Skip to content

Commit

Permalink
Skein/Skein2 SM 3.0 devices support
Browse files Browse the repository at this point in the history 
+ code cleanup

Signed-off-by: Tanguy Pruvot <tanguy.pruvot@gmail.com>
  • Loading branch information
@tpruvot
tpruvot committed Apr 14, 2015
1 parent d58d53f commit 163430d
Show file tree
Hide file tree
Showing 4 changed files with 261 additions and 99 deletions.
1 change: 1 addition & 0 deletions README.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -192,6 +192,7 @@ features.

Apr. 14th 2015 v1.6.1
Add the Double Skein Algo for Woodcoin
Skein/Skein2 SM 3.0 devices support

Mar. 27th 2015 v1.6.0
Add the ZR5 Algo for Ziftcoin
Expand Down
311 changes: 230 additions & 81 deletions quark/cuda_skein512.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -451,7 +451,7 @@ void quark_skein512_gpu_hash_64(uint32_t threads, uint32_t startNounce, uint64_t
}

__global__
void quark_skein512_gpu_hash_64_v30(uint32_t threads, uint32_t startNounce, uint64_t * const __restrict__ g_hash, uint32_t *g_nonceVector)
void quark_skein512_gpu_hash_64_sm3(uint32_t threads, uint32_t startNounce, uint64_t * const __restrict__ g_hash, uint32_t *g_nonceVector)
{
uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
Expand All @@ -464,7 +464,7 @@ void quark_skein512_gpu_hash_64_v30(uint32_t threads, uint32_t startNounce, uint
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 * 8];

// Init
h0 = 0x4903ADFF749C51CEull;
Expand Down Expand Up @@ -542,65 +542,14 @@ void quark_skein512_gpu_hash_64_v30(uint32_t threads, uint32_t startNounce, uint
TFBIG_ADDKEY(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], h, t, 18);

// output
uint64_t *outpHash = &g_hash[8 * hashPosition];
uint64_t *outpHash = &g_hash[hashPosition * 8];

#pragma unroll 8
for(int i=0; i<8; i++)
outpHash[i] = p[i];
}
}

__global__ __launch_bounds__(128,6)
void skein512_gpu_hash_close(uint32_t threads, uint32_t startNounce, uint64_t *g_hash)
{
uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
{
uint2 t0 = vectorize(8); // extra
uint2 t1 = vectorize(0xFF00000000000000ull); // etype
uint2 t2 = vectorize(0xB000000000000050ull);

uint64_t *state = &g_hash[8 * thread];
uint2 h0 = vectorize(state[0]);
uint2 h1 = vectorize(state[1]);
uint2 h2 = vectorize(state[2]);
uint2 h3 = vectorize(state[3]);
uint2 h4 = vectorize(state[4]);
uint2 h5 = vectorize(state[5]);
uint2 h6 = vectorize(state[6]);
uint2 h7 = vectorize(state[7]);
uint2 h8;
TFBIG_KINIT_UI2(h0, h1, h2, h3, h4, h5, h6, h7, h8, t0, t1, t2);

uint2 p[8] = { 0 };

TFBIG_4e_UI2(0);
TFBIG_4o_UI2(1);
TFBIG_4e_UI2(2);
TFBIG_4o_UI2(3);
TFBIG_4e_UI2(4);
TFBIG_4o_UI2(5);
TFBIG_4e_UI2(6);
TFBIG_4o_UI2(7);
TFBIG_4e_UI2(8);
TFBIG_4o_UI2(9);
TFBIG_4e_UI2(10);
TFBIG_4o_UI2(11);
TFBIG_4e_UI2(12);
TFBIG_4o_UI2(13);
TFBIG_4e_UI2(14);
TFBIG_4o_UI2(15);
TFBIG_4e_UI2(16);
TFBIG_4o_UI2(17);
TFBIG_ADDKEY_UI2(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], h, t, 18);

uint64_t *outpHash = state;
#pragma unroll 8
for (int i = 0; i < 8; i++)
outpHash[i] = devectorize(p[i]);
}
}

__global__ __launch_bounds__(128,5)
void skein512_gpu_hash_80(uint32_t threads, uint32_t startNounce, uint64_t *output64, int swap)
{
Expand Down Expand Up @@ -695,25 +644,219 @@ void skein512_gpu_hash_80(uint32_t threads, uint32_t startNounce, uint64_t *outp
TFBIG_4o_UI2(17);
TFBIG_ADDKEY_UI2(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], h, t, 18);

h0 = vectorize(c_PaddedMessage80[8]) ^ p[0];
h1 = nounce2 ^ p[1];
h2 = p[2];
h3 = p[3];
h4 = p[4];
h5 = p[5];
h6 = p[6];
h7 = p[7];
uint64_t *outpHash = &output64[thread * 8];
outpHash[0] = c_PaddedMessage80[8] ^ devectorize(p[0]);
outpHash[1] = devectorize(nounce2 ^ p[1]);
#pragma unroll
for(int i=2; i<8; i++)
outpHash[i] = devectorize(p[i]);
}
}

__global__
void skein512_gpu_hash_80_sm3(uint32_t threads, uint32_t startNounce, uint64_t *output64, int swap)
{
uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
{
uint64_t h0, h1, h2, h3, h4, h5, h6, h7, h8;
uint64_t t0, t1, t2;

// Init
h0 = 0x4903ADFF749C51CEull;
h1 = 0x0D95DE399746DF03ull;
h2 = 0x8FD1934127C79BCEull;
h3 = 0x9A255629FF352CB1ull;
h4 = 0x5DB62599DF6CA7B0ull;
h5 = 0xEABE394CA9D5C3F4ull;
h6 = 0x991112C71A75B523ull;
h7 = 0xAE18A40B660FCC33ull;

t0 = 64; // ptr
//t1 = vectorize(0xE0ull << 55); // etype
t1 = 0x7000000000000000ull;
TFBIG_KINIT(h0, h1, h2, h3, h4, h5, h6, h7, h8, t0, t1, t2);

uint64_t p[8];
#pragma unroll 8
for (int i = 0; i<8; i++)
p[i] = c_PaddedMessage80[i];

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 = c_PaddedMessage80[0] ^ p[0];
h1 = c_PaddedMessage80[1] ^ p[1];
h2 = c_PaddedMessage80[2] ^ p[2];
h3 = c_PaddedMessage80[3] ^ p[3];
h4 = c_PaddedMessage80[4] ^ p[4];
h5 = c_PaddedMessage80[5] ^ p[5];
h6 = c_PaddedMessage80[6] ^ p[6];
h7 = c_PaddedMessage80[7] ^ p[7];

uint32_t nonce = swap ? cuda_swab32(startNounce + thread) : startNounce + thread;
uint64_t nonce64 = MAKE_ULONGLONG(_LOWORD(c_PaddedMessage80[9]), nonce);

// skein_big_close -> etype = 0x160, ptr = 16, bcount = 1, extra = 16
p[0] = c_PaddedMessage80[8];
p[1] = nonce64;

#pragma unroll
for (int i = 2; i < 8; i++)
p[i] = 0ull;

t0 = 0x50ull; // SPH_T64(bcount << 6) + (sph_u64)(extra);
t1 = 0xB000000000000000ull; // (bcount >> 58) + ((sph_u64)(etype) << 55);

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);

// skein_big_close 2nd loop -> etype = 0x1fe, ptr = 8, bcount = 0
// output
uint64_t *outpHash = &output64[thread * 8];
outpHash[0] = devectorize(h0);
outpHash[1] = devectorize(h1);
outpHash[2] = devectorize(h2);
outpHash[3] = devectorize(h3);
outpHash[4] = devectorize(h4);
outpHash[5] = devectorize(h5);
outpHash[6] = devectorize(h6);
outpHash[7] = devectorize(h7);
outpHash[0] = c_PaddedMessage80[8] ^ p[0];
outpHash[1] = nonce64 ^ p[1];
#pragma unroll
for(int i=2; i<8; i++)
outpHash[i] = p[i];
}
}

__global__ __launch_bounds__(128,6)
void skein512_gpu_hash_close(uint32_t threads, uint32_t startNounce, uint64_t *g_hash)
{
uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
{
uint2 t0 = vectorize(8); // extra
uint2 t1 = vectorize(0xFF00000000000000ull); // etype
uint2 t2 = vectorize(0xB000000000000050ull);

uint64_t *state = &g_hash[thread * 8];
uint2 h0 = vectorize(state[0]);
uint2 h1 = vectorize(state[1]);
uint2 h2 = vectorize(state[2]);
uint2 h3 = vectorize(state[3]);
uint2 h4 = vectorize(state[4]);
uint2 h5 = vectorize(state[5]);
uint2 h6 = vectorize(state[6]);
uint2 h7 = vectorize(state[7]);
uint2 h8;
TFBIG_KINIT_UI2(h0, h1, h2, h3, h4, h5, h6, h7, h8, t0, t1, t2);

uint2 p[8] = { 0 };

TFBIG_4e_UI2(0);
TFBIG_4o_UI2(1);
TFBIG_4e_UI2(2);
TFBIG_4o_UI2(3);
TFBIG_4e_UI2(4);
TFBIG_4o_UI2(5);
TFBIG_4e_UI2(6);
TFBIG_4o_UI2(7);
TFBIG_4e_UI2(8);
TFBIG_4o_UI2(9);
TFBIG_4e_UI2(10);
TFBIG_4o_UI2(11);
TFBIG_4e_UI2(12);
TFBIG_4o_UI2(13);
TFBIG_4e_UI2(14);
TFBIG_4o_UI2(15);
TFBIG_4e_UI2(16);
TFBIG_4o_UI2(17);
TFBIG_ADDKEY_UI2(p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], h, t, 18);

uint64_t *outpHash = state;
#pragma unroll 8
for (int i = 0; i < 8; i++)
outpHash[i] = devectorize(p[i]);
}
}

__global__ __launch_bounds__(128,6)
void skein512_gpu_hash_close_sm3(uint32_t threads, uint32_t startNounce, uint64_t *g_hash)
{
uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
{
uint64_t t0 = 8ull; // extra
uint64_t t1 = 0xFF00000000000000ull; // etype
uint64_t t2 = 0xB000000000000050ull;

uint64_t *state = &g_hash[thread * 8];

uint64_t h0 = state[0];
uint64_t h1 = state[1];
uint64_t h2 = state[2];
uint64_t h3 = state[3];
uint64_t h4 = state[4];
uint64_t h5 = state[5];
uint64_t h6 = state[6];
uint64_t h7 = state[7];
uint64_t h8;
TFBIG_KINIT(h0, h1, h2, h3, h4, h5, h6, h7, h8, t0, t1, t2);

uint64_t p[8] = { 0 };

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);

uint64_t *outpHash = state;
#pragma unroll 8
for (int i = 0; i < 8; i++)
outpHash[i] = p[i];
}
}

Expand All @@ -738,33 +881,39 @@ void quark_skein512_cpu_hash_64(int thr_id, uint32_t threads, uint32_t startNoun
if (device_sm[dev_id] > 300 && cuda_arch[dev_id] > 300)
quark_skein512_gpu_hash_64 <<<grid, block>>> (threads, startNounce, (uint64_t*)d_hash, d_nonceVector);
else
quark_skein512_gpu_hash_64_v30 <<<grid, block>>> (threads, startNounce, (uint64_t*)d_hash, d_nonceVector);
quark_skein512_gpu_hash_64_sm3 <<<grid, block>>> (threads, startNounce, (uint64_t*)d_hash, d_nonceVector);

MyStreamSynchronize(NULL, order, thr_id);
}

/* skeincoin */
/* skein / skein2 */

__host__
void skein512_cpu_setBlock_80(void *pdata)
{
uint32_t PaddedMessage[32] = { 0 };
memcpy(&PaddedMessage[0], pdata, 80);
cudaMemcpyToSymbol(c_PaddedMessage80, pdata, 80, 0, cudaMemcpyHostToDevice);

CUDA_SAFE_CALL(
cudaMemcpyToSymbol(c_PaddedMessage80, PaddedMessage, sizeof(PaddedMessage), 0, cudaMemcpyHostToDevice)
);
CUDA_SAFE_CALL(cudaStreamSynchronize(NULL));
}

__host__
void skein512_cpu_hash_80(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *d_hash, int swap)
void skein512_cpu_hash_80(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *g_hash, int swap)
{
const uint32_t threadsperblock = 128;

dim3 grid((threads + threadsperblock-1)/threadsperblock);
dim3 block(threadsperblock);

// hash function is cut in 2 parts
skein512_gpu_hash_80 <<< grid, block >>> (threads, startNounce, (uint64_t*)d_hash, swap);
skein512_gpu_hash_close <<< grid, block >>> (threads, startNounce, (uint64_t*)d_hash);
int dev_id = device_map[thr_id];
uint64_t *d_hash = (uint64_t*) g_hash;

if (device_sm[dev_id] > 300 && cuda_arch[dev_id] > 300) {
// hash function is cut in 2 parts to reduce kernel size
skein512_gpu_hash_80 <<< grid, block >>> (threads, startNounce, d_hash, swap);
skein512_gpu_hash_close <<< grid, block >>> (threads, startNounce, d_hash);
} else {
// variant without uint2 variables
skein512_gpu_hash_80_sm3 <<< grid, block >>> (threads, startNounce, d_hash, swap);
skein512_gpu_hash_close_sm3 <<< grid, block >>> (threads, startNounce, d_hash);
}
}
Loading

0 comments on commit 163430d

Please sign in to comment.