Various bugfixes, still broken

exercises/src/gpu/grayscott.comp

@@ -174,8 +174,7 @@ void main() {
             }
         #endif
 
-        // Validated the tile selection, now make sure we fetch up-to-date data
-        // in process_tile.
+        // Make sure process_tile operates on up-to-date data
         memoryBarrierBuffer();
 
         // Process selected tile
@@ -209,7 +208,7 @@ void main() {
         #endif
 
         // Terminate work-group once it finishes a tile
-        if (s_output_buffer_idx == LAST_BUFFER_IDX) {
+        if (s_input_buffer_idx == LAST_BUFFER_IDX) {
             #if DEBUG
                 if (is_leader()) atomicAdd(g_metadata.num_normal_exit, 1);
             #endif

exercises/src/gpu/step.comp

@@ -22,10 +22,12 @@ bool leader_try_acquire_tile() {
         }
     #endif
 
+    const uint expected_input_idx = s_input_buffer_idx;
+    const uint expected_steal_count = s_expected_steal_count;
     do {
         // Try to signal other work-groups that we're working on this tile
-        const uint desired_steal_count = s_expected_steal_count + 1;
-        const uint desired_tile_status = encode_tile_status(s_input_buffer_idx,
+        const uint desired_steal_count = expected_steal_count + 1;
+        const uint desired_tile_status = encode_tile_status(expected_input_idx,
                                                             desired_steal_count,
                                                             s_expected_await_count);
         const uint status_before_cas = comp_swap_tile_status(
@@ -58,16 +60,14 @@ bool leader_try_acquire_tile() {
 
         // Update state variables, prepare to analyze the transition
         s_expected_tile_status = status_before_cas;
-        const uint old_input_idx = s_input_buffer_idx;
-        const uint old_steal_count = s_expected_steal_count;
         decode_tile_status(status_before_cas,
                            s_input_buffer_idx,
                            s_expected_steal_count,
                            s_expected_await_count);
 
         // Did another work-group steal our tile?
-        if ((s_input_buffer_idx != old_input_idx)
-            || (s_expected_steal_count != old_steal_count))
+        if ((s_input_buffer_idx != expected_input_idx)
+            || (s_expected_steal_count != expected_steal_count))
         {
             #if PROFILE
                 s_num_stolen_early += 1;
@@ -112,7 +112,9 @@ bool leader_try_finish_step() {
         if (status_before_cas == s_expected_tile_status) {
             s_expected_tile_status = desired_tile_status;
             s_input_buffer_idx = s_output_buffer_idx;
-            s_output_buffer_idx = output_buffer_idx(prev_input_idx);
+            if (s_input_buffer_idx < LAST_BUFFER_IDX) {
+                s_output_buffer_idx = output_buffer_idx(prev_input_idx);
+            }
             return true;
         }
 
@@ -129,7 +131,7 @@ bool leader_try_finish_step() {
             }
         #endif
 
-        // Update state variables, prepare to analyze the transition
+        // Update state variables
         s_expected_tile_status = status_before_cas;
         decode_tile_status(status_before_cas,
                            s_input_buffer_idx,

exercises/src/gpu/switch.comp

@@ -401,12 +401,6 @@ void switch_tile() {
         }
     #endif
 
-    // Release ownership of the current tile
-    if (s_acquired_tile) {
-        if (is_leader()) leader_release_tile();
-        barrier();
-    }
-
     // Start with a tile block surrounding our current tile
     //
     // Center it as much as possible to maximize the odds of finding a close
@@ -423,6 +417,10 @@ void switch_tile() {
         );
     }
 
+    // Release ownership of the current tile
+    if (is_leader() && s_acquired_tile) leader_release_tile();
+    barrier();
+
     // Repeatedly attempt to find a better tile
     uint attempts;
     for (attempts = 0; attempts < MAX_SWITCH_ATTEMPTS; ++attempts) {