From 433d6537231ebd01885e7ff6e927b769292c0fcb Mon Sep 17 00:00:00 2001 From: Christian Buchner Date: Thu, 1 May 2014 08:36:56 +0200 Subject: [PATCH] bump to revision 0.7 --- JHA/cuda_jha_compactionTest.cu | 255 +++++++++++++++++ JHA/cuda_jha_keccak512.cu | 12 +- JHA/jackpotcoin.cu | 167 +++++++----- Makefile.am | 9 +- README.txt | 6 +- ccminer.vcxproj | 5 + ccminer.vcxproj.filters | 15 + configure.ac | 2 +- cpu-miner.c | 2 +- cpuminer-config.h | 4 +- cuda_groestlcoin.cu | 4 +- quark/cuda_jh512.cu | 356 ++++++++++++++++++++++++ quark/cuda_quark_blake512.cu | 482 +++++++++++++++++++++++++++++++++ quark/cuda_quark_groestl512.cu | 397 +++++++++++++++++++++++++++ quark/cuda_skein512.cu | 450 ++++++++++++++++++++++++++++++ 15 files changed, 2078 insertions(+), 88 deletions(-) create mode 100644 JHA/cuda_jha_compactionTest.cu create mode 100644 quark/cuda_jh512.cu create mode 100644 quark/cuda_quark_blake512.cu create mode 100644 quark/cuda_quark_groestl512.cu create mode 100644 quark/cuda_skein512.cu diff --git a/JHA/cuda_jha_compactionTest.cu b/JHA/cuda_jha_compactionTest.cu new file mode 100644 index 0000000..bccbe12 --- /dev/null +++ b/JHA/cuda_jha_compactionTest.cu @@ -0,0 +1,255 @@ +#include +#include "cuda_runtime.h" +#include "device_launch_parameters.h" +#include "sm_30_intrinsics.h" + +#include +#include +#include + +// aus cpu-miner.c +extern int device_map[8]; + +// diese Struktur wird in der Init Funktion angefordert +static cudaDeviceProp props[8]; + +static uint32_t *d_tempBranch1Nonces[8]; +static uint32_t *d_tempBranch2Nonces[8]; +static size_t *d_numValid[8]; +static size_t *h_numValid[8]; + +static uint32_t *d_partSum1[8], *d_partSum2[8]; // 2x partielle summen +static uint32_t *d_validTemp1[8], *d_validTemp2[8]; + +// Zwischenspeicher +static uint32_t *d_tempBranchAllNonces[8]; + +// aus heavy.cu +extern cudaError_t MyStreamSynchronize(cudaStream_t stream, int situation, int thr_id); + + + +// Setup-Funktionen +__host__ void jackpot_compactTest_cpu_init(int thr_id, int threads) +{ + cudaGetDeviceProperties(&props[thr_id], device_map[thr_id]); + + // wir brauchen auch Speicherplatz auf dem Device + cudaMalloc(&d_tempBranchAllNonces[thr_id], sizeof(uint32_t) * threads); + cudaMalloc(&d_tempBranch1Nonces[thr_id], sizeof(uint32_t) * threads); + cudaMalloc(&d_tempBranch2Nonces[thr_id], sizeof(uint32_t) * threads); + cudaMalloc(&d_numValid[thr_id], 2*sizeof(size_t)); + cudaMallocHost(&h_numValid[thr_id], 2*sizeof(size_t)); + + uint32_t s1; + s1 = threads / 256; + + cudaMalloc(&d_partSum1[thr_id], sizeof(uint32_t) * s1); // BLOCKSIZE (Threads/Block) + cudaMalloc(&d_partSum2[thr_id], sizeof(uint32_t) * s1); // BLOCKSIZE (Threads/Block) + + cudaMalloc(&d_validTemp1[thr_id], sizeof(uint32_t) * threads); // BLOCKSIZE (Threads/Block) + cudaMalloc(&d_validTemp2[thr_id], sizeof(uint32_t) * threads); // BLOCKSIZE (Threads/Block) +} + +// Die Testfunktion (zum Erstellen der TestMap) +__global__ void jackpot_compactTest_gpu_TEST_64(int threads, uint32_t startNounce, uint32_t *inpHashes, uint32_t *d_noncesFull, + uint32_t *d_nonces1, uint32_t *d_nonces2, + uint32_t *d_validT1, uint32_t *d_validT2) +{ + int thread = (blockDim.x * blockIdx.x + threadIdx.x); + if (thread < threads) + { + // bestimme den aktuellen Zähler + uint32_t nounce = startNounce + thread; + uint32_t *inpHash = &inpHashes[16 * thread]; + + uint32_t tmp = inpHash[0] & 0x01; + uint32_t val1 = (tmp == 1); + uint32_t val2 = (tmp == 0); + + d_nonces1[thread] = val1; + d_validT1[thread] = val1; + d_nonces2[thread] = val2; + d_validT2[thread] = val2; + d_noncesFull[thread] = nounce; + } +} + +// Die Summenfunktion (vom NVIDIA SDK) +__global__ void jackpot_compactTest_gpu_SCAN(uint32_t *data, int width, uint32_t *partial_sums=NULL) +{ + extern __shared__ uint32_t sums[]; + int id = ((blockIdx.x * blockDim.x) + threadIdx.x); + //int lane_id = id % warpSize; + int lane_id = id % width; + // determine a warp_id within a block + //int warp_id = threadIdx.x / warpSize; + int warp_id = threadIdx.x / width; + + // Below is the basic structure of using a shfl instruction + // for a scan. + // Record "value" as a variable - we accumulate it along the way + uint32_t value = data[id]; + + // Now accumulate in log steps up the chain + // compute sums, with another thread's value who is + // distance delta away (i). Note + // those threads where the thread 'i' away would have + // been out of bounds of the warp are unaffected. This + // creates the scan sum. +#pragma unroll + + for (int i=1; i<=width; i*=2) + { + uint32_t n = __shfl_up((int)value, i, width); + + if (lane_id >= i) value += n; + } + + // value now holds the scan value for the individual thread + // next sum the largest values for each warp + + // write the sum of the warp to smem + //if (threadIdx.x % warpSize == warpSize-1) + if (threadIdx.x % width == width-1) + { + sums[warp_id] = value; + } + + __syncthreads(); + + // + // scan sum the warp sums + // the same shfl scan operation, but performed on warp sums + // + if (warp_id == 0) + { + uint32_t warp_sum = sums[lane_id]; + + for (int i=1; i<=width; i*=2) + { + uint32_t n = __shfl_up((int)warp_sum, i, width); + + if (lane_id >= i) warp_sum += n; + } + + sums[lane_id] = warp_sum; + } + + __syncthreads(); + + // perform a uniform add across warps in the block + // read neighbouring warp's sum and add it to threads value + uint32_t blockSum = 0; + + if (warp_id > 0) + { + blockSum = sums[warp_id-1]; + } + + value += blockSum; + + // Now write out our result + data[id] = value; + + // last thread has sum, write write out the block's sum + if (partial_sums != NULL && threadIdx.x == blockDim.x-1) + { + partial_sums[blockIdx.x] = value; + } +} + +// Uniform add: add partial sums array +__global__ void jackpot_compactTest_gpu_ADD(uint32_t *data, uint32_t *partial_sums, int len) +{ + __shared__ uint32_t buf; + int id = ((blockIdx.x * blockDim.x) + threadIdx.x); + + if (id > len) return; + + if (threadIdx.x == 0) + { + buf = partial_sums[blockIdx.x]; + } + + __syncthreads(); + data[id] += buf; +} + +// Der Scatter +__global__ void jackpot_compactTest_gpu_SCATTER(uint32_t *data, uint32_t *valid, uint32_t *sum, uint32_t *outp) +{ + int id = ((blockIdx.x * blockDim.x) + threadIdx.x); + if( valid[id] ) + { + int idx = sum[id]; + if(idx > 0) + outp[idx-1] = data[id]; + } +} + +////// ACHTUNG: Diese funktion geht aktuell nur mit threads > 65536 (Am besten 256 * 1024 oder 256*2048) +__host__ void jackpot_compactTest_cpu_dualCompaction(int thr_id, int threads, size_t *nrm, + uint32_t *d_nonces1, uint32_t *d_nonces2) +{ + // threadsPerBlock ausrechnen + int blockSize = 256; + int thr1 = threads / blockSize; + int thr2 = threads / (blockSize*blockSize); + + // 1 + jackpot_compactTest_gpu_SCAN<<>>(d_tempBranch1Nonces[thr_id], 32, d_partSum1[thr_id]); + jackpot_compactTest_gpu_SCAN<<>>(d_partSum1[thr_id], 32, d_partSum2[thr_id]); + jackpot_compactTest_gpu_SCAN<<<1, thr2, 8*sizeof(uint32_t)>>>(d_partSum2[thr_id], (thr2>32) ? 32 : thr2); + cudaStreamSynchronize(NULL); + cudaMemcpy(&nrm[0], &(d_partSum2[thr_id])[thr2-1], sizeof(uint32_t), cudaMemcpyDeviceToHost); + jackpot_compactTest_gpu_ADD<<>>(d_partSum1[thr_id]+blockSize, d_partSum2[thr_id], blockSize*thr2); + jackpot_compactTest_gpu_ADD<<>>(d_tempBranch1Nonces[thr_id]+blockSize, d_partSum1[thr_id], threads); + + // 2 + jackpot_compactTest_gpu_SCAN<<>>(d_tempBranch2Nonces[thr_id], 32, d_partSum1[thr_id]); + jackpot_compactTest_gpu_SCAN<<>>(d_partSum1[thr_id], 32, d_partSum2[thr_id]); + jackpot_compactTest_gpu_SCAN<<<1, thr2, 8*sizeof(uint32_t)>>>(d_partSum2[thr_id], (thr2>32) ? 32 : thr2); + cudaStreamSynchronize(NULL); + cudaMemcpy(&nrm[1], &(d_partSum2[thr_id])[thr2-1], sizeof(uint32_t), cudaMemcpyDeviceToHost); + jackpot_compactTest_gpu_ADD<<>>(d_partSum1[thr_id]+blockSize, d_partSum2[thr_id], blockSize*thr2); + jackpot_compactTest_gpu_ADD<<>>(d_tempBranch2Nonces[thr_id]+blockSize, d_partSum1[thr_id], threads); + + // Hier ist noch eine Besonderheit: in d_tempBranch1Nonces sind die element von 1...nrm1 die Interessanten + // Schritt 3: Scatter + jackpot_compactTest_gpu_SCATTER<<>>(d_tempBranchAllNonces[thr_id], d_validTemp1[thr_id], d_tempBranch1Nonces[thr_id], d_nonces1); + jackpot_compactTest_gpu_SCATTER<<>>(d_tempBranchAllNonces[thr_id], d_validTemp2[thr_id], d_tempBranch2Nonces[thr_id], d_nonces2); + cudaStreamSynchronize(NULL); +} + +__host__ void jackpot_compactTest_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *inpHashes, + uint32_t *d_nonces1, size_t *nrm1, + uint32_t *d_nonces2, size_t *nrm2, + int order) +{ + // Compute 3.x und 5.x Geräte am besten mit 768 Threads ansteuern, + // alle anderen mit 512 Threads. + //int threadsperblock = (props[thr_id].major >= 3) ? 768 : 512; + int threadsperblock = 256; + + // berechne wie viele Thread Blocks wir brauchen + dim3 grid((threads + threadsperblock-1)/threadsperblock); + dim3 block(threadsperblock); + + size_t shared_size = 0; + +// fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size); + + // Schritt 1: Prüfen der Bedingung und Speicherung in d_tempBranch1/2Nonces + jackpot_compactTest_gpu_TEST_64<<>>(threads, startNounce, inpHashes, d_tempBranchAllNonces[thr_id], + d_tempBranch1Nonces[thr_id], d_tempBranch2Nonces[thr_id], + d_validTemp1[thr_id], d_validTemp2[thr_id]); + + // Strategisches Sleep Kommando zur Senkung der CPU Last + jackpot_compactTest_cpu_dualCompaction(thr_id, threads, + h_numValid[thr_id], d_nonces1, d_nonces2); + + cudaStreamSynchronize(NULL); // Das original braucht zwar etwas CPU-Last, ist an dieser Stelle aber evtl besser + *nrm1 = h_numValid[thr_id][0]; + *nrm2 = h_numValid[thr_id][1]; +} diff --git a/JHA/cuda_jha_keccak512.cu b/JHA/cuda_jha_keccak512.cu index 62d143b..2906d47 100644 --- a/JHA/cuda_jha_keccak512.cu +++ b/JHA/cuda_jha_keccak512.cu @@ -1,3 +1,5 @@ + + #include #include "cuda_runtime.h" #include "device_launch_parameters.h" @@ -132,7 +134,7 @@ keccak_block(uint64_t *s, const uint32_t *in, const uint64_t *keccak_round_const } } -__global__ void jackpot_keccak512_gpu_hash_88(int threads, uint32_t startNounce, uint64_t *g_hash) +__global__ void jackpot_keccak512_gpu_hash(int threads, uint32_t startNounce, uint64_t *g_hash) { int thread = (blockDim.x * blockIdx.x + threadIdx.x); if (thread < threads) @@ -518,9 +520,9 @@ void KeccakF( tKeccakLane * state, const tKeccakLane *in, int laneCount ) } } -__host__ void jackpot_keccak512_cpu_setBlock_88(void *pdata) +// inlen kann 72...143 betragen +__host__ void jackpot_keccak512_cpu_setBlock(void *pdata, size_t inlen) { - unsigned long long inlen = 88; const unsigned char *in = (const unsigned char*)pdata; tKeccakLane state[5 * 5]; @@ -554,7 +556,7 @@ __host__ void jackpot_keccak512_cpu_setBlock_88(void *pdata) 0, cudaMemcpyHostToDevice); } -__host__ void jackpot_keccak512_cpu_hash_88(int thr_id, int threads, uint32_t startNounce, uint32_t *d_hash, int order) +__host__ void jackpot_keccak512_cpu_hash(int thr_id, int threads, uint32_t startNounce, uint32_t *d_hash, int order) { const int threadsperblock = 256; @@ -567,6 +569,6 @@ __host__ void jackpot_keccak512_cpu_hash_88(int thr_id, int threads, uint32_t st // fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size); - jackpot_keccak512_gpu_hash_88<<>>(threads, startNounce, (uint64_t*)d_hash); + jackpot_keccak512_gpu_hash<<>>(threads, startNounce, (uint64_t*)d_hash); MyStreamSynchronize(NULL, order, thr_id); } diff --git a/JHA/jackpotcoin.cu b/JHA/jackpotcoin.cu index a4242a8..3249ccd 100644 --- a/JHA/jackpotcoin.cu +++ b/JHA/jackpotcoin.cu @@ -6,9 +6,9 @@ extern "C" #include "sph/sph_groestl.h" #include "sph/sph_jh.h" #include "sph/sph_skein.h" +#include "miner.h" } -#include "miner.h" #include // aus cpu-miner.c @@ -19,13 +19,36 @@ extern bool opt_benchmark; static uint32_t *d_hash[8]; extern void jackpot_keccak512_cpu_init(int thr_id, int threads); -extern void jackpot_keccak512_cpu_setBlock_88(void *pdata); -extern void jackpot_keccak512_cpu_hash_88(int thr_id, int threads, uint32_t startNounce, uint32_t *d_hash, int order); +extern void jackpot_keccak512_cpu_setBlock(void *pdata, size_t inlen); +extern void jackpot_keccak512_cpu_hash(int thr_id, int threads, uint32_t startNounce, uint32_t *d_hash, int order); + +extern void quark_blake512_cpu_init(int thr_id, int threads); +extern void quark_blake512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order); + +extern void quark_groestl512_cpu_init(int thr_id, int threads); +extern void quark_groestl512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order); + +extern void quark_jh512_cpu_init(int thr_id, int threads); +extern void quark_jh512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order); + +extern void quark_skein512_cpu_init(int thr_id, int threads); +extern void quark_skein512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order); extern void quark_check_cpu_init(int thr_id, int threads); extern void quark_check_cpu_setTarget(const void *ptarget); extern uint32_t quark_check_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_inputHash, int order); +extern void jackpot_compactTest_cpu_init(int thr_id, int threads); +extern void jackpot_compactTest_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *inpHashes, + uint32_t *d_nonces1, size_t *nrm1, + uint32_t *d_nonces2, size_t *nrm2, + int order); + +// Speicher zur Generierung der Noncevektoren für die bedingten Hashes +static uint32_t *d_jackpotNonces[8]; +static uint32_t *d_branch1Nonces[8]; +static uint32_t *d_branch2Nonces[8]; + // Original jackpothash Funktion aus einem miner Quelltext inline unsigned int jackpothash(void *state, const void *input) { @@ -38,55 +61,38 @@ inline unsigned int jackpothash(void *state, const void *input) uint32_t hash[16]; sph_keccak512_init(&ctx_keccak); - sph_keccak512 (&ctx_keccak, input, 88); + sph_keccak512 (&ctx_keccak, input, 80); sph_keccak512_close(&ctx_keccak, hash); - unsigned int round_mask = ( - (unsigned int)(((unsigned char *)input)[84]) << 0 | - (unsigned int)(((unsigned char *)input)[85]) << 8 | - (unsigned int)(((unsigned char *)input)[86]) << 16 | - (unsigned int)(((unsigned char *)input)[87]) << 24 ); - unsigned int round_max = hash[0] & round_mask; unsigned int round; - for (round = 0; round < round_max; round++) { - switch (hash[0] & 3) { - case 0: - sph_blake512_init(&ctx_blake); - sph_blake512 (&ctx_blake, hash, 64); - sph_blake512_close(&ctx_blake, hash); - break; - case 1: - sph_groestl512_init(&ctx_groestl); - sph_groestl512 (&ctx_groestl, hash, 64); - sph_groestl512_close(&ctx_groestl, hash); - break; - case 2: - sph_jh512_init(&ctx_jh); - sph_jh512 (&ctx_jh, hash, 64); - sph_jh512_close(&ctx_jh, hash); - break; - case 3: - sph_skein512_init(&ctx_skein); - sph_skein512 (&ctx_skein, hash, 64); - sph_skein512_close(&ctx_skein, hash); - break; + for (round = 0; round < 3; round++) { + if (hash[0] & 0x01) { + sph_groestl512_init(&ctx_groestl); + sph_groestl512 (&ctx_groestl, (&hash), 64); + sph_groestl512_close(&ctx_groestl, (&hash)); + } + else { + sph_skein512_init(&ctx_skein); + sph_skein512 (&ctx_skein, (&hash), 64); + sph_skein512_close(&ctx_skein, (&hash)); + } + if (hash[0] & 0x01) { + sph_blake512_init(&ctx_blake); + sph_blake512 (&ctx_blake, (&hash), 64); + sph_blake512_close(&ctx_blake, (&hash)); + } + else { + sph_jh512_init(&ctx_jh); + sph_jh512 (&ctx_jh, (&hash), 64); + sph_jh512_close(&ctx_jh, (&hash)); } } memcpy(state, hash, 32); - return round_max; + return round; } -static int bit_population(uint32_t n){ - int c =0; - while(n){ - c += n&1; - n = n>>1; - } - return c; -} - extern "C" int scanhash_jackpot(int thr_id, uint32_t *pdata, const uint32_t *ptarget, uint32_t max_nonce, unsigned long *hashes_done) @@ -94,10 +100,8 @@ extern "C" int scanhash_jackpot(int thr_id, uint32_t *pdata, const uint32_t first_nonce = pdata[19]; // TODO: entfernen für eine Release! Ist nur zum Testen! - if (opt_benchmark) { - ((uint32_t*)ptarget)[7] = 0x00000f; - ((uint32_t*)pdata)[21] = 0x07000000; // round_mask von 7 vorgeben - } + if (opt_benchmark) + ((uint32_t*)ptarget)[7] = 0x0000ff; const uint32_t Htarg = ptarget[7]; @@ -111,7 +115,15 @@ extern "C" int scanhash_jackpot(int thr_id, uint32_t *pdata, // Konstanten kopieren, Speicher belegen cudaMalloc(&d_hash[thr_id], 16 * sizeof(uint32_t) * throughput); jackpot_keccak512_cpu_init(thr_id, throughput); + jackpot_compactTest_cpu_init(thr_id, throughput); + quark_blake512_cpu_init(thr_id, throughput); + quark_groestl512_cpu_init(thr_id, throughput); + quark_jh512_cpu_init(thr_id, throughput); + quark_skein512_cpu_init(thr_id, throughput); quark_check_cpu_init(thr_id, throughput); + cudaMalloc(&d_jackpotNonces[thr_id], sizeof(uint32_t)*throughput); + cudaMalloc(&d_branch1Nonces[thr_id], sizeof(uint32_t)*throughput); + cudaMalloc(&d_branch2Nonces[thr_id], sizeof(uint32_t)*throughput); init[thr_id] = true; } @@ -119,27 +131,41 @@ extern "C" int scanhash_jackpot(int thr_id, uint32_t *pdata, for (int k=0; k < 22; k++) be32enc(&endiandata[k], ((uint32_t*)pdata)[k]); - unsigned int round_mask = ( - (unsigned int)(((unsigned char *)endiandata)[84]) << 0 | - (unsigned int)(((unsigned char *)endiandata)[85]) << 8 | - (unsigned int)(((unsigned char *)endiandata)[86]) << 16 | - (unsigned int)(((unsigned char *)endiandata)[87]) << 24 ); - - // Zählen wie viele Bits in round_mask gesetzt sind - int bitcount = bit_population(round_mask); - - jackpot_keccak512_cpu_setBlock_88((void*)endiandata); + jackpot_keccak512_cpu_setBlock((void*)endiandata, 80); quark_check_cpu_setTarget(ptarget); do { int order = 0; - // erstes Blake512 Hash mit CUDA - jackpot_keccak512_cpu_hash_88(thr_id, throughput, pdata[19], d_hash[thr_id], order++); + // erstes Keccak512 Hash mit CUDA + jackpot_keccak512_cpu_hash(thr_id, throughput, pdata[19], d_hash[thr_id], order++); + + for (int round=0; round < 3; round++) + { + size_t nrm1, nrm2; + + // jackpotNonces in branch1/2 aufsplitten gemäss if (hash[0] & 0x01) + jackpot_compactTest_cpu_hash_64(thr_id, throughput, pdata[19], d_hash[thr_id], + d_branch1Nonces[thr_id], &nrm1, + d_branch2Nonces[thr_id], &nrm2, + order++); + + if (nrm1+nrm2 == throughput) { + quark_groestl512_cpu_hash_64(thr_id, nrm1, pdata[19], d_branch1Nonces[thr_id], d_hash[thr_id], order++); + quark_skein512_cpu_hash_64(thr_id, nrm2, pdata[19], d_branch2Nonces[thr_id], d_hash[thr_id], order++); + } + + // jackpotNonces in branch1/2 aufsplitten gemäss if (hash[0] & 0x01) + jackpot_compactTest_cpu_hash_64(thr_id, throughput, pdata[19], d_hash[thr_id], + d_branch1Nonces[thr_id], &nrm1, + d_branch2Nonces[thr_id], &nrm2, + order++); - // TODO: hier fehlen jetzt natürlich noch die anderen Hashrunden. - // bei round_mask=7 haben wir eine 1:8 Chance, dass das Hash dennoch - // die Kriterien erfüllt wenn hash[0] & round_mask zufällig 0 ist. + if (nrm1+nrm2 == throughput) { + quark_blake512_cpu_hash_64(thr_id, nrm1, pdata[19], d_branch1Nonces[thr_id], d_hash[thr_id], order++); + quark_jh512_cpu_hash_64(thr_id, nrm2, pdata[19], d_branch2Nonces[thr_id], d_hash[thr_id], order++); + } + } // Scan nach Gewinner Hashes auf der GPU uint32_t foundNonce = quark_check_cpu_hash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++); @@ -148,19 +174,16 @@ extern "C" int scanhash_jackpot(int thr_id, uint32_t *pdata, uint32_t vhash64[8]; be32enc(&endiandata[19], foundNonce); - // diese jackpothash Funktion gibt die Zahl der zusätzlichen Runden zurück + // diese jackpothash Funktion gibt die Zahl der Runden zurück unsigned int rounds = jackpothash(vhash64, endiandata); - // wir akzeptieren nur solche Hashes wo ausschliesslich Keccak verwendet wurde - if (rounds == 0) { - if ((vhash64[7]<=Htarg) && fulltest(vhash64, ptarget)) { + if ((vhash64[7]<=Htarg) && fulltest(vhash64, ptarget)) { - pdata[19] = foundNonce; - *hashes_done = (foundNonce - first_nonce + 1) / (1 << bitcount); - return 1; - } else { - applog(LOG_INFO, "GPU #%d: result for nonce $%08X does not validate on CPU (%d rounds)!", thr_id, foundNonce, rounds); - } + pdata[19] = foundNonce; + *hashes_done = (foundNonce - first_nonce + 1); + return 1; + } else { + applog(LOG_INFO, "GPU #%d: result for nonce $%08X does not validate on CPU (%d rounds)!", thr_id, foundNonce, rounds); } } @@ -168,6 +191,6 @@ extern "C" int scanhash_jackpot(int thr_id, uint32_t *pdata, } while (pdata[19] < max_nonce && !work_restart[thr_id].restart); - *hashes_done = (pdata[19] - first_nonce + 1) / (1 << bitcount); + *hashes_done = (pdata[19] - first_nonce + 1); return 0; } diff --git a/Makefile.am b/Makefile.am index 1a75f16..edb7f8f 100644 --- a/Makefile.am +++ b/Makefile.am @@ -16,8 +16,8 @@ bin_PROGRAMS = ccminer ccminer_SOURCES = elist.h miner.h compat.h \ compat/inttypes.h compat/stdbool.h compat/unistd.h \ compat/sys/time.h compat/getopt/getopt.h \ - cpu-miner.c util.c sph/blake.c sph/groestl.c sph/keccak.c hefty1.c scrypt.c sha2.c \ - sph/sph_blake.h sph/sph_groestl.h sph/sph_keccak.h sph/sph_types.h \ + cpu-miner.c util.c sph/blake.c sph/groestl.c sph/jh.c sph/keccak.c sph/skein.c hefty1.c scrypt.c sha2.c \ + sph/sph_blake.h sph/sph_groestl.h sph/sph_jh.h sph/sph_keccak.h sph/sph_skein.h sph/sph_types.h \ heavy.cu \ cuda_blake512.cu cuda_blake512.h \ cuda_combine.cu cuda_combine.h \ @@ -27,8 +27,9 @@ ccminer_SOURCES = elist.h miner.h compat.h \ cuda_sha256.cu cuda_sha256.h \ fuguecoin.cpp cuda_fugue256.cu sph/fugue.c sph/sph_fugue.h uint256.h \ groestlcoin.cpp cuda_groestlcoin.cu cuda_groestlcoin.h \ - JHA/jackpotcoin.cu JHA/cuda_jha_keccak512.cu sph/jh.c sph/skein.c \ - sph/sph_jh.h sph/sph_skein.h quark/cuda_quark_checkhash.cu \ + JHA/jackpotcoin.cu JHA/cuda_jha_keccak512.cu \ + JHA/cuda_jha_compactionTest.cu quark/cuda_quark_checkhash.cu \ + quark/cuda_jh512.cu quark/cuda_quark_blake512.cu quark/cuda_quark_groestl512.cu quark/cuda_skein512.cu \ myriadgroestl.cpp cuda_myriadgroestl.cu ccminer_LDFLAGS = $(PTHREAD_FLAGS) @CUDA_LDFLAGS@ diff --git a/README.txt b/README.txt index 9f81260..bc574e4 100644 --- a/README.txt +++ b/README.txt @@ -1,5 +1,5 @@ -ccMiner release 0.6 (April 27th 2014) - "Jackpot" +ccMiner release 0.7 (May 01th 2014) - "Jackpot II" ------------------------------------------------------------- *************************************************************** @@ -116,6 +116,10 @@ from your old clunkers. >>> RELEASE HISTORY <<< + May 1st 2014 adapt the Jackpot algorithms to changes made by the + coin developers. We keep our unique nVidia advantage + because we have a way to break up the divergence. + April, 27 2014 this release adds Myriad-Groestl and Jackpot Coin. we apply an optimization to Jackpot that turns this into a Keccak-only CUDA coin ;) Jackpot is tested with diff --git a/ccminer.vcxproj b/ccminer.vcxproj index c79e5ef..30bda9c 100644 --- a/ccminer.vcxproj +++ b/ccminer.vcxproj @@ -305,9 +305,14 @@ copy "$(CudaToolkitBinDir)\cudart*.dll" "$(OutDir)" + + + + + diff --git a/ccminer.vcxproj.filters b/ccminer.vcxproj.filters index d0e23e1..af1ee62 100644 --- a/ccminer.vcxproj.filters +++ b/ccminer.vcxproj.filters @@ -286,5 +286,20 @@ Source Files\CUDA + + Source Files\CUDA\JHA + + + Source Files\CUDA\quark + + + Source Files\CUDA\quark + + + Source Files\CUDA\quark + + + Source Files\CUDA\quark + \ No newline at end of file diff --git a/configure.ac b/configure.ac index 216a561..3b4e438 100644 --- a/configure.ac +++ b/configure.ac @@ -1,4 +1,4 @@ -AC_INIT([ccminer], [2014.04.27]) +AC_INIT([ccminer], [2014.05.01]) AC_PREREQ([2.59c]) AC_CANONICAL_SYSTEM diff --git a/cpu-miner.c b/cpu-miner.c index 583efd2..784aa1b 100644 --- a/cpu-miner.c +++ b/cpu-miner.c @@ -1404,7 +1404,7 @@ static void signal_handler(int sig) } #endif -#define PROGRAM_VERSION "0.6" +#define PROGRAM_VERSION "0.7" int main(int argc, char *argv[]) { struct thr_info *thr; diff --git a/cpuminer-config.h b/cpuminer-config.h index 0057c14..61cbc3a 100644 --- a/cpuminer-config.h +++ b/cpuminer-config.h @@ -152,7 +152,7 @@ #define PACKAGE_NAME "ccminer" /* Define to the full name and version of this package. */ -#define PACKAGE_STRING "ccminer 2014.04.27" +#define PACKAGE_STRING "ccminer 2014.05.01" /* Define to the one symbol short name of this package. */ #undef PACKAGE_TARNAME @@ -161,7 +161,7 @@ #undef PACKAGE_URL /* Define to the version of this package. */ -#define PACKAGE_VERSION "2014.04.27" +#define PACKAGE_VERSION "2014.05.01" /* If using the C implementation of alloca, define if you know the direction of stack growth for your system; otherwise it will be diff --git a/cuda_groestlcoin.cu b/cuda_groestlcoin.cu index 12de56d..258f11b 100644 --- a/cuda_groestlcoin.cu +++ b/cuda_groestlcoin.cu @@ -8,7 +8,7 @@ #include // it's unfortunate that this is a compile time constant. -#define MAXWELL_OR_FERMI 0 +#define MAXWELL_OR_FERMI 1 // aus cpu-miner.c extern int device_map[8]; @@ -50,7 +50,7 @@ __constant__ uint32_t groestlcoin_gpu_msg[32]; #define B32_3(x) __byte_perm(x, 0, 0x4443) //((x) >> 24) -#if MAXWELL_OR_FEMRI +#if MAXWELL_OR_FERMI #define USE_SHARED 1 // Maxwell and Fermi cards get the best speed with SHARED access it seems. #if USE_SHARED diff --git a/quark/cuda_jh512.cu b/quark/cuda_jh512.cu new file mode 100644 index 0000000..75508cf --- /dev/null +++ b/quark/cuda_jh512.cu @@ -0,0 +1,356 @@ +#include + +// aus heavy.cu +extern cudaError_t MyStreamSynchronize(cudaStream_t stream, int situation, int thr_id); + +typedef struct { + uint32_t x[8][4]; /*the 1024-bit state, ( x[i][0] || x[i][1] || x[i][2] || x[i][3] ) is the ith row of the state in the pseudocode*/ + uint32_t buffer[16]; /*the 512-bit message block to be hashed;*/ +} hashState; + +/*42 round constants, each round constant is 32-byte (256-bit)*/ +__constant__ uint32_t c_INIT_bitslice[8][4]; +__constant__ unsigned char c_E8_bitslice_roundconstant[42][32]; + +const uint32_t h_INIT_bitslice[8][4] = { + { 0x964bd16f, 0x17aa003e, 0x052e6a63, 0x43d5157a}, + { 0x8d5e228a, 0x0bef970c, 0x591234e9, 0x61c3b3f2}, + { 0xc1a01d89, 0x1e806f53, 0x6b05a92a, 0x806d2bea}, + { 0xdbcc8e58, 0xa6ba7520, 0x763a0fa9, 0xf73bf8ba}, + { 0x05e66901, 0x694ae341, 0x8e8ab546, 0x5ae66f2e}, + { 0xd0a74710, 0x243c84c1, 0xb1716e3b, 0x99c15a2d}, + { 0xecf657cf, 0x56f8b19d, 0x7c8806a7, 0x56b11657}, + { 0xdffcc2e3, 0xfb1785e6, 0x78465a54, 0x4bdd8ccc} }; + +const unsigned char h_E8_bitslice_roundconstant[42][32]={ +{0x72,0xd5,0xde,0xa2,0xdf,0x15,0xf8,0x67,0x7b,0x84,0x15,0xa,0xb7,0x23,0x15,0x57,0x81,0xab,0xd6,0x90,0x4d,0x5a,0x87,0xf6,0x4e,0x9f,0x4f,0xc5,0xc3,0xd1,0x2b,0x40}, +{0xea,0x98,0x3a,0xe0,0x5c,0x45,0xfa,0x9c,0x3,0xc5,0xd2,0x99,0x66,0xb2,0x99,0x9a,0x66,0x2,0x96,0xb4,0xf2,0xbb,0x53,0x8a,0xb5,0x56,0x14,0x1a,0x88,0xdb,0xa2,0x31}, +{0x3,0xa3,0x5a,0x5c,0x9a,0x19,0xe,0xdb,0x40,0x3f,0xb2,0xa,0x87,0xc1,0x44,0x10,0x1c,0x5,0x19,0x80,0x84,0x9e,0x95,0x1d,0x6f,0x33,0xeb,0xad,0x5e,0xe7,0xcd,0xdc}, +{0x10,0xba,0x13,0x92,0x2,0xbf,0x6b,0x41,0xdc,0x78,0x65,0x15,0xf7,0xbb,0x27,0xd0,0xa,0x2c,0x81,0x39,0x37,0xaa,0x78,0x50,0x3f,0x1a,0xbf,0xd2,0x41,0x0,0x91,0xd3}, +{0x42,0x2d,0x5a,0xd,0xf6,0xcc,0x7e,0x90,0xdd,0x62,0x9f,0x9c,0x92,0xc0,0x97,0xce,0x18,0x5c,0xa7,0xb,0xc7,0x2b,0x44,0xac,0xd1,0xdf,0x65,0xd6,0x63,0xc6,0xfc,0x23}, +{0x97,0x6e,0x6c,0x3,0x9e,0xe0,0xb8,0x1a,0x21,0x5,0x45,0x7e,0x44,0x6c,0xec,0xa8,0xee,0xf1,0x3,0xbb,0x5d,0x8e,0x61,0xfa,0xfd,0x96,0x97,0xb2,0x94,0x83,0x81,0x97}, +{0x4a,0x8e,0x85,0x37,0xdb,0x3,0x30,0x2f,0x2a,0x67,0x8d,0x2d,0xfb,0x9f,0x6a,0x95,0x8a,0xfe,0x73,0x81,0xf8,0xb8,0x69,0x6c,0x8a,0xc7,0x72,0x46,0xc0,0x7f,0x42,0x14}, +{0xc5,0xf4,0x15,0x8f,0xbd,0xc7,0x5e,0xc4,0x75,0x44,0x6f,0xa7,0x8f,0x11,0xbb,0x80,0x52,0xde,0x75,0xb7,0xae,0xe4,0x88,0xbc,0x82,0xb8,0x0,0x1e,0x98,0xa6,0xa3,0xf4}, +{0x8e,0xf4,0x8f,0x33,0xa9,0xa3,0x63,0x15,0xaa,0x5f,0x56,0x24,0xd5,0xb7,0xf9,0x89,0xb6,0xf1,0xed,0x20,0x7c,0x5a,0xe0,0xfd,0x36,0xca,0xe9,0x5a,0x6,0x42,0x2c,0x36}, +{0xce,0x29,0x35,0x43,0x4e,0xfe,0x98,0x3d,0x53,0x3a,0xf9,0x74,0x73,0x9a,0x4b,0xa7,0xd0,0xf5,0x1f,0x59,0x6f,0x4e,0x81,0x86,0xe,0x9d,0xad,0x81,0xaf,0xd8,0x5a,0x9f}, +{0xa7,0x5,0x6,0x67,0xee,0x34,0x62,0x6a,0x8b,0xb,0x28,0xbe,0x6e,0xb9,0x17,0x27,0x47,0x74,0x7,0x26,0xc6,0x80,0x10,0x3f,0xe0,0xa0,0x7e,0x6f,0xc6,0x7e,0x48,0x7b}, +{0xd,0x55,0xa,0xa5,0x4a,0xf8,0xa4,0xc0,0x91,0xe3,0xe7,0x9f,0x97,0x8e,0xf1,0x9e,0x86,0x76,0x72,0x81,0x50,0x60,0x8d,0xd4,0x7e,0x9e,0x5a,0x41,0xf3,0xe5,0xb0,0x62}, +{0xfc,0x9f,0x1f,0xec,0x40,0x54,0x20,0x7a,0xe3,0xe4,0x1a,0x0,0xce,0xf4,0xc9,0x84,0x4f,0xd7,0x94,0xf5,0x9d,0xfa,0x95,0xd8,0x55,0x2e,0x7e,0x11,0x24,0xc3,0x54,0xa5}, +{0x5b,0xdf,0x72,0x28,0xbd,0xfe,0x6e,0x28,0x78,0xf5,0x7f,0xe2,0xf,0xa5,0xc4,0xb2,0x5,0x89,0x7c,0xef,0xee,0x49,0xd3,0x2e,0x44,0x7e,0x93,0x85,0xeb,0x28,0x59,0x7f}, +{0x70,0x5f,0x69,0x37,0xb3,0x24,0x31,0x4a,0x5e,0x86,0x28,0xf1,0x1d,0xd6,0xe4,0x65,0xc7,0x1b,0x77,0x4,0x51,0xb9,0x20,0xe7,0x74,0xfe,0x43,0xe8,0x23,0xd4,0x87,0x8a}, +{0x7d,0x29,0xe8,0xa3,0x92,0x76,0x94,0xf2,0xdd,0xcb,0x7a,0x9,0x9b,0x30,0xd9,0xc1,0x1d,0x1b,0x30,0xfb,0x5b,0xdc,0x1b,0xe0,0xda,0x24,0x49,0x4f,0xf2,0x9c,0x82,0xbf}, +{0xa4,0xe7,0xba,0x31,0xb4,0x70,0xbf,0xff,0xd,0x32,0x44,0x5,0xde,0xf8,0xbc,0x48,0x3b,0xae,0xfc,0x32,0x53,0xbb,0xd3,0x39,0x45,0x9f,0xc3,0xc1,0xe0,0x29,0x8b,0xa0}, +{0xe5,0xc9,0x5,0xfd,0xf7,0xae,0x9,0xf,0x94,0x70,0x34,0x12,0x42,0x90,0xf1,0x34,0xa2,0x71,0xb7,0x1,0xe3,0x44,0xed,0x95,0xe9,0x3b,0x8e,0x36,0x4f,0x2f,0x98,0x4a}, +{0x88,0x40,0x1d,0x63,0xa0,0x6c,0xf6,0x15,0x47,0xc1,0x44,0x4b,0x87,0x52,0xaf,0xff,0x7e,0xbb,0x4a,0xf1,0xe2,0xa,0xc6,0x30,0x46,0x70,0xb6,0xc5,0xcc,0x6e,0x8c,0xe6}, +{0xa4,0xd5,0xa4,0x56,0xbd,0x4f,0xca,0x0,0xda,0x9d,0x84,0x4b,0xc8,0x3e,0x18,0xae,0x73,0x57,0xce,0x45,0x30,0x64,0xd1,0xad,0xe8,0xa6,0xce,0x68,0x14,0x5c,0x25,0x67}, +{0xa3,0xda,0x8c,0xf2,0xcb,0xe,0xe1,0x16,0x33,0xe9,0x6,0x58,0x9a,0x94,0x99,0x9a,0x1f,0x60,0xb2,0x20,0xc2,0x6f,0x84,0x7b,0xd1,0xce,0xac,0x7f,0xa0,0xd1,0x85,0x18}, +{0x32,0x59,0x5b,0xa1,0x8d,0xdd,0x19,0xd3,0x50,0x9a,0x1c,0xc0,0xaa,0xa5,0xb4,0x46,0x9f,0x3d,0x63,0x67,0xe4,0x4,0x6b,0xba,0xf6,0xca,0x19,0xab,0xb,0x56,0xee,0x7e}, +{0x1f,0xb1,0x79,0xea,0xa9,0x28,0x21,0x74,0xe9,0xbd,0xf7,0x35,0x3b,0x36,0x51,0xee,0x1d,0x57,0xac,0x5a,0x75,0x50,0xd3,0x76,0x3a,0x46,0xc2,0xfe,0xa3,0x7d,0x70,0x1}, +{0xf7,0x35,0xc1,0xaf,0x98,0xa4,0xd8,0x42,0x78,0xed,0xec,0x20,0x9e,0x6b,0x67,0x79,0x41,0x83,0x63,0x15,0xea,0x3a,0xdb,0xa8,0xfa,0xc3,0x3b,0x4d,0x32,0x83,0x2c,0x83}, +{0xa7,0x40,0x3b,0x1f,0x1c,0x27,0x47,0xf3,0x59,0x40,0xf0,0x34,0xb7,0x2d,0x76,0x9a,0xe7,0x3e,0x4e,0x6c,0xd2,0x21,0x4f,0xfd,0xb8,0xfd,0x8d,0x39,0xdc,0x57,0x59,0xef}, +{0x8d,0x9b,0xc,0x49,0x2b,0x49,0xeb,0xda,0x5b,0xa2,0xd7,0x49,0x68,0xf3,0x70,0xd,0x7d,0x3b,0xae,0xd0,0x7a,0x8d,0x55,0x84,0xf5,0xa5,0xe9,0xf0,0xe4,0xf8,0x8e,0x65}, +{0xa0,0xb8,0xa2,0xf4,0x36,0x10,0x3b,0x53,0xc,0xa8,0x7,0x9e,0x75,0x3e,0xec,0x5a,0x91,0x68,0x94,0x92,0x56,0xe8,0x88,0x4f,0x5b,0xb0,0x5c,0x55,0xf8,0xba,0xbc,0x4c}, +{0xe3,0xbb,0x3b,0x99,0xf3,0x87,0x94,0x7b,0x75,0xda,0xf4,0xd6,0x72,0x6b,0x1c,0x5d,0x64,0xae,0xac,0x28,0xdc,0x34,0xb3,0x6d,0x6c,0x34,0xa5,0x50,0xb8,0x28,0xdb,0x71}, +{0xf8,0x61,0xe2,0xf2,0x10,0x8d,0x51,0x2a,0xe3,0xdb,0x64,0x33,0x59,0xdd,0x75,0xfc,0x1c,0xac,0xbc,0xf1,0x43,0xce,0x3f,0xa2,0x67,0xbb,0xd1,0x3c,0x2,0xe8,0x43,0xb0}, +{0x33,0xa,0x5b,0xca,0x88,0x29,0xa1,0x75,0x7f,0x34,0x19,0x4d,0xb4,0x16,0x53,0x5c,0x92,0x3b,0x94,0xc3,0xe,0x79,0x4d,0x1e,0x79,0x74,0x75,0xd7,0xb6,0xee,0xaf,0x3f}, +{0xea,0xa8,0xd4,0xf7,0xbe,0x1a,0x39,0x21,0x5c,0xf4,0x7e,0x9,0x4c,0x23,0x27,0x51,0x26,0xa3,0x24,0x53,0xba,0x32,0x3c,0xd2,0x44,0xa3,0x17,0x4a,0x6d,0xa6,0xd5,0xad}, +{0xb5,0x1d,0x3e,0xa6,0xaf,0xf2,0xc9,0x8,0x83,0x59,0x3d,0x98,0x91,0x6b,0x3c,0x56,0x4c,0xf8,0x7c,0xa1,0x72,0x86,0x60,0x4d,0x46,0xe2,0x3e,0xcc,0x8,0x6e,0xc7,0xf6}, +{0x2f,0x98,0x33,0xb3,0xb1,0xbc,0x76,0x5e,0x2b,0xd6,0x66,0xa5,0xef,0xc4,0xe6,0x2a,0x6,0xf4,0xb6,0xe8,0xbe,0xc1,0xd4,0x36,0x74,0xee,0x82,0x15,0xbc,0xef,0x21,0x63}, +{0xfd,0xc1,0x4e,0xd,0xf4,0x53,0xc9,0x69,0xa7,0x7d,0x5a,0xc4,0x6,0x58,0x58,0x26,0x7e,0xc1,0x14,0x16,0x6,0xe0,0xfa,0x16,0x7e,0x90,0xaf,0x3d,0x28,0x63,0x9d,0x3f}, +{0xd2,0xc9,0xf2,0xe3,0x0,0x9b,0xd2,0xc,0x5f,0xaa,0xce,0x30,0xb7,0xd4,0xc,0x30,0x74,0x2a,0x51,0x16,0xf2,0xe0,0x32,0x98,0xd,0xeb,0x30,0xd8,0xe3,0xce,0xf8,0x9a}, +{0x4b,0xc5,0x9e,0x7b,0xb5,0xf1,0x79,0x92,0xff,0x51,0xe6,0x6e,0x4,0x86,0x68,0xd3,0x9b,0x23,0x4d,0x57,0xe6,0x96,0x67,0x31,0xcc,0xe6,0xa6,0xf3,0x17,0xa,0x75,0x5}, +{0xb1,0x76,0x81,0xd9,0x13,0x32,0x6c,0xce,0x3c,0x17,0x52,0x84,0xf8,0x5,0xa2,0x62,0xf4,0x2b,0xcb,0xb3,0x78,0x47,0x15,0x47,0xff,0x46,0x54,0x82,0x23,0x93,0x6a,0x48}, +{0x38,0xdf,0x58,0x7,0x4e,0x5e,0x65,0x65,0xf2,0xfc,0x7c,0x89,0xfc,0x86,0x50,0x8e,0x31,0x70,0x2e,0x44,0xd0,0xb,0xca,0x86,0xf0,0x40,0x9,0xa2,0x30,0x78,0x47,0x4e}, +{0x65,0xa0,0xee,0x39,0xd1,0xf7,0x38,0x83,0xf7,0x5e,0xe9,0x37,0xe4,0x2c,0x3a,0xbd,0x21,0x97,0xb2,0x26,0x1,0x13,0xf8,0x6f,0xa3,0x44,0xed,0xd1,0xef,0x9f,0xde,0xe7}, +{0x8b,0xa0,0xdf,0x15,0x76,0x25,0x92,0xd9,0x3c,0x85,0xf7,0xf6,0x12,0xdc,0x42,0xbe,0xd8,0xa7,0xec,0x7c,0xab,0x27,0xb0,0x7e,0x53,0x8d,0x7d,0xda,0xaa,0x3e,0xa8,0xde}, +{0xaa,0x25,0xce,0x93,0xbd,0x2,0x69,0xd8,0x5a,0xf6,0x43,0xfd,0x1a,0x73,0x8,0xf9,0xc0,0x5f,0xef,0xda,0x17,0x4a,0x19,0xa5,0x97,0x4d,0x66,0x33,0x4c,0xfd,0x21,0x6a}, +{0x35,0xb4,0x98,0x31,0xdb,0x41,0x15,0x70,0xea,0x1e,0xf,0xbb,0xed,0xcd,0x54,0x9b,0x9a,0xd0,0x63,0xa1,0x51,0x97,0x40,0x72,0xf6,0x75,0x9d,0xbf,0x91,0x47,0x6f,0xe2}}; + +/*swapping bit 2i with bit 2i+1 of 32-bit x*/ +#define SWAP1(x) (x) = ((((x) & 0x55555555UL) << 1) | (((x) & 0xaaaaaaaaUL) >> 1)); +/*swapping bits 4i||4i+1 with bits 4i+2||4i+3 of 32-bit x*/ +#define SWAP2(x) (x) = ((((x) & 0x33333333UL) << 2) | (((x) & 0xccccccccUL) >> 2)); +/*swapping bits 8i||8i+1||8i+2||8i+3 with bits 8i+4||8i+5||8i+6||8i+7 of 32-bit x*/ +#define SWAP4(x) (x) = ((((x) & 0x0f0f0f0fUL) << 4) | (((x) & 0xf0f0f0f0UL) >> 4)); +/*swapping bits 16i||16i+1||......||16i+7 with bits 16i+8||16i+9||......||16i+15 of 32-bit x*/ +#define SWAP8(x) (x) = ((((x) & 0x00ff00ffUL) << 8) | (((x) & 0xff00ff00UL) >> 8)); +/*swapping bits 32i||32i+1||......||32i+15 with bits 32i+16||32i+17||......||32i+31 of 32-bit x*/ +#define SWAP16(x) (x) = ((((x) & 0x0000ffffUL) << 16) | (((x) & 0xffff0000UL) >> 16)); + +/*The MDS transform*/ +#define L(m0,m1,m2,m3,m4,m5,m6,m7) \ + (m4) ^= (m1); \ + (m5) ^= (m2); \ + (m6) ^= (m0) ^ (m3); \ + (m7) ^= (m0); \ + (m0) ^= (m5); \ + (m1) ^= (m6); \ + (m2) ^= (m4) ^ (m7); \ + (m3) ^= (m4); + +/*The Sbox*/ +#define Sbox(m0,m1,m2,m3,cc) \ + m3 = ~(m3); \ + m0 ^= ((~(m2)) & (cc)); \ + temp0 = (cc) ^ ((m0) & (m1));\ + m0 ^= ((m2) & (m3)); \ + m3 ^= ((~(m1)) & (m2)); \ + m1 ^= ((m0) & (m2)); \ + m2 ^= ((m0) & (~(m3))); \ + m0 ^= ((m1) | (m3)); \ + m3 ^= ((m1) & (m2)); \ + m1 ^= (temp0 & (m0)); \ + m2 ^= temp0; + +__device__ __forceinline__ void Sbox_and_MDS_layer(hashState* state, uint32_t roundnumber) +{ + uint32_t temp0; + uint32_t cc0, cc1; + //Sbox and MDS layer +#pragma unroll 4 + for (int i = 0; i < 4; i++) { + cc0 = ((uint32_t*)c_E8_bitslice_roundconstant[roundnumber])[i]; + cc1 = ((uint32_t*)c_E8_bitslice_roundconstant[roundnumber])[i+4]; + Sbox(state->x[0][i],state->x[2][i], state->x[4][i], state->x[6][i], cc0); + Sbox(state->x[1][i],state->x[3][i], state->x[5][i], state->x[7][i], cc1); + L(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i]); + } +} + +__device__ __forceinline__ void RoundFunction0(hashState* state, uint32_t roundnumber) +{ + Sbox_and_MDS_layer(state, roundnumber); + +#pragma unroll 4 + for (int j = 1; j < 8; j = j+2) + { +#pragma unroll 4 + for (int i = 0; i < 4; i++) SWAP1(state->x[j][i]); + } +} + +__device__ __forceinline__ void RoundFunction1(hashState* state, uint32_t roundnumber) +{ + Sbox_and_MDS_layer(state, roundnumber); + +#pragma unroll 4 + for (int j = 1; j < 8; j = j+2) + { +#pragma unroll 4 + for (int i = 0; i < 4; i++) SWAP2(state->x[j][i]); + } +} + +__device__ __forceinline__ void RoundFunction2(hashState* state, uint32_t roundnumber) +{ + Sbox_and_MDS_layer(state, roundnumber); + +#pragma unroll 4 + for (int j = 1; j < 8; j = j+2) + { +#pragma unroll 4 + for (int i = 0; i < 4; i++) SWAP4(state->x[j][i]); + } +} + +__device__ __forceinline__ void RoundFunction3(hashState* state, uint32_t roundnumber) +{ + Sbox_and_MDS_layer(state, roundnumber); + +#pragma unroll 4 + for (int j = 1; j < 8; j = j+2) + { +#pragma unroll 4 + for (int i = 0; i < 4; i++) SWAP8(state->x[j][i]); + } +} + +__device__ __forceinline__ void RoundFunction4(hashState* state, uint32_t roundnumber) +{ + Sbox_and_MDS_layer(state, roundnumber); + +#pragma unroll 4 + for (int j = 1; j < 8; j = j+2) + { +#pragma unroll 4 + for (int i = 0; i < 4; i++) SWAP16(state->x[j][i]); + } +} + +__device__ __forceinline__ void RoundFunction5(hashState* state, uint32_t roundnumber) +{ + uint32_t temp0; + + Sbox_and_MDS_layer(state, roundnumber); + +#pragma unroll 4 + for (int j = 1; j < 8; j = j+2) + { +#pragma unroll 2 + for (int i = 0; i < 4; i = i+2) { + temp0 = state->x[j][i]; state->x[j][i] = state->x[j][i+1]; state->x[j][i+1] = temp0; + } + } +} + +__device__ __forceinline__ void RoundFunction6(hashState* state, uint32_t roundnumber) +{ + uint32_t temp0; + + Sbox_and_MDS_layer(state, roundnumber); + +#pragma unroll 4 + for (int j = 1; j < 8; j = j+2) + { +#pragma unroll 2 + for (int i = 0; i < 2; i++) { + temp0 = state->x[j][i]; state->x[j][i] = state->x[j][i+2]; state->x[j][i+2] = temp0; + } + } +} + +/*The bijective function E8, in bitslice form */ +__device__ __forceinline__ void E8(hashState *state) +{ + /*perform 6 rounds*/ +//#pragma unroll 6 + for (int i = 0; i < 42; i+=7) + { + RoundFunction0(state, i); + RoundFunction1(state, i+1); + RoundFunction2(state, i+2); + RoundFunction3(state, i+3); + RoundFunction4(state, i+4); + RoundFunction5(state, i+5); + RoundFunction6(state, i+6); + } +} + +/*The compression function F8 */ +__device__ __forceinline__ void F8(hashState *state) +{ + /*xor the 512-bit message with the fist half of the 1024-bit hash state*/ +#pragma unroll 16 + for (int i = 0; i < 16; i++) state->x[i >> 2][i & 3] ^= ((uint32_t*)state->buffer)[i]; + + /*the bijective function E8 */ + E8(state); + + /*xor the 512-bit message with the second half of the 1024-bit hash state*/ +#pragma unroll 16 + for (int i = 0; i < 16; i++) state->x[(16+i) >> 2][(16+i) & 3] ^= ((uint32_t*)state->buffer)[i]; +} + + +__device__ __forceinline__ void JHHash(const uint32_t *data, uint32_t *hashval) +{ + hashState state; + + /*load the intital hash value H0 into state*/ + /* + #define INIT(a,b,c,d) ((a) | ((b)<<8) | ((c)<<16) | ((d)<<24)) + state.x[0][0] = INIT(0x6f,0xd1,0x4b,0x96); + state.x[0][1] = INIT(0x3e,0x00,0xaa,0x17); + state.x[0][2] = INIT(0x63,0x6a,0x2e,0x05); + state.x[0][3] = INIT(0x7a,0x15,0xd5,0x43); + state.x[1][0] = INIT(0x8a,0x22,0x5e,0x8d); + state.x[1][1] = INIT(0x0c,0x97,0xef,0x0b); + state.x[1][2] = INIT(0xe9,0x34,0x12,0x59); + state.x[1][3] = INIT(0xf2,0xb3,0xc3,0x61); + state.x[2][0] = INIT(0x89,0x1d,0xa0,0xc1); + state.x[2][1] = INIT(0x53,0x6f,0x80,0x1e); + state.x[2][2] = INIT(0x2a,0xa9,0x05,0x6b); + state.x[2][3] = INIT(0xea,0x2b,0x6d,0x80); + state.x[3][0] = INIT(0x58,0x8e,0xcc,0xdb); + state.x[3][1] = INIT(0x20,0x75,0xba,0xa6); + state.x[3][2] = INIT(0xa9,0x0f,0x3a,0x76); + state.x[3][3] = INIT(0xba,0xf8,0x3b,0xf7); + state.x[4][0] = INIT(0x01,0x69,0xe6,0x05); + state.x[4][1] = INIT(0x41,0xe3,0x4a,0x69); + state.x[4][2] = INIT(0x46,0xb5,0x8a,0x8e); + state.x[4][3] = INIT(0x2e,0x6f,0xe6,0x5a); + state.x[5][0] = INIT(0x10,0x47,0xa7,0xd0); + state.x[5][1] = INIT(0xc1,0x84,0x3c,0x24); + state.x[5][2] = INIT(0x3b,0x6e,0x71,0xb1); + state.x[5][3] = INIT(0x2d,0x5a,0xc1,0x99); + state.x[6][0] = INIT(0xcf,0x57,0xf6,0xec); + state.x[6][1] = INIT(0x9d,0xb1,0xf8,0x56); + state.x[6][2] = INIT(0xa7,0x06,0x88,0x7c); + state.x[6][3] = INIT(0x57,0x16,0xb1,0x56); + state.x[7][0] = INIT(0xe3,0xc2,0xfc,0xdf); + state.x[7][1] = INIT(0xe6,0x85,0x17,0xfb); + state.x[7][2] = INIT(0x54,0x5a,0x46,0x78); + state.x[7][3] = INIT(0xcc,0x8c,0xdd,0x4b); + */ +#pragma unroll 8 + for(int j=0;j<8;j++) + { +#pragma unroll 4 + for(int i=0;i<4;i++) + state.x[j][i] = c_INIT_bitslice[j][i]; + } + +#pragma unroll 16 + for (int i=0; i < 16; ++i) state.buffer[i] = data[i]; + F8(&state); + + /*pad the message when databitlen is multiple of 512 bits, then process the padded block*/ + state.buffer[0] = 0x80; +#pragma unroll 14 + for (int i=1; i < 15; i++) state.buffer[i] = 0; + state.buffer[15] = 0x00020000; + F8(&state); + + /*truncating the final hash value to generate the message digest*/ +#pragma unroll 16 + for (int i=0; i < 16; ++i) hashval[i] = state.x[4][i]; +} + +// Die Hash-Funktion +__global__ void quark_jh512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_hash, uint32_t *g_nonceVector) +{ + int thread = (blockDim.x * blockIdx.x + threadIdx.x); + if (thread < threads) + { + uint32_t nounce = (g_nonceVector != NULL) ? g_nonceVector[thread] : (startNounce + thread); + + int hashPosition = nounce - startNounce; + uint32_t *Hash = (uint32_t*)&g_hash[8 * hashPosition]; + + JHHash(Hash, Hash); + } +} + + +// Setup-Funktionen +__host__ void quark_jh512_cpu_init(int thr_id, int threads) +{ + + cudaMemcpyToSymbol( c_E8_bitslice_roundconstant, + h_E8_bitslice_roundconstant, + sizeof(h_E8_bitslice_roundconstant), + 0, cudaMemcpyHostToDevice); + + cudaMemcpyToSymbol( c_INIT_bitslice, + h_INIT_bitslice, + sizeof(h_INIT_bitslice), + 0, cudaMemcpyHostToDevice); +} + +__host__ void quark_jh512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order) +{ + const int threadsperblock = 256; + + // berechne wie viele Thread Blocks wir brauchen + dim3 grid((threads + threadsperblock-1)/threadsperblock); + dim3 block(threadsperblock); + + // Größe des dynamischen Shared Memory Bereichs + size_t shared_size = 0; + +// fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size); + + quark_jh512_gpu_hash_64<<>>(threads, startNounce, (uint64_t*)d_hash, d_nonceVector); + MyStreamSynchronize(NULL, order, thr_id); +} + diff --git a/quark/cuda_quark_blake512.cu b/quark/cuda_quark_blake512.cu new file mode 100644 index 0000000..0e8cb3b --- /dev/null +++ b/quark/cuda_quark_blake512.cu @@ -0,0 +1,482 @@ +#include +#include "cuda_runtime.h" +#include "device_launch_parameters.h" + +#include +#include + +#define USE_SHUFFLE 0 + +// Folgende Definitionen später durch header ersetzen +typedef unsigned char uint8_t; +typedef unsigned int uint32_t; +typedef unsigned long long uint64_t; + +// aus heavy.cu +extern cudaError_t MyStreamSynchronize(cudaStream_t stream, int situation, int thr_id); + +// die Message it Padding zur Berechnung auf der GPU +__constant__ uint64_t c_PaddedMessage80[16]; // padded message (80 bytes + padding) + +// ---------------------------- BEGIN CUDA quark_blake512 functions ------------------------------------ + +__constant__ uint8_t c_sigma[16][16]; + +const uint8_t host_sigma[16][16] = +{ + { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 }, + {14, 10, 4, 8, 9, 15, 13, 6, 1, 12, 0, 2, 11, 7, 5, 3 }, + {11, 8, 12, 0, 5, 2, 15, 13, 10, 14, 3, 6, 7, 1, 9, 4 }, + { 7, 9, 3, 1, 13, 12, 11, 14, 2, 6, 5, 10, 4, 0, 15, 8 }, + { 9, 0, 5, 7, 2, 4, 10, 15, 14, 1, 11, 12, 6, 8, 3, 13 }, + { 2, 12, 6, 10, 0, 11, 8, 3, 4, 13, 7, 5, 15, 14, 1, 9 }, + {12, 5, 1, 15, 14, 13, 4, 10, 0, 7, 6, 3, 9, 2, 8, 11 }, + {13, 11, 7, 14, 12, 1, 3, 9, 5, 0, 15, 4, 8, 6, 2, 10 }, + { 6, 15, 14, 9, 11, 3, 0, 8, 12, 2, 13, 7, 1, 4, 10, 5 }, + {10, 2, 8, 4, 7, 6, 1, 5, 15, 11, 9, 14, 3, 12, 13 , 0 }, + { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 }, + {14, 10, 4, 8, 9, 15, 13, 6, 1, 12, 0, 2, 11, 7, 5, 3 }, + {11, 8, 12, 0, 5, 2, 15, 13, 10, 14, 3, 6, 7, 1, 9, 4 }, + { 7, 9, 3, 1, 13, 12, 11, 14, 2, 6, 5, 10, 4, 0, 15, 8 }, + { 9, 0, 5, 7, 2, 4, 10, 15, 14, 1, 11, 12, 6, 8, 3, 13 }, + { 2, 12, 6, 10, 0, 11, 8, 3, 4, 13, 7, 5, 15, 14, 1, 9 } +}; + +// das Hi Word aus einem 64 Bit Typen extrahieren +static __device__ uint32_t HIWORD(const uint64_t &x) { +#if __CUDA_ARCH__ >= 130 + return (uint32_t)__double2hiint(__longlong_as_double(x)); +#else + return (uint32_t)(x >> 32); +#endif +} + +// das Hi Word in einem 64 Bit Typen ersetzen +static __device__ uint64_t REPLACE_HIWORD(const uint64_t &x, const uint32_t &y) { + return (x & 0xFFFFFFFFULL) | (((uint64_t)y) << 32ULL); +} + +// das Lo Word aus einem 64 Bit Typen extrahieren +static __device__ uint32_t LOWORD(const uint64_t &x) { +#if __CUDA_ARCH__ >= 130 + return (uint32_t)__double2loint(__longlong_as_double(x)); +#else + return (uint32_t)(x & 0xFFFFFFFFULL); +#endif +} + +// das Lo Word in einem 64 Bit Typen ersetzen +static __device__ uint64_t REPLACE_LOWORD(const uint64_t &x, const uint32_t &y) { + return (x & 0xFFFFFFFF00000000ULL) | ((uint64_t)y); +} + +/* +#define SWAP32(x) \ + ((((x) << 24) & 0xff000000u) | (((x) << 8) & 0x00ff0000u) | \ + (((x) >> 8) & 0x0000ff00u) | (((x) >> 24) & 0x000000ffu)) + +#define SWAP64(x) \ + ((uint64_t)((((uint64_t)(x) & 0xff00000000000000ULL) >> 56) | \ + (((uint64_t)(x) & 0x00ff000000000000ULL) >> 40) | \ + (((uint64_t)(x) & 0x0000ff0000000000ULL) >> 24) | \ + (((uint64_t)(x) & 0x000000ff00000000ULL) >> 8) | \ + (((uint64_t)(x) & 0x00000000ff000000ULL) << 8) | \ + (((uint64_t)(x) & 0x0000000000ff0000ULL) << 24) | \ + (((uint64_t)(x) & 0x000000000000ff00ULL) << 40) | \ + (((uint64_t)(x) & 0x00000000000000ffULL) << 56))) +*/ + +/* +__device__ __forceinline__ void SWAP32(uint32_t *x) +{ + // Input: 33221100 + // Output: 00112233 + x[0] = __byte_perm(x[0], 0, 0x0123); +} +*/ +__device__ __forceinline__ uint64_t SWAP64(uint64_t x) +{ + // Input: 77665544 33221100 + // Output: 00112233 44556677 + uint64_t temp[2]; + temp[0] = __byte_perm(HIWORD(x), 0, 0x0123); + temp[1] = __byte_perm(LOWORD(x), 0, 0x0123); + + return temp[0] | (temp[1]<<32); +} + +__constant__ uint64_t c_u512[16]; + +const uint64_t host_u512[16] = +{ + 0x243f6a8885a308d3ULL, 0x13198a2e03707344ULL, + 0xa4093822299f31d0ULL, 0x082efa98ec4e6c89ULL, + 0x452821e638d01377ULL, 0xbe5466cf34e90c6cULL, + 0xc0ac29b7c97c50ddULL, 0x3f84d5b5b5470917ULL, + 0x9216d5d98979fb1bULL, 0xd1310ba698dfb5acULL, + 0x2ffd72dbd01adfb7ULL, 0xb8e1afed6a267e96ULL, + 0xba7c9045f12c7f99ULL, 0x24a19947b3916cf7ULL, + 0x0801f2e2858efc16ULL, 0x636920d871574e69ULL +}; + + +// diese 64 Bit Rotates werden unter Compute 3.5 (und besser) mit dem Funnel Shifter beschleunigt +#if __CUDA_ARCH__ >= 350 +__forceinline__ __device__ uint64_t ROTR(const uint64_t value, const int offset) { + uint2 result; + if(offset < 32) { + asm("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(result.x) : "r"(__double2loint(__longlong_as_double(value))), "r"(__double2hiint(__longlong_as_double(value))), "r"(offset)); + asm("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(result.y) : "r"(__double2hiint(__longlong_as_double(value))), "r"(__double2loint(__longlong_as_double(value))), "r"(offset)); + } else { + asm("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(result.x) : "r"(__double2hiint(__longlong_as_double(value))), "r"(__double2loint(__longlong_as_double(value))), "r"(offset)); + asm("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(result.y) : "r"(__double2loint(__longlong_as_double(value))), "r"(__double2hiint(__longlong_as_double(value))), "r"(offset)); + } + return __double_as_longlong(__hiloint2double(result.y, result.x)); +} +#else +#define ROTR(x, n) (((x) >> (n)) | ((x) << (64 - (n)))) +#endif + +#define G(a,b,c,d,e) \ + v[a] += (m[sigma[i][e]] ^ u512[sigma[i][e+1]]) + v[b];\ + v[d] = ROTR( v[d] ^ v[a],32); \ + v[c] += v[d]; \ + v[b] = ROTR( v[b] ^ v[c],25); \ + v[a] += (m[sigma[i][e+1]] ^ u512[sigma[i][e]])+v[b]; \ + v[d] = ROTR( v[d] ^ v[a],16); \ + v[c] += v[d]; \ + v[b] = ROTR( v[b] ^ v[c],11); + + +__device__ void quark_blake512_compress( uint64_t *h, const uint64_t *block, const uint8_t ((*sigma)[16]), const uint64_t *u512, const int bits ) +{ + uint64_t v[16], m[16], i; + +#pragma unroll 16 + for( i = 0; i < 16; ++i ) + { + m[i] = SWAP64(block[i]); + } + +#pragma unroll 8 + for( i = 0; i < 8; ++i ) v[i] = h[i]; + + v[ 8] = u512[0]; + v[ 9] = u512[1]; + v[10] = u512[2]; + v[11] = u512[3]; + v[12] = u512[4]; + v[13] = u512[5]; + v[14] = u512[6]; + v[15] = u512[7]; + + v[12] ^= bits; + v[13] ^= bits; + +//#pragma unroll 16 + for( i = 0; i < 16; ++i ) + { + /* column step */ + G( 0, 4, 8, 12, 0 ); + G( 1, 5, 9, 13, 2 ); + G( 2, 6, 10, 14, 4 ); + G( 3, 7, 11, 15, 6 ); + /* diagonal step */ + G( 0, 5, 10, 15, 8 ); + G( 1, 6, 11, 12, 10 ); + G( 2, 7, 8, 13, 12 ); + G( 3, 4, 9, 14, 14 ); + } + +#pragma unroll 16 + for( i = 0; i < 16; ++i ) h[i % 8] ^= v[i]; +} + +// Endian Drehung für 32 Bit Typen + +static __device__ uint32_t cuda_swab32(uint32_t x) +{ + return __byte_perm(x, 0, 0x0123); + /* + return (((x << 24) & 0xff000000u) | ((x << 8) & 0x00ff0000u) + | ((x >> 8) & 0x0000ff00u) | ((x >> 24) & 0x000000ffu)); + */ +} +/* +// Endian Drehung für 64 Bit Typen +static __device__ uint64_t cuda_swab64(uint64_t x) { + uint32_t h = (x >> 32); + uint32_t l = (x & 0xFFFFFFFFULL); + return (((uint64_t)cuda_swab32(l)) << 32) | ((uint64_t)cuda_swab32(h)); +} +*/ + +static __constant__ uint64_t d_constMem[8]; +static const uint64_t h_constMem[8] = { + 0x6a09e667f3bcc908ULL, + 0xbb67ae8584caa73bULL, + 0x3c6ef372fe94f82bULL, + 0xa54ff53a5f1d36f1ULL, + 0x510e527fade682d1ULL, + 0x9b05688c2b3e6c1fULL, + 0x1f83d9abfb41bd6bULL, + 0x5be0cd19137e2179ULL }; + +// Hash-Padding +static __constant__ uint64_t d_constHashPadding[8]; +static const uint64_t h_constHashPadding[8] = { + 0x0000000000000080ull, + 0, + 0, + 0, + 0, + 0x0100000000000000ull, + 0, + 0x0002000000000000ull }; + +__global__ void quark_blake512_gpu_hash_64(int threads, uint32_t startNounce, uint32_t *g_nonceVector, uint64_t *g_hash) +{ + int thread = (blockDim.x * blockIdx.x + threadIdx.x); + +#if USE_SHUFFLE + const int warpID = threadIdx.x & 0x0F; // 16 warps + const int warpBlockID = (thread + 15)>>4; // aufrunden auf volle Warp-Blöcke + const int maxHashPosition = thread<<3; +#endif + +#if USE_SHUFFLE + if (warpBlockID < ( (threads+15)>>4 )) +#else + if (thread < threads) +#endif + { + // bestimme den aktuellen Zähler + uint32_t nounce = (g_nonceVector != NULL) ? g_nonceVector[thread] : (startNounce + thread); + + int hashPosition = nounce - startNounce; + //uint64_t *inpHash = &g_hash[8 * hashPosition]; + uint64_t *inpHash = &g_hash[hashPosition<<3]; + + // State vorbereiten + uint64_t h[8]; + /* + h[0] = 0x6a09e667f3bcc908ULL; + h[1] = 0xbb67ae8584caa73bULL; + h[2] = 0x3c6ef372fe94f82bULL; + h[3] = 0xa54ff53a5f1d36f1ULL; + h[4] = 0x510e527fade682d1ULL; + h[5] = 0x9b05688c2b3e6c1fULL; + h[6] = 0x1f83d9abfb41bd6bULL; + h[7] = 0x5be0cd19137e2179ULL; + */ +#pragma unroll 8 + for(int i=0;i<8;i++) + h[i] = d_constMem[i]; + + // 128 Byte für die Message + uint64_t buf[16]; + + // Message für die erste Runde in Register holen +#pragma unroll 8 + for (int i=0; i < 8; ++i) buf[i] = inpHash[i]; + + /* + buf[ 8] = 0x0000000000000080ull; + buf[ 9] = 0; + buf[10] = 0; + buf[11] = 0; + buf[12] = 0; + buf[13] = 0x0100000000000000ull; + buf[14] = 0; + buf[15] = 0x0002000000000000ull; + */ +#pragma unroll 8 + for(int i=0;i<8;i++) + buf[i+8] = d_constHashPadding[i]; + + // die einzige Hashing-Runde + quark_blake512_compress( h, buf, c_sigma, c_u512, 512 ); + + // Hash rauslassen +#if __CUDA_ARCH__ >= 130 + // ausschliesslich 32 bit Operationen sofern die SM1.3 double intrinsics verfügbar sind + uint32_t *outHash = (uint32_t*)&g_hash[8 * hashPosition]; +#pragma unroll 8 + for (int i=0; i < 8; ++i) { + outHash[2*i+0] = cuda_swab32( HIWORD(h[i]) ); + outHash[2*i+1] = cuda_swab32( LOWORD(h[i]) ); + } +#else + // in dieser Version passieren auch ein paar 64 Bit Shifts + uint64_t *outHash = &g_hash[8 * hashPosition]; +#pragma unroll 8 + for (int i=0; i < 8; ++i) + { + //outHash[i] = cuda_swab64( h[i] ); + outHash[i] = SWAP64(h[i]); + } +#endif + } +} + +__global__ void quark_blake512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHash) +{ + int thread = (blockDim.x * blockIdx.x + threadIdx.x); + if (thread < threads) + { + // bestimme den aktuellen Zähler + uint32_t nounce = startNounce + thread; + + // State vorbereiten + uint64_t h[8]; + /* + h[0] = 0x6a09e667f3bcc908ULL; + h[1] = 0xbb67ae8584caa73bULL; + h[2] = 0x3c6ef372fe94f82bULL; + h[3] = 0xa54ff53a5f1d36f1ULL; + h[4] = 0x510e527fade682d1ULL; + h[5] = 0x9b05688c2b3e6c1fULL; + h[6] = 0x1f83d9abfb41bd6bULL; + h[7] = 0x5be0cd19137e2179ULL; + */ +#pragma unroll 8 + for(int i=0;i<8;i++) + h[i] = d_constMem[i]; + // 128 Byte für die Message + uint64_t buf[16]; + + // Message für die erste Runde in Register holen +#pragma unroll 16 + for (int i=0; i < 16; ++i) buf[i] = c_PaddedMessage80[i]; + + // die Nounce durch die thread-spezifische ersetzen + buf[9] = REPLACE_HIWORD(buf[9], cuda_swab32(nounce)); + + // die einzige Hashing-Runde + quark_blake512_compress( h, buf, c_sigma, c_u512, 640 ); + + // Hash rauslassen +#if __CUDA_ARCH__ >= 130 + // ausschliesslich 32 bit Operationen sofern die SM1.3 double intrinsics verfügbar sind + uint32_t *outHash = (uint32_t *)outputHash + 16 * thread; +#pragma unroll 8 + for (int i=0; i < 8; ++i) { + outHash[2*i+0] = cuda_swab32( HIWORD(h[i]) ); + outHash[2*i+1] = cuda_swab32( LOWORD(h[i]) ); + } +#else + // in dieser Version passieren auch ein paar 64 Bit Shifts + uint64_t *outHash = (uint64_t *)outputHash + 8 * thread; +#pragma unroll 8 + for (int i=0; i < 8; ++i) + { + //outHash[i] = cuda_swab64( h[i] ); + outHash[i] = SWAP64(h[i]); + } +#endif + } +} + + +// ---------------------------- END CUDA quark_blake512 functions ------------------------------------ + +// Setup-Funktionen +__host__ void quark_blake512_cpu_init(int thr_id, int threads) +{ + // Kopiere die Hash-Tabellen in den GPU-Speicher + cudaMemcpyToSymbol( c_sigma, + host_sigma, + sizeof(host_sigma), + 0, cudaMemcpyHostToDevice); + + cudaMemcpyToSymbol( c_u512, + host_u512, + sizeof(host_u512), + 0, cudaMemcpyHostToDevice); + + cudaMemcpyToSymbol( d_constMem, + h_constMem, + sizeof(h_constMem), + 0, cudaMemcpyHostToDevice); + + cudaMemcpyToSymbol( d_constHashPadding, + h_constHashPadding, + sizeof(h_constHashPadding), + 0, cudaMemcpyHostToDevice); +} + +// Blake512 für 80 Byte grosse Eingangsdaten +__host__ void quark_blake512_cpu_setBlock_80(void *pdata) +{ + // Message mit Padding bereitstellen + // lediglich die korrekte Nonce ist noch ab Byte 76 einzusetzen. + unsigned char PaddedMessage[128]; + memcpy(PaddedMessage, pdata, 80); + memset(PaddedMessage+80, 0, 48); + PaddedMessage[80] = 0x80; + PaddedMessage[111] = 1; + PaddedMessage[126] = 0x02; + PaddedMessage[127] = 0x80; + + // die Message zur Berechnung auf der GPU + cudaMemcpyToSymbol( c_PaddedMessage80, PaddedMessage, 16*sizeof(uint64_t), 0, cudaMemcpyHostToDevice); +} + +#if 0 +// Blake512 für 64 Byte grosse Eingangsdaten +// evtl. macht es gar keinen Sinn, das alles ins Constant Memory to schicken. Es sind hier sowieso +// nur die letzten 64 Bytes des Blocks konstant, und die meisten Bytes davon sind 0. Das kann mnan +// auch im Kernel initialisieren. +__host__ void quark_blake512_cpu_setBlock_64(void *pdata) +{ + // Message mit Padding bereitstellen + unsigned char PaddedMessage[128]; + memcpy(PaddedMessage, pdata, 64); // Hinweis: diese 64 Bytes sind nonce-spezifisch und ändern sich KOMPLETT für jede Nonce! + memset(PaddedMessage+64, 0, 64); + PaddedMessage[64] = 0x80; + PaddedMessage[111] = 1; + PaddedMessage[126] = 0x02; + PaddedMessage[127] = 0x00; + + // die Message zur Berechnung auf der GPU + cudaMemcpyToSymbol( c_PaddedMessage80, PaddedMessage, 16*sizeof(uint64_t), 0, cudaMemcpyHostToDevice); +} +#endif + +__host__ void quark_blake512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_outputHash, int order) +{ + const int threadsperblock = 256; + + // berechne wie viele Thread Blocks wir brauchen + dim3 grid((threads + threadsperblock-1)/threadsperblock); + dim3 block(threadsperblock); + + // Größe des dynamischen Shared Memory Bereichs + size_t shared_size = 0; + +// fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size); + + quark_blake512_gpu_hash_64<<>>(threads, startNounce, d_nonceVector, (uint64_t*)d_outputHash); + + // Strategisches Sleep Kommando zur Senkung der CPU Last + MyStreamSynchronize(NULL, order, thr_id); +} + +__host__ void quark_blake512_cpu_hash_80(int thr_id, int threads, uint32_t startNounce, uint32_t *d_outputHash, int order) +{ + const int threadsperblock = 256; + + // berechne wie viele Thread Blocks wir brauchen + dim3 grid((threads + threadsperblock-1)/threadsperblock); + dim3 block(threadsperblock); + + // Größe des dynamischen Shared Memory Bereichs + size_t shared_size = 0; + +// fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size); + + quark_blake512_gpu_hash_80<<>>(threads, startNounce, d_outputHash); + + // Strategisches Sleep Kommando zur Senkung der CPU Last + MyStreamSynchronize(NULL, order, thr_id); +} diff --git a/quark/cuda_quark_groestl512.cu b/quark/cuda_quark_groestl512.cu new file mode 100644 index 0000000..311abc7 --- /dev/null +++ b/quark/cuda_quark_groestl512.cu @@ -0,0 +1,397 @@ +// Auf QuarkCoin spezialisierte Version von Groestl + +#include +#include "cuda_runtime.h" +#include "device_launch_parameters.h" + +#include +#include + +// it's unfortunate that this is a compile time constant. +#define MAXWELL_OR_FERMI 0 + +// aus cpu-miner.c +extern int device_map[8]; + +// aus heavy.cu +extern cudaError_t MyStreamSynchronize(cudaStream_t stream, int situation, int thr_id); + +// Folgende Definitionen später durch header ersetzen +typedef unsigned char uint8_t; +typedef unsigned int uint32_t; +typedef unsigned long long uint64_t; + +// diese Struktur wird in der Init Funktion angefordert +static cudaDeviceProp props[8]; + +#define SPH_C32(x) ((uint32_t)(x ## U)) +#define SPH_T32(x) ((x) & SPH_C32(0xFFFFFFFF)) + +#define PC32up(j, r) ((uint32_t)((j) + (r))) +#define PC32dn(j, r) 0 +#define QC32up(j, r) 0xFFFFFFFF +#define QC32dn(j, r) (((uint32_t)(r) << 24) ^ SPH_T32(~((uint32_t)(j) << 24))) + +#define B32_0(x) __byte_perm(x, 0, 0x4440) +//((x) & 0xFF) +#define B32_1(x) __byte_perm(x, 0, 0x4441) +//(((x) >> 8) & 0xFF) +#define B32_2(x) __byte_perm(x, 0, 0x4442) +//(((x) >> 16) & 0xFF) +#define B32_3(x) __byte_perm(x, 0, 0x4443) +//((x) >> 24) + +#if MAXWELL_OR_FEMRI +#define USE_SHARED 1 +// Maxwell and Fermi cards get the best speed with SHARED access it seems. +#if USE_SHARED +#define T0up(x) (*((uint32_t*)mixtabs + ( (x)))) +#define T0dn(x) (*((uint32_t*)mixtabs + (256+(x)))) +#define T1up(x) (*((uint32_t*)mixtabs + (512+(x)))) +#define T1dn(x) (*((uint32_t*)mixtabs + (768+(x)))) +#define T2up(x) (*((uint32_t*)mixtabs + (1024+(x)))) +#define T2dn(x) (*((uint32_t*)mixtabs + (1280+(x)))) +#define T3up(x) (*((uint32_t*)mixtabs + (1536+(x)))) +#define T3dn(x) (*((uint32_t*)mixtabs + (1792+(x)))) +#else +#define T0up(x) tex1Dfetch(t0up1, x) +#define T0dn(x) tex1Dfetch(t0dn1, x) +#define T1up(x) tex1Dfetch(t1up1, x) +#define T1dn(x) tex1Dfetch(t1dn1, x) +#define T2up(x) tex1Dfetch(t2up1, x) +#define T2dn(x) tex1Dfetch(t2dn1, x) +#define T3up(x) tex1Dfetch(t3up1, x) +#define T3dn(x) tex1Dfetch(t3dn1, x) +#endif +#else +#define USE_SHARED 1 +// a healthy mix between shared and textured access provides the highest speed on Compute 3.0 and 3.5! +#define T0up(x) (*((uint32_t*)mixtabs + ( (x)))) +#define T0dn(x) tex1Dfetch(t0dn1, x) +#define T1up(x) tex1Dfetch(t1up1, x) +#define T1dn(x) (*((uint32_t*)mixtabs + (768+(x)))) +#define T2up(x) tex1Dfetch(t2up1, x) +#define T2dn(x) (*((uint32_t*)mixtabs + (1280+(x)))) +#define T3up(x) (*((uint32_t*)mixtabs + (1536+(x)))) +#define T3dn(x) tex1Dfetch(t3dn1, x) +#endif + +texture t0up1; +texture t0dn1; +texture t1up1; +texture t1dn1; +texture t2up1; +texture t2dn1; +texture t3up1; +texture t3dn1; + +extern uint32_t T0up_cpu[]; +extern uint32_t T0dn_cpu[]; +extern uint32_t T1up_cpu[]; +extern uint32_t T1dn_cpu[]; +extern uint32_t T2up_cpu[]; +extern uint32_t T2dn_cpu[]; +extern uint32_t T3up_cpu[]; +extern uint32_t T3dn_cpu[]; + +__device__ __forceinline__ void quark_groestl512_perm_P(uint32_t *a, char *mixtabs) +{ + uint32_t t[32]; + +//#pragma unroll 14 + for(int r=0;r<14;r++) + { + switch(r) + { + case 0: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 0); break; + case 1: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 1); break; + case 2: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 2); break; + case 3: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 3); break; + case 4: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 4); break; + case 5: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 5); break; + case 6: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 6); break; + case 7: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 7); break; + case 8: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 8); break; + case 9: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 9); break; + case 10: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 10); break; + case 11: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 11); break; + case 12: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 12); break; + case 13: +#pragma unroll 16 + for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 13); break; + } + + // RBTT +#pragma unroll 16 + for(int k=0;k<32;k+=2) + { + uint32_t t0_0 = B32_0(a[(k ) & 0x1f]), t9_0 = B32_0(a[(k + 9) & 0x1f]); + uint32_t t2_1 = B32_1(a[(k + 2) & 0x1f]), t11_1 = B32_1(a[(k + 11) & 0x1f]); + uint32_t t4_2 = B32_2(a[(k + 4) & 0x1f]), t13_2 = B32_2(a[(k + 13) & 0x1f]); + uint32_t t6_3 = B32_3(a[(k + 6) & 0x1f]), t23_3 = B32_3(a[(k + 23) & 0x1f]); + + t[k + 0] = T0up( t0_0 ) ^ T1up( t2_1 ) ^ T2up( t4_2 ) ^ T3up( t6_3 ) ^ + T0dn( t9_0 ) ^ T1dn( t11_1 ) ^ T2dn( t13_2 ) ^ T3dn( t23_3 ); + + t[k + 1] = T0dn( t0_0 ) ^ T1dn( t2_1 ) ^ T2dn( t4_2 ) ^ T3dn( t6_3 ) ^ + T0up( t9_0 ) ^ T1up( t11_1 ) ^ T2up( t13_2 ) ^ T3up( t23_3 ); + } +#pragma unroll 32 + for(int k=0;k<32;k++) + a[k] = t[k]; + } +} + +__device__ __forceinline__ void quark_groestl512_perm_Q(uint32_t *a, char *mixtabs) +{ +//#pragma unroll 14 + for(int r=0;r<14;r++) + { + uint32_t t[32]; + + switch(r) + { + case 0: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 0); a[(k*2)+1] ^= QC32dn(k<< 4, 0);} break; + case 1: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 1); a[(k*2)+1] ^= QC32dn(k<< 4, 1);} break; + case 2: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 2); a[(k*2)+1] ^= QC32dn(k<< 4, 2);} break; + case 3: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 3); a[(k*2)+1] ^= QC32dn(k<< 4, 3);} break; + case 4: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 4); a[(k*2)+1] ^= QC32dn(k<< 4, 4);} break; + case 5: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 5); a[(k*2)+1] ^= QC32dn(k<< 4, 5);} break; + case 6: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 6); a[(k*2)+1] ^= QC32dn(k<< 4, 6);} break; + case 7: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 7); a[(k*2)+1] ^= QC32dn(k<< 4, 7);} break; + case 8: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 8); a[(k*2)+1] ^= QC32dn(k<< 4, 8);} break; + case 9: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 9); a[(k*2)+1] ^= QC32dn(k<< 4, 9);} break; + case 10: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 10); a[(k*2)+1] ^= QC32dn(k<< 4, 10);} break; + case 11: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 11); a[(k*2)+1] ^= QC32dn(k<< 4, 11);} break; + case 12: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 12); a[(k*2)+1] ^= QC32dn(k<< 4, 12);} break; + case 13: + #pragma unroll 16 + for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 13); a[(k*2)+1] ^= QC32dn(k<< 4, 13);} break; + } + + // RBTT +#pragma unroll 16 + for(int k=0;k<32;k+=2) + { + uint32_t t2_0 = B32_0(a[(k + 2) & 0x1f]), t1_0 = B32_0(a[(k + 1) & 0x1f]); + uint32_t t6_1 = B32_1(a[(k + 6) & 0x1f]), t5_1 = B32_1(a[(k + 5) & 0x1f]); + uint32_t t10_2 = B32_2(a[(k + 10) & 0x1f]), t9_2 = B32_2(a[(k + 9) & 0x1f]); + uint32_t t22_3 = B32_3(a[(k + 22) & 0x1f]), t13_3 = B32_3(a[(k + 13) & 0x1f]); + + t[k + 0] = T0up( t2_0 ) ^ T1up( t6_1 ) ^ T2up( t10_2 ) ^ T3up( t22_3 ) ^ + T0dn( t1_0 ) ^ T1dn( t5_1 ) ^ T2dn( t9_2 ) ^ T3dn( t13_3 ); + + t[k + 1] = T0dn( t2_0 ) ^ T1dn( t6_1 ) ^ T2dn( t10_2 ) ^ T3dn( t22_3 ) ^ + T0up( t1_0 ) ^ T1up( t5_1 ) ^ T2up( t9_2 ) ^ T3up( t13_3 ); + } +#pragma unroll 32 + for(int k=0;k<32;k++) + a[k] = t[k]; + } +} +__global__ void quark_groestl512_gpu_hash_64(int threads, uint32_t startNounce, uint32_t *g_hash, uint32_t *g_nonceVector) +{ +#if USE_SHARED + extern __shared__ char mixtabs[]; + + if (threadIdx.x < 256) + { + *((uint32_t*)mixtabs + ( threadIdx.x)) = tex1Dfetch(t0up1, threadIdx.x); + *((uint32_t*)mixtabs + (256+threadIdx.x)) = tex1Dfetch(t0dn1, threadIdx.x); + *((uint32_t*)mixtabs + (512+threadIdx.x)) = tex1Dfetch(t1up1, threadIdx.x); + *((uint32_t*)mixtabs + (768+threadIdx.x)) = tex1Dfetch(t1dn1, threadIdx.x); + *((uint32_t*)mixtabs + (1024+threadIdx.x)) = tex1Dfetch(t2up1, threadIdx.x); + *((uint32_t*)mixtabs + (1280+threadIdx.x)) = tex1Dfetch(t2dn1, threadIdx.x); + *((uint32_t*)mixtabs + (1536+threadIdx.x)) = tex1Dfetch(t3up1, threadIdx.x); + *((uint32_t*)mixtabs + (1792+threadIdx.x)) = tex1Dfetch(t3dn1, threadIdx.x); + } + + __syncthreads(); +#endif + + int thread = (blockDim.x * blockIdx.x + threadIdx.x); + if (thread < threads) + { + // GROESTL + uint32_t message[32]; + uint32_t state[32]; + + uint32_t nounce = (g_nonceVector != NULL) ? g_nonceVector[thread] : (startNounce + thread); + + int hashPosition = nounce - startNounce; + uint32_t *inpHash = &g_hash[16 * hashPosition]; + +#pragma unroll 16 + for(int k=0;k<16;k++) message[k] = inpHash[k]; +#pragma unroll 14 + for(int k=1;k<15;k++) + message[k+16] = 0; + + message[16] = 0x80; + message[31] = 0x01000000; + +#pragma unroll 32 + for(int u=0;u<32;u++) state[u] = message[u]; + state[31] ^= 0x20000; + + // Perm +#if USE_SHARED + quark_groestl512_perm_P(state, mixtabs); + state[31] ^= 0x20000; + quark_groestl512_perm_Q(message, mixtabs); +#else + quark_groestl512_perm_P(state, NULL); + state[31] ^= 0x20000; + quark_groestl512_perm_Q(message, NULL); +#endif +#pragma unroll 32 + for(int u=0;u<32;u++) state[u] ^= message[u]; + +#pragma unroll 32 + for(int u=0;u<32;u++) message[u] = state[u]; + +#if USE_SHARED + quark_groestl512_perm_P(message, mixtabs); +#else + quark_groestl512_perm_P(message, NULL); +#endif + +#pragma unroll 32 + for(int u=0;u<32;u++) state[u] ^= message[u]; + // Erzeugten Hash rausschreiben + + uint32_t *outpHash = &g_hash[16 * hashPosition]; + +#pragma unroll 16 + for(int k=0;k<16;k++) outpHash[k] = state[k+16]; + } +} + +#define texDef(texname, texmem, texsource, texsize) \ + unsigned int *texmem; \ + cudaMalloc(&texmem, texsize); \ + cudaMemcpy(texmem, texsource, texsize, cudaMemcpyHostToDevice); \ + texname.normalized = 0; \ + texname.filterMode = cudaFilterModePoint; \ + texname.addressMode[0] = cudaAddressModeClamp; \ + { cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc(); \ + cudaBindTexture(NULL, &texname, texmem, &channelDesc, texsize ); } \ + +// Setup-Funktionen +__host__ void quark_groestl512_cpu_init(int thr_id, int threads) +{ + cudaGetDeviceProperties(&props[thr_id], device_map[thr_id]); + +// Texturen mit obigem Makro initialisieren + texDef(t0up1, d_T0up, T0up_cpu, sizeof(uint32_t)*256); + texDef(t0dn1, d_T0dn, T0dn_cpu, sizeof(uint32_t)*256); + texDef(t1up1, d_T1up, T1up_cpu, sizeof(uint32_t)*256); + texDef(t1dn1, d_T1dn, T1dn_cpu, sizeof(uint32_t)*256); + texDef(t2up1, d_T2up, T2up_cpu, sizeof(uint32_t)*256); + texDef(t2dn1, d_T2dn, T2dn_cpu, sizeof(uint32_t)*256); + texDef(t3up1, d_T3up, T3up_cpu, sizeof(uint32_t)*256); + texDef(t3dn1, d_T3dn, T3dn_cpu, sizeof(uint32_t)*256); +} + +__host__ void quark_groestl512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order) +{ + // Compute 3.5 und 5.x Geräte am besten mit 768 Threads ansteuern, + // alle anderen mit 512 Threads. + int threadsperblock = ((props[thr_id].major == 3 && props[thr_id].minor == 5) || props[thr_id].major > 3) ? 768 : 512; + + // berechne wie viele Thread Blocks wir brauchen + dim3 grid((threads + threadsperblock-1)/threadsperblock); + dim3 block(threadsperblock); + + // Größe des dynamischen Shared Memory Bereichs +#if USE_SHARED + size_t shared_size = 8 * 256 * sizeof(uint32_t); +#else + size_t shared_size = 0; +#endif + +// fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size); + //fprintf(stderr, "ThrID: %d\n", thr_id); + quark_groestl512_gpu_hash_64<<>>(threads, startNounce, d_hash, d_nonceVector); + + // Strategisches Sleep Kommando zur Senkung der CPU Last + MyStreamSynchronize(NULL, order, thr_id); +} + +__host__ void quark_doublegroestl512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order) +{ + // Compute 3.5 und 5.x Geräte am besten mit 768 Threads ansteuern, + // alle anderen mit 512 Threads. + int threadsperblock = ((props[thr_id].major == 3 && props[thr_id].minor == 5) || props[thr_id].major > 3) ? 768 : 512; + + // berechne wie viele Thread Blocks wir brauchen + dim3 grid((threads + threadsperblock-1)/threadsperblock); + dim3 block(threadsperblock); + + // Größe des dynamischen Shared Memory Bereichs +#if USE_SHARED + size_t shared_size = 8 * 256 * sizeof(uint32_t); +#else + size_t shared_size = 0; +#endif + +// fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size); + //fprintf(stderr, "ThrID: %d\n", thr_id); + quark_groestl512_gpu_hash_64<<>>(threads, startNounce, d_hash, d_nonceVector); + quark_groestl512_gpu_hash_64<<>>(threads, startNounce, d_hash, d_nonceVector); + + // Strategisches Sleep Kommando zur Senkung der CPU Last + MyStreamSynchronize(NULL, order, thr_id); +} diff --git a/quark/cuda_skein512.cu b/quark/cuda_skein512.cu new file mode 100644 index 0000000..7e807ea --- /dev/null +++ b/quark/cuda_skein512.cu @@ -0,0 +1,450 @@ +#include +#include "cuda_runtime.h" +#include "device_launch_parameters.h" + +#include +#include + +// Folgende Definitionen später durch header ersetzen +typedef unsigned char uint8_t; +typedef unsigned int uint32_t; +typedef unsigned long long uint64_t; + +#define SPH_C64(x) ((uint64_t)(x ## ULL)) + +// aus cpu-miner.c +extern "C" extern int device_map[8]; +// aus heavy.cu +extern cudaError_t MyStreamSynchronize(cudaStream_t stream, int situation, int thr_id); + +// Take a look at: https://www.schneier.com/skein1.3.pdf + +#if __CUDA_ARCH__ >= 350 +__forceinline__ __device__ uint64_t ROTL64(const uint64_t value, const int offset) { + uint2 result; + if(offset >= 32) { + asm("shf.l.wrap.b32 %0, %1, %2, %3;" : "=r"(result.x) : "r"(__double2loint(__longlong_as_double(value))), "r"(__double2hiint(__longlong_as_double(value))), "r"(offset)); + asm("shf.l.wrap.b32 %0, %1, %2, %3;" : "=r"(result.y) : "r"(__double2hiint(__longlong_as_double(value))), "r"(__double2loint(__longlong_as_double(value))), "r"(offset)); + } else { + asm("shf.l.wrap.b32 %0, %1, %2, %3;" : "=r"(result.x) : "r"(__double2hiint(__longlong_as_double(value))), "r"(__double2loint(__longlong_as_double(value))), "r"(offset)); + asm("shf.l.wrap.b32 %0, %1, %2, %3;" : "=r"(result.y) : "r"(__double2loint(__longlong_as_double(value))), "r"(__double2hiint(__longlong_as_double(value))), "r"(offset)); + } + return __double_as_longlong(__hiloint2double(result.y, result.x)); +} +#else +#define ROTL64(x, n) (((x) << (n)) | ((x) >> (64 - (n)))) +#endif +#define SHL(x, n) ((x) << (n)) +#define SHR(x, n) ((x) >> (n)) + +// Zum testen Hostcode... +/* Hier erstmal die Tabelle mit den Konstanten für die Mix-Funktion. Kann später vll. + mal direkt in den Code eingesetzt werden +*/ + +/* + * M9_ ## s ## _ ## i evaluates to s+i mod 9 (0 <= s <= 18, 0 <= i <= 7). + */ + +#define M9_0_0 0 +#define M9_0_1 1 +#define M9_0_2 2 +#define M9_0_3 3 +#define M9_0_4 4 +#define M9_0_5 5 +#define M9_0_6 6 +#define M9_0_7 7 + +#define M9_1_0 1 +#define M9_1_1 2 +#define M9_1_2 3 +#define M9_1_3 4 +#define M9_1_4 5 +#define M9_1_5 6 +#define M9_1_6 7 +#define M9_1_7 8 + +#define M9_2_0 2 +#define M9_2_1 3 +#define M9_2_2 4 +#define M9_2_3 5 +#define M9_2_4 6 +#define M9_2_5 7 +#define M9_2_6 8 +#define M9_2_7 0 + +#define M9_3_0 3 +#define M9_3_1 4 +#define M9_3_2 5 +#define M9_3_3 6 +#define M9_3_4 7 +#define M9_3_5 8 +#define M9_3_6 0 +#define M9_3_7 1 + +#define M9_4_0 4 +#define M9_4_1 5 +#define M9_4_2 6 +#define M9_4_3 7 +#define M9_4_4 8 +#define M9_4_5 0 +#define M9_4_6 1 +#define M9_4_7 2 + +#define M9_5_0 5 +#define M9_5_1 6 +#define M9_5_2 7 +#define M9_5_3 8 +#define M9_5_4 0 +#define M9_5_5 1 +#define M9_5_6 2 +#define M9_5_7 3 + +#define M9_6_0 6 +#define M9_6_1 7 +#define M9_6_2 8 +#define M9_6_3 0 +#define M9_6_4 1 +#define M9_6_5 2 +#define M9_6_6 3 +#define M9_6_7 4 + +#define M9_7_0 7 +#define M9_7_1 8 +#define M9_7_2 0 +#define M9_7_3 1 +#define M9_7_4 2 +#define M9_7_5 3 +#define M9_7_6 4 +#define M9_7_7 5 + +#define M9_8_0 8 +#define M9_8_1 0 +#define M9_8_2 1 +#define M9_8_3 2 +#define M9_8_4 3 +#define M9_8_5 4 +#define M9_8_6 5 +#define M9_8_7 6 + +#define M9_9_0 0 +#define M9_9_1 1 +#define M9_9_2 2 +#define M9_9_3 3 +#define M9_9_4 4 +#define M9_9_5 5 +#define M9_9_6 6 +#define M9_9_7 7 + +#define M9_10_0 1 +#define M9_10_1 2 +#define M9_10_2 3 +#define M9_10_3 4 +#define M9_10_4 5 +#define M9_10_5 6 +#define M9_10_6 7 +#define M9_10_7 8 + +#define M9_11_0 2 +#define M9_11_1 3 +#define M9_11_2 4 +#define M9_11_3 5 +#define M9_11_4 6 +#define M9_11_5 7 +#define M9_11_6 8 +#define M9_11_7 0 + +#define M9_12_0 3 +#define M9_12_1 4 +#define M9_12_2 5 +#define M9_12_3 6 +#define M9_12_4 7 +#define M9_12_5 8 +#define M9_12_6 0 +#define M9_12_7 1 + +#define M9_13_0 4 +#define M9_13_1 5 +#define M9_13_2 6 +#define M9_13_3 7 +#define M9_13_4 8 +#define M9_13_5 0 +#define M9_13_6 1 +#define M9_13_7 2 + +#define M9_14_0 5 +#define M9_14_1 6 +#define M9_14_2 7 +#define M9_14_3 8 +#define M9_14_4 0 +#define M9_14_5 1 +#define M9_14_6 2 +#define M9_14_7 3 + +#define M9_15_0 6 +#define M9_15_1 7 +#define M9_15_2 8 +#define M9_15_3 0 +#define M9_15_4 1 +#define M9_15_5 2 +#define M9_15_6 3 +#define M9_15_7 4 + +#define M9_16_0 7 +#define M9_16_1 8 +#define M9_16_2 0 +#define M9_16_3 1 +#define M9_16_4 2 +#define M9_16_5 3 +#define M9_16_6 4 +#define M9_16_7 5 + +#define M9_17_0 8 +#define M9_17_1 0 +#define M9_17_2 1 +#define M9_17_3 2 +#define M9_17_4 3 +#define M9_17_5 4 +#define M9_17_6 5 +#define M9_17_7 6 + +#define M9_18_0 0 +#define M9_18_1 1 +#define M9_18_2 2 +#define M9_18_3 3 +#define M9_18_4 4 +#define M9_18_5 5 +#define M9_18_6 6 +#define M9_18_7 7 + +/* + * M3_ ## s ## _ ## i evaluates to s+i mod 3 (0 <= s <= 18, 0 <= i <= 1). + */ + +#define M3_0_0 0 +#define M3_0_1 1 +#define M3_1_0 1 +#define M3_1_1 2 +#define M3_2_0 2 +#define M3_2_1 0 +#define M3_3_0 0 +#define M3_3_1 1 +#define M3_4_0 1 +#define M3_4_1 2 +#define M3_5_0 2 +#define M3_5_1 0 +#define M3_6_0 0 +#define M3_6_1 1 +#define M3_7_0 1 +#define M3_7_1 2 +#define M3_8_0 2 +#define M3_8_1 0 +#define M3_9_0 0 +#define M3_9_1 1 +#define M3_10_0 1 +#define M3_10_1 2 +#define M3_11_0 2 +#define M3_11_1 0 +#define M3_12_0 0 +#define M3_12_1 1 +#define M3_13_0 1 +#define M3_13_1 2 +#define M3_14_0 2 +#define M3_14_1 0 +#define M3_15_0 0 +#define M3_15_1 1 +#define M3_16_0 1 +#define M3_16_1 2 +#define M3_17_0 2 +#define M3_17_1 0 +#define M3_18_0 0 +#define M3_18_1 1 + +#define XCAT(x, y) XCAT_(x, y) +#define XCAT_(x, y) x ## y + +#define SKBI(k, s, i) XCAT(k, XCAT(XCAT(XCAT(M9_, s), _), i)) +#define SKBT(t, s, v) XCAT(t, XCAT(XCAT(XCAT(M3_, s), _), v)) + +#define TFBIG_KINIT(k0, k1, k2, k3, k4, k5, k6, k7, k8, t0, t1, t2) { \ + k8 = ((k0 ^ k1) ^ (k2 ^ k3)) ^ ((k4 ^ k5) ^ (k6 ^ k7)) \ + ^ SPH_C64(0x1BD11BDAA9FC1A22); \ + t2 = t0 ^ t1; \ + } + +#define TFBIG_ADDKEY(w0, w1, w2, w3, w4, w5, w6, w7, k, t, s) { \ + w0 = (w0 + SKBI(k, s, 0)); \ + w1 = (w1 + SKBI(k, s, 1)); \ + w2 = (w2 + SKBI(k, s, 2)); \ + w3 = (w3 + SKBI(k, s, 3)); \ + w4 = (w4 + SKBI(k, s, 4)); \ + w5 = (w5 + SKBI(k, s, 5) + SKBT(t, s, 0)); \ + w6 = (w6 + SKBI(k, s, 6) + SKBT(t, s, 1)); \ + w7 = (w7 + SKBI(k, s, 7) + (uint64_t)s); \ + } + +#define TFBIG_MIX(x0, x1, rc) { \ + x0 = x0 + x1; \ + x1 = ROTL64(x1, rc) ^ x0; \ + } + +#define TFBIG_MIX8(w0, w1, w2, w3, w4, w5, w6, w7, rc0, rc1, rc2, rc3) { \ + TFBIG_MIX(w0, w1, rc0); \ + TFBIG_MIX(w2, w3, rc1); \ + TFBIG_MIX(w4, w5, rc2); \ + TFBIG_MIX(w6, w7, rc3); \ + } + +#define TFBIG_4e(s) { \ + TFBIG_ADDKEY(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], h, t, s); \ + TFBIG_MIX8(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 46, 36, 19, 37); \ + TFBIG_MIX8(p[2], p[1], p[4], p[7], p[6], p[5], p[0], p[3], 33, 27, 14, 42); \ + TFBIG_MIX8(p[4], p[1], p[6], p[3], p[0], p[5], p[2], p[7], 17, 49, 36, 39); \ + TFBIG_MIX8(p[6], p[1], p[0], p[7], p[2], p[5], p[4], p[3], 44, 9, 54, 56); \ + } + +#define TFBIG_4o(s) { \ + TFBIG_ADDKEY(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], h, t, s); \ + TFBIG_MIX8(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 39, 30, 34, 24); \ + TFBIG_MIX8(p[2], p[1], p[4], p[7], p[6], p[5], p[0], p[3], 13, 50, 10, 17); \ + TFBIG_MIX8(p[4], p[1], p[6], p[3], p[0], p[5], p[2], p[7], 25, 29, 39, 43); \ + TFBIG_MIX8(p[6], p[1], p[0], p[7], p[2], p[5], p[4], p[3], 8, 35, 56, 22); \ + } + +static __constant__ uint64_t d_constMem[8]; +static uint64_t h_constMem[8] = { + SPH_C64(0x4903ADFF749C51CE), + SPH_C64(0x0D95DE399746DF03), + SPH_C64(0x8FD1934127C79BCE), + SPH_C64(0x9A255629FF352CB1), + SPH_C64(0x5DB62599DF6CA7B0), + SPH_C64(0xEABE394CA9D5C3F4), + SPH_C64(0x991112C71A75B523), + SPH_C64(0xAE18A40B660FCC33) }; + +__global__ void quark_skein512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_hash, uint32_t *g_nonceVector) +{ + int thread = (blockDim.x * blockIdx.x + threadIdx.x); + if (thread < threads) + { + // Skein + uint64_t p[8]; + uint64_t h0, h1, h2, h3, h4, h5, h6, h7, h8; + uint64_t t0, t1, t2; + + uint32_t nounce = (g_nonceVector != NULL) ? g_nonceVector[thread] : (startNounce + thread); + + int hashPosition = nounce - startNounce; + uint64_t *inpHash = &g_hash[8 * hashPosition]; + + // Initialisierung + h0 = d_constMem[0]; + h1 = d_constMem[1]; + h2 = d_constMem[2]; + h3 = d_constMem[3]; + h4 = d_constMem[4]; + h5 = d_constMem[5]; + h6 = d_constMem[6]; + h7 = d_constMem[7]; + + // 1. Runde -> etype = 480, ptr = 64, bcount = 0, data = msg +#pragma unroll 8 + for(int i=0;i<8;i++) + p[i] = inpHash[i]; + + t0 = 64; // ptr + t1 = 480ull << 55; // etype + TFBIG_KINIT(h0, h1, h2, h3, h4, h5, h6, h7, h8, t0, t1, t2); + TFBIG_4e(0); + TFBIG_4o(1); + TFBIG_4e(2); + TFBIG_4o(3); + TFBIG_4e(4); + TFBIG_4o(5); + TFBIG_4e(6); + TFBIG_4o(7); + TFBIG_4e(8); + TFBIG_4o(9); + TFBIG_4e(10); + TFBIG_4o(11); + TFBIG_4e(12); + TFBIG_4o(13); + TFBIG_4e(14); + TFBIG_4o(15); + TFBIG_4e(16); + TFBIG_4o(17); + TFBIG_ADDKEY(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], h, t, 18); + + h0 = inpHash[0] ^ p[0]; + h1 = inpHash[1] ^ p[1]; + h2 = inpHash[2] ^ p[2]; + h3 = inpHash[3] ^ p[3]; + h4 = inpHash[4] ^ p[4]; + h5 = inpHash[5] ^ p[5]; + h6 = inpHash[6] ^ p[6]; + h7 = inpHash[7] ^ p[7]; + + // 2. Runde -> etype = 510, ptr = 8, bcount = 0, data = 0 +#pragma unroll 8 + for(int i=0;i<8;i++) + p[i] = 0; + + t0 = 8; // ptr + t1 = 510ull << 55; // etype + TFBIG_KINIT(h0, h1, h2, h3, h4, h5, h6, h7, h8, t0, t1, t2); + TFBIG_4e(0); + TFBIG_4o(1); + TFBIG_4e(2); + TFBIG_4o(3); + TFBIG_4e(4); + TFBIG_4o(5); + TFBIG_4e(6); + TFBIG_4o(7); + TFBIG_4e(8); + TFBIG_4o(9); + TFBIG_4e(10); + TFBIG_4o(11); + TFBIG_4e(12); + TFBIG_4o(13); + TFBIG_4e(14); + TFBIG_4o(15); + TFBIG_4e(16); + TFBIG_4o(17); + TFBIG_ADDKEY(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], h, t, 18); + + // fertig + uint64_t *outpHash = &g_hash[8 * hashPosition]; + +#pragma unroll 8 + for(int i=0;i<8;i++) + outpHash[i] = p[i]; + } +} + +// Setup-Funktionen +__host__ void quark_skein512_cpu_init(int thr_id, int threads) +{ + // nix zu tun ;-) + cudaMemcpyToSymbol( d_constMem, + h_constMem, + sizeof(h_constMem), + 0, cudaMemcpyHostToDevice); +} + +__host__ void quark_skein512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order) +{ + const int threadsperblock = 256; + + // berechne wie viele Thread Blocks wir brauchen + dim3 grid((threads + threadsperblock-1)/threadsperblock); + dim3 block(threadsperblock); + + // Größe des dynamischen Shared Memory Bereichs + size_t shared_size = 0; + +// fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size); + quark_skein512_gpu_hash_64<<>>(threads, startNounce, (uint64_t*)d_hash, d_nonceVector); + + // Strategisches Sleep Kommando zur Senkung der CPU Last + MyStreamSynchronize(NULL, order, thr_id); +}