Optimized computation of Sobol dim2, +1% overall perf

RenderKit · Jan 8, 2025 · d57b226 · d57b226
1 parent 0ee3e72
commit d57b226
Show file tree

Hide file tree

Showing 2 changed files with 28 additions and 51 deletions.
diff --git a/modules/cpu/math/sobol.ih b/modules/cpu/math/sobol.ih
@@ -30,19 +30,30 @@ OSPRAY_BEGIN_ISPC_NAMESPACE
 #ifdef OSPRAY_TARGET_SYCL
 #include "sobol.inl"
 #else
-extern const uniform unsigned int Sobol_revMatrices[4 * 32];
+extern const uniform unsigned int Sobol_revMatrices[3 * 32];
 extern const uniform unsigned int Sobol_matrices[5 * 52];
 #endif
 
+// compute component 1 of the Sobol'-sequence (count starting from 0)
+// Ahmed, "An Implementation Algorithm of 2D Sobol Sequence Fast, Elegant, and
+// Compact", EGSR 2024, https://doi.org/10.2312/sr.20241147
+inline uint32 Sobol_revSample1(uint32 v)
+{
+  v ^= v << 16;
+  v ^= (v & 0x00FF00FF) << 8;
+  v ^= (v & 0x0F0F0F0F) << 4;
+  v ^= (v & 0x33333333) << 2;
+  v ^= (v & 0x55555555) << 1;
+  return reverseBits(v);
+}
+
 // compute components 1 and 2 of the Sobol'-sequence (count starting from 0)
 inline vec2ui Sobol_revSample2(uint32 revIndex)
 {
-  vec2ui result = make_vec2ui(0);
-  for (uniform uint32 i = 0; revIndex; revIndex <<= 1, i += 4) {
-    if (revIndex & 0x80000000u) {
-      result.x ^= Sobol_revMatrices[i];
-      result.y ^= Sobol_revMatrices[i + 1];
-    }
+  vec2ui result = make_vec2ui(Sobol_revSample1(revIndex), 0);
+  for (uniform uint32 i = 0; revIndex; revIndex <<= 1, i += 3) {
+    if (revIndex & 0x80000000u)
+      result.y ^= Sobol_revMatrices[i];
   }
 
   return result;
@@ -51,12 +62,11 @@ inline vec2ui Sobol_revSample2(uint32 revIndex)
 // compute components 1 to 3 of the Sobol'-sequence (count starting from 0)
 inline vec3ui Sobol_revSample3(uint32 revIndex)
 {
-  vec3ui result = make_vec3ui(0);
-  for (uniform uint32 i = 0; revIndex; revIndex <<= 1, i += 4) {
+  vec3ui result = make_vec3ui(Sobol_revSample1(revIndex), 0, 0);
+  for (uniform uint32 i = 0; revIndex; revIndex <<= 1, i += 3) {
     if (revIndex & 0x80000000u) {
-      result.x ^= Sobol_revMatrices[i];
-      result.y ^= Sobol_revMatrices[i + 1];
-      result.z ^= Sobol_revMatrices[i + 2];
+      result.y ^= Sobol_revMatrices[i];
+      result.z ^= Sobol_revMatrices[i + 1];
     }
   }
 
@@ -66,13 +76,12 @@ inline vec3ui Sobol_revSample3(uint32 revIndex)
 // compute components 1 to 4 of the Sobol'-sequence (count starting from 0)
 inline vec4ui Sobol_revSample4(uint32 revIndex)
 {
-  vec4ui result = make_vec4ui(0);
-  for (uniform uint32 i = 0; revIndex; revIndex <<= 1, i += 4) {
+  vec4ui result = make_vec4ui(Sobol_revSample1(revIndex), 0, 0, 0);
+  for (uniform uint32 i = 0; revIndex; revIndex <<= 1, i += 3) {
     if (revIndex & 0x80000000u) {
-      result.x ^= Sobol_revMatrices[i];
-      result.y ^= Sobol_revMatrices[i + 1];
-      result.z ^= Sobol_revMatrices[i + 2];
-      result.w ^= Sobol_revMatrices[i + 3];
+      result.y ^= Sobol_revMatrices[i];
+      result.z ^= Sobol_revMatrices[i + 1];
+      result.w ^= Sobol_revMatrices[i + 2];
     }
   }
 

diff --git a/modules/cpu/math/sobol.inl b/modules/cpu/math/sobol.inl
@@ -30,163 +30,131 @@
 // http://web.maths.unsw.edu.au/~fkuo/sobol/new-joe-kuo-6.21201
 
 // bit-reversed and interleaved
-OSPRAY_GLOBAL const unsigned int Sobol_revMatrices[4 * 32] = {
-    0x1u,
+OSPRAY_GLOBAL const unsigned int Sobol_revMatrices[3 * 32] = {
     0x1u,
     0x1u,
     0x1u,
 
-    0x3u,
     0x3u,
     0x3u,
     0x2u,
 
-    0x5u,
     0x6u,
     0x4u,
     0x4u,
 
-    0xfu,
     0x9u,
     0xau,
     0xdu,
 
-    0x11u,
     0x17u,
     0x1fu,
     0x1fu,
 
-    0x33u,
     0x3au,
     0x2eu,
     0x3bu,
 
-    0x55u,
     0x71u,
     0x45u,
     0x5eu,
 
-    0xffu,
     0xa3u,
     0xc9u,
     0xb9u,
 
-    0x101u,
     0x116u,
     0x11bu,
     0x15au,
 
-    0x303u,
     0x339u,
     0x2a4u,
     0x3f4u,
 
-    0x505u,
     0x677u,
     0x79au,
     0x685u,
 
-    0xf0fu,
     0x9aau,
     0xb67u,
     0xd0fu,
 
-    0x1111u,
     0x1601u,
     0x101eu,
     0x115bu,
 
-    0x3333u,
     0x3903u,
     0x302du,
     0x23f6u,
 
-    0x5555u,
     0x7706u,
     0x4041u,
     0x4681u,
 
-    0xffffu,
     0xaa09u,
     0xa0c3u,
     0xdd02u,
 
-    0x10001u,
     0x10117u,
     0x1f104u,
     0x1e144u,
 
-    0x30003u,
     0x3033au,
     0x2e28au,
     0x393cdu,
 
-    0x50005u,
     0x60671u,
     0x457dfu,
     0x5a6dfu,
 
-    0xf000fu,
     0x909a3u,
     0xc9baeu,
     0xb4dbbu,
 
-    0x110011u,
     0x171616u,
     0x11a105u,
     0x14401eu,
 
-    0x330033u,
     0x3a3939u,
     0x2a7289u,
     0x3cd039u,
 
-    0x550055u,
     0x717777u,
     0x79e7dbu,
     0x6df05au,
 
-    0xff00ffu,
     0xa3aaaau,
     0xb6dba4u,
     0xdbb0b4u,
 
-    0x1010101u,
     0x1170001u,
     0x100011au,
     0x101e145u,
 
-    0x3030303u,
     0x33a0003u,
     0x30002a7u,
     0x20393cfu,
 
-    0x5050505u,
     0x6710006u,
     0x400079eu,
     0x405a6dbu,
 
-    0xf0f0f0fu,
     0x9a30009u,
     0xa000b6du,
     0xd0b4db6u,
 
-    0x11111111u,
     0x16160017u,
     0x1f001001u,
     0x1f144001u,
 
-    0x33333333u,
     0x3939003au,
     0x2e003003u,
     0x3b3cd002u,
 
-    0x55555555u,
     0x77770071u,
     0x45004004u,
     0x5e6df004u,
 
-    0xffffffffu,
     0xaaaa00a3u,
     0xc900a00au,
     0xb9dbb00du,