Merge remote-tracking branch 'upstream/3.4' into merge-3.4

modules/core/CMakeLists.txt

@@ -2,7 +2,7 @@ set(the_description "The Core Functionality")
 
 ocv_add_dispatched_file(mathfuncs_core SSE2 AVX AVX2)
 ocv_add_dispatched_file(stat SSE4_2 AVX2)
-ocv_add_dispatched_file(arithm SSE2 SSE4_1 AVX2)
+ocv_add_dispatched_file(arithm SSE2 SSE4_1 AVX2 VSX3)
 
 # dispatching for accuracy tests
 ocv_add_dispatched_file_force_all(test_intrin128 TEST SSE2 SSE3 SSSE3 SSE4_1 SSE4_2 AVX FP16 AVX2)

modules/core/include/opencv2/core/cv_cpu_dispatch.h

@@ -107,7 +107,7 @@
 #  include <arm_neon.h>
 #endif
 
-#if defined(__VSX__) && defined(__PPC64__) && defined(__LITTLE_ENDIAN__)
+#ifdef CV_CPU_COMPILE_VSX
 #  include <altivec.h>
 #  undef vector
 #  undef pixel
@@ -115,6 +115,10 @@
 #  define CV_VSX 1
 #endif
 
+#ifdef CV_CPU_COMPILE_VSX3
+#  define CV_VSX3 1
+#endif
+
 #endif // CV_ENABLE_INTRINSICS && !CV_DISABLE_OPTIMIZATION && !__CUDACC__
 
 #if defined CV_CPU_COMPILE_AVX && !defined CV_CPU_BASELINE_COMPILE_AVX
@@ -237,3 +241,7 @@ struct VZeroUpperGuard {
 #ifndef CV_VSX
 #  define CV_VSX 0
 #endif
+
+#ifndef CV_VSX3
+#  define CV_VSX3 0
+#endif

modules/core/include/opencv2/core/cv_cpu_helper.h

@@ -315,5 +315,26 @@
 #endif
 #define __CV_CPU_DISPATCH_CHAIN_VSX(fn, args, mode, ...)  CV_CPU_CALL_VSX(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
 
+#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_VSX3
+#  define CV_TRY_VSX3 1
+#  define CV_CPU_FORCE_VSX3 1
+#  define CV_CPU_HAS_SUPPORT_VSX3 1
+#  define CV_CPU_CALL_VSX3(fn, args) return (cpu_baseline::fn args)
+#  define CV_CPU_CALL_VSX3_(fn, args) return (opt_VSX3::fn args)
+#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_VSX3
+#  define CV_TRY_VSX3 1
+#  define CV_CPU_FORCE_VSX3 0
+#  define CV_CPU_HAS_SUPPORT_VSX3 (cv::checkHardwareSupport(CV_CPU_VSX3))
+#  define CV_CPU_CALL_VSX3(fn, args) if (CV_CPU_HAS_SUPPORT_VSX3) return (opt_VSX3::fn args)
+#  define CV_CPU_CALL_VSX3_(fn, args) if (CV_CPU_HAS_SUPPORT_VSX3) return (opt_VSX3::fn args)
+#else
+#  define CV_TRY_VSX3 0
+#  define CV_CPU_FORCE_VSX3 0
+#  define CV_CPU_HAS_SUPPORT_VSX3 0
+#  define CV_CPU_CALL_VSX3(fn, args)
+#  define CV_CPU_CALL_VSX3_(fn, args)
+#endif
+#define __CV_CPU_DISPATCH_CHAIN_VSX3(fn, args, mode, ...)  CV_CPU_CALL_VSX3(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
+
 #define CV_CPU_CALL_BASELINE(fn, args) return (cpu_baseline::fn args)
 #define __CV_CPU_DISPATCH_CHAIN_BASELINE(fn, args, mode, ...)  CV_CPU_CALL_BASELINE(fn, args) /* last in sequence */

modules/core/include/opencv2/core/cvdef.h

@@ -226,9 +226,10 @@ namespace cv { namespace debug_build_guard { } using namespace debug_build_guard
 #define CV_CPU_AVX_512VBMI      20
 #define CV_CPU_AVX_512VL        21
 
-#define CV_CPU_NEON   100
+#define CV_CPU_NEON             100
 
-#define CV_CPU_VSX 200
+#define CV_CPU_VSX              200
+#define CV_CPU_VSX3             201
 
 // CPU features groups
 #define CV_CPU_AVX512_SKX       256
@@ -266,6 +267,7 @@ enum CpuFeatures {
     CPU_NEON            = 100,
 
     CPU_VSX             = 200,
+    CPU_VSX3            = 201,
 
     CPU_AVX512_SKX      = 256, //!< Skylake-X with AVX-512F/CD/BW/DQ/VL
 

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

@@ -905,6 +905,11 @@ OPENCV_HAL_IMPL_AVX_CMP_OP_64BIT(v_int64x4)
 OPENCV_HAL_IMPL_AVX_CMP_OP_FLT(v_float32x8, ps)
 OPENCV_HAL_IMPL_AVX_CMP_OP_FLT(v_float64x4, pd)
 
+inline v_float32x8 v_not_nan(const v_float32x8& a)
+{ return v_float32x8(_mm256_cmp_ps(a.val, a.val, _CMP_ORD_Q)); }
+inline v_float64x4 v_not_nan(const v_float64x4& a)
+{ return v_float64x4(_mm256_cmp_pd(a.val, a.val, _CMP_ORD_Q)); }
+
 /** min/max **/
 OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_uint8x32,  _mm256_min_epu8)
 OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_uint8x32,  _mm256_max_epu8)

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

@@ -683,6 +683,25 @@ OPENCV_HAL_IMPL_CMP_OP(==)
 For all types except 64-bit integer values. */
 OPENCV_HAL_IMPL_CMP_OP(!=)
 
+template<int n>
+inline v_reg<float, n> v_not_nan(const v_reg<float, n>& a)
+{
+typedef typename V_TypeTraits<float>::int_type itype;
+v_reg<float, n> c;
+for (int i = 0; i < n; i++)
+    c.s[i] = V_TypeTraits<float>::reinterpret_from_int((itype)-(int)(a.s[i] == a.s[i]));
+    return c;
+}
+template<int n>
+inline v_reg<double, n> v_not_nan(const v_reg<double, n>& a)
+{
+    typedef typename V_TypeTraits<double>::int_type itype;
+    v_reg<double, n> c;
+    for (int i = 0; i < n; i++)
+        c.s[i] = V_TypeTraits<double>::reinterpret_from_int((itype)-(int)(a.s[i] == a.s[i]));
+    return c;
+}
+
 //! @brief Helper macro
 //! @ingroup core_hal_intrin_impl
 #define OPENCV_HAL_IMPL_ARITHM_OP(func, bin_op, cast_op, _Tp2) \

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

@@ -764,6 +764,13 @@ OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_int64x2, vreinterpretq_s64_u64, s64, u64)
 OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_float64x2, vreinterpretq_f64_u64, f64, u64)
 #endif
 
+inline v_float32x4 v_not_nan(const v_float32x4& a)
+{ return v_float32x4(vreinterpretq_f32_u32(vceqq_f32(a.val, a.val))); }
+#if CV_SIMD128_64F
+inline v_float64x2 v_not_nan(const v_float64x2& a)
+{ return v_float64x2(vreinterpretq_f64_u64(vceqq_f64(a.val, a.val))); }
+#endif
+
 OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint8x16, v_add_wrap, vaddq_u8)
 OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int8x16, v_add_wrap, vaddq_s8)
 OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint16x8, v_add_wrap, vaddq_u16)

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

@@ -1041,6 +1041,11 @@ inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
 OPENCV_HAL_IMPL_SSE_64BIT_CMP_OP(v_uint64x2, v_reinterpret_as_u64)
 OPENCV_HAL_IMPL_SSE_64BIT_CMP_OP(v_int64x2, v_reinterpret_as_s64)
 
+inline v_float32x4 v_not_nan(const v_float32x4& a)
+{ return v_float32x4(_mm_cmpord_ps(a.val, a.val)); }
+inline v_float64x2 v_not_nan(const v_float64x2& a)
+{ return v_float64x2(_mm_cmpord_pd(a.val, a.val)); }
+
 OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint8x16, v_add_wrap, _mm_add_epi8)
 OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_int8x16, v_add_wrap, _mm_add_epi8)
 OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint16x8, v_add_wrap, _mm_add_epi16)

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

@@ -607,6 +607,11 @@ OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_float64x2)
 OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_uint64x2)
 OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_int64x2)
 
+inline v_float32x4 v_not_nan(const v_float32x4& a)
+{ return v_float32x4(vec_cmpeq(a.val, a.val)); }
+inline v_float64x2 v_not_nan(const v_float64x2& a)
+{ return v_float64x2(vec_cmpeq(a.val, a.val)); }
+
 /** min/max **/
 OPENCV_HAL_IMPL_VSX_BIN_FUNC(v_min, vec_min)
 OPENCV_HAL_IMPL_VSX_BIN_FUNC(v_max, vec_max)

modules/core/include/opencv2/core/types.hpp

@@ -1941,8 +1941,11 @@ Rect_<_Tp>& operator += ( Rect_<_Tp>& a, const Size_<_Tp>& b )
 template<typename _Tp> static inline
 Rect_<_Tp>& operator -= ( Rect_<_Tp>& a, const Size_<_Tp>& b )
 {
-    a.width -= b.width;
-    a.height -= b.height;
+    const _Tp width = a.width - b.width;
+    const _Tp height = a.height - b.height;
+    CV_DbgAssert(width >= 0 && height >= 0);
+    a.width = width;
+    a.height = height;
     return a;
 }
 
@@ -2007,6 +2010,15 @@ Rect_<_Tp> operator + (const Rect_<_Tp>& a, const Size_<_Tp>& b)
     return Rect_<_Tp>( a.x, a.y, a.width + b.width, a.height + b.height );
 }
 
+template<typename _Tp> static inline
+Rect_<_Tp> operator - (const Rect_<_Tp>& a, const Size_<_Tp>& b)
+{
+    const _Tp width = a.width - b.width;
+    const _Tp height = a.height - b.height;
+    CV_DbgAssert(width >= 0 && height >= 0);
+    return Rect_<_Tp>( a.x, a.y, width, height );
+}
+
 template<typename _Tp> static inline
 Rect_<_Tp> operator & (const Rect_<_Tp>& a, const Rect_<_Tp>& b)
 {

modules/core/src/parallel.cpp

@@ -454,7 +454,7 @@ static inline int _initMaxThreads()
     {
         omp_set_dynamic(maxThreads);
     }
-    return numThreads;
+    return maxThreads;
 }
 static int numThreadsMax = _initMaxThreads();
 #elif defined HAVE_GCD

modules/core/src/system.cpp

@@ -107,15 +107,14 @@ void* allocSingletonBuffer(size_t size) { return fastMalloc(size); }
 #  include <cpu-features.h>
 #endif
 
-#ifndef __VSX__
-# if defined __PPC64__ && defined __linux__
-#   include "sys/auxv.h"
-#   ifndef AT_HWCAP2
-#     define AT_HWCAP2 26
-#   endif
-#   ifndef PPC_FEATURE2_ARCH_2_07
-#     define PPC_FEATURE2_ARCH_2_07 0x80000000
-#   endif
+
+#if CV_VSX && defined __linux__
+# include "sys/auxv.h"
+# ifndef AT_HWCAP2
+#   define AT_HWCAP2 26
+# endif
+# ifndef PPC_FEATURE2_ARCH_3_00
+#   define PPC_FEATURE2_ARCH_3_00 0x00800000
 # endif
 #endif
 
@@ -359,6 +358,7 @@ struct HWFeatures
         g_hwFeatureNames[CPU_NEON] = "NEON";
 
         g_hwFeatureNames[CPU_VSX] = "VSX";
+        g_hwFeatureNames[CPU_VSX3] = "VSX3";
 
         g_hwFeatureNames[CPU_AVX512_SKX] = "AVX512-SKX";
     }
@@ -513,14 +513,14 @@ struct HWFeatures
     #endif
     #endif
 
-    #ifdef __VSX__
-        have[CV_CPU_VSX] = true;
-    #elif (defined __PPC64__ && defined __linux__)
-        uint64 hwcaps = getauxval(AT_HWCAP);
+    // there's no need to check VSX availability in runtime since it's always available on ppc64le CPUs
+    have[CV_CPU_VSX] = (CV_VSX);
+    // TODO: Check VSX3 availability in runtime for other platforms
+    #if CV_VSX && defined __linux__
         uint64 hwcap2 = getauxval(AT_HWCAP2);
-        have[CV_CPU_VSX] = (hwcaps & PPC_FEATURE_PPC_LE && hwcaps & PPC_FEATURE_HAS_VSX && hwcap2 & PPC_FEATURE2_ARCH_2_07);
+        have[CV_CPU_VSX3] = (hwcap2 & PPC_FEATURE2_ARCH_3_00);
     #else
-        have[CV_CPU_VSX] = false;
+        have[CV_CPU_VSX3] = (CV_VSX3);
     #endif
 
         int baseline_features[] = { CV_CPU_BASELINE_FEATURES };

modules/core/test/test_operations.cpp

@@ -972,6 +972,13 @@ bool CV_OperationsTest::operations1()
         if (sz.width != 10 || sz.height != 20) throw test_excep();
         if (cvSize(sz).width != 10 || cvSize(sz).height != 20) throw test_excep();
 
+        Rect r1(0, 0, 10, 20);
+        Size sz1(5, 10);
+        r1 -= sz1;
+        if (r1.size().width != 5 || r1.size().height != 10) throw test_excep();
+        Rect r2 = r1 - sz1;
+        if (r2.size().width != 0 || r2.size().height != 0) throw test_excep();
+
         Vec<double, 5> v5d(1, 1, 1, 1, 1);
         Vec<double, 6> v6d(1, 1, 1, 1, 1, 1);
         Vec<double, 7> v7d(1, 1, 1, 1, 1, 1, 1);