Party swapping was broken

src/cg.c

@@ -272,11 +272,29 @@ void FromPToCg(const size_t a_out, const size_t b_out, const size_t a_in,
   }
 }
 
+// To perform a swap for every combination, we need only to consider tuples
+//
+//    (oA, oB, iA, iB)
+//
+// such that, when lexicographically ordered (and we take this order to be
+// right-to-left, in the sense that a greater iB places an entry later, then
+// a greater iA, and so on), we have that
+//
+//    (oA, oB, iA, iB) < (oB, oA, iB, iA)
+//
+// meaning the left-hand side comes before, in the ordering, to the right-hand
+// side.
+// (The case where the two sides are equal is not covered, but that case does
+//  not require a swap.)
+//
+// This can easily be seen to mean that we require
+//
+//    iB < iA  OR  (iB = iA  AND  oB < oA)
 void PSwapParties(const size_t out, const size_t in, data_t *p) {
-  for (size_t a_out_i = 0; a_out_i < out; a_out_i++) {
-    for (size_t b_out_i = a_out_i + 1; b_out_i < out; b_out_i++) {
-      for (size_t a_in_i = 0; a_in_i < in; a_in_i++) {
-        for (size_t b_in_i = a_in_i + 1; b_in_i < in; b_in_i++) {
+  for (size_t a_in_i = 0; a_in_i < in; a_in_i++) {
+    for (size_t b_in_i = 0; b_in_i < a_in_i; b_in_i++) {
+      for (size_t a_out_i = 0; a_out_i < out; a_out_i++) {
+        for (size_t b_out_i = 0; b_out_i < out; b_out_i++) {
           size_t u = PIndex(out, out, in, in, a_out_i, b_out_i, a_in_i, b_in_i);
           size_t v = PIndex(out, out, in, in, b_out_i, a_out_i, b_in_i, a_in_i);
           data_t tmp = p[u];
@@ -285,6 +303,16 @@ void PSwapParties(const size_t out, const size_t in, data_t *p) {
         }
       }
     }
+
+    for (size_t a_out_i = 0; a_out_i < out; a_out_i++) {
+      for (size_t b_out_i = 0; b_out_i < a_out_i; b_out_i++) {
+          size_t u = PIndex(out, out, in, in, a_out_i, b_out_i, a_in_i, a_in_i);
+          size_t v = PIndex(out, out, in, in, b_out_i, a_out_i, a_in_i, a_in_i);
+          data_t tmp = p[u];
+          p[u] = p[v];
+          p[v] = tmp;
+      }
+    }
   }
 }