提交 66a8acfd 编写于 作者: A Alexander Karsakov

Optimization for HoughLinesP

上级 eaf5a163
......@@ -117,7 +117,7 @@ namespace cv { namespace cuda { namespace device
if (dir.x < 0)
dir = -dir;
}
else if (pb[1].x == cols - 1 && (pb[0].y >= 0 && pb[0].y < rows))
else if (pb[1].x == cols - 1 && (pb[1].y >= 0 && pb[1].y < rows))
{
p0 = pb[1];
if (dir.x > 0)
......
......@@ -221,7 +221,7 @@ HoughLinesSDiv( const Mat& img,
std::vector<hough_index> lst;
CV_Assert( img.type() == CV_8UC1 );
CV_Assert( linesMax > 0 && rho > 0 && theta > 0 );
CV_Assert( linesMax > 0 );
threshold = MIN( threshold, 255 );
......@@ -655,6 +655,8 @@ HoughLinesProbabilistic( Mat& image,
#ifdef HAVE_OPENCL
#define OCL_MAX_LINES 4096
static bool ocl_makePointsList(InputArray _src, OutputArray _pointsList, InputOutputArray _counters)
{
UMat src = _src.getUMat();
......@@ -702,7 +704,7 @@ static bool ocl_fillAccum(InputArray _pointsList, OutputArray _accum, int total_
if (fillAccumKernel.empty())
return false;
globalThreads[0] = workgroup_size; globalThreads[1] = numangle;
fillAccumKernel.args(ocl::KernelArg::ReadOnlyNoSize(pointsList), ocl::KernelArg::WriteOnly(accum),
fillAccumKernel.args(ocl::KernelArg::ReadOnlyNoSize(pointsList), ocl::KernelArg::WriteOnlyNoSize(accum),
total_points, irho, (float) theta, numrho, numangle);
return fillAccumKernel.run(2, globalThreads, NULL, false);
}
......@@ -714,7 +716,7 @@ static bool ocl_fillAccum(InputArray _pointsList, OutputArray _accum, int total_
return false;
localThreads[0] = workgroup_size; localThreads[1] = 1;
globalThreads[0] = workgroup_size; globalThreads[1] = numangle+2;
fillAccumKernel.args(ocl::KernelArg::ReadOnlyNoSize(pointsList), ocl::KernelArg::WriteOnly(accum),
fillAccumKernel.args(ocl::KernelArg::ReadOnlyNoSize(pointsList), ocl::KernelArg::WriteOnlyNoSize(accum),
total_points, irho, (float) theta, numrho, numangle);
return fillAccumKernel.run(2, globalThreads, localThreads, false);
}
......@@ -731,6 +733,9 @@ static bool ocl_HoughLines(InputArray _src, OutputArray _lines, double rho, doub
if (min_theta < 0 || min_theta > max_theta ) {
CV_Error( CV_StsBadArg, "min_theta must fall between 0 and max_theta" );
}
if (!(rho > 0 && theta > 0)) {
CV_Error( CV_StsBadArg, "rho and theta must be greater 0" );
}
UMat src = _src.getUMat();
int numangle = cvRound((max_theta - min_theta) / theta);
......@@ -759,8 +764,7 @@ static bool ocl_HoughLines(InputArray _src, OutputArray _lines, double rho, doub
if (getLinesKernel.empty())
return false;
// TODO: investigate other strategies to choose linesMax
int linesMax = min(total_points*numangle/threshold, 4096);
int linesMax = threshold > 0 ? min(total_points*numangle/threshold, OCL_MAX_LINES) : OCL_MAX_LINES;
UMat lines(linesMax, 1, CV_32FC2);
getLinesKernel.args(ocl::KernelArg::ReadOnly(accum), ocl::KernelArg::WriteOnlyNoSize(lines),
......@@ -783,6 +787,10 @@ static bool ocl_HoughLinesP(InputArray _src, OutputArray _lines, double rho, dou
{
CV_Assert(_src.type() == CV_8UC1);
if (!(rho > 0 && theta > 0)) {
CV_Error( CV_StsBadArg, "rho and theta must be greater 0" );
}
UMat src = _src.getUMat();
int numangle = cvRound(CV_PI / theta);
int numrho = cvRound(((src.cols + src.rows) * 2 + 1) / rho);
......@@ -809,8 +817,7 @@ static bool ocl_HoughLinesP(InputArray _src, OutputArray _lines, double rho, dou
if (getLinesKernel.empty())
return false;
// TODO: investigate other strategies to choose linesMax
int linesMax = min(total_points*numangle/threshold, 4096);
int linesMax = threshold > 0 ? min(total_points*numangle/threshold, OCL_MAX_LINES) : OCL_MAX_LINES;
UMat lines(linesMax, 1, CV_32SC4);
getLinesKernel.args(ocl::KernelArg::ReadOnly(accum), ocl::KernelArg::ReadOnly(src),
......
......@@ -5,6 +5,8 @@
// Copyright (C) 2014, Itseez, Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
#define ACCUM(ptr) *((__global int*)(ptr))
#ifdef MAKE_POINTS_LIST
__kernel void make_point_list(__global const uchar * src_ptr, int src_step, int src_offset, int src_rows, int src_cols,
......@@ -25,11 +27,13 @@ __kernel void make_point_list(__global const uchar * src_ptr, int src_step, int
if (y < src_rows)
{
y <<= 16;
for (int i=x; i < src_cols; i+=GROUP_SIZE)
{
if (src[i])
{
int val = (y << 16) | i;
int val = y | i;
int index = atomic_inc(&l_index);
l_points[index] = val;
}
......@@ -53,7 +57,7 @@ __kernel void make_point_list(__global const uchar * src_ptr, int src_step, int
#elif defined FILL_ACCUM_GLOBAL
__kernel void fill_accum_global(__global const uchar * list_ptr, int list_step, int list_offset,
__global uchar * accum_ptr, int accum_step, int accum_offset, int accum_rows, int accum_cols,
__global uchar * accum_ptr, int accum_step, int accum_offset,
int total_points, float irho, float theta, int numrho, int numangle)
{
int theta_idx = get_global_id(1);
......@@ -76,7 +80,7 @@ __kernel void fill_accum_global(__global const uchar * list_ptr, int list_step,
const int x = (val & 0xFFFF);
const int y = (val >> 16) & 0xFFFF;
int r = convert_int_rte(x * cosVal + y * sinVal) + shift;
int r = convert_int_rte(mad(x, cosVal, y * sinVal)) + shift;
atomic_inc(accum + r + 1);
}
}
......@@ -85,7 +89,7 @@ __kernel void fill_accum_global(__global const uchar * list_ptr, int list_step,
#elif defined FILL_ACCUM_LOCAL
__kernel void fill_accum_local(__global const uchar * list_ptr, int list_step, int list_offset,
__global uchar * accum_ptr, int accum_step, int accum_offset, int accum_rows, int accum_cols,
__global uchar * accum_ptr, int accum_step, int accum_offset,
int total_points, float irho, float theta, int numrho, int numangle)
{
int theta_idx = get_group_id(1);
......@@ -133,15 +137,13 @@ __kernel void fill_accum_local(__global const uchar * list_ptr, int list_step, i
#elif defined GET_LINES
#define ACCUM(ptr) *((__global int*)(ptr))
__kernel void get_lines(__global uchar * accum_ptr, int accum_step, int accum_offset, int accum_rows, int accum_cols,
__global uchar * lines_ptr, int lines_step, int lines_offset, __global int* lines_index_ptr,
int linesMax, int threshold, float rho, float theta)
{
int x0 = get_global_id(0);
int y = get_global_id(1);
int gl_size = get_global_size(0);
int glob_size = get_global_size(0);
if (y < accum_rows-2)
{
......@@ -149,7 +151,7 @@ __kernel void get_lines(__global uchar * accum_ptr, int accum_step, int accum_of
__global float2* lines = (__global float2*)(lines_ptr + lines_offset);
__global int* lines_index = lines_index_ptr + 1;
for (int x=x0; x<accum_cols-2; x+=gl_size)
for (int x=x0; x<accum_cols-2; x+=glob_size)
{
int curVote = ACCUM(accum);
......@@ -167,15 +169,13 @@ __kernel void get_lines(__global uchar * accum_ptr, int accum_step, int accum_of
}
}
accum += gl_size * (int) sizeof(int);
accum += glob_size * (int) sizeof(int);
}
}
}
#elif GET_LINES_PROBABOLISTIC
#define ACCUM(ptr) *((__global int*)(ptr))
__kernel void get_lines(__global const uchar * accum_ptr, int accum_step, int accum_offset, int accum_rows, int accum_cols,
__global const uchar * src_ptr, int src_step, int src_offset, int src_rows, int src_cols,
__global uchar * lines_ptr, int lines_step, int lines_offset, __global int* lines_index_ptr,
......@@ -222,6 +222,7 @@ __kernel void get_lines(__global const uchar * accum_ptr, int accum_step, int ac
pb[1].x = src_cols - 1;
pb[1].y = p0.y + a * dir.y;
}
if (dir.y != 0)
{
a = -p0.y / dir.y;
......@@ -239,7 +240,7 @@ __kernel void get_lines(__global const uchar * accum_ptr, int accum_step, int ac
if (dir.x < 0)
dir = -dir;
}
else if (pb[1].x == src_cols - 1 && (pb[0].y >= 0 && pb[0].y < src_rows))
else if (pb[1].x == src_cols - 1 && (pb[1].y >= 0 && pb[1].y < src_rows))
{
p0 = pb[1];
if (dir.x > 0)
......@@ -258,41 +259,30 @@ __kernel void get_lines(__global const uchar * accum_ptr, int accum_step, int ac
dir = -dir;
}
float2 d;
if (fabs(dir.x) > fabs(dir.y))
{
d.x = dir.x > 0 ? 1 : -1;
d.y = dir.y / fabs(dir.x);
}
else
{
d.x = dir.x / fabs(dir.y);
d.y = dir.y > 0 ? 1 : -1;
}
dir /= max(fabs(dir.x), fabs(dir.y));
float2 line_end[2];
int gap;
bool inLine = false;
float2 p1 = p0;
if (p1.x < 0 || p1.x >= src_cols || p1.y < 0 || p1.y >= src_rows)
if (p0.x < 0 || p0.x >= src_cols || p0.y < 0 || p0.y >= src_rows)
return;
for (;;)
{
if (*(src_ptr + mad24(p1.y, src_step, p1.x + src_offset)))
if (*(src_ptr + mad24(p0.y, src_step, p0.x + src_offset)))
{
gap = 0;
if (!inLine)
{
line_end[0] = p1;
line_end[1] = p1;
line_end[0] = p0;
line_end[1] = p0;
inLine = true;
}
else
{
line_end[1] = p1;
line_end[1] = p0;
}
}
else if (inLine)
......@@ -314,8 +304,8 @@ __kernel void get_lines(__global const uchar * accum_ptr, int accum_step, int ac
}
}
p1 = p1 + d;
if (p1.x < 0 || p1.x >= src_cols || p1.y < 0 || p1.y >= src_rows)
p0 = p0 + dir;
if (p0.x < 0 || p0.x >= src_cols || p0.y < 0 || p0.y >= src_rows)
{
if (inLine)
{
......
......@@ -168,7 +168,7 @@ OCL_TEST_P(HoughLinesP, RealImage)
OCL_OFF(cv::HoughLinesP(src, dst, rhoStep, thetaStep, threshold, minLineLength, maxGap));
OCL_ON(cv::HoughLinesP(usrc, udst, rhoStep, thetaStep, threshold, minLineLength, maxGap));
Near(0.2);
Near(0.25);
}
OCL_INSTANTIATE_TEST_CASE_P(Imgproc, HoughLines, Combine(Values(1, 0.5), // rhoStep
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册