Use mixed atomics / non-atomics

This is probably the version of the test I want to promote.

Author: raphlinus

compute-shader-hello/src/shader.wgsl

@@ -14,9 +14,14 @@
 //
 // Also licensed under MIT license, at your choice.
 
+struct Element {
+    data: u32;
+    flag: atomic<u32>;
+};
+
 [[block]]
 struct DataBuf {
-    data: [[stride(4)]] array<atomic<u32>>;
+    data: [[stride(8)]] array<Element>;
 };
 
 [[block]]
@@ -48,23 +53,17 @@ fn main([[builtin(global_invocation_id)]] global_id: vec3<u32>) {
     var failures = 0u;
     for (var i: u32 = 0u; i < n_iter; i = i + 1u) {
         let wr_flag_ix = permute_flag_ix(ix);
-        atomicStore(&data_buf.data[ix * 2u], i + 1u);
+        data_buf.data[ix].data = i + 1u;
         storageBarrier(); // release semantics for writing flag
-        atomicStore(&data_buf.data[wr_flag_ix * 2u + 1u], i + 1u);
+        atomicStore(&data_buf.data[wr_flag_ix].flag, i + 1u);
 
         // Read from a different workgroup
         let read_ix = ((ix & 0xffu) << 8u) | (ix >> 8u);
         let read_flag_ix = permute_flag_ix(read_ix);
 
-        let flag = atomicLoad(&data_buf.data[read_flag_ix * 2u + 1u]);
-        //let flag = atomicOr(&data_buf.data[read_flag_ix * 2u + 1u], 0u);
+        let flag = atomicLoad(&data_buf.data[read_flag_ix].flag);
         storageBarrier(); // acquire semantics for reading flag
-        var data = 0u;
-        if (strategy == 0u) {
-            data = atomicLoad(&data_buf.data[read_ix * 2u]);
-        } else {
-            data = atomicOr(&data_buf.data[read_ix * 2u], 0u);
-        }
+        let data = data_buf.data[read_ix].data;
         if (flag > data) {
             failures = failures + 1u;
         }