From c48839d5f3af9fd0147dcfdc16fe46e171037275 Mon Sep 17 00:00:00 2001
From: Nathan Vegdahl <cessen@cessen.com>
Date: Sat, 23 Jan 2021 00:55:54 +0900
Subject: [PATCH] An even better Owen scrambling hash.

---
 sub_crates/sobol/src/lib.rs  | 12 +++----
 sub_crates/sobol/src/wide.rs | 65 ++++++++++++++++++++++++++++++------
 2 files changed, 60 insertions(+), 17 deletions(-)

diff --git a/sub_crates/sobol/src/lib.rs b/sub_crates/sobol/src/lib.rs
index b149ea8..64ca5ec 100644
--- a/sub_crates/sobol/src/lib.rs
+++ b/sub_crates/sobol/src/lib.rs
@@ -66,10 +66,10 @@ pub fn sample_4d(sample_index: u32, dimension_set: u32, seed: u32) -> [f32; 4] {
 fn lk_scramble(mut n: u32, scramble: u32) -> u32 {
     let scramble = hash(scramble);
 
-    n = n.wrapping_add(scramble);
-    n ^= n.wrapping_mul(0x3354734a);
     n = n.wrapping_add(n << 2);
-    n ^= n.wrapping_mul(scramble & !1);
+    n ^= n.wrapping_mul(0xfe9b5742);
+    n = n.wrapping_add(scramble);
+    n = n.wrapping_mul(scramble | 1);
 
     n
 }
@@ -79,10 +79,10 @@ fn lk_scramble(mut n: u32, scramble: u32) -> u32 {
 fn lk_scramble_int4(mut n: Int4, scramble: u32) -> Int4 {
     let scramble = hash_int4([scramble; 4].into());
 
-    n += scramble;
-    n ^= n * [0x3354734a; 4].into();
     n += n << 2;
-    n ^= n * (scramble & [!1; 4].into());
+    n ^= n * [0xfe9b5742; 4].into();
+    n += scramble;
+    n *= scramble | [1; 4].into();
 
     n
 }
diff --git a/sub_crates/sobol/src/wide.rs b/sub_crates/sobol/src/wide.rs
index 831648d..f90940f 100644
--- a/sub_crates/sobol/src/wide.rs
+++ b/sub_crates/sobol/src/wide.rs
@@ -145,6 +145,40 @@ pub(crate) mod sse {
         }
     }
 
+    impl std::ops::BitAnd for Int4 {
+        type Output = Int4;
+
+        #[inline(always)]
+        fn bitand(self, other: Self) -> Int4 {
+            Int4 {
+                v: unsafe { _mm_and_si128(self.v, other.v) },
+            }
+        }
+    }
+
+    impl std::ops::BitAndAssign for Int4 {
+        fn bitand_assign(&mut self, other: Self) {
+            *self = *self & other;
+        }
+    }
+
+    impl std::ops::BitOr for Int4 {
+        type Output = Int4;
+
+        #[inline(always)]
+        fn bitor(self, other: Self) -> Int4 {
+            Int4 {
+                v: unsafe { _mm_or_si128(self.v, other.v) },
+            }
+        }
+    }
+
+    impl std::ops::BitOrAssign for Int4 {
+        fn bitor_assign(&mut self, other: Self) {
+            *self = *self | other;
+        }
+    }
+
     impl std::ops::BitXor for Int4 {
         type Output = Int4;
 
@@ -163,17 +197,6 @@ pub(crate) mod sse {
         }
     }
 
-    impl std::ops::BitAnd for Int4 {
-        type Output = Int4;
-
-        #[inline(always)]
-        fn bitand(self, other: Self) -> Int4 {
-            Int4 {
-                v: unsafe { _mm_and_si128(self.v, other.v) },
-            }
-        }
-    }
-
     impl std::ops::Shl<i32> for Int4 {
         type Output = Int4;
 
@@ -301,6 +324,26 @@ pub(crate) mod fallback {
         }
     }
 
+    impl std::ops::BitOr for Int4 {
+        type Output = Int4;
+        fn bitor(self, other: Self) -> Int4 {
+            Int4 {
+                v: [
+                    self.v[0] | other.v[0],
+                    self.v[1] | other.v[1],
+                    self.v[2] | other.v[2],
+                    self.v[3] | other.v[3],
+                ],
+            }
+        }
+    }
+
+    impl std::ops::BitOrAssign for Int4 {
+        fn bitor_assign(&mut self, other: Self) {
+            *self = *self | other;
+        }
+    }
+
     impl std::ops::BitXor for Int4 {
         type Output = Int4;
         fn bitxor(self, other: Self) -> Int4 {