sse2neon: update library

fix: double calculation of angle vectors with sse instructions fix: cr::abs behavior for floating point arguments
yapb · Oct 27, 2024 · 63cfde1 · 63cfde1
1 parent c0d1357
commit 63cfde1
Show file tree

Hide file tree

Showing 7 changed files with 413 additions and 304 deletions.
diff --git a/crlib/basic.h b/crlib/basic.h
@@ -25,10 +25,6 @@ template <typename T, size_t N> constexpr size_t bufsize (const T (&)[N]) {
    return N - 1;
 }
 
-template <typename T> constexpr T abs (const T &a) {
-   return a > 0 ? a : -a;
-}
-
 template <typename T> constexpr T bit (const T &a) {
    return static_cast <T> (1ULL << a);
 }
@@ -41,7 +37,6 @@ template <typename T> constexpr T max (const T &a, const T &b) {
    return a > b ? a : b;
 }
 
-
 template <typename T> constexpr T clamp (const T &x, const T &a, const T &b) {
    return min (max (x, a), b);
 }

diff --git a/crlib/hashmap.h b/crlib/hashmap.h
@@ -141,7 +141,7 @@ template <typename K, typename V, typename H = Hash <K>> class HashMap {
 
       ++length_;
 
-      const auto length = static_cast<float> (length_);
+      const auto length = static_cast <float> (length_);
       const auto capacity = static_cast <float> (contents_.length ());
 
       if (!contents_.empty () && length / capacity >= kLoadFactor) {

diff --git a/crlib/lambda.h b/crlib/lambda.h
@@ -34,7 +34,7 @@ template <typename R, typename ...Args> class Lambda <R (Args...)> final {
 
    public:
       constexpr LambdaFunctor (const LambdaFunctor &rhs) : callee_ { rhs.callee_ } {}
-      constexpr LambdaFunctor (LambdaFunctor &&rhs) : callee_ { cr::move (rhs.callee_) } {}
+      constexpr LambdaFunctor (LambdaFunctor &&rhs) noexcept : callee_ { cr::move (rhs.callee_) } {}
       constexpr LambdaFunctor (const T &callee) : callee_ { callee } {}
       constexpr LambdaFunctor (T &&callee) : callee_ { cr::move (callee) } {}
 

diff --git a/crlib/mathlib.h b/crlib/mathlib.h
@@ -24,9 +24,20 @@ constexpr float kRadiansToDegree = 180.0f / kMathPi;
 CR_NAMESPACE_END
 
 #include <crlib/simd.h>
+#include <crlib/traits.h>
 
 CR_NAMESPACE_BEGIN
 
+// abs() overload
+template <typename T> constexpr T abs (const T &a) {
+   if constexpr (is_same <T, float>::value) {
+      return ::fabs (a);
+   }
+   else if constexpr (is_same <T, int>::value) {
+      return ::abs (a);
+   }
+}
+
 #if defined (CR_HAS_SIMD_SSE)
 template <> CR_FORCE_INLINE float abs (const float &x) {
    return _mm_cvtss_f32 (simd::fabs_ps (_mm_load_ss (&x)));
@@ -214,7 +225,7 @@ namespace detail {
 #endif
 
    template <int D> CR_FORCE_INLINE float scalar_wrapAngleFn (float x) {
-      return x - 2.0f * static_cast <float> (D) * cr::floorf (x / (2.0f * static_cast <float> (D)) + 0.5f);
+      return x - 2.0f * static_cast <float> (D) * ::floorf (x / (2.0f * static_cast <float> (D)) + 0.5f);
    }
 
    template <int D> CR_FORCE_INLINE float _wrapAngleFn (float x) {

diff --git a/crlib/platform.h b/crlib/platform.h
@@ -250,6 +250,7 @@ struct Platform : public Singleton <Platform> {
    bool x64 = false;
    bool arm = false;
    bool ppc = false;
+   bool simd = false;
 
    char appName[64] = {};
 
@@ -281,6 +282,10 @@ struct Platform : public Singleton <Platform> {
 #if defined(CR_ARCH_PPC)
       ppc = true;
 #endif
+
+#if !defined(CR_DISABLE_SIMD)
+      simd = true;
+#endif
    }
 
    // set the app name
@@ -304,7 +309,7 @@ struct Platform : public Singleton <Platform> {
       MEMORY_BASIC_INFORMATION mbi {};
 
       if (VirtualQuery (reinterpret_cast <LPVOID> (ptr), &mbi, sizeof (mbi))) {
-         auto result = (mbi.Protect & (PAGE_READONLY | PAGE_READWRITE | PAGE_WRITECOPY | PAGE_EXECUTE_READ | PAGE_EXECUTE_READWRITE | PAGE_EXECUTE_WRITECOPY));
+         auto result = !!(mbi.Protect & (PAGE_READONLY | PAGE_READWRITE | PAGE_WRITECOPY | PAGE_EXECUTE_READ | PAGE_EXECUTE_READWRITE | PAGE_EXECUTE_WRITECOPY));
 
          if (mbi.Protect & (PAGE_GUARD | PAGE_NOACCESS)) {
             result = false;