From 4ea8d2d92af80d64a16b3d20f899b22921aa42d0 Mon Sep 17 00:00:00 2001 From: ladevieq Date: Fri, 20 Aug 2021 22:53:57 +0200 Subject: Use simd for vec3 computations --- include/aabb.hpp | 53 +++++++++++++++++++++++++++++----------------------- include/bvh-node.hpp | 12 ++++++------ include/utils.hpp | 1 + include/vec3.hpp | 41 ++++++++++++++++++++-------------------- 4 files changed, 58 insertions(+), 49 deletions(-) (limited to 'include') diff --git a/include/aabb.hpp b/include/aabb.hpp index cdbb79a..5f819db 100644 --- a/include/aabb.hpp +++ b/include/aabb.hpp @@ -16,41 +16,47 @@ class aabb { auto &v2 = tri.v2.position; auto &v3 = tri.v3.position; - minimum.x = fmin(fmin(v1.x, v2.x), v3.x); - minimum.y = fmin(fmin(v1.y, v2.y), v3.y); - minimum.z = fmin(fmin(v1.z, v2.z), v3.z); + // minimum.x = fmin(fmin(v1.x, v2.x), v3.x); + // minimum.y = fmin(fmin(v1.y, v2.y), v3.y); + // minimum.z = fmin(fmin(v1.z, v2.z), v3.z); - maximum.x = fmax(fmax(v1.x, v2.x), v3.x); - maximum.y = fmax(fmax(v1.y, v2.y), v3.y); - maximum.z = fmax(fmax(v1.z, v2.z), v3.z); + minimum.v[0] = fmin(fmin(v1.v[0], v2.v[0]), v3.v[0]); + minimum.v[1] = fmin(fmin(v1.v[1], v2.v[1]), v3.v[1]); + minimum.v[2] = fmin(fmin(v1.v[2], v2.v[2]), v3.v[2]); + + // maximum.x = fmax(fmax(v1.x, v2.x), v3.x); + // maximum.y = fmax(fmax(v1.y, v2.y), v3.y); + // maximum.z = fmax(fmax(v1.z, v2.z), v3.z); + + maximum.v[0] = fmax(fmax(v1.v[0], v2.v[0]), v3.v[0]); + maximum.v[1] = fmax(fmax(v1.v[1], v2.v[1]), v3.v[1]); + maximum.v[2] = fmax(fmax(v1.v[2], v2.v[2]), v3.v[2]); } void union_with(const aabb &box) { - minimum.x = minimum.x < box.minimum.x ? minimum.x : box.minimum.x; - minimum.y = minimum.y < box.minimum.y ? minimum.y : box.minimum.y; - minimum.z = minimum.z < box.minimum.z ? minimum.z : box.minimum.z; - - maximum.x = maximum.x > box.maximum.x ? maximum.x : box.maximum.x; - maximum.y = maximum.y > box.maximum.y ? maximum.y : box.maximum.y; - maximum.z = maximum.z > box.maximum.z ? maximum.z : box.maximum.z; + minimum.min(box.minimum); + maximum.max(box.maximum); } void union_with(const point3 &point) { - minimum.x = minimum.x < point.x ? minimum.x : point.x; - minimum.y = minimum.y < point.y ? minimum.y : point.y; - minimum.z = minimum.z < point.z ? minimum.z : point.z; - - maximum.x = maximum.x > point.x ? maximum.x : point.x; - maximum.y = maximum.y > point.y ? maximum.y : point.y; - maximum.z = maximum.z > point.z ? maximum.z : point.z; + minimum.min(point); + maximum.max(point); } int32_t maximum_axis() { vec3 diag = diagonal(); - if (diag.x > diag.y && diag.x > diag.z) { + // if (diag.x > diag.y && diag.x > diag.z) { + // return 0; + // } else if (diag.y > diag.z) { + // return 1; + // } else { + // return 2; + // } + + if (diag.v[0] > diag.v[1] && diag.v[0] > diag.v[2]) { return 0; - } else if (diag.y > diag.z) { + } else if (diag.v[1] > diag.v[2]) { return 1; } else { return 2; @@ -67,7 +73,8 @@ class aabb { float surface_area() { auto diag = diagonal(); - return 2.f * (diag.x * diag.y + diag.x * diag.z + diag.y * diag.z); + // return 2.f * (diag.x * diag.y + diag.x * diag.z + diag.y * diag.z); + return 2.f * (diag.v[0] * diag.v[1] + diag.v[0] * diag.v[2] + diag.v[1] * diag.v[2]); } vec3 minimum = vec3(std::numeric_limits::max()); diff --git a/include/bvh-node.hpp b/include/bvh-node.hpp index 09c9420..698977a 100644 --- a/include/bvh-node.hpp +++ b/include/bvh-node.hpp @@ -91,13 +91,13 @@ public: node.primitive_id = old_node.primitive_id; - node.min[0] = old_node.bounding_box.minimum.x; - node.min[1] = old_node.bounding_box.minimum.y; - node.min[2] = old_node.bounding_box.minimum.z; + node.min[0] = old_node.bounding_box.minimum.v[0]; + node.min[1] = old_node.bounding_box.minimum.v[1]; + node.min[2] = old_node.bounding_box.minimum.v[2]; - node.max[0] = old_node.bounding_box.maximum.x; - node.max[1] = old_node.bounding_box.maximum.y; - node.max[2] = old_node.bounding_box.maximum.z; + node.max[0] = old_node.bounding_box.maximum.v[0]; + node.max[1] = old_node.bounding_box.maximum.v[1]; + node.max[2] = old_node.bounding_box.maximum.v[2]; if (old_node.next_id != -1) { node.next_id = temp_nodes[old_node.next_id].df_id; diff --git a/include/utils.hpp b/include/utils.hpp index e7178d8..8548a5f 100644 --- a/include/utils.hpp +++ b/include/utils.hpp @@ -9,6 +9,7 @@ std::vector read_file(const char* path); std::vector read_file(std::string& path); #define PI 3.14159265359 +#define EPSILON 0.000001 inline float deg_to_rad(float deg) { return deg * PI / 180.0; diff --git a/include/vec3.hpp b/include/vec3.hpp index 6291e30..429d747 100644 --- a/include/vec3.hpp +++ b/include/vec3.hpp @@ -2,6 +2,7 @@ #define __VEC3_HPP_ #include +#include class vec3 { public: @@ -9,10 +10,15 @@ class vec3 { vec3(float x); vec3(float x, float y, float z); + vec3(__m128& v); + vec3(__m128&& v); + vec3(const vec3& vec); vec3 operator-() const; - vec3& operator+=(vec3 vec); + vec3& operator=(const vec3& vec); + + vec3& operator+=(const vec3& vec); vec3& operator*=(float scale); @@ -28,7 +34,11 @@ class vec3 { vec3 cross(const vec3& vec) const; - vec3 unit() const; + vec3& min(const vec3& vec); + + vec3& max(const vec3& vec); + + vec3& normalize(); bool near_zero() const; @@ -36,38 +46,29 @@ class vec3 { static vec3 random(float min, float max); - float x, y, z; - float padding; + __m128 v; }; -vec3 lerp(vec3 u, vec3 v, float t); +vec3 lerp(const vec3& u, const vec3& v, float t); -vec3 operator+(vec3 u, vec3 v); +vec3 operator+(const vec3& u, const vec3& v); -vec3 operator-(vec3 u, vec3 v); +vec3 operator-(const vec3& u, const vec3& v); -vec3 operator*(vec3 u, vec3 v); +vec3 operator*(const vec3& u, const vec3& v); -vec3 operator*(vec3 u, float scale); +vec3 operator*(const vec3& u, float scale); -vec3 operator*(float scale, vec3 u); +vec3 operator*(float scale, const vec3& u); -vec3 operator/(vec3 u, float scale); +vec3 operator/(const vec3& u, float scale); -vec3 reflect(vec3 v, vec3 n); +vec3 reflect(const vec3& v, const vec3& n); vec3 refract(const vec3& uv, const vec3& n, float etai_over_etat); std::ostream& operator<<(std::ostream& out, vec3 vec); -vec3 random_in_unit_sphere(); - -vec3 random_unit_vector(); - -vec3 random_in_hemisphere(const vec3& normal); - -vec3 random_in_unit_disk(); - using point3 = vec3; using color = vec3; -- cgit v1.3