simd: have a "structure of arrays" layout for all spheres data, chang…

…e HitSphere -> HitSpheres - PC 171 -> 184 Mray/s - Mac 42.6 -> 48.1 Mray/s
aras-p · Apr 10, 2018 · b4ef99d · b4ef99d
1 parent bcfca46
commit b4ef99d
Show file tree

Hide file tree

Showing 3 changed files with 81 additions and 41 deletions.
diff --git a/Cpp/Source/Maths.cpp b/Cpp/Source/Maths.cpp
@@ -47,34 +47,46 @@ float3 RandomUnitVector(uint32_t& state)
 }
 
 
-bool HitSphere(const Ray& r, const Sphere& s, float tMin, float tMax, Hit& outHit)
+int HitSpheres(const Ray& r, const SpheresSoA& spheres, float tMin, float tMax, Hit& outHit)
 {
-    assert(s.invRadius == 1.0f/s.radius);
-    AssertUnit(r.dir);
-    float3 oc = r.orig - s.center;
-    float b = dot(oc, r.dir);
-    float c = dot(oc, oc) - s.radius*s.radius;
-    float discr = b*b - c;
-    if (discr > 0)
+    Hit tmpHit;
+    int id = -1;
+    float closest = tMax;
+    for (int i = 0; i < spheres.count; ++i)
     {
-        float discrSq = sqrtf(discr);
-
-        float t = (-b - discrSq);
-        if (t < tMax && t > tMin)
+        float ocX = r.orig.getX() - spheres.centerX[i];
+        float ocY = r.orig.getY() - spheres.centerY[i];
+        float ocZ = r.orig.getZ() - spheres.centerZ[i];
+        float b = ocX * r.dir.getX() + ocY * r.dir.getY() + ocZ * r.dir.getZ();
+        float c = ocX * ocX + ocY * ocY + ocZ * ocZ - spheres.radius[i] * spheres.radius[i];
+        float discr = b * b - c;
+        if (discr > 0)
         {
-            outHit.pos = r.pointAt(t);
-            outHit.normal = (outHit.pos - s.center) * s.invRadius;
-            outHit.t = t;
-            return true;
-        }
-        t = (-b + discrSq);
-        if (t < tMax && t > tMin)
-        {
-            outHit.pos = r.pointAt(t);
-            outHit.normal = (outHit.pos - s.center) * s.invRadius;
-            outHit.t = t;
-            return true;
+            float discrSq = sqrtf(discr);
+
+            float t = (-b - discrSq);
+            if (t > tMin && t < tMax && t < closest)
+            {
+                id = i;
+                closest = t;
+                tmpHit.pos = r.pointAt(t);
+                tmpHit.normal = (tmpHit.pos - float3(spheres.centerX[i], spheres.centerY[i], spheres.centerZ[i])) * spheres.invRadius[i];
+                tmpHit.t = t;
+            }
+            else
+            {
+                t = (-b + discrSq);
+                if (t > tMin && t < tMax && t < closest)
+                {
+                    id = i;
+                    closest = t;
+                    tmpHit.pos = r.pointAt(t);
+                    tmpHit.normal = (tmpHit.pos - float3(spheres.centerX[i], spheres.centerY[i], spheres.centerZ[i])) * spheres.invRadius[i];
+                    tmpHit.t = t;
+                }
+            }
         }
     }
-    return false;
+    outHit = tmpHit;
+    return id;
 }
diff --git a/Cpp/Source/Maths.h b/Cpp/Source/Maths.h
@@ -222,7 +222,37 @@ struct Sphere
 };
 
 
-bool HitSphere(const Ray& r, const Sphere& s, float tMin, float tMax, Hit& outHit);
+// data for all spheres in a "structure of arrays" layout
+struct SpheresSoA
+{
+    SpheresSoA(int c)
+        : count(c)
+    {
+        centerX = new float[c];
+        centerY = new float[c];
+        centerZ = new float[c];
+        radius = new float[c];
+        invRadius = new float[c];
+    }
+    ~SpheresSoA()
+    {
+        delete[] centerX;
+        delete[] centerY;
+        delete[] centerZ;
+        delete[] radius;
+        delete[] invRadius;
+    }
+
+    float* centerX;
+    float* centerY;
+    float* centerZ;
+    float* radius;
+    float* invRadius;
+    int count;
+};
+
+
+int HitSpheres(const Ray& r, const SpheresSoA& spheres, float tMin, float tMax, Hit& outHit);
 
 float RandomFloat01(uint32_t& state);
 float3 RandomInUnitDisk(uint32_t& state);

diff --git a/Cpp/Source/Test.cpp b/Cpp/Source/Test.cpp
@@ -19,6 +19,8 @@ static Sphere s_Spheres[] =
 };
 const int kSphereCount = sizeof(s_Spheres) / sizeof(s_Spheres[0]);
 
+static SpheresSoA s_SpheresSoA(kSphereCount);
+
 struct Material
 {
     enum Type { Lambert, Metal, Dielectric };
@@ -51,20 +53,8 @@ const int kMaxDepth = 10;
 
 bool HitWorld(const Ray& r, float tMin, float tMax, Hit& outHit, int& outID)
 {
-    Hit tmpHit;
-    bool anything = false;
-    float closest = tMax;
-    for (int i = 0; i < kSphereCount; ++i)
-    {
-        if (HitSphere(r, s_Spheres[i], tMin, closest, tmpHit))
-        {
-            anything = true;
-            closest = tmpHit.t;
-            outHit = tmpHit;
-            outID = i;
-        }
-    }
-    return anything;
+    outID = HitSpheres(r, s_SpheresSoA, tMin, tMax, outHit);
+    return outID != -1;
 }
 
 
@@ -295,7 +285,15 @@ void UpdateTest(float time, int frameCount, int screenWidth, int screenHeight)
 #endif
 
     for (int i = 0; i < kSphereCount; ++i)
-        s_Spheres[i].UpdateDerivedData();
+    {
+        Sphere& s = s_Spheres[i];
+        s.UpdateDerivedData();
+        s_SpheresSoA.centerX[i] = s.center.getX();
+        s_SpheresSoA.centerY[i] = s.center.getY();
+        s_SpheresSoA.centerZ[i] = s.center.getZ();
+        s_SpheresSoA.radius[i] = s.radius;
+        s_SpheresSoA.invRadius[i] = s.invRadius;
+    }
 
     s_Cam = Camera(lookfrom, lookat, float3(0, 1, 0), 60, float(screenWidth) / float(screenHeight), aperture, distToFocus);
 }