Function REGISTER_TEST

test_conformance/SVM/test_byte_granularity.cpp:51–162

REGISTER_TEST(svm_byte_granularity)
{
    clContextWrapper contextWrapper;
    clProgramWrapper program;
    clKernelWrapper k1, k2;
    clCommandQueueWrapper queues[MAXQ];

    cl_uint num_devices = 0;
    cl_int err = CL_SUCCESS;

    err = create_cl_objects(device, &byte_manipulation_kernels[0],
                            &contextWrapper, &program, &queues[0], &num_devices,
                            CL_DEVICE_SVM_FINE_GRAIN_BUFFER);
    context = contextWrapper;
    if (err == 1)
        return 0; // no devices capable of requested SVM level, so don't execute
                  // but count test as passing.
    if (err < 0) return -1; // fail test.

    cl_uint num_devices_plus_host = num_devices + 1;

    k1 = clCreateKernel(program, "write_owned_locations", &err);
    test_error(err, "clCreateKernel failed");
    k2 = clCreateKernel(program, "sum_neighbor_locations", &err);
    test_error(err, "clCreateKernel failed");


    cl_char *pA = (cl_char *)clSVMAlloc(
        context, CL_MEM_READ_WRITE | CL_MEM_SVM_FINE_GRAIN_BUFFER,
        sizeof(cl_char) * num_elements, 0);

    cl_uint **error_counts = (cl_uint **)malloc(sizeof(void *) * num_devices);

    for (cl_uint i = 0; i < num_devices; i++)
    {
        error_counts[i] = (cl_uint *)clSVMAlloc(
            context, CL_MEM_READ_WRITE | CL_MEM_SVM_FINE_GRAIN_BUFFER,
            sizeof(cl_uint), 0);
        *error_counts[i] = 0;
    }
    for (int i = 0; i < num_elements; i++) pA[i] = -1;

    err |= clSetKernelArgSVMPointer(k1, 0, pA);
    err |= clSetKernelArg(k1, 1, sizeof(cl_uint), &num_devices_plus_host);
    test_error(err, "clSetKernelArg failed");

    // get all the devices going simultaneously
    size_t element_num = num_elements;
    for (cl_uint d = 0; d < num_devices; d++) // device ids starting at 1.
    {
        err = clSetKernelArg(k1, 2, sizeof(cl_uint), &d);
        test_error(err, "clSetKernelArg failed");
        err = clEnqueueNDRangeKernel(queues[d], k1, 1, NULL, &element_num, NULL,
                                     0, NULL, NULL);
        test_error(err, "clEnqueueNDRangeKernel failed");
    }

    for (cl_uint d = 0; d < num_devices; d++) clFlush(queues[d]);