diff --git a/pocketfft_hdronly.h b/pocketfft_hdronly.h
index 0b955ba5407b7efdd4e0d9bb3ba6ad7150aad18b..d1eec5d1c9f16f905f589f416dd0bcdce4de2e82 100644
--- a/pocketfft_hdronly.h
+++ b/pocketfft_hdronly.h
@@ -2927,7 +2927,7 @@ POCKETFFT_NOINLINE void general_nd(const cndarr<T> &in, ndarr<T> &out,
         {
         it.advance(vlen);
         auto tdatav = reinterpret_cast<add_vec_t<T> *>(storage.data());
-        exec(it, in, out, tdatav, *plan, fct);
+        exec(it, tin, out, tdatav, *plan, fct);
         }
 #endif
     while (it.remaining()>0)
@@ -2935,7 +2935,7 @@ POCKETFFT_NOINLINE void general_nd(const cndarr<T> &in, ndarr<T> &out,
       it.advance(1);
       auto buf = allow_inplace && it.stride_out() == sizeof(T) ?
         &out[it.oofs(0)] : reinterpret_cast<T *>(storage.data());
-      exec(it, in, out, buf, *plan, fct);
+      exec(it, tin, out, buf, *plan, fct);
       }
 } // end of parallel region
     fct = T0(1); // factor has been applied, use 1 for remaining axes