SSE resize Lanczos

diff --git a/modules/imgproc/src/imgwarp.cpp b/modules/imgproc/src/imgwarp.cpp
index d952b60..8a581fc 100644 (file)
--- a/modules/imgproc/src/imgwarp.cpp
+++ b/modules/imgproc/src/imgwarp.cpp
@@ -894,9 +894,139 @@ struct VResizeCubicVec_32f
     }
 };
 
+#if CV_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;
+    }
+};
+
+#else
+
 typedef VResizeNoVec VResizeLanczos4Vec_32f16u;
-typedef VResizeNoVec VResizeLanczos4Vec_32f16s;
-typedef VResizeNoVec VResizeLanczos4Vec_32f;
+
+#endif
+
+struct VResizeLanczos4Vec_32f16s
+{
+    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_packs_epi32(v_dsti0, v_dsti1));
+        }
+
+        return x;
+    }
+};
+
+
+struct VResizeLanczos4Vec_32f
+{
+    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];
+        float* dst = (float*)_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 - 4; x += 4 )
+        {
+            __m128 v_dst = _mm_mul_ps(v_b0, _mm_loadu_ps(S0 + x));
+            v_dst = _mm_add_ps(v_dst, _mm_mul_ps(v_b1, _mm_loadu_ps(S1 + x)));
+            v_dst = _mm_add_ps(v_dst, _mm_mul_ps(v_b2, _mm_loadu_ps(S2 + x)));
+            v_dst = _mm_add_ps(v_dst, _mm_mul_ps(v_b3, _mm_loadu_ps(S3 + x)));
+            v_dst = _mm_add_ps(v_dst, _mm_mul_ps(v_b4, _mm_loadu_ps(S4 + x)));
+            v_dst = _mm_add_ps(v_dst, _mm_mul_ps(v_b5, _mm_loadu_ps(S5 + x)));
+            v_dst = _mm_add_ps(v_dst, _mm_mul_ps(v_b6, _mm_loadu_ps(S6 + x)));
+            v_dst = _mm_add_ps(v_dst, _mm_mul_ps(v_b7, _mm_loadu_ps(S7 + x)));
+
+            _mm_storeu_ps(dst + x, v_dst);
+        }
+
+        return x;
+    }
+};
+
 
 #elif CV_NEON