I have a question about wrapping a Halide::Image object around an already allocated C++ array that defines a 2-D double precision floating point image.
I've already looked at:
C++ array to Halide Image (and back)
This is close to what I want to do, but I'm confused by uint8_t type of the host member of buffer_t in halide and how you work with existing images that are not uint8_t.
I see that in the blur app which uses aot, the example allocates a Halide Image, and then copies elements into this halide image. I would like to do this, but without paying for the copy.
It's not an option for me to use load_image, I need to work with existing, already allocated memory defined by a double *.
Image<uint16_t> input(6408, 4802);
for (int y = 0; y < input.height(); y++) {
for (int x = 0; x < input.width(); x++) {
input(x, y) = rand() & 0xfff;
}
}
The uint8_t* type of the host field of buffer_t is just a pointer to any array of data. You are free to point it to a float or double array. It's basically a void* pointer that gets reinterpreted by the actual pipeline code. Its interpretation is determined by the combination of the Halide program that loads from it, and the elem_size field (which should be, e.g., sizeof(double) or sizeof(float) for floating point data of different kinds).
Related
I need to transform the coordinates from spherical to Cartesian space using the Eigen C++ Library. The following code serves the purpose.
const int size = 1000;
Eigen::Array<std::pair<float, float>, Eigen::Dynamic, 1> direction(size);
for(int i=0; i<direction.size();i++)
{
direction(i).first = (i+10)%360; // some value for this example (denoting the azimuth angle)
direction(i).second = (i+20)%360; // some value for this example (denoting the elevation angle)
}
SSPL::MatrixX<T1> transformedMatrix(3, direction.size());
for(int i=0; i<transformedMatrix.cols(); i++)
{
const T1 azimuthAngle = direction(i).first*M_PI/180; //converting to radians
const T1 elevationAngle = direction(i).second*M_PI/180; //converting to radians
transformedMatrix(0,i) = std::cos(azimuthAngle)*std::cos(elevationAngle);
transformedMatrix(1,i) = std::sin(azimuthAngle)*std::cos(elevationAngle);
transformedMatrix(2,i) = std::sin(elevationAngle);
}
I would like to know a better implementation is possible to improve the speed.
I know that Eigen has supporting functions for Geometrical transformations. But I am yet to see a clear example to implement the same.
Is it also possible to vectorize the code to improve the performance?
You could at least use the vectorized versions of sine/cosine:
void dir2vector2(Eigen::Matrix3Xf& out, const Eigen::Array2Xf& in){
Eigen::Array2Xf sine = sin(in * (M_PI/180));
Eigen::Array2Xf cosi = cos(in * (M_PI/180));
out.resize(3, in.cols());
out << cosi.row(0) * cosi.row(1),
sine.row(0) * cosi.row(1),
sine.row(1);
}
There would still be a lot of optimization potential, e.g., calculating both sine and cosine of the same angle could share a lot of computation. And it is technically not necessary to store sine and cosi explicitly into temporaries (but Eigen is currently not able to automatically re-use common-sub expressions).
Also, the multiplication at the end could be vectorized better, if you store your input and output in row-major format (though the Eigen comma-initializer currently does not well with vectorization, it seems).
I want to convert dp to px in my C# code in xamarin.android, but all I could find were java codes in android studio that have some problems in xamarin. I tried to use equivalent like using Resources instead of getResources() and I could solve some little problems, but there are some problems yet that I couldn't find any equivalent for them. here are original codes, my codes, and my codes problems in xamarin:
First code
(found from Programatically set height on LayoutParams as density-independent pixels)
java code
int height = (int)TypedValue.applyDimension(TypedValue.COMPLEX_UNIT_DIP, < HEIGHT >, getResources().getDisplayMetrics());
C# code
int height = (int)TypedValue.ApplyDimension(TypedValue.COMPLEX_UNIT_DIP, < HEIGHT >, Resources.DisplayMetrics);
problems:
'TypedValue' does not contain a definition for 'COMPLEX_UNIT_DIP'
Invalid expression term < (The same error for >)
The name 'HEIGHT' does not exist in the current context
Second code
(found from Formula px to dp, dp to px android)
java code
DisplayMetrics displayMetrics = getContext().getResources().getDisplayMetrics();
int px = Math.round(dp * (displayMetrics.xdpi / DisplayMetrics.DENSITY_DEFAULT));
C# code
DisplayMetrics displayMetrics = Application.Context.Resources.DisplayMetrics;
int pixel = Math.Round(dp * (displayMetrics.Xdpi / DisplayMetrics.DensityDefault));
problem
Operator '/' cannot be applied to operands of type 'float' and 'DisplayMetricsDensity'
Now I have actually two questions. Which code is more proper? What's equivalent code for them in xamarin.android?
Thanks in advance.
Solution to "First code":
Xamarin tends to move constants into their own enums. COMPLEX_UNIT_DIP can be found on ComplexUnitType enum. Also you cannot have < HEIGHT > in your code you actually need to pass in dips to get the equivalent pixel value. in the example below I am getting pixels in 100 dips.
var dp = 100;
int pixel = (int)TypedValue.ApplyDimension(ComplexUnitType.Dip, dp, Context.Resources.DisplayMetrics);
Solution to "Second code":
You need to explicitly cast 'DisplayMetrics.DensityDefault' to a float and the entire round to an int:
int pixel = (int)System.Math.Round(dp * (displayMetrics.Xdpi / (float)DisplayMetrics.DensityDefault));
I prefer the first approach as the second code is specifically for calculating along the "x dimension":
Per Android docs and Xamarin.Android docs, the Xdpi property is
"The exact physical pixels per inch of the screen in the X dimension."
These are from the project that I am currently working on..
public static float pxFromDp(Context context, float dp)
{
return dp * context.Resources.DisplayMetrics.Density;
}
public static float dpFromPx(Context context, float px)
{
return px / context.Resources.DisplayMetrics.Density;
}
From android source code textview settextsize
var convertToDp = TypedValue.ApplyDimension( ComplexUnitType.Dip, size, context.Resouces.DisplayMetrics);
dp to px:
DisplayMetrics displayMetrics = Application.Context.Resources.DisplayMetrics;
double pixel = Math.Round((dp * DisplayMetrics.DensityDefault) + 0.5);
Answer taken from here : https://stackoverflow.com/a/8490361/6949388
Sorry, not enough rep to comment.
I am trying to access the image colors in a QImage.
The method that I found most in docs is based on the scanline function...
I tried and it worked... on RGB32 images. I had surprising - and unpleasant results when using the exact method to get color data for 8 bit indexed or monochrome images.
This was my code:
// note RGBTriple is a struct containing unsigned R, G, B
// rgbImage.pixels is a RGBTriple* array
RGBTriple* pTriple = rgbImage.pixels;
for (int y = 0; y < source.height(); y++)
{
const unsigned char* pScanLine = source.scanLine(y);
for (int x = 0; x < source.width(); x++)
{
QRgb* color = (QRgb*)pScanLine;
pTriple->R = qRed(*color);
pTriple->G = qGreen(*color);
pTriple->B = qBlue(*color);
++pTriple;
pScanLine += 4;
}
}
Running the same code with images 8bit indexed or monochrome, I got errors in creating getting colors. The documentation says that scanline is aligned to multiples of 32b - but since that is a multiple of 8 and 2 I didn't think it would be a problem.
Once I found out that I am not getting correct results for all types of input images, I changed it to
RGBTriple* pTriple = rgbImage.pixels;
for (int y = 0; y < source.height(); y++)
{
for (int x = 0; x < source.width(); x++)
{
pTriple->R = qRed(source.pixel(x, y));
pTriple->G = qGreen(source.pixel(x, y));
pTriple->B = qBlue(source.pixel(x, y));
++pTriple;
}
}
Works perfectly... I wonder if it is slower or will have other unexpected behavior ? After all, I am using the pixel() function - even on indexed images - to get color information, which actually should be stored differently... that seems like it should fail...
Is there a way to make the first version, using scanline, work for other image types ?
Why does it seem like using scanline to get the data is the preferred method ?
I tried and it worked... on RGB32 images. I had surprising - and
unpleasant results when using the exact method to get color data for 8
bit indexed or monochrome images.
You should not be surprised because the indexed and monochrome images are different formats. The first code snippet you posted is based on the knowledge on how RGB32 (and RGB32 only) is layed out in memory.
Think about it. In a monochrome image R=G=B. So only one channel need to be saved in memory.
If your goal is to obtain an rgb image inside rgbImage.pixels use QImage::convertToFormat() :
QImage source;
QImage dest = source.convertToFormat( QImage::Format_RGB888 );
memcpy(rgbImage.pixels, dest.bits(),dest.byteCount () );
I am doing several steps of reprojections of a point cloud (around 40 Million points initially, ~20 Million while processing). The Programm crashes at seemingly random points at one of these 2 loops. If I run it with a smaller subset (~10 Million Points) everything works fine.
//Projection of Point Cloud into a sphere
pcl::PointCloud<pcl::PointXYZ>::Ptr projSphere(pcl::PointCloud<pcl::PointXYZ>::Ptr cloud,int radius)
{
//output cloud
pcl::PointCloud<pcl::PointXYZ>::Ptr output(new pcl::PointCloud<pcl::PointXYZ>);
//time marker
int startTime = time(NULL);
cout<<"Start Sphere Projection"<<endl;
//factor by which each Point Vector ist multiplied to get a distance of radius to the origin
float scalar;
for (int i=0;i<cloud->size();i++)
{
if (i%1000000==0) cout<<i<<endl;
//P
pcl::PointXYZ tmpin=cloud->points.at(i);
//P'
pcl::PointXYZ tmpout;
scalar=radius/(sqrt(pow(tmpin.x,2)+pow(tmpin.y,2)+pow(tmpin.z,2)));
tmpout.x=tmpin.x*scalar;
tmpout.y=tmpin.y*scalar;
tmpout.z=tmpin.z*scalar;
//Adding P' to the output cloud
output->push_back(tmpout);
}
cout<<"Finished projection of "<<output->size()<<" points in "<<time(NULL)-startTime<<" seconds"<<endl;
return(output);
}
//Stereographic Projection
pcl::PointCloud<pcl::PointXYZ>::Ptr projStereo(pcl::PointCloud<pcl::PointXYZ>::Ptr cloud)
{
//output cloud
pcl::PointCloud<pcl::PointXYZ>::Ptr outputSt(new pcl::PointCloud<pcl::PointXYZ>);
//time marker
int startTime = time(NULL);
cout<<"Start Stereographic Projection"<<endl;
for (int i=0;i<cloud->size();i++)
{
//P
if (i%1000000==0) cout<<i<<endl;
pcl::PointXYZ tmpin=cloud->points.at(i);
//P'
pcl::PointXYZ tmpout;
//equation
tmpout.x=tmpin.x/(1.0+tmpin.z);
tmpout.y=tmpin.y/(1.0+tmpin.z);
tmpout.z=0;
//Adding P' to the output cloud
outputSt->push_back(tmpout);
}
cout<<"Finished projection of"<<outputSt->size()<<" points in "<<time(NULL)-startTime<<" seconds"<<endl;
return(outputSt);
}
If I do all the steps independently by saving/loading the pointclouds on the harddisk and rerunning the program for each step it also works fine. I'd like to provie the entire source files but I'm not sure how/if it's neccessary.
Thanks in advance
Edit:1
After about a week I have still no idea what might be the issue here, since the crashes are somewhat random, but not really? I tried to test the programm with a different system workload (freshly rebooted, with heavy duty programs loaded etc.) makes no apparent difference. Since I thought it's maybe a memory issue, I tried o move the large objects from stack to heap (initialising them with new), did also make no difference. By far the largest object is the raw input file, which I open and close by:
ifstream file;
file.open(infile);
/*......*/
file.close();
delete file;
Is that properly done, so that after the method is completed the memory is released?
Edit again:
So I try further and further, and finally I managed to put all the steps into one function like this:
void stereoTiffI(string infile, string outfile, int length)
{
//set up file input
cout<<"Opening file: "<< infile<<endl;
ifstream file;
file.open(infile);
string line;
//skip first lines
for (int i=0;i<9;i++)
{
getline(file,line);
}
//output cloud
pcl::PointCloud<pcl::PointXYZ> cloud;
getline(file,line);
//indexes for string parsing, coordinates and starting Timer
int i=0;
int j=0;
int k=0;
float x=0;
float y=0;
float z=0;
float intensity=0;
float scalar=0;
int startTime = time(NULL);
pcl::PointXYZ tmp;
//begin loop
cout<<"Begin reading and projecting"<< infile<<endl;
while (!file.eof())
{
getline(file,line);
i=0;
j=line.find(" ");
x=atof(line.substr(i,j).c_str());
i=line.find(" ",i)+1;
j=line.find(" ",i)-i;
y=atof(line.substr(i,j).c_str());
i=line.find(" ",i)+1;
j=line.find(" ",i)-i;
z=atof(line.substr(i,j).c_str());
//i=line.find(" ",i)+1;
//j=line.find(" ",i)-i;
//intensity=atof(line.substr(i,j).c_str());
//leave out points below scanner height
if (z>0)
{
//projection to a hemisphere with radius 1
scalar=1/(sqrt(pow(x,2)+pow(y,2)+pow(z,2)));
x=x*scalar;
y=y*scalar;
z=z*scalar;
//stereographic projection
x=x/(1.0+z);
y=y/(1.0+z);
z=0;
tmp.x=x;
tmp.y=y;
tmp.z=z;
//tmp.intensity=intensity;
cloud.push_back(tmp);
k++;
if (k%1000000==0)cout<<k<<endl;
}
}
cout<<"Finished producing projected cloud in: "<<time(NULL)-startTime<<" with "<<cloud.size()<<" points."<<endl;
And this actually works quit nicely and quickly. In a next step I tried to use Pointtype XYZI because I need to also get the intensity of the scanned points. And guess what, the program crashes at around 17000000 again, and again I have no idea why. Please help
Ok, I solved it. Dr. Memory gave me the right hint by giving me a heap allocation error. After a bit of googling I enabled Large Addresses in Visual Studio (Properties -> Linker -> System)
Everything works like a charm.
I am pretty new to Objective C and working with Cocoa Framework. I want to read an image and then extract the image data (just pixel data and not the header) and then write the data to a binary file. I am kind of stuck with this, I was going through the methods of NSImage but I couldn't find a suitable one. Can anyone suggest me some other ways of doing this?
Cocoa-wise, the easiest approach is to use the NSBitmapImageRep class. Once initialized with a NSData object, for example, you can access the color value at any coordinate as a NSColor object using the -setColor:atX:y: and -colorAtX:y: methods. Note that if you call these methods in tight loops, you may suffer a performance hit from objc_msg_send. You could consider accessing the raw bitmap data as C array via the -bitmapData method. When dealing with a RGB image, for example, the color values for each channel are stored at offsets of 3.
For example:
color values: [R,G,B][R,G,B][R,G,B]
indices: [0,1,2, 3,4,5, 6,7,8]
To loop through each pixel in the image and extract the RGB components:
unsigned char *bitmapData = [bitmapRep bitmapData];
if ([bitmapRep samplesPerPixel] == 3) {
for (i = 0; i < [image size].width * [image size].height; i++) {
int base = (i * 3);
// these range from 0-255
unsigned char red = bitmapData[base + 0];
unsigned char green = bitmapData[base + 1];
unsigned char blue = bitmapData[base + 2];
}
}