Hi,
I’m using ;
Developer Drivers for Linux (260.24)
cudatoolkit_3.2.9_linux_64_ubuntu10.04.run
consider this kernel :
__kernel void render(__read_only image3d_t volume)
{
// read from 3D texture
float4 voxel = read_imagef(volume, volumeSampler, (float4)(1.0f,1.0f,1.0f,1.0f));
// whatever
}
If read_image_f is not called, ciErrNum = clBuildProgram(cpProgram, 0, NULL, NULL, NULL, NULL); segfault!
Hi,
I’m using ;
Developer Drivers for Linux (260.24)
cudatoolkit_3.2.9_linux_64_ubuntu10.04.run
consider this kernel :
__kernel void render(__read_only image3d_t volume)
{
// read from 3D texture
float4 voxel = read_imagef(volume, volumeSampler, (float4)(1.0f,1.0f,1.0f,1.0f));
// whatever
}
If read_image_f is not called, ciErrNum = clBuildProgram(cpProgram, 0, NULL, NULL, NULL, NULL); segfault!
Unfortunately NVidia OpenCL compiler doesn’t give you error messages what’s wrong with your code, it only crashes :angry: Change your code to become syntax right. Define volumeSampler and compilation runs fine after that.
consider this kernel :
__kernel void render(__read_only image3d_t volume)
{
const volumeSampler = CLK_FILTER_NEAREST;
// read from 3D texture
float4 voxel = read_imagef(volume, volumeSampler, (float4)(1.0f,1.0f,1.0f,1.0f));
// whatever
}
Unfortunately NVidia OpenCL compiler doesn’t give you error messages what’s wrong with your code, it only crashes :angry: Change your code to become syntax right. Define volumeSampler and compilation runs fine after that.
consider this kernel :
__kernel void render(__read_only image3d_t volume)
{
const volumeSampler = CLK_FILTER_NEAREST;
// read from 3D texture
float4 voxel = read_imagef(volume, volumeSampler, (float4)(1.0f,1.0f,1.0f,1.0f));
// whatever
}
Oups, I’m sorry, Of course, volumeSample is a kernel parameter.
__kernel void render(__read_only image3d_t volume, sampler_t volumeSampler)
{
//
}
I wanted to simplify but did it too much.
Oups, I’m sorry, Of course, volumeSample is a kernel parameter.
__kernel void render(__read_only image3d_t volume, sampler_t volumeSampler)
{
//
}
I wanted to simplify but did it too much.
Well, after that I can’t reproduce clBuildProgram segfaulting. I’m using driver version 256.40 and Ubuntu toolkit version, though.
Well, after that I can’t reproduce clBuildProgram segfaulting. I’m using driver version 256.40 and Ubuntu toolkit version, though.
ok, thanks.
So maybe I should not use the 3.2rc…
ok, thanks.
So maybe I should not use the 3.2rc…
Sorry for misleading you about error messages from nvidia OpenCL compiler - register callback function at clBuildProgram and in that callback function call clGetProgramBuildInfo with CL_PROGRAM_BUILD_LOG . I must have been blinded, great function!
Try to read the CL_PROGRAM_BUILD_LOG first before you change your driver to older one.
Sorry for misleading you about error messages from nvidia OpenCL compiler - register callback function at clBuildProgram and in that callback function call clGetProgramBuildInfo with CL_PROGRAM_BUILD_LOG . I must have been blinded, great function!
Try to read the CL_PROGRAM_BUILD_LOG first before you change your driver to older one.
I did not know that, and that’s very very very what I need.
Unfortunately, it segfauts before !
But, well… Now that I know that I must use at leat one time read_imagef for the compiler to be happy, I do it once at the begining of the kernel
Eric
I did not know that, and that’s very very very what I need.
Unfortunately, it segfauts before !
But, well… Now that I know that I must use at leat one time read_imagef for the compiler to be happy, I do it once at the begining of the kernel
Eric
Strange you aren’t able to read the build log. Although I sometimes end with segmentation fault in clBuildProgram I always get that info, because the callback function is called in those cases as well.
Fine you found a workaround, but it is rather strange. I wish NVidia developers answer our questions. (AMD forums are much better and there are questions quickly solved.)
Strange you aren’t able to read the build log. Although I sometimes end with segmentation fault in clBuildProgram I always get that info, because the callback function is called in those cases as well.
Fine you found a workaround, but it is rather strange. I wish NVidia developers answer our questions. (AMD forums are much better and there are questions quickly solved.)
I wanted to confirm that I have seen the same issue.
Ubuntu 10.10 64-bit
GeForce GTX 460
NVIDIA developer driver 3.2 Linux 64 260.19.26
This code
const samper_t imgSamp = CLK_NORMALIZED_COORDS_TRUE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;
__kernel void aKernel(read_only image3d_t vol)
{
}
will cause clBuildProgram to segfault (before the build log can even be retrieved). Adding a call to read_image, though, like so
const samper_t imgSamp = CLK_NORMALIZED_COORDS_TRUE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;
__kernel void aKernel(read_only image3d_t vol)
{
float4 voxel = read_imagef(vol, imgSamp, (float4)(0.f));
}
will compile just fine, and with no errors in the build log.
The NVIDIA OpenCL implementation has some serious bugs in it with image objects. I personally would consider the image extension ‘beta’ in all current drivers.