diff --git a/src/nmod_mat.h b/src/nmod_mat.h
index 0b14b3c441..23322dd283 100644
--- a/src/nmod_mat.h
+++ b/src/nmod_mat.h
@@ -137,15 +137,8 @@ void nmod_mat_neg(nmod_mat_t B, const nmod_mat_t A);
 
 /* Matrix-scalar arithmetic */
 
-void _nmod_mat_scalar_mul_generic(nmod_mat_t B, const nmod_mat_t A, ulong c);
-void _nmod_mat_scalar_mul_precomp(nmod_mat_t B, const nmod_mat_t A, ulong c, ulong c_pr);
 void nmod_mat_scalar_mul(nmod_mat_t B, const nmod_mat_t A, ulong c);
 void nmod_mat_scalar_mul_fmpz(nmod_mat_t B, const nmod_mat_t A, const fmpz_t c);
-
-void _nmod_mat_scalar_addmul_ui_generic(nmod_mat_t C, const nmod_mat_t A, const nmod_mat_t B,
-                                       const ulong c);
-void _nmod_mat_scalar_addmul_ui_precomp(nmod_mat_t C, const nmod_mat_t A, const nmod_mat_t B,
-                                       const ulong c, const ulong c_pr);
 void nmod_mat_scalar_addmul_ui(nmod_mat_t C,
                        const nmod_mat_t A, const nmod_mat_t B, const ulong c);
 
diff --git a/src/nmod_mat/profile/p-scalar_addmul.c b/src/nmod_mat/profile/p-scalar_addmul.c
index 8ab8b1dcd1..79e244cca4 100644
--- a/src/nmod_mat/profile/p-scalar_addmul.c
+++ b/src/nmod_mat/profile/p-scalar_addmul.c
@@ -13,6 +13,11 @@
 #include "ulong_extras.h"
 #include "nmod_mat.h"
 
+void _nmod_mat_scalar_addmul_ui_generic(nmod_mat_t C, const nmod_mat_t A, const nmod_mat_t B,
+                                       const ulong c);
+void _nmod_mat_scalar_addmul_ui_precomp(nmod_mat_t C, const nmod_mat_t A, const nmod_mat_t B,
+                                       const ulong c, const ulong c_pr);
+
 typedef struct
 {
     flint_bitcnt_t bits;
diff --git a/src/nmod_mat/profile/p-scalar_mul.c b/src/nmod_mat/profile/p-scalar_mul.c
index 60e9379d3e..1420cb62f5 100644
--- a/src/nmod_mat/profile/p-scalar_mul.c
+++ b/src/nmod_mat/profile/p-scalar_mul.c
@@ -13,6 +13,9 @@
 #include "ulong_extras.h"
 #include "nmod_mat.h"
 
+void _nmod_mat_scalar_mul_generic(nmod_mat_t B, const nmod_mat_t A, ulong c);
+void _nmod_mat_scalar_mul_precomp(nmod_mat_t B, const nmod_mat_t A, ulong c, ulong c_pr);
+
 typedef struct
 {
     flint_bitcnt_t bits;