diff --git a/include/bx/float4_neon.h b/include/bx/float4_neon.h
index 479bcba..2331309 100644
--- a/include/bx/float4_neon.h
+++ b/include/bx/float4_neon.h
@@ -222,6 +222,8 @@ namespace bx
 
 #define float4_div_nr float4_div_nr_ni
 #define float4_div float4_div_nr_ni
+#define float4_ceil float4_ceil_ni
+#define float4_floor float4_floor_ni
 #include "float4_ni.h"
 
 #endif // __BX_FLOAT4_NEON_H__
diff --git a/include/bx/float4_ni.h b/include/bx/float4_ni.h
index 328c51c..08e4c55 100644
--- a/include/bx/float4_ni.h
+++ b/include/bx/float4_ni.h
@@ -402,6 +402,30 @@ namespace bx
 		return result;
 	}
 
+	BX_FLOAT4_INLINE float4_t float4_ceil_ni(float4_t _a)
+	{
+		const float4_t tmp0   = float4_ftoi(_a);
+		const float4_t tmp1   = float4_itof(tmp0);
+		const float4_t mask   = float4_cmplt(tmp1, _a);
+		const float4_t one    = float4_one();
+		const float4_t tmp2   = float4_and(one, mask);
+		const float4_t result = float4_add(tmp1, tmp2);
+
+		return result;
+	}
+
+	BX_FLOAT4_INLINE float4_t float4_floor_ni(float4_t _a)
+	{
+		const float4_t tmp0   = float4_ftoi(_a);
+		const float4_t tmp1   = float4_itof(tmp0);
+		const float4_t mask   = float4_cmpgt(tmp1, _a);
+		const float4_t one    = float4_one();
+		const float4_t tmp2   = float4_and(one, mask);
+		const float4_t result = float4_sub(tmp1, tmp2);
+
+		return result;
+	}
+
 } // namespace bx
 
 #endif // __BX_FLOAT4_NI_H__
diff --git a/include/bx/float4_ref.h b/include/bx/float4_ref.h
index bf74b40..4359119 100644
--- a/include/bx/float4_ref.h
+++ b/include/bx/float4_ref.h
@@ -517,6 +517,8 @@ IMPLEMENT_TEST(xyzw , 0xf);
 #define float4_normalize3 float4_normalize3_ni
 #define float4_dot3 float4_dot3_ni
 #define float4_dot float4_dot_ni
+#define float4_ceil float4_ceil_ni
+#define float4_floor float4_floor_ni
 #include "float4_ni.h"
 
 #endif // __BX_FLOAT4_REF_H__
diff --git a/include/bx/float4_sse.h b/include/bx/float4_sse.h
index ff14710..239fdaa 100644
--- a/include/bx/float4_sse.h
+++ b/include/bx/float4_sse.h
@@ -395,6 +395,8 @@ IMPLEMENT_TEST(xyzw , 0xf);
 #define float4_dot3 float4_dot3_ni
 #define float4_dot float4_dot_ni
 #endif // defined(__SSE4_1__)
+#define float4_ceil float4_ceil_ni
+#define float4_floor float4_floor_ni
 #include "float4_ni.h"
 
 #endif // __FLOAT4_SSE_H__
diff --git a/include/bx/float4_t.h b/include/bx/float4_t.h
index bdeeb91..9b251f8 100644
--- a/include/bx/float4_t.h
+++ b/include/bx/float4_t.h
@@ -11,7 +11,7 @@
 
 #define BX_FLOAT4_INLINE BX_FORCE_INLINE
 
-#if 0 // defined(__SSE2__)
+#if defined(__SSE2__)
 #	include "float4_sse.h"
 #elif 0 // __ARM_NEON__
 #	include "float4_neon.h"