diff --git a/src/test_vec.c b/src/test_vec.c
index bb7c852..ba33469 100644
--- a/src/test_vec.c
+++ b/src/test_vec.c
@@ -18,6 +18,8 @@
 #define vec_sigmoid vec_sigmoid_fast
 #define sgemv_accum16 sgemv_accum16_fast
 #define sparse_sgemv_accum16 sparse_sgemv_accum16_fast
+#define sgemv_accum8x4 sgemv_accum8x4_fast
+#define sparse_sgemv_accum8x4 sparse_sgemv_accum8x4_fast
 
 #ifdef __AVX__
 #include "vec_avx.h"
@@ -34,7 +36,6 @@ const char simd[]="SSE";
 const char simd[]="NEON";
 #else
 const char simd[]="none";
-#include "vec.h"
 #endif
 
 #undef celt_exp2
@@ -45,6 +46,12 @@ const char simd[]="none";
 #undef vec_sigmoid
 #undef sgemv_accum16
 #undef sparse_sgemv_accum16
+#undef sgemv_accum8x4 
+#undef sparse_sgemv_accum8x4 
+
+/* Override auto-detection of architecture so we have the non-optimized versions too. */
+#define NO_OPTIMIZATIONS
+#include "vec.h"
 
 #define ROW_STEP 16
 #define ROWS     ROW_STEP*10
diff --git a/src/vec.h b/src/vec.h
index d1f82b5..21b4484 100644
--- a/src/vec.h
+++ b/src/vec.h
@@ -35,14 +35,15 @@
 #include "arch.h"
 
 
-#if defined(__AVX__) || defined(__SSE2__)
+#if !defined(NO_OPTIMIZATIONS) && (defined(__AVX__) || defined(__SSE2__))
 #include "vec_avx.h"
-#elif defined(__ARM_NEON__) || defined(__ARM_NEON)
+#elif !defined(NO_OPTIMIZATIONS) && (defined(__ARM_NEON__) || defined(__ARM_NEON))
 #include "vec_neon.h"
 #else
 
 #define MAX_INPUTS (2048)
 
+#undef NO_OPTIMIZATIONS
 #define NO_OPTIMIZATIONS
 
 #ifndef DISABLE_DOT_PROD