diff --git a/drama/src/rev-mc.c b/drama/src/rev-mc.c
index 4c765a9..3471d0f 100644
--- a/drama/src/rev-mc.c
+++ b/drama/src/rev-mc.c
@@ -10,9 +10,14 @@
 #include <time.h>
 
 #include <algorithm>
+#include <atomic>
 #include <bitset>
+#include <condition_variable>
 #include <functional>
+#include <mutex>
+#include <queue>
 #include <set>
+#include <thread>
 #include <vector>
 
 #include "rev-mc.h"
@@ -295,29 +300,51 @@ find_row_mask(std::vector<set_t> &sets,
         }
     }
 
-    uint64_t row_mask  = LS_BITMASK(16);// use 16 bits for the row
-    uint64_t last_mask = (row_mask << (40 - 16));
+    uint64_t row_mask        = LS_BITMASK(16);// use 16 bits for the row
+    const uint64_t last_mask = (row_mask << (40 - 16));
     row_mask <<= CL_SHIFT;// skip the lowest 6 bits since they're used for CL addressing
 
-    while (row_mask < last_mask) {
-        uint64_t next_row_mask = next_bit_permutation(row_mask);
-        if (row_mask & LS_BITMASK(CL_SHIFT)) {
-            row_mask = next_row_mask;
-            continue;
-        }
-
+    auto resolve = [=](uint64_t row_mask) -> bool {
+        if (row_mask & LS_BITMASK(CL_SHIFT)) { return false; }
         for (auto addr_pool : same_row_sets) {
             addr_tuple base_addr = addr_pool[0];
             for (int i = 1; i < addr_pool.size(); i++) {
                 addr_tuple tmp = addr_pool[i];
-                if ((tmp.p_addr & row_mask) != (base_addr.p_addr & row_mask)) { goto next_mask; }
+                if ((tmp.p_addr & row_mask) != (base_addr.p_addr & row_mask)) { return false; }
             }
         }
+        return true;
+    };
 
-        break;
+    std::atomic_bool found{false};
+    std::mutex lock;
+    std::queue<uint64_t> mask_task;
 
-    next_mask:
-        row_mask = next_row_mask;
+    while (row_mask < last_mask) {
+        mask_task.push(row_mask);
+        row_mask = next_bit_permutation(row_mask);
+    }
+
+    std::vector<std::thread> worker;
+    for (int i = 0; i < 8; ++i) {
+        worker.emplace_back([&] {
+            while (!found) {
+                uint64_t cur_mask;
+                {
+                    std::lock_guard<std::mutex> _(lock);
+                    if (mask_task.empty()) { break; }
+                    cur_mask = mask_task.front();
+                    mask_task.pop();
+                }
+
+                if (resolve(cur_mask)) {
+                    std::lock_guard<std::mutex> _(lock);
+                    found.store(true);
+                    row_mask = cur_mask;
+                    while (!mask_task.empty()) mask_task.pop();
+                }
+            }
+        });
     }
 
     // super hackish way to recover the real row mask