Use threadfence_block instead of sync_threads inside of chacha_blocks!

Use threadfence_block to ensure that ChaCha rounds that writes to memory
performed by the parallelized ChaCha rounds are observed by all threads
in the same block before switching from columnar -> diagonal rounds or
diagonal -> columnar rounds. This function previously used sync_threads
to accomplish the same goal, which appears to have added roughly 2x or
so overhead.

Author: kernelmethod

src/ChaCha.jl

@@ -220,11 +220,11 @@ function _cuda_chacha_rounds!(state, doublerounds)
     for _ = 1:doublerounds
         # Columnar rounds
         _QR!(state_slice, i, i + 4, i + 8, i + 12)
-        CUDA.sync_threads()
+        CUDA.threadfence_block()
 
         # Diagonal rounds
         _QR!(state_slice, dgc1, dgc2, dgc3, dgc4)
-        CUDA.sync_threads()
+        CUDA.threadfence_block()
     end
 
     nothing