From 0ba14879e0e2beb18113276d11edeb23dc2c7d6e Mon Sep 17 00:00:00 2001
From: assiduous <assiduous@diligentgraphics.com>
Date: Fri, 10 Apr 2020 17:02:02 -0700
Subject: MathLib: added FastFloor and FastCeil functions

---
 Common/interface/BasicMath.hpp      | 81 ++++++++++++++++++++++++++++++++++++-
 Common/interface/FastRand.hpp       |  2 +-
 Common/interface/FilteringTools.hpp |  2 +-
 3 files changed, 82 insertions(+), 3 deletions(-)

(limited to 'Common/interface')

diff --git a/Common/interface/BasicMath.hpp b/Common/interface/BasicMath.hpp
index 867c91cd..757244e1 100644
--- a/Common/interface/BasicMath.hpp
+++ b/Common/interface/BasicMath.hpp
@@ -1942,6 +1942,86 @@ inline Uint32 F4Color_To_RGBA8Unorm(const float4& f4Color)
     return RGBA8U;
 }
 
+
+// At least on MSVC std::floor is an actual function call into ucrtbase.dll.
+// For floats/doubles that fit into Int64 representable range we can do much better.
+template <typename T>
+T FastFloor(T x)
+{
+    auto i   = static_cast<Int64>(x);
+    auto flr = static_cast<T>(i);
+    //   x         flr    floor(x)  flr <= x
+    // +1.0   ->   1.0      1.0       true
+    // +0.5   ->   0.0      0.0       true
+    //  0.0   ->   0.0      0.0       true
+    // -0.5   ->   0.0     -1.0      false
+    // -1.0   ->  -1.0     -1.0       true
+
+    return flr <= x ? flr : flr - 1;
+}
+
+template <typename T>
+T FastCeil(T x)
+{
+    return -FastFloor(-x);
+}
+
+
+template <typename T>
+Diligent::Vector2<T> FastFloor(const Diligent::Vector2<T>& vec)
+{
+    return Diligent::Vector2<T>{
+        FastFloor(vec.x),
+        FastFloor(vec.y)};
+}
+
+template <typename T>
+Diligent::Vector3<T> FastFloor(const Diligent::Vector3<T>& vec)
+{
+    return Diligent::Vector3<T>{
+        FastFloor(vec.x),
+        FastFloor(vec.y),
+        FastFloor(vec.z)};
+}
+
+template <typename T>
+Diligent::Vector4<T> FastFloor(const Diligent::Vector4<T>& vec)
+{
+    return Diligent::Vector4<T>{
+        FastFloor(vec.x),
+        FastFloor(vec.y),
+        FastFloor(vec.z),
+        FastFloor(vec.w)};
+}
+
+
+template <typename T>
+Diligent::Vector2<T> FastCeil(const Diligent::Vector2<T>& vec)
+{
+    return Diligent::Vector2<T>{
+        FastCeil(vec.x),
+        FastCeil(vec.y)};
+}
+
+template <typename T>
+Diligent::Vector3<T> FastCeil(const Diligent::Vector3<T>& vec)
+{
+    return Diligent::Vector3<T>{
+        FastCeil(vec.x),
+        FastCeil(vec.y),
+        FastCeil(vec.z)};
+}
+
+template <typename T>
+Diligent::Vector4<T> FastCeil(const Diligent::Vector4<T>& vec)
+{
+    return Diligent::Vector4<T>{
+        FastCeil(vec.x),
+        FastCeil(vec.y),
+        FastCeil(vec.z),
+        FastCeil(vec.w)};
+}
+
 } // namespace Diligent
 
 
@@ -2003,7 +2083,6 @@ Diligent::Vector4<T> min(const Diligent::Vector4<T>& Left, const Diligent::Vecto
         std::min(Left.w, Right.w));
 }
 
-
 template <typename T>
 Diligent::Vector2<T> floor(const Diligent::Vector2<T>& vec)
 {
diff --git a/Common/interface/FastRand.hpp b/Common/interface/FastRand.hpp
index 2494fc83..546da633 100644
--- a/Common/interface/FastRand.hpp
+++ b/Common/interface/FastRand.hpp
@@ -93,7 +93,7 @@ public:
         Range{_Max - _Min + 1}
     {
         VERIFY_EXPR(_Max > _Min);
-        VERIFY(Range < FastRand::Max / 8, "Range is too large");
+        VERIFY(Range <= FastRand::Max, "Range is too large");
     }
 
     Type operator()()
diff --git a/Common/interface/FilteringTools.hpp b/Common/interface/FilteringTools.hpp
index e256f88d..73871093 100644
--- a/Common/interface/FilteringTools.hpp
+++ b/Common/interface/FilteringTools.hpp
@@ -100,7 +100,7 @@ template <TEXTURE_ADDRESS_MODE AddressMode, bool IsNormalizedCoord>
 LinearTexFilterSampleInfo GetLinearTexFilterSampleInfo(Uint32 Width, float u)
 {
     float x  = IsNormalizedCoord ? u * static_cast<float>(Width) : u;
-    float x0 = std::floor(x - 0.5f);
+    float x0 = FastFloor(x - 0.5f);
 
     // clang-format off
     LinearTexFilterSampleInfo SampleInfo
-- 
cgit v1.2.3