Accessing GpuMat on another GPU

2018-04-17 05:53 Ken Y-N imported from Stackoverflow

I'm pretty sure this is extremely dodgy code, but I just want to make sure. Assume I have two GPUs, and let us skip error checking:

cudaSetDevice(0);
cv::cuda::GpuMat foo(5, 5, CV_8UC1);
foo.at<uchar>(2, 2) = 128;
cudaSetDevice(1);
foo.at<uchar>(4, 4) = 192;

Is GpuMat or the CUDA runtime smart enough to know that foo data pointer points to GPU 0, or am I (as I suspect) accessing potentially unowned memory on GPU 1 with the foo.at<uchar>(4, 4) = 192; line? I'm using OpenCV 3.2 and CUDA 8.0 on Ubuntu 16.04.

See also questions close to this topic

Qt: How should I route a signal to the appropriate object dynamically?
I am writing an application in Qt. I am trying to create a system of signals and slots to connect a single controller thread to multiple threads which represent real-life devices. The number of devices is known at compile-time, but I would like it to be as easy as possible to change the number of them. My initial approach used templates on the signals, something like this:
```
signals:
    template<int whichOne> updateDoohickeyState(dooHickeyState newState);
```
My hope was that I could then connect these to the devices like so:
```
connect(doohickeyController, doohickeyController::updateDoohickeyState<0>,
        doohickeys[0], doohickeyObject::updateState,
        Qt::QueuedConnection);

connect(doohickeyController, doohickeyController::updateDoohickeyState<1>,
        doohickeys[1], doohickeyObject::updateState,
        Qt::QueuedConnection);
// Etc...
```
Then, when I wanted to signal a device, I could do something like:
```
emit updateDoohickeyState<0>(doohickeyState);
emit updateDoohickeyState<1>(anotherDoohickeyState);
```
Unfortunately, Qt's MOC does not support templates on signals and slots; therefore, this does not work.

So, how else can I implement this? I have one controller which needs to signal to particular one of multiple identical devices in other threads. I would ideally like to do so without filtering slot-side (inefficient) or adding another class in the middle.

Why does while loop skips a for loop?

I am having some mystery problem. I have written a path planning code and for some reason, in my while loop, the for loop is not being executed at all!

Here is the while loop code:

while (!dijkstra.empty()) {
        node* n0 = new node(dijkstra.begin()->getxPos(), dijkstra.begin()->getyPos(), dijkstra.begin()->getDir(), dijkstra.begin()->getTurns(), dijkstra.begin()->getDist());
        dijkstra.erase(dijkstra.begin());

        int x = n0->getxPos();
        int y = n0->getyPos();
        int dir = n0->getDir();
        int turns = n0->getTurns();
        int dist = n0->getDist();
        ROS_INFO("X : %d, Y: %d", x, y);


        if (x == x_goal && y == y_goal) {
            geometry_msgs::PoseStamped temp;
            std::vector<geometry_msgs::PoseStamped> temp_list;
            nav_msgs::Path path_to_pub;
            ROS_INFO("HEllo, %s", "loop");
            double x_world, y_world;
            for (;;) {
                if (x == x_start && y == y_start) break;
                int new_x = prev_state[x][y][dir].x;
                int new_y = prev_state[x][y][dir].y;
                int new_dir = prev_state[x][y][dir].dir;
                x = new_x;
                y = new_y;
                dir = new_dir;
                temp.header.stamp = ros::Time::now();
                temp.header.frame_id = "map";
                x_world = mapToWorld(x, true);
                y_world = mapToWorld(y, false);
                temp.pose.position.x = x_world;
                temp.pose.position.y = y_world;
                temp_list.push_back(temp);
            }

            path_to_pub.poses.resize(temp_list.size());
            reverse(temp_list.begin(), temp_list.end());
            if (!temp_list.empty()) {
                path_to_pub.header.frame_id = "map";
                path_to_pub.header.stamp = temp_list[0].header.stamp;
            }

            for (unsigned int i = 0; i < temp_list.size(); i++) {
                path_to_pub.poses[i] = temp_list[i];
            }

            plan = temp_list;
            path_pub_.publish(path_to_pub);
            return true;
        }

        for (int i = 0; i < dir; i++) {
            ROS_INFO("FOOR %s", "LOOP");
            int to_x = x + dx[i];
            int to_y = y + dy[i];
            ROS_INFO("To X: %d, to Y: %d", to_x, to_y);

            if (!isValid(mapToWorld(to_x, 1), mapToWorld(to_y, 0))) {
                continue;
            }

            if (to_x < 0 || to_x >= n - 1 || to_y < 0 || to_y >= m - 1) {
                continue;
            }

            std::pair<int, int> to_cost = std::make_pair(turns + (dir != i), dist + 1);

            if (to_cost < best_cost[to_x][to_y][i]) {
                prev_state[to_x][to_y][i].dir = (i + dir / 2) % dir;
                best_cost[to_x][to_y][i] = to_cost;
                node* to_erase = new node(to_x, to_y, i, best_cost[to_x][to_y][i].first, best_cost[to_x][to_y][i].second);
                dijkstra.erase(*to_erase);
                node* m0 = new node(to_x, to_y, i, to_cost.first, to_cost.second);
                dijkstra.insert(*m0);
            }

        }

    }

    ROS_INFO("Returning %s", "False");
    return false;

}

As you can see, I have put some debugging messages along the lines (ROS_INFO) to see where the problem might exist. It starts okay and I can see the correct values at the very beginning (where ROS_INFO("X : , Y: )). But right after that, the while loop just skips the for loop right at the bottom. Other debugging messages are never executed other than "Returning False" at the very bottom. The size of the dijkstra is 0 but shouldn't the code run at least 1 for loop?

Can someone see why does for loop never execute in my code?

Limited visibility window on a dynamically update graph

I have this non trivial challenge (at least for me) to build an API to work with a data set that are sort of a graph which needs to be dynamically updated (changes in amount of vertices their properties and connections between them).

Data set has geospatial data. Connections between vertices are handled via object ids and their relations. So, it's a PostGIS database in a C++ interpretation.

An acceptable limitation is that we allow updates only for data that are "far" from currently observed location. This assumes that API limits visibility of a graph to a certain area. This limited visibility window is moving along the known trajectory.

Now, the question is how an API and internal implementation should look like in order to provide access to data (vertices, their properties and connections) in the way that we don't have to keep the trace of once initialized (and exposed via API) data for as long as program runs. Basically, a user should be able to work with data exposed via current limited visibility window and should not have access to data that have been accessible via old window but not in the current one.

Simple example of an API that would fail: expose direct access to data in graph. A user can save a raw pointer to data. Once window is moved, that pointer might become invalid but user would never know about it. Some of user pointers will still point of existing data (available via current window) when other will point to un-managed memory (cause old data has been release).

Things like shared pointers are not applicable because graph has circular connections.
Slice an opencv image based on list of list

I have an OpenCV image and I need to slice it based on indices of list of list. Sample data of list of list

So that 2D list means take pixel at 162,162, slice until the two x,y are equal. make an image out of that slice..

I need the result images of slicing that 2D List.

[[162, 162], [162, 163], [162, 164], [162, 165], [169, 169], [169, 170], [169, 171], [169, 172], [197, 197], [197, 198], [197, 199], [197, 200], [214, 214], [214, 215], [214, 216], [214, 217], [214, 218], [214, 219], [214, 220], [228, 228], [228, 229], [228, 230], [228, 231], [228, 232], [228, 233], [228, 234], [228, 235], [243, 243], [243, 244], [243, 245], [243, 246], [256, 256], [256, 257], [256, 258], [256, 259], [256, 260], [270, 270], [270, 271]]

OpenCV distance calculate

I have created a C++ file in which I can track an object and find the area.

Now when it takes the area i want to print the distance. So I tried to mark the area of a known object, respectively, and then measure the distance with the tape measure from the camera.

My problem is that now when I say to print the distance it does not print anything.

Here is the piece of code:

void trackFilteredObject(int &x, int &y, Mat threshold, Mat &cameraFeed){

    Mat temp;
    threshold.copyTo(temp);
    //these two vectors needed for output of findContours
    std::vector< std::vector<Point> > contours;
    std::vector<Vec4i> hierarchy;
    //find contours of filtered image using openCV findContours function
    findContours(temp,contours,hierarchy,CV_RETR_CCOMP,CV_CHAIN_APPROX_SIMPLE );
    //use moments method to find our filtered object
    double refArea = 0;
    bool objectFound = false;
    if (hierarchy.size() > 0) {
        int numObjects = hierarchy.size();
        //if number of objects greater than MAX_NUM_OBJECTS we have a noisy filter
        if(numObjects<MAX_NUM_OBJECTS){
            for (int index = 0; index >= 0; index = hierarchy[index][0]) {

                Moments moment = moments((cv::Mat)contours[index]);
                double area = moment.m00;

                //if the area is less than 20 px by 20px then it is probably just noise
                //if the area is the same as the 3/2 of the image size, probably just a bad filter
                //we only want the object with the largest area so we safe a reference area each
                //iteration and compare it to the area in the next iteration.
                if(area>MIN_OBJECT_AREA && area<MAX_OBJECT_AREA && area>refArea){
                    x = moment.m10/area;
                    y = moment.m01/area;
                    xCoord=x;
                    yCoord=y;
                    objectFound = true;
                    refArea = area;
                    std::cout<<area<<std::endl;

                    //DISTANCE HERE
                    double dist;
                    if (refArea==0) {
                    cout << "\nArea is egual to 0" << endl;}
                    else if (refArea>0 && refArea<=295){
                      dist = 50;
                      cout << "\nThe distance of the object is " << dist << " cm\n" <<endl;
                    }
                    else if (refArea>295 && refArea<=320) {
                    dist = 65;
                      cout << "\nThe distance of the object is " << dist << " cm\n" <<endl;
                    }
                    else if (refArea>320 && refArea<=335){
                    dist = 95;
                    cout << "\nThe distance of the object is " << dist << " cm\n" <<endl;
                    }
                    else
                    cout << "\nDistance no determinate" <<endl;
                        // end of distance 
                }else objectFound = false;


            }
            //let user know you found an object
            if(objectFound ==true){
                putText(cameraFeed,"Tracking Object",Point(0,50),2,1,Scalar(0,255,0),2);
                //draw object location on screen
                drawObject(x,y,cameraFeed);}

        }else putText(cameraFeed,"TOO MUCH NOISE! ADJUST FILTER",Point(0,50),1,2,Scalar(0,0,255),2);
    }
}

LINK FOR THE CODE

IndexError: tuple index out of range - calculating mean RGB

I am attempting to calculate the mean RGB values for the an image using python. I found an acceptable method using numpy here.

How to find the average colour of an image in Python with OpenCV?

I am able to calculate the mean RGB values using the example image in the above answer, but I am not able to calculate the mean RGB values of my image (bellow).

import os
import cv2
import numpy as np


path = ('C:/images')

img = cv2.imread(path + '/1049.jpg', 0)
img = np.array(img)
average = img.mean(axis=0).mean(axis =0)
print(average)

I get the following error.

    Traceback (most recent call last):
  File "c:/Users/isaac_madsen/Google Drive/Rhizoc_2018/image_stats.py", line 21, in <module>
    average = img.mean(axis=0).mean(axis =0)
  File "C:\Users\isaac_madsen\AppData\Local\Continuum\anaconda3\lib\site-packages\numpy\core\_methods.py", line 57, in _mean
    rcount = _count_reduce_items(arr, axis)
  File "C:\Users\isaac_madsen\AppData\Local\Continuum\anaconda3\lib\site-packages\numpy\core\_methods.py", line 50, in _count_reduce_items
    items *= arr.shape[ax]
IndexError: tuple index out of range

I found this answer to a similar question on asymmetric arrays, but I am unsure how to implement the solution in my particular case, or if I am actually dealing with an asymmetric array.

Means of asymmetric arrays in numpy

Warp shuffle - conflicting documentation description
How does __shfl_sync() (and similar) functions behave when exceeding maximum threadIdx? Or attempting to read threads that have exited? In the documentation I can read two conflicting sentences that leave me puzzled. The following statements are taken from the CUDA C Programming Guide

Statement 1:

If the source lane ID is out of range or the source thread has exited, the calling thread's own var is returned.

Statement 2:

Threads may only read data from another thread which is actively participating in the __shfl_sync() command. If the target thread is inactive, the retrieved value is undefined.

But the thread that has existed, must have diverged and is not actively participating in the __shfl_sync().

For example, consider:
```
__global__ void kernel() {
    int val = threadIdx.x;
    if (threadIdx.x>5)
        return;
    __syncthreads();
    val = __shfl_down_sync(0x7fffffff, val, +5);
    printf("thread[%d].val = %d\n", threadIdx.x, val);
}
.....
kernel<<<1,32>>>()
```
Is the resulting val unchanged or undefined?

When I actually tested, I got the output:
```
thread[0].val = 5
thread[1].val = 0
thread[2].val = 0
thread[3].val = 0
thread[4].val = 0
thread[5].val = 0
```
which would suggest that the value is in fact undefined. But then what it means that a thread has exited?

I get the same result when I remove the conditional return, but invoke the kernel with just 6 threads. So ID that are too high still give the undefined value?

So is the Statement 1 completely wrong? Or I misunderstand it somehow?
CUDA GPU capability issue

I am trying to use numba vectorization in anaconda but I'm getting strange error. I installed CUDA 8.0 which supports my GPU (GTX 460) but when I am trying to execute simple code I am getting error:

NvvmSupportError: GPU compute capability 2.1 is not supported (requires >=2.0)

Since 2.1 >= 2.0 I do not know what is wrong.
vs2015 cuda9.0 linked SHA1_Init with CUDA implement instead of openssl cpu libs
I am a beginner to cuda, c++ and I am trying to move openssl sha1 cpu code to cuda c,but I ran into a weired problem.

here is the minimum code that can reproduce the problem. There are three files in this vs2015 cuda9.0 project. They are main.cpp ,sha1.cu and sha1.h
```
//main.cpp
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include "openssl\sha.h"
int main()
{
    SHA_CTX ctx;
    SHA1_Init(&ctx);
    return 0;
}

//sha1.h
#ifndef SHA1_H
#define SHA1_H
#include <stdint.h>
#include <cuda_runtime.h>
namespace cudatest {
#ifdef __cplusplus
extern "C" {
#endif
        typedef struct
        {
            uint32_t state[5];
            uint32_t count[2];
            unsigned char buffer[64];
        } SHA1_CTX;
#define SHA_CTX SHA1_CTX
#define SHA_DIGEST_LENGTH 20
        __device__  void SHA1_Init(SHA1_CTX * context);
#ifdef __cplusplus
    }
#endif
}
#endif /* SHA1_H */


//sha1.cu
#include <cuda_runtime.h>
#include "sha1.h"
namespace cudatest {
    __device__ void SHA1_Init(SHA1_CTX * context)
    {

    }
}
```
The main.cpp uses C/C++ compiler and sha1.cu uses CUDA C/C++

And I add openssl headers into the AdditionalIncludeDirectories,set directory which contains ssleay32.lib and libeay32.lib to library path,set AdditionalDependencies with ssleay32.lib, libeay32.lib .

Then the project built with no error and no warning. But when I run it or debug it,I found the function SHA1_Init runs into device code and the program crashed immediately.

why the compiler linked function SHA1_Init with the cuda device SHA1_Init implement which has a namespace cudatest wrapped instead of a ssleay32.lib, libeay32.lib CPU implement?

Related questions

Accessing GpuMat on another GPU

See also questions close to this topic