From bf17f34001ecdf232f49d7b777d74dad4f0cd1ae Mon Sep 17 00:00:00 2001 From: Tanguy Pruvot Date: Sun, 17 Jul 2016 15:21:57 +0200 Subject: [PATCH] preview 3, with alexis78 touch --- lbry/cuda_sha256_lbry.cu | 408 ++++++++++++++++++--------------------- lbry/cuda_sha512_lbry.cu | 4 +- lbry/lbry.cu | 31 ++- 3 files changed, 197 insertions(+), 246 deletions(-) diff --git a/lbry/cuda_sha256_lbry.cu b/lbry/cuda_sha256_lbry.cu index 52cddcd..1af73f4 100644 --- a/lbry/cuda_sha256_lbry.cu +++ b/lbry/cuda_sha256_lbry.cu @@ -1,24 +1,38 @@ /* - * sha256 CUDA implementation. + * sha256 + ripemd CUDA implementation. + * tpruvot and alexis78 */ + #include #include #include #include +#include + #include -__constant__ static uint32_t __align__(8) c_midstate112[8]; -__constant__ static uint32_t __align__(8) c_dataEnd112[12]; +__constant__ static uint32_t _ALIGN(8) c_midstate112[8]; +__constant__ static uint32_t _ALIGN(8) c_midbuffer112[8]; +__constant__ static uint32_t _ALIGN(8) c_dataEnd112[12]; -const __constant__ uint32_t __align__(8) c_H256[8] = { +__constant__ const uint32_t c_H256[8] = { 0x6A09E667U, 0xBB67AE85U, 0x3C6EF372U, 0xA54FF53AU, 0x510E527FU, 0x9B05688CU, 0x1F83D9ABU, 0x5BE0CD19U }; -__constant__ static uint32_t __align__(8) c_K[64]; +__constant__ static uint32_t _ALIGN(16) c_K[64] = { + 0x428A2F98, 0x71374491, 0xB5C0FBCF, 0xE9B5DBA5, 0x3956C25B, 0x59F111F1, 0x923F82A4, 0xAB1C5ED5, + 0xD807AA98, 0x12835B01, 0x243185BE, 0x550C7DC3, 0x72BE5D74, 0x80DEB1FE, 0x9BDC06A7, 0xC19BF174, + 0xE49B69C1, 0xEFBE4786, 0x0FC19DC6, 0x240CA1CC, 0x2DE92C6F, 0x4A7484AA, 0x5CB0A9DC, 0x76F988DA, + 0x983E5152, 0xA831C66D, 0xB00327C8, 0xBF597FC7, 0xC6E00BF3, 0xD5A79147, 0x06CA6351, 0x14292967, + 0x27B70A85, 0x2E1B2138, 0x4D2C6DFC, 0x53380D13, 0x650A7354, 0x766A0ABB, 0x81C2C92E, 0x92722C85, + 0xA2BFE8A1, 0xA81A664B, 0xC24B8B70, 0xC76C51A3, 0xD192E819, 0xD6990624, 0xF40E3585, 0x106AA070, + 0x19A4C116, 0x1E376C08, 0x2748774C, 0x34B0BCB5, 0x391C0CB3, 0x4ED8AA4A, 0x5B9CCA4F, 0x682E6FF3, + 0x748F82EE, 0x78A5636F, 0x84C87814, 0x8CC70208, 0x90BEFFFA, 0xA4506CEB, 0xBEF9A3F7, 0xC67178F2 +}; static __thread uint32_t* d_resNonces; -__constant__ static uint32_t __align__(8) c_target[2]; +__constant__ static uint32_t _ALIGN(8) c_target[2]; __device__ uint64_t d_target[1]; #ifdef __INTELLISENSE__ @@ -43,32 +57,24 @@ static const uint32_t cpu_K[64] = { 0x748F82EE, 0x78A5636F, 0x84C87814, 0x8CC70208, 0x90BEFFFA, 0xA4506CEB, 0xBEF9A3F7, 0xC67178F2 }; -#define ROTR ROTR32 - __host__ static void sha256_step1_host(uint32_t a, uint32_t b, uint32_t c, uint32_t &d, - uint32_t e, uint32_t f, uint32_t g, uint32_t &h, - uint32_t in, const uint32_t Kshared) + uint32_t e, uint32_t f, uint32_t g, uint32_t &h, uint32_t in, const uint32_t Kshared) { - uint32_t t1,t2; uint32_t vxandx = (((f) ^ (g)) & (e)) ^ (g); // xandx(e, f, g); - uint32_t bsg21 = ROTR(e, 6) ^ ROTR(e, 11) ^ ROTR(e, 25); // bsg2_1(e); - uint32_t bsg20 = ROTR(a, 2) ^ ROTR(a, 13) ^ ROTR(a, 22); //bsg2_0(a); + uint32_t bsg21 = ROTR32(e, 6) ^ ROTR32(e, 11) ^ ROTR32(e, 25); // bsg2_1(e); + uint32_t bsg20 = ROTR32(a, 2) ^ ROTR32(a, 13) ^ ROTR32(a, 22); //bsg2_0(a); uint32_t andorv = ((b) & (c)) | (((b) | (c)) & (a)); //andor32(a,b,c); - - t1 = h + bsg21 + vxandx + Kshared + in; - t2 = bsg20 + andorv; + uint32_t t1 = h + bsg21 + vxandx + Kshared + in; + uint32_t t2 = bsg20 + andorv; d = d + t1; h = t1 + t2; } __host__ static void sha256_step2_host(uint32_t a, uint32_t b, uint32_t c, uint32_t &d, - uint32_t e, uint32_t f, uint32_t g, uint32_t &h, - uint32_t* in, uint32_t pc, const uint32_t Kshared) + uint32_t e, uint32_t f, uint32_t g, uint32_t &h, uint32_t* in, uint32_t pc, const uint32_t Kshared) { - uint32_t t1,t2; - int pcidx1 = (pc-2) & 0xF; int pcidx2 = (pc-7) & 0xF; int pcidx3 = (pc-15) & 0xF; @@ -78,12 +84,13 @@ static void sha256_step2_host(uint32_t a, uint32_t b, uint32_t c, uint32_t &d, uint32_t inx2 = in[pcidx2]; uint32_t inx3 = in[pcidx3]; - uint32_t ssg21 = ROTR(inx1, 17) ^ ROTR(inx1, 19) ^ SPH_T32((inx1) >> 10); //ssg2_1(inx1); - uint32_t ssg20 = ROTR(inx3, 7) ^ ROTR(inx3, 18) ^ SPH_T32((inx3) >> 3); //ssg2_0(inx3); + uint32_t ssg21 = ROTR32(inx1, 17) ^ ROTR32(inx1, 19) ^ SPH_T32((inx1) >> 10); //ssg2_1(inx1); + uint32_t ssg20 = ROTR32(inx3, 7) ^ ROTR32(inx3, 18) ^ SPH_T32((inx3) >> 3); //ssg2_0(inx3); uint32_t vxandx = (((f) ^ (g)) & (e)) ^ (g); // xandx(e, f, g); - uint32_t bsg21 = ROTR(e, 6) ^ ROTR(e, 11) ^ ROTR(e, 25); // bsg2_1(e); - uint32_t bsg20 = ROTR(a, 2) ^ ROTR(a, 13) ^ ROTR(a, 22); //bsg2_0(a); + uint32_t bsg21 = ROTR32(e, 6) ^ ROTR32(e, 11) ^ ROTR32(e, 25); // bsg2_1(e); + uint32_t bsg20 = ROTR32(a, 2) ^ ROTR32(a, 13) ^ ROTR32(a, 22); //bsg2_0(a); uint32_t andorv = ((b) & (c)) | (((b) | (c)) & (a)); //andor32(a,b,c); + uint32_t t1,t2; in[pc] = ssg21 + inx2 + ssg20 + inx0; @@ -152,55 +159,32 @@ static void sha256_round_body_host(uint32_t* in, uint32_t* state, const uint32_t state[7] += h; } -__device__ __forceinline__ -uint32_t xor3b(const uint32_t a, const uint32_t b, const uint32_t c) { - uint32_t result; -#if __CUDA_ARCH__ >= 500 && CUDA_VERSION >= 7050 - asm ("lop3.b32 %0, %1, %2, %3, 0x96; // xor3b" //0x96 = 0xF0 ^ 0xCC ^ 0xAA - : "=r"(result) : "r"(a), "r"(b),"r"(c)); -#else - result = a^b^c; -#endif - return result; -} +#define xor3b(a,b,c) ((a ^ b) ^ c) __device__ __forceinline__ uint32_t bsg2_0(const uint32_t x) { - uint32_t r1 = ROTR32(x,2); - uint32_t r2 = ROTR32(x,13); - uint32_t r3 = ROTR32(x,22); - return xor3b(r1,r2,r3); + return xor3b(ROTR32(x,2),ROTR32(x,13),ROTR32(x,22)); } __device__ __forceinline__ uint32_t bsg2_1(const uint32_t x) { - uint32_t r1 = ROTR32(x,6); - uint32_t r2 = ROTR32(x,11); - uint32_t r3 = ROTR32(x,25); - return xor3b(r1,r2,r3); + return xor3b(ROTR32(x,6),ROTR32(x,11),ROTR32(x,25)); } __device__ __forceinline__ uint32_t ssg2_0(const uint32_t x) { - uint64_t r1 = ROTR32(x,7); - uint64_t r2 = ROTR32(x,18); - uint64_t r3 = shr_t32(x,3); - return xor3b(r1,r2,r3); + return xor3b(ROTR32(x,7),ROTR32(x,18),(x>>3)); } __device__ __forceinline__ uint32_t ssg2_1(const uint32_t x) { - uint64_t r1 = ROTR32(x,17); - uint64_t r2 = ROTR32(x,19); - uint64_t r3 = shr_t32(x,10); - return xor3b(r1,r2,r3); + return xor3b(ROTR32(x,17),ROTR32(x,19),(x>>10)); } __device__ __forceinline__ uint32_t andor32(const uint32_t a, const uint32_t b, const uint32_t c) { uint32_t result; - asm("{\n\t" - ".reg .u32 m,n,o;\n\t" + asm("{ .reg .u32 m,n,o; // andor32 \n\t" "and.b32 m, %1, %2;\n\t" " or.b32 n, %1, %2;\n\t" "and.b32 o, n, %3;\n\t" @@ -210,27 +194,21 @@ __device__ __forceinline__ uint32_t andor32(const uint32_t a, const uint32_t b, return result; } -__device__ __forceinline__ uint2 vectorizeswap(uint64_t v) { +__device__ __forceinline__ uint2 vectorizeswap(uint64_t v) +{ uint2 result; - asm("mov.b64 {%0,%1},%2; \n\t" + asm("mov.b64 {%0,%1},%2; // vectorizeswap \n\t" : "=r"(result.y), "=r"(result.x) : "l"(v)); return result; } __device__ -static void sha2_step1(uint32_t a, uint32_t b, uint32_t c, uint32_t &d, uint32_t e, uint32_t f, uint32_t g, uint32_t &h, - uint32_t in, const uint32_t Kshared) +__forceinline__ +static void sha2_step1(uint32_t a, uint32_t b, uint32_t c, uint32_t &d, uint32_t e, uint32_t f, uint32_t g, uint32_t &h, uint32_t in, const uint32_t Kshared) { - uint32_t t1,t2; - uint32_t vxandx = xandx(e, f, g); - uint32_t bsg21 = bsg2_1(e); - uint32_t bsg20 = bsg2_0(a); - uint32_t andorv = andor32(a,b,c); - - t1 = h + bsg21 + vxandx + Kshared + in; - t2 = bsg20 + andorv; - d = d + t1; - h = t1 + t2; + uint32_t t1 = bsg2_1(e) + ((((f) ^ (g)) & (e)) ^ (g)) + Kshared + in; + d = d + h + t1; + h += t1 + bsg2_0(a) + (((b) & (c)) | (((b) | (c)) & (a))); } __device__ @@ -263,6 +241,56 @@ static void sha2_step2(uint32_t a, uint32_t b, uint32_t c, uint32_t &d, uint32_t h = t1 + t2; } +__device__ __forceinline__ +static void sha256_round_first(uint32_t* in,uint32_t *buf, uint32_t* state, uint32_t* const Kshared) +{ + uint32_t a = buf[0]; + uint32_t b = buf[1]; + uint32_t c = buf[2]; + uint32_t d = buf[3]; + uint32_t e = buf[4]; + uint32_t f = buf[5]; + uint32_t g = buf[6]; + uint32_t h = buf[7]; + + // 10 first steps made on host + sha2_step1(f,g,h,a,b,c,d,e,in[11],Kshared[11]); + sha2_step1(e,f,g,h,a,b,c,d,in[12],Kshared[12]); + sha2_step1(d,e,f,g,h,a,b,c,in[13],Kshared[13]); + sha2_step1(c,d,e,f,g,h,a,b,in[14],Kshared[14]); + sha2_step1(b,c,d,e,f,g,h,a,in[15],Kshared[15]); + + #pragma unroll + for (int i=0; i<3; i++) + { + sha2_step2(a,b,c,d,e,f,g,h,in,0, Kshared[16+16*i]); + sha2_step2(h,a,b,c,d,e,f,g,in,1, Kshared[17+16*i]); + sha2_step2(g,h,a,b,c,d,e,f,in,2, Kshared[18+16*i]); + sha2_step2(f,g,h,a,b,c,d,e,in,3, Kshared[19+16*i]); + sha2_step2(e,f,g,h,a,b,c,d,in,4, Kshared[20+16*i]); + sha2_step2(d,e,f,g,h,a,b,c,in,5, Kshared[21+16*i]); + sha2_step2(c,d,e,f,g,h,a,b,in,6, Kshared[22+16*i]); + sha2_step2(b,c,d,e,f,g,h,a,in,7, Kshared[23+16*i]); + sha2_step2(a,b,c,d,e,f,g,h,in,8, Kshared[24+16*i]); + sha2_step2(h,a,b,c,d,e,f,g,in,9, Kshared[25+16*i]); + sha2_step2(g,h,a,b,c,d,e,f,in,10,Kshared[26+16*i]); + sha2_step2(f,g,h,a,b,c,d,e,in,11,Kshared[27+16*i]); + sha2_step2(e,f,g,h,a,b,c,d,in,12,Kshared[28+16*i]); + sha2_step2(d,e,f,g,h,a,b,c,in,13,Kshared[29+16*i]); + sha2_step2(c,d,e,f,g,h,a,b,in,14,Kshared[30+16*i]); + sha2_step2(b,c,d,e,f,g,h,a,in,15,Kshared[31+16*i]); + } + + buf[ 0] = state[0] + a; + buf[ 1] = state[1] + b; + buf[ 2] = state[2] + c; + buf[ 3] = state[3] + d; + buf[ 4] = state[4] + e; + buf[ 5] = state[5] + f; + buf[ 6] = state[6] + g; + buf[ 7] = state[7] + h; +} + __device__ static void sha256_round_body(uint32_t* in, uint32_t* state, uint32_t* const Kshared) { @@ -328,166 +356,70 @@ uint64_t cuda_swab32ll(uint64_t x) { return MAKE_ULONGLONG(cuda_swab32(_LODWORD(x)), cuda_swab32(_HIDWORD(x))); } -#if 0 -__global__ -void lbry_sha256_gpu_hash_112(const uint32_t threads, const uint32_t startNonce, const bool swabNonce, uint64_t *outputHash) -{ - const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); - if (thread < threads) - { - const uint32_t nonce = startNonce + thread; - - uint32_t dat[16]; - #pragma unroll - for (int i=0;i<11;i++) dat[i] = c_dataEnd112[i]; // pre "swabed" - dat[11] = swabNonce ? cuda_swab32(nonce) : nonce; - dat[12] = 0x80000000; - dat[13] = 0; - dat[14] = 0; - dat[15] = 0x380; - - uint32_t __align__(8) buf[8]; - #pragma unroll - for (int i=0;i<8;i++) buf[i] = c_midstate112[i]; - - sha256_round_body(dat, buf, c_K); - - // output - uint2* output = (uint2*) (&outputHash[thread * 8U]); - #pragma unroll - for (int i=0;i<4;i++) { - //output[i] = vectorize(cuda_swab32ll(((uint64_t*)buf)[i])); - output[i] = vectorize(((uint64_t*)buf)[i]); // out without swap, new sha256 after - } - } -} - -__global__ -void lbry_sha256_gpu_hash_32(uint32_t threads, uint64_t *Hash512) -{ - const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); - if (thread < threads) - { - uint32_t __align__(8) buf[8]; // align for vectorize - #pragma unroll - for (int i=0; i<8; i++) buf[i] = c_H256[i]; - - uint32_t* input = (uint32_t*) (&Hash512[thread * 8U]); - - uint32_t dat[16]; - #pragma unroll - //for (int i=0;i<8;i++) dat[i] = cuda_swab32(input[i]); - for (int i=0; i<8; i++) dat[i] = input[i]; - dat[8] = 0x80000000; - #pragma unroll - for (int i=9; i<15; i++) dat[i] = 0; - dat[15] = 0x100; - - sha256_round_body(dat, buf, c_K); - - // output - uint2* output = (uint2*) input; - #pragma unroll - for (int i=0;i<4;i++) { - //output[i] = vectorize(cuda_swab32ll(((uint64_t*)buf)[i])); - output[i] = vectorizeswap(((uint64_t*)buf)[i]); - } -#ifdef PAD_ZEROS - #pragma unroll - for (int i=4; i<8; i++) output[i] = vectorize(0); -#endif - } -} - -__host__ -void lbry_sha256_hash_112(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_outputHash, bool swabNonce, cudaStream_t stream) -{ - const int threadsperblock = 256; - - dim3 grid(threads/threadsperblock); - dim3 block(threadsperblock); - - lbry_sha256_gpu_hash_112 <<>> (threads, startNonce, swabNonce, (uint64_t*) d_outputHash); - cudaGetLastError(); -} - -__host__ -void lbry_sha256_hash_32(int thr_id, uint32_t threads, uint32_t *d_Hash, cudaStream_t stream) -{ - const int threadsperblock = 256; - - dim3 grid(threads/threadsperblock); - dim3 block(threadsperblock); - - lbry_sha256_gpu_hash_32 <<>> (threads, (uint64_t*) d_Hash); -} -#endif - __global__ -void lbry_sha256d_gpu_hash_112(const uint32_t threads, const uint32_t startNonce, const bool swabNonce, uint64_t *outputHash) +__launch_bounds__(512,2) /* to force 64 regs */ +void lbry_sha256d_gpu_hash_112(const uint32_t threads, const uint32_t startNonce, uint64_t *outputHash) { const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); extern __shared__ uint32_t s_K[]; - //s_K[thread & 63] = c_K[thread & 63]; if (threadIdx.x < 64U) s_K[threadIdx.x] = c_K[threadIdx.x]; + //__threadfence_block(); if (thread < threads) { - const uint32_t nonce = startNonce + thread; - uint32_t dat[16]; #pragma unroll for (int i=0; i<11; i++) dat[i] = c_dataEnd112[i]; - dat[11] = swabNonce ? cuda_swab32(nonce) : nonce; + dat[11] = startNonce + thread; dat[12] = 0x80000000; dat[13] = 0; dat[14] = 0; dat[15] = 0x380; - uint32_t __align__(8) buf[8]; - #pragma unroll - for (int i=0;i<8;i++) buf[i] = c_midstate112[i]; + uint32_t __align__(8) buf[8], state[8]; - sha256_round_body(dat, buf, s_K); + *(uint2x4*)&state[0] = *(uint2x4*)&c_midstate112[0]; + *(uint2x4*)&buf[0] = *(uint2x4*)&c_midbuffer112[0]; + + sha256_round_first(dat, buf, state, c_K); // no shared mem here // second sha256 - #pragma unroll - for (int i=0; i<8; i++) dat[i] = buf[i]; + *(uint2x4*)&dat[0] = *(uint2x4*)&buf[0]; + dat[8] = 0x80000000; + #pragma unroll for (int i=9; i<15; i++) dat[i] = 0; dat[15] = 0x100; - #pragma unroll - for (int i=0; i<8; i++) buf[i] = c_H256[i]; + *(uint2x4*)&buf[0] = *(uint2x4*)&c_H256[0]; sha256_round_body(dat, buf, s_K); // output - uint2* output = (uint2*) (&outputHash[thread * 8U]); - #pragma unroll - for (int i=0;i<4;i++) { - // //output[i] = vectorize(cuda_swab32ll(((uint64_t*)buf)[i])); - output[i] = vectorizeswap(((uint64_t*)buf)[i]); - } + *(uint2*)&buf[0] = vectorizeswap(((uint64_t*)buf)[0]); + *(uint2*)&buf[2] = vectorizeswap(((uint64_t*)buf)[1]); + *(uint2*)&buf[4] = vectorizeswap(((uint64_t*)buf)[2]); + *(uint2*)&buf[6] = vectorizeswap(((uint64_t*)buf)[3]); + + *(uint2x4*)&outputHash[thread*8U] = *(uint2x4*)&buf[0]; } } __host__ -void lbry_sha256d_hash_112(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_outputHash, bool swabNonce, cudaStream_t stream) +void lbry_sha256d_hash_112(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_outputHash) { - const int threadsperblock = 256; + const int threadsperblock = 512; dim3 grid(threads/threadsperblock); dim3 block(threadsperblock); - lbry_sha256d_gpu_hash_112 <<>> (threads, startNonce, swabNonce, (uint64_t*) d_outputHash); + lbry_sha256d_gpu_hash_112 <<>> (threads, startNonce, (uint64_t*) d_outputHash); } __host__ void lbry_sha256_init(int thr_id) { - //cudaMemcpyToSymbol(c_H256, cpu_H256, sizeof(cpu_H256), 0, cudaMemcpyHostToDevice); cudaMemcpyToSymbol(c_K, cpu_K, sizeof(cpu_K), 0, cudaMemcpyHostToDevice); CUDA_SAFE_CALL(cudaMalloc(&d_resNonces, 4*sizeof(uint32_t))); } @@ -507,21 +439,53 @@ void lbry_sha256_setBlock_112(uint32_t *pdata, uint32_t *ptarget) for (int i=0;i<11;i++) end[i] = cuda_swab32(pdata[16+i]); sha256_round_body_host(in, buf, cpu_K); - CUDA_SAFE_CALL(cudaMemcpyToSymbol(c_midstate112, buf, 32, 0, cudaMemcpyHostToDevice)); - CUDA_SAFE_CALL(cudaMemcpyToSymbol(c_dataEnd112, end, sizeof(end), 0, cudaMemcpyHostToDevice)); - CUDA_SAFE_CALL(cudaMemcpyToSymbol(c_target, &ptarget[6], sizeof(uint64_t), 0, cudaMemcpyHostToDevice)); - CUDA_SAFE_CALL(cudaMemcpyToSymbol(d_target, &ptarget[6], sizeof(uint64_t), 0, cudaMemcpyHostToDevice)); + cudaMemcpyToSymbol(c_midstate112, buf, 32, 0, cudaMemcpyHostToDevice); + cudaMemcpyToSymbol(c_dataEnd112, end, sizeof(end), 0, cudaMemcpyHostToDevice); + + uint32_t a = buf[0]; + uint32_t b = buf[1]; + uint32_t c = buf[2]; + uint32_t d = buf[3]; + uint32_t e = buf[4]; + uint32_t f = buf[5]; + uint32_t g = buf[6]; + uint32_t h = buf[7]; + + sha256_step1_host(a,b,c,d,e,f,g,h,end[0], cpu_K[0]); + sha256_step1_host(h,a,b,c,d,e,f,g,end[1], cpu_K[1]); + sha256_step1_host(g,h,a,b,c,d,e,f,end[2], cpu_K[2]); + sha256_step1_host(f,g,h,a,b,c,d,e,end[3], cpu_K[3]); + sha256_step1_host(e,f,g,h,a,b,c,d,end[4], cpu_K[4]); + sha256_step1_host(d,e,f,g,h,a,b,c,end[5], cpu_K[5]); + sha256_step1_host(c,d,e,f,g,h,a,b,end[6], cpu_K[6]); + sha256_step1_host(b,c,d,e,f,g,h,a,end[7], cpu_K[7]); + sha256_step1_host(a,b,c,d,e,f,g,h,end[8], cpu_K[8]); + sha256_step1_host(h,a,b,c,d,e,f,g,end[9], cpu_K[9]); + sha256_step1_host(g,h,a,b,c,d,e,f,end[10],cpu_K[10]); + + buf[0] = a; + buf[1] = b; + buf[2] = c; + buf[3] = d; + buf[4] = e; + buf[5] = f; + buf[6] = g; + buf[7] = h; + + cudaMemcpyToSymbol(c_midbuffer112, buf, 32, 0, cudaMemcpyHostToDevice); + + cudaMemcpyToSymbol(c_target, &ptarget[6], sizeof(uint64_t), 0, cudaMemcpyHostToDevice); + cudaMemcpyToSymbol(d_target, &ptarget[6], sizeof(uint64_t), 0, cudaMemcpyHostToDevice); } // ------------------------------------------------------------------------------------------ - static __constant__ uint32_t c_IV[5] = { 0x67452301u, 0xEFCDAB89u, 0x98BADCFEu, 0x10325476u, 0xC3D2E1F0u }; /* - * Round functions for RIPEMD-128 and RIPEMD-160. + * Round functions for RIPEMD-160. */ #if 1 #define F1(x, y, z) ((x) ^ (y) ^ (z)) @@ -753,19 +717,21 @@ static __constant__ uint32_t c_IV[5] = { } __global__ -__launch_bounds__(256,3) -void lbry_sha256d_gpu_hash_final(const uint32_t threads, const uint32_t startNonce, uint64_t *Hash512, uint32_t *resNonces) +__launch_bounds__(640,2) /* 640,2 <= 48 regs, 512,2 <= 64 */ +void lbry_sha256d_gpu_hash_final(const uint32_t threads, uint64_t *Hash512, uint32_t *resNonces) { const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); + extern __shared__ uint32_t s_K[]; + if (threadIdx.x < 64U) s_K[threadIdx.x] = c_K[threadIdx.x]; + //__threadfence_block(); if (thread < threads) { uint32_t* input = (uint32_t*) (&Hash512[thread * 8U]); - uint32_t __align__(8) buf[8]; // align for vectorize - uint32_t dat[16]; - #pragma unroll - for (int i=0; i<8; i++) - dat[i] = (input[i]); + uint32_t __align__(8) dat[16]; + + *(uint2x4*)&dat[0] = *(uint2x4*)&input[0]; + dat[8] = 0x80; #pragma unroll @@ -780,15 +746,15 @@ void lbry_sha256d_gpu_hash_final(const uint32_t threads, const uint32_t startNon RIPEMD160_ROUND_BODY(dat, h); + uint32_t __align__(8) buf[8]; // align for vectorize #pragma unroll for (int i=0; i<5; i++) buf[i] = h[i]; // second 32 bytes block hash - #pragma unroll - for (int i=0; i<8; i++) - dat[i] = (input[8+i]); + *(uint2x4*)&dat[0] = *(uint2x4*)&input[8]; + dat[8] = 0x80; #pragma unroll @@ -810,56 +776,50 @@ void lbry_sha256d_gpu_hash_final(const uint32_t threads, const uint32_t startNon for (int i=0;i<5;i++) dat[i+5] = cuda_swab32(h[i]); dat[10] = 0x80000000; #pragma unroll - for (int i=11;i<15;i++) dat[i] = 0; + for (int i=11; i<15; i++) dat[i] = 0; dat[15] = 0x140; - #pragma unroll - for (int i=0;i<8;i++) buf[i] = c_H256[i]; + *(uint2x4*)&buf[0] = *(uint2x4*)&c_H256[0]; - sha256_round_body(dat, buf, c_K); + sha256_round_body(dat, buf, c_K); // s_K uses too much regs // second sha256 - #pragma unroll - for (int i=0;i<8;i++) dat[i] = buf[i]; + *(uint2x4*)&dat[0] = *(uint2x4*)&buf[0]; dat[8] = 0x80000000; #pragma unroll - for (int i=9;i<15;i++) dat[i] = 0; + for (int i=9; i<15; i++) dat[i] = 0; dat[15] = 0x100; - #pragma unroll - for (int i=0;i<8;i++) buf[i] = c_H256[i]; + *(uint2x4*)&buf[0] = *(uint2x4*)&c_H256[0]; - sha256_round_body(dat, buf, c_K); + sha256_round_body(dat, buf, s_K); // valid nonces - uint64_t high = cuda_swab32ll(((uint64_t*)buf)[3]); + const uint64_t high = cuda_swab32ll(((uint64_t*)buf)[3]); if (high <= d_target[0]) { - // printf("%08x %08x - %016llx %016llx - %08x %08x\n", buf[7], buf[6], high, d_target[0], c_target[1], c_target[0]); - uint32_t nonce = startNonce + thread; - resNonces[1] = atomicExch(resNonces, nonce); + resNonces[1] = atomicExch(resNonces, thread); d_target[0] = high; } } } __host__ -void lbry_sha256d_hash_final(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_inputHash, uint32_t *resNonces, cudaStream_t stream) +void lbry_sha256d_hash_final(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_inputHash, uint32_t *resNonces) { - const int threadsperblock = 256; + const int threadsperblock = 512; dim3 grid(threads/threadsperblock); dim3 block(threadsperblock); - CUDA_SAFE_CALL(cudaMemset(d_resNonces, 0xFF, 2 * sizeof(uint32_t))); - cudaThreadSynchronize(); + cudaMemset(d_resNonces, 0xFF, 2 * sizeof(uint32_t)); - lbry_sha256d_gpu_hash_final <<>> (threads, startNonce, (uint64_t*) d_inputHash, d_resNonces); + lbry_sha256d_gpu_hash_final <<>> (threads, (uint64_t*) d_inputHash, d_resNonces); - cudaThreadSynchronize(); - - CUDA_SAFE_CALL(cudaMemcpy(resNonces, d_resNonces, 2 * sizeof(uint32_t), cudaMemcpyDeviceToHost)); + cudaMemcpy(resNonces, d_resNonces, 2 * sizeof(uint32_t), cudaMemcpyDeviceToHost); if (resNonces[0] == resNonces[1]) { resNonces[1] = UINT32_MAX; } -} \ No newline at end of file + if (resNonces[0] != UINT32_MAX) resNonces[0] += startNonce; + if (resNonces[1] != UINT32_MAX) resNonces[1] += startNonce; +} diff --git a/lbry/cuda_sha512_lbry.cu b/lbry/cuda_sha512_lbry.cu index 79c0906..e4fdd25 100644 --- a/lbry/cuda_sha512_lbry.cu +++ b/lbry/cuda_sha512_lbry.cu @@ -161,7 +161,7 @@ void lbry_sha512_gpu_hash_32(const uint32_t threads, uint64_t *g_hash) } __host__ -void lbry_sha512_hash_32(int thr_id, uint32_t threads, uint32_t *d_hash, cudaStream_t stream) +void lbry_sha512_hash_32(int thr_id, uint32_t threads, uint32_t *d_hash) { const int threadsperblock = 256; @@ -169,7 +169,7 @@ void lbry_sha512_hash_32(int thr_id, uint32_t threads, uint32_t *d_hash, cudaStr dim3 block(threadsperblock); size_t shared_size = 0; - lbry_sha512_gpu_hash_32 <<>> (threads, (uint64_t*)d_hash); + lbry_sha512_gpu_hash_32 <<>> (threads, (uint64_t*)d_hash); } /**************************************************************************************************/ diff --git a/lbry/lbry.cu b/lbry/lbry.cu index ecf272b..b86d66a 100644 --- a/lbry/lbry.cu +++ b/lbry/lbry.cu @@ -65,12 +65,10 @@ extern "C" void lbry_hash(void* output, const void* input) extern void lbry_sha256_init(int thr_id); extern void lbry_sha256_free(int thr_id); extern void lbry_sha256_setBlock_112(uint32_t *pdata, uint32_t *ptarget); -extern void lbry_sha256_hash_112(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_outputHash, bool swabNonce, cudaStream_t stream); -extern void lbry_sha256d_hash_112(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_outputHash, bool swabNonce, cudaStream_t stream); -extern void lbry_sha256_hash_32(int thr_id, uint32_t threads, uint32_t *d_hash, cudaStream_t stream); +extern void lbry_sha256d_hash_112(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_outputHash); extern void lbry_sha512_init(int thr_id); -extern void lbry_sha512_hash_32(int thr_id, uint32_t threads, uint32_t *d_hash, cudaStream_t stream); -extern void lbry_sha256d_hash_final(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_inputHash, uint32_t *resNonces, cudaStream_t stream); +extern void lbry_sha512_hash_32(int thr_id, uint32_t threads, uint32_t *d_hash); +extern void lbry_sha256d_hash_final(int thr_id, uint32_t threads, uint32_t startNonce, uint32_t *d_inputHash, uint32_t *resNonces); static __inline uint32_t swab32_if(uint32_t val, bool iftrue) { return iftrue ? swab32(val) : val; @@ -91,7 +89,7 @@ extern "C" int scanhash_lbry(int thr_id, struct work *work, uint32_t max_nonce, uint32_t *ptarget = work->target; const uint32_t first_nonce = pdata[LBC_NONCE_OFT32]; - const int swap = 0; // to toggle nonce endian + const int swap = 0; // to toggle nonce endian (need kernel change) const int dev_id = device_map[thr_id]; int intensity = (device_sm[dev_id] > 500 && !is_windows()) ? 22 : 20; @@ -99,7 +97,7 @@ extern "C" int scanhash_lbry(int thr_id, struct work *work, uint32_t max_nonce, if (device_sm[dev_id] < 350) intensity = 18; uint32_t throughput = cuda_default_throughput(thr_id, 1U << intensity); - if (init[thr_id]) throughput = min(throughput, max_nonce - first_nonce); + //if (init[thr_id]) throughput = min(throughput, max_nonce - first_nonce); if (opt_benchmark) { ptarget[7] = 0xf; @@ -118,7 +116,6 @@ extern "C" int scanhash_lbry(int thr_id, struct work *work, uint32_t max_nonce, lbry_sha256_init(thr_id); lbry_sha512_init(thr_id); - cuda_check_cpu_init(thr_id, throughput); CUDA_LOG_ERROR(); init[thr_id] = true; @@ -129,23 +126,18 @@ extern "C" int scanhash_lbry(int thr_id, struct work *work, uint32_t max_nonce, } lbry_sha256_setBlock_112(endiandata, ptarget); - cuda_check_cpu_setTarget(ptarget); do { - // Hash with CUDA - #if 0 - lbry_sha256_hash_112(thr_id, throughput, pdata[LBC_NONCE_OFT32], d_hash[thr_id], swap, 0); - lbry_sha256_hash_32(thr_id, throughput, d_hash[thr_id], 0); - #else - lbry_sha256d_hash_112(thr_id, throughput, pdata[LBC_NONCE_OFT32], d_hash[thr_id], swap, 0); - #endif + lbry_sha256d_hash_112(thr_id, throughput, pdata[LBC_NONCE_OFT32], d_hash[thr_id]); CUDA_LOG_ERROR(); - lbry_sha512_hash_32(thr_id, throughput, d_hash[thr_id], 0); + lbry_sha512_hash_32(thr_id, throughput, d_hash[thr_id]); + CUDA_LOG_ERROR(); uint32_t resNonces[2] = { UINT32_MAX, UINT32_MAX }; - lbry_sha256d_hash_final(thr_id, throughput, pdata[LBC_NONCE_OFT32], d_hash[thr_id], resNonces, 0); + lbry_sha256d_hash_final(thr_id, throughput, pdata[LBC_NONCE_OFT32], d_hash[thr_id], resNonces); + CUDA_LOG_ERROR(); uint32_t foundNonce = resNonces[0]; *hashes_done = pdata[LBC_NONCE_OFT32] - first_nonce + throughput; @@ -165,11 +157,11 @@ extern "C" int scanhash_lbry(int thr_id, struct work *work, uint32_t max_nonce, gpulog(LOG_BLUE, thr_id, "found second nonce %08x", swab32(secNonce)); endiandata[LBC_NONCE_OFT32] = swab32_if(secNonce, !swap); lbry_hash(vhash, endiandata); + work->nonces[1] = swab32_if(secNonce, swap); if (bn_hash_target_ratio(vhash, ptarget) > work->shareratio) { work_set_target_ratio(work, vhash); xchg(work->nonces[0], work->nonces[1]); } - work->nonces[1] = swab32_if(secNonce, swap); res++; } pdata[LBC_NONCE_OFT32] = work->nonces[0]; @@ -204,7 +196,6 @@ void free_lbry(int thr_id) cudaFree(d_hash[thr_id]); lbry_sha256_free(thr_id); - cuda_check_cpu_free(thr_id); init[thr_id] = false; cudaDeviceSynchronize();