Optimize initUndistortRectifyMap with AVX2

diff --git a/modules/imgproc/src/undistort.cpp b/modules/imgproc/src/undistort.cpp
index 127481f..e08cace 100644 (file)
--- a/modules/imgproc/src/undistort.cpp
+++ b/modules/imgproc/src/undistort.cpp
@@ -136,6 +136,10 @@ void cv::initUndistortRectifyMap( InputArray _cameraMatrix, InputArray _distCoef
     cv::Matx33d matTilt = cv::Matx33d::eye();
     cv::detail::computeTiltProjectionMatrix(tauX, tauY, &matTilt);
 
+#if CV_AVX2
+    bool USE_AVX2 = cv::checkHardwareSupport(CV_CPU_AVX2);
+#endif
+
     for( int i = 0; i < size.height; i++ )
     {
         float* m1f = map1.ptr<float>(i);
@@ -144,7 +148,150 @@ void cv::initUndistortRectifyMap( InputArray _cameraMatrix, InputArray _distCoef
         ushort* m2 = (ushort*)m2f;
         double _x = i*ir[1] + ir[2], _y = i*ir[4] + ir[5], _w = i*ir[7] + ir[8];
 
-        for( int j = 0; j < size.width; j++, _x += ir[0], _y += ir[3], _w += ir[6] )
+        int j = 0;
+#if CV_AVX2
+if( USE_AVX2 )
+{
+        static const __m256d __one = _mm256_set1_pd(1.0);
+        static const __m256d __two = _mm256_set1_pd(2.0);
+
+        const __m256d __matTilt_00 = _mm256_set1_pd(matTilt(0, 0));
+        const __m256d __matTilt_10 = _mm256_set1_pd(matTilt(1, 0));
+        const __m256d __matTilt_20 = _mm256_set1_pd(matTilt(2, 0));
+
+        const __m256d __matTilt_01 = _mm256_set1_pd(matTilt(0, 1));
+        const __m256d __matTilt_11 = _mm256_set1_pd(matTilt(1, 1));
+        const __m256d __matTilt_21 = _mm256_set1_pd(matTilt(2, 1));
+
+        const __m256d __matTilt_02 = _mm256_set1_pd(matTilt(0, 2));
+        const __m256d __matTilt_12 = _mm256_set1_pd(matTilt(1, 2));
+        const __m256d __matTilt_22 = _mm256_set1_pd(matTilt(2, 2));
+
+        for( ; j <= size.width - 4; j += 4, _x += 4 * ir[0], _y += 4 * ir[3], _w += 4 * ir[6] )
+        {
+            // Question: Should we load the constants first?
+            __m256d __w = _mm256_div_pd(__one, _mm256_set_pd(_w + 3 * ir[6], _w + 2 * ir[6], _w + ir[6], _w));
+            __m256d __x = _mm256_mul_pd(_mm256_set_pd(_x + 3 * ir[0], _x + 2 * ir[0], _x + ir[0], _x), __w);
+            __m256d __y = _mm256_mul_pd(_mm256_set_pd(_y + 3 * ir[3], _y + 2 * ir[3], _y + ir[3], _y), __w);
+            __m256d __x2 = _mm256_mul_pd(__x, __x);
+            __m256d __y2 = _mm256_mul_pd(__y, __y);
+            __m256d __r2 = _mm256_add_pd(__x2, __y2);
+            __m256d __2xy = _mm256_mul_pd(__two, _mm256_mul_pd(__x, __y));
+            __m256d __kr = _mm256_div_pd(
+#if CV_FMA3
+                _mm256_fmadd_pd(_mm256_fmadd_pd(_mm256_fmadd_pd(_mm256_set1_pd(k3), __r2, _mm256_set1_pd(k2)), __r2, _mm256_set1_pd(k1)), __r2, __one),
+                _mm256_fmadd_pd(_mm256_fmadd_pd(_mm256_fmadd_pd(_mm256_set1_pd(k6), __r2, _mm256_set1_pd(k5)), __r2, _mm256_set1_pd(k4)), __r2, __one)
+#else
+                _mm256_add_pd(__one, _mm256_mul_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set1_pd(k3), __r2), _mm256_set1_pd(k2)), __r2), _mm256_set1_pd(k1)), __r2)),
+                _mm256_add_pd(__one, _mm256_mul_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_set1_pd(k6), __r2), _mm256_set1_pd(k5)), __r2), _mm256_set1_pd(k4)), __r2))
+#endif
+            );
+            __m256d __r22 = _mm256_mul_pd(__r2, __r2);
+#if CV_FMA3
+            __m256d __xd = _mm256_fmadd_pd(__x, __kr,
+                _mm256_add_pd(
+                    _mm256_fmadd_pd(_mm256_set1_pd(p1), __2xy, _mm256_mul_pd(_mm256_set1_pd(p2), _mm256_fmadd_pd(__two, __x2, __r2))),
+                    _mm256_fmadd_pd(_mm256_set1_pd(s1), __r2, _mm256_mul_pd(_mm256_set1_pd(s2), __r22))));
+            __m256d __yd = _mm256_fmadd_pd(__y, __kr,
+                _mm256_add_pd(
+                    _mm256_fmadd_pd(_mm256_set1_pd(p1), _mm256_fmadd_pd(__two, __y2, __r2), _mm256_mul_pd(_mm256_set1_pd(p2), __2xy)),
+                    _mm256_fmadd_pd(_mm256_set1_pd(s3), __r2, _mm256_mul_pd(_mm256_set1_pd(s4), __r22))));
+
+            __m256d __vecTilt2 = _mm256_fmadd_pd(__matTilt_20, __xd, _mm256_fmadd_pd(__matTilt_21, __yd, __matTilt_22));
+#else
+            __m256d __xd = _mm256_add_pd(
+                _mm256_mul_pd(__x, __kr),
+                _mm256_add_pd(
+                    _mm256_add_pd(
+                        _mm256_mul_pd(_mm256_set1_pd(p1), __2xy),
+                        _mm256_mul_pd(_mm256_set1_pd(p2), _mm256_add_pd(__r2, _mm256_mul_pd(__two, __x2)))),
+                    _mm256_add_pd(
+                        _mm256_mul_pd(_mm256_set1_pd(s1), __r2),
+                        _mm256_mul_pd(_mm256_set1_pd(s2), __r22))));
+            __m256d __yd = _mm256_add_pd(
+                _mm256_mul_pd(__y, __kr),
+                _mm256_add_pd(
+                    _mm256_add_pd(
+                        _mm256_mul_pd(_mm256_set1_pd(p1), _mm256_add_pd(__r2, _mm256_mul_pd(__two, __y2))),
+                        _mm256_mul_pd(_mm256_set1_pd(p2), __2xy)),
+                    _mm256_add_pd(
+                        _mm256_mul_pd(_mm256_set1_pd(s3), __r2),
+                        _mm256_mul_pd(_mm256_set1_pd(s4), __r22))));
+
+            __m256d __vecTilt2 = _mm256_add_pd(_mm256_add_pd(
+                _mm256_mul_pd(__matTilt_20, __xd), _mm256_mul_pd(__matTilt_21, __yd)), __matTilt_22);
+#endif
+            __m256d __invProj = _mm256_blendv_pd(
+                __one, _mm256_div_pd(__one, __vecTilt2),
+                _mm256_cmp_pd(__vecTilt2, _mm256_setzero_pd(), _CMP_EQ_OQ));
+
+#if CV_FMA3
+            __m256d __u = _mm256_fmadd_pd(__matTilt_00, __xd, _mm256_fmadd_pd(__matTilt_01, __yd, __matTilt_02));
+            __u = _mm256_fmadd_pd(_mm256_mul_pd(_mm256_set1_pd(fx), __invProj), __u, _mm256_set1_pd(u0));
+
+            __m256d __v = _mm256_fmadd_pd(__matTilt_10, __xd, _mm256_fmadd_pd(__matTilt_11, __yd, __matTilt_12));
+            __v = _mm256_fmadd_pd(_mm256_mul_pd(_mm256_set1_pd(fy), __invProj), __v, _mm256_set1_pd(v0));
+#else
+            __m256d __u = _mm256_add_pd(_mm256_add_pd(
+                _mm256_mul_pd(__matTilt_00, __xd), _mm256_mul_pd(__matTilt_01, __yd)), __matTilt_02);
+            __u = _mm256_add_pd(_mm256_mul_pd(_mm256_mul_pd(_mm256_set1_pd(fx), __invProj), __u), _mm256_set1_pd(u0));
+
+            __m256d __v = _mm256_add_pd(_mm256_add_pd(
+                _mm256_mul_pd(__matTilt_10, __xd), _mm256_mul_pd(__matTilt_11, __yd)), __matTilt_12);
+            __v = _mm256_add_pd(_mm256_mul_pd(_mm256_mul_pd(_mm256_set1_pd(fy), __invProj), __v), _mm256_set1_pd(v0));
+#endif
+
+            if ( m1type == CV_32FC1 )
+            {
+                _mm_storeu_ps(&m1f[j], _mm256_cvtpd_ps(__u));
+                _mm_storeu_ps(&m2f[j], _mm256_cvtpd_ps(__v));
+            }
+            else if ( m1type == CV_32FC2 )
+            {
+                __m128 __u_float = _mm256_cvtpd_ps(__u);
+                __m128 __v_float = _mm256_cvtpd_ps(__v);
+
+                _mm_storeu_ps(&m1f[j*2], _mm_unpacklo_ps(__u_float, __v_float));
+                _mm_storeu_ps(&m1f[j*2 + 4], _mm_unpackhi_ps(__u_float, __v_float));
+            }
+            else // m1type == CV_16SC2
+            {
+                __u = _mm256_mul_pd(__u, _mm256_set1_pd(INTER_TAB_SIZE));
+                __v = _mm256_mul_pd(__v, _mm256_set1_pd(INTER_TAB_SIZE));
+
+                __m128 __u_float = _mm256_cvtpd_ps(__u);
+                __m128 __v_float = _mm256_cvtpd_ps(__v);
+                _mm256_zeroupper();
+                static const __m128 __int_max = _mm_set1_ps(std::numeric_limits<int>::max());
+                static const __m128 __int_min = _mm_set1_ps(std::numeric_limits<int>::min());
+                __u_float = _mm_max_ps(_mm_min_ps(__u_float, __int_max), __int_min);
+                __v_float = _mm_max_ps(_mm_min_ps(__v_float, __int_max), __int_min);
+
+                __m128i __iu = _mm_cvtps_epi32(__u_float);
+                __m128i __iv = _mm_cvtps_epi32(__v_float);
+
+                static const __m128i __INTER_TAB_SIZE_m1 = _mm_set1_epi32(INTER_TAB_SIZE-1);
+                __m128i __m2 = _mm_add_epi32(
+                    _mm_mul_epi32(_mm_and_si128(__iv, __INTER_TAB_SIZE_m1), _mm_set1_epi32(INTER_TAB_SIZE)),
+                    _mm_and_si128(__iu, __INTER_TAB_SIZE_m1));
+                __m2 = _mm_packus_epi16(__m2, __m2);
+                _mm_maskstore_epi64((long long int*) &m2[j], _mm_set_epi32(0, 0, 0xFFFFFFFF, 0xFFFFFFFF), __m2);
+
+                // gcc4.9 does not support _mm256_set_m128
+                // __m256i __m1 = _mm256_set_m128i(__iv, __iu);
+                __m256i __m1;
+                __m1 = _mm256_inserti128_si256(__m1, __iu, 0);
+                __m1 = _mm256_inserti128_si256(__m1, __iv, 1);
+                __m1 = _mm256_srli_epi32(__m1, INTER_BITS); // v3 v2 v1 v0 u3 u2 u1 u0 (int32_t)
+                static const __m256i __permute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1 ,4, 0);
+                __m1 = _mm256_permutevar8x32_epi32(__m1, __permute_mask); // v3 u3 v2 u2 v1 u1 v0 u0 (int32_t)
+                __m1 = _mm256_packs_epi32(__m1, __m1); // x x x x v3 u3 v2 u2 x x x x v1 u1 v0 u0 (int16_t)
+                _mm_storeu_si128((__m128i*) &m1[j*2], _mm256_extracti128_si256(_mm256_permute4x64_epi64(__m1, (2 << 2) + 0), 0));
+            }
+        }
+}
+#endif
+        for( ; j < size.width; j++, _x += ir[0], _y += ir[3], _w += ir[6] )
         {
             double w = 1./_w, x = _x*w, y = _y*w;
             double x2 = x*x, y2 = y*y;