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

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

@@ -220,6 +220,11 @@ struct VZeroUpperGuard {
 #  define CV_VSX 1
 #endif
 
+#ifdef __F16C__
+#  include <immintrin.h>
+#  define CV_FP16 1
+#endif
+
 #endif // !__OPENCV_BUILD && !__CUDACC (Compatibility code)
 
 

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

@@ -844,7 +844,7 @@ protected:
     float16_t() : w(0) {}
     explicit float16_t(float x)
     {
-    #if CV_AVX2
+    #if CV_FP16
         __m128 v = _mm_load_ss(&x);
         w = (ushort)_mm_cvtsi128_si32(_mm_cvtps_ph(v, 0));
     #else
@@ -875,7 +875,7 @@ protected:
 
     operator float() const
     {
-    #if CV_AVX2
+    #if CV_FP16
         float f;
         _mm_store_ss(&f, _mm_cvtph_ps(_mm_cvtsi32_si128(w)));
         return f;

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

@@ -3121,18 +3121,39 @@ OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_float32x8, float, f32, v_uint32x8, un
 OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_int64x4, int64, s64, v_uint64x4, uint64, u64)
 OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_float64x4, double, f64, v_uint64x4, uint64, u64)
 
+//
 // FP16
+//
+
 inline v_float32x8 v256_load_expand(const float16_t* ptr)
 {
+#if CV_FP16
     return v_float32x8(_mm256_cvtph_ps(_mm_loadu_si128((const __m128i*)ptr)));
+#else
+    float CV_DECL_ALIGNED(32) buf[8];
+    for (int i = 0; i < 8; i++)
+        buf[i] = (float)ptr[i];
+    return v256_load_aligned(buf);
+#endif
 }
 
 inline void v_pack_store(float16_t* ptr, const v_float32x8& a)
 {
+#if CV_FP16
     __m128i ah = _mm256_cvtps_ph(a.val, 0);
     _mm_storeu_si128((__m128i*)ptr, ah);
+#else
+    float CV_DECL_ALIGNED(32) buf[8];
+    v_store_aligned(buf, a);
+    for (int i = 0; i < 8; i++)
+        ptr[i] = float16_t(buf[i]);
+#endif
 }
 
+//
+// end of FP16
+//
+
 inline void v256_cleanup() { _mm256_zeroall(); }
 
 CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END