convertTo from 8s

diff --git a/modules/core/src/convert.cpp b/modules/core/src/convert.cpp
index a92ff47..5dae935 100644 (file)
--- a/modules/core/src/convert.cpp
+++ b/modules/core/src/convert.cpp
@@ -1571,6 +1571,8 @@ struct cvtScale_SIMD
 
 #if CV_SSE2
 
+// from uchar
+
 template <>
 struct cvtScale_SIMD<uchar, uchar, float>
 {
@@ -1801,6 +1803,238 @@ struct cvtScale_SIMD<uchar, double, float>
     }
 };
 
+// from schar
+
+template <>
+struct cvtScale_SIMD<schar, uchar, float>
+{
+    int operator () (const schar * src, uchar * dst, int width, float scale, float shift) const
+    {
+        int x = 0;
+
+        if (!USE_SSE2)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+        __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift);
+
+        for ( ; x <= width - 8; x += 8)
+        {
+            __m128i v_src = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _mm_loadl_epi64((__m128i const *)(src + x))), 8);
+            __m128 v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src), 16));
+            __m128 v_dst_0 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src), 16));
+            __m128 v_dst_1 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            __m128i v_dst = _mm_packs_epi32(_mm_cvtps_epi32(v_dst_0),
+                                            _mm_cvtps_epi32(v_dst_1));
+            _mm_storel_epi64((__m128i *)(dst + x), _mm_packus_epi16(v_dst, v_zero));
+        }
+
+        return x;
+    }
+};
+
+template <>
+struct cvtScale_SIMD<schar, schar, float>
+{
+    int operator () (const schar * src, schar * dst, int width, float scale, float shift) const
+    {
+        int x = 0;
+
+        if (!USE_SSE2)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+        __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift);
+
+        for ( ; x <= width - 8; x += 8)
+        {
+            __m128i v_src = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _mm_loadl_epi64((__m128i const *)(src + x))), 8);
+            __m128 v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src), 16));
+            __m128 v_dst_0 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src), 16));
+            __m128 v_dst_1 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            __m128i v_dst = _mm_packs_epi32(_mm_cvtps_epi32(v_dst_0),
+                                            _mm_cvtps_epi32(v_dst_1));
+            _mm_storel_epi64((__m128i *)(dst + x), _mm_packs_epi16(v_dst, v_zero));
+        }
+
+        return x;
+    }
+};
+
+#if CV_SSE4_1
+
+template <>
+struct cvtScale_SIMD<schar, ushort, float>
+{
+    cvtScale_SIMD()
+    {
+        haveSSE = checkHardwareSupport(CV_CPU_SSE4_1);
+    }
+
+    int operator () (const schar * src, ushort * dst, int width, float scale, float shift) const
+    {
+        int x = 0;
+
+        if (!haveSSE)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+        __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift);
+
+        for ( ; x <= width - 8; x += 8)
+        {
+            __m128i v_src = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _mm_loadl_epi64((__m128i const *)(src + x))), 8);
+            __m128 v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src), 16));
+            __m128 v_dst_0 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src), 16));
+            __m128 v_dst_1 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            __m128i v_dst = _mm_packus_epi32(_mm_cvtps_epi32(v_dst_0),
+                                             _mm_cvtps_epi32(v_dst_1));
+            _mm_storeu_si128((__m128i *)(dst + x), v_dst);
+        }
+
+        return x;
+    }
+
+    bool haveSSE;
+};
+
+#endif
+
+template <>
+struct cvtScale_SIMD<schar, short, float>
+{
+    int operator () (const schar * src, short * dst, int width, float scale, float shift) const
+    {
+        int x = 0;
+
+        if (!USE_SSE2)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+        __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift);
+
+        for ( ; x <= width - 8; x += 8)
+        {
+            __m128i v_src = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _mm_loadl_epi64((__m128i const *)(src + x))), 8);
+            __m128 v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src), 16));
+            __m128 v_dst_0 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src), 16));
+            __m128 v_dst_1 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            __m128i v_dst = _mm_packs_epi32(_mm_cvtps_epi32(v_dst_0),
+                                            _mm_cvtps_epi32(v_dst_1));
+            _mm_storeu_si128((__m128i *)(dst + x), v_dst);
+        }
+
+        return x;
+    }
+};
+
+template <>
+struct cvtScale_SIMD<schar, int, float>
+{
+    int operator () (const schar * src, int * dst, int width, float scale, float shift) const
+    {
+        int x = 0;
+
+        if (!USE_SSE2)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+        __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift);
+
+        for ( ; x <= width - 8; x += 8)
+        {
+            __m128i v_src = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _mm_loadl_epi64((__m128i const *)(src + x))), 8);
+            __m128 v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src), 16));
+            __m128 v_dst_0 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src), 16));
+            __m128 v_dst_1 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            _mm_storeu_si128((__m128i *)(dst + x), _mm_cvtps_epi32(v_dst_0));
+            _mm_storeu_si128((__m128i *)(dst + x + 4), _mm_cvtps_epi32(v_dst_1));
+        }
+
+        return x;
+    }
+};
+
+template <>
+struct cvtScale_SIMD<schar, float, float>
+{
+    int operator () (const schar * src, float * dst, int width, float scale, float shift) const
+    {
+        int x = 0;
+
+        if (!USE_SSE2)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+        __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift);
+
+        for ( ; x <= width - 8; x += 8)
+        {
+            __m128i v_src = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _mm_loadl_epi64((__m128i const *)(src + x))), 8);
+            __m128 v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src), 16));
+            __m128 v_dst_0 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src), 16));
+            __m128 v_dst_1 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            _mm_storeu_ps(dst + x, v_dst_0);
+            _mm_storeu_ps(dst + x + 4, v_dst_1);
+        }
+
+        return x;
+    }
+};
+
+template <>
+struct cvtScale_SIMD<schar, double, float>
+{
+    int operator () (const schar * src, double * dst, int width, float scale, float shift) const
+    {
+        int x = 0;
+
+        if (!USE_SSE2)
+            return x;
+
+        __m128i v_zero = _mm_setzero_si128();
+        __m128 v_scale = _mm_set1_ps(scale), v_shift = _mm_set1_ps(shift);
+
+        for ( ; x <= width - 8; x += 8)
+        {
+            __m128i v_src = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _mm_loadl_epi64((__m128i const *)(src + x))), 8);
+            __m128 v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src), 16));
+            __m128 v_dst_0 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            v_src_f = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src), 16));
+            __m128 v_dst_1 = _mm_add_ps(_mm_mul_ps(v_src_f, v_scale), v_shift);
+
+            _mm_storeu_pd(dst + x, _mm_cvtps_pd(v_dst_0));
+            _mm_storeu_pd(dst + x + 4, _mm_cvtps_pd(_mm_castsi128_ps(
+                _mm_srli_si128(_mm_castps_si128(v_dst_0), 16))));
+
+            _mm_storeu_pd(dst + x + 8, _mm_cvtps_pd(v_dst_1));
+            _mm_storeu_pd(dst + x + 12, _mm_cvtps_pd(_mm_castsi128_ps(
+                _mm_srli_si128(_mm_castps_si128(v_dst_1), 16))));
+        }
+
+        return x;
+    }
+};
+
 #elif CV_NEON
 
 // from uchar