Double Precision not working

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I got some problems when i try to build double precision kernels with OpenCL:

this one works:

const char *gpcAddFloatSrc =

					"__kernel void addiere(__global float *gpfA, __global float *gpfB, __global float *gpfC, int iCount)\

					{\

						int liIDX = get_global_id(0);\

						if (liIDX < iCount)\

						{\

 							gpfC[liIDX] = gpfA[liIDX] + gpfB[liIDX];\

						}\

					}";

this one returns a program build failure:

const char *gpcAddDoubleSrc =

					"__kernel void addiere(__global double *gpdA, __global double *gpdB, __global double *gpdC, int iCount)\

					{\

						int liIDX = get_global_id(0);\

						if (liIDX < iCount)\

						{\

 							gpdC[liIDX] = gpdA[liIDX] + gpdB[liIDX];\

						}\

					}";

Of course I tried to enable the fp64 with

#pragma OPENCL EXTENSION cl_khr_fp64 : enable

My compiler (vc90) reclaims that this pragma is unknown, but that seems to be normal.

A Cuda version of that code (it’s only a test) runs without problems on my gt540m.

Can anyone help me? :(

2

#pragma OPENCL - it is for OpenCL source code, which is compiled by OpenCL compiler. Generic C(++) compiler cannot do anything with this pragma.

3

so it has to look like this?:

const char *gpcAddDoubleSrc =

                                        "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\

                                        __kernel void addiere(__global double *gpdA, __global double *gpdB, __global double *gpdC, int iCount)\

                                        {\

                                                int liIDX = get_global_id(0);\

                                                if (liIDX < iCount)\

                                                {\

                                                        gpdC[liIDX] = gpdA[liIDX] + gpdB[liIDX];\

                                                }\

                                        }";

I just tested this one…now the clBuildProgram Function says the Kernel name (“addiere”) is invalid …

In case the Compiler compiled the float version before, does the Compiler still remember that and actually claims that the name allready exists?

4

I don’t know… Maybe you put linefeed (\n) after the whole line with pragma declaration?

5

I need some Coffee…you’re a lifesaver.

Works<