From 1161a4114dde2fe0de23cc920a64c64c06f171f4 Mon Sep 17 00:00:00 2001
From: Wojtek Figat <wojtek@figat.pl>
Date: Thu, 26 Sep 2024 18:56:22 +0200
Subject: [PATCH] Minor tweaks to simd codebase

---
 Source/Engine/Core/Math/BoundingFrustum.h | 18 +++++-----
 Source/Engine/Core/SIMD.h                 | 40 ++++++++++++++++++-----
 Source/Engine/Platform/Defines.h          |  1 +
 3 files changed, 41 insertions(+), 18 deletions(-)

diff --git a/Source/Engine/Core/Math/BoundingFrustum.h b/Source/Engine/Core/Math/BoundingFrustum.h
index f98bcafae..554019079 100644
--- a/Source/Engine/Core/Math/BoundingFrustum.h
+++ b/Source/Engine/Core/Math/BoundingFrustum.h
@@ -101,7 +101,7 @@ public:
     }
 
     /// <summary>
-    /// Gets the far plane of the BoundingFrustum.
+    /// Gets the far plane of the frustum.
     /// </summary>
     FORCE_INLINE Plane GetFar() const
     {
@@ -109,7 +109,7 @@ public:
     }
 
     /// <summary>
-    /// Gets the left plane of the BoundingFrustum.
+    /// Gets the left plane of the frustum.
     /// </summary>
     FORCE_INLINE Plane GetLeft() const
     {
@@ -117,7 +117,7 @@ public:
     }
 
     /// <summary>
-    /// Gets the right plane of the BoundingFrustum.
+    /// Gets the right plane of the frustum.
     /// </summary>
     FORCE_INLINE Plane GetRight() const
     {
@@ -125,7 +125,7 @@ public:
     }
 
     /// <summary>
-    /// Gets the top plane of the BoundingFrustum.
+    /// Gets the top plane of the frustum.
     /// </summary>
     FORCE_INLINE Plane GetTop() const
     {
@@ -133,7 +133,7 @@ public:
     }
 
     /// <summary>
-    /// Gets the bottom plane of the BoundingFrustum.
+    /// Gets the bottom plane of the frustum.
     /// </summary>
     FORCE_INLINE Plane GetBottom() const
     {
@@ -230,17 +230,17 @@ public:
     ContainmentType Contains(const BoundingSphere& sphere) const;
 
     /// <summary>
-    /// Checks whether the current BoundingFrustum intersects a BoundingSphere.
+    /// Checks whether the current frustum intersects a sphere.
     /// </summary>
     /// <param name="sphere">The sphere.</param>
-    /// <returns>True if the current BoundingFrustum intersects a BoundingSphere, otherwise false.</returns>
+    /// <returns>True if the current frustum intersects a sphere, otherwise false.</returns>
     bool Intersects(const BoundingSphere& sphere) const;
 
     /// <summary>
-    /// Checks whether the current BoundingFrustum intersects a BoundingBox.
+    /// Checks whether the current frustum intersects a box.
     /// </summary>
     /// <param name="box">The box</param>
-    /// <returns>True if the current BoundingFrustum intersects a BoundingBox, otherwise false.</returns>
+    /// <returns>True if the current frustum intersects a box, otherwise false.</returns>
     FORCE_INLINE bool Intersects(const BoundingBox& box) const
     {
         return CollisionsHelper::FrustumContainsBox(*this, box) != ContainmentType::Disjoint;
diff --git a/Source/Engine/Core/SIMD.h b/Source/Engine/Core/SIMD.h
index 0ce58f7e7..b1d0db3a7 100644
--- a/Source/Engine/Core/SIMD.h
+++ b/Source/Engine/Core/SIMD.h
@@ -13,7 +13,7 @@
 
 #if PLATFORM_SIMD_SSE2
 
-// Vector of four floating point values stored in vector register.
+// Vector of four floating point values stored in a vector register.
 typedef __m128 SimdVector4;
 
 namespace SIMD
@@ -28,9 +28,15 @@ namespace SIMD
         return _mm_set_ps(w, z, y, x);
     }
 
-    FORCE_INLINE SimdVector4 Load(const void* src)
+    FORCE_INLINE SimdVector4 Load(const float* __restrict src)
     {
-        return _mm_load_ps((const float*)(src));
+        return _mm_loadu_ps(src);
+    }
+
+    FORCE_INLINE SimdVector4 LoadAligned(const float* __restrict src)
+    {
+        ASSERT_LOW_LAYER(((uintptr)src & 15) == 0);
+        return _mm_load_ps(src);
     }
 
     FORCE_INLINE SimdVector4 Splat(float value)
@@ -38,9 +44,15 @@ namespace SIMD
         return _mm_set_ps1(value);
     }
 
-    FORCE_INLINE void Store(void* dst, SimdVector4 src)
+    FORCE_INLINE void Store(float* __restrict dst, SimdVector4 src)
     {
-        _mm_store_ps((float*)dst, src);
+        _mm_storeu_ps(dst, src);
+    }
+
+    FORCE_INLINE void StoreAligned(float* __restrict dst, SimdVector4 src)
+    {
+        ASSERT_LOW_LAYER(((uintptr)dst & 15) == 0);
+        _mm_store_ps(dst, src);
     }
 
     FORCE_INLINE int MoveMask(SimdVector4 a)
@@ -113,7 +125,12 @@ namespace SIMD
 		return { x, y, z, w };
     }
 
-	FORCE_INLINE SimdVector4 Load(const void* src)
+	FORCE_INLINE SimdVector4 Load(const float* __restrict src)
+	{
+		return *(const SimdVector4*)src;
+	}
+
+	FORCE_INLINE SimdVector4 LoadAligned(const float* __restrict src)
 	{
 		return *(const SimdVector4*)src;
 	}
@@ -123,10 +140,15 @@ namespace SIMD
 		return { value, value, value, value };
 	}
 
-	FORCE_INLINE void Store(void* dst, SimdVector4 src)
-	{
+    FORCE_INLINE void Store(float* __restrict dst, SimdVector4 src)
+    {
 		(*(SimdVector4*)dst) = src;
-	}
+    }
+
+    FORCE_INLINE void StoreAligned(float* __restrict dst, SimdVector4 src)
+    {
+		(*(SimdVector4*)dst) = src;
+    }
 
 	FORCE_INLINE int MoveMask(SimdVector4 a)
 	{
diff --git a/Source/Engine/Platform/Defines.h b/Source/Engine/Platform/Defines.h
index 40b250983..1f12f63d2 100644
--- a/Source/Engine/Platform/Defines.h
+++ b/Source/Engine/Platform/Defines.h
@@ -228,6 +228,7 @@ API_ENUM() enum class ArchitectureType
 #if defined(_M_PPC) || defined(__CELLOS_LV2__)
 #define PLATFORM_SIMD_VMX 1
 #endif
+#define PLATFORM_SIMD (PLATFORM_SIMD_SSE2 || PLATFORM_SIMD_SSE3 || PLATFORM_SIMD_SSE4 || PLATFORM_SIMD_NEON || PLATFORM_SIMD_VMX)
 
 // Unicode text macro
 #if !defined(TEXT)