keypoints = min(keypointsRatio * img.size().area(), 65535)
UMat sum, intBuffer;
UMat det, trace;
UMat maxPosBuffer;
int img_cols, img_rows;
int maxCandidates;
int maxFeatures;
UMat img, counters;
// texture buffers
ocl::Image2D imgTex, sumTex;
bool haveImageSupport;
String kerOpts;
int status;
};
/*
template<typename _Tp> void copyVectorToUMat(const std::vector<_Tp>& v, UMat& um)
{
if(v.empty())
um.release();
else
Mat(1, (int)(v.size()*sizeof(v[0])), CV_8U, (void*)&v[0]).copyTo(um);
}
template<typename _Tp> void copyUMatToVector(const UMat& um, std::vector<_Tp>& v)
{
if(um.empty())
v.clear();
else
{
size_t sz = um.total()*um.elemSize();
CV_Assert(um.isContinuous() && (sz % sizeof(_Tp) == 0));
v.resize(sz/sizeof(_Tp));
Mat m(um.size(), um.type(), &v[0]);
um.copyTo(m);
}
}*/
}
#endif
......@@ -43,42 +43,16 @@
//
//M*/
#include "precomp.hpp"
#include "surf.hpp"
#ifdef HAVE_OPENCV_OCL
#include <cstdio>
#include <sstream>
#include "opencl_kernels.hpp"
using namespace cv;
using namespace cv::ocl;
static ProgramEntry surfprog = cv::ocl::nonfree::surf;
namespace cv
{
namespace ocl
{
// The number of degrees between orientation samples in calcOrientation
const static int ORI_SEARCH_INC = 5;
// The local size of the calcOrientation kernel
const static int ORI_LOCAL_SIZE = (360 / ORI_SEARCH_INC);
static void openCLExecuteKernelSURF(Context *clCxt, const cv::ocl::ProgramEntry* source, String kernelName, size_t globalThreads[3],
size_t localThreads[3], std::vector< std::pair<size_t, const void *> > &args, int channels, int depth)
{
std::stringstream optsStr;
optsStr << "-D ORI_LOCAL_SIZE=" << ORI_LOCAL_SIZE << " ";
optsStr << "-D ORI_SEARCH_INC=" << ORI_SEARCH_INC << " ";
cl_kernel kernel;
kernel = openCLGetKernelFromSource(clCxt, source, kernelName, optsStr.str().c_str());
size_t wave_size = queryWaveFrontSize(kernel);
CV_Assert(clReleaseKernel(kernel) == CL_SUCCESS);
optsStr << "-D WAVE_SIZE=" << wave_size;
openCLExecuteKernel(clCxt, source, kernelName, globalThreads, localThreads, args, channels, depth, optsStr.str().c_str());
}
}
}
enum { ORI_SEARCH_INC=5, ORI_LOCAL_SIZE=(360 / ORI_SEARCH_INC) };
static inline int calcSize(int octave, int layer)
{
......@@ -96,223 +70,208 @@ static inline int calcSize(int octave, int layer)
}
class SURF_OCL_Invoker
SURF_OCL::SURF_OCL()
{
public:
// facilities
void bindImgTex(const oclMat &img, cl_mem &texture);
//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 declarations
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, int maxCounter,
oclMat &keypoints, oclMat &counters, int octave, int layer_rows, int maxFeatures);
void icvCalcOrientation_gpu(const oclMat &keypoints, int nFeatures);
void icvSetUpright_gpu(const oclMat &keypoints, int nFeatures);
void compute_descriptors_gpu(const oclMat &descriptors, const oclMat &keypoints, int nFeatures);
// end of kernel callers declarations
img_cols = img_rows = maxCandidates = maxFeatures = 0;
haveImageSupport = false;
status = -1;
}
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()), counters(oclMat()),
imgTex(NULL), sumTex(NULL), maskSumTex(NULL), _img(img)
bool SURF_OCL::init(const SURF* p)
{
params = p;
if(status < 0)
{
status = 0;
if(ocl::haveOpenCL())
{
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 ocl::Device& dev = ocl::Device::getDefault();
if( dev.type() == ocl::Device::TYPE_CPU || dev.doubleFPConfig() == 0 )
return false;
haveImageSupport = false;//dev.imageSupport();
kerOpts = haveImageSupport ? "-D HAVE_IMAGE2D -D DOUBLE_SUPPORT" : "";
status = 1;
}
}
return status > 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;
bool SURF_OCL::setImage(InputArray _img, InputArray _mask)
{
if( status <= 0 )
return false;
if( !_mask.empty())
return false;
int imgtype = _img.type();
CV_Assert(!_img.empty());
CV_Assert(params && params->nOctaves > 0 && params->nOctaveLayers > 0);
int min_size = calcSize(params->nOctaves - 1, 0);
Size sz = _img.size();
img_cols = sz.width;
img_rows = sz.height;
CV_Assert(img_rows >= min_size && img_cols >= min_size);
const int layer_rows = img_rows >> (params->nOctaves - 1);
const int layer_cols = img_cols >> (params->nOctaves - 1);
const int min_margin = ((calcSize((params->nOctaves - 1), 2) >> 1) >> (params->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);
maxFeatures = std::min(static_cast<int>(img_cols*img_rows * 0.01f), 65535);
maxCandidates = std::min(static_cast<int>(1.5 * maxFeatures), 65535);
CV_Assert(maxFeatures > 0);
counters.create(1, surf_.nOctaves + 1, CV_32SC1);
counters.create(1, params->nOctaves + 1, CV_32SC1);
counters.setTo(Scalar::all(0));
integral(img, surf_.sum);
bindImgTex(img, imgTex);
bindImgTex(surf_.sum, sumTex);
finish();
img.release();
if(_img.isUMat() && imgtype == CV_8UC1)
img = _img.getUMat();
else if( imgtype == CV_8UC1 )
_img.copyTo(img);
else
cvtColor(_img, img, COLOR_BGR2GRAY);
maskSumTex = 0;
integral(img, sum);
if (use_mask)
if(haveImageSupport)
{
CV_Error(Error::StsBadFunc, "Masked SURF detector is not implemented yet");
//!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);
//bindImgTex(surf_.maskSum, maskSumTex);
}
imgTex = ocl::Image2D(img);
sumTex = ocl::Image2D(sum);
}
void detectKeypoints(oclMat &keypoints)
{
return true;
}
bool SURF_OCL::detectKeypoints(UMat &keypoints)
{
// create image pyramid buffers
// different layers have same sized buffers, but they are sampled from Gaussian kernel.
ensureSizeIsEnough(img_rows * (surf_.nOctaveLayers + 2), img_cols, CV_32FC1, surf_.det);
ensureSizeIsEnough(img_rows * (surf_.nOctaveLayers + 2), img_cols, CV_32FC1, surf_.trace);
det.create(img_rows * (params->nOctaveLayers + 2), img_cols, CV_32F);
trace.create(img_rows * (params->nOctaveLayers + 2), img_cols, CV_32FC1);
ensureSizeIsEnough(1, maxCandidates, CV_32SC4, surf_.maxPosBuffer);
ensureSizeIsEnough(SURF_OCL::ROWS_COUNT, maxFeatures, CV_32FC1, keypoints);
maxPosBuffer.create(1, maxCandidates, CV_32SC4);
keypoints.create(SURF_OCL::ROWS_COUNT, maxFeatures, CV_32F);
keypoints.setTo(Scalar::all(0));
Mat cpuCounters;
for (int octave = 0; octave < surf_.nOctaves; ++octave)
for (int octave = 0; octave < params->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);
if(!calcLayerDetAndTrace(octave, layer_rows))
return false;
icvFindMaximaInLayer_gpu(surf_.det, surf_.trace, surf_.maxPosBuffer, counters, 1 + octave,
octave, use_mask, surf_.nOctaveLayers, layer_rows, layer_cols);
if(!findMaximaInLayer(1 + octave, octave, layer_rows, layer_cols))
return false;
int maxCounter = ((Mat)counters).at<int>(1 + octave);
maxCounter = std::min(maxCounter, static_cast<int>(maxCandidates));
cpuCounters = counters.getMat(ACCESS_READ);
int maxCounter = cpuCounters.at<int>(1 + octave);
maxCounter = std::min(maxCounter, maxCandidates);
cpuCounters.release();
if (maxCounter > 0)
{
icvInterpolateKeypoint_gpu(surf_.det, surf_.maxPosBuffer, maxCounter,
keypoints, counters, octave, layer_rows, maxFeatures);
if(!interpolateKeypoint(maxCounter, keypoints, octave, layer_rows, maxFeatures))
return false;
}
}
int featureCounter = Mat(counters).at<int>(0);
featureCounter = std::min(featureCounter, static_cast<int>(maxFeatures));
keypoints.cols = featureCounter;
cpuCounters = counters.getMat(ACCESS_READ);
int featureCounter = cpuCounters.at<int>(0);
featureCounter = std::min(featureCounter, maxFeatures);
cpuCounters.release();
if (surf_.upright)
{
//keypoints.row(SURF_OCL::ANGLE_ROW).setTo(Scalar::all(90.0));
setUpright(keypoints);
}
keypoints = UMat(keypoints, Rect(0, 0, featureCounter, keypoints.rows));
if (params->upright)
return setUpRight(keypoints);
else
{
findOrientation(keypoints);
}
}
return calcOrientation(keypoints);
}
void setUpright(oclMat &keypoints)
{
const int nFeatures = keypoints.cols;
if(nFeatures > 0)
{
icvSetUpright_gpu(keypoints, keypoints.cols);
}
}
void findOrientation(oclMat &keypoints)
{
const int nFeatures = keypoints.cols;
if (nFeatures > 0)
bool SURF_OCL::setUpRight(UMat &keypoints)
{
int nFeatures = keypoints.cols;
if( nFeatures == 0 )
return true;
size_t globalThreads[3] = {nFeatures, 1};
ocl::Kernel kerUpRight("SURF_setUpRight", ocl::nonfree::surf_oclsrc, kerOpts);
return kerUpRight.args(ocl::KernelArg::ReadWrite(keypoints)).run(2, globalThreads, 0, true);
}
bool SURF_OCL::computeDescriptors(const UMat &keypoints, OutputArray _descriptors)
{
int dsize = params->descriptorSize();
int nFeatures = keypoints.cols;
if (nFeatures == 0)
{
icvCalcOrientation_gpu(keypoints, nFeatures);
}
_descriptors.release();
return true;
}
_descriptors.create(nFeatures, dsize, CV_32F);
UMat descriptors;
if( _descriptors.isUMat() )
descriptors = _descriptors.getUMat();
else
descriptors.create(nFeatures, dsize, CV_32F);
void computeDescriptors(const oclMat &keypoints, oclMat &descriptors, int descriptorSize)
{
const int nFeatures = keypoints.cols;
if (nFeatures > 0)
ocl::Kernel kerCalcDesc, kerNormDesc;
if( dsize == 64 )
{
ensureSizeIsEnough(nFeatures, descriptorSize, CV_32F, descriptors);
compute_descriptors_gpu(descriptors, keypoints, nFeatures);
}
kerCalcDesc.create("SURF_computeDescriptors64", ocl::nonfree::surf_oclsrc, kerOpts);
kerNormDesc.create("SURF_normalizeDescriptors64", ocl::nonfree::surf_oclsrc, kerOpts);
}
~SURF_OCL_Invoker()
else
{
if(imgTex)
openCLFree(imgTex);
if(sumTex)
openCLFree(sumTex);
if(maskSumTex)
openCLFree(maskSumTex);
CV_Assert(dsize == 128);
kerCalcDesc.create("SURF_computeDescriptors128", ocl::nonfree::surf_oclsrc, kerOpts);
kerNormDesc.create("SURF_normalizeDescriptors128", ocl::nonfree::surf_oclsrc, kerOpts);
}
private:
SURF_OCL &surf_;
int img_cols, img_rows;
bool use_mask;
size_t localThreads[] = {6, 6};
size_t globalThreads[] = {nFeatures*localThreads[0], localThreads[1]};
int maxCandidates;
int maxFeatures;
oclMat counters;
// texture buffers
cl_mem imgTex;
cl_mem sumTex;
cl_mem maskSumTex;
const oclMat _img; // make a copy for non-image2d_t supported platform
SURF_OCL_Invoker &operator= (const SURF_OCL_Invoker &right)
if(haveImageSupport)
{
(*this) = right;
return *this;
} // remove warning C4512
};
kerCalcDesc.args(imgTex,
img_rows, img_cols,
ocl::KernelArg::ReadOnlyNoSize(keypoints),
ocl::KernelArg::WriteOnlyNoSize(descriptors));
}
else
{
kerCalcDesc.args(ocl::KernelArg::ReadOnlyNoSize(img),
img_rows, img_cols,
ocl::KernelArg::ReadOnlyNoSize(keypoints),
ocl::KernelArg::WriteOnlyNoSize(descriptors));
}
cv::ocl::SURF_OCL::SURF_OCL()
{
hessianThreshold = 100.0f;
extended = true;
nOctaves = 4;
nOctaveLayers = 2;
keypointsRatio = 0.01f;
upright = false;
}
if(!kerCalcDesc.run(2, globalThreads, localThreads, true))
return false;
cv::ocl::SURF_OCL::SURF_OCL(double _threshold, int _nOctaves, int _nOctaveLayers, bool _extended, float _keypointsRatio, bool _upright)
{
hessianThreshold = saturate_cast<float>(_threshold);
extended = _extended;
nOctaves = _nOctaves;
nOctaveLayers = _nOctaveLayers;
keypointsRatio = _keypointsRatio;
upright = _upright;
}
size_t localThreads_n[] = {dsize, 1};
size_t globalThreads_n[] = {nFeatures*localThreads_n[0], localThreads_n[1]};
int cv::ocl::SURF_OCL::descriptorSize() const
{
return extended ? 128 : 64;
globalThreads[0] = nFeatures * localThreads[0];
globalThreads[1] = localThreads[1];
bool ok = kerNormDesc.args(ocl::KernelArg::ReadWriteNoSize(descriptors)).
run(2, globalThreads_n, localThreads_n, true);
if(ok && !_descriptors.isUMat())
descriptors.copyTo(_descriptors);
return ok;
}
int cv::ocl::SURF_OCL::defaultNorm() const
{
return NORM_L2;
}
void cv::ocl::SURF_OCL::uploadKeypoints(const std::vector<KeyPoint> &keypoints, oclMat &keypointsGPU)
void SURF_OCL::uploadKeypoints(const std::vector<KeyPoint> &keypoints, UMat &keypointsGPU)
{
if (keypoints.empty())
keypointsGPU.release();
......@@ -340,11 +299,11 @@ void cv::ocl::SURF_OCL::uploadKeypoints(const std::vector<KeyPoint> &keypoints,
kp_laplacian[i] = 1;
}
keypointsGPU.upload(keypointsCPU);
keypointsCPU.copyTo(keypointsGPU);
}
}
void cv::ocl::SURF_OCL::downloadKeypoints(const oclMat &keypointsGPU, std::vector<KeyPoint> &keypoints)
void SURF_OCL::downloadKeypoints(const UMat &keypointsGPU, std::vector<KeyPoint> &keypoints)
{
const int nFeatures = keypointsGPU.cols;
......@@ -354,8 +313,7 @@ void cv::ocl::SURF_OCL::downloadKeypoints(const oclMat &keypointsGPU, std::vecto
{
CV_Assert(keypointsGPU.type() == CV_32FC1 && keypointsGPU.rows == ROWS_COUNT);
Mat keypointsCPU(keypointsGPU);
Mat keypointsCPU = keypointsGPU.getMat(ACCESS_READ);
keypoints.resize(nFeatures);
float *kp_x = keypointsCPU.ptr<float>(SURF_OCL::X_ROW);
......@@ -380,354 +338,122 @@ void cv::ocl::SURF_OCL::downloadKeypoints(const oclMat &keypointsGPU, std::vecto
}
}
void cv::ocl::SURF_OCL::downloadDescriptors(const oclMat &descriptorsGPU, std::vector<float> &descriptors)
bool SURF_OCL::detect(InputArray _img, InputArray _mask, UMat& keypoints)
{
if (descriptorsGPU.empty())
descriptors.clear();
else
{
CV_Assert(descriptorsGPU.type() == CV_32F);
if( !setImage(_img, _mask) )
return false;
descriptors.resize(descriptorsGPU.rows * descriptorsGPU.cols);
Mat descriptorsCPU(descriptorsGPU.size(), CV_32F, &descriptors[0]);
descriptorsGPU.download(descriptorsCPU);
}
return detectKeypoints(keypoints);
}
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)
bool SURF_OCL::detectAndCompute(InputArray _img, InputArray _mask, UMat& keypoints,
OutputArray _descriptors, bool useProvidedKeypoints )
{
if (!img.empty())
{
SURF_OCL_Invoker surf(*this, img, mask);
if( !setImage(_img, _mask) )
return false;
if (!useProvidedKeypoints)
surf.detectKeypoints(keypoints);
else if (!upright)
{
surf.findOrientation(keypoints);
}
if( !useProvidedKeypoints && !detectKeypoints(keypoints) )
return false;
surf.computeDescriptors(keypoints, descriptors, descriptorSize());
}
return computeDescriptors(keypoints, _descriptors);
}
void cv::ocl::SURF_OCL::operator()(const oclMat &img, const oclMat &mask, std::vector<KeyPoint> &keypoints)
{
oclMat keypointsGPU;
(*this)(img, mask, keypointsGPU);
downloadKeypoints(keypointsGPU, keypoints);
}
void cv::ocl::SURF_OCL::operator()(const oclMat &img, const oclMat &mask, std::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, std::vector<KeyPoint> &keypoints,
std::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();
}
// bind source buffer to image oject.
void SURF_OCL_Invoker::bindImgTex(const oclMat &img, cl_mem &texture)
{
if(texture)
{
openCLFree(texture);
}
texture = bindTexture(img);
}
inline int divUp(int a, int b) { return (a + b-1)/b; }
////////////////////////////
// kernel caller definitions
void SURF_OCL_Invoker::icvCalcLayerDetAndTrace_gpu(oclMat &det, oclMat &trace, int octave, int nOctaveLayers, int c_layer_rows)
bool SURF_OCL::calcLayerDetAndTrace(int octave, int c_layer_rows)
{
int nOctaveLayers = params->nOctaveLayers;
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";
std::vector< std::pair<size_t, const void *> > args;
if(sumTex)
size_t localThreads[] = {16, 16};
size_t globalThreads[] =
{
divUp(max_samples_j, (int)localThreads[0]) * localThreads[0],
divUp(max_samples_i, (int)localThreads[1]) * localThreads[1] * (nOctaveLayers + 2)
};
ocl::Kernel kerCalcDetTrace("SURF_calcLayerDetAndTrace", ocl::nonfree::surf_oclsrc, kerOpts);
if(haveImageSupport)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&sumTex));
kerCalcDetTrace.args(sumTex,
img_rows, img_cols, nOctaveLayers,
octave, c_layer_rows,
ocl::KernelArg::WriteOnlyNoSize(det),
ocl::KernelArg::WriteOnlyNoSize(trace));
}
else
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&surf_.sum.data)); // if image2d is not supported
kerCalcDetTrace.args(ocl::KernelArg::ReadOnlyNoSize(sum),
img_rows, img_cols, nOctaveLayers,
octave, c_layer_rows,
ocl::KernelArg::WriteOnlyNoSize(det),
ocl::KernelArg::WriteOnlyNoSize(trace));
}
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&det.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&trace.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&det.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&trace.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&img_rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&img_cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&nOctaveLayers));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&octave));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&c_layer_rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&surf_.sum.step));
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
};
openCLExecuteKernelSURF(clCxt, &surfprog, kernelName, globalThreads, localThreads, args, -1, -1);
return kerCalcDetTrace.run(2, globalThreads, localThreads, true);
}
void SURF_OCL_Invoker::icvFindMaximaInLayer_gpu(const oclMat &det, const oclMat &trace, oclMat &maxPosBuffer, oclMat &maxCounter, int counterOffset,
int octave, bool useMask, int nLayers, int layer_rows, int layer_cols)
bool SURF_OCL::findMaximaInLayer(int counterOffset, int octave,
int layer_rows, int layer_cols)
{
const int min_margin = ((calcSize(octave, 2) >> 1) >> octave) + 1;
int nOctaveLayers = params->nOctaveLayers;
Context *clCxt = det.clCxt;
String kernelName = use_mask ? "icvFindMaximaInLayer_withmask" : "icvFindMaximaInLayer";
std::vector< std::pair<size_t, const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&det.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&trace.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&maxPosBuffer.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&maxCounter.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&counterOffset));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&det.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&trace.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&img_rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&img_cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&nLayers));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&octave));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&layer_rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&layer_cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&maxCandidates));
args.push_back( std::make_pair( sizeof(cl_float), (void *)&surf_.hessianThreshold));
if(useMask)
{
if(maskSumTex)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&maskSumTex));
}
else
size_t localThreads[3] = {16, 16};
size_t globalThreads[3] =
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&surf_.maskSum.data));
}
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&surf_.maskSum.step));
}
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
divUp(layer_cols - 2 * min_margin, (int)localThreads[0] - 2) * localThreads[0],
divUp(layer_rows - 2 * min_margin, (int)localThreads[1] - 2) * nOctaveLayers * localThreads[1]
};
openCLExecuteKernelSURF(clCxt, &surfprog, kernelName, globalThreads, localThreads, args, -1, -1);
ocl::Kernel kerFindMaxima("SURF_findMaximaInLayer", ocl::nonfree::surf_oclsrc, kerOpts);
return kerFindMaxima.args(ocl::KernelArg::ReadOnlyNoSize(det),
ocl::KernelArg::ReadOnlyNoSize(trace),
ocl::KernelArg::PtrReadWrite(maxPosBuffer),
ocl::KernelArg::PtrReadWrite(counters),
counterOffset, img_rows, img_cols,
octave, nOctaveLayers,
layer_rows, layer_cols,
maxCandidates,
(float)params->hessianThreshold).run(2, globalThreads, localThreads, true);
}
void SURF_OCL_Invoker::icvInterpolateKeypoint_gpu(const oclMat &det, const oclMat &maxPosBuffer, int maxCounter,
oclMat &keypoints, oclMat &counters_, int octave, int layer_rows, int max_features)
bool SURF_OCL::interpolateKeypoint(int maxCounter, UMat &keypoints, int octave, int layer_rows, int max_features)
{
Context *clCxt = det.clCxt;
String kernelName = "icvInterpolateKeypoint";
std::vector< std::pair<size_t, const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&det.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&maxPosBuffer.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&counters_.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&det.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&keypoints.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&img_rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&img_cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&octave));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&layer_rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&max_features));
size_t localThreads[3] = {3, 3, 3};
size_t globalThreads[3] = {maxCounter *localThreads[0], localThreads[1], 1};
size_t globalThreads[3] = {maxCounter*localThreads[0], localThreads[1], 3};
openCLExecuteKernelSURF(clCxt, &surfprog, kernelName, globalThreads, localThreads, args, -1, -1);
}
ocl::Kernel kerInterp("SURF_interpolateKeypoint", ocl::nonfree::surf_oclsrc, kerOpts);
void SURF_OCL_Invoker::icvCalcOrientation_gpu(const oclMat &keypoints, int nFeatures)
{
Context *clCxt = counters.clCxt;
String kernelName = "icvCalcOrientation";
std::vector< std::pair<size_t, const void *> > args;
if(sumTex)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&sumTex));
}
else
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&surf_.sum.data)); // if image2d is not supported
}
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&keypoints.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&img_rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&img_cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&surf_.sum.step));
size_t localThreads[3] = {ORI_LOCAL_SIZE, 1, 1};
size_t globalThreads[3] = {nFeatures * localThreads[0], 1, 1};
openCLExecuteKernelSURF(clCxt, &surfprog, kernelName, globalThreads, localThreads, args, -1, -1);
return kerInterp.args(ocl::KernelArg::ReadOnlyNoSize(det),
ocl::KernelArg::PtrReadOnly(maxPosBuffer),
ocl::KernelArg::ReadWriteNoSize(keypoints),
ocl::KernelArg::PtrReadWrite(counters),
img_rows, img_cols, octave, layer_rows, max_features).
run(3, globalThreads, localThreads, true);
}
void SURF_OCL_Invoker::icvSetUpright_gpu(const oclMat &keypoints, int nFeatures)
bool SURF_OCL::calcOrientation(UMat &keypoints)
{
Context *clCxt = counters.clCxt;
String kernelName = "icvSetUpright";
std::vector< std::pair<size_t, const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&keypoints.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&nFeatures));
size_t localThreads[3] = {256, 1, 1};
size_t globalThreads[3] = {saturate_cast<size_t>(nFeatures), 1, 1};
openCLExecuteKernelSURF(clCxt, &surfprog, 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;
std::vector< std::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();
if(imgTex)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&imgTex));
}
int nFeatures = keypoints.cols;
if( nFeatures == 0 )
return true;
ocl::Kernel kerOri("SURF_calcOrientation", ocl::nonfree::surf_oclsrc, kerOpts);
if( haveImageSupport )
kerOri.args(sumTex, img_rows, img_cols,
ocl::KernelArg::ReadWriteNoSize(keypoints));
else
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&_img.data));
}
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&descriptors.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&descriptors.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&keypoints.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&_img.rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&_img.cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&_img.step));
openCLExecuteKernelSURF(clCxt, &surfprog, 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( std::make_pair( sizeof(cl_mem), (void *)&descriptors.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&descriptors.step));
kerOri.args(ocl::KernelArg::ReadOnlyNoSize(sum),
img_rows, img_cols,
ocl::KernelArg::ReadWriteNoSize(keypoints));
openCLExecuteKernelSURF(clCxt, &surfprog, 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();
if(imgTex)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&imgTex));
}
else
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&_img.data));
}
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&descriptors.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&keypoints.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&descriptors.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&keypoints.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&_img.rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&_img.cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&_img.step));
openCLExecuteKernelSURF(clCxt, &surfprog, 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( std::make_pair( sizeof(cl_mem), (void *)&descriptors.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&descriptors.step));
openCLExecuteKernelSURF(clCxt, &surfprog, kernelName, globalThreads, localThreads, args, -1, -1);
}
size_t localThreads[3] = {ORI_LOCAL_SIZE, 1};
size_t globalThreads[3] = {nFeatures * localThreads[0], 1};
return kerOri.run(2, globalThreads, localThreads, true);
}
#endif //HAVE_OPENCV_OCL
}
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