Add SSE2 fallbacks for max_epi32 and min_epi32

Author: squk

si_normalmap.h

@@ -141,8 +141,23 @@ sinm_composite(const uint32_t *in1, const uint32_t *in2, uint32_t *out, int32_t
 #define simd__andnot_ps(a, b) simd_prefix_float(andnot_ps(a, b))
 #define simd__add_epi32(a, b) simd_prefix_float(add_epi32(a, b))
 #define simd__sub_epi32(a, b) simd_prefix_float(sub_epi32(a, b))
+
+#if defined(__AVX__) || defined(__SSE4_1__)
 #define simd__max_epi32(a, b) simd_prefix_float(max_epi32(a, b))
 #define simd__min_epi32(a, b) simd_prefix_float(min_epi32(a, b))
+#else
+static sinm__inline __m128i sinm__sse2_max_epi32(__m128i a, __m128i b) {
+    __m128i mask = _mm_cmpgt_epi32(a, b);
+    return _mm_or_si128(_mm_and_si128(mask, a), _mm_andnot_si128(mask, b));
+}
+static sinm__inline __m128i sinm__sse2_min_epi32(__m128i a, __m128i b) {
+    __m128i mask = _mm_cmpgt_epi32(b, a);
+    return _mm_or_si128(_mm_and_si128(mask, a), _mm_andnot_si128(mask, b));
+}
+#define simd__max_epi32(a, b) sinm__sse2_max_epi32(a, b)
+#define simd__min_epi32(a, b) sinm__sse2_min_epi32(a, b)
+#endif
+
 #define simd__loadu_ps(a) simd_prefix_float(loadu_ps(a))
 #define simd__srli_epi32(a, i) simd_prefix_float(srli_epi32(a, i))
 #define simd__slli_epi32(a, i) simd_prefix_float(slli_epi32(a, i))