Ask Your Question
0

Use Opencl to access Umat data

asked 2020-11-21 04:41:25 -0600

AndyK gravatar image

I want to use the opencl function atan2(y, x) which can take vectors for x and Y and perform a per-element operation.

Is there a way I can pass in pointers to two Umats (2 dimensional, float) and get a per-element atan2 operation?

This would obviously save having to copy from Gpu to CPU and back.

edit retag flag offensive close merge delete

Comments

opencl function atan2(y, x) which can take vectors for x and Y

where did you find this ? (i don't think such a thing exists)

pointers to two Umats

you should not use pointers at all here

berak gravatar imageberak ( 2020-11-21 05:49:23 -0600 )edit
1

Use OpenGL, it's more than capable.

sjhalayka gravatar imagesjhalayka ( 2020-11-21 10:23:49 -0600 )edit

3 answers

Sort by ยป oldest newest most voted
2

answered 2020-11-26 13:38:03 -0600

AndyK gravatar image

Thanks for the responses.

cartToPolar only calculates the angle to an accuracy of 0.3 degrees, which is no good for me.

However, I solved this by implementing my own atan2(x,y) [note the switched function signature!] function using OpenCL accelerated OpenCV. There are one or two unnecessary copies so it needs some tidying, but its fast and produces the same result as std::atan2. It also risks divide-by-zero so I need to implement a mask to prevent this (unless there is some sort of default protection against div 0 in OpenCV divide()?Its based on the reference implementation of nVidia code from https://developer.download.nvidia.com... but I'm still curious to know if there is a cleaner way to just do this using the native OpenCL atan2(y,x)(https://www.khronos.org/registry/Open...)

and feeding in the handle to the UMats?

Anyway, here is my messy, but working, solution:

   void gpu_atan2(UMat& x_in, UMat& y_in, UMat& mask, UMat& t3)
{
    // Calculate a good approximation of atan2 on the gpu using openCV UMats
    // Formula adapted from (https://developer.download.nvidia.com/cg/atan2.html)
    // This seems to be accurate (compared to atan2 CPU) to about 0.0002 degrees

    UMat t0, t1, t2,t3_mul, t4, t3_abs, t1_abs;

    absdiff(x_in, (float)0.0, t3_abs); /* A trick to get abs(x_in) */
    absdiff(y_in,(float)0.0, t1_abs); /* A trick to get abs(y_in) */    
    max(t3_abs, t1_abs, t0); 

    // We have to divide by t0 to calculate t3
    // Lets create a t0 mask that masks out any zero values
    UMat t0_mask(t0.size(), CV_8UC1, cv::Scalar(0));
    compare(t0, (float)0.0, t0_mask, cv::CMP_NE); /* Sets anything not equal to zero to 255 */
    // TODO - use this mask....

    min(t3_abs, t1_abs, t1);
    divide((float)1.0, t0, t3);
    multiply(t1, t3, t3_mul);

    cv::pow(t3_mul, 2, t4);
    //multiply(t3_mul, t3_mul, t4);

    t0.setTo((float)-0.013480470);

    UMat t0_mul(t0.size(), t0.type());
    multiply(t0, t4, t0_mul);
    add(t0_mul, (float)0.057477314, t0);
    multiply(t0, t4, t0_mul);
    subtract(t0_mul, (float)0.121239071, t0);
    multiply(t0, t4, t0_mul);
    add(t0_mul, (float)0.195635925, t0);
    multiply(t0, t4, t0_mul);
    subtract(t0_mul, (float)0.332994597, t0);
    multiply(t0, t4, t0_mul);
    add(t0_mul, (float)0.999995630, t0);

    multiply(t0, t3_mul, t3);


    UMat t1_gt_t3(t1_abs.size(), CV_8UC1, cv::Scalar(0));
    compare(t1_abs, t3_abs, t1_gt_t3, cv::CMP_GT);
    UMat sub1;
    subtract(Scalar((float)1.570796327), t3, sub1);
    //subtract(Scalar((float)1.570796327), t3, t3_mul, t1_gt_t3);
    sub1.copyTo(t3, t1_gt_t3);

    UMat x_in_lt_zero(x_in.size(), CV_8UC1, cv::Scalar(0));
    compare(x_in, Scalar(0.0), x_in_lt_zero, cv::CMP_LT);
    subtract(Scalar((float)3.141592654), t3, sub1);
    //subtract(Scalar((float)3.141592654), t3_mul, t3, x_in_lt_zero);
    sub1.copyTo(t3, x_in_lt_zero);

    UMat y_in_lt_zero(y_in.size(), CV_8UC1, cv::Scalar(0));
    compare(y_in, Scalar((float)0.0), y_in_lt_zero, cv::CMP_LT);
    UMat t3_all_negated(y_in.size(), y_in.type());
    multiply(t3,Scalar((float)-1.0), t3_all_negated); /* Annoyingly we cannot use a mask with multiply in OpenCV */
    t3_all_negated.copyTo(t3, y_in_lt_zero);


}
edit flag offensive delete link more
1

answered 2020-11-21 14:31:00 -0600

crackwitz gravatar image

you mention OpenCL. are you writing an OpenCL kernel?

if you have a cv::UMat(), you can call the .handle() method to get an OpenCL buffer handle: https://docs.opencv.org/master/d7/d45...

edit flag offensive delete link more
0

answered 2020-11-21 14:40:50 -0600

berak gravatar image

if you can live with it also calculating the magnitude, cartToPolar has ocl support, just give it UMat's as in/output

edit flag offensive delete link more

Question Tools

1 follower

Stats

Asked: 2020-11-21 04:41:25 -0600

Seen: 994 times

Last updated: Nov 26 '20