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

Author: alalek

CMakeLists.txt

@@ -323,7 +323,6 @@ OCV_OPTION(ENABLE_PROFILING           "Enable profiling in the GCC compiler (Add
 OCV_OPTION(ENABLE_COVERAGE            "Enable coverage collection with  GCov"                    OFF  IF CV_GCC )
 OCV_OPTION(ENABLE_OMIT_FRAME_POINTER  "Enable -fomit-frame-pointer for GCC"                      ON   IF CV_GCC )
 OCV_OPTION(ENABLE_POWERPC             "Enable PowerPC for GCC"                                   ON   IF (CV_GCC AND CMAKE_SYSTEM_PROCESSOR MATCHES powerpc.*) )
-OCV_OPTION(ENABLE_VSX                 "Enable POWER8 and above VSX (64-bit little-endian)"       ON   IF ((CV_GCC OR CV_CLANG) AND PPC64LE) )
 OCV_OPTION(ENABLE_FAST_MATH           "Enable -ffast-math (not recommended for GCC 4.6.x)"       OFF  IF (CV_GCC AND (X86 OR X86_64)) )
 if(NOT IOS AND (NOT ANDROID OR OPENCV_ANDROID_USE_LEGACY_FLAGS))  # Use CPU_BASELINE instead
 OCV_OPTION(ENABLE_NEON                "Enable NEON instructions"                                 (NEON OR ANDROID_ARM_NEON OR AARCH64) IF (CV_GCC OR CV_CLANG) AND (ARM OR AARCH64 OR IOS) )

cmake/OpenCVCompilerOptimizations.cmake

@@ -5,6 +5,10 @@
 # AVX / AVX2 / AVX_512F
 # FMA3
 
+# ppc64le arch:
+# VSX  (always available on Power8)
+# VSX3 (always available on Power9)
+
 # CPU_{opt}_SUPPORTED=ON/OFF - compiler support (possibly with additional flag)
 # CPU_{opt}_IMPLIES=<list>
 # CPU_{opt}_FORCE=<list> - subset of "implies" list
@@ -29,7 +33,7 @@
 
 set(CPU_ALL_OPTIMIZATIONS "SSE;SSE2;SSE3;SSSE3;SSE4_1;SSE4_2;POPCNT;AVX;FP16;AVX2;FMA3;AVX_512F;AVX512_SKX")
 list(APPEND CPU_ALL_OPTIMIZATIONS NEON VFPV3 FP16)
-list(APPEND CPU_ALL_OPTIMIZATIONS VSX)
+list(APPEND CPU_ALL_OPTIMIZATIONS VSX VSX3)
 list(REMOVE_DUPLICATES CPU_ALL_OPTIMIZATIONS)
 
 ocv_update(CPU_VFPV3_FEATURE_ALIAS "")
@@ -81,7 +85,7 @@ ocv_optimization_process_obsolete_option(ENABLE_FMA3 FMA3 ON)
 ocv_optimization_process_obsolete_option(ENABLE_VFPV3 VFPV3 OFF)
 ocv_optimization_process_obsolete_option(ENABLE_NEON NEON OFF)
 
-ocv_optimization_process_obsolete_option(ENABLE_VSX VSX OFF)
+ocv_optimization_process_obsolete_option(ENABLE_VSX VSX ON)
 
 macro(ocv_is_optimization_in_list resultvar check_opt)
   set(__checked "")
@@ -289,14 +293,24 @@ elseif(ARM OR AARCH64)
     set(CPU_BASELINE "NEON;FP16" CACHE STRING "${HELP_CPU_BASELINE}")
   endif()
 elseif(PPC64LE)
-  ocv_update(CPU_KNOWN_OPTIMIZATIONS "VSX")
+  ocv_update(CPU_KNOWN_OPTIMIZATIONS "VSX;VSX3")
   ocv_update(CPU_VSX_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_vsx.cpp")
+  ocv_update(CPU_VSX3_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_vsx3.cpp")
+
+  if(NOT OPENCV_CPU_OPT_IMPLIES_IGNORE)
+    ocv_update(CPU_VSX3_IMPLIES "VSX")
+  endif()
 
   if(CV_CLANG AND (NOT ${CMAKE_CXX_COMPILER} MATCHES "xlc"))
     ocv_update(CPU_VSX_FLAGS_ON "-mvsx -maltivec")
+    ocv_update(CPU_VSX3_FLAGS_ON "-mpower9-vector")
   else()
     ocv_update(CPU_VSX_FLAGS_ON "-mcpu=power8")
+    ocv_update(CPU_VSX3_FLAGS_ON "-mcpu=power9 -mtune=power9")
   endif()
+
+  set(CPU_DISPATCH "VSX3" CACHE STRING "${HELP_CPU_DISPATCH}")
+  set(CPU_BASELINE "VSX" CACHE STRING "${HELP_CPU_BASELINE}")
 endif()
 
 # Helper values for cmake-gui

cmake/checks/cpu_vsx.cpp

@@ -1,8 +1,12 @@
-# if defined(__VSX__)
-#   include <altivec.h>
-# else
-#   error "VSX is not supported"
-# endif
+#if defined(__VSX__)
+    #if defined(__PPC64__) && defined(__LITTLE_ENDIAN__)
+        #include <altivec.h>
+    #else
+        #error "OpenCV only supports little-endian mode"
+    #endif
+#else
+    #error "VSX is not supported"
+#endif
 
 int main()
 {

cmake/checks/cpu_vsx3.cpp

@@ -0,0 +1,17 @@
+#if defined(__VSX__)
+    #if defined(__PPC64__) && defined(__LITTLE_ENDIAN__)
+        #include <altivec.h>
+    #else
+        #error "OpenCV only supports little-endian mode"
+    #endif
+#else
+    #error "VSX3 is not supported"
+#endif
+
+int main()
+{
+    __vector unsigned char a = vec_splats((unsigned char)1);
+    __vector unsigned char b = vec_splats((unsigned char)2);
+    __vector unsigned char r = vec_absd(a, b);
+    return 0;
+}

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

@@ -107,14 +107,18 @@
 #  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
 #  undef bool
 #  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) \