From ebf885d4826d4310a0711e9304116323cb249b6b Mon Sep 17 00:00:00 2001 From: pallas1 Date: Sat, 2 Apr 2016 15:28:38 +0200 Subject: [PATCH] ~10% speedup --- Algo256/decred.cu | 256 +++++++++++++++++++++++++--------------------- 1 file changed, 138 insertions(+), 118 deletions(-) diff --git a/Algo256/decred.cu b/Algo256/decred.cu index 770cc38..fe6092c 100644 --- a/Algo256/decred.cu +++ b/Algo256/decred.cu @@ -2,6 +2,8 @@ * Blake-256 Decred 180-Bytes input Cuda Kernel (Tested on SM 5/5.2) * * Tanguy Pruvot - Feb 2016 + * + * Revised for optimisation by pallas @ bitcointalk - Apr 2016 */ #include @@ -14,7 +16,9 @@ extern "C" { } /* threads per block */ -#define TPB 256 +#define TPB 512 +/* nonces per round */ +#define NPR 128 /* hash by cpu with blake 256 */ extern "C" void decred_hash(void *output, const void *input) @@ -46,6 +50,7 @@ static uint32_t *h_resNonce[MAX_GPUS]; static uint32_t extra_results[NBN] = { UINT32_MAX }; #endif + /* ############################################################################################################################### */ #define GSPREC(a,b,c,d,x,y) { \ @@ -59,107 +64,106 @@ static uint32_t extra_results[NBN] = { UINT32_MAX }; v[b] = SPH_ROTR32(v[b] ^ v[c], 7); \ } -__device__ __forceinline__ -void blake256_compress_14(uint32_t *h, const uint32_t nonce, const uint32_t T0) -{ - uint32_t v[16]; - - #pragma unroll 8 - for(uint32_t i = 0; i < 8; i++) - v[i] = h[i]; - const uint32_t c_u256[16] = { - 0x243F6A88, 0x85A308D3, 0x13198A2E, 0x03707344, - 0xA4093822, 0x299F31D0, 0x082EFA98, 0xEC4E6C89, - 0x452821E6, 0x38D01377, 0xBE5466CF, 0x34E90C6C, - 0xC0AC29B7, 0xC97C50DD, 0x3F84D5B5, 0xB5470917 - }; +#define GSPREC4(a0,b0,c0,d0,x0,y0,a1,b1,c1,d1,x1,y1,a2,b2,c2,d2,x2,y2,a3,b3,c3,d3,x3,y3) { \ + v[a0] += (m[x0] ^ c_u256[y0]) + v[b0]; \ + v[a1] += (m[x1] ^ c_u256[y1]) + v[b1]; \ + v[a2] += (m[x2] ^ c_u256[y2]) + v[b2]; \ + v[a3] += (m[x3] ^ c_u256[y3]) + v[b3]; \ + v[d0] = __byte_perm(v[d0] ^ v[a0], 0, 0x1032); \ + v[d1] = __byte_perm(v[d1] ^ v[a1], 0, 0x1032); \ + v[d2] = __byte_perm(v[d2] ^ v[a2], 0, 0x1032); \ + v[d3] = __byte_perm(v[d3] ^ v[a3], 0, 0x1032); \ + v[c0] += v[d0]; \ + v[c1] += v[d1]; \ + v[c2] += v[d2]; \ + v[c3] += v[d3]; \ + v[b0] = SPH_ROTR32(v[b0] ^ v[c0], 12); \ + v[b1] = SPH_ROTR32(v[b1] ^ v[c1], 12); \ + v[b2] = SPH_ROTR32(v[b2] ^ v[c2], 12); \ + v[b3] = SPH_ROTR32(v[b3] ^ v[c3], 12); \ + v[a0] += (m[y0] ^ c_u256[x0]) + v[b0]; \ + v[a1] += (m[y1] ^ c_u256[x1]) + v[b1]; \ + v[a2] += (m[y2] ^ c_u256[x2]) + v[b2]; \ + v[a3] += (m[y3] ^ c_u256[x3]) + v[b3]; \ + v[d0] = __byte_perm(v[d0] ^ v[a0], 0, 0x0321); \ + v[d1] = __byte_perm(v[d1] ^ v[a1], 0, 0x0321); \ + v[d2] = __byte_perm(v[d2] ^ v[a2], 0, 0x0321); \ + v[d3] = __byte_perm(v[d3] ^ v[a3], 0, 0x0321); \ + v[c0] += v[d0]; \ + v[c1] += v[d1]; \ + v[c2] += v[d2]; \ + v[c3] += v[d3]; \ + v[b0] = SPH_ROTR32(v[b0] ^ v[c0], 7); \ + v[b1] = SPH_ROTR32(v[b1] ^ v[c1], 7); \ + v[b2] = SPH_ROTR32(v[b2] ^ v[c2], 7); \ + v[b3] = SPH_ROTR32(v[b3] ^ v[c3], 7); \ +} - v[ 8] = c_u256[0]; - v[ 9] = c_u256[1]; - v[10] = c_u256[2]; - v[11] = c_u256[3]; - v[12] = c_u256[4] ^ T0; - v[13] = c_u256[5] ^ T0; - v[14] = c_u256[6]; - v[15] = c_u256[7]; +static const __constant__ uint32_t c_u256[16] = { + 0x243F6A88, 0x85A308D3, 0x13198A2E, 0x03707344, + 0xA4093822, 0x299F31D0, 0x082EFA98, 0xEC4E6C89, + 0x452821E6, 0x38D01377, 0xBE5466CF, 0x34E90C6C, + 0xC0AC29B7, 0xC97C50DD, 0x3F84D5B5, 0xB5470917 +}; - uint32_t m[16]; - m[0] = d_data[8]; - m[1] = d_data[9]; - m[2] = d_data[10]; - m[3] = nonce; +__device__ __forceinline__ +uint32_t blake256_compress_14(uint32_t *m, uint32_t *v_init, uint32_t d_data6, uint32_t d_data7) +{ + uint32_t v[16]; #pragma unroll - for (uint32_t i = 4; i < 16; i++) { - m[i] = d_data[i+8U]; - } + for (uint32_t i = 0; i < 16; i++) v[i] = v_init[i]; + // these two are not modified: + v[ 9] = 0x85A308D3; + v[13] = 0x299F31D0 ^ (180U*8U); - // round 1 - GSPREC(0, 4, 0x8, 0xC, 0, 1); + // round 1 with nonce GSPREC(1, 5, 0x9, 0xD, 2, 3); - GSPREC(2, 6, 0xA, 0xE, 4, 5); - GSPREC(3, 7, 0xB, 0xF, 6, 7); GSPREC(0, 5, 0xA, 0xF, 8, 9); GSPREC(1, 6, 0xB, 0xC, 10, 11); GSPREC(2, 7, 0x8, 0xD, 12, 13); GSPREC(3, 4, 0x9, 0xE, 14, 15); // round 2 - GSPREC(0, 4, 0x8, 0xC, 14, 10); - GSPREC(1, 5, 0x9, 0xD, 4, 8); - GSPREC(2, 6, 0xA, 0xE, 9, 15); - GSPREC(3, 7, 0xB, 0xF, 13, 6); - GSPREC(0, 5, 0xA, 0xF, 1, 12); - GSPREC(1, 6, 0xB, 0xC, 0, 2); - GSPREC(2, 7, 0x8, 0xD, 11, 7); - GSPREC(3, 4, 0x9, 0xE, 5, 3); + GSPREC4(0, 4, 0x8, 0xC, 14, 10, 1, 5, 0x9, 0xD, 4, 8, 2, 6, 0xA, 0xE, 9, 15, 3, 7, 0xB, 0xF, 13, 6); + GSPREC4(0, 5, 0xA, 0xF, 1, 12, 1, 6, 0xB, 0xC, 0, 2, 2, 7, 0x8, 0xD, 11, 7, 3, 4, 0x9, 0xE, 5, 3); // round 3 - GSPREC(0, 4, 0x8, 0xC, 11, 8); - GSPREC(1, 5, 0x9, 0xD, 12, 0); - GSPREC(2, 6, 0xA, 0xE, 5, 2); - GSPREC(3, 7, 0xB, 0xF, 15, 13); - GSPREC(0, 5, 0xA, 0xF, 10, 14); - GSPREC(1, 6, 0xB, 0xC, 3, 6); - GSPREC(2, 7, 0x8, 0xD, 7, 1); - GSPREC(3, 4, 0x9, 0xE, 9, 4); + GSPREC4(0, 4, 0x8, 0xC, 11, 8, 1, 5, 0x9, 0xD, 12, 0, 2, 6, 0xA, 0xE, 5, 2, 3, 7, 0xB, 0xF, 15, 13); + GSPREC4(0, 5, 0xA, 0xF, 10, 14, 1, 6, 0xB, 0xC, 3, 6, 2, 7, 0x8, 0xD, 7, 1, 3, 4, 0x9, 0xE, 9, 4); // round 4 - GSPREC(0, 4, 0x8, 0xC, 7, 9); - GSPREC(1, 5, 0x9, 0xD, 3, 1); - GSPREC(2, 6, 0xA, 0xE, 13, 12); - GSPREC(3, 7, 0xB, 0xF, 11, 14); - GSPREC(0, 5, 0xA, 0xF, 2, 6); - GSPREC(1, 6, 0xB, 0xC, 5, 10); - GSPREC(2, 7, 0x8, 0xD, 4, 0); - GSPREC(3, 4, 0x9, 0xE, 15, 8); + GSPREC4(0, 4, 0x8, 0xC, 7, 9, 1, 5, 0x9, 0xD, 3, 1, 2, 6, 0xA, 0xE, 13, 12, 3, 7, 0xB, 0xF, 11, 14); + GSPREC4(0, 5, 0xA, 0xF, 2, 6, 1, 6, 0xB, 0xC, 5, 10, 2, 7, 0x8, 0xD, 4, 0, 3, 4, 0x9, 0xE, 15, 8); // round 5 - GSPREC(0, 4, 0x8, 0xC, 9, 0); - GSPREC(1, 5, 0x9, 0xD, 5, 7); - GSPREC(2, 6, 0xA, 0xE, 2, 4); - GSPREC(3, 7, 0xB, 0xF, 10, 15); - GSPREC(0, 5, 0xA, 0xF, 14, 1); - GSPREC(1, 6, 0xB, 0xC, 11, 12); - GSPREC(2, 7, 0x8, 0xD, 6, 8); - GSPREC(3, 4, 0x9, 0xE, 3, 13); + GSPREC4(0, 4, 0x8, 0xC, 9, 0, 1, 5, 0x9, 0xD, 5, 7, 2, 6, 0xA, 0xE, 2, 4, 3, 7, 0xB, 0xF, 10, 15); + GSPREC4(0, 5, 0xA, 0xF, 14, 1, 1, 6, 0xB, 0xC, 11, 12, 2, 7, 0x8, 0xD, 6, 8, 3, 4, 0x9, 0xE, 3, 13); // round 6 - GSPREC(0, 4, 0x8, 0xC, 2, 12); - GSPREC(1, 5, 0x9, 0xD, 6, 10); - GSPREC(2, 6, 0xA, 0xE, 0, 11); - GSPREC(3, 7, 0xB, 0xF, 8, 3); - GSPREC(0, 5, 0xA, 0xF, 4, 13); - GSPREC(1, 6, 0xB, 0xC, 7, 5); - GSPREC(2, 7, 0x8, 0xD, 15,14); - GSPREC(3, 4, 0x9, 0xE, 1, 9); + GSPREC4(0, 4, 0x8, 0xC, 2, 12, 1, 5, 0x9, 0xD, 6, 10, 2, 6, 0xA, 0xE, 0, 11, 3, 7, 0xB, 0xF, 8, 3); + GSPREC4(0, 5, 0xA, 0xF, 4, 13, 1, 6, 0xB, 0xC, 7, 5, 2, 7, 0x8, 0xD, 15,14, 3, 4, 0x9, 0xE, 1, 9); // round 7 - GSPREC(0, 4, 0x8, 0xC, 12, 5); - GSPREC(1, 5, 0x9, 0xD, 1, 15); - GSPREC(2, 6, 0xA, 0xE, 14,13); - GSPREC(3, 7, 0xB, 0xF, 4, 10); - GSPREC(0, 5, 0xA, 0xF, 0, 7); - GSPREC(1, 6, 0xB, 0xC, 6, 3); - GSPREC(2, 7, 0x8, 0xD, 9, 2); - GSPREC(3, 4, 0x9, 0xE, 8, 11); + GSPREC4(0, 4, 0x8, 0xC, 12, 5, 1, 5, 0x9, 0xD, 1, 15, 2, 6, 0xA, 0xE, 14,13, 3, 7, 0xB, 0xF, 4, 10); + GSPREC4(0, 5, 0xA, 0xF, 0, 7, 1, 6, 0xB, 0xC, 6, 3, 2, 7, 0x8, 0xD, 9, 2, 3, 4, 0x9, 0xE, 8, 11); + /* + // round 8 + GSPREC4(0, 4, 0x8, 0xC, 13,11, 1, 5, 0x9, 0xD, 7, 14, 2, 6, 0xA, 0xE, 12, 1, 3, 7, 0xB, 0xF, 3, 9); + GSPREC4(0, 5, 0xA, 0xF, 5, 0, 1, 6, 0xB, 0xC, 15, 4, 2, 7, 0x8, 0xD, 8, 6, 3, 4, 0x9, 0xE, 2, 10); + // round 9 + GSPREC4(0, 4, 0x8, 0xC, 6, 15, 1, 5, 0x9, 0xD, 14, 9, 2, 6, 0xA, 0xE, 11, 3, 3, 7, 0xB, 0xF, 0, 8); + GSPREC4(0, 5, 0xA, 0xF, 12, 2, 1, 6, 0xB, 0xC, 13, 7, 2, 7, 0x8, 0xD, 1, 4, 3, 4, 0x9, 0xE, 10, 5); + // round 10 + GSPREC4(0, 4, 0x8, 0xC, 10, 2, 1, 5, 0x9, 0xD, 8, 4, 2, 6, 0xA, 0xE, 7, 6, 3, 7, 0xB, 0xF, 1, 5); + GSPREC4(0, 5, 0xA, 0xF, 15,11, 1, 6, 0xB, 0xC, 9, 14, 2, 7, 0x8, 0xD, 3, 12, 3, 4, 0x9, 0xE, 13, 0); + // round 11 + GSPREC4(0, 4, 0x8, 0xC, 0, 1, 1, 5, 0x9, 0xD, 2, 3, 2, 6, 0xA, 0xE, 4, 5, 3, 7, 0xB, 0xF, 6, 7); + GSPREC4(0, 5, 0xA, 0xF, 8, 9, 1, 6, 0xB, 0xC, 10,11, 2, 7, 0x8, 0xD, 12,13, 3, 4, 0x9, 0xE, 14,15); + // round 12 + GSPREC4(0, 4, 0x8, 0xC, 14,10, 1, 5, 0x9, 0xD, 4, 8, 2, 6, 0xA, 0xE, 9, 15, 3, 7, 0xB, 0xF, 13, 6); + GSPREC4(0, 5, 0xA, 0xF, 1, 12, 1, 6, 0xB, 0xC, 0, 2, 2, 7, 0x8, 0xD, 11, 7, 3, 4, 0x9, 0xE, 5, 3); + // round 13 + GSPREC4(0, 4, 0x8, 0xC, 11, 8, 1, 5, 0x9, 0xD, 12, 0, 2, 6, 0xA, 0xE, 5, 2, 3, 7, 0xB, 0xF, 15,13); + GSPREC4(0, 5, 0xA, 0xF, 10,14, 1, 6, 0xB, 0xC, 3, 6, 2, 7, 0x8, 0xD, 7, 1, 3, 4, 0x9, 0xE, 9, 4); + */ // round 8 GSPREC(0, 4, 0x8, 0xC, 13,11); GSPREC(1, 5, 0x9, 0xD, 7, 14); @@ -220,68 +224,81 @@ void blake256_compress_14(uint32_t *h, const uint32_t nonce, const uint32_t T0) GSPREC(2, 6, 0xA, 0xE, 13,12); GSPREC(3, 7, 0xB, 0xF, 11,14); GSPREC(0, 5, 0xA, 0xF, 2, 6); - GSPREC(1, 6, 0xB, 0xC, 5, 10); GSPREC(2, 7, 0x8, 0xD, 4, 0); - //GSPREC(3, 4, 0x9, 0xE, 15, 8); - - v[3] += (m[15] ^ c_u256[8]) + v[4]; - v[14] = __byte_perm(v[14] ^ v[3], 0, 0x1032); - v[9] += v[14]; \ - v[4] = SPH_ROTR32(v[4] ^ v[9], 12); - v[3] += (m[8] ^ c_u256[15]) + v[4]; - v[14] = __byte_perm(v[14] ^ v[3], 0, 0x0321); - - // only compute h6 & 7 - h[6] ^= v[6] ^ v[14]; - h[7] ^= v[7] ^ v[15]; + + if ((d_data7 ^ v[7] ^ v[15]) == 0) { + GSPREC(1, 6, 0xB, 0xC, 5, 10); + GSPREC(3, 4, 0x9, 0xE, 15, 8); + return (d_data6 ^ v[6] ^ v[14]); + } + return UINT32_MAX; } + /* ############################################################################################################################### */ +// ------ Close: Last 52/64 bytes ------ __global__ void blake256_gpu_hash_nonce(const uint32_t threads, const uint32_t startNonce, uint32_t *resNonce, const uint64_t highTarget) { uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); - if (thread < threads) +// if (thread < threads) { - const uint32_t nonce = startNonce + thread; - uint32_t h[8]; + const uint32_t nonce = startNonce + thread * NPR; + uint32_t m[16], v[16], temp; + const uint32_t d_data6 = d_data[6], d_data7 = d_data[7]; #pragma unroll - for(int i=0; i < 8; i++) { - h[i] = d_data[i]; - } - - // ------ Close: Last 52/64 bytes ------ - - blake256_compress_14(h, nonce, (180U*8U)); + for(int i = 0; i < 8; i++) v[i] = d_data[i]; - if (h[7] == 0 && cuda_swab32(h[6]) <= highTarget) { -#if NBN == 2 - if (resNonce[0] != UINT32_MAX) - resNonce[1] = nonce; - else - resNonce[0] = nonce; -#else - resNonce[0] = nonce; -#endif + #pragma unroll + for (uint32_t i = 0; i < 16; i++) m[i] = d_data[i+8U]; + + v[ 8] = 0x243F6A88; + v[ 9] = 0x85A308D3; + v[10] = 0x13198A2E; + v[11] = 0x03707344; + + v[12] = 0xA4093822 ^ (180U*8U); + v[13] = 0x299F31D0 ^ (180U*8U); + v[14] = 0x082EFA98; + v[15] = 0xEC4E6C89; + + // round 1 without nonce + GSPREC(0, 4, 0x8, 0xC, 0, 1); + GSPREC(2, 6, 0xA, 0xE, 4, 5); + GSPREC(3, 7, 0xB, 0xF, 6, 7); + + for (m[3] = nonce; m[3] < nonce + NPR; m[3]++) { + temp = blake256_compress_14(m, v, d_data6, d_data7); + + if (temp != UINT32_MAX && cuda_swab32(temp) <= highTarget) { + #if NBN == 2 + if (resNonce[0] != UINT32_MAX) resNonce[1] = m[3]; + else resNonce[0] = m[3]; + #else + resNonce[0] = m[3]; + #endif + } } } } + __host__ static uint32_t decred_cpu_hash_nonce(const int thr_id, const uint32_t threads, const uint32_t startNonce, const uint64_t highTarget) { uint32_t result = UINT32_MAX; + const uint32_t real_threads = threads / NPR; - dim3 grid((threads + TPB-1)/TPB); + dim3 grid((real_threads + TPB-1)/TPB); dim3 block(TPB); /* Check error on Ctrl+C or kill to prevent segfaults on exit */ if (cudaMemset(d_resNonce[thr_id], 0xff, NBN*sizeof(uint32_t)) != cudaSuccess) return result; - blake256_gpu_hash_nonce <<>> (threads, startNonce, d_resNonce[thr_id], highTarget); + blake256_gpu_hash_nonce <<>> (real_threads, startNonce, d_resNonce[thr_id], highTarget); cudaThreadSynchronize(); if (cudaSuccess == cudaMemcpy(h_resNonce[thr_id], d_resNonce[thr_id], NBN*sizeof(uint32_t), cudaMemcpyDeviceToHost)) { @@ -294,6 +311,7 @@ static uint32_t decred_cpu_hash_nonce(const int thr_id, const uint32_t threads, return result; } + __host__ static void decred_midstate_128(uint32_t *output, const uint32_t *input) { @@ -307,6 +325,7 @@ static void decred_midstate_128(uint32_t *output, const uint32_t *input) memcpy(output, (void*)ctx.H, 32); } + __host__ void decred_cpu_setBlock_52(uint32_t *penddata, const uint32_t *midstate, const uint32_t *ptarget) { @@ -321,6 +340,7 @@ void decred_cpu_setBlock_52(uint32_t *penddata, const uint32_t *midstate, const CUDA_SAFE_CALL(cudaMemcpyToSymbol(d_data, data, 32 + 64, 0, cudaMemcpyHostToDevice)); } + /* ############################################################################################################################### */ static bool init[MAX_GPUS] = { 0 }; @@ -425,6 +445,7 @@ extern "C" int scanhash_decred(int thr_id, struct work* work, uint32_t max_nonce return rc; } + // cleanup extern "C" void free_decred(int thr_id) { @@ -440,4 +461,3 @@ extern "C" void free_decred(int thr_id) cudaDeviceSynchronize(); } -