SSE mul

diff --git a/modules/core/src/arithm.cpp b/modules/core/src/arithm.cpp
index 68c8979..d207463 100644 (file)
--- a/modules/core/src/arithm.cpp
+++ b/modules/core/src/arithm.cpp
@@ -2355,6 +2355,165 @@ struct Mul_SIMD<float, float>
     }
 };
 
+#elif CV_SSE2
+
+#if CV_SSE4_1
+
+template <>
+struct Mul_SIMD<ushort, float>
+{
+    Mul_SIMD()
+    {
+        haveSSE = checkHardwareSupport(CV_CPU_SSE4_1);
+    }
+
+    int operator() (const ushort * src1, const ushort * src2, ushort * dst, int width, float scale) const
+    {
+        int x = 0;
+
+        if (!haveSSE)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+
+        if( scale != 1.0f )
+        {
+            __m128 v_scale = _mm_set1_ps(scale);
+            for ( ; x <= width - 8; x += 8)
+            {
+                __m128i v_src1 = _mm_loadu_si128((__m128i const *)(src1 + x));
+                __m128i v_src2 = _mm_loadu_si128((__m128i const *)(src2 + x));
+
+                __m128 v_dst1 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(v_src1, v_zero)),
+                                           _mm_cvtepi32_ps(_mm_unpacklo_epi16(v_src2, v_zero)));
+                v_dst1 = _mm_mul_ps(v_dst1, v_scale);
+
+                __m128 v_dst2 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(v_src1, v_zero)),
+                                           _mm_cvtepi32_ps(_mm_unpackhi_epi16(v_src2, v_zero)));
+                v_dst2 = _mm_mul_ps(v_dst2, v_scale);
+
+                __m128i v_dsti = _mm_packus_epi32(_mm_cvtps_epi32(v_dst1), _mm_cvtps_epi32(v_dst2));
+                _mm_storeu_si128((__m128i *)(dst + x), v_dsti);
+            }
+        }
+
+        return x;
+    }
+
+    bool haveSSE;
+};
+
+#endif
+
+template <>
+struct Mul_SIMD<schar, float>
+{
+    Mul_SIMD()
+    {
+        haveSSE = checkHardwareSupport(CV_CPU_SSE2);
+    }
+
+    int operator() (const schar * src1, const schar * src2, schar * dst, int width, float scale) const
+    {
+        int x = 0;
+
+        if (!haveSSE)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+
+        if( scale == 1.0f )
+            for ( ; x <= width - 8; x += 8)
+            {
+                __m128i v_src1 = _mm_loadl_epi64((__m128i const *)(src1 + x));
+                __m128i v_src2 = _mm_loadl_epi64((__m128i const *)(src2 + x));
+
+                v_src1 = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, v_src1), 8);
+                v_src2 = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, v_src2), 8);
+
+                __m128 v_dst1 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src1), 16)),
+                                           _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src2), 16)));
+
+                __m128 v_dst2 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src1), 16)),
+                                           _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src2), 16)));
+
+                __m128i v_dsti = _mm_packs_epi32(_mm_cvtps_epi32(v_dst1), _mm_cvtps_epi32(v_dst2));
+                _mm_storel_epi64((__m128i *)(dst + x), _mm_packs_epi16(v_dsti, v_zero));
+            }
+        else
+        {
+            __m128 v_scale = _mm_set1_ps(scale);
+            for ( ; x <= width - 8; x += 8)
+            {
+                __m128i v_src1 = _mm_loadl_epi64((__m128i const *)(src1 + x));
+                __m128i v_src2 = _mm_loadl_epi64((__m128i const *)(src2 + x));
+
+                v_src1 = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, v_src1), 8);
+                v_src2 = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, v_src2), 8);
+
+                __m128 v_dst1 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src1), 16)),
+                                           _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src2), 16)));
+                v_dst1 = _mm_mul_ps(v_dst1, v_scale);
+
+                __m128 v_dst2 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src1), 16)),
+                                           _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src2), 16)));
+                v_dst2 = _mm_mul_ps(v_dst2, v_scale);
+
+                __m128i v_dsti = _mm_packs_epi32(_mm_cvtps_epi32(v_dst1), _mm_cvtps_epi32(v_dst2));
+                _mm_storel_epi64((__m128i *)(dst + x), _mm_packs_epi16(v_dsti, v_zero));
+            }
+        }
+
+        return x;
+    }
+
+    bool haveSSE;
+};
+
+template <>
+struct Mul_SIMD<short, float>
+{
+    Mul_SIMD()
+    {
+        haveSSE = checkHardwareSupport(CV_CPU_SSE2);
+    }
+
+    int operator() (const short * src1, const short * src2, short * dst, int width, float scale) const
+    {
+        int x = 0;
+
+        if (!haveSSE)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+
+        if( scale != 1.0f )
+        {
+            __m128 v_scale = _mm_set1_ps(scale);
+            for ( ; x <= width - 8; x += 8)
+            {
+                __m128i v_src1 = _mm_loadu_si128((__m128i const *)(src1 + x));
+                __m128i v_src2 = _mm_loadu_si128((__m128i const *)(src2 + x));
+
+                __m128 v_dst1 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src1), 16)),
+                                           _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src2), 16)));
+                v_dst1 = _mm_mul_ps(v_dst1, v_scale);
+
+                __m128 v_dst2 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src1), 16)),
+                                           _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src2), 16)));
+                v_dst2 = _mm_mul_ps(v_dst2, v_scale);
+
+                __m128i v_dsti = _mm_packs_epi32(_mm_cvtps_epi32(v_dst1), _mm_cvtps_epi32(v_dst2));
+                _mm_storeu_si128((__m128i *)(dst + x), v_dsti);
+            }
+        }
+
+        return x;
+    }
+
+    bool haveSSE;
+};
+
 #endif
 
 template<typename T, typename WT> static void