MinMax

[EJainDev]

This resolves #568 although the test doesn't pass. Any ideas on why it doesn't pass? Most of the kernel implementation is xamax but changed for max and min.

[EJainDev]

I tested a dummy script, it seems to be a problem in the test implementation rather than in the actual code implementation, do you know what the error is?

EDIT: The problem is with testblas.cpp because I haven't added the min argument yet into it.
EDIT2: Fixed some more (potential) errors but it seems I'm still missing something.

[EJainDev]

I think the issue was with local memory. It was using to much local memory and then when I fixed that, WGS1 or WGS2 was not divisible by 2. I will add a Xaminmax version of this, and then I think it will be mergeable.

[EJainDev]

Should be ready for merge but the routines are incompatible with WGS1 or WGS2 not being divisible by 2.

[CNugteren]

I left some minor comments on the new code.

Regarding the issue with the local memory size (#634 (comment)) it seems you solved it by halving the sizes, but now the new issue that WGS needs to be a multiple of 2. I think this is a valid assumption, since https://github.com/CNugteren/CLBlast/blob/master/src/tuning/kernels/xdot.hpp#L64 only tests with those values anyway. Perhaps you can add a comment on the place where you divide WGS by 2 that refers to this line in the tuner?

Other than that I think my previous question remains unanswered: As for the tests that you did, I assume you testing all of the following:

• Run with the -full_test option?
• Run the above with a variety of values for WGS1 and WGS2?

[EJainDev]

I did run both full test and a variety of values for WGS1 and WGS2.

[CNugteren]

I did run both full test and a variety of values for WGS1 and WGS2.

Good. I'm assuming this includes all the values/combinations that the tuner searches for (apart from those that don't fit on your test device of course).

I think the only other thing left is the suggestion above regarding adding a link to the code in two places:

(...) Perhaps you can add a comment on the place where you divide WGS by 2 that refers to this line in the tuner?

After that I'll do a local test and then we should be ready to go.

[CNugteren]

I compiled and ran locally, but on my test system I do get one error for the double-complex case (or 4 when I run the full_test). Can you check? Might just be an error in the test itself?

* Running on OpenCL device 'Intel(R) UHD Graphics 620 [0x3ea0]'.
* Starting tests for the 'iZMINMAX' routine. Legend:
   : -> Test produced correct results
   . -> Test returned the correct error code
   X -> Test produced incorrect results
   / -> Test returned an incorrect error code
   \ -> Test not executed: OpenCL-kernel compilation error
   o -> Test not executed: Unsupported precision
   - -> Test not completed: Reference CBLAS doesn't output error codes
* Testing with error margins of 0.5% (relative) and 0.001 (absolute)
* Testing 'regular behaviour' for 'default':
   Testing: n=7 offx=0 incx=1 [CLBlast] [CPU BLAS] -> :
   Testing: n=7 offx=0 incx=2 [CLBlast] [CPU BLAS] -> :
   Testing: n=7 offx=0 incx=7 [CLBlast] [CPU BLAS] -> :
   Testing: n=64 offx=0 incx=1 [CLBlast] [CPU BLAS] -> :
   Testing: n=64 offx=0 incx=2 [CLBlast] [CPU BLAS] -> 
   Buffer overflow index 1:  38.00+0.00i (reference) versus  14.00+0.00i (CLBlast)
   Combined average L2 error: 0.00e+00
   X
   Testing: n=64 offx=0 incx=7 [CLBlast] [CPU BLAS] -> :
   Error rate 0.78%: n=64 offx=0 incx=2 
   Pass rate  83.3%: 5 passed / 0 skipped / 1 failed
* Completed all test-cases for this routine. Results:
   5 test(s) passed
   0 test(s) skipped
   1 test(s) failed

Let me know if you want more info.

[EJainDev]

Not done yet, got to test on another device, and that has a lot of errors. Probably something to do with the testing system...