nano-pow 2.0.0 → 3.0.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +38 -16
- package/dist/main.min.js +487 -329
- package/dist/types.d.ts +18 -1
- package/package.json +5 -5
- package/dist/classes/gl.d.ts +0 -20
- package/dist/classes/gl.js +0 -232
- package/dist/classes/gpu.d.ts +0 -24
- package/dist/classes/gpu.js +0 -236
- package/dist/classes/index.d.ts +0 -4
- package/dist/classes/index.js +0 -23
- package/dist/global.d.ts +0 -1
- package/dist/global.js +0 -6
- package/dist/global.min.js +0 -801
- package/dist/main.d.ts +0 -3
- package/dist/main.js +0 -5
- package/dist/shaders/gl-fragment.d.ts +0 -1
- package/dist/shaders/gl-fragment.js +0 -206
- package/dist/shaders/gl-vertex.d.ts +0 -1
- package/dist/shaders/gl-vertex.js +0 -16
- package/dist/shaders/index.d.ts +0 -4
- package/dist/shaders/index.js +0 -6
package/dist/main.min.js
CHANGED
|
@@ -1,36 +1,77 @@
|
|
|
1
1
|
// src/shaders/compute.wgsl
|
|
2
|
-
var compute_default = "struct UBO{blockhash:array<vec4<u32>,2>,random:vec2<u32>,threshold:u32};@group(0)@binding(0)var<uniform> ubo:UBO;struct WORK{nonce:vec2<u32>,found:atomic<u32>};@group(0)@binding(1)var<storage,read_write>work:WORK;const BLAKE2B_IV32_1:u32=0x6A09E667u;const ROTATE_1=vec2(1u,1u);const ROTATE_8=vec2(8u,8u);const ROTATE_16=vec2(16u,16u);const ROTATE_24=vec2(24u,24u);const ROTATE_31=vec2(31u,31u);@compute @workgroup_size(8,8)fn search(@builtin(global_invocation_id)global_id:vec3<u32>){if(atomicLoad(&work.found)!=0u){return;}main(global_id);}@compute @workgroup_size(1)fn validate(@builtin(global_invocation_id)global_id:vec3<u32>){if(atomicLoad(&work.found)!=0u){return;}main(global_id);}fn main(id:vec3<u32>){let threshold:u32=ubo.threshold;let m0:vec2<u32>=ubo.random ^ id.xy;let m1:vec2<u32>=ubo.blockhash[0u].xy;let m2:vec2<u32>=ubo.blockhash[0u].zw;let m3:vec2<u32>=ubo.blockhash[1u].xy;let m4:vec2<u32>=ubo.blockhash[1u].zw;var v0:vec2<u32>=vec2(0xF2BDC900u,0x6A09E667u);var v1:vec2<u32>=vec2(0x84CAA73Bu,0xBB67AE85u);var v2:vec2<u32>=vec2(0xFE94F82Bu,0x3C6EF372u);var v3:vec2<u32>=vec2(0x5F1D36F1u,0xA54FF53Au);var v4:vec2<u32>=vec2(0xADE682D1u,0x510E527Fu);var v5:vec2<u32>=vec2(0x2B3E6C1Fu,0x9B05688Cu);var v6:vec2<u32>=vec2(0xFB41BD6Bu,0x1F83D9ABu);var v7:vec2<u32>=vec2(0x137E2179u,0x5BE0CD19u);var v8:vec2<u32>=vec2(0xF3BCC908u,0x6A09E667u);var v9:vec2<u32>=vec2(0x84CAA73Bu,0xBB67AE85u);var v10:vec2<u32>=vec2(0xFE94F82Bu,0x3C6EF372u);var v11:vec2<u32>=vec2(0x5F1D36F1u,0xA54FF53Au);var v12:vec2<u32>=vec2(0xADE682F9u,0x510E527Fu);var v13:vec2<u32>=vec2(0x2B3E6C1Fu,0x9B05688Cu);var v14:vec2<u32>=vec2(0x04BE4294u,0xE07C2654u);var v15:vec2<u32>=vec2(0x137E2179u,0x5BE0CD19u);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v0=v0+m0+vec2(0u,u32(v0.x+m0.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v0=v0+m1+vec2(0u,u32(v0.x+m1.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m2+vec2(0u,u32(v1.x+m2.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m3+vec2(0u,u32(v1.x+m3.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v2=v2+m4+vec2(0u,u32(v2.x+m4.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m4+vec2(0u,u32(v1.x+m4.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v0=v0+m1+vec2(0u,u32(v0.x+m1.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v1=v1+m0+vec2(0u,u32(v1.x+m0.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v1=v1+m2+vec2(0u,u32(v1.x+m2.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v3=v3+m3+vec2(0u,u32(v3.x+m3.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m0+vec2(0u,u32(v1.x+m0.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v2=v2+m2+vec2(0u,u32(v2.x+m2.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v1=v1+m3+vec2(0u,u32(v1.x+m3.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v2=v2+m1+vec2(0u,u32(v2.x+m1.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v3=v3+m4+vec2(0u,u32(v3.x+m4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m3+vec2(0u,u32(v1.x+m3.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m1+vec2(0u,u32(v1.x+m1.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v0=v0+m2+vec2(0u,u32(v0.x+m2.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v2=v2+m4+vec2(0u,u32(v2.x+m4.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v2=v2+m0+vec2(0u,u32(v2.x+m0.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v0=v0+m0+vec2(0u,u32(v0.x+m0.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v2=v2+m2+vec2(0u,u32(v2.x+m2.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v2=v2+m4+vec2(0u,u32(v2.x+m4.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v0=v0+m1+vec2(0u,u32(v0.x+m1.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v3=v3+m3+vec2(0u,u32(v3.x+m3.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v0=v0+m2+vec2(0u,u32(v0.x+m2.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v2=v2+m0+vec2(0u,u32(v2.x+m0.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v3=v3+m3+vec2(0u,u32(v3.x+m3.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v0=v0+m4+vec2(0u,u32(v0.x+m4.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v3=v3+m1+vec2(0u,u32(v3.x+m1.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m1+vec2(0u,u32(v1.x+m1.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v3=v3+m4+vec2(0u,u32(v3.x+m4.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v0=v0+m0+vec2(0u,u32(v0.x+m0.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v1=v1+m3+vec2(0u,u32(v1.x+m3.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v2=v2+m2+vec2(0u,u32(v2.x+m2.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v2=v2+m1+vec2(0u,u32(v2.x+m1.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v3=v3+m3+vec2(0u,u32(v3.x+m3.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v0=v0+m0+vec2(0u,u32(v0.x+m0.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v1=v1+m4+vec2(0u,u32(v1.x+m4.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v3=v3+m2+vec2(0u,u32(v3.x+m2.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v2=v2+m3+vec2(0u,u32(v2.x+m3.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v3=v3+m0+vec2(0u,u32(v3.x+m0.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v0=v0+m2+vec2(0u,u32(v0.x+m2.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v2=v2+m1+vec2(0u,u32(v2.x+m1.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v2=v2+m4+vec2(0u,u32(v2.x+m4.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v0=v0+m2+vec2(0u,u32(v0.x+m2.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m4+vec2(0u,u32(v1.x+m4.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v3=v3+m1+vec2(0u,u32(v3.x+m1.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v2=v2+m3+vec2(0u,u32(v2.x+m3.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v3=v3+m0+vec2(0u,u32(v3.x+m0.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v0=v0+m0+vec2(0u,u32(v0.x+m0.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v0=v0+m1+vec2(0u,u32(v0.x+m1.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy<<ROTATE_1)|((v4 ^ v8).yx>>ROTATE_31);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m2+vec2(0u,u32(v1.x+m2.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m3+vec2(0u,u32(v1.x+m3.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v2=v2+m4+vec2(0u,u32(v2.x+m4.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy<<ROTATE_1)|((v6 ^ v10).yx>>ROTATE_31);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v15=((v15 ^ v0).xy>>ROTATE_16)|((v15 ^ v0).yx<<ROTATE_16);v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy<<ROTATE_1)|((v5 ^ v10).yx>>ROTATE_31);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=v12.yx ^ v1.yx;v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy>>ROTATE_24)|((v6 ^ v11).yx<<ROTATE_8);v1=v1+v6+vec2(0u,u32(v1.x+v6.x<v1.x));v12=((v12 ^ v1).xy>>ROTATE_16)|((v12 ^ v1).yx<<ROTATE_16);v11=v11+v12+vec2(0u,u32(v11.x+v12.x<v11.x));v6=((v6 ^ v11).xy<<ROTATE_1)|((v6 ^ v11).yx>>ROTATE_31);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy<<ROTATE_1)|((v7 ^ v8).yx>>ROTATE_31);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=v14.yx ^ v3.yx;v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy>>ROTATE_24)|((v4 ^ v9).yx<<ROTATE_8);v3=v3+v4+vec2(0u,u32(v3.x+v4.x<v3.x));v14=((v14 ^ v3).xy>>ROTATE_16)|((v14 ^ v3).yx<<ROTATE_16);v9=v9+v14+vec2(0u,u32(v9.x+v14.x<v9.x));v4=((v4 ^ v9).xy<<ROTATE_1)|((v4 ^ v9).yx>>ROTATE_31);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=v12.yx ^ v0.yx;v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v4=((v4 ^ v8).xy>>ROTATE_24)|((v4 ^ v8).yx<<ROTATE_8);v0=v0+v4+vec2(0u,u32(v0.x+v4.x<v0.x));v12=((v12 ^ v0).xy>>ROTATE_16)|((v12 ^ v0).yx<<ROTATE_16);v8=v8+v12+vec2(0u,u32(v8.x+v12.x<v8.x));v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v1=v1+m4+vec2(0u,u32(v1.x+m4.x<v1.x));v13=v13.yx ^ v1.yx;v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy>>ROTATE_24)|((v5 ^ v9).yx<<ROTATE_8);v1=v1+v5+vec2(0u,u32(v1.x+v5.x<v1.x));v13=((v13 ^ v1).xy>>ROTATE_16)|((v13 ^ v1).yx<<ROTATE_16);v9=v9+v13+vec2(0u,u32(v9.x+v13.x<v9.x));v5=((v5 ^ v9).xy<<ROTATE_1)|((v5 ^ v9).yx>>ROTATE_31);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=v14.yx ^ v2.yx;v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v6=((v6 ^ v10).xy>>ROTATE_24)|((v6 ^ v10).yx<<ROTATE_8);v2=v2+v6+vec2(0u,u32(v2.x+v6.x<v2.x));v14=((v14 ^ v2).xy>>ROTATE_16)|((v14 ^ v2).yx<<ROTATE_16);v10=v10+v14+vec2(0u,u32(v10.x+v14.x<v10.x));v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=v15.yx ^ v3.yx;v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy>>ROTATE_24)|((v7 ^ v11).yx<<ROTATE_8);v3=v3+v7+vec2(0u,u32(v3.x+v7.x<v3.x));v15=((v15 ^ v3).xy>>ROTATE_16)|((v15 ^ v3).yx<<ROTATE_16);v11=v11+v15+vec2(0u,u32(v11.x+v15.x<v11.x));v7=((v7 ^ v11).xy<<ROTATE_1)|((v7 ^ v11).yx>>ROTATE_31);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v0=v0+m1+vec2(0u,u32(v0.x+m1.x<v0.x));v15=v15.yx ^ v0.yx;v10=v10+v15+vec2(0u,u32(v10.x+v15.x<v10.x));v5=((v5 ^ v10).xy>>ROTATE_24)|((v5 ^ v10).yx<<ROTATE_8);v0=v0+v5+vec2(0u,u32(v0.x+v5.x<v0.x));v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=v13.yx ^ v2.yx;v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));v7=((v7 ^ v8).xy>>ROTATE_24)|((v7 ^ v8).yx<<ROTATE_8);v2=v2+v7+vec2(0u,u32(v2.x+v7.x<v2.x));v13=((v13 ^ v2).xy>>ROTATE_16)|((v13 ^ v2).yx<<ROTATE_16);v8=v8+v13+vec2(0u,u32(v8.x+v13.x<v8.x));if((BLAKE2B_IV32_1 ^ v0.y ^ v8.y)>threshold&&atomicLoad(&work.found)==0u){atomicStore(&work.found,1u);work.nonce=m0;}return;}";
|
|
2
|
+
var compute_default = "struct UBO{blockhash:array<vec4<u32>,2>,seed:vec2<u32>,threshold:u32};@group(0)@binding(0)var<uniform> ubo:UBO;struct WORK{nonce:vec2<u32>,found:atomic<u32>};@group(0)@binding(1)var<storage,read_write>work:WORK;const BLAKE2B_IV_0=vec2(0xF2BDC900u,0x6A09E667u);const Z=vec2(0u);const CARRY=vec4(1u,0u,1u,0u);const ROTATE_1=vec4(1u);const ROTATE_8=vec4(8u);const ROTATE_16=vec4(16u);const ROTATE_24=vec4(24u);const ROTATE_31=vec4(31u);var<workgroup> found:bool;@compute @workgroup_size(32)fn search(@builtin(global_invocation_id)global_id:vec3<u32>,@builtin(local_invocation_id)local_id:vec3<u32>){found=(local_id.x==0u&&atomicLoad(&work.found)!=0u);workgroupBarrier();if(found){return;}main(global_id);}@compute @workgroup_size(1)fn validate(@builtin(global_invocation_id)global_id:vec3<u32>){main(global_id);}fn main(id:vec3<u32>){let m0:vec2<u32>=ubo.seed ^ id.xy;let m1:vec2<u32>=ubo.blockhash[0u].xy;let m2:vec2<u32>=ubo.blockhash[0u].zw;let m3:vec2<u32>=ubo.blockhash[1u].xy;let m4:vec2<u32>=ubo.blockhash[1u].zw;var v01:vec4<u32>=vec4(BLAKE2B_IV_0,0x84CAA73Bu,0xBB67AE85u);var v23:vec4<u32>=vec4(0xFE94F82Bu,0x3C6EF372u,0x5F1D36F1u,0xA54FF53Au);var v45:vec4<u32>=vec4(0xADE682D1u,0x510E527Fu,0x2B3E6C1Fu,0x9B05688Cu);var v67:vec4<u32>=vec4(0xFB41BD6Bu,0x1F83D9ABu,0x137E2179u,0x5BE0CD19u);var v89:vec4<u32>=vec4(0xF3BCC908u,0x6A09E667u,0x84CAA73Bu,0xBB67AE85u);var vAB:vec4<u32>=vec4(0xFE94F82Bu,0x3C6EF372u,0x5F1D36F1u,0xA54FF53Au);var vCD:vec4<u32>=vec4(0xADE682F9u,0x510E527Fu,0x2B3E6C1Fu,0x9B05688Cu);var vEF:vec4<u32>=vec4(0x04BE4294u,0xE07C2654u,0x137E2179u,0x5BE0CD19u);var v56:vec4<u32>;var vFC:vec4<u32>;var v74:vec4<u32>;var vDE:vec4<u32>;var s0:vec4<u32>;var s1:vec4<u32>;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s0=v01+vec4(m0,m2);v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;v23+=vec4(m4,Z);v23.y+=u32(v23.x<m4.x);vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s0=v01+vec4(m1,m3);v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(Z,m4);v01.w+=u32(v01.z<m4.x);vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s0=v01+vec4(m1,m0);v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(Z,m2);v01.w+=u32(v01.z<m2.x);v23+=vec4(Z,m3);v23.w+=u32(v23.z<m3.x);vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s0=v01+vec4(Z,m0);v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+vec4(m2,Z);v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01.z+=m3.x;v01.w+=m3.y+u32(v01.z<m3.x);vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(m1,m4)+vec4(0u,u32(v23.x+vec4(m1,m4).x<v23.x),0u,u32(v23.z+vec4(m1,m4).z<v23.z));vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(Z,m3)+vec4(Z,0u,u32(v01.z+vec4(Z,m3).z<v01.z));vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(Z,m1)+vec4(Z,0u,u32(v01.z+vec4(Z,m1).z<v01.z));vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(m2,Z)+vec4(0u,u32(v01.x+vec4(m2,Z).x<v01.x),Z);v23+=vec4(m4,Z)+vec4(0u,u32(v23.x+vec4(m4,Z).x<v23.x),Z);vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(m0,Z)+vec4(0u,u32(v23.x+vec4(m0,Z).x<v23.x),Z);vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s1=v23+vec4(m2,Z);v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(m0,Z)+vec4(0u,u32(v01.x+vec4(m0,Z).x<v01.x),Z);v23+=vec4(m4,Z)+vec4(0u,u32(v23.x+vec4(m4,Z).x<v23.x),Z);vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s1=v23+vec4(Z,m3);v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(m1,Z)+vec4(0u,u32(v01.x+vec4(m1,Z).x<v01.x),Z);vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(m2,Z)+vec4(0u,u32(v01.x+vec4(m2,Z).x<v01.x),Z);v23+=vec4(m0,Z)+vec4(0u,u32(v23.x+vec4(m0,Z).x<v23.x),Z);vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s1=v23+vec4(Z,m3);v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(m4,Z)+vec4(0u,u32(v01.x+vec4(m4,Z).x<v01.x),Z);v23+=vec4(Z,m1)+vec4(Z,0u,u32(v23.z+vec4(Z,m1).z<v23.z));vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(Z,m1)+vec4(Z,0u,u32(v01.z+vec4(Z,m1).z<v01.z));v23+=vec4(Z,m4)+vec4(Z,0u,u32(v23.z+vec4(Z,m4).z<v23.z));vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(m0,Z)+vec4(0u,u32(v01.x+vec4(m0,Z).x<v01.x),Z);vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(Z,m3)+vec4(Z,0u,u32(v01.z+vec4(Z,m3).z<v01.z));s1=v23+vec4(m2,Z);v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s1=v23+vec4(Z,m3);v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(m1,Z)+vec4(0u,u32(v23.x+vec4(m1,Z).x<v23.x),Z);vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(Z,m2)+vec4(Z,0u,u32(v23.z+vec4(Z,m2).z<v23.z));vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(m0,m4)+vec4(0u,u32(v01.x+vec4(m0,m4).x<v01.x),0u,u32(v01.z+vec4(m0,m4).z<v01.z));vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(Z,m0);v23.w+=u32(v23.z<m0.x);vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(m3,Z)+vec4(0u,u32(v23.x+vec4(m3,Z).x<v23.x),Z);vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(m1,Z)+vec4(0u,u32(v23.x+vec4(m1,Z).x<v23.x),Z);vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(m2,Z)+vec4(0u,u32(v01.x+vec4(m2,Z).x<v01.x),Z);v23+=vec4(m4,Z)+vec4(0u,u32(v23.x+vec4(m4,Z).x<v23.x),Z);vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(Z,m1);v23.w+=u32(v23.z<m1.x);vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s0=v01+vec4(m2,m4);v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(m3,Z);v23.y+=u32(v23.x<m3.x);vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v23+=vec4(Z,m0);v23.w+=u32(v23.z<m0.x);vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s0=v01+vec4(m0,m2);v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;v23+=vec4(m4,Z);v23.y+=u32(v23.x<m4.x);vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s0=v01+vec4(m1,m3);v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vFC ^=v01;vFC=(vFC>>ROTATE_16)|(vFC<<ROTATE_16).yxwz;vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_31).yxwz|(v56<<ROTATE_1);v74 ^=v89;v74=(v74>>ROTATE_31).yxwz|(v74<<ROTATE_1);v45=vec4(v74.zw,v56.xy);v67=vec4(v56.zw,v74.xy);vCD=vec4(vFC.zw,vDE.xy);vEF=vec4(vDE.zw,vFC.xy);s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(Z,m4);v01.w+=u32(v01.z<m4.x);vCD=(vCD ^ v01).yxwz;vEF=(vEF ^ v23).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_24)|(v45<<ROTATE_8).yxwz;v67 ^=vAB;v67=(v67>>ROTATE_24)|(v67<<ROTATE_8).yxwz;s0=v01+v45;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v67;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;vCD ^=v01;vCD=(vCD>>ROTATE_16)|(vCD<<ROTATE_16).yxwz;vEF ^=v23;vEF=(vEF>>ROTATE_16)|(vEF<<ROTATE_16).yxwz;s0=v89+vCD;v89=s0+(vec4<u32>(s0<v89)&CARRY).yxwz;s1=vAB+vEF;vAB=s1+(vec4<u32>(s1<vAB)&CARRY).yxwz;v45 ^=v89;v45=(v45>>ROTATE_31).yxwz|(v45<<ROTATE_1);v67 ^=vAB;v67=(v67>>ROTATE_31).yxwz|(v67<<ROTATE_1);v56=vec4(v45.zw,v67.xy);v74=vec4(v67.zw,v45.xy);vFC=vec4(vEF.zw,vCD.xy);vDE=vec4(vCD.zw,vEF.xy);s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;s0=v01+vec4(m1,m0);v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;vFC=(vFC ^ v01).yxwz;vDE=(vDE ^ v23).yxwz;s0=vAB+vFC;vAB=s0+(vec4<u32>(s0<vAB)&CARRY).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89)&CARRY).yxwz;v56 ^=vAB;v56=(v56>>ROTATE_24)|(v56<<ROTATE_8).yxwz;v74 ^=v89;v74=(v74>>ROTATE_24)|(v74<<ROTATE_8).yxwz;s0=v01+v56;v01=s0+(vec4<u32>(s0<v01)&CARRY).yxwz;s1=v23+v74;v23=s1+(vec4<u32>(s1<v23)&CARRY).yxwz;v01+=vec4(Z,m2);v01.w+=u32(v01.z<m2.x);v23+=vec4(Z,m3);v23.w+=u32(v23.z<m3.x);vDE ^=v23;vDE=(vDE>>ROTATE_16)|(vDE<<ROTATE_16).yxwz;s1=v89+vDE;v89=s1+(vec4<u32>(s1<v89).yxwz&CARRY);if((BLAKE2B_IV_0.y ^ v01.y ^ v89.y)>ubo.threshold&&atomicLoad(&work.found)==0u){atomicStore(&work.found,1u);work.nonce=m0;}return;}";
|
|
3
3
|
|
|
4
|
-
// src/shaders/gl-
|
|
5
|
-
var
|
|
4
|
+
// src/shaders/gl-downsample.ts
|
|
5
|
+
var NanoPowGlDownsampleShader = `#version 300 es
|
|
6
6
|
#pragma vscode_glsllint_stage: frag
|
|
7
|
+
#ifdef GL_FRAGMENT_PRECISION_HIGH
|
|
7
8
|
precision highp float;
|
|
9
|
+
#else
|
|
10
|
+
precision mediump float;
|
|
11
|
+
#endif
|
|
8
12
|
precision highp int;
|
|
9
13
|
|
|
10
|
-
|
|
11
|
-
out vec4 fragColor;
|
|
14
|
+
out uvec4 nonce;
|
|
12
15
|
|
|
13
|
-
//
|
|
16
|
+
// source texture to be downsampled
|
|
17
|
+
uniform highp usampler2D src;
|
|
18
|
+
|
|
19
|
+
void main() {
|
|
20
|
+
nonce = uvec4(0u);
|
|
21
|
+
vec2 inputSize = vec2(textureSize(src, 0));
|
|
22
|
+
vec2 texel = vec2(1.0) / inputSize;
|
|
23
|
+
vec2 blockCoord = (floor(gl_FragCoord.xy) * 2.0 + vec2(0.5)) / inputSize;
|
|
24
|
+
|
|
25
|
+
uvec4 pixel = texture(src, blockCoord);
|
|
26
|
+
nonce = pixel.x == 0u ? nonce : pixel;
|
|
27
|
+
|
|
28
|
+
pixel = texture(src, blockCoord + vec2(texel.x, 0.0));
|
|
29
|
+
nonce = pixel.x == 0u ? nonce : pixel;
|
|
30
|
+
|
|
31
|
+
pixel = texture(src, blockCoord + vec2(0.0, texel.y));
|
|
32
|
+
nonce = pixel.x == 0u ? nonce : pixel;
|
|
33
|
+
|
|
34
|
+
pixel = texture(src, blockCoord + vec2(texel.x, texel.y));
|
|
35
|
+
nonce = pixel.x == 0u ? nonce : pixel;
|
|
36
|
+
}
|
|
37
|
+
`;
|
|
38
|
+
|
|
39
|
+
// src/shaders/gl-draw.ts
|
|
40
|
+
var NanoPowGlDrawShader = `#version 300 es
|
|
41
|
+
#pragma vscode_glsllint_stage: frag
|
|
42
|
+
#ifdef GL_FRAGMENT_PRECISION_HIGH
|
|
43
|
+
precision highp float;
|
|
44
|
+
#else
|
|
45
|
+
precision mediump float;
|
|
46
|
+
#endif
|
|
47
|
+
|
|
48
|
+
out uvec4 nonce;
|
|
49
|
+
|
|
50
|
+
// blockhash - Array of precalculated block hash components
|
|
14
51
|
// threshold - 0xfffffff8 for send/change blocks, 0xfffffe00 for all else
|
|
15
|
-
//
|
|
52
|
+
// search - Checks all pixels if true, else only checks 1 pixel to validate
|
|
16
53
|
layout(std140) uniform UBO {
|
|
17
54
|
uint blockhash[8];
|
|
18
55
|
uint threshold;
|
|
19
|
-
|
|
56
|
+
bool search;
|
|
20
57
|
};
|
|
21
58
|
|
|
22
|
-
// Random work values
|
|
23
|
-
// First 2 bytes will be overwritten by texture pixel position
|
|
24
|
-
// Second 2 bytes will be modified if the canvas size is greater than 256x256
|
|
25
|
-
// Last 4 bytes remain as generated externally
|
|
59
|
+
// Random work seed values
|
|
26
60
|
layout(std140) uniform WORK {
|
|
27
|
-
|
|
61
|
+
uvec2 seed;
|
|
28
62
|
};
|
|
29
63
|
|
|
30
|
-
// Defined separately from uint v[
|
|
64
|
+
// Defined separately from uint v[0].y below as the original value is required
|
|
31
65
|
// to calculate the second uint32 of the digest for threshold comparison
|
|
32
66
|
const uint BLAKE2B_IV32_1 = 0x6A09E667u;
|
|
33
67
|
|
|
68
|
+
// Used during G for vector bit rotations
|
|
69
|
+
const uvec4 ROTATE_1 = uvec4(1u);
|
|
70
|
+
const uvec4 ROTATE_8 = uvec4(8u);
|
|
71
|
+
const uvec4 ROTATE_16 = uvec4(16u);
|
|
72
|
+
const uvec4 ROTATE_24 = uvec4(24u);
|
|
73
|
+
const uvec4 ROTATE_31 = uvec4(31u);
|
|
74
|
+
|
|
34
75
|
// Both buffers represent 16 uint64s as 32 uint32s
|
|
35
76
|
// because that's what GLSL offers, just like Javascript
|
|
36
77
|
|
|
@@ -39,168 +80,165 @@ const uint BLAKE2B_IV32_1 = 0x6A09E667u;
|
|
|
39
80
|
// OUTLEN is constant 8 bytes
|
|
40
81
|
// v[0] ^= 0x01010000u ^ uint(OUTLEN);
|
|
41
82
|
// INLEN is constant 40 bytes: work value (8) + block hash (32)
|
|
42
|
-
// v[
|
|
83
|
+
// v[12] ^= uint(INLEN);
|
|
43
84
|
// It's always the "last" compression at this INLEN
|
|
44
|
-
// v[
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
0xFE94F82Bu, 0x3C6EF372u,
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
)
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
const uint SIGMA82[192] = uint[192](
|
|
63
|
-
0u,2u,4u,6u,8u,10u,12u,14u,16u,18u,20u,22u,24u,26u,28u,30u,
|
|
64
|
-
28u,20u,8u,16u,18u,30u,26u,12u,2u,24u,0u,4u,22u,14u,10u,6u,
|
|
65
|
-
22u,16u,24u,0u,10u,4u,30u,26u,20u,28u,6u,12u,14u,2u,18u,8u,
|
|
66
|
-
14u,18u,6u,2u,26u,24u,22u,28u,4u,12u,10u,20u,8u,0u,30u,16u,
|
|
67
|
-
18u,0u,10u,14u,4u,8u,20u,30u,28u,2u,22u,24u,12u,16u,6u,26u,
|
|
68
|
-
4u,24u,12u,20u,0u,22u,16u,6u,8u,26u,14u,10u,30u,28u,2u,18u,
|
|
69
|
-
24u,10u,2u,30u,28u,26u,8u,20u,0u,14u,12u,6u,18u,4u,16u,22u,
|
|
70
|
-
26u,22u,14u,28u,24u,2u,6u,18u,10u,0u,30u,8u,16u,12u,4u,20u,
|
|
71
|
-
12u,30u,28u,18u,22u,6u,0u,16u,24u,4u,26u,14u,2u,8u,20u,10u,
|
|
72
|
-
20u,4u,16u,8u,14u,12u,2u,10u,30u,22u,18u,28u,6u,24u,26u,0u,
|
|
73
|
-
0u,2u,4u,6u,8u,10u,12u,14u,16u,18u,20u,22u,24u,26u,28u,30u,
|
|
74
|
-
28u,20u,8u,16u,18u,30u,26u,12u,2u,24u,0u,4u,22u,14u,10u,6u
|
|
85
|
+
// v[14] = ~v[14];
|
|
86
|
+
const uvec2 blake2b_iv[16] = uvec2[16](
|
|
87
|
+
uvec2(0xF2BDC900u, 0x6A09E667u),
|
|
88
|
+
uvec2(0x84CAA73Bu, 0xBB67AE85u),
|
|
89
|
+
uvec2(0xFE94F82Bu, 0x3C6EF372u),
|
|
90
|
+
uvec2(0x5F1D36F1u, 0xA54FF53Au),
|
|
91
|
+
uvec2(0xADE682D1u, 0x510E527Fu),
|
|
92
|
+
uvec2(0x2B3E6C1Fu, 0x9B05688Cu),
|
|
93
|
+
uvec2(0xFB41BD6Bu, 0x1F83D9ABu),
|
|
94
|
+
uvec2(0x137E2179u, 0x5BE0CD19u),
|
|
95
|
+
uvec2(0xF3BCC908u, 0x6A09E667u),
|
|
96
|
+
uvec2(0x84CAA73Bu, 0xBB67AE85u),
|
|
97
|
+
uvec2(0xFE94F82Bu, 0x3C6EF372u),
|
|
98
|
+
uvec2(0x5F1D36F1u, 0xA54FF53Au),
|
|
99
|
+
uvec2(0xADE682F9u, 0x510E527Fu),
|
|
100
|
+
uvec2(0x2B3E6C1Fu, 0x9B05688Cu),
|
|
101
|
+
uvec2(0x04BE4294u, 0xE07C2654u),
|
|
102
|
+
uvec2(0x137E2179u, 0x5BE0CD19u)
|
|
75
103
|
);
|
|
76
104
|
|
|
77
|
-
//
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
v[
|
|
91
|
-
v[
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
v[
|
|
113
|
-
v[
|
|
114
|
-
|
|
115
|
-
// b = rotr64(b ^ c, 24);
|
|
116
|
-
xor0 = v[b] ^ v[c];
|
|
117
|
-
xor1 = v[b+1u] ^ v[c+1u];
|
|
118
|
-
v[b] = (xor0 >> 24u) ^ (xor1 << 8u);
|
|
119
|
-
v[b+1u] = (xor1 >> 24u) ^ (xor0 << 8u);
|
|
120
|
-
|
|
121
|
-
// a = a + b;
|
|
122
|
-
o0 = v[a] + v[b];
|
|
123
|
-
o1 = v[a+1u] + v[b+1u];
|
|
124
|
-
carry = o0 < v[a];
|
|
125
|
-
o1 = o1 + uint(mix(0.0, 1.0, float(carry)));
|
|
126
|
-
v[a] = o0;
|
|
127
|
-
v[a+1u] = o1;
|
|
128
|
-
|
|
129
|
-
// a = a + m[sigma[r][2*i+1]];
|
|
130
|
-
o0 = v[a] + m[iy];
|
|
131
|
-
o1 = v[a+1u] + m[iy+1u];
|
|
132
|
-
carry = o0 < v[a];
|
|
133
|
-
o1 = o1 + uint(mix(0.0, 1.0, float(carry)));
|
|
134
|
-
v[a] = o0;
|
|
135
|
-
v[a+1u] = o1;
|
|
136
|
-
|
|
137
|
-
// d = rotr64(d ^ a, 16)
|
|
138
|
-
xor0 = v[d] ^ v[a];
|
|
139
|
-
xor1 = v[d+1u] ^ v[a+1u];
|
|
140
|
-
v[d] = (xor0 >> 16u) ^ (xor1 << 16u);
|
|
141
|
-
v[d+1u] = (xor1 >> 16u) ^ (xor0 << 16u);
|
|
142
|
-
|
|
143
|
-
// c = c + d;
|
|
144
|
-
o0 = v[c] + v[d];
|
|
145
|
-
o1 = v[c+1u] + v[d+1u];
|
|
146
|
-
carry = o0 < v[c];
|
|
147
|
-
o1 = o1 + uint(mix(0.0, 1.0, float(carry)));
|
|
148
|
-
v[c] = o0;
|
|
149
|
-
v[c+1u] = o1;
|
|
150
|
-
|
|
151
|
-
// b = rotr64(b ^ c, 63)
|
|
152
|
-
xor0 = v[b] ^ v[c];
|
|
153
|
-
xor1 = v[b+1u] ^ v[c+1u];
|
|
154
|
-
v[b] = (xor1 >> 31u) ^ (xor0 << 1u);
|
|
155
|
-
v[b+1u] = (xor0 >> 31u) ^ (xor1 << 1u);
|
|
105
|
+
// Iterated initialization vector
|
|
106
|
+
uvec2 v[16];
|
|
107
|
+
|
|
108
|
+
// Input data buffer
|
|
109
|
+
uvec2 m[16];
|
|
110
|
+
|
|
111
|
+
// G mixing function, compressing two subprocesses into one
|
|
112
|
+
void G (
|
|
113
|
+
uint a0, uint b0, uint c0, uint d0, uvec2 x0, uvec2 y0,
|
|
114
|
+
uint a1, uint b1, uint c1, uint d1, uvec2 x1, uvec2 y1
|
|
115
|
+
) {
|
|
116
|
+
uvec4 a = uvec4(v[a0], v[a1]);
|
|
117
|
+
uvec4 b = uvec4(v[b0], v[b1]);
|
|
118
|
+
uvec4 c = uvec4(v[c0], v[c1]);
|
|
119
|
+
uvec4 d = uvec4(v[d0], v[d1]);
|
|
120
|
+
uvec4 mx = uvec4(x0, x1);
|
|
121
|
+
uvec4 my = uvec4(y0, y1);
|
|
122
|
+
|
|
123
|
+
a = a + b + uvec4(0u, uint(a.x + b.x < a.x), 0u, uint(a.z + b.z < a.z));
|
|
124
|
+
a = a + mx + uvec4(0u, uint(a.x + mx.x < a.x), 0u, uint(a.z + mx.z < a.z));
|
|
125
|
+
d = (d ^ a).yxwz;
|
|
126
|
+
c = c + d + uvec4(0u, uint(c.x + d.x < c.x), 0u, uint(c.z + d.z < c.z));
|
|
127
|
+
b = ((b ^ c) >> ROTATE_24) | ((b ^ c) << ROTATE_8).yxwz;
|
|
128
|
+
a = a + b + uvec4(0u, uint(a.x + b.x < b.x), 0u, uint(a.z + b.z < b.z));
|
|
129
|
+
a = a + my + uvec4(0u, uint(a.x + my.x < a.x), 0u, uint(a.z + my.z < a.z));
|
|
130
|
+
d = ((d ^ a) >> ROTATE_16) | ((d ^ a) << ROTATE_16).yxwz;
|
|
131
|
+
c = c + d + uvec4(0u, uint(c.x + d.x < c.x), 0u, uint(c.z + d.z < c.z));
|
|
132
|
+
b = ((b ^ c) >> ROTATE_31).yxwz | ((b ^ c) << ROTATE_1);
|
|
133
|
+
|
|
134
|
+
v[a0] = a.xy;
|
|
135
|
+
v[b0] = b.xy;
|
|
136
|
+
v[c0] = c.xy;
|
|
137
|
+
v[d0] = d.xy;
|
|
138
|
+
v[a1] = a.zw;
|
|
139
|
+
v[b1] = b.zw;
|
|
140
|
+
v[c1] = c.zw;
|
|
141
|
+
v[d1] = d.zw;
|
|
156
142
|
}
|
|
157
143
|
|
|
158
144
|
void main() {
|
|
159
|
-
|
|
160
|
-
|
|
161
|
-
uvec4 u_work1 = work[1u];
|
|
162
|
-
uint uv_x = uint(uv_pos.x * workload);
|
|
163
|
-
uint uv_y = uint(uv_pos.y * workload);
|
|
164
|
-
uint x_pos = uv_x % 256u;
|
|
165
|
-
uint y_pos = uv_y % 256u;
|
|
166
|
-
uint x_index = (uv_x - x_pos) / 256u;
|
|
167
|
-
uint y_index = (uv_y - y_pos) / 256u;
|
|
168
|
-
|
|
169
|
-
// First 2 work bytes are the x,y pos within the 256x256 area, the next
|
|
170
|
-
// two bytes are modified from the random generated value, XOR'd with
|
|
171
|
-
// the x,y area index of where this pixel is located
|
|
172
|
-
m[0u] = (x_pos ^ (y_pos << 8u) ^ ((u_work0.b ^ x_index) << 16u) ^ ((u_work0.a ^ y_index) << 24u));
|
|
173
|
-
|
|
174
|
-
// Remaining bytes are un-modified from the random generated value
|
|
175
|
-
m[1u] = (u_work1.r ^ (u_work1.g << 8u) ^ (u_work1.b << 16u) ^ (u_work1.a << 24u));
|
|
145
|
+
// Initialize fragment output
|
|
146
|
+
nonce = uvec4(0u);
|
|
176
147
|
|
|
177
|
-
//
|
|
178
|
-
|
|
179
|
-
m[i+2u] = blockhash[i];
|
|
180
|
-
}
|
|
148
|
+
// Nonce uniquely differentiated by pixel location
|
|
149
|
+
m[0u] = seed ^ uvec2(gl_FragCoord);
|
|
181
150
|
|
|
182
|
-
//
|
|
183
|
-
|
|
184
|
-
|
|
185
|
-
|
|
186
|
-
|
|
187
|
-
|
|
188
|
-
|
|
189
|
-
|
|
190
|
-
|
|
191
|
-
|
|
151
|
+
// Block hash
|
|
152
|
+
m[1u] = uvec2(blockhash[0u], blockhash[1u]);
|
|
153
|
+
m[2u] = uvec2(blockhash[2u], blockhash[3u]);
|
|
154
|
+
m[3u] = uvec2(blockhash[4u], blockhash[5u]);
|
|
155
|
+
m[4u] = uvec2(blockhash[6u], blockhash[7u]);
|
|
156
|
+
|
|
157
|
+
// Reset v
|
|
158
|
+
v = blake2b_iv;
|
|
159
|
+
|
|
160
|
+
// Twelve rounds of G mixing
|
|
161
|
+
|
|
162
|
+
// Round 0
|
|
163
|
+
G(0u, 4u, 8u, 12u, m[0u], m[1u], 1u, 5u, 9u, 13u, m[2u], m[3u]);
|
|
164
|
+
G(2u, 6u, 10u, 14u, m[4u], m[5u], 3u, 7u, 11u, 15u, m[6u], m[7u]);
|
|
165
|
+
G(0u, 5u, 10u, 15u, m[8u], m[9u], 1u, 6u, 11u, 12u, m[10u], m[11u]);
|
|
166
|
+
G(2u, 7u, 8u, 13u, m[12u], m[13u], 3u, 4u, 9u, 14u, m[14u], m[15u]);
|
|
167
|
+
|
|
168
|
+
// Round 1
|
|
169
|
+
G(0u, 4u, 8u, 12u, m[14u], m[10u], 1u, 5u, 9u, 13u, m[4u], m[8u]);
|
|
170
|
+
G(2u, 6u, 10u, 14u, m[9u], m[15u], 3u, 7u, 11u, 15u, m[13u], m[6u]);
|
|
171
|
+
G(0u, 5u, 10u, 15u, m[1u], m[12u], 1u, 6u, 11u, 12u, m[0u], m[2u]);
|
|
172
|
+
G(2u, 7u, 8u, 13u, m[11u], m[7u], 3u, 4u, 9u, 14u, m[5u], m[3u]);
|
|
173
|
+
|
|
174
|
+
// Round 2
|
|
175
|
+
G(0u, 4u, 8u, 12u, m[11u], m[8u], 1u, 5u, 9u, 13u, m[12u], m[0u]);
|
|
176
|
+
G(2u, 6u, 10u, 14u, m[5u], m[2u], 3u, 7u, 11u, 15u, m[15u], m[13u]);
|
|
177
|
+
G(0u, 5u, 10u, 15u, m[10u], m[14u], 1u, 6u, 11u, 12u, m[3u], m[6u]);
|
|
178
|
+
G(2u, 7u, 8u, 13u, m[7u], m[1u], 3u, 4u, 9u, 14u, m[9u], m[4u]);
|
|
179
|
+
|
|
180
|
+
// Round 3
|
|
181
|
+
G(0u, 4u, 8u, 12u, m[7u], m[9u], 1u, 5u, 9u, 13u, m[3u], m[1u]);
|
|
182
|
+
G(2u, 6u, 10u, 14u, m[13u], m[12u], 3u, 7u, 11u, 15u, m[11u], m[14u]);
|
|
183
|
+
G(0u, 5u, 10u, 15u, m[2u], m[6u], 1u, 6u, 11u, 12u, m[5u], m[10u]);
|
|
184
|
+
G(2u, 7u, 8u, 13u, m[4u], m[0u], 3u, 4u, 9u, 14u, m[15u], m[8u]);
|
|
185
|
+
|
|
186
|
+
// Round 4
|
|
187
|
+
G(0u, 4u, 8u, 12u, m[9u], m[0u], 1u, 5u, 9u, 13u, m[5u], m[7u]);
|
|
188
|
+
G(2u, 6u, 10u, 14u, m[2u], m[4u], 3u, 7u, 11u, 15u, m[10u], m[15u]);
|
|
189
|
+
G(0u, 5u, 10u, 15u, m[14u], m[1u], 1u, 6u, 11u, 12u, m[11u], m[12u]);
|
|
190
|
+
G(2u, 7u, 8u, 13u, m[6u], m[8u], 3u, 4u, 9u, 14u, m[3u], m[13u]);
|
|
191
|
+
|
|
192
|
+
// Round 5
|
|
193
|
+
G(0u, 4u, 8u, 12u, m[2u], m[12u], 1u, 5u, 9u, 13u, m[6u], m[10u]);
|
|
194
|
+
G(2u, 6u, 10u, 14u, m[0u], m[11u], 3u, 7u, 11u, 15u, m[8u], m[3u]);
|
|
195
|
+
G(0u, 5u, 10u, 15u, m[4u], m[13u], 1u, 6u, 11u, 12u, m[7u], m[5u]);
|
|
196
|
+
G(2u, 7u, 8u, 13u, m[15u], m[14u], 3u, 4u, 9u, 14u, m[1u], m[9u]);
|
|
197
|
+
|
|
198
|
+
// Round 6
|
|
199
|
+
G(0u, 4u, 8u, 12u, m[12u], m[5u], 1u, 5u, 9u, 13u, m[1u], m[15u]);
|
|
200
|
+
G(2u, 6u, 10u, 14u, m[14u], m[13u], 3u, 7u, 11u, 15u, m[4u], m[10u]);
|
|
201
|
+
G(0u, 5u, 10u, 15u, m[0u], m[7u], 1u, 6u, 11u, 12u, m[6u], m[3u]);
|
|
202
|
+
G(2u, 7u, 8u, 13u, m[9u], m[2u], 3u, 4u, 9u, 14u, m[8u], m[11u]);
|
|
203
|
+
|
|
204
|
+
// Round 7
|
|
205
|
+
G(0u, 4u, 8u, 12u, m[13u], m[11u], 1u, 5u, 9u, 13u, m[7u], m[14u]);
|
|
206
|
+
G(2u, 6u, 10u, 14u, m[12u], m[1u], 3u, 7u, 11u, 15u, m[3u], m[9u]);
|
|
207
|
+
G(0u, 5u, 10u, 15u, m[5u], m[0u], 1u, 6u, 11u, 12u, m[15u], m[4u]);
|
|
208
|
+
G(2u, 7u, 8u, 13u, m[8u], m[6u], 3u, 4u, 9u, 14u, m[2u], m[10u]);
|
|
209
|
+
|
|
210
|
+
// Round 8
|
|
211
|
+
G(0u, 4u, 8u, 12u, m[6u], m[15u], 1u, 5u, 9u, 13u, m[14u], m[9u]);
|
|
212
|
+
G(2u, 6u, 10u, 14u, m[11u], m[3u], 3u, 7u, 11u, 15u, m[0u], m[8u]);
|
|
213
|
+
G(0u, 5u, 10u, 15u, m[12u], m[2u], 1u, 6u, 11u, 12u, m[13u], m[7u]);
|
|
214
|
+
G(2u, 7u, 8u, 13u, m[1u], m[4u], 3u, 4u, 9u, 14u, m[10u], m[5u]);
|
|
215
|
+
|
|
216
|
+
// Round 9
|
|
217
|
+
G(0u, 4u, 8u, 12u, m[10u], m[2u], 1u, 5u, 9u, 13u, m[8u], m[4u]);
|
|
218
|
+
G(2u, 6u, 10u, 14u, m[7u], m[6u], 3u, 7u, 11u, 15u, m[1u], m[5u]);
|
|
219
|
+
G(0u, 5u, 10u, 15u, m[15u], m[11u], 1u, 6u, 11u, 12u, m[9u], m[14u]);
|
|
220
|
+
G(2u, 7u, 8u, 13u, m[3u], m[12u], 3u, 4u, 9u, 14u, m[13u], m[0u]);
|
|
221
|
+
|
|
222
|
+
// Round 10
|
|
223
|
+
G(0u, 4u, 8u, 12u, m[0u], m[1u], 1u, 5u, 9u, 13u, m[2u], m[3u]);
|
|
224
|
+
G(2u, 6u, 10u, 14u, m[4u], m[5u], 3u, 7u, 11u, 15u, m[6u], m[7u]);
|
|
225
|
+
G(0u, 5u, 10u, 15u, m[8u], m[9u], 1u, 6u, 11u, 12u, m[10u], m[11u]);
|
|
226
|
+
G(2u, 7u, 8u, 13u, m[12u], m[13u], 3u, 4u, 9u, 14u, m[14u], m[15u]);
|
|
227
|
+
|
|
228
|
+
// Round 11
|
|
229
|
+
G(0u, 4u, 8u, 12u, m[14u], m[10u], 1u, 5u, 9u, 13u, m[4u], m[8u]);
|
|
230
|
+
G(2u, 6u, 10u, 14u, m[9u], m[15u], 3u, 7u, 11u, 15u, m[13u], m[6u]);
|
|
231
|
+
G(0u, 5u, 10u, 15u, m[1u], m[12u], 1u, 6u, 11u, 12u, m[0u], m[2u]);
|
|
232
|
+
G(2u, 7u, 8u, 13u, m[11u], m[7u], 3u, 4u, 9u, 14u, m[5u], m[3u]);
|
|
233
|
+
|
|
234
|
+
// Pixel data set from work seed values
|
|
235
|
+
// Finalize digest from high bits, low bits can be safely ignored
|
|
236
|
+
if ((BLAKE2B_IV32_1 ^ v[0u].y ^ v[8u].y) >= threshold && (search || uvec2(gl_FragCoord) == uvec2(0u))) {
|
|
237
|
+
nonce = uvec4(1u, m[0u].y, m[0u].x, (uint(gl_FragCoord.x) << 16u) | uint(gl_FragCoord.y));
|
|
192
238
|
}
|
|
193
239
|
|
|
194
|
-
//
|
|
195
|
-
|
|
196
|
-
if ((BLAKE2B_IV32_1 ^ v[1u] ^ v[17u]) > threshold) {
|
|
197
|
-
fragColor = vec4(
|
|
198
|
-
float(x_index + 1u)/255.0, // +1 to distinguish from 0 (unsuccessful) pixels
|
|
199
|
-
float(y_index + 1u)/255.0, // Same as previous
|
|
200
|
-
float(x_pos)/255.0, // Return the 2 custom bytes used in work value
|
|
201
|
-
float(y_pos)/255.0 // Second custom byte
|
|
202
|
-
);
|
|
203
|
-
} else {
|
|
240
|
+
// Valid nonce not found
|
|
241
|
+
if (nonce.x == 0u) {
|
|
204
242
|
discard;
|
|
205
243
|
}
|
|
206
244
|
}
|
|
@@ -209,14 +247,15 @@ void main() {
|
|
|
209
247
|
// src/shaders/gl-vertex.ts
|
|
210
248
|
var NanoPowGlVertexShader = `#version 300 es
|
|
211
249
|
#pragma vscode_glsllint_stage: vert
|
|
250
|
+
#ifdef GL_FRAGMENT_PRECISION_HIGH
|
|
212
251
|
precision highp float;
|
|
213
|
-
|
|
214
|
-
|
|
252
|
+
#else
|
|
253
|
+
precision mediump float;
|
|
254
|
+
#endif
|
|
215
255
|
|
|
216
|
-
|
|
256
|
+
layout (location=0) in vec4 position;
|
|
217
257
|
|
|
218
258
|
void main() {
|
|
219
|
-
uv_pos = uv;
|
|
220
259
|
gl_Position = position;
|
|
221
260
|
}
|
|
222
261
|
`;
|
|
@@ -224,59 +263,40 @@ void main() {
|
|
|
224
263
|
// src/classes/gl.ts
|
|
225
264
|
var NanoPowGl = class _NanoPowGl {
|
|
226
265
|
static #busy = false;
|
|
227
|
-
|
|
228
|
-
static #
|
|
229
|
-
|
|
230
|
-
|
|
231
|
-
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
return
|
|
266
|
+
static #debug = false;
|
|
267
|
+
static #raf = 0;
|
|
268
|
+
/** Used to set canvas size. */
|
|
269
|
+
static #cores = Math.max(1, Math.floor(navigator.hardwareConcurrency));
|
|
270
|
+
static #WORKLOAD = 256 * this.#cores;
|
|
271
|
+
static #canvas = new OffscreenCanvas(this.#WORKLOAD, this.#WORKLOAD);
|
|
272
|
+
static get size() {
|
|
273
|
+
return this.#gl?.drawingBufferWidth;
|
|
235
274
|
}
|
|
236
275
|
static #gl;
|
|
237
|
-
static #
|
|
276
|
+
static #drawProgram;
|
|
277
|
+
static #downsampleProgram;
|
|
238
278
|
static #vertexShader;
|
|
239
|
-
static #
|
|
279
|
+
static #drawShader;
|
|
280
|
+
static #downsampleShader;
|
|
240
281
|
static #positionBuffer;
|
|
241
|
-
static #
|
|
282
|
+
static #drawFbo;
|
|
283
|
+
static #downsampleFbos = [];
|
|
284
|
+
static #downsampleSrcLocation;
|
|
242
285
|
static #uboBuffer;
|
|
243
|
-
static #
|
|
286
|
+
static #uboView = new DataView(new ArrayBuffer(144));
|
|
287
|
+
static #seedBuffer;
|
|
288
|
+
static #seed = new BigUint64Array(1);
|
|
244
289
|
static #query;
|
|
245
290
|
static #pixels;
|
|
246
291
|
/**Vertex Positions, 2 triangles */
|
|
247
292
|
static #positions = new Float32Array([
|
|
248
293
|
-1,
|
|
249
294
|
-1,
|
|
250
|
-
0,
|
|
251
|
-
-1,
|
|
252
|
-
1,
|
|
253
|
-
0,
|
|
254
|
-
1,
|
|
255
|
-
1,
|
|
256
|
-
0,
|
|
257
295
|
1,
|
|
258
296
|
-1,
|
|
259
|
-
0,
|
|
260
297
|
1,
|
|
261
298
|
1,
|
|
262
|
-
0,
|
|
263
|
-
-1,
|
|
264
299
|
-1,
|
|
265
|
-
0
|
|
266
|
-
]);
|
|
267
|
-
/** Texture Positions */
|
|
268
|
-
static #uvPosArray = new Float32Array([
|
|
269
|
-
1,
|
|
270
|
-
1,
|
|
271
|
-
1,
|
|
272
|
-
0,
|
|
273
|
-
0,
|
|
274
|
-
0,
|
|
275
|
-
0,
|
|
276
|
-
1,
|
|
277
|
-
0,
|
|
278
|
-
0,
|
|
279
|
-
1,
|
|
280
300
|
1
|
|
281
301
|
]);
|
|
282
302
|
/** Compile */
|
|
@@ -284,70 +304,139 @@ var NanoPowGl = class _NanoPowGl {
|
|
|
284
304
|
if (this.#busy) return;
|
|
285
305
|
this.#busy = true;
|
|
286
306
|
try {
|
|
287
|
-
this.#
|
|
307
|
+
this.#canvas.addEventListener("webglcontextlost", (event) => {
|
|
308
|
+
event.preventDefault();
|
|
309
|
+
console.warn("WebGL context lost. Waiting for it to be restored...");
|
|
310
|
+
cancelAnimationFrame(this.#raf);
|
|
311
|
+
}, false);
|
|
312
|
+
this.#canvas.addEventListener("webglcontextrestored", (event) => {
|
|
313
|
+
console.warn("WebGL context restored. Reinitializing...");
|
|
314
|
+
_NanoPowGl.init();
|
|
315
|
+
}, false);
|
|
316
|
+
this.#gl = this.#canvas.getContext("webgl2");
|
|
288
317
|
if (this.#gl == null) throw new Error("WebGL 2 is required");
|
|
289
|
-
this.#gl.
|
|
290
|
-
this.#
|
|
291
|
-
if (this.#program == null) throw new Error("Failed to create shader program");
|
|
318
|
+
this.#drawProgram = this.#gl.createProgram();
|
|
319
|
+
if (this.#drawProgram == null) throw new Error("Failed to create shader program");
|
|
292
320
|
this.#vertexShader = this.#gl.createShader(this.#gl.VERTEX_SHADER);
|
|
293
321
|
if (this.#vertexShader == null) throw new Error("Failed to create vertex shader");
|
|
294
322
|
this.#gl.shaderSource(this.#vertexShader, NanoPowGlVertexShader);
|
|
295
323
|
this.#gl.compileShader(this.#vertexShader);
|
|
296
324
|
if (!this.#gl.getShaderParameter(this.#vertexShader, this.#gl.COMPILE_STATUS))
|
|
297
325
|
throw new Error(this.#gl.getShaderInfoLog(this.#vertexShader) ?? `Failed to compile vertex shader`);
|
|
298
|
-
this.#
|
|
299
|
-
if (this.#
|
|
300
|
-
this.#gl.shaderSource(this.#
|
|
301
|
-
this.#gl.compileShader(this.#
|
|
302
|
-
if (!this.#gl.getShaderParameter(this.#
|
|
303
|
-
throw new Error(this.#gl.getShaderInfoLog(this.#
|
|
304
|
-
this.#gl.attachShader(this.#
|
|
305
|
-
this.#gl.attachShader(this.#
|
|
306
|
-
this.#gl.linkProgram(this.#
|
|
307
|
-
if (!this.#gl.getProgramParameter(this.#
|
|
308
|
-
throw new Error(this.#gl.getProgramInfoLog(this.#
|
|
309
|
-
this.#gl.
|
|
326
|
+
this.#drawShader = this.#gl.createShader(this.#gl.FRAGMENT_SHADER);
|
|
327
|
+
if (this.#drawShader == null) throw new Error("Failed to create fragment shader");
|
|
328
|
+
this.#gl.shaderSource(this.#drawShader, NanoPowGlDrawShader);
|
|
329
|
+
this.#gl.compileShader(this.#drawShader);
|
|
330
|
+
if (!this.#gl.getShaderParameter(this.#drawShader, this.#gl.COMPILE_STATUS))
|
|
331
|
+
throw new Error(this.#gl.getShaderInfoLog(this.#drawShader) ?? `Failed to compile fragment shader`);
|
|
332
|
+
this.#gl.attachShader(this.#drawProgram, this.#vertexShader);
|
|
333
|
+
this.#gl.attachShader(this.#drawProgram, this.#drawShader);
|
|
334
|
+
this.#gl.linkProgram(this.#drawProgram);
|
|
335
|
+
if (!this.#gl.getProgramParameter(this.#drawProgram, this.#gl.LINK_STATUS))
|
|
336
|
+
throw new Error(this.#gl.getProgramInfoLog(this.#drawProgram) ?? `Failed to link program`);
|
|
337
|
+
this.#downsampleProgram = this.#gl.createProgram();
|
|
338
|
+
if (this.#downsampleProgram == null) throw new Error("Failed to create downsample program");
|
|
339
|
+
this.#downsampleShader = this.#gl.createShader(this.#gl.FRAGMENT_SHADER);
|
|
340
|
+
if (this.#downsampleShader == null) throw new Error("Failed to create downsample shader");
|
|
341
|
+
this.#gl.shaderSource(this.#downsampleShader, NanoPowGlDownsampleShader);
|
|
342
|
+
this.#gl.compileShader(this.#downsampleShader);
|
|
343
|
+
if (!this.#gl.getShaderParameter(this.#downsampleShader, this.#gl.COMPILE_STATUS))
|
|
344
|
+
throw new Error(this.#gl.getShaderInfoLog(this.#downsampleShader) ?? `Failed to compile downsample shader`);
|
|
345
|
+
this.#gl.attachShader(this.#downsampleProgram, this.#vertexShader);
|
|
346
|
+
this.#gl.attachShader(this.#downsampleProgram, this.#downsampleShader);
|
|
347
|
+
this.#gl.linkProgram(this.#downsampleProgram);
|
|
348
|
+
if (!this.#gl.getProgramParameter(this.#downsampleProgram, this.#gl.LINK_STATUS))
|
|
349
|
+
throw new Error(this.#gl.getProgramInfoLog(this.#downsampleProgram) ?? `Failed to link program`);
|
|
350
|
+
this.#gl.useProgram(this.#drawProgram);
|
|
310
351
|
const triangleArray = this.#gl.createVertexArray();
|
|
311
352
|
this.#gl.bindVertexArray(triangleArray);
|
|
312
353
|
this.#positionBuffer = this.#gl.createBuffer();
|
|
313
354
|
this.#gl.bindBuffer(this.#gl.ARRAY_BUFFER, this.#positionBuffer);
|
|
314
355
|
this.#gl.bufferData(this.#gl.ARRAY_BUFFER, this.#positions, this.#gl.STATIC_DRAW);
|
|
315
|
-
this.#gl.vertexAttribPointer(0,
|
|
356
|
+
this.#gl.vertexAttribPointer(0, 2, this.#gl.FLOAT, false, 0, 0);
|
|
316
357
|
this.#gl.enableVertexAttribArray(0);
|
|
317
|
-
this.#
|
|
318
|
-
this.#gl.
|
|
319
|
-
this.#gl.
|
|
320
|
-
this.#gl.
|
|
321
|
-
this.#gl.
|
|
358
|
+
this.#gl.bindBuffer(this.#gl.ARRAY_BUFFER, null);
|
|
359
|
+
const texture = this.#gl.createTexture();
|
|
360
|
+
this.#gl.bindTexture(this.#gl.TEXTURE_2D, texture);
|
|
361
|
+
this.#gl.texImage2D(this.#gl.TEXTURE_2D, 0, this.#gl.RGBA32UI, this.#gl.drawingBufferWidth, this.#gl.drawingBufferHeight, 0, this.#gl.RGBA_INTEGER, this.#gl.UNSIGNED_INT, null);
|
|
362
|
+
this.#gl.texParameteri(this.#gl.TEXTURE_2D, this.#gl.TEXTURE_MIN_FILTER, this.#gl.NEAREST);
|
|
363
|
+
this.#gl.texParameteri(this.#gl.TEXTURE_2D, this.#gl.TEXTURE_MAG_FILTER, this.#gl.NEAREST);
|
|
364
|
+
const framebuffer = this.#gl.createFramebuffer();
|
|
365
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, framebuffer);
|
|
366
|
+
this.#gl.framebufferTexture2D(this.#gl.FRAMEBUFFER, this.#gl.COLOR_ATTACHMENT0, this.#gl.TEXTURE_2D, texture, 0);
|
|
367
|
+
if (this.#gl.checkFramebufferStatus(this.#gl.FRAMEBUFFER) !== this.#gl.FRAMEBUFFER_COMPLETE)
|
|
368
|
+
throw new Error(`Failed to create drawing framebuffer`);
|
|
369
|
+
this.#drawFbo = { texture, framebuffer, size: { x: this.#gl.drawingBufferWidth, y: this.#gl.drawingBufferHeight } };
|
|
370
|
+
for (let i = 1; i <= 4; i++) {
|
|
371
|
+
const width = this.#gl.drawingBufferWidth / 2 ** i;
|
|
372
|
+
const height = this.#gl.drawingBufferHeight / 2 ** i;
|
|
373
|
+
const texture2 = this.#gl.createTexture();
|
|
374
|
+
this.#gl.bindTexture(this.#gl.TEXTURE_2D, texture2);
|
|
375
|
+
this.#gl.texImage2D(this.#gl.TEXTURE_2D, 0, this.#gl.RGBA32UI, width, height, 0, this.#gl.RGBA_INTEGER, this.#gl.UNSIGNED_INT, null);
|
|
376
|
+
this.#gl.texParameteri(this.#gl.TEXTURE_2D, this.#gl.TEXTURE_MIN_FILTER, this.#gl.NEAREST);
|
|
377
|
+
this.#gl.texParameteri(this.#gl.TEXTURE_2D, this.#gl.TEXTURE_MAG_FILTER, this.#gl.NEAREST);
|
|
378
|
+
const framebuffer2 = this.#gl.createFramebuffer();
|
|
379
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, framebuffer2);
|
|
380
|
+
this.#gl.framebufferTexture2D(this.#gl.FRAMEBUFFER, this.#gl.COLOR_ATTACHMENT0, this.#gl.TEXTURE_2D, texture2, 0);
|
|
381
|
+
if (this.#gl.checkFramebufferStatus(this.#gl.FRAMEBUFFER) !== this.#gl.FRAMEBUFFER_COMPLETE)
|
|
382
|
+
throw new Error(`Failed to create downsampling framebuffer ${i}`);
|
|
383
|
+
this.#downsampleFbos.push({ texture: texture2, framebuffer: framebuffer2, size: { x: width, y: height } });
|
|
384
|
+
}
|
|
385
|
+
this.#downsampleSrcLocation = this.#gl.getUniformLocation(this.#downsampleProgram, "src");
|
|
386
|
+
this.#gl.bindTexture(this.#gl.TEXTURE_2D, null);
|
|
387
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, null);
|
|
322
388
|
this.#uboBuffer = this.#gl.createBuffer();
|
|
323
389
|
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, this.#uboBuffer);
|
|
324
390
|
this.#gl.bufferData(this.#gl.UNIFORM_BUFFER, 144, this.#gl.DYNAMIC_DRAW);
|
|
325
391
|
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, null);
|
|
326
392
|
this.#gl.bindBufferBase(this.#gl.UNIFORM_BUFFER, 0, this.#uboBuffer);
|
|
327
|
-
this.#gl.uniformBlockBinding(this.#
|
|
328
|
-
this.#
|
|
329
|
-
this.#
|
|
330
|
-
this.#gl.
|
|
393
|
+
this.#gl.uniformBlockBinding(this.#drawProgram, this.#gl.getUniformBlockIndex(this.#drawProgram, "UBO"), 0);
|
|
394
|
+
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, null);
|
|
395
|
+
this.#seedBuffer = this.#gl.createBuffer();
|
|
396
|
+
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, this.#seedBuffer);
|
|
397
|
+
this.#gl.bufferData(this.#gl.UNIFORM_BUFFER, 16, this.#gl.DYNAMIC_DRAW);
|
|
398
|
+
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, null);
|
|
399
|
+
this.#gl.bindBufferBase(this.#gl.UNIFORM_BUFFER, 1, this.#seedBuffer);
|
|
400
|
+
this.#gl.uniformBlockBinding(this.#drawProgram, this.#gl.getUniformBlockIndex(this.#drawProgram, "WORK"), 1);
|
|
331
401
|
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, null);
|
|
332
|
-
this.#gl.bindBufferBase(this.#gl.UNIFORM_BUFFER, 1, this.#workBuffer);
|
|
333
|
-
this.#gl.uniformBlockBinding(this.#program, this.#gl.getUniformBlockIndex(this.#program, "WORK"), 1);
|
|
334
|
-
this.#pixels = new Uint8Array(this.#gl.drawingBufferWidth * this.#gl.drawingBufferHeight * 4);
|
|
335
402
|
this.#query = this.#gl.createQuery();
|
|
403
|
+
this.#pixels = new Uint32Array(this.#gl.drawingBufferWidth * this.#gl.drawingBufferHeight * 4);
|
|
404
|
+
console.log(`NanoPow WebGL initialized at ${this.#gl.drawingBufferWidth}x${this.#gl.drawingBufferHeight}. Maximum nonces checked per frame: ${this.#gl.drawingBufferWidth * this.#gl.drawingBufferHeight}`);
|
|
336
405
|
} catch (err) {
|
|
337
|
-
throw new Error(
|
|
406
|
+
throw new Error("WebGL initialization failed.", { cause: err });
|
|
338
407
|
} finally {
|
|
339
408
|
this.#busy = false;
|
|
340
409
|
}
|
|
341
410
|
}
|
|
342
411
|
static reset() {
|
|
412
|
+
cancelAnimationFrame(_NanoPowGl.#raf);
|
|
413
|
+
_NanoPowGl.#gl?.deleteQuery(_NanoPowGl.#query);
|
|
343
414
|
_NanoPowGl.#query = null;
|
|
344
|
-
_NanoPowGl.#
|
|
415
|
+
_NanoPowGl.#gl?.deleteBuffer(_NanoPowGl.#seedBuffer);
|
|
416
|
+
_NanoPowGl.#seedBuffer = null;
|
|
417
|
+
_NanoPowGl.#gl?.deleteBuffer(_NanoPowGl.#uboBuffer);
|
|
345
418
|
_NanoPowGl.#uboBuffer = null;
|
|
346
|
-
_NanoPowGl.#
|
|
419
|
+
for (const fbo of _NanoPowGl.#downsampleFbos) {
|
|
420
|
+
_NanoPowGl.#gl?.deleteFramebuffer(fbo.framebuffer);
|
|
421
|
+
_NanoPowGl.#gl?.deleteTexture(fbo.texture);
|
|
422
|
+
}
|
|
423
|
+
_NanoPowGl.#downsampleFbos = [];
|
|
424
|
+
_NanoPowGl.#gl?.deleteShader(_NanoPowGl.#downsampleShader);
|
|
425
|
+
_NanoPowGl.#downsampleShader = null;
|
|
426
|
+
_NanoPowGl.#gl?.deleteProgram(_NanoPowGl.#downsampleProgram);
|
|
427
|
+
_NanoPowGl.#downsampleProgram = null;
|
|
428
|
+
_NanoPowGl.#gl?.deleteFramebuffer(_NanoPowGl.#drawFbo?.framebuffer ?? null);
|
|
429
|
+
_NanoPowGl.#drawFbo = null;
|
|
430
|
+
_NanoPowGl.#gl?.deleteTexture(_NanoPowGl.#drawFbo);
|
|
431
|
+
_NanoPowGl.#drawFbo = null;
|
|
432
|
+
_NanoPowGl.#gl?.deleteBuffer(_NanoPowGl.#positionBuffer);
|
|
347
433
|
_NanoPowGl.#positionBuffer = null;
|
|
348
|
-
_NanoPowGl.#
|
|
434
|
+
_NanoPowGl.#gl?.deleteShader(_NanoPowGl.#drawShader);
|
|
435
|
+
_NanoPowGl.#drawShader = null;
|
|
436
|
+
_NanoPowGl.#gl?.deleteShader(_NanoPowGl.#vertexShader);
|
|
349
437
|
_NanoPowGl.#vertexShader = null;
|
|
350
|
-
_NanoPowGl.#
|
|
438
|
+
_NanoPowGl.#gl?.deleteProgram(_NanoPowGl.#drawProgram);
|
|
439
|
+
_NanoPowGl.#drawProgram = null;
|
|
351
440
|
_NanoPowGl.#gl = null;
|
|
352
441
|
_NanoPowGl.#busy = false;
|
|
353
442
|
_NanoPowGl.init();
|
|
@@ -374,34 +463,41 @@ var NanoPowGl = class _NanoPowGl {
|
|
|
374
463
|
"Harmonic Mean (ms)": count / reciprocals,
|
|
375
464
|
"Geometric Mean (ms)": Math.exp(logarithms / count)
|
|
376
465
|
};
|
|
466
|
+
console.log(`Averages: ${JSON.stringify(averages)}`);
|
|
377
467
|
console.table(averages);
|
|
378
468
|
}
|
|
379
|
-
static #draw(
|
|
469
|
+
static #draw(seed) {
|
|
380
470
|
if (this.#gl == null || this.#query == null) throw new Error("WebGL 2 is required to draw and query pixels");
|
|
381
|
-
if (this.#
|
|
382
|
-
this.#
|
|
383
|
-
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, this.#
|
|
384
|
-
this.#gl.bufferSubData(this.#gl.UNIFORM_BUFFER, 0,
|
|
471
|
+
if (this.#drawFbo == null) throw new Error("FBO is required to draw");
|
|
472
|
+
if (this.#seed[0] == null || this.#seedBuffer == null) throw new Error("Seed is required to draw");
|
|
473
|
+
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, this.#seedBuffer);
|
|
474
|
+
this.#gl.bufferSubData(this.#gl.UNIFORM_BUFFER, 0, seed);
|
|
385
475
|
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, null);
|
|
476
|
+
this.#gl.useProgram(this.#drawProgram);
|
|
477
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, this.#drawFbo.framebuffer);
|
|
478
|
+
this.#gl.activeTexture(this.#gl.TEXTURE0);
|
|
479
|
+
this.#gl.bindTexture(this.#gl.TEXTURE_2D, this.#drawFbo.texture);
|
|
386
480
|
this.#gl.beginQuery(this.#gl.ANY_SAMPLES_PASSED_CONSERVATIVE, this.#query);
|
|
387
|
-
this.#gl.
|
|
481
|
+
this.#gl.viewport(0, 0, this.#drawFbo.size.x, this.#drawFbo.size.y);
|
|
482
|
+
this.#gl.drawArrays(this.#gl.TRIANGLES, 0, 4);
|
|
483
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, null);
|
|
388
484
|
this.#gl.endQuery(this.#gl.ANY_SAMPLES_PASSED_CONSERVATIVE);
|
|
389
485
|
}
|
|
390
486
|
static async #checkQueryResult() {
|
|
391
487
|
return new Promise((resolve, reject) => {
|
|
392
|
-
|
|
393
|
-
|
|
394
|
-
|
|
395
|
-
|
|
396
|
-
|
|
397
|
-
|
|
398
|
-
|
|
399
|
-
|
|
400
|
-
|
|
488
|
+
function check() {
|
|
489
|
+
try {
|
|
490
|
+
if (_NanoPowGl.#gl == null || _NanoPowGl.#query == null) throw new Error("WebGL 2 is required to check query results");
|
|
491
|
+
if (_NanoPowGl.#gl.getQueryParameter(_NanoPowGl.#query, _NanoPowGl.#gl.QUERY_RESULT_AVAILABLE)) {
|
|
492
|
+
resolve(!!_NanoPowGl.#gl.getQueryParameter(_NanoPowGl.#query, _NanoPowGl.#gl.QUERY_RESULT));
|
|
493
|
+
} else {
|
|
494
|
+
_NanoPowGl.#raf = requestAnimationFrame(check);
|
|
495
|
+
}
|
|
496
|
+
} catch (err) {
|
|
497
|
+
reject(err);
|
|
401
498
|
}
|
|
402
|
-
} catch (err) {
|
|
403
|
-
reject(err);
|
|
404
499
|
}
|
|
500
|
+
check();
|
|
405
501
|
});
|
|
406
502
|
}
|
|
407
503
|
/**
|
|
@@ -409,21 +505,40 @@ var NanoPowGl = class _NanoPowGl {
|
|
|
409
505
|
* byte, converts the subsequent 3 pixels with the nonce byte values to a hex
|
|
410
506
|
* string, and returns the result.
|
|
411
507
|
*
|
|
412
|
-
* @param workBytes - Buffer with the original random nonce value
|
|
413
508
|
* @param workHex - Original nonce if provided for a validation call
|
|
414
509
|
* @returns Nonce as an 8-byte (16-char) hexadecimal string
|
|
415
510
|
*/
|
|
416
|
-
static #readResult(
|
|
511
|
+
static #readResult(workHex) {
|
|
417
512
|
if (this.#gl == null) throw new Error("WebGL 2 is required to read pixels");
|
|
418
|
-
this.#
|
|
419
|
-
|
|
513
|
+
if (this.#drawFbo == null) throw new Error("Source FBO is required to downsample");
|
|
514
|
+
let source = this.#drawFbo;
|
|
515
|
+
let pixelCount;
|
|
516
|
+
const start = performance.now();
|
|
517
|
+
if (workHex != null) {
|
|
518
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, source.framebuffer);
|
|
519
|
+
this.#gl.readPixels(0, 0, 1, 1, this.#gl.RGBA_INTEGER, this.#gl.UNSIGNED_INT, this.#pixels);
|
|
520
|
+
pixelCount = 4;
|
|
521
|
+
} else {
|
|
522
|
+
this.#gl.useProgram(this.#downsampleProgram);
|
|
523
|
+
for (const fbo of this.#downsampleFbos) {
|
|
524
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, fbo.framebuffer);
|
|
525
|
+
this.#gl.activeTexture(this.#gl.TEXTURE0);
|
|
526
|
+
this.#gl.bindTexture(this.#gl.TEXTURE_2D, source.texture);
|
|
527
|
+
this.#gl.uniform1i(this.#downsampleSrcLocation, 0);
|
|
528
|
+
this.#gl.viewport(0, 0, fbo.size.x, fbo.size.y);
|
|
529
|
+
this.#gl.drawArrays(this.#gl.TRIANGLES, 0, 4);
|
|
530
|
+
source = fbo;
|
|
531
|
+
}
|
|
532
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, source.framebuffer);
|
|
533
|
+
this.#gl.readPixels(0, 0, source.size.x, source.size.y, this.#gl.RGBA_INTEGER, this.#gl.UNSIGNED_INT, this.#pixels);
|
|
534
|
+
pixelCount = source.size.x * source.size.y * 4;
|
|
535
|
+
}
|
|
536
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, null);
|
|
537
|
+
for (let i = 0; i < pixelCount; i += 4) {
|
|
420
538
|
if (this.#pixels[i] !== 0) {
|
|
421
|
-
|
|
422
|
-
|
|
423
|
-
|
|
424
|
-
workBytes[2] ^ this.#pixels[i] - 1,
|
|
425
|
-
workBytes[3] ^ this.#pixels[i + 1] - 1
|
|
426
|
-
]);
|
|
539
|
+
if (this.#debug) console.log(`readResults (${performance.now() - start} ms)`);
|
|
540
|
+
if (this.#debug) console.log(`Pixel: rgba(${this.#pixels[i]}, ${this.#pixels[i + 1]}, ${this.#pixels[i + 2]}, ${this.#pixels[i + 3].toString(16).padStart(8, "0")})`);
|
|
541
|
+
const hex = `${this.#pixels[i + 1].toString(16).padStart(8, "0")}${this.#pixels[i + 2].toString(16).padStart(8, "0")}`;
|
|
427
542
|
if (workHex == null || workHex == hex) return hex;
|
|
428
543
|
}
|
|
429
544
|
}
|
|
@@ -437,6 +552,7 @@ var NanoPowGl = class _NanoPowGl {
|
|
|
437
552
|
*/
|
|
438
553
|
static async search(hash, options) {
|
|
439
554
|
if (this.#busy) {
|
|
555
|
+
console.log("NanoPowGl is busy. Retrying search...");
|
|
440
556
|
return new Promise((resolve) => {
|
|
441
557
|
setTimeout(async () => {
|
|
442
558
|
const result = this.search(hash, options);
|
|
@@ -445,42 +561,53 @@ var NanoPowGl = class _NanoPowGl {
|
|
|
445
561
|
});
|
|
446
562
|
}
|
|
447
563
|
this.#busy = true;
|
|
448
|
-
if (_NanoPowGl.#gl == null) throw new Error("WebGL 2 is required");
|
|
449
|
-
if (this.#gl == null) throw new Error("WebGL 2 is required");
|
|
450
564
|
if (!/^[A-Fa-f0-9]{64}$/.test(hash)) throw new Error(`Invalid hash ${hash}`);
|
|
451
565
|
const threshold = typeof options?.threshold !== "number" || options.threshold < 0 || options.threshold > 4294967295 ? 4294967288 : options.threshold;
|
|
452
|
-
const effort = typeof options?.effort !== "number" || options.effort < 1 || options.effort >
|
|
453
|
-
|
|
566
|
+
const effort = typeof options?.effort !== "number" || options.effort < 1 || options.effort > 32 ? this.#cores : options.effort;
|
|
567
|
+
this.#debug = !!options?.debug;
|
|
454
568
|
if (this.#WORKLOAD !== 256 * effort) {
|
|
455
569
|
this.#WORKLOAD = 256 * effort;
|
|
570
|
+
this.#canvas.height = this.#WORKLOAD;
|
|
571
|
+
this.#canvas.width = this.#WORKLOAD;
|
|
456
572
|
this.reset();
|
|
457
573
|
}
|
|
458
|
-
|
|
574
|
+
if (_NanoPowGl.#gl == null) throw new Error("WebGL 2 is required");
|
|
575
|
+
if (this.#gl == null) throw new Error("WebGL 2 is required");
|
|
576
|
+
if (this.#drawFbo == null) throw new Error("WebGL framebuffer is required");
|
|
577
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, this.#drawFbo.framebuffer);
|
|
578
|
+
this.#gl.clearBufferuiv(this.#gl.COLOR, 0, [0, 0, 0, 0]);
|
|
579
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, null);
|
|
580
|
+
for (let i = 0; i < this.#uboView.byteLength; i++) this.#uboView.setUint8(i, 0);
|
|
459
581
|
for (let i = 0; i < 64; i += 8) {
|
|
460
582
|
const uint32 = hash.slice(i, i + 8);
|
|
461
|
-
uboView.setUint32(i * 2, parseInt(uint32, 16));
|
|
583
|
+
this.#uboView.setUint32(i * 2, parseInt(uint32, 16));
|
|
462
584
|
}
|
|
463
|
-
uboView.setUint32(128, threshold, true);
|
|
464
|
-
uboView.
|
|
465
|
-
|
|
466
|
-
|
|
467
|
-
|
|
585
|
+
this.#uboView.setUint32(128, threshold, true);
|
|
586
|
+
this.#uboView.setUint32(132, 1, true);
|
|
587
|
+
if (this.#debug) console.log("UBO", this.#uboView.buffer.slice(0));
|
|
588
|
+
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, this.#uboBuffer);
|
|
589
|
+
this.#gl.bufferSubData(this.#gl.UNIFORM_BUFFER, 0, this.#uboView);
|
|
590
|
+
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, null);
|
|
468
591
|
let times = [];
|
|
469
592
|
let start = performance.now();
|
|
470
593
|
let nonce = null;
|
|
471
|
-
|
|
594
|
+
if (this.#debug) console.groupCollapsed("Seeds (click to view)");
|
|
472
595
|
while (nonce == null) {
|
|
473
596
|
start = performance.now();
|
|
474
|
-
|
|
475
|
-
|
|
597
|
+
const random0 = Math.floor(Math.random() * 4294967295);
|
|
598
|
+
const random1 = Math.floor(Math.random() * 4294967295);
|
|
599
|
+
this.#seed[0] = BigInt(random0) << 32n | BigInt(random1);
|
|
600
|
+
if (this.#debug) console.log("Seed", this.#seed);
|
|
601
|
+
this.#draw(this.#seed);
|
|
476
602
|
const found = await this.#checkQueryResult();
|
|
477
603
|
times.push(performance.now() - start);
|
|
478
604
|
if (found) {
|
|
479
|
-
|
|
605
|
+
if (this.#debug) console.groupEnd();
|
|
606
|
+
nonce = this.#readResult();
|
|
480
607
|
}
|
|
481
608
|
}
|
|
482
609
|
this.#busy = false;
|
|
483
|
-
if (debug) this.#logAverages(times);
|
|
610
|
+
if (this.#debug) this.#logAverages(times);
|
|
484
611
|
return nonce;
|
|
485
612
|
}
|
|
486
613
|
/**
|
|
@@ -492,6 +619,7 @@ var NanoPowGl = class _NanoPowGl {
|
|
|
492
619
|
*/
|
|
493
620
|
static async validate(work, hash, options) {
|
|
494
621
|
if (this.#busy) {
|
|
622
|
+
console.log("NanoPowGl is busy. Retrying validate...");
|
|
495
623
|
return new Promise((resolve) => {
|
|
496
624
|
setTimeout(async () => {
|
|
497
625
|
const result = this.validate(work, hash, options);
|
|
@@ -500,31 +628,35 @@ var NanoPowGl = class _NanoPowGl {
|
|
|
500
628
|
});
|
|
501
629
|
}
|
|
502
630
|
this.#busy = true;
|
|
503
|
-
if (_NanoPowGl.#gl == null) throw new Error("WebGL 2 is required");
|
|
504
|
-
if (this.#gl == null) throw new Error("WebGL 2 is required");
|
|
505
631
|
if (!/^[A-Fa-f0-9]{16}$/.test(work)) throw new Error(`Invalid work ${work}`);
|
|
506
632
|
if (!/^[A-Fa-f0-9]{64}$/.test(hash)) throw new Error(`Invalid hash ${hash}`);
|
|
507
633
|
const threshold = typeof options?.threshold !== "number" || options.threshold < 0 || options.threshold > 4294967295 ? 4294967288 : options.threshold;
|
|
508
|
-
|
|
509
|
-
|
|
634
|
+
this.#debug = !!options?.debug;
|
|
635
|
+
if (_NanoPowGl.#gl == null) throw new Error("WebGL 2 is required");
|
|
636
|
+
if (this.#gl == null) throw new Error("WebGL 2 is required");
|
|
637
|
+
if (this.#drawFbo == null) throw new Error("WebGL framebuffer is required");
|
|
638
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, this.#drawFbo.framebuffer);
|
|
639
|
+
this.#gl.clearBufferuiv(this.#gl.COLOR, 0, [0, 0, 0, 0]);
|
|
640
|
+
this.#gl.bindFramebuffer(this.#gl.FRAMEBUFFER, null);
|
|
641
|
+
for (let i = 0; i < this.#uboView.byteLength; i++) this.#uboView.setUint8(i, 0);
|
|
510
642
|
for (let i = 0; i < 64; i += 8) {
|
|
511
643
|
const uint32 = hash.slice(i, i + 8);
|
|
512
|
-
uboView.setUint32(i * 2, parseInt(uint32, 16));
|
|
644
|
+
this.#uboView.setUint32(i * 2, parseInt(uint32, 16));
|
|
513
645
|
}
|
|
514
|
-
uboView.setUint32(128, threshold, true);
|
|
515
|
-
uboView.
|
|
516
|
-
|
|
517
|
-
|
|
518
|
-
|
|
646
|
+
this.#uboView.setUint32(128, threshold, true);
|
|
647
|
+
this.#uboView.setUint32(132, 0, true);
|
|
648
|
+
if (this.#debug) console.log("UBO", this.#uboView.buffer.slice(0));
|
|
649
|
+
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, this.#uboBuffer);
|
|
650
|
+
this.#gl.bufferSubData(this.#gl.UNIFORM_BUFFER, 0, this.#uboView);
|
|
651
|
+
this.#gl.bindBuffer(this.#gl.UNIFORM_BUFFER, null);
|
|
519
652
|
let nonce = null;
|
|
520
|
-
|
|
521
|
-
|
|
522
|
-
|
|
523
|
-
this.#draw(seed);
|
|
653
|
+
this.#seed[0] = BigInt(`0x${work}`);
|
|
654
|
+
if (this.#debug) console.log("Work", this.#seed);
|
|
655
|
+
this.#draw(this.#seed);
|
|
524
656
|
let found = await this.#checkQueryResult();
|
|
525
657
|
if (found) {
|
|
526
658
|
try {
|
|
527
|
-
nonce = this.#readResult(
|
|
659
|
+
nonce = this.#readResult(work);
|
|
528
660
|
} catch (err) {
|
|
529
661
|
found = false;
|
|
530
662
|
}
|
|
@@ -539,10 +671,13 @@ var NanoPowGl = class _NanoPowGl {
|
|
|
539
671
|
var NanoPowGpu = class _NanoPowGpu {
|
|
540
672
|
// Initialize WebGPU
|
|
541
673
|
static #busy = false;
|
|
674
|
+
static #debug = false;
|
|
542
675
|
static #device = null;
|
|
543
|
-
static #
|
|
676
|
+
static #gpuBufferReset = new BigUint64Array([0n, 0n]);
|
|
544
677
|
static #gpuBuffer;
|
|
545
678
|
static #cpuBuffer;
|
|
679
|
+
static #uboBuffer;
|
|
680
|
+
static #uboView;
|
|
546
681
|
static #bindGroupLayout;
|
|
547
682
|
static #searchPipeline;
|
|
548
683
|
static #validatePipeline;
|
|
@@ -560,17 +695,13 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
560
695
|
this.#device = device;
|
|
561
696
|
this.setup();
|
|
562
697
|
} catch (err) {
|
|
563
|
-
throw new Error(
|
|
698
|
+
throw new Error("WebGPU initialization failed.", { cause: err });
|
|
564
699
|
} finally {
|
|
565
700
|
this.#busy = false;
|
|
566
701
|
}
|
|
567
702
|
}
|
|
568
703
|
static setup() {
|
|
569
704
|
if (this.#device == null) throw new Error(`WebGPU device failed to load.`);
|
|
570
|
-
this.#uboBuffer = this.#device.createBuffer({
|
|
571
|
-
size: 48,
|
|
572
|
-
usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST
|
|
573
|
-
});
|
|
574
705
|
this.#gpuBuffer = this.#device.createBuffer({
|
|
575
706
|
size: 16,
|
|
576
707
|
usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST | GPUBufferUsage.COPY_SRC
|
|
@@ -579,6 +710,11 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
579
710
|
size: 16,
|
|
580
711
|
usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ
|
|
581
712
|
});
|
|
713
|
+
this.#uboBuffer = this.#device.createBuffer({
|
|
714
|
+
size: 48,
|
|
715
|
+
usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST
|
|
716
|
+
});
|
|
717
|
+
this.#uboView = new DataView(new ArrayBuffer(48));
|
|
582
718
|
this.#bindGroupLayout = this.#device.createBindGroupLayout({
|
|
583
719
|
entries: [
|
|
584
720
|
{
|
|
@@ -614,6 +750,7 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
614
750
|
module: shaderModule
|
|
615
751
|
}
|
|
616
752
|
});
|
|
753
|
+
console.log(`NanoPow WebGPU initialized. Recommended effort: ${Math.max(1, Math.floor(navigator.hardwareConcurrency / 2))}`);
|
|
617
754
|
}
|
|
618
755
|
static reset() {
|
|
619
756
|
console.warn(`GPU device lost. Reinitializing...`);
|
|
@@ -624,7 +761,7 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
624
761
|
_NanoPowGpu.init();
|
|
625
762
|
}
|
|
626
763
|
static #logAverages(times) {
|
|
627
|
-
let count = times.length,
|
|
764
|
+
let count = times.length, truncatedCount = 0, truncated = 0, sum = 0, reciprocals = 0, logarithms = 0, min = Number.MAX_SAFE_INTEGER, max = 0, median = 0, rate = 0;
|
|
628
765
|
times.sort();
|
|
629
766
|
for (let i = 0; i < count; i++) {
|
|
630
767
|
sum += times[i];
|
|
@@ -632,16 +769,23 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
632
769
|
logarithms += Math.log(times[i]);
|
|
633
770
|
min = Math.min(min, times[i]);
|
|
634
771
|
max = Math.max(max, times[i]);
|
|
635
|
-
if (
|
|
772
|
+
if (i === Math.ceil(count / 2)) {
|
|
773
|
+
median = times[i];
|
|
774
|
+
}
|
|
775
|
+
if (count < 3 || i > 0.1 * count && i < 0.9 * (count - 1)) {
|
|
776
|
+
truncated += times[i];
|
|
777
|
+
truncatedCount++;
|
|
778
|
+
}
|
|
636
779
|
}
|
|
637
780
|
const averages = {
|
|
638
781
|
"Count (dispatches)": count,
|
|
639
782
|
"Total (ms)": sum,
|
|
640
|
-
"Rate (d/s)": 1e3 *
|
|
783
|
+
"Rate (d/s)": 1e3 * truncatedCount / (truncated || sum),
|
|
641
784
|
"Minimum (ms)": min,
|
|
642
785
|
"Maximum (ms)": max,
|
|
786
|
+
"Median (ms)": median,
|
|
643
787
|
"Arithmetic Mean (ms)": sum / count,
|
|
644
|
-
"Truncated Mean (ms)": truncated /
|
|
788
|
+
"Truncated Mean (ms)": truncated / truncatedCount,
|
|
645
789
|
"Harmonic Mean (ms)": count / reciprocals,
|
|
646
790
|
"Geometric Mean (ms)": Math.exp(logarithms / count)
|
|
647
791
|
};
|
|
@@ -649,15 +793,16 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
649
793
|
}
|
|
650
794
|
static async #dispatch(pipeline, seed, hash, threshold, passes) {
|
|
651
795
|
if (this.#device == null) throw new Error(`WebGPU device failed to load.`);
|
|
652
|
-
|
|
653
|
-
for (let i = 0; i < 64; i +=
|
|
654
|
-
const
|
|
655
|
-
uboView.
|
|
796
|
+
for (let i = 0; i < this.#uboView.byteLength; i++) this.#uboView.setUint8(i, 0);
|
|
797
|
+
for (let i = 0; i < 64; i += 16) {
|
|
798
|
+
const u64 = hash.slice(i, i + 16);
|
|
799
|
+
this.#uboView.setBigUint64(i / 2, BigInt(`0x${u64}`));
|
|
656
800
|
}
|
|
657
|
-
uboView.setBigUint64(32, seed, true);
|
|
658
|
-
uboView.setUint32(40, threshold, true);
|
|
659
|
-
this.#
|
|
660
|
-
this.#device.queue.writeBuffer(this.#
|
|
801
|
+
this.#uboView.setBigUint64(32, seed, true);
|
|
802
|
+
this.#uboView.setUint32(40, threshold, true);
|
|
803
|
+
if (this.#debug) console.log("UBO", this.#uboView);
|
|
804
|
+
this.#device.queue.writeBuffer(this.#uboBuffer, 0, this.#uboView);
|
|
805
|
+
this.#device.queue.writeBuffer(this.#gpuBuffer, 0, this.#gpuBufferReset);
|
|
661
806
|
const bindGroup = this.#device.createBindGroup({
|
|
662
807
|
layout: this.#bindGroupLayout,
|
|
663
808
|
entries: [
|
|
@@ -693,6 +838,7 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
693
838
|
console.warn(`Error getting data from GPU. ${err}`);
|
|
694
839
|
return this.#dispatch(pipeline, seed, hash, threshold, passes);
|
|
695
840
|
}
|
|
841
|
+
if (this.#debug) console.log("gpuBuffer data", data);
|
|
696
842
|
if (data == null) throw new Error(`Failed to get data from buffer.`);
|
|
697
843
|
return data;
|
|
698
844
|
}
|
|
@@ -703,7 +849,9 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
703
849
|
* @param {NanoPowOptions} options - Used to configure search execution
|
|
704
850
|
*/
|
|
705
851
|
static async search(hash, options) {
|
|
852
|
+
if (!/^[A-Fa-f0-9]{64}$/.test(hash)) throw new TypeError(`Invalid hash ${hash}`);
|
|
706
853
|
if (this.#busy) {
|
|
854
|
+
console.log("NanoPowGpu is busy. Retrying search...");
|
|
707
855
|
return new Promise((resolve) => {
|
|
708
856
|
setTimeout(async () => {
|
|
709
857
|
const result = this.search(hash, options);
|
|
@@ -712,30 +860,34 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
712
860
|
});
|
|
713
861
|
}
|
|
714
862
|
this.#busy = true;
|
|
715
|
-
if (!/^[A-Fa-f0-9]{64}$/.test(hash)) throw new TypeError(`Invalid hash ${hash}`);
|
|
716
863
|
const threshold = typeof options?.threshold !== "number" || options.threshold < 0 || options.threshold > 4294967295 ? 4294967288 : options.threshold;
|
|
717
864
|
const effort = typeof options?.effort !== "number" || options.effort < 1 || options.effort > 32 ? 2048 : options.effort * 256;
|
|
718
|
-
|
|
865
|
+
this.#debug = !!options?.debug;
|
|
719
866
|
let loads = 0;
|
|
720
867
|
while (this.#device == null && loads < 20) {
|
|
721
868
|
await new Promise((resolve) => {
|
|
722
869
|
setTimeout(resolve, 500);
|
|
723
870
|
});
|
|
724
871
|
}
|
|
725
|
-
if (this.#device == null)
|
|
872
|
+
if (this.#device == null) {
|
|
873
|
+
this.#busy = false;
|
|
874
|
+
throw new Error(`WebGPU device failed to load.`);
|
|
875
|
+
}
|
|
726
876
|
let times = [];
|
|
727
877
|
let start = performance.now();
|
|
728
878
|
let nonce = 0n;
|
|
729
879
|
do {
|
|
730
880
|
start = performance.now();
|
|
731
|
-
const
|
|
732
|
-
const
|
|
881
|
+
const random0 = Math.floor(Math.random() * 4294967295);
|
|
882
|
+
const random1 = Math.floor(Math.random() * 4294967295);
|
|
883
|
+
const seed = BigInt(random0) << 32n | BigInt(random1);
|
|
884
|
+
if (this.#debug) console.log(`seed: ${seed}`);
|
|
733
885
|
const data = await this.#dispatch(this.#searchPipeline, seed, hash, threshold, effort);
|
|
734
886
|
nonce = data.getBigUint64(0, true);
|
|
735
887
|
this.#busy = !data.getUint32(8);
|
|
736
888
|
times.push(performance.now() - start);
|
|
737
889
|
} while (this.#busy);
|
|
738
|
-
if (debug) this.#logAverages(times);
|
|
890
|
+
if (this.#debug) this.#logAverages(times);
|
|
739
891
|
return nonce.toString(16).padStart(16, "0");
|
|
740
892
|
}
|
|
741
893
|
/**
|
|
@@ -746,7 +898,10 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
746
898
|
* @param {NanoPowOptions} options - Options used to configure search execution
|
|
747
899
|
*/
|
|
748
900
|
static async validate(work, hash, options) {
|
|
901
|
+
if (!/^[A-Fa-f0-9]{16}$/.test(work)) throw new TypeError(`Invalid work ${work}`);
|
|
902
|
+
if (!/^[A-Fa-f0-9]{64}$/.test(hash)) throw new TypeError(`Invalid hash ${hash}`);
|
|
749
903
|
if (this.#busy) {
|
|
904
|
+
console.log("NanoPowGpu is busy. Retrying validate...");
|
|
750
905
|
return new Promise((resolve) => {
|
|
751
906
|
setTimeout(async () => {
|
|
752
907
|
const result = this.validate(work, hash, options);
|
|
@@ -755,9 +910,7 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
755
910
|
});
|
|
756
911
|
}
|
|
757
912
|
this.#busy = true;
|
|
758
|
-
|
|
759
|
-
if (!/^[A-Fa-f0-9]{64}$/.test(hash)) throw new TypeError(`Invalid hash ${hash}`);
|
|
760
|
-
const debug = !!options?.debug;
|
|
913
|
+
this.#debug = !!options?.debug;
|
|
761
914
|
const threshold = typeof options?.threshold !== "number" || options.threshold < 0 || options.threshold > 4294967295 ? 4294967288 : options.threshold;
|
|
762
915
|
let loads = 0;
|
|
763
916
|
while (this.#device == null && loads < 20) {
|
|
@@ -765,10 +918,15 @@ var NanoPowGpu = class _NanoPowGpu {
|
|
|
765
918
|
setTimeout(resolve, 500);
|
|
766
919
|
});
|
|
767
920
|
}
|
|
768
|
-
if (this.#device == null)
|
|
921
|
+
if (this.#device == null) {
|
|
922
|
+
this.#busy = false;
|
|
923
|
+
throw new Error(`WebGPU device failed to load.`);
|
|
924
|
+
}
|
|
769
925
|
const seed = BigInt(`0x${work}`);
|
|
926
|
+
if (this.#debug) console.log(`work: ${work}`);
|
|
770
927
|
const data = await this.#dispatch(this.#validatePipeline, seed, hash, threshold, 1);
|
|
771
928
|
const nonce = data.getBigUint64(0, true).toString(16).padStart(16, "0");
|
|
929
|
+
if (this.#debug) console.log(`nonce: ${nonce}`);
|
|
772
930
|
const found = !!data.getUint32(8);
|
|
773
931
|
this.#busy = false;
|
|
774
932
|
if (found && work !== nonce) throw new Error(`Nonce (${nonce}) found but does not match work (${work})`);
|
|
@@ -783,14 +941,14 @@ try {
|
|
|
783
941
|
await NanoPowGpu.init();
|
|
784
942
|
isGpuSupported = true;
|
|
785
943
|
} catch (err) {
|
|
786
|
-
console.warn(
|
|
944
|
+
console.warn("WebGPU is not supported in this environment.\n", err);
|
|
787
945
|
isGpuSupported = false;
|
|
788
946
|
}
|
|
789
947
|
try {
|
|
790
948
|
await NanoPowGl.init();
|
|
791
949
|
isGlSupported = true;
|
|
792
950
|
} catch (err) {
|
|
793
|
-
console.warn(
|
|
951
|
+
console.warn("WebGL is not supported in this environment.\n", err);
|
|
794
952
|
isGlSupported = false;
|
|
795
953
|
}
|
|
796
954
|
var NanoPow = isGpuSupported ? NanoPowGpu : isGlSupported ? NanoPowGl : null;
|