diff --git a/LeopardFF8.cpp b/LeopardFF8.cpp
index d29f6b4..dd36778 100644
--- a/LeopardFF8.cpp
+++ b/LeopardFF8.cpp
@@ -846,14 +846,18 @@ void Encode(
     // work <- data
 
     // FIXME: Unroll first loop to eliminate this
-    for (unsigned i = 0; i < original_count; ++i)
+    unsigned first_end = m;
+    if (original_count < m)
+    {
+        first_end = original_count;
+        for (unsigned i = original_count; i < m; ++i)
+            memset(work[i], 0, buffer_bytes);
+    }
+    for (unsigned i = 0; i < first_end; ++i)
         memcpy(work[i], data[i], buffer_bytes);
-    for (unsigned i = original_count; i < m; ++i)
-        memset(work[i], 0, buffer_bytes);
 
     // work <- IFFT(data, m, m)
 
-    const unsigned first_end = (original_count < m) ? original_count : m;
     for (unsigned width = 1; width < m; width <<= 1)
     {
         const unsigned range = width << 1;
diff --git a/tests/benchmark.cpp b/tests/benchmark.cpp
index dc04e6e..2f42b66 100644
--- a/tests/benchmark.cpp
+++ b/tests/benchmark.cpp
@@ -389,11 +389,11 @@ struct TestParameters
     unsigned original_count = 1000; // under 65536
     unsigned recovery_count = 100; // under 65536 - original_count
 #else
-    unsigned original_count = 128; // under 65536
-    unsigned recovery_count = 128; // under 65536 - original_count
+    unsigned original_count = 200; // under 65536
+    unsigned recovery_count = 20; // under 65536 - original_count
 #endif
     unsigned buffer_bytes = 64000; // multiple of 64 bytes
-    unsigned loss_count = 128; // some fraction of original_count
+    unsigned loss_count = 20; // some fraction of original_count
     unsigned seed = 0;
     bool multithreaded = true;
 };