From 464e12baa035b0c352064d65dea8b9d4e8a927f4 Mon Sep 17 00:00:00 2001 From: Chris Duncan Date: Sun, 5 Jan 2025 02:47:09 -0800 Subject: [PATCH] Unroll blake compression rounds. --- src/lib/workers/powgpu.ts | 129 +++++++++++++++++++++++++++++++++++--- 1 file changed, 119 insertions(+), 10 deletions(-) diff --git a/src/lib/workers/powgpu.ts b/src/lib/workers/powgpu.ts index 6ade30d..37acec9 100644 --- a/src/lib/workers/powgpu.ts +++ b/src/lib/workers/powgpu.ts @@ -200,16 +200,125 @@ export class PowGpu extends WorkerInterface { /** * Twelve rounds of mixing as part of BLAKE2b compression step */ - for (var r: u32 = 0u; r < 12u; r = r + 1u) { - G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[r * 16u + 0u], SIGMA82[r * 16u + 1u]); - G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[r * 16u + 2u], SIGMA82[r * 16u + 3u]); - G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[r * 16u + 4u], SIGMA82[r * 16u + 5u]); - G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[r * 16u + 6u], SIGMA82[r * 16u + 7u]); - G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[r * 16u + 8u], SIGMA82[r * 16u + 9u]); - G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[r * 16u + 10u], SIGMA82[r * 16u + 11u]); - G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[r * 16u + 12u], SIGMA82[r * 16u + 13u]); - G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[r * 16u + 14u], SIGMA82[r * 16u + 15u]); - } + // ROUND(0) + G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[0u], SIGMA82[1u]); + G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[2u], SIGMA82[3u]); + G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[4u], SIGMA82[5u]); + G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[6u], SIGMA82[7u]); + G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[8u], SIGMA82[9u]); + G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[10u], SIGMA82[11u]); + G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[12u], SIGMA82[13u]); + G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[14u], SIGMA82[15u]); + + // ROUND(1) + G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[16u], SIGMA82[17u]); + G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[18u], SIGMA82[19u]); + G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[20u], SIGMA82[21u]); + G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[22u], SIGMA82[23u]); + G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[24u], SIGMA82[25u]); + G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[26u], SIGMA82[27u]); + G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[28u], SIGMA82[29u]); + G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[30u], SIGMA82[31u]); + + // ROUND(2) + G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[32u], SIGMA82[33u]); + G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[34u], SIGMA82[35u]); + G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[36u], SIGMA82[37u]); + G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[38u], SIGMA82[39u]); + G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[40u], SIGMA82[41u]); + G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[42u], SIGMA82[43u]); + G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[44u], SIGMA82[45u]); + G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[46u], SIGMA82[47u]); + + // ROUND(3) + G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[48u], SIGMA82[49u]); + G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[50u], SIGMA82[51u]); + G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[52u], SIGMA82[53u]); + G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[54u], SIGMA82[55u]); + G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[56u], SIGMA82[57u]); + G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[58u], SIGMA82[59u]); + G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[60u], SIGMA82[61u]); + G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[62u], SIGMA82[63u]); + + // ROUND(4) + G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[64u], SIGMA82[65u]); + G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[66u], SIGMA82[67u]); + G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[68u], SIGMA82[69u]); + G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[70u], SIGMA82[71u]); + G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[72u], SIGMA82[73u]); + G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[74u], SIGMA82[75u]); + G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[76u], SIGMA82[77u]); + G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[78u], SIGMA82[79u]); + + // ROUND(5) + G(&v, &m, 0u, 8u, 16u, 24u, SIGMA82[80u], SIGMA82[81u]); + G(&v, &m, 2u, 10u, 18u, 26u, SIGMA82[82u], SIGMA82[83u]); + G(&v, &m, 4u, 12u, 20u, 28u, SIGMA82[84u], SIGMA82[85u]); + G(&v, &m, 6u, 14u, 22u, 30u, SIGMA82[86u], SIGMA82[87u]); + G(&v, &m, 0u, 10u, 20u, 30u, SIGMA82[88u], SIGMA82[89u]); + G(&v, &m, 2u, 12u, 22u, 24u, SIGMA82[90u], SIGMA82[91u]); + G(&v, &m, 4u, 14u, 16u, 26u, SIGMA82[92u], SIGMA82[93u]); + G(&v, &m, 6u, 8u, 18u, 28u, SIGMA82[94u], SIGMA82[95u]); + + // 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]); + + // 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]); + + // 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]); + + // 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]); + + // 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]); + + // 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]); /** * Set nonce if it passes the threshold and no other thread has set it -- 2.34.1