Merge pull request #13334 from terfendail:histogram_wintr

* added performance test for compareHist

* compareHist reworked to use wide universal intrinsics

* Disabled vectorization for CV_COMP_CORREL and CV_COMP_BHATTACHARYYA if f64 is unsupported

modules/core/include/opencv2/core/hal/intrin_avx.hpp

@@ -1125,6 +1125,12 @@ inline float v_reduce_sum(const v_float32x8& a)
     return _mm_cvtss_f32(s1);
 }
 
+inline double v_reduce_sum(const v_float64x4& a)
+{
+    __m256d s0 = _mm256_hadd_pd(a.val, a.val);
+    return _mm_cvtsd_f64(_mm_add_pd(_v256_extract_low(s0), _v256_extract_high(s0)));
+}
+
 inline v_float32x8 v_reduce_sum4(const v_float32x8& a, const v_float32x8& b,
                                  const v_float32x8& c, const v_float32x8& d)
 {

modules/core/include/opencv2/core/hal/intrin_neon.hpp

@@ -984,6 +984,13 @@ OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_float32x4, float32x2, float, sum, add, f32)
 OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_float32x4, float32x2, float, max, max, f32)
 OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_float32x4, float32x2, float, min, min, f32)
 
+#if CV_SIMD128_64F
+inline double v_reduce_sum(const v_float64x2& a)
+{
+    return vgetq_lane_f64(a.val, 0) + vgetq_lane_f64(a.val, 1);
+}
+#endif
+
 inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
                                  const v_float32x4& c, const v_float32x4& d)
 {

modules/core/include/opencv2/core/hal/intrin_sse.hpp

@@ -1456,6 +1456,13 @@ OPENCV_HAL_IMPL_SSE_REDUCE_OP_4_SUM(v_uint32x4, unsigned, __m128i, epi32, OPENCV
 OPENCV_HAL_IMPL_SSE_REDUCE_OP_4_SUM(v_int32x4, int, __m128i, epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP, si128_si32)
 OPENCV_HAL_IMPL_SSE_REDUCE_OP_4_SUM(v_float32x4, float, __m128, ps, _mm_castps_si128, _mm_castsi128_ps, ss_f32)
 
+inline double v_reduce_sum(const v_float64x2& a)
+{
+    double CV_DECL_ALIGNED(32) idx[2];
+    v_store_aligned(idx, a);
+    return idx[0] + idx[1];
+}
+
 inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
                                  const v_float32x4& c, const v_float32x4& d)
 {

modules/core/include/opencv2/core/hal/intrin_vsx.hpp

@@ -716,6 +716,11 @@ OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_float32x4, vec_float4, float, sum, vec_add)
 OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_float32x4, vec_float4, float, max, vec_max)
 OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_float32x4, vec_float4, float, min, vec_min)
 
+inline double v_reduce_sum(const v_float64x2& a)
+{
+    return vec_extract(vec_add(a.val, vec_sld(a.val, a.val, 8)), 0);
+}
+
 #define OPENCV_HAL_IMPL_VSX_REDUCE_OP_8(_Tpvec, _Tpvec2, scalartype, suffix, func) \
 inline scalartype v_reduce_##suffix(const _Tpvec& a)                               \
 {                                                                                  \

modules/imgproc/perf/perf_histogram.cpp

@@ -116,6 +116,31 @@ PERF_TEST_P(MatSize, equalizeHist,
 }
 #undef MatSize
 
+typedef TestBaseWithParam< tuple<int, int> > Dim_Cmpmethod;
+PERF_TEST_P(Dim_Cmpmethod, compareHist,
+            testing::Combine(testing::Values(1, 3),
+                             testing::Values(HISTCMP_CORREL, HISTCMP_CHISQR, HISTCMP_INTERSECT, HISTCMP_BHATTACHARYYA, HISTCMP_CHISQR_ALT, HISTCMP_KL_DIV))
+            )
+{
+    int dims = get<0>(GetParam());
+    int method = get<1>(GetParam());
+    int histSize[] = { 2048, 128, 64 };
+
+    Mat hist1(dims, histSize, CV_32FC1);
+    Mat hist2(dims, histSize, CV_32FC1);
+    randu(hist1, 0, 256);
+    randu(hist2, 0, 256);
+
+    declare.in(hist1.reshape(1, 256), hist2.reshape(1, 256));
+
+    TEST_CYCLE()
+    {
+        compareHist(hist1, hist2, method);
+    }
+
+    SANITY_CHECK_NOTHING();
+}
+
 typedef tuple<Size, double> Sz_ClipLimit_t;
 typedef TestBaseWithParam<Sz_ClipLimit_t> Sz_ClipLimit;
 

modules/imgproc/src/histogram.cpp

@@ -41,6 +41,7 @@
 
 #include "precomp.hpp"
 #include "opencl_kernels_imgproc.hpp"
+#include "opencv2/core/hal/intrin.hpp"
 
 #include "opencv2/core/openvx/ovx_defs.hpp"
 
@@ -1938,10 +1939,6 @@ double cv::compareHist( InputArray _H1, InputArray _H2, int method )
 
     CV_Assert( it.planes[0].isContinuous() && it.planes[1].isContinuous() );
 
-#if CV_SSE2
-    bool haveSIMD = checkHardwareSupport(CV_CPU_SSE2);
-#endif
-
     for( size_t i = 0; i < it.nplanes; i++, ++it )
     {
         const float* h1 = it.planes[0].ptr<float>();
@@ -1961,50 +1958,63 @@ double cv::compareHist( InputArray _H1, InputArray _H2, int method )
         }
         else if( method == CV_COMP_CORREL )
         {
-            #if CV_SSE2
-            if (haveSIMD)
+#if CV_SIMD_64F
+            v_float64 v_s1 = vx_setzero_f64();
+            v_float64 v_s2 = vx_setzero_f64();
+            v_float64 v_s11 = vx_setzero_f64();
+            v_float64 v_s12 = vx_setzero_f64();
+            v_float64 v_s22 = vx_setzero_f64();
+            for ( ; j <= len - v_float32::nlanes; j += v_float32::nlanes)
             {
-                __m128d v_s1 = _mm_setzero_pd(), v_s2 = v_s1;
-                __m128d v_s11 = v_s1, v_s22 = v_s1, v_s12 = v_s1;
-
-                for ( ; j <= len - 4; j += 4)
-                {
-                    __m128 v_a = _mm_loadu_ps(h1 + j);
-                    __m128 v_b = _mm_loadu_ps(h2 + j);
-
-                    // 0-1
-                    __m128d v_ad = _mm_cvtps_pd(v_a);
-                    __m128d v_bd = _mm_cvtps_pd(v_b);
-                    v_s12 = _mm_add_pd(v_s12, _mm_mul_pd(v_ad, v_bd));
-                    v_s11 = _mm_add_pd(v_s11, _mm_mul_pd(v_ad, v_ad));
-                    v_s22 = _mm_add_pd(v_s22, _mm_mul_pd(v_bd, v_bd));
-                    v_s1 = _mm_add_pd(v_s1, v_ad);
-                    v_s2 = _mm_add_pd(v_s2, v_bd);
-
-                    // 2-3
-                    v_ad = _mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_a), 8)));
-                    v_bd = _mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_b), 8)));
-                    v_s12 = _mm_add_pd(v_s12, _mm_mul_pd(v_ad, v_bd));
-                    v_s11 = _mm_add_pd(v_s11, _mm_mul_pd(v_ad, v_ad));
-                    v_s22 = _mm_add_pd(v_s22, _mm_mul_pd(v_bd, v_bd));
-                    v_s1 = _mm_add_pd(v_s1, v_ad);
-                    v_s2 = _mm_add_pd(v_s2, v_bd);
-                }
-
-                double CV_DECL_ALIGNED(16) ar[10];
-                _mm_store_pd(ar, v_s12);
-                _mm_store_pd(ar + 2, v_s11);
-                _mm_store_pd(ar + 4, v_s22);
-                _mm_store_pd(ar + 6, v_s1);
-                _mm_store_pd(ar + 8, v_s2);
-
-                s12 += ar[0] + ar[1];
-                s11 += ar[2] + ar[3];
-                s22 += ar[4] + ar[5];
-                s1 += ar[6] + ar[7];
-                s2 += ar[8] + ar[9];
+                v_float32 v_a = vx_load(h1 + j);
+                v_float32 v_b = vx_load(h2 + j);
+
+                // 0-1
+                v_float64 v_ad = v_cvt_f64(v_a);
+                v_float64 v_bd = v_cvt_f64(v_b);
+                v_s12 = v_muladd(v_ad, v_bd, v_s12);
+                v_s11 = v_muladd(v_ad, v_ad, v_s11);
+                v_s22 = v_muladd(v_bd, v_bd, v_s22);
+                v_s1 += v_ad;
+                v_s2 += v_bd;
+
+                // 2-3
+                v_ad = v_cvt_f64_high(v_a);
+                v_bd = v_cvt_f64_high(v_b);
+                v_s12 = v_muladd(v_ad, v_bd, v_s12);
+                v_s11 = v_muladd(v_ad, v_ad, v_s11);
+                v_s22 = v_muladd(v_bd, v_bd, v_s22);
+                v_s1 += v_ad;
+                v_s2 += v_bd;
             }
-            #endif
+            s12 += v_reduce_sum(v_s12);
+            s11 += v_reduce_sum(v_s11);
+            s22 += v_reduce_sum(v_s22);
+            s1 += v_reduce_sum(v_s1);
+            s2 += v_reduce_sum(v_s2);
+#elif CV_SIMD && 0 //Disable vectorization for CV_COMP_CORREL if f64 is unsupported due to low precision
+            v_float32 v_s1 = vx_setzero_f32();
+            v_float32 v_s2 = vx_setzero_f32();
+            v_float32 v_s11 = vx_setzero_f32();
+            v_float32 v_s12 = vx_setzero_f32();
+            v_float32 v_s22 = vx_setzero_f32();
+            for (; j <= len - v_float32::nlanes; j += v_float32::nlanes)
+            {
+                v_float32 v_a = vx_load(h1 + j);
+                v_float32 v_b = vx_load(h2 + j);
+
+                v_s12 = v_muladd(v_a, v_b, v_s12);
+                v_s11 = v_muladd(v_a, v_a, v_s11);
+                v_s22 = v_muladd(v_b, v_b, v_s22);
+                v_s1 += v_a;
+                v_s2 += v_b;
+            }
+            s12 += v_reduce_sum(v_s12);
+            s11 += v_reduce_sum(v_s11);
+            s22 += v_reduce_sum(v_s22);
+            s1 += v_reduce_sum(v_s1);
+            s2 += v_reduce_sum(v_s2);
+#endif
             for( ; j < len; j++ )
             {
                 double a = h1[j];
@@ -2019,67 +2029,68 @@ double cv::compareHist( InputArray _H1, InputArray _H2, int method )
         }
         else if( method == CV_COMP_INTERSECT )
         {
-            #if CV_NEON
-            float32x4_t v_result = vdupq_n_f32(0.0f);
-            for( ; j <= len - 4; j += 4 )
-                v_result = vaddq_f32(v_result, vminq_f32(vld1q_f32(h1 + j), vld1q_f32(h2 + j)));
-            float CV_DECL_ALIGNED(16) ar[4];
-            vst1q_f32(ar, v_result);
-            result += ar[0] + ar[1] + ar[2] + ar[3];
-            #elif CV_SSE2
-            if (haveSIMD)
+#if CV_SIMD_64F
+            v_float64 v_result = vx_setzero_f64();
+            for ( ; j <= len - v_float32::nlanes; j += v_float32::nlanes)
             {
-                __m128d v_result = _mm_setzero_pd();
-                for ( ; j <= len - 4; j += 4)
-                {
-                    __m128 v_src = _mm_min_ps(_mm_loadu_ps(h1 + j),
-                                              _mm_loadu_ps(h2 + j));
-                    v_result = _mm_add_pd(v_result, _mm_cvtps_pd(v_src));
-                    v_src = _mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_src), 8));
-                    v_result = _mm_add_pd(v_result, _mm_cvtps_pd(v_src));
-                }
-
-                double CV_DECL_ALIGNED(16) ar[2];
-                _mm_store_pd(ar, v_result);
-                result += ar[0] + ar[1];
+                v_float32 v_src = v_min(vx_load(h1 + j), vx_load(h2 + j));
+                v_result += v_cvt_f64(v_src) + v_cvt_f64_high(v_src);
+            }
+            result += v_reduce_sum(v_result);
+#elif CV_SIMD
+            v_float32 v_result = vx_setzero_f32();
+            for (; j <= len - v_float32::nlanes; j += v_float32::nlanes)
+            {
+                v_float32 v_src = v_min(vx_load(h1 + j), vx_load(h2 + j));
+                v_result += v_src;
             }
-            #endif
+            result += v_reduce_sum(v_result);
+#endif
             for( ; j < len; j++ )
                 result += std::min(h1[j], h2[j]);
         }
         else if( method == CV_COMP_BHATTACHARYYA )
         {
-            #if CV_SSE2
-            if (haveSIMD)
+#if CV_SIMD_64F
+            v_float64 v_s1 = vx_setzero_f64();
+            v_float64 v_s2 = vx_setzero_f64();
+            v_float64 v_result = vx_setzero_f64();
+            for ( ; j <= len - v_float32::nlanes; j += v_float32::nlanes)
             {
-                __m128d v_s1 = _mm_setzero_pd(), v_s2 = v_s1, v_result = v_s1;
-                for ( ; j <= len - 4; j += 4)
-                {
-                    __m128 v_a = _mm_loadu_ps(h1 + j);
-                    __m128 v_b = _mm_loadu_ps(h2 + j);
-
-                    __m128d v_ad = _mm_cvtps_pd(v_a);
-                    __m128d v_bd = _mm_cvtps_pd(v_b);
-                    v_s1 = _mm_add_pd(v_s1, v_ad);
-                    v_s2 = _mm_add_pd(v_s2, v_bd);
-                    v_result = _mm_add_pd(v_result, _mm_sqrt_pd(_mm_mul_pd(v_ad, v_bd)));
-
-                    v_ad = _mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_a), 8)));
-                    v_bd = _mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_b), 8)));
-                    v_s1 = _mm_add_pd(v_s1, v_ad);
-                    v_s2 = _mm_add_pd(v_s2, v_bd);
-                    v_result = _mm_add_pd(v_result, _mm_sqrt_pd(_mm_mul_pd(v_ad, v_bd)));
-                }
-
-                double CV_DECL_ALIGNED(16) ar[6];
-                _mm_store_pd(ar, v_s1);
-                _mm_store_pd(ar + 2, v_s2);
-                _mm_store_pd(ar + 4, v_result);
-                s1 += ar[0] + ar[1];
-                s2 += ar[2] + ar[3];
-                result += ar[4] + ar[5];
+                v_float32 v_a = vx_load(h1 + j);
+                v_float32 v_b = vx_load(h2 + j);
+
+                v_float64 v_ad = v_cvt_f64(v_a);
+                v_float64 v_bd = v_cvt_f64(v_b);
+                v_s1 += v_ad;
+                v_s2 += v_bd;
+                v_result += v_sqrt(v_ad * v_bd);
+
+                v_ad = v_cvt_f64_high(v_a);
+                v_bd = v_cvt_f64_high(v_b);
+                v_s1 += v_ad;
+                v_s2 += v_bd;
+                v_result += v_sqrt(v_ad * v_bd);
             }
-            #endif
+            s1 += v_reduce_sum(v_s1);
+            s2 += v_reduce_sum(v_s2);
+            result += v_reduce_sum(v_result);
+#elif CV_SIMD && 0 //Disable vectorization for CV_COMP_BHATTACHARYYA if f64 is unsupported due to low precision
+            v_float32 v_s1 = vx_setzero_f32();
+            v_float32 v_s2 = vx_setzero_f32();
+            v_float32 v_result = vx_setzero_f32();
+            for (; j <= len - v_float32::nlanes; j += v_float32::nlanes)
+            {
+                v_float32 v_a = vx_load(h1 + j);
+                v_float32 v_b = vx_load(h2 + j);
+                v_s1 += v_a;
+                v_s2 += v_b;
+                v_result += v_sqrt(v_a * v_b);
+            }
+            s1 += v_reduce_sum(v_s1);
+            s2 += v_reduce_sum(v_s2);
+            result += v_reduce_sum(v_result);
+#endif
             for( ; j < len; j++ )
             {
                 double a = h1[j];