Modernization of conversions test

[shajder]

.. in connection to preparations to handle cl_khr_fp16 extension which was already introduced in original PR. It was separated due to vendor's request.

for consistency below are my original remarks for original PR:

• In the process of adding cl_khr_fp16 support source code was unified and as result heavily refactored.

• Some pieces of code (eg. InitCL fuction) stayed as it was previously for the sake of comparison purposes. For consistency I would suggest to move it to methods (eg. ConversionsTest::SetUp).

• Please take a closer look at method DataInfoSpec<>::init under condition is_in_half(). There is new code to initialize array of cl_half values for conversions. It was source of the problem due to some special values which brought undefined results. ATM this code avoids those problematic values.

• Capacity of test was decreased. I had some difficult time trying to figure out why this number is so extremely big. I will appreciate your feedback. I limited number of tests to 16 777 216 conversions (uint64_t lastCase = 1000000ULL;)

• At my machine intel driver do not implement some of kernel functions which I think should be available:

#pragma OPENCL EXTENSION cl_khr_fp16 : enable
__kernel void test_convert_half_sat_uchar( __global uchar *src, __global half *dest )
{
   size_t i = get_global_id(0);
   dest[i] = convert_half_sat( src[i] );
}
Build not successful for device "Intel(R) UHD Graphics [0x9a60]", status: CL_BUILD_ERROR
Build log for device "Intel(R) UHD Graphics [0x9a60]" is: ------------
3632:5:14: error: implicit declaration of function 'convert_half_sat' is invalid in OpenCL
   dest[i] = convert_half_sat( src[i] );

[lakshmih]

Looks good.

[svenvh]

I found this quite difficult to review, due to large pieces being moved around while functionality is being changed simultaneously. I've left some initial comments, but I'm not sure I'll be able to say I'm confident with the PR. Any attempts to break this into smaller PRs would be appreciated, if possible. :)

[shajder]

I found this quite difficult to review, due to large pieces being moved around

This is a consequence of removing huge duplication of code which was requested in an issue discussion. According to my calculation this PR removes 3700 lines of clang formatted code, I think it is worth reviewing.

while functionality is being changed simultaneously.

This wasn't my intention, functionality should stay intact and if you spot such case I will greatly appreciate you pointing it out, thanks.

[svenvh]

I've found another issue with this PR: running a specific test (e.g. float_rte_char or char_sat_uchar) causes tests with other saturation/rounding modes to be run as well.

[shajder]

I've found another issue with this PR: running a specific test (e.g. float_rte_char or char_sat_uchar) causes tests with other saturation/rounding modes to be run as well.

Corrected, thanks!

[lakshmih]

Reviewed an earlier PS however would like to test new commits and need some time because of how long this test takes to complete.