[OpenCL] Failure writing the same vector type from different threads

[sanguinariojoe]

I faced this bug on OpenCL, but it is probably affecting some other tools.

As confirmed here, the following kernel should be working just fine:

__kernel void test_barrier(__global int2* a)
{
   const int i = get_global_id(0);
   a[i].y = i;
   a[i + 1].x = i;
}

However, as discussed here, clang is failing.

[llvmbot]

@llvm/issue-subscribers-opencl

Author: Jose Luis Cercós Pita (sanguinariojoe)

I faced this bug on OpenCL, but it is probably affecting some other tools.

As confirmed here, the following kernel should be working just fine:

__kernel void test_barrier(__global int2* a)
{
   const int i = get_global_id(0);
   a[i].y = i;
   a[i + 1].x = i;
}

However, as discussed here, clang is failing.

[svenvh]

This seems to be a long-standing issue, see also KhronosGroup/OpenCL-Docs#581. I think we may need a specification clarification first.

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Jose Luis Cercós Pita (sanguinariojoe)

I faced this bug on OpenCL, but it is probably affecting some other tools.

As confirmed here, the following kernel should be working just fine:

__kernel void test_barrier(__global int2* a)
{
   const int i = get_global_id(0);
   a[i].y = i;
   a[i + 1].x = i;
}

However, as discussed here, clang is failing.

[sanguinariojoe]

It seems indeed diplicated... Sorry to bother, feel free to close this then