Fix indexing for gpu rand

[christiangnrd]

Finish #615. The issue was that JLArrays wasn't limiting the number of workgroup threads so it was trying to access rand state past 256.

Closes #614
Closes #615

[maleadt]

Thanks for following up! Annoyingly though, test failures seem related.

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic master) to apply these changes.

Click here to view the suggested changes.

diff --git a/test/testsuite/random.jl b/test/testsuite/random.jl
index 6ecd235..bb493da 100644
--- a/test/testsuite/random.jl
+++ b/test/testsuite/random.jl
@@ -24,7 +24,7 @@
             rand!(rng, A)
             Random.seed!(rng, 1)
             rand!(rng, B)
-            @test Array(A) == Array(B) broken=SEEDING_BROKEN && (prod(d) > length(rng.state))
+            @test Array(A) == Array(B) broken = SEEDING_BROKEN && (prod(d) > length(rng.state))
 
             if rng != cpu_rng
                 rand!(cpu_rng, A)
@@ -64,7 +64,7 @@
             randn!(rng, A)
             Random.seed!(rng, 1)
             randn!(rng, B)
-            @test Array(A) == Array(B) broken=SEEDING_BROKEN && (prod(d) > (2 * length(rng.state)))
+            @test Array(A) == Array(B) broken = SEEDING_BROKEN && (prod(d) > (2 * length(rng.state)))
 
             if rng != cpu_rng
                 randn!(cpu_rng, A)

[christiangnrd]

The fix resurfaced #530, which wasn't detected before because we weren't testing on arrays with length greater than the maximum workgroup size. I think this should be merged once I find a way to properly mark the broken tests because that'll bring us back to where we were before the switch to KA introduced the bug.

[christiangnrd]

Failing Metal tests are related but I'll push a fix to mark the test broken once this is merged