From ffd010767fe401f87a747c0e7f935ddb7da4c63b Mon Sep 17 00:00:00 2001
From: Vincent Rabaud <vrabaud@google.com>
Date: Thu, 4 Nov 2021 22:48:10 +0100
Subject: [PATCH] Only use fma functions when CV_FMA3 is set.

In practice, processors offering AVX2/AVX512 also FMA, that is why it got unnoticed.
---
 .../core/include/opencv2/core/hal/intrin_avx.hpp   | 14 +++++++++++++-
 .../include/opencv2/core/hal/intrin_avx512.hpp     | 14 +++++++++++++-
 2 files changed, 26 insertions(+), 2 deletions(-)

diff --git a/modules/core/include/opencv2/core/hal/intrin_avx.hpp b/modules/core/include/opencv2/core/hal/intrin_avx.hpp
index 54e8927192..7fbf6f1314 100644
--- a/modules/core/include/opencv2/core/hal/intrin_avx.hpp
+++ b/modules/core/include/opencv2/core/hal/intrin_avx.hpp
@@ -1390,11 +1390,21 @@ OPENCV_HAL_IMPL_AVX_CHECK_SHORT(v_int16x16)
 ////////// Other math /////////
 
 /** Some frequent operations **/
+#if CV_FMA3
 #define OPENCV_HAL_IMPL_AVX_MULADD(_Tpvec, suffix)                            \
     inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)    \
     { return _Tpvec(_mm256_fmadd_##suffix(a.val, b.val, c.val)); }            \
     inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c) \
-    { return _Tpvec(_mm256_fmadd_##suffix(a.val, b.val, c.val)); }            \
+    { return _Tpvec(_mm256_fmadd_##suffix(a.val, b.val, c.val)); }
+#else
+#define OPENCV_HAL_IMPL_AVX_MULADD(_Tpvec, suffix)                                    \
+    inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)            \
+    { return _Tpvec(_mm256_add_##suffix(_mm256_mul_##suffix(a.val, b.val), c.val)); } \
+    inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)         \
+    { return _Tpvec(_mm256_add_##suffix(_mm256_mul_##suffix(a.val, b.val), c.val)); }
+#endif
+
+#define OPENCV_HAL_IMPL_AVX_MISC(_Tpvec, suffix)                              \
     inline _Tpvec v_sqrt(const _Tpvec& x)                                     \
     { return _Tpvec(_mm256_sqrt_##suffix(x.val)); }                           \
     inline _Tpvec v_sqr_magnitude(const _Tpvec& a, const _Tpvec& b)           \
@@ -1404,6 +1414,8 @@ OPENCV_HAL_IMPL_AVX_CHECK_SHORT(v_int16x16)
 
 OPENCV_HAL_IMPL_AVX_MULADD(v_float32x8, ps)
 OPENCV_HAL_IMPL_AVX_MULADD(v_float64x4, pd)
+OPENCV_HAL_IMPL_AVX_MISC(v_float32x8, ps)
+OPENCV_HAL_IMPL_AVX_MISC(v_float64x4, pd)
 
 inline v_int32x8 v_fma(const v_int32x8& a, const v_int32x8& b, const v_int32x8& c)
 {
diff --git a/modules/core/include/opencv2/core/hal/intrin_avx512.hpp b/modules/core/include/opencv2/core/hal/intrin_avx512.hpp
index 75a3bd4b85..d20d6dd1ff 100644
--- a/modules/core/include/opencv2/core/hal/intrin_avx512.hpp
+++ b/modules/core/include/opencv2/core/hal/intrin_avx512.hpp
@@ -1385,11 +1385,21 @@ inline v_uint64x8  v_popcount(const v_uint64x8&  a) { return v_popcount(v_reinte
 ////////// Other math /////////
 
 /** Some frequent operations **/
+#if CV_FMA3
 #define OPENCV_HAL_IMPL_AVX512_MULADD(_Tpvec, suffix)                         \
     inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)    \
     { return _Tpvec(_mm512_fmadd_##suffix(a.val, b.val, c.val)); }            \
     inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c) \
-    { return _Tpvec(_mm512_fmadd_##suffix(a.val, b.val, c.val)); }            \
+    { return _Tpvec(_mm512_fmadd_##suffix(a.val, b.val, c.val)); }
+#else
+#define OPENCV_HAL_IMPL_AVX512_MULADD(_Tpvec, suffix)                                 \
+    inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)            \
+    { return _Tpvec(_mm512_add_##suffix(_mm512_mul_##suffix(a.val, b.val), c.val)); } \
+    inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)         \
+    { return _Tpvec(_mm512_add_##suffix(_mm512_mul_##suffix(a.val, b.val), c.val)); }
+#endif
+
+#define OPENCV_HAL_IMPL_AVX512_MISC(_Tpvec, suffix)                           \
     inline _Tpvec v_sqrt(const _Tpvec& x)                                     \
     { return _Tpvec(_mm512_sqrt_##suffix(x.val)); }                           \
     inline _Tpvec v_sqr_magnitude(const _Tpvec& a, const _Tpvec& b)           \
@@ -1399,6 +1409,8 @@ inline v_uint64x8  v_popcount(const v_uint64x8&  a) { return v_popcount(v_reinte
 
 OPENCV_HAL_IMPL_AVX512_MULADD(v_float32x16, ps)
 OPENCV_HAL_IMPL_AVX512_MULADD(v_float64x8,  pd)
+OPENCV_HAL_IMPL_AVX512_MISC(v_float32x16, ps)
+OPENCV_HAL_IMPL_AVX512_MISC(v_float64x8,  pd)
 
 inline v_int32x16 v_fma(const v_int32x16& a, const v_int32x16& b, const v_int32x16& c)
 { return a * b + c; }
-- 
GitLab