From d25345aa07ef7ca19b02c545642b868f28be565d Mon Sep 17 00:00:00 2001
From: scorpioblood <77296181+scorpioblood@users.noreply.github.com>
Date: Tue, 28 May 2024 22:01:51 +0200
Subject: [PATCH] More SIMD.
---
.../Core/public/Math/SIMD/PhanesSIMDTypes.h | 2 +
.../public/Math/SIMD/PhanesVectorMathSSE.hpp | 74 ++++++++++++-
.../Runtime/Core/public/Math/SIMD/Platform.h | 24 +++-
.../Runtime/Core/public/Math/Vector4.hpp | 12 +-
.../Runtime/Core/public/Math/Vector4.inl | 103 +++++++++++++++++-
5 files changed, 193 insertions(+), 22 deletions(-)
diff --git a/Engine/Source/Runtime/Core/public/Math/SIMD/PhanesSIMDTypes.h b/Engine/Source/Runtime/Core/public/Math/SIMD/PhanesSIMDTypes.h
index b614e0c..6ff9116 100644
--- a/Engine/Source/Runtime/Core/public/Math/SIMD/PhanesSIMDTypes.h
+++ b/Engine/Source/Runtime/Core/public/Math/SIMD/PhanesSIMDTypes.h
@@ -1,3 +1,5 @@
+#pragma once
+
// This file includes the necessary header for vectorization intrinsics. If no specifics are defined SSE4.2 is used.
//
// ARM is not supported.
diff --git a/Engine/Source/Runtime/Core/public/Math/SIMD/PhanesVectorMathSSE.hpp b/Engine/Source/Runtime/Core/public/Math/SIMD/PhanesVectorMathSSE.hpp
index 5901cb3..f0815c7 100644
--- a/Engine/Source/Runtime/Core/public/Math/SIMD/PhanesVectorMathSSE.hpp
+++ b/Engine/Source/Runtime/Core/public/Math/SIMD/PhanesVectorMathSSE.hpp
@@ -1,10 +1,30 @@
#pragma once
-#include "Core/public/Math/Boilerplate.h"
#include <nmmintrin.h>
+#include "Core/public/Math/SIMD/PhanesSIMDTypes.h"
+#include "Core/public/Math/Boilerplate.h"
+#include "Core/public/Math/MathCommon.hpp"
+
+#include <iostream>
+
// -> For IntelliSense
+#include "Core/public/Math/Vector4.hpp"
+
+// ========== //
+// Common //
+// ========== //
+
+
+Phanes::Core::Types::Vec4f32Reg p_vec4_abs(const Phanes::Core::Types::Vec4f32Reg& v)
+{
+ return _mm_and_ps(v, _mm_castsi128_ps(_mm_set1_epi32(0x7FFFFFFF)));
+}
+
+
+
+
// ============ //
// TVector4 //
// ============ //
@@ -22,4 +42,54 @@ namespace Phanes::Core::Math::Detail
r.comp = _mm_add_ps(v1.comp, v2.comp);
}
};
-}
\ No newline at end of file
+
+ template<>
+ struct compute_vec4_sub<float, true>
+ {
+ static FORCEINLINE void map(Phanes::Core::Math::TVector4<float, true>& r, const Phanes::Core::Math::TVector4<float, true>& v1, const Phanes::Core::Math::TVector4<float, true>& v2)
+ {
+ r.comp = _mm_sub_ps(v1.comp, v2.comp);
+ }
+ };
+
+ template<>
+ struct compute_vec4_mul<float, true>
+ {
+ static FORCEINLINE void map(Phanes::Core::Math::TVector4<float, true>& r, const Phanes::Core::Math::TVector4<float, true>& v1, const Phanes::Core::Math::TVector4<float, true>& v2)
+ {
+ r.comp = _mm_mul_ps(v1.comp, v2.comp);
+ }
+ };
+
+ template<>
+ struct compute_vec4_div<float, true>
+ {
+ static FORCEINLINE void map(Phanes::Core::Math::TVector4<float, true>& r, const Phanes::Core::Math::TVector4<float, true>& v1, const Phanes::Core::Math::TVector4<float, true>& v2)
+ {
+ r.comp = _mm_div_ps(v1.comp, v2.comp);
+ }
+ };
+
+ template<>
+ struct compute_vec4_eq<float, true>
+ {
+ static FORCEINLINE bool map(const Phanes::Core::Math::TVector4<float, true>& v1, const Phanes::Core::Math::TVector4<float, true>& v2)
+ {
+ float r;
+ _mm_store_ps1(&r, _mm_cmpeq_ps(v1.comp, v2.comp));
+ return (r == 0xffffffff) ? true : false;
+ }
+ };
+
+ template<>
+ struct compute_vec4_ieq<float, true>
+ {
+ static FORCEINLINE bool map(const Phanes::Core::Math::TVector4<float, true>& v1, const Phanes::Core::Math::TVector4<float, true>& v2)
+ {
+ float r;
+ _mm_store_ps1(&r, _mm_cmpneq_ps(v1.comp, v2.comp));
+ return (r == 0xffffffff) ? true : false;
+ }
+ };
+}
+
diff --git a/Engine/Source/Runtime/Core/public/Math/SIMD/Platform.h b/Engine/Source/Runtime/Core/public/Math/SIMD/Platform.h
index c15b890..a2b8b61 100644
--- a/Engine/Source/Runtime/Core/public/Math/SIMD/Platform.h
+++ b/Engine/Source/Runtime/Core/public/Math/SIMD/Platform.h
@@ -254,11 +254,6 @@
// Define also supported instruction sets for Visual Studio, as it only defines the latest (e.g. only __AVX__ not __SSE4__ ...).
-#define P_AVX2__ 0
-#define P_AVX__ 0
-#define P_SSE__ 0
-#define P_NEON__ 0
-
#ifdef P_FORCE_INTRINSICS
# undef __AVX2__
@@ -272,7 +267,7 @@
#else
# ifdef __AVX2__
-# define P_AVX2__ 1
+# define P_AVX2__ 1
# elif defined(__AVX__)
# define P_AVX__ 1
# elif defined(__SSE__)
@@ -290,6 +285,23 @@
# define P_SSE__ 1
#endif
+// Deactivate unset SIMD
+#ifndef P_AVX2__
+# define P_AVX2__ 0
+#endif
+
+// Deactivate unset SIMD
+#ifndef P_AVX__
+# define P_AVX__ 0
+#endif
+
+#ifndef P_SSE__
+# define P_SSE__ 0
+#endif
+
+#ifndef P_NEON__
+# define P_NEON__ 0
+#endif
#define P_INTRINSICS_FPU 0
#define P_INTRINSICS_SSE 1
diff --git a/Engine/Source/Runtime/Core/public/Math/Vector4.hpp b/Engine/Source/Runtime/Core/public/Math/Vector4.hpp
index 63bcd2f..81f771d 100644
--- a/Engine/Source/Runtime/Core/public/Math/Vector4.hpp
+++ b/Engine/Source/Runtime/Core/public/Math/Vector4.hpp
@@ -90,13 +90,6 @@ namespace Phanes::Core::Math
/// <param name="comp">Array of at least 4 components</param>
TVector4(const Real* comp);
- /// <summary>
- /// Construct the vector, by calculating the way between two points.
- /// </summary>
- /// <param name="start">Starting point of the vector.</param>
- /// <param name="end">End point of the vector.</param>
- TVector4(const TPoint4<Real>& start, const TPoint4<Real>& end);
-
};
// ===================== //
@@ -698,8 +691,5 @@ namespace Phanes::Core::Math
TVector4<T, A> PrespectiveDivideV(TVector4<T, A>& v1);
}
-// No SIMD
-#include "Core/public/Math/Vector4.inl"
-// SIMD
-#include "Core/public/Math/SIMD/SIMDIntrinsics.h"
\ No newline at end of file
+#include "Core/public/Math/Vector4.inl"
diff --git a/Engine/Source/Runtime/Core/public/Math/Vector4.inl b/Engine/Source/Runtime/Core/public/Math/Vector4.inl
index 4f2cfcc..f419752 100644
--- a/Engine/Source/Runtime/Core/public/Math/Vector4.inl
+++ b/Engine/Source/Runtime/Core/public/Math/Vector4.inl
@@ -3,8 +3,13 @@
#include "Core/public/Math/Boilerplate.h"
#include "Core/public/Math/Detail/Vector4Decl.inl"
+#include "Core/public/Math/SIMD/SIMDIntrinsics.h"
+
#include "Core/public/Math/Vector4.hpp"
+
+#include "Core/public/Math/SIMD/PhanesSIMDTypes.h"
+
#include <stdio.h>
namespace Phanes::Core::Math
@@ -25,6 +30,30 @@ namespace Phanes::Core::Math
w(_w)
{}
+ template<RealType T, bool A>
+ Phanes::Core::Math::TVector4<T, A>::TVector4(Real s) :
+ x(s),
+ y(s),
+ z(s),
+ w(s)
+ {}
+
+ template<RealType T, bool A>
+ Phanes::Core::Math::TVector4<T, A>::TVector4(const TVector2<Real>& v1, const TVector2<Real>& v2) :
+ x(v1.x),
+ y(v1.y),
+ z(v2.x),
+ w(v2.y)
+ {}
+
+ template<RealType T, bool A>
+ Phanes::Core::Math::TVector4<T, A>::TVector4(const Real* comp) :
+ x(comp[0]),
+ y(comp[1]),
+ z(comp[2]),
+ w(comp[3])
+ {}
+
template<RealType T, bool A>
TVector4<T, A> operator+=(TVector4<T, A>& v1, const TVector4<T, A>& v2)
{
@@ -159,19 +188,87 @@ namespace Phanes::Core::Math
return Detail::compute_vec4_ieq<T, A>::map(v1, v2);
}
+
+
+ // Inc- / Decrement
+ template<RealType T, bool A>
+ TVector4<T, A>& operator++(TVector4<T, A>& v1)
+ {
+ ++v1.x;
+ ++v1.y;
+ ++v1.z;
+ ++v1.w;
- // SIMD
+ return v1;
+ }
+ template<RealType T, bool A>
+ TVector4<T, A>& operator--(TVector4<T, A>& v1)
+ {
+ --v1.x;
+ --v1.y;
+ --v1.z;
+ --v1.w;
+
+ return v1;
+ }
+
+ template<RealType T, bool A>
+ TVector4<T, A>& operator++(TVector4<T, A>& v1, int)
+ {
+ return ++v1;
+ }
+
+ template<RealType T, bool A>
+ TVector4<T, A>& operator--(TVector4<T, A>& v1, int)
+ {
+ return --v1;
+ }
+
+
+ // SIMD constructor
template<>
- TVector4<float, true>::TVector4(Real _x, Real _y, Real _z, Real _w) :
+ TVector4<float, true>::TVector4(const TVector4<float, true>& v)
+ {
+ this->comp = _mm_load_ps(reinterpret_cast<const float*>(&v));
+ }
+
+ template<>
+ TVector4<float, true>::TVector4(float _x, float _y, float _z, float _w) :
x(_x),
y(_y),
z(_z),
w(_w)
{
- this->comp = _mm_load_ps(reinterpret_cast<float*>(&x));
+ this->comp = _mm_load_ps(reinterpret_cast<float*>(&this->x));
+ }
+
+ template<>
+ TVector4<float, true>::TVector4(float s)
+ {
+ this->comp = _mm_load_ps1(&s);
+ }
+
+ template<>
+ TVector4<float, true>::TVector4(const TVector2<float>& v1, const TVector2<float>& v2) :
+ x(v1.x),
+ y(v1.y),
+ z(v2.x),
+ w(v2.y)
+ {
+ this->comp = _mm_load_ps(reinterpret_cast<float*>(&this->x));
+ }
+
+ template<>
+ TVector4<float, true>::TVector4(const float* comp) :
+ x(comp[0]),
+ y(comp[1]),
+ z(comp[2]),
+ w(comp[3])
+ {
+ this->comp = _mm_load_ps(reinterpret_cast<float*>(&this->x));
}
}
\ No newline at end of file