fixes radix sort error

[rosenrodt]

insert memory fence to avoid AMD OpenCL compiler reordering the read/writes; fixes issue #811

CTest log for Vega56

99% tests passed, 2 tests failed out of 169
 
Total Test time (real) =  85.07 sec

The following tests FAILED:
 	143 - misc.amd_cpp_kernel_language (Failed)
 	148 - example.amd_cpp_kernel (Exit code 0xc0000409)
Errors while running CTest

[rosenrodt]

I am closing this PR because some issues still spill over after a clean rebuild on my test machine. Will reopen when it's done

[coveralls]

Coverage remained the same at 84.02% when pulling 81dc8c7 on rosenrodt:sort-failure-on-amd-vega into 2135633 on boostorg:develop.

[jszuppe]

You reopened the PR, but the code is the same. How is mem_fence related to what you wrote in #811 (comment)?

[rosenrodt]

I closed and reopened the PR because I was unsure of the cause of failed assertions when invoking is_sorted() on sorted arrays, but later figured it was totally unrelated to radix sort. Details were described in my comment here

The PR is to make sure the memory access is in correct order and not mangled up by the compiler, therefore the mem_fence in line 179. But I rarely encounter such a case where compiler is not respecting obvious variable dependencies. So this is just an ad hoc fix until maybe we can get AMD to look into this, especially by the fact that it only happens on AMD Vega/RX5xx and not the older chips nor chips from other vendors.

Another mem_fence is inserted for no apparent reason because it has to be like this to work on Vegas.

[jszuppe]

Hmm.. At least a fence for global memory would make more sense, does that option work too?

[rosenrodt]

Fences for global, local or both also work. So it won't make a difference. But I think you're right on that: global fencing is more appropriate in literal sense

[jszuppe]

If I'm correct this kernel is run with one thread (via enqueue_task), so those fences should not have any performance cost. Because of that it shouldn't be a problem to include that workaround after:

• You change fence flag to global memory
• In those lines add reference to Sort algorithms fail running on AMD Radeon RX Vega 56 #811 bug, for example // workaround for Radeon RX Vega bug, see #811.

[rosenrodt]

bump 😃
@jszuppe anything else that needs to be done?

[jszuppe]

No, it's ok. I just didn't have time to look at it.