From c3c9158e28803d4c15cc1f8929566e101a7dd93e Mon Sep 17 00:00:00 2001
From: Chris Duncan <chris@zoso.dev>
Date: Thu, 16 Jan 2025 05:31:17 -0800
Subject: [PATCH] Replace 16 and 63 bit rotations with swizzle version.

---
 src/shaders/compute.wgsl | 290 +++++++++++++++++++--------------------
 1 file changed, 144 insertions(+), 146 deletions(-)

diff --git a/src/shaders/compute.wgsl b/src/shaders/compute.wgsl
index b63d95d..8cee7b9 100644
--- a/src/shaders/compute.wgsl
+++ b/src/shaders/compute.wgsl
@@ -178,16 +178,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	// v12 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	// v4 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -222,14 +220,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -264,14 +262,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -306,14 +304,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -355,7 +353,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -390,14 +388,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -432,7 +430,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -520,14 +518,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -562,14 +560,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -604,14 +602,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -646,14 +644,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -695,7 +693,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -730,14 +728,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -772,7 +770,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -860,14 +858,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -902,14 +900,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -944,14 +942,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -986,14 +984,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1035,7 +1033,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1070,14 +1068,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1112,7 +1110,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -1200,14 +1198,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1242,14 +1240,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1284,14 +1282,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1326,14 +1324,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1375,7 +1373,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1410,14 +1408,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1452,7 +1450,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -1540,14 +1538,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1582,14 +1580,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1624,14 +1622,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1666,14 +1664,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1715,7 +1713,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1750,14 +1748,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1792,7 +1790,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -1880,14 +1878,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1922,14 +1920,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -1964,14 +1962,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2006,14 +2004,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2055,7 +2053,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2090,14 +2088,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2132,7 +2130,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -2220,14 +2218,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2262,14 +2260,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2304,14 +2302,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2346,14 +2344,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2395,7 +2393,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2430,14 +2428,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2472,7 +2470,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -2560,14 +2558,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2602,14 +2600,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2644,14 +2642,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2686,14 +2684,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2735,7 +2733,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2770,14 +2768,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2812,7 +2810,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -2900,14 +2898,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2942,14 +2940,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -2984,14 +2982,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3026,14 +3024,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3075,7 +3073,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3110,14 +3108,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3152,7 +3150,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -3240,14 +3238,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3282,14 +3280,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3324,14 +3322,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3366,14 +3364,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3415,7 +3413,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3450,14 +3448,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3492,7 +3490,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -3580,14 +3578,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3622,14 +3620,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3664,14 +3662,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3706,14 +3704,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3755,7 +3753,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3790,14 +3788,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3832,7 +3830,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
@@ -3920,14 +3918,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v1;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v13 + vec2(0u, u32(v9.x + v13.x < v9.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v9;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -3962,14 +3960,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v2;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v14 + vec2(0u, u32(v10.x + v14.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v10;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -4004,14 +4002,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v3;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v11 = v11 + v15 + vec2(0u, u32(v11.x + v15.x < v11.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v11;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -4046,14 +4044,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v15 ^ v0;
-	v15 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v15 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v10 = v10 + v15 + vec2(0u, u32(v10.x + v15.x < v10.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v5 ^ v10;
-	v5 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v5 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -4095,7 +4093,7 @@ fn main(id: vec3<u32>) {
 
 	// b = rotr64(b ^ c, 63)
 	xor = v6 ^ v11;
-	v6 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v6 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -4130,14 +4128,14 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v13 ^ v2;
-	v13 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v13 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v8 = v8 + v13 + vec2(0u, u32(v8.x + v13.x < v8.x));
 
 	// b = rotr64(b ^ c, 63)
 	xor = v7 ^ v8;
-	v7 = vec2((xor.x << 1u) | (xor.y >> 31u), (xor.y << 1u) | (xor.x >> 31u));
+	v7 = vec2((xor.xy << vec2(1u, 1u)) | (xor.yx >> vec2(31u, 31u)));
 
 
 
@@ -4172,7 +4170,7 @@ fn main(id: vec3<u32>) {
 
 	// d = rotr64(d ^ a, 16)
 	xor = v14 ^ v3;
-	v14 = vec2((xor.x >> 16u) | (xor.y << 16u), (xor.y >> 16u) | (xor.x << 16u));
+	v14 = vec2((xor.xy >> vec2(16u, 16u)) | (xor.yx << vec2(16u, 16u)));
 
 	// c = c + d
 	v9 = v9 + v14 + vec2(0u, u32(v9.x + v14.x < v9.x));
-- 
2.34.1