I am trying to optimize my OpenCL code running on a Samsung Galaxy S7 (SM-G930F) with an Exynos Octa 8890 and a Mali-T880 using Android. According to OpenCL-Z the gpu supports OpenCL 1.2 full profile. I have a lot of OpenCL code correctly compiling and running. However, I cannot use any of the half_* or native_* functions such as: half_sin or native_sin. As soon as I add one of them I get a CL_BUILD_FAILURE when compiling the kernel. This is one kernel where such behaviour happens:
#pragma OPENCL EXTENSION cl_khr_fp16 : enable half distancePointToEllipse(short2 point, half2 ecenter, half phi, half a, half b) { half2 t = convert_half2(point) - ecenter; half angle = atan2(t.y, t.x) - phi; half cos1, sin1; //cos1 = native_cos(angle); //sin1 = native_sin(angle); sin1 = sincos(angle, &cos1); half tx = a * a * cos1 * cos1; half ty = b * b * sin1 * sin1; //half c = native_sqrt(tx + ty); //half d = native_sqrt(t.x * t.x + t.y * t.y); half c = sqrt(tx + ty); half d = hypot(t.x, t.y); return fabs(d - c); }
And this is the code used to compile it:
auto program_RANSACEllipses = cl::Program(context, CL_RANSAC_ELLIPSES_KERNEL, true); program_RANSACEllipses.build("-cl-std=CL1.2"); Kernels::RANSACEllipses = cl::Kernel(program_RANSACEllipses, "RANSAC");
I have downloaded the headers for OpenCL 1.2 from khronos.
Do you know what am I doing wrong?
Hi mompes,
Thanks for the question - I've raised this with the driver team and hope to have an answer for you soon.
Regards, Pete
Thank you very much. But I already found the problem, I was getting a compiling error which actually makes sense but I wasn't displaying compiling errors... my bad. I will write back in a while with more details in case someone has the same problem in the future.