Added cl_khr_fp16 extension support for test_explicit_s2v from basic

[shajder]

According to issue description (basic section):

#142

Additional remark (1): original test skips conversions between different types, in other words supported conversions were eg. char scalar->char vector, int scalar->int vector, float scalar->float vector and so on. This PR extend the scope of original test to integral scalar types->integral vector types, floating scalar types->floating vector types

Additional remark (2): test fails with my Intel device and kernel with half conversions:
ERROR: Explicit cast of scalar float to vector half2 FAILED; skipping other half vector tests
I will appreciate additional attention when it comes to testing cl_khr_fp16 with this test

[EwanC]

Regarding

Additional remark (2): test fails with my Intel device and kernel with half conversions:
ERROR: Explicit cast of scalar float to vector half2 FAILED; skipping other half vector tests
I will appreciate additional attention when it comes to testing cl_khr_fp16 with this test

I'm also seeing this error locally:

ERROR: Output value 0:0 does not validate for size 2:2!
       Input:   0x4dbe2551
       Actual:  0x7c007c00 0x7c007c00 0x7c007c00 0x7c007c00
ERROR: Explicit cast of scalar float to vector half2 FAILED; skipping other half vector tests

0x4dbe2551 is 398764576.0f which is outside the range of half, which is why the half2 is (0x7c00, 0x7c00), i.e INF values. So think we might need to check for overflow during the verification.

[shajder]

@EwanC I just added missing part of convert_explicit_value to support kHalf ExplicitType. Test passes now but I will appreciate reviewing it more carefully - I am not quite sure it is 100% correct

[EwanC]

Passes for me now 🙂 HALF_ROUND_CASE has a lot of code duplication with FLOAT_ROUND_CASE & DOUBLE_ROUND_CASE which would be nice to improve upon but i'm okay with the change without it.

[lakshmih]

Reviewing

[lakshmih]

This test is failing (on below kernel) with an error that cl_khr_fp16 is not enabled. Can we add the pragma for kernels that use half?

__kernel void test_conversion(__global float *sourceValues, __global half2 *destValues )
{
int tid = get_global_id(0);
float src = sourceValues[tid];

     destValues[tid] = (half2)src;

 }

[lakshmih]

Needs "#pragma OPENCL EXTENSION cl_khr_fp16 : enable\n" here.

[shajder]

Corrected - good find, thanks!

[bcalidas]

We are seeing failures with the latest commit. Debugging at our end. Need some more time.
Thanks,
Balaji

[lakshmih]

Added a patch that does the NAN check.

[bashbaug]

Seems OK but I have a few questions, mostly around NaN handling.

The additional test coverage does increase the runtime of this sub-test, but it still runs in ~30 seconds on my system so probably not worth worrying about (yet).

[bashbaug]

Discussed in the October 3rd teleconference. Will wait a week to address the comments above.

[bashbaug]

LGTM with one small nice-to-have - thanks!

[bashbaug]

Merging as discussed in the October 17th teleconference.