From c5208ad17a472db5ff3d3af6e0a3a68f94feb6f8 Mon Sep 17 00:00:00 2001
From: Chris Duncan <chris@zoso.dev>
Date: Sat, 21 Dec 2024 07:25:35 -0800
Subject: [PATCH] Combine work bytes for upload to GPU and split them in the
 shader. Clear performance measure once it's been saved to frameTimes.

---
 src/lib/workers/powgl.ts | 38 ++++++++++++++++++--------------------
 1 file changed, 18 insertions(+), 20 deletions(-)

diff --git a/src/lib/workers/powgl.ts b/src/lib/workers/powgl.ts
index 973213d..af0f312 100644
--- a/src/lib/workers/powgl.ts
+++ b/src/lib/workers/powgl.ts
@@ -62,12 +62,6 @@ precision highp int;
 in vec2 uv_pos;
 out vec4 fragColor;
 
-// Random work values
-// First 2 bytes will be overwritten by texture pixel position
-// Second 2 bytes will be modified if the canvas size is greater than 256x256
-uniform uvec4 u_work0;
-// Last 4 bytes remain as generated externally
-uniform uvec4 u_work1;
 // Precalculated block hash components
 uniform uint blockHash[8];
 // Threshold is 0xfffffff8 for send/change blocks and 0xfffffe for all else
@@ -75,6 +69,12 @@ uniform uint threshold;
 // Defines canvas size
 uniform float workload;
 
+// Random work values
+// First 2 bytes will be overwritten by texture pixel position
+// Second 2 bytes will be modified if the canvas size is greater than 256x256
+// Last 4 bytes remain as generated externally
+uniform uvec4 work[2];
+
 // Defined separately from uint v[32] below as the original value is required
 // to calculate the second uint32 of the digest for threshold comparison
 #define BLAKE2B_IV32_1 0x6A09E667u
@@ -174,6 +174,8 @@ void B2B_G (int a, int b, int c, int d, int ix, int iy) {
 
 void main() {
 	int i;
+	uvec4 u_work0 = work[0];
+	uvec4 u_work1 = work[1];
 	uint uv_x = uint(uv_pos.x * workload);
 	uint uv_y = uint(uv_pos.y * workload);
 	uint x_pos = uv_x % 256u;
@@ -221,7 +223,7 @@ void main() {
 }`
 
 	/** Used to set canvas size. Must be a multiple of 256. */
-	static #WORKLOAD: number = 256 * Math.max(1, Math.floor(navigator.hardwareConcurrency) / 2)
+	static #WORKLOAD: number = 256 * Math.max(1, Math.floor(navigator.hardwareConcurrency))
 
 	static #hexify (arr: number[] | Uint8Array): string {
 		let out = ''
@@ -233,8 +235,7 @@ void main() {
 
 	static #gl: WebGL2RenderingContext | null
 	static #program: WebGLProgram | null
-	static #work0Location: WebGLUniformLocation | null
-	static #work1Location: WebGLUniformLocation | null
+	static #workLocation: WebGLUniformLocation | null
 	static #blockHashLocation: WebGLUniformLocation | null
 	static #thresholdLocation: WebGLUniformLocation | null
 	static #workloadLocation: WebGLUniformLocation | null
@@ -303,8 +304,7 @@ void main() {
 		this.#pixels = new Uint8Array(this.#gl.drawingBufferWidth * this.#gl.drawingBufferHeight * 4)
 		this.#query = this.#gl.createQuery()
 
-		this.#work0Location = this.#gl.getUniformLocation(this.#program, 'u_work0')
-		this.#work1Location = this.#gl.getUniformLocation(this.#program, 'u_work1')
+		this.#workLocation = this.#gl.getUniformLocation(this.#program, 'work')
 		this.#blockHashLocation = this.#gl.getUniformLocation(this.#program, "blockHash")
 		this.#thresholdLocation = this.#gl.getUniformLocation(this.#program, "threshold")
 		this.#workloadLocation = this.#gl.getUniformLocation(this.#program, "workload")
@@ -326,8 +326,7 @@ void main() {
 		Pow.#gl.uniform1uiv(Pow.#blockHashLocation, hashBytes)
 		Pow.#gl.uniform1ui(Pow.#thresholdLocation, threshold)
 		Pow.#gl.uniform1f(Pow.#workloadLocation, Pow.#WORKLOAD - 1)
-		const work0 = new Uint8Array(4)
-		const work1 = new Uint8Array(4)
+		const work = new Uint8Array(8)
 
 		// Draw output until success or progressCallback says to stop
 		let n = 0
@@ -337,12 +336,10 @@ void main() {
 			if (Pow.#gl == null) throw new Error('WebGL 2 is required')
 			if (Pow.#query == null) throw new Error('WebGL 2 is required to run queries')
 			performance.mark('start')
-			crypto.getRandomValues(work0)
-			crypto.getRandomValues(work1)
+			crypto.getRandomValues(work)
 			Pow.#gl.clear(Pow.#gl.COLOR_BUFFER_BIT)
 
-			Pow.#gl.uniform4uiv(Pow.#work0Location, work0)
-			Pow.#gl.uniform4uiv(Pow.#work1Location, work1)
+			Pow.#gl.uniform4uiv(Pow.#workLocation, work)
 
 			Pow.#gl.beginQuery(Pow.#gl.ANY_SAMPLES_PASSED_CONSERVATIVE, Pow.#query)
 			Pow.#gl.drawArrays(Pow.#gl.TRIANGLES, 0, 6)
@@ -363,6 +360,7 @@ void main() {
 					performance.mark('end')
 					frameTimes.push(performance.measure('draw', 'start', 'end').duration)
 					performance.clearMarks()
+					performance.clearMeasures()
 					// No valid nonce found, start the next draw call
 					requestAnimationFrame(draw)
 				}
@@ -392,11 +390,11 @@ void main() {
 					console.log(`Average: ${sum / count} ms`)
 					console.log(`Harmonic: ${count / reciprocals} ms`)
 					console.log(`Geometric: ${Math.pow(product, 1 / count)} ms`)
-					const hex = Pow.#hexify(work1) + Pow.#hexify([
+					const hex = Pow.#hexify(work.subarray(4, 8)) + Pow.#hexify([
 						Pow.#pixels[i + 2],
 						Pow.#pixels[i + 3],
-						work0[2] ^ (Pow.#pixels[i] - 1),
-						work0[3] ^ (Pow.#pixels[i + 1] - 1)
+						work[2] ^ (Pow.#pixels[i] - 1),
+						work[3] ^ (Pow.#pixels[i + 1] - 1)
 					])
 					// Return the work value with the custom bits
 					typeof callback === 'function' && callback(hex)
-- 
2.34.1