Remove subgroup_uniform_control_flow test.

[luciechoi]

Update 11/03/2025
We would like to remove this test and add similar tests when porting vulkan conformance tests. Currently, there's not a good way to specify the subgroup size inside the tests.

Fixes #490

[luciechoi]

cc: @s-perron

[s-perron]

I think the test needs to be updated again.

I'm also not sure what this is trying to test anymore. Before, the code in the if and else block were the "same", so a compiler might try to remove the if-statement:

if (c) dosomething();
else dosomething();

is equivalent to

dosomething()

for CPUs. In that vain, I could see some compiler optimizations being done that would merge the execution of the calls to WaveActiveMax(TID.x). Looking at this new test, each call is distinct. Can you add a comment explaining what you are trying to check in this case?

[s-perron]

You need to do something that will work for all wave sizes. This does not work for wave sizes 1 and 2.

[luciechoi]

Hi Steven, I think the minimum wave size we can get is 4 from the spec

But please let me know if there are preferred ways to test different wave size (and get same result!).

[s-perron]

I don't know if that is true in Vulkan.

https://docs.vulkan.org/spec/latest/chapters/interfaces.html#interfaces-builtin-variables-sgs

The minimum size is implementation defined by the minSubgroupSize, and it must be at least 1: https://docs.vulkan.org/spec/latest/chapters/devsandqueues.html#limits-minSubgroupSize

[luciechoi]

(I guess it still doesn't work if the wave size is an odd number.. 5, 7)

[s-perron]

But please let me know if there are preferred ways to test different wave size (and get same result!).

I'm not sure how to do that. We can work on it.

[damyanp]

(I guess it still doesn't work if the wave size is an odd number.. 5, 7)

Wave size (& I'm pretty sure subgroup size in vulkan) will always be a power of 2.

[s-perron]

Correct it has to be a power of two in vulkan

[luciechoi]

Sure, added a comment. The test example is from the first scenario described in this article, where waves are expected to reconverge after the nested non-uniform branch with the flag (Block B).

Sorry, I understood your comment. Changed the test to have the same block, to trigger compiler optimization.

I think the test needs to be updated again.

I'm also not sure what this is trying to test anymore. Before, the code in the if and else block were the "same", so a compiler might try to remove the if-statement:

if (c) dosomething();
else dosomething();

is equivalent to

dosomething()

for CPUs. In that vain, I could see some compiler optimizations being done that would merge the execution of the calls to WaveActiveMax(TID.x). Looking at this new test, each call is distinct. Can you add a comment explaining what you are trying to check in this case?

[luciechoi]

@s-perron @damyanp

(Updated the PR to remove the test as discussed today)