[opencv] 18/251: SSE4_1 optimized implementation of resize and warp functions migrated to separate file

Nobuhiro Iwamatsu iwamatsu at moszumanska.debian.org
Sun Aug 27 23:27:19 UTC 2017


This is an automated email from the git hooks/post-receive script.

iwamatsu pushed a commit to annotated tag 3.3.0
in repository opencv.

commit fadf25acd6e255ed03303d9354b0e0bd93325e3c
Author: Vitaly Tuzov <terfendail at mediana.jetos.com>
Date:   Mon Jul 3 18:21:22 2017 +0300

    SSE4_1 optimized implementation of resize and warp functions migrated to separate file
---
 modules/imgproc/src/imgwarp.avx2.cpp   |  12 +-
 modules/imgproc/src/imgwarp.cpp        | 435 +++--------------------------
 modules/imgproc/src/imgwarp.hpp        |  18 +-
 modules/imgproc/src/imgwarp.sse4_1.cpp | 489 ++++++++++++++++++++++++++++++++-
 4 files changed, 548 insertions(+), 406 deletions(-)

diff --git a/modules/imgproc/src/imgwarp.avx2.cpp b/modules/imgproc/src/imgwarp.avx2.cpp
index 321957f..d434807 100644
--- a/modules/imgproc/src/imgwarp.avx2.cpp
+++ b/modules/imgproc/src/imgwarp.avx2.cpp
@@ -83,7 +83,9 @@ public:
                 uchar* Dstart = D;
                 int sy = std::min(cvFloor(y*ify), ssize.height-1);
                 const uchar* S = src.data + sy*src.step;
+#ifdef CV_ICC
 #pragma unroll(4)
+#endif
                 for(x = 0; x < avxWidth; x += 8)
                 {
                     const __m256i CV_DECL_ALIGNED(64) *addr = (__m256i*)(x_ofs + x);
@@ -106,7 +108,9 @@ public:
                 uchar* Dstart = D;
                 int sy = std::min(cvFloor(y*ify), ssize.height-1);
                 const uchar* S = src.data + sy*src.step;
+#ifdef CV_ICC
 #pragma unroll(4)
+#endif
                 for(x = 0; x < avxWidth; x += 8)
                 {
                     const __m256i CV_DECL_ALIGNED(64) *addr = (__m256i*)(x_ofs + x);
@@ -157,8 +161,8 @@ public:
         const __m256i CV_DECL_ALIGNED(64) shuffle_mask = _mm256_set_epi8(15,14,11,10,13,12,9,8,7,6,3,2,5,4,1,0,
                                                                          15,14,11,10,13,12,9,8,7,6,3,2,5,4,1,0);
         const __m256i CV_DECL_ALIGNED(64) permute_mask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
-        const __m256i CV_DECL_ALIGNED(64) shift_shuffle_mask = _mm256_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2,
-                                                                               13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2);
+        //const __m256i CV_DECL_ALIGNED(64) shift_shuffle_mask = _mm256_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2,
+        //                                                                       13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2);
         if(((int64)(dst.data + dst.step) & 0x1f) == 0)
         {
             for(y = range.start; y < range.end; y++)
@@ -168,7 +172,9 @@ public:
                 int sy = std::min(cvFloor(y*ify), ssize.height-1);
                 const uchar* S = src.data + sy*src.step;
                 const uchar* S2 = S - 2;
+#ifdef CV_ICC
 #pragma unroll(4)
+#endif
                 for(x = 0; x < avxWidth; x += 16)
                 {
                     const __m256i CV_DECL_ALIGNED(64) *addr = (__m256i*)(x_ofs + x);
@@ -200,7 +206,9 @@ public:
                 int sy = std::min(cvFloor(y*ify), ssize.height-1);
                 const uchar* S = src.data + sy*src.step;
                 const uchar* S2 = S - 2;
+#ifdef CV_ICC
 #pragma unroll(4)
+#endif
                 for(x = 0; x < avxWidth; x += 16)
                 {
                     const __m256i CV_DECL_ALIGNED(64) *addr = (__m256i*)(x_ofs + x);
diff --git a/modules/imgproc/src/imgwarp.cpp b/modules/imgproc/src/imgwarp.cpp
index 4f040a9..a960c01 100644
--- a/modules/imgproc/src/imgwarp.cpp
+++ b/modules/imgproc/src/imgwarp.cpp
@@ -450,9 +450,9 @@ resizeNN( const Mat& src, Mat& dst, double fx, double fy )
     if(CV_CPU_HAS_SUPPORT_SSE4_1 && ((pix_size == 2) || (pix_size == 4)))
     {
         if(pix_size == 2)
-            opt_SSE41::resizeNN2_SSE4_1(range, src, dst, x_ofs, pix_size4, ify);
+            opt_SSE4_1::resizeNN2_SSE4_1(range, src, dst, x_ofs, pix_size4, ify);
         else
-            opt_SSE41::resizeNN4_SSE4_1(range, src, dst, x_ofs, pix_size4, ify);
+            opt_SSE4_1::resizeNN4_SSE4_1(range, src, dst, x_ofs, pix_size4, ify);
     }
     else
 #endif
@@ -916,50 +916,14 @@ struct VResizeCubicVec_32f
     }
 };
 
-#if CV_SSE4_1
+#if CV_TRY_SSE4_1
 
 struct VResizeLanczos4Vec_32f16u
 {
     int operator()(const uchar** _src, uchar* _dst, const uchar* _beta, int width ) const
     {
-        const float** src = (const float**)_src;
-        const float* beta = (const float*)_beta;
-        const float *S0 = src[0], *S1 = src[1], *S2 = src[2], *S3 = src[3],
-                    *S4 = src[4], *S5 = src[5], *S6 = src[6], *S7 = src[7];
-        short * dst = (short*)_dst;
-        int x = 0;
-        __m128 v_b0 = _mm_set1_ps(beta[0]), v_b1 = _mm_set1_ps(beta[1]),
-               v_b2 = _mm_set1_ps(beta[2]), v_b3 = _mm_set1_ps(beta[3]),
-               v_b4 = _mm_set1_ps(beta[4]), v_b5 = _mm_set1_ps(beta[5]),
-               v_b6 = _mm_set1_ps(beta[6]), v_b7 = _mm_set1_ps(beta[7]);
-
-        for( ; x <= width - 8; x += 8 )
-        {
-            __m128 v_dst0 = _mm_mul_ps(v_b0, _mm_loadu_ps(S0 + x));
-            v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b1, _mm_loadu_ps(S1 + x)));
-            v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b2, _mm_loadu_ps(S2 + x)));
-            v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b3, _mm_loadu_ps(S3 + x)));
-            v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b4, _mm_loadu_ps(S4 + x)));
-            v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b5, _mm_loadu_ps(S5 + x)));
-            v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b6, _mm_loadu_ps(S6 + x)));
-            v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b7, _mm_loadu_ps(S7 + x)));
-
-            __m128 v_dst1 = _mm_mul_ps(v_b0, _mm_loadu_ps(S0 + x + 4));
-            v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b1, _mm_loadu_ps(S1 + x + 4)));
-            v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b2, _mm_loadu_ps(S2 + x + 4)));
-            v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b3, _mm_loadu_ps(S3 + x + 4)));
-            v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b4, _mm_loadu_ps(S4 + x + 4)));
-            v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b5, _mm_loadu_ps(S5 + x + 4)));
-            v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b6, _mm_loadu_ps(S6 + x + 4)));
-            v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b7, _mm_loadu_ps(S7 + x + 4)));
-
-            __m128i v_dsti0 = _mm_cvtps_epi32(v_dst0);
-            __m128i v_dsti1 = _mm_cvtps_epi32(v_dst1);
-
-            _mm_storeu_si128((__m128i *)(dst + x), _mm_packus_epi32(v_dsti0, v_dsti1));
-        }
-
-        return x;
+        if (CV_CPU_HAS_SUPPORT_SSE4_1) return opt_SSE4_1::VResizeLanczos4Vec_32f16u_SSE41(_src, _dst, _beta, width);
+        else return 0;
     }
 };
 
@@ -5149,8 +5113,8 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
 #if CV_SSE2
     bool useSSE2 = checkHardwareSupport(CV_CPU_SSE2);
 #endif
-#if CV_SSE4_1
-    bool useSSE4_1 = checkHardwareSupport(CV_CPU_SSE4_1);
+#if CV_TRY_SSE4_1
+    bool useSSE4_1 = CV_CPU_HAS_SUPPORT_SSE4_1;
 #endif
 
     const float scale = 1.f/INTER_TAB_SIZE;
@@ -5183,29 +5147,10 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
 
                     vst2q_s16(dst1 + (x << 1), v_dst);
                 }
-                #elif CV_SSE4_1
+                #elif CV_TRY_SSE4_1
                 if (useSSE4_1)
-                {
-                    for( ; x <= size.width - 16; x += 16 )
-                    {
-                        __m128i v_dst0 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src1f + x)),
-                                                         _mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 4)));
-                        __m128i v_dst1 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 8)),
-                                                         _mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 12)));
-
-                        __m128i v_dst2 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src2f + x)),
-                                                         _mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 4)));
-                        __m128i v_dst3 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 8)),
-                                                         _mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 12)));
-
-                        _mm_interleave_epi16(v_dst0, v_dst1, v_dst2, v_dst3);
-
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst0);
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 8), v_dst1);
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 16), v_dst2);
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 24), v_dst3);
-                    }
-                }
+                    opt_SSE4_1::convertMaps_nninterpolate32f1c16s_SSE41(src1f, src2f, dst1, size.width);
+                else
                 #endif
                 for( ; x < size.width; x++ )
                 {
@@ -5240,52 +5185,10 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
                                                               vandq_s32(v_ix1, v_mask)));
                     vst1q_u16(dst2 + x, vcombine_u16(v_dst0, v_dst1));
                 }
-                #elif CV_SSE4_1
+                #elif CV_TRY_SSE4_1
                 if (useSSE4_1)
-                {
-                    __m128 v_its = _mm_set1_ps(INTER_TAB_SIZE);
-                    __m128i v_its1 = _mm_set1_epi32(INTER_TAB_SIZE-1);
-
-                    for( ; x <= size.width - 16; x += 16 )
-                    {
-                        __m128i v_ix0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x), v_its));
-                        __m128i v_ix1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 4), v_its));
-                        __m128i v_iy0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x), v_its));
-                        __m128i v_iy1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 4), v_its));
-
-                        __m128i v_dst10 = _mm_packs_epi32(_mm_srai_epi32(v_ix0, INTER_BITS),
-                                                          _mm_srai_epi32(v_ix1, INTER_BITS));
-                        __m128i v_dst12 = _mm_packs_epi32(_mm_srai_epi32(v_iy0, INTER_BITS),
-                                                          _mm_srai_epi32(v_iy1, INTER_BITS));
-                        __m128i v_dst20 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy0, v_its1), INTER_BITS),
-                                                        _mm_and_si128(v_ix0, v_its1));
-                        __m128i v_dst21 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy1, v_its1), INTER_BITS),
-                                                        _mm_and_si128(v_ix1, v_its1));
-                        _mm_storeu_si128((__m128i *)(dst2 + x), _mm_packus_epi32(v_dst20, v_dst21));
-
-                        v_ix0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 8), v_its));
-                        v_ix1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 12), v_its));
-                        v_iy0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 8), v_its));
-                        v_iy1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 12), v_its));
-
-                        __m128i v_dst11 = _mm_packs_epi32(_mm_srai_epi32(v_ix0, INTER_BITS),
-                                                          _mm_srai_epi32(v_ix1, INTER_BITS));
-                        __m128i v_dst13 = _mm_packs_epi32(_mm_srai_epi32(v_iy0, INTER_BITS),
-                                                          _mm_srai_epi32(v_iy1, INTER_BITS));
-                        v_dst20 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy0, v_its1), INTER_BITS),
-                                                _mm_and_si128(v_ix0, v_its1));
-                        v_dst21 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy1, v_its1), INTER_BITS),
-                                                _mm_and_si128(v_ix1, v_its1));
-                        _mm_storeu_si128((__m128i *)(dst2 + x + 8), _mm_packus_epi32(v_dst20, v_dst21));
-
-                        _mm_interleave_epi16(v_dst10, v_dst11, v_dst12, v_dst13);
-
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst10);
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 8), v_dst11);
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 16), v_dst12);
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 24), v_dst13);
-                    }
-                }
+                    opt_SSE4_1::convertMaps_32f1c16s_SSE41(src1f, src2f, dst1, dst2, size.width);
+                else
                 #endif
                 for( ; x < size.width; x++ )
                 {
@@ -5346,30 +5249,10 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
                                                               vandq_s32(v_ix1, v_mask)));
                     vst1q_u16(dst2 + x, vcombine_u16(v_dst0, v_dst1));
                 }
-                #elif CV_SSE4_1
+                #elif CV_TRY_SSE4_1
                 if (useSSE4_1)
-                {
-                    __m128 v_its = _mm_set1_ps(INTER_TAB_SIZE);
-                    __m128i v_its1 = _mm_set1_epi32(INTER_TAB_SIZE-1);
-                    __m128i v_y_mask = _mm_set1_epi32((INTER_TAB_SIZE-1) << 16);
-
-                    for( ; x <= size.width - 4; x += 4 )
-                    {
-                        __m128i v_src0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x * 2), v_its));
-                        __m128i v_src1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x * 2 + 4), v_its));
-
-                        __m128i v_dst1 = _mm_packs_epi32(_mm_srai_epi32(v_src0, INTER_BITS),
-                                                         _mm_srai_epi32(v_src1, INTER_BITS));
-                        _mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst1);
-
-                        // x0 y0 x1 y1 . . .
-                        v_src0 = _mm_packs_epi32(_mm_and_si128(v_src0, v_its1),
-                                                 _mm_and_si128(v_src1, v_its1));
-                        __m128i v_dst2 = _mm_or_si128(_mm_srli_epi32(_mm_and_si128(v_src0, v_y_mask), 16 - INTER_BITS), // y0 0 y1 0 . . .
-                                                      _mm_and_si128(v_src0, v_its1)); // 0 x0 0 x1 . . .
-                        _mm_storel_epi64((__m128i *)(dst2 + x), _mm_packus_epi32(v_dst2, v_dst2));
-                    }
-                }
+                    opt_SSE4_1::convertMaps_32f2c16s_SSE41(src1f, dst1, dst2, size.width);
+                else
                 #endif
                 for( ; x < size.width; x++ )
                 {
@@ -5557,8 +5440,8 @@ public:
     #if CV_SSE2
         bool useSSE2 = checkHardwareSupport(CV_CPU_SSE2);
     #endif
-    #if CV_SSE4_1
-        bool useSSE4_1 = checkHardwareSupport(CV_CPU_SSE4_1);
+    #if CV_TRY_SSE4_1
+        bool useSSE4_1 = CV_CPU_HAS_SUPPORT_SSE4_1;
     #endif
 
         int bh0 = std::min(BLOCK_SZ/2, dst.rows);
@@ -5596,31 +5479,10 @@ public:
 
                             vst2q_s16(xy + (x1 << 1), v_dst);
                         }
-                        #elif CV_SSE4_1
+                        #elif CV_TRY_SSE4_1
                         if (useSSE4_1)
-                        {
-                            __m128i v_X0 = _mm_set1_epi32(X0);
-                            __m128i v_Y0 = _mm_set1_epi32(Y0);
-                            for ( ; x1 <= bw - 16; x1 += 16)
-                            {
-                                __m128i v_x0 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1))), AB_BITS),
-                                                               _mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 4))), AB_BITS));
-                                __m128i v_x1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 8))), AB_BITS),
-                                                               _mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 12))), AB_BITS));
-
-                                __m128i v_y0 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1))), AB_BITS),
-                                                               _mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 4))), AB_BITS));
-                                __m128i v_y1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 8))), AB_BITS),
-                                                               _mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 12))), AB_BITS));
-
-                                _mm_interleave_epi16(v_x0, v_x1, v_y0, v_y1);
-
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2), v_x0);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_x1);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_y0);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_y1);
-                            }
-                        }
+                            opt_SSE4_1::WarpAffineInvoker_Blockline_SSE41(adelta + x, bdelta + x, xy, X0, Y0, bw);
+                        else
                         #endif
                         for( ; x1 < bw; x1++ )
                         {
@@ -6132,18 +5994,10 @@ public:
         int bw0 = std::min(BLOCK_SZ*BLOCK_SZ/bh0, width);
         bh0 = std::min(BLOCK_SZ*BLOCK_SZ/bw0, height);
 
-        #if CV_SSE4_1
-        bool haveSSE4_1 = checkHardwareSupport(CV_CPU_SSE4_1);
-        __m128d v_M0 = _mm_set1_pd(M[0]);
-        __m128d v_M3 = _mm_set1_pd(M[3]);
-        __m128d v_M6 = _mm_set1_pd(M[6]);
-        __m128d v_intmax = _mm_set1_pd((double)INT_MAX);
-        __m128d v_intmin = _mm_set1_pd((double)INT_MIN);
-        __m128d v_2 = _mm_set1_pd(2),
-                v_zero = _mm_setzero_pd(),
-                v_1 = _mm_set1_pd(1),
-                v_its = _mm_set1_pd(INTER_TAB_SIZE);
-        __m128i v_itsi1 = _mm_set1_epi32(INTER_TAB_SIZE - 1);
+        #if CV_TRY_SSE4_1
+        Ptr<opt_SSE4_1::WarpPerspectiveLine_SSE4> pwarp_impl_sse4;
+        if(CV_CPU_HAS_SUPPORT_SSE4_1)
+            pwarp_impl_sse4 = opt_SSE4_1::WarpPerspectiveLine_SSE4::getImpl(M);
         #endif
 
         for( y = range.start; y < range.end; y += bh0 )
@@ -6167,116 +6021,11 @@ public:
                     {
                         x1 = 0;
 
-                        #if CV_SSE4_1
-                        if (haveSSE4_1)
-                        {
-                            __m128d v_X0d = _mm_set1_pd(X0);
-                            __m128d v_Y0d = _mm_set1_pd(Y0);
-                            __m128d v_W0 = _mm_set1_pd(W0);
-                            __m128d v_x1 = _mm_set_pd(1, 0);
-
-                            for( ; x1 <= bw - 16; x1 += 16 )
-                            {
-                                // 0-3
-                                __m128i v_X0, v_Y0;
-                                {
-                                    __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
-                                    __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
-                                    __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_X0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
-                                    v_Y0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
-                                }
-
-                                // 4-8
-                                __m128i v_X1, v_Y1;
-                                {
-                                    __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
-                                    __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
-                                    __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
-                                    v_Y1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
-                                }
-
-                                // 8-11
-                                __m128i v_X2, v_Y2;
-                                {
-                                    __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
-                                    __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
-                                    __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_X2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
-                                    v_Y2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
-                                }
-
-                                // 12-15
-                                __m128i v_X3, v_Y3;
-                                {
-                                    __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
-                                    __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
-                                    __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_X3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
-                                    v_Y3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
-                                }
-
-                                // convert to 16s
-                                v_X0 = _mm_packs_epi32(v_X0, v_X1);
-                                v_X1 = _mm_packs_epi32(v_X2, v_X3);
-                                v_Y0 = _mm_packs_epi32(v_Y0, v_Y1);
-                                v_Y1 = _mm_packs_epi32(v_Y2, v_Y3);
-
-                                _mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
-
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2), v_X0);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_Y0);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_Y1);
-                            }
-                        }
+                        #if CV_TRY_SSE4_1
+                        if (pwarp_impl_sse4)
+                            pwarp_impl_sse4->processNN(M, xy, X0, Y0, W0, bw);
+                        else
                         #endif
-
                         for( ; x1 < bw; x1++ )
                         {
                             double W = W0 + M[6]*x1;
@@ -6295,129 +6044,11 @@ public:
                         short* alpha = A + y1*bw;
                         x1 = 0;
 
-                        #if CV_SSE4_1
-                        if (haveSSE4_1)
-                        {
-                            __m128d v_X0d = _mm_set1_pd(X0);
-                            __m128d v_Y0d = _mm_set1_pd(Y0);
-                            __m128d v_W0 = _mm_set1_pd(W0);
-                            __m128d v_x1 = _mm_set_pd(1, 0);
-
-                            for( ; x1 <= bw - 16; x1 += 16 )
-                            {
-                                // 0-3
-                                __m128i v_X0, v_Y0;
-                                {
-                                    __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
-                                    __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
-                                    __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_X0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
-                                    v_Y0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
-                                }
-
-                                // 4-8
-                                __m128i v_X1, v_Y1;
-                                {
-                                    __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
-                                    __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
-                                    __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
-                                    v_Y1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
-                                }
-
-                                // 8-11
-                                __m128i v_X2, v_Y2;
-                                {
-                                    __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
-                                    __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
-                                    __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_X2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
-                                    v_Y2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
-                                }
-
-                                // 12-15
-                                __m128i v_X3, v_Y3;
-                                {
-                                    __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
-                                    __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
-                                    v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
-                                    __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
-                                    __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
-                                    v_x1 = _mm_add_pd(v_x1, v_2);
-
-                                    v_X3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
-                                    v_Y3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
-                                                                          _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
-                                }
-
-                                // store alpha
-                                __m128i v_alpha0 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y0, v_itsi1), INTER_BITS),
-                                                                 _mm_and_si128(v_X0, v_itsi1));
-                                __m128i v_alpha1 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y1, v_itsi1), INTER_BITS),
-                                                                 _mm_and_si128(v_X1, v_itsi1));
-                                _mm_storeu_si128((__m128i *)(alpha + x1), _mm_packs_epi32(v_alpha0, v_alpha1));
-
-                                v_alpha0 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y2, v_itsi1), INTER_BITS),
-                                                         _mm_and_si128(v_X2, v_itsi1));
-                                v_alpha1 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y3, v_itsi1), INTER_BITS),
-                                                         _mm_and_si128(v_X3, v_itsi1));
-                                _mm_storeu_si128((__m128i *)(alpha + x1 + 8), _mm_packs_epi32(v_alpha0, v_alpha1));
-
-                                // convert to 16s
-                                v_X0 = _mm_packs_epi32(_mm_srai_epi32(v_X0, INTER_BITS), _mm_srai_epi32(v_X1, INTER_BITS));
-                                v_X1 = _mm_packs_epi32(_mm_srai_epi32(v_X2, INTER_BITS), _mm_srai_epi32(v_X3, INTER_BITS));
-                                v_Y0 = _mm_packs_epi32(_mm_srai_epi32(v_Y0, INTER_BITS), _mm_srai_epi32(v_Y1, INTER_BITS));
-                                v_Y1 = _mm_packs_epi32(_mm_srai_epi32(v_Y2, INTER_BITS), _mm_srai_epi32(v_Y3, INTER_BITS));
-
-                                _mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
-
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2), v_X0);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_Y0);
-                                _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_Y1);
-                            }
-                        }
+                        #if CV_TRY_SSE4_1
+                        if (pwarp_impl_sse4)
+                            pwarp_impl_sse4->process(M, xy, alpha, X0, Y0, W0, bw);
+                        else
                         #endif
-
                         for( ; x1 < bw; x1++ )
                         {
                             double W = W0 + M[6]*x1;
diff --git a/modules/imgproc/src/imgwarp.hpp b/modules/imgproc/src/imgwarp.hpp
index 428bcc5..ed1146d 100644
--- a/modules/imgproc/src/imgwarp.hpp
+++ b/modules/imgproc/src/imgwarp.hpp
@@ -61,11 +61,27 @@ void resizeNN4_AVX2(const Range&, const Mat&, Mat&, int*, int, double);
 int warpAffineBlockline(int *adelta, int *bdelta, short* xy, short* alpha, int X0, int Y0, int bw);
 #endif
 }
-namespace opt_SSE41
+
+namespace opt_SSE4_1
 {
 #if CV_TRY_SSE4_1
 void resizeNN2_SSE4_1(const Range&, const Mat&, Mat&, int*, int, double);
 void resizeNN4_SSE4_1(const Range&, const Mat&, Mat&, int*, int, double);
+
+int VResizeLanczos4Vec_32f16u_SSE41(const uchar** _src, uchar* _dst, const uchar* _beta, int width);
+void convertMaps_nninterpolate32f1c16s_SSE41(const float* src1f, const float* src2f, short* dst1, int width);
+void convertMaps_32f1c16s_SSE41(const float* src1f, const float* src2f, short* dst1, ushort* dst2, int width);
+void convertMaps_32f2c16s_SSE41(const float* src1f, short* dst1, ushort* dst2, int width);
+void WarpAffineInvoker_Blockline_SSE41(int *adelta, int *bdelta, short* xy, int X0, int Y0, int bw);
+
+class WarpPerspectiveLine_SSE4
+{
+public:
+    static Ptr<WarpPerspectiveLine_SSE4> getImpl(const double *M);
+    virtual void processNN(const double *M, short* xy, double X0, double Y0, double W0, int bw) = 0;
+    virtual void process(const double *M, short* xy, short* alpha, double X0, double Y0, double W0, int bw) = 0;
+    virtual ~WarpPerspectiveLine_SSE4() {};
+};
 #endif
 }
 }
diff --git a/modules/imgproc/src/imgwarp.sse4_1.cpp b/modules/imgproc/src/imgwarp.sse4_1.cpp
index b326ffc..79137d1 100644
--- a/modules/imgproc/src/imgwarp.sse4_1.cpp
+++ b/modules/imgproc/src/imgwarp.sse4_1.cpp
@@ -52,7 +52,7 @@
 
 namespace cv
 {
-namespace opt_SSE41
+namespace opt_SSE4_1
 {
 
 class resizeNNInvokerSSE2 :
@@ -186,7 +186,494 @@ void resizeNN4_SSE4_1(const Range& range, const Mat& src, Mat &dst, int *x_ofs,
     parallel_for_(range, invoker, dst.total() / (double)(1 << 16));
 }
 
+int VResizeLanczos4Vec_32f16u_SSE41(const uchar** _src, uchar* _dst, const uchar* _beta, int width)
+{
+    const float** src = (const float**)_src;
+    const float* beta = (const float*)_beta;
+    const float *S0 = src[0], *S1 = src[1], *S2 = src[2], *S3 = src[3],
+        *S4 = src[4], *S5 = src[5], *S6 = src[6], *S7 = src[7];
+    short * dst = (short*)_dst;
+    int x = 0;
+    __m128 v_b0 = _mm_set1_ps(beta[0]), v_b1 = _mm_set1_ps(beta[1]),
+        v_b2 = _mm_set1_ps(beta[2]), v_b3 = _mm_set1_ps(beta[3]),
+        v_b4 = _mm_set1_ps(beta[4]), v_b5 = _mm_set1_ps(beta[5]),
+        v_b6 = _mm_set1_ps(beta[6]), v_b7 = _mm_set1_ps(beta[7]);
+
+    for (; x <= width - 8; x += 8)
+    {
+        __m128 v_dst0 = _mm_mul_ps(v_b0, _mm_loadu_ps(S0 + x));
+        v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b1, _mm_loadu_ps(S1 + x)));
+        v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b2, _mm_loadu_ps(S2 + x)));
+        v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b3, _mm_loadu_ps(S3 + x)));
+        v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b4, _mm_loadu_ps(S4 + x)));
+        v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b5, _mm_loadu_ps(S5 + x)));
+        v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b6, _mm_loadu_ps(S6 + x)));
+        v_dst0 = _mm_add_ps(v_dst0, _mm_mul_ps(v_b7, _mm_loadu_ps(S7 + x)));
+
+        __m128 v_dst1 = _mm_mul_ps(v_b0, _mm_loadu_ps(S0 + x + 4));
+        v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b1, _mm_loadu_ps(S1 + x + 4)));
+        v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b2, _mm_loadu_ps(S2 + x + 4)));
+        v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b3, _mm_loadu_ps(S3 + x + 4)));
+        v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b4, _mm_loadu_ps(S4 + x + 4)));
+        v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b5, _mm_loadu_ps(S5 + x + 4)));
+        v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b6, _mm_loadu_ps(S6 + x + 4)));
+        v_dst1 = _mm_add_ps(v_dst1, _mm_mul_ps(v_b7, _mm_loadu_ps(S7 + x + 4)));
+
+        __m128i v_dsti0 = _mm_cvtps_epi32(v_dst0);
+        __m128i v_dsti1 = _mm_cvtps_epi32(v_dst1);
+
+        _mm_storeu_si128((__m128i *)(dst + x), _mm_packus_epi32(v_dsti0, v_dsti1));
+    }
+
+    return x;
 }
+
+void convertMaps_nninterpolate32f1c16s_SSE41(const float* src1f, const float* src2f, short* dst1, int width)
+{
+    int x = 0;
+    for (; x <= width - 16; x += 16)
+    {
+        __m128i v_dst0 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src1f + x)),
+            _mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 4)));
+        __m128i v_dst1 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 8)),
+            _mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 12)));
+
+        __m128i v_dst2 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src2f + x)),
+            _mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 4)));
+        __m128i v_dst3 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 8)),
+            _mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 12)));
+
+        _mm_interleave_epi16(v_dst0, v_dst1, v_dst2, v_dst3);
+
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst0);
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 8), v_dst1);
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 16), v_dst2);
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 24), v_dst3);
+    }
+
+    for (; x < width; x++)
+    {
+        dst1[x * 2] = saturate_cast<short>(src1f[x]);
+        dst1[x * 2 + 1] = saturate_cast<short>(src2f[x]);
+    }
+}
+
+void convertMaps_32f1c16s_SSE41(const float* src1f, const float* src2f, short* dst1, ushort* dst2, int width)
+{
+    int x = 0;
+    __m128 v_its = _mm_set1_ps(INTER_TAB_SIZE);
+    __m128i v_its1 = _mm_set1_epi32(INTER_TAB_SIZE - 1);
+
+    for (; x <= width - 16; x += 16)
+    {
+        __m128i v_ix0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x), v_its));
+        __m128i v_ix1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 4), v_its));
+        __m128i v_iy0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x), v_its));
+        __m128i v_iy1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 4), v_its));
+
+        __m128i v_dst10 = _mm_packs_epi32(_mm_srai_epi32(v_ix0, INTER_BITS),
+            _mm_srai_epi32(v_ix1, INTER_BITS));
+        __m128i v_dst12 = _mm_packs_epi32(_mm_srai_epi32(v_iy0, INTER_BITS),
+            _mm_srai_epi32(v_iy1, INTER_BITS));
+        __m128i v_dst20 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy0, v_its1), INTER_BITS),
+            _mm_and_si128(v_ix0, v_its1));
+        __m128i v_dst21 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy1, v_its1), INTER_BITS),
+            _mm_and_si128(v_ix1, v_its1));
+        _mm_storeu_si128((__m128i *)(dst2 + x), _mm_packus_epi32(v_dst20, v_dst21));
+
+        v_ix0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 8), v_its));
+        v_ix1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 12), v_its));
+        v_iy0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 8), v_its));
+        v_iy1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 12), v_its));
+
+        __m128i v_dst11 = _mm_packs_epi32(_mm_srai_epi32(v_ix0, INTER_BITS),
+            _mm_srai_epi32(v_ix1, INTER_BITS));
+        __m128i v_dst13 = _mm_packs_epi32(_mm_srai_epi32(v_iy0, INTER_BITS),
+            _mm_srai_epi32(v_iy1, INTER_BITS));
+        v_dst20 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy0, v_its1), INTER_BITS),
+            _mm_and_si128(v_ix0, v_its1));
+        v_dst21 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy1, v_its1), INTER_BITS),
+            _mm_and_si128(v_ix1, v_its1));
+        _mm_storeu_si128((__m128i *)(dst2 + x + 8), _mm_packus_epi32(v_dst20, v_dst21));
+
+        _mm_interleave_epi16(v_dst10, v_dst11, v_dst12, v_dst13);
+
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst10);
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 8), v_dst11);
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 16), v_dst12);
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2 + 24), v_dst13);
+    }
+    for (; x < width; x++)
+    {
+        int ix = saturate_cast<int>(src1f[x] * INTER_TAB_SIZE);
+        int iy = saturate_cast<int>(src2f[x] * INTER_TAB_SIZE);
+        dst1[x * 2] = saturate_cast<short>(ix >> INTER_BITS);
+        dst1[x * 2 + 1] = saturate_cast<short>(iy >> INTER_BITS);
+        dst2[x] = (ushort)((iy & (INTER_TAB_SIZE - 1))*INTER_TAB_SIZE + (ix & (INTER_TAB_SIZE - 1)));
+    }
 }
 
+void convertMaps_32f2c16s_SSE41(const float* src1f, short* dst1, ushort* dst2, int width)
+{
+    int x = 0;
+    __m128 v_its = _mm_set1_ps(INTER_TAB_SIZE);
+    __m128i v_its1 = _mm_set1_epi32(INTER_TAB_SIZE - 1);
+    __m128i v_y_mask = _mm_set1_epi32((INTER_TAB_SIZE - 1) << 16);
+
+    for (; x <= width - 4; x += 4)
+    {
+        __m128i v_src0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x * 2), v_its));
+        __m128i v_src1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x * 2 + 4), v_its));
+
+        __m128i v_dst1 = _mm_packs_epi32(_mm_srai_epi32(v_src0, INTER_BITS),
+            _mm_srai_epi32(v_src1, INTER_BITS));
+        _mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst1);
+
+        // x0 y0 x1 y1 . . .
+        v_src0 = _mm_packs_epi32(_mm_and_si128(v_src0, v_its1),
+            _mm_and_si128(v_src1, v_its1));
+        __m128i v_dst2 = _mm_or_si128(_mm_srli_epi32(_mm_and_si128(v_src0, v_y_mask), 16 - INTER_BITS), // y0 0 y1 0 . . .
+            _mm_and_si128(v_src0, v_its1)); // 0 x0 0 x1 . . .
+        _mm_storel_epi64((__m128i *)(dst2 + x), _mm_packus_epi32(v_dst2, v_dst2));
+    }
+    for (; x < width; x++)
+    {
+        int ix = saturate_cast<int>(src1f[x * 2] * INTER_TAB_SIZE);
+        int iy = saturate_cast<int>(src1f[x * 2 + 1] * INTER_TAB_SIZE);
+        dst1[x * 2] = saturate_cast<short>(ix >> INTER_BITS);
+        dst1[x * 2 + 1] = saturate_cast<short>(iy >> INTER_BITS);
+        dst2[x] = (ushort)((iy & (INTER_TAB_SIZE - 1))*INTER_TAB_SIZE + (ix & (INTER_TAB_SIZE - 1)));
+    }
+}
+
+void WarpAffineInvoker_Blockline_SSE41(int *adelta, int *bdelta, short* xy, int X0, int Y0, int bw)
+{
+    const int AB_BITS = MAX(10, (int)INTER_BITS);
+    int x1 = 0;
+
+    __m128i v_X0 = _mm_set1_epi32(X0);
+    __m128i v_Y0 = _mm_set1_epi32(Y0);
+    for (; x1 <= bw - 16; x1 += 16)
+    {
+        __m128i v_x0 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x1))), AB_BITS),
+            _mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x1 + 4))), AB_BITS));
+        __m128i v_x1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x1 + 8))), AB_BITS),
+            _mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x1 + 12))), AB_BITS));
+
+        __m128i v_y0 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x1))), AB_BITS),
+            _mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x1 + 4))), AB_BITS));
+        __m128i v_y1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x1 + 8))), AB_BITS),
+            _mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x1 + 12))), AB_BITS));
+
+        _mm_interleave_epi16(v_x0, v_x1, v_y0, v_y1);
+
+        _mm_storeu_si128((__m128i *)(xy + x1 * 2), v_x0);
+        _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_x1);
+        _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_y0);
+        _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_y1);
+    }
+    for (; x1 < bw; x1++)
+    {
+        int X = (X0 + adelta[x1]) >> AB_BITS;
+        int Y = (Y0 + bdelta[x1]) >> AB_BITS;
+        xy[x1 * 2] = saturate_cast<short>(X);
+        xy[x1 * 2 + 1] = saturate_cast<short>(Y);
+    }
+}
+
+
+class WarpPerspectiveLine_SSE4_Impl: public WarpPerspectiveLine_SSE4
+{
+public:
+    WarpPerspectiveLine_SSE4_Impl(const double *M)
+    {
+        v_M0 = _mm_set1_pd(M[0]);
+        v_M3 = _mm_set1_pd(M[3]);
+        v_M6 = _mm_set1_pd(M[6]);
+        v_intmax = _mm_set1_pd((double)INT_MAX);
+        v_intmin = _mm_set1_pd((double)INT_MIN);
+        v_2 = _mm_set1_pd(2);
+        v_zero = _mm_setzero_pd();
+        v_1 = _mm_set1_pd(1);
+        v_its = _mm_set1_pd(INTER_TAB_SIZE);
+        v_itsi1 = _mm_set1_epi32(INTER_TAB_SIZE - 1);
+    }
+    virtual void processNN(const double *M, short* xy, double X0, double Y0, double W0, int bw)
+    {
+        int x1 = 0;
+        __m128d v_X0d = _mm_set1_pd(X0);
+        __m128d v_Y0d = _mm_set1_pd(Y0);
+        __m128d v_W0 = _mm_set1_pd(W0);
+        __m128d v_x1 = _mm_set_pd(1, 0);
+
+        for (; x1 <= bw - 16; x1 += 16)
+        {
+            // 0-3
+            __m128i v_X0, v_Y0;
+            {
+                __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
+                __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
+                __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_X0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
+                v_Y0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
+            }
+
+            // 4-8
+            __m128i v_X1, v_Y1;
+            {
+                __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
+                __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
+                __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
+                v_Y1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
+            }
+
+            // 8-11
+            __m128i v_X2, v_Y2;
+            {
+                __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
+                __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
+                __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_X2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
+                v_Y2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
+            }
+
+            // 12-15
+            __m128i v_X3, v_Y3;
+            {
+                __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
+                __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
+                __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_X3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
+                v_Y3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
+            }
+
+            // convert to 16s
+            v_X0 = _mm_packs_epi32(v_X0, v_X1);
+            v_X1 = _mm_packs_epi32(v_X2, v_X3);
+            v_Y0 = _mm_packs_epi32(v_Y0, v_Y1);
+            v_Y1 = _mm_packs_epi32(v_Y2, v_Y3);
+
+            _mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
+
+            _mm_storeu_si128((__m128i *)(xy + x1 * 2), v_X0);
+            _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
+            _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_Y0);
+            _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_Y1);
+        }
+
+        for (; x1 < bw; x1++)
+        {
+            double W = W0 + M[6] * x1;
+            W = W ? 1. / W : 0;
+            double fX = std::max((double)INT_MIN, std::min((double)INT_MAX, (X0 + M[0] * x1)*W));
+            double fY = std::max((double)INT_MIN, std::min((double)INT_MAX, (Y0 + M[3] * x1)*W));
+            int X = saturate_cast<int>(fX);
+            int Y = saturate_cast<int>(fY);
+
+            xy[x1 * 2] = saturate_cast<short>(X);
+            xy[x1 * 2 + 1] = saturate_cast<short>(Y);
+        }
+    }
+    virtual void process(const double *M, short* xy, short* alpha, double X0, double Y0, double W0, int bw)
+    {
+        int x1 = 0;
+
+        __m128d v_X0d = _mm_set1_pd(X0);
+        __m128d v_Y0d = _mm_set1_pd(Y0);
+        __m128d v_W0 = _mm_set1_pd(W0);
+        __m128d v_x1 = _mm_set_pd(1, 0);
+
+        for (; x1 <= bw - 16; x1 += 16)
+        {
+            // 0-3
+            __m128i v_X0, v_Y0;
+            {
+                __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
+                __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
+                __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_X0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
+                v_Y0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
+            }
+
+            // 4-8
+            __m128i v_X1, v_Y1;
+            {
+                __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
+                __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
+                __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
+                v_Y1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
+            }
+
+            // 8-11
+            __m128i v_X2, v_Y2;
+            {
+                __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
+                __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
+                __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_X2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
+                v_Y2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
+            }
+
+            // 12-15
+            __m128i v_X3, v_Y3;
+            {
+                __m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
+                __m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
+                v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
+                __m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
+                __m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
+                v_x1 = _mm_add_pd(v_x1, v_2);
+
+                v_X3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
+                v_Y3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
+                    _mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
+            }
+
+            // store alpha
+            __m128i v_alpha0 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y0, v_itsi1), INTER_BITS),
+                _mm_and_si128(v_X0, v_itsi1));
+            __m128i v_alpha1 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y1, v_itsi1), INTER_BITS),
+                _mm_and_si128(v_X1, v_itsi1));
+            _mm_storeu_si128((__m128i *)(alpha + x1), _mm_packs_epi32(v_alpha0, v_alpha1));
+
+            v_alpha0 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y2, v_itsi1), INTER_BITS),
+                _mm_and_si128(v_X2, v_itsi1));
+            v_alpha1 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y3, v_itsi1), INTER_BITS),
+                _mm_and_si128(v_X3, v_itsi1));
+            _mm_storeu_si128((__m128i *)(alpha + x1 + 8), _mm_packs_epi32(v_alpha0, v_alpha1));
+
+            // convert to 16s
+            v_X0 = _mm_packs_epi32(_mm_srai_epi32(v_X0, INTER_BITS), _mm_srai_epi32(v_X1, INTER_BITS));
+            v_X1 = _mm_packs_epi32(_mm_srai_epi32(v_X2, INTER_BITS), _mm_srai_epi32(v_X3, INTER_BITS));
+            v_Y0 = _mm_packs_epi32(_mm_srai_epi32(v_Y0, INTER_BITS), _mm_srai_epi32(v_Y1, INTER_BITS));
+            v_Y1 = _mm_packs_epi32(_mm_srai_epi32(v_Y2, INTER_BITS), _mm_srai_epi32(v_Y3, INTER_BITS));
+
+            _mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
+
+            _mm_storeu_si128((__m128i *)(xy + x1 * 2), v_X0);
+            _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
+            _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_Y0);
+            _mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_Y1);
+        }
+        for (; x1 < bw; x1++)
+        {
+            double W = W0 + M[6] * x1;
+            W = W ? INTER_TAB_SIZE / W : 0;
+            double fX = std::max((double)INT_MIN, std::min((double)INT_MAX, (X0 + M[0] * x1)*W));
+            double fY = std::max((double)INT_MIN, std::min((double)INT_MAX, (Y0 + M[3] * x1)*W));
+            int X = saturate_cast<int>(fX);
+            int Y = saturate_cast<int>(fY);
+
+            xy[x1 * 2] = saturate_cast<short>(X >> INTER_BITS);
+            xy[x1 * 2 + 1] = saturate_cast<short>(Y >> INTER_BITS);
+            alpha[x1] = (short)((Y & (INTER_TAB_SIZE - 1))*INTER_TAB_SIZE +
+                (X & (INTER_TAB_SIZE - 1)));
+        }
+    }
+    virtual ~WarpPerspectiveLine_SSE4_Impl() {};
+private:
+    __m128d v_M0;
+    __m128d v_M3;
+    __m128d v_M6;
+    __m128d v_intmax;
+    __m128d v_intmin;
+    __m128d v_2,
+            v_zero,
+            v_1,
+            v_its;
+    __m128i v_itsi1;
+};
+
+Ptr<WarpPerspectiveLine_SSE4> WarpPerspectiveLine_SSE4::getImpl(const double *M)
+{
+    return Ptr<WarpPerspectiveLine_SSE4>(new WarpPerspectiveLine_SSE4_Impl(M));
+}
+
+}
+}
 /* End of file. */

-- 
Alioth's /usr/local/bin/git-commit-notice on /srv/git.debian.org/git/debian-science/packages/opencv.git



More information about the debian-science-commits mailing list