From be478bd72576ec9ee7ca37580a8e362c8a04d442 Mon Sep 17 00:00:00 2001 From: Tanguy Pruvot Date: Tue, 12 May 2015 02:35:56 +0200 Subject: [PATCH] groestl: tabs to space + arch check --- cuda_groestlcoin.cu | 213 ++++++++++++++++++++++---------------------- 1 file changed, 108 insertions(+), 105 deletions(-) diff --git a/cuda_groestlcoin.cu b/cuda_groestlcoin.cu index 6d57fe8..27da418 100644 --- a/cuda_groestlcoin.cu +++ b/cuda_groestlcoin.cu @@ -4,9 +4,9 @@ #include #include "cuda_helper.h" -#include -// globaler Speicher für alle HeftyHashes aller Threads +#include "miner.h" + __constant__ uint32_t pTarget[8]; // Single GPU __constant__ uint32_t groestlcoin_gpu_msg[32]; @@ -24,135 +24,138 @@ __global__ __launch_bounds__(256, 4) void groestlcoin_gpu_hash_quad(uint32_t threads, uint32_t startNounce, uint32_t *resNounce) { #if __CUDA_ARCH__ >= 300 - // durch 4 dividieren, weil jeweils 4 Threads zusammen ein Hash berechnen - uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x) / 4; - if (thread < threads) - { - // GROESTL - uint32_t paddedInput[8]; - - #pragma unroll 8 - for(int k=0;k<8;k++) paddedInput[k] = groestlcoin_gpu_msg[4*k+threadIdx.x%4]; - - uint32_t nounce = startNounce + thread; - if ((threadIdx.x % 4) == 3) - paddedInput[4] = SWAB32(nounce); // 4*4+3 = 19 - - uint32_t msgBitsliced[8]; - to_bitslice_quad(paddedInput, msgBitsliced); - - uint32_t state[8]; - for (int round=0; round<2; round++) - { - groestl512_progressMessage_quad(state, msgBitsliced); - - if (round < 1) - { - // Verkettung zweier Runden inclusive Padding. - msgBitsliced[ 0] = __byte_perm(state[ 0], 0x00800100, 0x4341 + ((threadIdx.x%4)==3)*0x2000); - msgBitsliced[ 1] = __byte_perm(state[ 1], 0x00800100, 0x4341); - msgBitsliced[ 2] = __byte_perm(state[ 2], 0x00800100, 0x4341); - msgBitsliced[ 3] = __byte_perm(state[ 3], 0x00800100, 0x4341); - msgBitsliced[ 4] = __byte_perm(state[ 4], 0x00800100, 0x4341); - msgBitsliced[ 5] = __byte_perm(state[ 5], 0x00800100, 0x4341); - msgBitsliced[ 6] = __byte_perm(state[ 6], 0x00800100, 0x4341); - msgBitsliced[ 7] = __byte_perm(state[ 7], 0x00800100, 0x4341 + ((threadIdx.x%4)==0)*0x0010); - } - } - - // Nur der erste von jeweils 4 Threads bekommt das Ergebns-Hash - uint32_t out_state[16]; - from_bitslice_quad(state, out_state); - - if (threadIdx.x % 4 == 0) - { - int i, position = -1; - bool rc = true; - - #pragma unroll 8 - for (i = 7; i >= 0; i--) { - if (out_state[i] > pTarget[i]) { - if(position < i) { - position = i; - rc = false; - } - } - if (out_state[i] < pTarget[i]) { - if(position < i) { - position = i; - rc = true; - } - } - } - - if(rc == true) - if(resNounce[0] > nounce) - resNounce[0] = nounce; - } - } + // durch 4 dividieren, weil jeweils 4 Threads zusammen ein Hash berechnen + uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x) / 4; + if (thread < threads) + { + // GROESTL + uint32_t paddedInput[8]; + + #pragma unroll 8 + for(int k=0;k<8;k++) paddedInput[k] = groestlcoin_gpu_msg[4*k+threadIdx.x%4]; + + uint32_t nounce = startNounce + thread; + if ((threadIdx.x % 4) == 3) + paddedInput[4] = SWAB32(nounce); // 4*4+3 = 19 + + uint32_t msgBitsliced[8]; + to_bitslice_quad(paddedInput, msgBitsliced); + + uint32_t state[8]; + for (int round=0; round<2; round++) + { + groestl512_progressMessage_quad(state, msgBitsliced); + + if (round < 1) + { + // Verkettung zweier Runden inclusive Padding. + msgBitsliced[ 0] = __byte_perm(state[ 0], 0x00800100, 0x4341 + ((threadIdx.x%4)==3)*0x2000); + msgBitsliced[ 1] = __byte_perm(state[ 1], 0x00800100, 0x4341); + msgBitsliced[ 2] = __byte_perm(state[ 2], 0x00800100, 0x4341); + msgBitsliced[ 3] = __byte_perm(state[ 3], 0x00800100, 0x4341); + msgBitsliced[ 4] = __byte_perm(state[ 4], 0x00800100, 0x4341); + msgBitsliced[ 5] = __byte_perm(state[ 5], 0x00800100, 0x4341); + msgBitsliced[ 6] = __byte_perm(state[ 6], 0x00800100, 0x4341); + msgBitsliced[ 7] = __byte_perm(state[ 7], 0x00800100, 0x4341 + ((threadIdx.x%4)==0)*0x0010); + } + } + + // Nur der erste von jeweils 4 Threads bekommt das Ergebns-Hash + uint32_t out_state[16]; + from_bitslice_quad(state, out_state); + + if (threadIdx.x % 4 == 0) + { + int i, position = -1; + bool rc = true; + + #pragma unroll 8 + for (i = 7; i >= 0; i--) { + if (out_state[i] > pTarget[i]) { + if(position < i) { + position = i; + rc = false; + } + } + if (out_state[i] < pTarget[i]) { + if(position < i) { + position = i; + rc = true; + } + } + } + + if(rc == true) + if(resNounce[0] > nounce) + resNounce[0] = nounce; + } + } #endif } __host__ void groestlcoin_cpu_init(int thr_id, uint32_t threads) { - cudaMalloc(&d_resultNonce[thr_id], sizeof(uint32_t)); + // to check if the binary supports SM3+ + cuda_get_arch(thr_id); + + cudaMalloc(&d_resultNonce[thr_id], sizeof(uint32_t)); } __host__ void groestlcoin_cpu_setBlock(int thr_id, void *data, void *pTargetIn) { - // Nachricht expandieren und setzen - uint32_t msgBlock[32]; + uint32_t msgBlock[32]; - memset(msgBlock, 0, sizeof(uint32_t) * 32); - memcpy(&msgBlock[0], data, 80); + memset(msgBlock, 0, sizeof(uint32_t) * 32); + memcpy(&msgBlock[0], data, 80); - // Erweitere die Nachricht auf den Nachrichtenblock (padding) - // Unsere Nachricht hat 80 Byte - msgBlock[20] = 0x80; - msgBlock[31] = 0x01000000; + // Erweitere die Nachricht auf den Nachrichtenblock (padding) + // Unsere Nachricht hat 80 Byte + msgBlock[20] = 0x80; + msgBlock[31] = 0x01000000; - // groestl512 braucht hierfür keinen CPU-Code (die einzige Runde wird - // auf der GPU ausgeführt) + // groestl512 braucht hierfür keinen CPU-Code (die einzige Runde wird + // auf der GPU ausgeführt) - // Blockheader setzen (korrekte Nonce und Hefty Hash fehlen da drin noch) - cudaMemcpyToSymbol( groestlcoin_gpu_msg, - msgBlock, - 128); + // Blockheader setzen (korrekte Nonce und Hefty Hash fehlen da drin noch) + cudaMemcpyToSymbol( groestlcoin_gpu_msg, + msgBlock, + 128); - cudaMemset(d_resultNonce[thr_id], 0xFF, sizeof(uint32_t)); - cudaMemcpyToSymbol( pTarget, - pTargetIn, - sizeof(uint32_t) * 8 ); + cudaMemset(d_resultNonce[thr_id], 0xFF, sizeof(uint32_t)); + cudaMemcpyToSymbol( pTarget, + pTargetIn, + sizeof(uint32_t) * 8 ); } __host__ void groestlcoin_cpu_hash(int thr_id, uint32_t threads, uint32_t startNounce, void *outputHashes, uint32_t *nounce) { - uint32_t threadsperblock = 256; + uint32_t threadsperblock = 256; - // Compute 3.0 benutzt die registeroptimierte Quad Variante mit Warp Shuffle - // mit den Quad Funktionen brauchen wir jetzt 4 threads pro Hash, daher Faktor 4 bei der Blockzahl - int factor = 4; + // Compute 3.0 benutzt die registeroptimierte Quad Variante mit Warp Shuffle + // mit den Quad Funktionen brauchen wir jetzt 4 threads pro Hash, daher Faktor 4 bei der Blockzahl + int factor = 4; - // berechne wie viele Thread Blocks wir brauchen - dim3 grid(factor*((threads + threadsperblock-1)/threadsperblock)); - dim3 block(threadsperblock); + // berechne wie viele Thread Blocks wir brauchen + dim3 grid(factor*((threads + threadsperblock-1)/threadsperblock)); + dim3 block(threadsperblock); - // Größe des dynamischen Shared Memory Bereichs - size_t shared_size = 0; + // Größe des dynamischen Shared Memory Bereichs + size_t shared_size = 0; - if (device_sm[device_map[thr_id]] < 300) { - printf("Sorry, This algo is not supported by this GPU arch (SM 3.0 required)"); - return; - } + int dev_id = device_map[thr_id]; + if (device_sm[dev_id] < 300 || cuda_arch[dev_id] < 300) { + printf("Sorry, This algo is not supported by this GPU arch (SM 3.0 required)"); + proper_exit(EXIT_CODE_CUDA_ERROR); + } - cudaMemset(d_resultNonce[thr_id], 0xFF, sizeof(uint32_t)); - groestlcoin_gpu_hash_quad<<>>(threads, startNounce, d_resultNonce[thr_id]); + cudaMemset(d_resultNonce[thr_id], 0xFF, sizeof(uint32_t)); + groestlcoin_gpu_hash_quad<<>>(threads, startNounce, d_resultNonce[thr_id]); - // Strategisches Sleep Kommando zur Senkung der CPU Last - MyStreamSynchronize(NULL, 0, thr_id); + // Strategisches Sleep Kommando zur Senkung der CPU Last + MyStreamSynchronize(NULL, 0, thr_id); - cudaMemcpy(nounce, d_resultNonce[thr_id], sizeof(uint32_t), cudaMemcpyDeviceToHost); + cudaMemcpy(nounce, d_resultNonce[thr_id], sizeof(uint32_t), cudaMemcpyDeviceToHost); }