fix sort bug

[xaellison]

No description provided.

[maleadt]

You can reproduce this locally by restricting the number of threads kernel1 can use while setting the number for kernel2 higher than that (which shouldn't matter since it's the blocksize of a kernel we won't execute):

diff --git a/src/sorting.jl b/src/sorting.jl
index 7dd563831..3568aaafe 100644
--- a/src/sorting.jl
+++ b/src/sorting.jl
@@ -909,7 +909,7 @@ function bitonic_sort!(c; by = identity, lt = isless, rev = false, dims=1)
     # compile kernels (using Int32 for indexing, if possible, yielding a 70% speedup)
     I = c_len <= typemax(Int32) ? Int32 : Int
     args1 = (c, I(c_len), one(I), one(I), one(I), by, lt, Val(rev), Val(dims))
-    kernel1 = @cuda launch=false comparator_small_kernel(args1...)
+    kernel1 = @cuda maxthreads=896 launch=false comparator_small_kernel(args1...)

     config1 = launch_configuration(kernel1.fun, shmem = threads -> bitonic_shmem(c, threads))
     args2 = (c, I(c_len), one(I), one(I), by, lt, Val(rev), Val(dims))
@@ -917,6 +917,7 @@ function bitonic_sort!(c; by = identity, lt = isless, rev = false, dims=1)
     config2 = launch_configuration(kernel2.fun, shmem = threads -> bitonic_shmem(c, threads))
     # blocksize for kernel2 MUST be a power of 2
     threads2 = prevpow(2, config2.threads)
+    threads2 = 1024 # doesn't matter since we'll pick kernel1

     # determines cutoff for when to use kernel1 vs kernel2
     log_threads = threads2 |> log2 |> Int

julia> CUDA.bitonic_sort!(CUDA.rand(Int32, (2, 2, 50000)); dims=3)
ERROR: Number of threads per block exceeds kernel limit (1024 > 896).

[maleadt]

#2353