That test is failing using clvk and various Vulkan drivers (llvmpipe, swiftshader, mesa-based drivers).

I am seeing multiple issues in the test itself:

1. OpenCL-CTS/test_conformance/profiling/execute_multipass.cpp

   Line 107 in 5b35180

   err = clGetDeviceInfo( device, CL_DEVICE_MAX_WORK_ITEM_SIZES, sizeof( cl_uint ), (size_t*)localThreads, NULL );

   
   localThreads type is size_t[3]. Thus the size given is not right, and we proper implementation should end up with undefined values in localThreads[1] & localThreads[2] at least. Most probably also localThreads[0] as cl_uint might be smaller than size_t...

2. each element of localThreads is clamp to its max value, except for localThreads[2]. I'm not sure I understand why, and would also clamp it

3. We end up with a workgroup using the maximum work-item sizes per dimensions (CL_DEVICE_MAX_WORK_ITEM_SIZES). This can be higher than CL_DEVICE_MAX_WORK_GROUP_SIZE on some devices, thus we would also need to make sure not to get higher than it.

4. The input and output buffer are declared as cl_uchar[w * h * d * nChannels]. But
   4.1.

   OpenCL-CTS/test_conformance/profiling/execute_multipass.cpp

   Line 136 in 5b35180

   sizeof(cl_float) * w * h * d * nChannels, NULL, &err);

   
   The output is created with a bigger size (*sizeof(cl_float))
   4.2.

   OpenCL-CTS/test_conformance/profiling/execute_multipass.cpp

   Line 240 in 5b35180

   err = clEnqueueReadBuffer(queue, memobjs[1], CL_TRUE, 0, w*h*d*nChannels*4, outptr, 0, NULL, NULL);

   
   The output is read with a bigger size (*4)

This is trigger segfault because the host buffer where the output should be stored is smaller than the CL buffer.