From d3cf2e37bb6ad32740359b51c7e50f87432df2cd Mon Sep 17 00:00:00 2001 From: Chris Duncan Date: Mon, 6 Jan 2025 05:27:21 -0800 Subject: [PATCH] Substitute remaining SIGMA82 array lookups with literal values. --- src/lib/workers/powgpu.ts | 150 ++++++++++++++------------------------ 1 file changed, 56 insertions(+), 94 deletions(-) diff --git a/src/lib/workers/powgpu.ts b/src/lib/workers/powgpu.ts index e265695..06e1340 100644 --- a/src/lib/workers/powgpu.ts +++ b/src/lib/workers/powgpu.ts @@ -86,20 +86,6 @@ export class PowGpu extends WorkerInterface { * 28u,20u,8u,16u,18u,30u,26u,12u,2u,24u,0u,4u,22u,14u,10u,6u * ); */ - const SIGMA82: array = array( - 0u,2u,4u,6u,8u,10u,12u,14u,16u,18u,20u,22u,24u,26u,28u,30u, - 28u,20u,8u,16u,18u,30u,26u,12u,2u,24u,0u,4u,22u,14u,10u,6u, - 22u,16u,24u,0u,10u,4u,30u,26u,20u,28u,6u,12u,14u,2u,18u,8u, - 14u,18u,6u,2u,26u,24u,22u,28u,4u,12u,10u,20u,8u,0u,30u,16u, - 18u,0u,10u,14u,4u,8u,20u,30u,28u,2u,22u,24u,12u,16u,6u,26u, - 4u,24u,12u,20u,0u,22u,16u,6u,8u,26u,14u,10u,30u,28u,2u,18u, - 24u,10u,2u,30u,28u,26u,8u,20u,0u,14u,12u,6u,18u,4u,16u,22u, - 26u,22u,14u,28u,24u,2u,6u,18u,10u,0u,30u,8u,16u,12u,4u,20u, - 12u,30u,28u,18u,22u,6u,0u,16u,24u,4u,26u,14u,2u,8u,20u,10u, - 20u,4u,16u,8u,14u,12u,2u,10u,30u,22u,18u,28u,6u,24u,26u,0u, - 0u,2u,4u,6u,8u,10u,12u,14u,16u,18u,20u,22u,24u,26u,28u,30u, - 28u,20u,8u,16u,18u,30u,26u,12u,2u,24u,0u,4u,22u,14u,10u,6u - ); /** * G Mixing function @@ -306,14 +292,14 @@ export class PowGpu extends WorkerInterface { G(&v, &m, 6u, 8u, 18u, 28u, 28u, 30u); // ROUND(1) - G(&v, &m, 0u, 8u, 16u, 24u, 28u, 20u); - G(&v, &m, 2u, 10u, 18u, 26u, 8u, 16u); - G(&v, &m, 4u, 12u, 20u, 28u, 18u, 30u); - G(&v, &m, 6u, 14u, 22u, 30u, 26u, 12u); - G(&v, &m, 0u, 10u, 20u, 30u, 2u, 24u); - G(&v, &m, 2u, 12u, 22u, 24u, 0u, 4u); - G(&v, &m, 4u, 14u, 16u, 26u, 22u, 14u); - G(&v, &m, 6u, 8u, 18u, 28u, 10u, 6u); + G(&v, &m, 0u, 8u, 16u, 24u, 28u, 20u); + G(&v, &m, 2u, 10u, 18u, 26u, 8u, 16u); + G(&v, &m, 4u, 12u, 20u, 28u, 18u, 30u); + G(&v, &m, 6u, 14u, 22u, 30u, 26u, 12u); + G(&v, &m, 0u, 10u, 20u, 30u, 2u, 24u); + G(&v, &m, 2u, 12u, 22u, 24u, 0u, 4u); + G(&v, &m, 4u, 14u, 16u, 26u, 22u, 14u); + G(&v, &m, 6u, 8u, 18u, 28u, 10u, 6u); // ROUND(2) G(&v, &m, 0u, 8u, 16u, 24u, 22u, 16u); @@ -355,89 +341,65 @@ export class PowGpu extends WorkerInterface { G(&v, &m, 4u, 14u, 16u, 26u, 30u, 28u); G(&v, &m, 6u, 8u, 18u, 28u, 2u, 18u); - - - // 96-111 24u,10u,2u,30u,28u,26u,8u,20u,0u,14u,12u,6u,18u,4u,16u,22u - // ROUND(6) - G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[96u], SIGMA82[97u]); - G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[98u], SIGMA82[99u]); - G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[100u], SIGMA82[101u]); - G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[102u], SIGMA82[103u]); - G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[104u], SIGMA82[105u]); - G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[106u], SIGMA82[107u]); - G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[108u], SIGMA82[109u]); - G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[110u], SIGMA82[111u]); - - - - // 112-127 26u,22u,14u,28u,24u,2u,6u,18u,10u,0u,30u,8u,16u,12u,4u,20u + G(&v, &m, 0u, 8u, 16u, 24u, 24u, 10u); + G(&v, &m, 2u, 10u, 18u, 26u, 2u, 30u); + G(&v, &m, 4u, 12u, 20u, 28u, 28u, 26u); + G(&v, &m, 6u, 14u, 22u, 30u, 8u, 20u); + G(&v, &m, 0u, 10u, 20u, 30u, 0u, 14u); + G(&v, &m, 2u, 12u, 22u, 24u, 12u, 6u); + G(&v, &m, 4u, 14u, 16u, 26u, 18u, 4u); + G(&v, &m, 6u, 8u, 18u, 28u, 16u, 22u); // ROUND(7) - G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[112u], SIGMA82[113u]); - G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[114u], SIGMA82[115u]); - G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[116u], SIGMA82[117u]); - G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[118u], SIGMA82[119u]); - G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[120u], SIGMA82[121u]); - G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[122u], SIGMA82[123u]); - G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[124u], SIGMA82[125u]); - G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[126u], SIGMA82[127u]); - - - - // 128-143 12u,30u,28u,18u,22u,6u,0u,16u,24u,4u,26u,14u,2u,8u,20u,10u + G(&v, &m, 0u, 8u, 16u, 24u, 26u, 22u); + G(&v, &m, 2u, 10u, 18u, 26u, 14u, 28u); + G(&v, &m, 4u, 12u, 20u, 28u, 24u, 2u); + G(&v, &m, 6u, 14u, 22u, 30u, 6u, 18u); + G(&v, &m, 0u, 10u, 20u, 30u, 10u, 0u); + G(&v, &m, 2u, 12u, 22u, 24u, 30u, 8u); + G(&v, &m, 4u, 14u, 16u, 26u, 16u, 12u); + G(&v, &m, 6u, 8u, 18u, 28u, 4u, 20u); // ROUND(8) - G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[128u], SIGMA82[129u]); - G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[130u], SIGMA82[131u]); - G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[132u], SIGMA82[133u]); - G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[134u], SIGMA82[135u]); - G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[136u], SIGMA82[137u]); - G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[138u], SIGMA82[139u]); - G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[140u], SIGMA82[141u]); - G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[142u], SIGMA82[143u]); - - - - // 144-159 20u,4u,16u,8u,14u,12u,2u,10u,30u,22u,18u,28u,6u,24u,26u,0u + G(&v, &m, 0u, 8u, 16u, 24u, 12u, 30u); + G(&v, &m, 2u, 10u, 18u, 26u, 28u, 18u); + G(&v, &m, 4u, 12u, 20u, 28u, 22u, 6u); + G(&v, &m, 6u, 14u, 22u, 30u, 0u, 16u); + G(&v, &m, 0u, 10u, 20u, 30u, 24u, 4u); + G(&v, &m, 2u, 12u, 22u, 24u, 26u, 14u); + G(&v, &m, 4u, 14u, 16u, 26u, 2u, 8u); + G(&v, &m, 6u, 8u, 18u, 28u, 20u, 10u); // ROUND(9) - G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[144u], SIGMA82[145u]); - G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[146u], SIGMA82[147u]); - G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[148u], SIGMA82[149u]); - G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[150u], SIGMA82[151u]); - G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[152u], SIGMA82[153u]); - G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[154u], SIGMA82[155u]); - G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[156u], SIGMA82[157u]); - G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[158u], SIGMA82[159u]); - - - - // 160-175 0u,2u,4u,6u,8u,10u,12u,14u,16u,18u,20u,22u,24u,26u,28u,30u + G(&v, &m, 0u, 8u, 16u, 24u, 20u, 4u); + G(&v, &m, 2u, 10u, 18u, 26u, 16u, 8u); + G(&v, &m, 4u, 12u, 20u, 28u, 14u, 12u); + G(&v, &m, 6u, 14u, 22u, 30u, 2u, 10u); + G(&v, &m, 0u, 10u, 20u, 30u, 30u, 22u); + G(&v, &m, 2u, 12u, 22u, 24u, 18u, 28u); + G(&v, &m, 4u, 14u, 16u, 26u, 6u, 24u); + G(&v, &m, 6u, 8u, 18u, 28u, 26u, 0u); // ROUND(10) - G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[160u], SIGMA82[161u]); - G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[162u], SIGMA82[163u]); - G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[164u], SIGMA82[165u]); - G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[166u], SIGMA82[167u]); - G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[168u], SIGMA82[169u]); - G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[170u], SIGMA82[171u]); - G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[172u], SIGMA82[173u]); - G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[174u], SIGMA82[175u]); - - - - // 176-191 28u,20u,8u,16u,18u,30u,26u,12u,2u,24u,0u,4u,22u,14u,10u,6u + G(&v, &m, 0u, 8u, 16u, 24u, 0u, 2u); + G(&v, &m, 2u, 10u, 18u, 26u, 4u, 6u); + G(&v, &m, 4u, 12u, 20u, 28u, 8u, 10u); + G(&v, &m, 6u, 14u, 22u, 30u, 12u, 14u); + G(&v, &m, 0u, 10u, 20u, 30u, 16u, 18u); + G(&v, &m, 2u, 12u, 22u, 24u, 20u, 22u); + G(&v, &m, 4u, 14u, 16u, 26u, 24u, 26u); + G(&v, &m, 6u, 8u, 18u, 28u, 28u, 30u); // ROUND(11) - G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[176u], SIGMA82[177u]); - G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[178u], SIGMA82[179u]); - G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[180u], SIGMA82[181u]); - G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[182u], SIGMA82[183u]); - G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[184u], SIGMA82[185u]); - G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[186u], SIGMA82[187u]); - G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[188u], SIGMA82[189u]); - G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[190u], SIGMA82[191u]); + G(&v, &m, 0u, 8u, 16u, 24u, 28u, 20u); + G(&v, &m, 2u, 10u, 18u, 26u, 8u, 16u); + G(&v, &m, 4u, 12u, 20u, 28u, 18u, 30u); + G(&v, &m, 6u, 14u, 22u, 30u, 26u, 12u); + G(&v, &m, 0u, 10u, 20u, 30u, 2u, 24u); + G(&v, &m, 2u, 12u, 22u, 24u, 0u, 4u); + G(&v, &m, 4u, 14u, 16u, 26u, 22u, 14u); + G(&v, &m, 6u, 8u, 18u, 28u, 10u, 6u); /** * Set nonce if it passes the threshold and no other thread has set it -- 2.34.1