Switch from murmur32 to xxhash32 for object-overlays (#89)

crocdialer · Feb 16, 2025 · 77d73fb · 77d73fb
1 parent 67499ec
commit 77d73fb
Show file tree

Hide file tree

Showing 10 changed files with 51 additions and 23 deletions.
diff --git a/include/vierkant/hash.hpp b/include/vierkant/hash.hpp
@@ -127,8 +127,8 @@ static inline uint32_t murmur3_32(const K &key, uint32_t seed)
 // https://jcgt.org/published/0009/03/02/
 inline uint32_t xxhash32(uint32_t lhs, uint32_t rhs)
 {
-    const uint32_t PRIME32_2 = 2246822519U, PRIME32_3 = 3266489917U;
-    const uint32_t PRIME32_4 = 668265263U, PRIME32_5 = 374761393U;
+    constexpr uint32_t PRIME32_2 = 2246822519U, PRIME32_3 = 3266489917U;
+    constexpr uint32_t PRIME32_4 = 668265263U, PRIME32_5 = 374761393U;
     uint32_t h32 = lhs + PRIME32_5 + rhs * PRIME32_3;
     h32 = PRIME32_4 * ((h32 << 17) | (h32 >> (32 - 17)));
     h32 = PRIME32_2 * (h32 ^ (h32 >> 15));
@@ -141,6 +141,32 @@ inline uint32_t hash_combine32(uint32_t lhs, uint32_t rhs)
     return lhs ^ (rhs + 0x9e3779b9 + (lhs << 6U) + (lhs >> 2U));
 }
 
+template<typename K>
+static inline uint32_t xxhash32(const K &key, uint32_t seed)
+{
+    constexpr uint32_t num_hashes = sizeof(K) / sizeof(uint32_t);
+    constexpr uint32_t num_excess_bytes = sizeof(K) % sizeof(uint32_t);
+    uint32_t h = seed;
+
+    if constexpr(num_hashes)
+    {
+        auto ptr = reinterpret_cast<const uint32_t *>(&key), end = ptr + num_hashes;
+        for(; ptr < end; ++ptr) { h = xxhash32(*ptr, h); }
+    }
+    if constexpr(num_excess_bytes)
+    {
+        auto end_u8 = reinterpret_cast<const uint8_t *>(&key) + sizeof(uint32_t) * num_hashes;
+        uint32_t k = 0;
+        for(uint32_t i = num_excess_bytes; i; i--)
+        {
+            k <<= 8;
+            k |= end_u8[i - 1];
+        }
+        h = xxhash32(k, h);
+    }
+    return h;
+}
+
 template<class T>
 inline void hash_combine(std::size_t &seed, const T &v)
 {

diff --git a/include/vierkant/linear_hashmap.hpp b/include/vierkant/linear_hashmap.hpp
@@ -201,7 +201,7 @@ class linear_hashmap
     uint64_t m_capacity = 0;
     uint64_t m_num_elements = 0;
     std::unique_ptr<storage_item_t[]> m_storage;
-    hash32_fn m_hash_fn = std::bind(murmur3_32<key_t>, std::placeholders::_1, 0);
+    hash32_fn m_hash_fn = std::bind(xxhash32<key_t>, std::placeholders::_1, 0);
 
     // reasonably low load-factor to keep average probe-lengths low
     float m_max_load_factor = 0.5f;
@@ -408,7 +408,7 @@ class linear_hashmap_mt
     uint64_t m_capacity{};
     std::atomic<uint64_t> m_num_elements;
     std::unique_ptr<storage_item_t[]> m_storage;
-    hash32_fn m_hash_fn = std::bind(vierkant::murmur3_32<key_t>, std::placeholders::_1, 0);
+    hash32_fn m_hash_fn = std::bind(vierkant::xxhash32<key_t>, std::placeholders::_1, 0);
     mutable std::shared_mutex m_mutex;
 
     // reasonably low load-factor to keep average probe-lengths low

diff --git a/shaders/fullscreen/ao_screenspace.frag b/shaders/fullscreen/ao_screenspace.frag
@@ -7,6 +7,7 @@
 
 // rnd(state)
 #include "../utils/random.glsl"
+#include "../utils/hash.glsl"
 
 // bsdf utils
 #include "../utils/bsdf.glsl"

diff --git a/shaders/ray/raygen.rgen b/shaders/ray/raygen.rgen
@@ -3,6 +3,7 @@
 #extension GL_GOOGLE_include_directive : enable
 
 #include "../utils/random.glsl"
+#include "../utils/hash.glsl"
 #include "../utils/bsdf.glsl"
 #include "ray_common.glsl"
 #include "reservoir.glsl"

diff --git a/shaders/renderer/object_overlay.comp b/shaders/renderer/object_overlay.comp
@@ -21,7 +21,7 @@ struct hashmap_t
 
 uint hash(uint v)
 {
-    return murmur3_32(v, 0);
+    return xxhash32(v, 0);
 }
 
 uint get(const hashmap_t hashmap, const uint key)

diff --git a/shaders/utils/hash.glsl b/shaders/utils/hash.glsl
@@ -55,4 +55,18 @@ uint hash_combine32(uint lhs, uint rhs)
     return lhs ^ (rhs + 0x9e3779b9 + (lhs << 6U) + (lhs >> 2U));
 }
 
+// Generate a random unsigned int from two unsigned int values
+// @see: "Mark Jarzynski and Marc Olano, Hash Functions for GPU Rendering, Journal of Computer Graphics Techniques (JCGT), vol. 9, no. 3, 21-38, 2020"
+// https://jcgt.org/published/0009/03/02/
+uint xxhash32(uint lhs, uint rhs)
+{
+    const uint PRIME32_2 = 2246822519U, PRIME32_3 = 3266489917U;
+    const uint PRIME32_4 = 668265263U, PRIME32_5 = 374761393U;
+    uint h32 = lhs + PRIME32_5 + rhs * PRIME32_3;
+    h32 = PRIME32_4 * ((h32 << 17) | (h32 >> (32 - 17)));
+    h32 = PRIME32_2 * (h32 ^ (h32 >> 15));
+    h32 = PRIME32_3 * (h32 ^ (h32 >> 13));
+    return h32 ^ (h32 >> 16);
+}
+
 #endif // UTILS_HASH_GLSL
diff --git a/shaders/utils/random.glsl b/shaders/utils/random.glsl
@@ -22,20 +22,6 @@ uint tea(uint lhs, uint rhs)
     return v0;
 }
 
-// Generate a random unsigned int from two unsigned int values
-// @see: "Mark Jarzynski and Marc Olano, Hash Functions for GPU Rendering, Journal of Computer Graphics Techniques (JCGT), vol. 9, no. 3, 21-38, 2020"
-// https://jcgt.org/published/0009/03/02/
-uint xxhash32(uint lhs, uint rhs)
-{
-    const uint PRIME32_2 = 2246822519U, PRIME32_3 = 3266489917U;
-    const uint PRIME32_4 = 668265263U, PRIME32_5 = 374761393U;
-    uint h32 = lhs + PRIME32_5 + rhs * PRIME32_3;
-    h32 = PRIME32_4 * ((h32 << 17) | (h32 >> (32 - 17)));
-    h32 = PRIME32_2 * (h32 ^ (h32 >> 15));
-    h32 = PRIME32_3 * (h32 ^ (h32 >> 13));
-    return h32 ^ (h32 >> 16);
-}
-
 // Generate a random unsigned int in [0, 2^24) given the previous RNG state
 // using the Numerical Recipes linear congruential generator
 uint lcg(inout uint prev)

diff --git a/submodules/JoltPhysics b/submodules/JoltPhysics
diff --git a/submodules/meshoptimizer b/submodules/meshoptimizer
diff --git a/tests/TestLinearHashmap.cpp b/tests/TestLinearHashmap.cpp
@@ -81,8 +81,8 @@ void test_probe_length()
     // test a load-factor of 0.25
     hashmap.max_load_factor(0.25f);
 
-    constexpr uint32_t test_capacity = 512;
-    constexpr uint32_t num_insertions = 128;
+    constexpr uint32_t test_capacity = 4096;
+    constexpr uint32_t num_insertions = 1024;
     hashmap.reserve(test_capacity);
 
     float probe_length_sum = 0.f;
+1 −0		CMakeLists.txt
+1 −3		README.md
+75 −31		demo/main.cpp
+211 −224		demo/nanite.cpp
+76 −0		demo/tests.cpp
+42 −11		extern/cgltf.h
+4 −4		gltf/write.cpp
+1 −1		js/package.json
+74 −8		src/clusterizer.cpp
+34 −22		src/meshoptimizer.h
+422 −0		src/partition.cpp