From 2b43d57d42590e080f6add5c711d47859c49d80b Mon Sep 17 00:00:00 2001 From: Tanguy Pruvot Date: Fri, 5 Jun 2015 23:12:17 +0200 Subject: [PATCH] lyra2: simplify skein code (no perf changes) --- Algo256/cuda_skein256.cu | 223 ++++++++++++++++++--------------------- 1 file changed, 103 insertions(+), 120 deletions(-) diff --git a/Algo256/cuda_skein256.cu b/Algo256/cuda_skein256.cu index bc9acef..6902509 100644 --- a/Algo256/cuda_skein256.cu +++ b/Algo256/cuda_skein256.cu @@ -20,21 +20,17 @@ static __constant__ uint2 vSKEIN_IV512_256[8] = { { 0x33EDFC13, 0x3EEDBA18 } }; -static __constant__ int ROT256[8][4] = -{ - 46,36, 19, 37, - 33,27, 14, 42, - 17,49, 36, 39, - 44, 9, 54, 56, - 39,30, 34, 24, - 13,50, 10, 17, - 25,29, 39, 43, - 8, 35, 56, 22, +static __constant__ uint8_t ROT256[8][4] = { + 46, 36, 19, 37, + 33, 27, 14, 42, + 17, 49, 36, 39, + 44, 9, 54, 56, + 39, 30, 34, 24, + 13, 50, 10, 17, + 25, 29, 39, 43, + 8, 35, 56, 22, }; -static __constant__ uint2 skein_ks_parity = { 0xA9FC1A22,0x1BD11BDA}; -static __constant__ uint64_t skein_ks_parity64 = 0x1BD11BDAA9FC1A22ull; - static __constant__ uint2 t12[6] = { { 0x20, 0 }, { 0, 0xf0000000 }, @@ -44,25 +40,19 @@ static __constant__ uint2 t12[6] = { { 0x08, 0xff000000 } }; -static __constant__ uint64_t t12_30[6] = { - 0x20, - 0xf000000000000000, - 0xf000000000000020, - 0x08, - 0xff00000000000000, - 0xff00000000000008 -}; +static __constant__ uint2 skein_ks_parity = { 0xA9FC1A22, 0x1BD11BDA }; +static __constant__ uint64_t skein_ks_parity64 = 0x1BD11BDAA9FC1A22ull; + static __forceinline__ __device__ void Round512v35(uint2 &p0, uint2 &p1, uint2 &p2, uint2 &p3, uint2 &p4, uint2 &p5, uint2 &p6, uint2 &p7, int ROT) { - p0 += p1; p1 = ROL2(p1, ROT256[ROT][0]); p1 ^= p0; - p2 += p3; p3 = ROL2(p3, ROT256[ROT][1]); p3 ^= p2; - p4 += p5; p5 = ROL2(p5, ROT256[ROT][2]); p5 ^= p4; - p6 += p7; p7 = ROL2(p7, ROT256[ROT][3]); p7 ^= p6; + p0 += p1; p1 = ROL2(p1, ROT256[ROT][0]) ^ p0; + p2 += p3; p3 = ROL2(p3, ROT256[ROT][1]) ^ p2; + p4 += p5; p5 = ROL2(p5, ROT256[ROT][2]) ^ p4; + p6 += p7; p7 = ROL2(p7, ROT256[ROT][3]) ^ p6; } - static __forceinline__ __device__ void Round_8_512v35(uint2 *ks, uint2 *ts, uint2 &p0, uint2 &p1, uint2 &p2, uint2 &p3, @@ -72,39 +62,36 @@ void Round_8_512v35(uint2 *ks, uint2 *ts, Round512v35(p2, p1, p4, p7, p6, p5, p0, p3, 1); Round512v35(p4, p1, p6, p3, p0, p5, p2, p7, 2); Round512v35(p6, p1, p0, p7, p2, p5, p4, p3, 3); - p0 += ks[((R)+0) % 9]; /* inject the key schedule value */ - p1 += ks[((R)+1) % 9]; - p2 += ks[((R)+2) % 9]; - p3 += ks[((R)+3) % 9]; - p4 += ks[((R)+4) % 9]; - p5 += ks[((R)+5) % 9] + ts[((R)+0) % 3]; - p6 += ks[((R)+6) % 9] + ts[((R)+1) % 3]; - p7 += ks[((R)+7) % 9] + make_uint2((R),0); + p0 += ks[(R+0) % 9]; /* inject the key schedule value */ + p1 += ks[(R+1) % 9]; + p2 += ks[(R+2) % 9]; + p3 += ks[(R+3) % 9]; + p4 += ks[(R+4) % 9]; + p5 += ks[(R+5) % 9] + ts[(R+0) % 3]; + p6 += ks[(R+6) % 9] + ts[(R+1) % 3]; + p7 += ks[(R+7) % 9] + make_uint2((R),0); Round512v35(p0, p1, p2, p3, p4, p5, p6, p7, 4); Round512v35(p2, p1, p4, p7, p6, p5, p0, p3, 5); Round512v35(p4, p1, p6, p3, p0, p5, p2, p7, 6); Round512v35(p6, p1, p0, p7, p2, p5, p4, p3, 7); - p0 += ks[((R)+1) % 9]; /* inject the key schedule value */ - p1 += ks[((R)+2) % 9]; - p2 += ks[((R)+3) % 9]; - p3 += ks[((R)+4) % 9]; - p4 += ks[((R)+5) % 9]; - p5 += ks[((R)+6) % 9] + ts[((R)+1) % 3]; - p6 += ks[((R)+7) % 9] + ts[((R)+2) % 3]; - p7 += ks[((R)+8) % 9] + make_uint2((R)+1, 0); + p0 += ks[(R+1) % 9]; /* inject the key schedule value */ + p1 += ks[(R+2) % 9]; + p2 += ks[(R+3) % 9]; + p3 += ks[(R+4) % 9]; + p4 += ks[(R+5) % 9]; + p5 += ks[(R+6) % 9] + ts[(R+1) % 3]; + p6 += ks[(R+7) % 9] + ts[(R+2) % 3]; + p7 += ks[(R+8) % 9] + make_uint2(R+1, 0); } __global__ __launch_bounds__(256,3) void skein256_gpu_hash_32(uint32_t threads, uint32_t startNounce, uint64_t *outputHash) { - uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); + const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); if (thread < threads) { - uint2 h[9]; - uint2 t[3]; - uint2 dt0,dt1,dt2,dt3; - uint2 p0, p1, p2, p3, p4, p5, p6, p7; + uint2 h[9], t[3] = { t12[0], t12[1], t12[2] }; h[8] = skein_ks_parity; for (int i = 0; i<8; i++) { @@ -112,27 +99,23 @@ void skein256_gpu_hash_32(uint32_t threads, uint32_t startNounce, uint64_t *outp h[8] ^= h[i]; } - t[0]=t12[0]; - t[1]=t12[1]; - t[2]=t12[2]; - - LOHI(dt0.x,dt0.y,outputHash[thread]); - LOHI(dt1.x,dt1.y,outputHash[threads+thread]); - LOHI(dt2.x,dt2.y,outputHash[2*threads+thread]); - LOHI(dt3.x,dt3.y,outputHash[3*threads+thread]); - - p0 = h[0] + dt0; - p1 = h[1] + dt1; - p2 = h[2] + dt2; - p3 = h[3] + dt3; - p4 = h[4]; - p5 = h[5] + t[0]; - p6 = h[6] + t[1]; - p7 = h[7]; - - #pragma unroll + uint2 dt0 = vectorize(outputHash[thread]); + uint2 dt1 = vectorize(outputHash[threads + thread]); + uint2 dt2 = vectorize(outputHash[threads*2 + thread]); + uint2 dt3 = vectorize(outputHash[threads*3 + thread]); + + uint2 p0 = h[0] + dt0; + uint2 p1 = h[1] + dt1; + uint2 p2 = h[2] + dt2; + uint2 p3 = h[3] + dt3; + uint2 p4 = h[4]; + uint2 p5 = h[5] + t[0]; + uint2 p6 = h[6] + t[1]; + uint2 p7 = h[7]; + + #pragma unroll 9 for (int i = 1; i<19; i+=2) { - Round_8_512v35(h,t,p0,p1,p2,p3,p4,p5,p6,p7,i); + Round_8_512v35(h, t, p0, p1, p2, p3, p4, p5, p6, p7, i); } p0 ^= dt0; @@ -158,18 +141,19 @@ void skein256_gpu_hash_32(uint32_t threads, uint32_t startNounce, uint64_t *outp t[0] = t12[3]; t[1] = t12[4]; t[2] = t12[5]; - p5 += t[0]; //p5 already equal h[5] + + p5 += t[0]; p6 += t[1]; - #pragma unroll + #pragma unroll 9 for (int i = 1; i<19; i+=2) { Round_8_512v35(h, t, p0, p1, p2, p3, p4, p5, p6, p7, i); } - outputHash[thread] = devectorize(p0); - outputHash[threads+thread] = devectorize(p1); - outputHash[2*threads+thread] = devectorize(p2); - outputHash[3*threads+thread] = devectorize(p3); + outputHash[thread] = devectorize(p0); + outputHash[threads + thread] = devectorize(p1); + outputHash[threads*2 + thread] = devectorize(p2); + outputHash[threads*3 + thread] = devectorize(p3); } } @@ -177,10 +161,10 @@ static __forceinline__ __device__ void Round512v30(uint64_t &p0, uint64_t &p1, uint64_t &p2, uint64_t &p3, uint64_t &p4, uint64_t &p5, uint64_t &p6, uint64_t &p7, int ROT) { - p0 += p1; p1 = ROTL64(p1, ROT256[ROT][0]); p1 ^= p0; - p2 += p3; p3 = ROTL64(p3, ROT256[ROT][1]); p3 ^= p2; - p4 += p5; p5 = ROTL64(p5, ROT256[ROT][2]); p5 ^= p4; - p6 += p7; p7 = ROTL64(p7, ROT256[ROT][3]); p7 ^= p6; + p0 += p1; p1 = ROTL64(p1, ROT256[ROT][0]) ^ p0; + p2 += p3; p3 = ROTL64(p3, ROT256[ROT][1]) ^ p2; + p4 += p5; p5 = ROTL64(p5, ROT256[ROT][2]) ^ p4; + p6 += p7; p7 = ROTL64(p7, ROT256[ROT][3]) ^ p6; } static __forceinline__ __device__ @@ -191,38 +175,36 @@ void Round_8_512v30(uint64_t *ks, uint64_t *ts, uint64_t &p0, uint64_t &p1, uint Round512v30(p2, p1, p4, p7, p6, p5, p0, p3, 1); Round512v30(p4, p1, p6, p3, p0, p5, p2, p7, 2); Round512v30(p6, p1, p0, p7, p2, p5, p4, p3, 3); - p0 += ks[((R)+0) % 9]; /* inject the key schedule value */ - p1 += ks[((R)+1) % 9]; - p2 += ks[((R)+2) % 9]; - p3 += ks[((R)+3) % 9]; - p4 += ks[((R)+4) % 9]; - p5 += ks[((R)+5) % 9] + ts[((R)+0) % 3]; - p6 += ks[((R)+6) % 9] + ts[((R)+1) % 3]; - p7 += ks[((R)+7) % 9] + R; + p0 += ks[(R+0) % 9]; /* inject the key schedule value */ + p1 += ks[(R+1) % 9]; + p2 += ks[(R+2) % 9]; + p3 += ks[(R+3) % 9]; + p4 += ks[(R+4) % 9]; + p5 += ks[(R+5) % 9] + ts[(R+0) % 3]; + p6 += ks[(R+6) % 9] + ts[(R+1) % 3]; + p7 += ks[(R+7) % 9] + R; Round512v30(p0, p1, p2, p3, p4, p5, p6, p7, 4); Round512v30(p2, p1, p4, p7, p6, p5, p0, p3, 5); Round512v30(p4, p1, p6, p3, p0, p5, p2, p7, 6); Round512v30(p6, p1, p0, p7, p2, p5, p4, p3, 7); - p0 += ks[((R)+1) % 9]; /* inject the key schedule value */ - p1 += ks[((R)+2) % 9]; - p2 += ks[((R)+3) % 9]; - p3 += ks[((R)+4) % 9]; - p4 += ks[((R)+5) % 9]; - p5 += ks[((R)+6) % 9] + ts[((R)+1) % 3]; - p6 += ks[((R)+7) % 9] + ts[((R)+2) % 3]; - p7 += ks[((R)+8) % 9] + (R)+1; + p0 += ks[(R+1) % 9]; /* inject the key schedule value */ + p1 += ks[(R+2) % 9]; + p2 += ks[(R+3) % 9]; + p3 += ks[(R+4) % 9]; + p4 += ks[(R+5) % 9]; + p5 += ks[(R+6) % 9] + ts[(R+1) % 3]; + p6 += ks[(R+7) % 9] + ts[(R+2) % 3]; + p7 += ks[(R+8) % 9] + R+1; } __global__ __launch_bounds__(256, 3) void skein256_gpu_hash_32_v30(uint32_t threads, uint32_t startNounce, uint64_t *outputHash) { - uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); + const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x); if (thread < threads) { - uint64_t h[9]; - uint64_t t[3]; - uint64_t dt0, dt1, dt2, dt3; - uint64_t p0, p1, p2, p3, p4, p5, p6, p7; + uint64_t h[9], t[3]; + h[8] = skein_ks_parity64; for (int i = 0; i<8; i++) { h[i] = SKEIN_IV512_256[i]; @@ -233,21 +215,21 @@ void skein256_gpu_hash_32_v30(uint32_t threads, uint32_t startNounce, uint64_t * t[1] = devectorize(t12[1]); t[2] = devectorize(t12[2]); - dt0 = outputHash[thread]; - dt1 = outputHash[threads+thread]; - dt2 = outputHash[2*threads+thread]; - dt3 = outputHash[3*threads+thread]; - - p0 = h[0] + dt0; - p1 = h[1] + dt1; - p2 = h[2] + dt2; - p3 = h[3] + dt3; - p4 = h[4]; - p5 = h[5] + t[0]; - p6 = h[6] + t[1]; - p7 = h[7]; - - #pragma unroll + uint64_t dt0 = outputHash[thread]; + uint64_t dt1 = outputHash[threads + thread]; + uint64_t dt2 = outputHash[threads*2 + thread]; + uint64_t dt3 = outputHash[threads*3 + thread]; + + uint64_t p0 = h[0] + dt0; + uint64_t p1 = h[1] + dt1; + uint64_t p2 = h[2] + dt2; + uint64_t p3 = h[3] + dt3; + uint64_t p4 = h[4]; + uint64_t p5 = h[5] + t[0]; + uint64_t p6 = h[6] + t[1]; + uint64_t p7 = h[7]; + + #pragma unroll 9 for (int i = 1; i<19; i += 2) { Round_8_512v30(h, t, p0, p1, p2, p3, p4, p5, p6, p7, i); } @@ -272,22 +254,22 @@ void skein256_gpu_hash_32_v30(uint32_t threads, uint32_t startNounce, uint64_t * h[8] ^= h[i]; } - t[0] = t12_30[3]; - t[1] = t12_30[4]; - t[2] = t12_30[5]; + t[0] = devectorize(t12[3]); + t[1] = devectorize(t12[4]); + t[2] = devectorize(t12[5]); p5 += t[0]; //p5 already equal h[5] p6 += t[1]; - #pragma unroll + #pragma unroll 9 for (int i = 1; i<19; i += 2) { Round_8_512v30(h, t, p0, p1, p2, p3, p4, p5, p6, p7, i); } outputHash[thread] = p0; - outputHash[threads + thread] = p1; - outputHash[2 * threads + thread] = p2; - outputHash[3 * threads + thread] = p3; + outputHash[threads + thread] = p1; + outputHash[threads*2 + thread] = p2; + outputHash[threads*3 + thread] = p3; } //thread } @@ -306,6 +288,7 @@ void skein256_cpu_hash_32(int thr_id, uint32_t threads, uint32_t startNounce, ui dim3 grid((threads + threadsperblock - 1) / threadsperblock); dim3 block(threadsperblock); + // only 1kH/s perf change between kernels on a 960... if (device_sm[device_map[thr_id]] > 300 && cuda_arch[device_map[thr_id]] > 300) skein256_gpu_hash_32<<>>(threads, startNounce, d_outputHash); else