From dbe894fea373e7048a8ed8d7c3606a794fe57705 Mon Sep 17 00:00:00 2001
From: Matthias Hochsteger <mhochsteger@cerbsim.com>
Date: Tue, 15 Dec 2020 10:12:30 +0100
Subject: [PATCH 1/4] Support for Apple M1

---
 libsrc/core/ngcore_api.hpp   |  8 ++++
 libsrc/core/simd.hpp         | 14 ++++---
 libsrc/core/simd_generic.hpp | 78 +++++++++++++++++++++++++-----------
 libsrc/core/taskmanager.cpp  |  8 +++-
 libsrc/core/utils.hpp        | 23 +++++++++--
 5 files changed, 96 insertions(+), 35 deletions(-)

diff --git a/libsrc/core/ngcore_api.hpp b/libsrc/core/ngcore_api.hpp
index b6412157..330e7e33 100644
--- a/libsrc/core/ngcore_api.hpp
+++ b/libsrc/core/ngcore_api.hpp
@@ -67,6 +67,14 @@
   #endif
 #endif
 
+#if defined(__amd64__) || defined(_M_AMD64)
+#define NETGEN_ARCH_AMD64
+#endif
+
+#if defined(__arm64__) || defined(_M_ARM64)
+#define NETGEN_ARCH_ARM
+#endif
+
 #ifdef __MAC_OS_X_VERSION_MIN_REQUIRED
 #if __MAC_OS_X_VERSION_MIN_REQUIRED < 101400
 // The c++ standard library on MacOS 10.13 and earlier has no aligned new operator,
diff --git a/libsrc/core/simd.hpp b/libsrc/core/simd.hpp
index 277dd851..0d69dec1 100644
--- a/libsrc/core/simd.hpp
+++ b/libsrc/core/simd.hpp
@@ -11,7 +11,7 @@
 
 #include "simd_generic.hpp"
 
-#if (defined(_M_AMD64) || defined(_M_X64) || defined(__SSE__))
+#ifdef NETGEN_ARCH_AMD64
 #ifndef __SSE__
 #define __SSE__
 #endif
@@ -28,6 +28,7 @@
 
 namespace ngcore
 {
+#ifdef NETGEN_ARCH_AMD64
   NETGEN_INLINE auto HSum (SIMD<double,2> v1, SIMD<double,2> v2, SIMD<double,2> v3, SIMD<double,2> v4)
   {
     SIMD<double,2> hsum1 = my_mm_hadd_pd (v1.Data(), v2.Data());
@@ -35,6 +36,12 @@ namespace ngcore
     return SIMD<double,4> (hsum1, hsum2);
   }
 
+  NETGEN_INLINE auto GetMaskFromBits( unsigned int i )
+  {
+    return SIMD<mask64>::GetMaskFromBits(i);
+  }
+#endif
+
 
   NETGEN_INLINE void SIMDTranspose (SIMD<double,4> a1, SIMD<double,4> a2, SIMD <double,4> a3, SIMD<double,4> a4,
                              SIMD<double,4> & b1, SIMD<double,4> & b2, SIMD<double,4> & b3, SIMD<double,4> & b4)
@@ -59,11 +66,6 @@ namespace ngcore
   {
     return SIMD<double,4>(HSum(s1), HSum(s2), HSum(s3), HSum(s4));
   }
-
-  NETGEN_INLINE auto GetMaskFromBits( unsigned int i )
-  {
-    return SIMD<mask64>::GetMaskFromBits(i);
-  }
 }
 
 #endif // NETGEN_CORE_SIMD_HPP
diff --git a/libsrc/core/simd_generic.hpp b/libsrc/core/simd_generic.hpp
index 8ebd399a..c83b5348 100644
--- a/libsrc/core/simd_generic.hpp
+++ b/libsrc/core/simd_generic.hpp
@@ -21,10 +21,10 @@ namespace ngcore
     return 8;
 #elif defined __AVX__
     return 4;
-#elif (defined(_M_AMD64) || defined(_M_X64) || defined(__SSE__))
+#elif defined NETGEN_ARCH_AMD64
     return 2;
 #else
-    return 1;
+    return 2;
 #endif
   }
 
@@ -104,8 +104,10 @@ namespace ngcore
     SIMD () {}
     SIMD (const SIMD &) = default;
     SIMD & operator= (const SIMD &) = default;
-    SIMD (int64_t val) { data = val; }
-    SIMD (std::array<int64_t, 1> arr)
+    SIMD (int val) : data{val} {}
+    SIMD (int64_t val) : data{val} {}
+    SIMD (size_t val) : data(val) {}
+    explicit SIMD (std::array<int64_t, 1> arr)
         : data{arr[0]}
     {}
 
@@ -136,16 +138,18 @@ namespace ngcore
     SIMD (const SIMD &) = default;
     SIMD & operator= (const SIMD &) = default;
 
+    SIMD (int val) : lo{val}, high{val} { ; }
     SIMD (int64_t val) : lo{val}, high{val} { ; }
+    SIMD (size_t val) : lo{val}, high{val} { ; }
     SIMD (SIMD<int64_t,N1> lo_, SIMD<int64_t,N2> high_) : lo(lo_), high(high_) { ; }
 
-    SIMD( std::array<int64_t, N> arr )
+    explicit SIMD( std::array<int64_t, N> arr )
         : lo(detail::array_range<N1>(arr, 0)),
           high(detail::array_range<N2>(arr, N1))
       {}
 
     template<typename ...T>
-    SIMD(const T... vals)
+    explicit SIMD(const T... vals)
     : lo(detail::array_range<N1>(std::array<int64_t, N>{vals...}, 0)),
       high(detail::array_range<N2>(std::array<int64_t, N>{vals...}, N1))
       {
@@ -204,7 +208,7 @@ namespace ngcore
     SIMD (size_t val) { data = val; }
     SIMD (double const * p) { data = *p; }
     SIMD (double const * p, SIMD<mask64,1> mask) { data = mask.Data() ? *p : 0.0; }
-    SIMD (std::array<double, 1> arr)
+    explicit SIMD (std::array<double, 1> arr)
         : data{arr[0]}
     {}
 
@@ -253,19 +257,17 @@ namespace ngcore
     template <typename T, typename std::enable_if<std::is_convertible<T,std::function<double(int)>>::value,int>::type = 0>
     SIMD (const T & func)
     {
-      for(auto i : IntRange(N1))
-          lo[i] = func(i);
-      for(auto i : IntRange(N2))
-          high[i] = func(N1+i);
+      double  *p = (double*)this;
+      for(auto i : IntRange(N))
+          p[i] = func(i);
     }
 
     template <typename T, typename std::enable_if<std::is_convertible<T,std::function<double(int)>>::value,int>::type = 0>
     SIMD & operator= (const T & func)
     {
-      for(auto i : IntRange(N1))
-          lo[i] = func(i);
-      for(auto i : IntRange(N2))
-          high[i] = func(N1+i);
+      double  *p = (double*)this;
+      for(auto i : IntRange(N))
+          p[i] = func(i);
       return *this;
     }
 
@@ -285,13 +287,13 @@ namespace ngcore
         : lo{p, mask.Lo()}, high{p+N1, mask.Hi()}
       { }
 
-    SIMD( std::array<double, N> arr )
+    explicit SIMD( std::array<double, N> arr )
         : lo(detail::array_range<N1>(arr, 0)),
           high(detail::array_range<N2>(arr, N1))
       {}
 
     template<typename ...T>
-    SIMD(const T... vals)
+    explicit SIMD(const T... vals)
     : lo(detail::array_range<N1>(std::array<double, N>{vals...}, 0)),
       high(detail::array_range<N2>(std::array<double, N>{vals...}, N1))
       {
@@ -312,7 +314,10 @@ namespace ngcore
 
     template<typename=std::enable_if<N==2>>
     operator std::tuple<double&,double&> ()
-    { return std::tuple<double&,double&>((*this)[0], (*this)[1]); }
+    { 
+	double *p = (double*)this;
+	return std::tuple<double&,double&>(p[0], p[1]); 
+    }
 
     template<typename=std::enable_if<N==4>>
     operator std::tuple<double&,double&,double&,double&> ()
@@ -325,6 +330,7 @@ namespace ngcore
       if constexpr(I<N1) return lo.template Get<I>();
       else               return high.template Get<I-N1>();
     }
+    auto Data() const { return *this; }
   };
 
 
@@ -359,42 +365,42 @@ namespace ngcore
   }
 
   template <typename T, int N>
-  NETGEN_INLINE SIMD<mask64,N> operator< (SIMD<T,N> & a, SIMD<T,N> b)
+  NETGEN_INLINE SIMD<mask64,N> operator< (SIMD<T,N> a, SIMD<T,N> b)
     {
       if constexpr(N==1) return a.Data() < b.Data();
       else               return { a.Lo()<b.Lo(), a.Hi()<b.Hi() };
     }
 
   template <typename T, int N>
-  NETGEN_INLINE SIMD<mask64,N> operator<= (SIMD<T,N> & a, SIMD<T,N> b)
+  NETGEN_INLINE SIMD<mask64,N> operator<= (SIMD<T,N> a, SIMD<T,N> b)
     {
       if constexpr(N==1) return a.Data() <= b.Data();
       else               return { a.Lo()<=b.Lo(), a.Hi()<=b.Hi() };
     }
 
   template <typename T, int N>
-  NETGEN_INLINE SIMD<mask64,N> operator> (SIMD<T,N> & a, SIMD<T,N> b)
+  NETGEN_INLINE SIMD<mask64,N> operator> (SIMD<T,N> a, SIMD<T,N> b)
     {
       if constexpr(N==1) return a.Data() > b.Data();
       else               return { a.Lo()>b.Lo(), a.Hi()>b.Hi() };
     }
 
   template <typename T, int N>
-  NETGEN_INLINE SIMD<mask64,N> operator>= (SIMD<T,N> & a, SIMD<T,N> b)
+  NETGEN_INLINE SIMD<mask64,N> operator>= (SIMD<T,N> a, SIMD<T,N> b)
     {
       if constexpr(N==1) return a.Data() >= b.Data();
       else               return { a.Lo()>=b.Lo(), a.Hi()>=b.Hi() };
     }
 
   template <typename T, int N>
-  NETGEN_INLINE SIMD<mask64,N> operator== (SIMD<T,N> & a, SIMD<T,N> b)
+  NETGEN_INLINE SIMD<mask64,N> operator== (SIMD<T,N> a, SIMD<T,N> b)
     {
       if constexpr(N==1) return a.Data() == b.Data();
       else               return { a.Lo()==b.Lo(), a.Hi()==b.Hi() };
     }
 
   template <typename T, int N>
-  NETGEN_INLINE SIMD<mask64,N> operator!= (SIMD<T,N> & a, SIMD<T,N> b)
+  NETGEN_INLINE SIMD<mask64,N> operator!= (SIMD<T,N> a, SIMD<T,N> b)
     {
       if constexpr(N==1) return a.Data() != b.Data();
       else               return { a.Lo()!=b.Lo(), a.Hi()!=b.Hi() };
@@ -547,6 +553,30 @@ namespace ngcore
     return ost;
   }
 
+  using std::sqrt;
+  template <int N>
+  NETGEN_INLINE ngcore::SIMD<double,N> sqrt (ngcore::SIMD<double,N> a) {
+    return ngcore::SIMD<double>([a](int i)->double { return sqrt(a[i]); } );
+  }
+
+  using std::fabs;
+  template <int N>
+  NETGEN_INLINE ngcore::SIMD<double,N> fabs (ngcore::SIMD<double,N> a) {
+    return ngcore::SIMD<double>([a](int i)->double { return fabs(a[i]); } );
+  }
+
+  using std::floor;
+  template <int N>
+  NETGEN_INLINE ngcore::SIMD<double,N> floor (ngcore::SIMD<double,N> a) {
+    return ngcore::SIMD<double>([a](int i)->double { return floor(a[i]); } );
+  }
+
+  using std::ceil;
+  template <int N>
+  NETGEN_INLINE ngcore::SIMD<double,N> ceil (ngcore::SIMD<double,N> a) {
+    return ngcore::SIMD<double>([a](int i)->double { return ceil(a[i]); } );
+  }
+
   using std::exp;
   template <int N>
   NETGEN_INLINE ngcore::SIMD<double,N> exp (ngcore::SIMD<double,N> a) {
diff --git a/libsrc/core/taskmanager.cpp b/libsrc/core/taskmanager.cpp
index a1049a1c..1d88b766 100644
--- a/libsrc/core/taskmanager.cpp
+++ b/libsrc/core/taskmanager.cpp
@@ -201,14 +201,14 @@ namespace ngcore
       ;
   }
 
-  static size_t calibrate_init_tsc = __rdtsc();
+  static size_t calibrate_init_tsc = GetTimeCounter();
   typedef std::chrono::system_clock TClock;
   static TClock::time_point calibrate_init_clock = TClock::now();
   
   void TaskManager :: StopWorkers()
   {
     done = true;
-    double delta_tsc = __rdtsc()-calibrate_init_tsc;
+    double delta_tsc = GetTimeCounter()-calibrate_init_tsc;
     double delta_sec = std::chrono::duration<double>(TClock::now()-calibrate_init_clock).count();
     double frequ = (delta_sec != 0) ? delta_tsc/delta_sec : 2.7e9;
     
@@ -421,7 +421,11 @@ namespace ngcore
       if (workers_on_node[j])
         {
           while (complete[j] != jobnr)
+          {
+#ifdef NETGEN_ARCH_AMD64
             _mm_pause();
+#endif // NETGEN_ARCH_AMD64
+          }
         }
 
     func = nullptr;
diff --git a/libsrc/core/utils.hpp b/libsrc/core/utils.hpp
index 81b0073f..ca015ae3 100644
--- a/libsrc/core/utils.hpp
+++ b/libsrc/core/utils.hpp
@@ -8,13 +8,19 @@
 #include <sstream>
 #include <string>
 
+#include "ngcore_api.hpp"       // for NGCORE_API and CPU arch macros
+
+#if defined(__APPLE__) && defined(NETGEN_ARCH_ARM)
+#include <mach/mach_time.h>
+#endif
+
+#ifdef NETGEN_ARCH_AMD64
 #ifdef WIN32
 #include <intrin.h>   // for __rdtsc()  CPU time step counter
 #else
 #include <x86intrin.h>   // for __rdtsc()  CPU time step counter
 #endif // WIN32
-
-#include "ngcore_api.hpp"       // for NGCORE_API
+#endif // NETGEN_ARCH_AMD64
 
 namespace ngcore
 {
@@ -52,7 +58,16 @@ namespace ngcore
 
   inline TTimePoint GetTimeCounter() noexcept
   {
-      return TTimePoint(__rdtsc());
+#if defined(__APPLE__) && defined(NETGEN_ARCH_ARM)
+    return mach_absolute_time();
+#elif defined(NETGEN_ARCH_AMD64)
+    return __rdtsc();
+#elif defined(NETGEN_ARCH_ARM)
+    return __builtin_readcyclecounter();
+#else
+#warning "Unsupported CPU architecture"
+    return 0;
+#endif
   }
 
   template <class T>
@@ -161,7 +176,9 @@ namespace ngcore
       while (!m.compare_exchange_weak(should, true))
         {
           should = false;
+#ifdef NETGEN_ARCH_AMD64
           _mm_pause();
+#endif // NETGEN_ARCH_AMD64
         }
     }
     void unlock()

From 9c0dbec8c99288138a1dbaadc4461d0e6980e041 Mon Sep 17 00:00:00 2001
From: Matthias Hochsteger <mhochsteger@cerbsim.com>
Date: Tue, 15 Dec 2020 15:31:17 +0100
Subject: [PATCH 2/4] Fix SIMD<mask64> ctor and Unpack

---
 libsrc/core/simd_generic.hpp | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/libsrc/core/simd_generic.hpp b/libsrc/core/simd_generic.hpp
index c83b5348..5846ad4c 100644
--- a/libsrc/core/simd_generic.hpp
+++ b/libsrc/core/simd_generic.hpp
@@ -76,7 +76,7 @@ namespace ngcore
     SIMD<mask64,N2> hi;
   public:
 
-    SIMD (int i) : lo(i), hi(i-N1) { ; }
+    SIMD (size_t i) : lo(i), hi(i>N1 ? i-N1 : 0) { ; }
     SIMD (SIMD<mask64,N1> lo_, SIMD<mask64,N2> hi_) : lo(lo_), hi(hi_) { ; }
     SIMD<mask64,N1> Lo() const { return lo; }
     SIMD<mask64,N2> Hi() const { return hi; }
@@ -664,6 +664,11 @@ namespace ngcore
       {
         return std::make_tuple(SIMD<double,N>{a.Data()}, SIMD<double,N>{b.Data()} );
       }
+    else if constexpr(N==2)
+      {
+        return std::make_tuple(SIMD<double,N>{ a.Lo(), b.Lo() },
+            SIMD<double,N>{ a.Hi(), b.Hi() });
+      }
     else
       {
         auto [a1,b1] = Unpack(a.Lo(), b.Lo());

From d97a9a65949f99075b7b56549b339ac2f6be5c8a Mon Sep 17 00:00:00 2001
From: Matthias Hochsteger <mhochsteger@cerbsim.com>
Date: Wed, 16 Dec 2020 17:20:18 +0100
Subject: [PATCH 3/4] Alignment for generic SIMD classes

---
 libsrc/core/simd_generic.hpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/libsrc/core/simd_generic.hpp b/libsrc/core/simd_generic.hpp
index 5846ad4c..36f44d85 100644
--- a/libsrc/core/simd_generic.hpp
+++ b/libsrc/core/simd_generic.hpp
@@ -67,7 +67,7 @@ namespace ngcore
 
 
   template <int N>
-  class SIMD<mask64,N>
+  class alignas(GetDefaultSIMDSize()*sizeof(int64_t)) SIMD<mask64,N>
   {
     static constexpr int N1 = std::min(GetDefaultSIMDSize(), N/2);
     static constexpr int N2 = N-N1;
@@ -123,7 +123,7 @@ namespace ngcore
   };
 
   template<int N>
-  class SIMD<int64_t,N>
+  class alignas(GetDefaultSIMDSize()*sizeof(int64_t)) SIMD<int64_t,N>
   {
     static constexpr int N1 = std::min(GetDefaultSIMDSize(), N/2);
     static constexpr int N2 = N-N1;
@@ -240,7 +240,7 @@ namespace ngcore
 
 
   template<int N>
-  class  SIMD<double, N>
+  class alignas(GetDefaultSIMDSize()*sizeof(double)) SIMD<double, N>
   {
     static constexpr int N1 = std::min(GetDefaultSIMDSize(), N/2);
     static constexpr int N2 = N-N1;

From eb6ac164e7fe94af30348e6b3e447c775bac7230 Mon Sep 17 00:00:00 2001
From: Matthias Hochsteger <mhochsteger@cerbsim.com>
Date: Wed, 16 Dec 2020 21:00:12 +0100
Subject: [PATCH 4/4] int64_t for masks

---
 libsrc/core/simd_avx.hpp     | 2 +-
 libsrc/core/simd_generic.hpp | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/libsrc/core/simd_avx.hpp b/libsrc/core/simd_avx.hpp
index f089a0b2..09f5d7de 100644
--- a/libsrc/core/simd_avx.hpp
+++ b/libsrc/core/simd_avx.hpp
@@ -33,7 +33,7 @@ namespace ngcore
   {
     __m256i mask;
   public:
-    SIMD (size_t i)
+    SIMD (int64_t i)
       : mask(my_mm256_cmpgt_epi64(_mm256_set1_epi64x(i),
                                   _mm256_set_epi64x(3, 2, 1, 0)))
     { ; }
diff --git a/libsrc/core/simd_generic.hpp b/libsrc/core/simd_generic.hpp
index 36f44d85..849e0922 100644
--- a/libsrc/core/simd_generic.hpp
+++ b/libsrc/core/simd_generic.hpp
@@ -58,7 +58,7 @@ namespace ngcore
   {
     int64_t mask;
   public:
-    SIMD (size_t i)
+    SIMD (int64_t i)
       : mask(i > 0 ? -1 : 0) { ; }
     bool Data() const { return mask; }
     static constexpr int Size() { return 1; }
@@ -76,7 +76,7 @@ namespace ngcore
     SIMD<mask64,N2> hi;
   public:
 
-    SIMD (size_t i) : lo(i), hi(i>N1 ? i-N1 : 0) { ; }
+    SIMD (int64_t i) : lo(i), hi(i-N1 ) { ; }
     SIMD (SIMD<mask64,N1> lo_, SIMD<mask64,N2> hi_) : lo(lo_), hi(hi_) { ; }
     SIMD<mask64,N1> Lo() const { return lo; }
     SIMD<mask64,N2> Hi() const { return hi; }