Skip to content

Commit

Permalink
blake: tune up and cleanup, ~100 MH/s on a normal 750Ti
Browse files Browse the repository at this point in the history 
tested on linux and windows (x86 binary)...

but there is a high number of duplicated shares... weird
  • Loading branch information
@tpruvot
tpruvot committed Sep 1, 2014
1 parent 4a52d05 commit 0aeac87
Show file tree
Hide file tree
Showing 2 changed files with 59 additions and 144 deletions.
197 changes: 55 additions & 142 deletions blake32.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -25,14 +25,6 @@ extern "C" void blake32hash(void *output, const void *input)

#include "cuda_helper.h"

#if __CUDA_ARCH__ < 350
// Kepler (Compute 3.0) + Host
#define ROTR32(x, n) (((x) >> (n)) | ((x) << (32 - (n))))
#else
// Kepler (Compute 3.5 / 5.0)
#define ROTR32(x, n) __funnelshift_r( (x), (x), (n) )
#endif

// in cpu-miner.c
extern bool opt_benchmark;
extern bool opt_debug;
Expand All @@ -47,9 +39,11 @@ __constant__
static uint32_t pTarget[8];

__constant__
static uint32_t c_PaddedMessage80[32]; // padded message (80 bytes + padding)
static uint32_t __align__(32) c_PaddedMessage80[32]; // padded message (80 bytes + padding)

static uint32_t *d_resNounce[8];
static uint32_t *h_resNounce[8];
static bool init_made = false;

__constant__
static uint8_t c_sigma[16][16];
Expand Down Expand Up @@ -120,14 +114,14 @@ static const uint32_t c_u256[16] = {

#define GS(a,b,c,d,e) { \
v[a] += (m[sigma[i][e]] ^ u256[sigma[i][e+1]]) + v[b]; \
v[d] = ROTR32(v[d] ^ v[a], 16); \
v[d] = SPH_ROTR32(v[d] ^ v[a], 16); \
v[c] += v[d]; \
v[b] = ROTR32(v[b] ^ v[c], 12); \
v[b] = SPH_ROTR32(v[b] ^ v[c], 12); \
\
v[a] += (m[sigma[i][e+1]] ^ u256[sigma[i][e]]) + v[b]; \
v[d] = ROTR32(v[d] ^ v[a], 8); \
v[d] = SPH_ROTR32(v[d] ^ v[a], 8); \
v[c] += v[d]; \
v[b] = ROTR32(v[b] ^ v[c], 7); \
v[b] = SPH_ROTR32(v[b] ^ v[c], 7); \
}

__device__ static
Expand All @@ -138,11 +132,10 @@ void blake256_compress(uint32_t *h, uint32_t *block, uint8_t ((*sigma)[16]), con

//#pragma unroll
for (int i = 0; i < 16; ++i) {
m[i] = cuda_swab32(block[i]);
//m[i] = block[i];
m[i] = block[i];
}

#pragma unroll
#pragma unroll 8
for(int i = 0; i < 8; i++)
v[i] = h[i];

Expand All @@ -156,10 +149,6 @@ void blake256_compress(uint32_t *h, uint32_t *block, uint8_t ((*sigma)[16]), con
v[14] = u256[6];
v[15] = u256[7];

// on a 80-bytes null buffer :
// first : v = {0x6a09e667, 0xbb67ae85, 0x3c6ef372, 0xa54ff53a, ...}
// second : v = {0xb5bfb2f9, 0x14cfcc63, 0xb85c549c, 0xc9b4184e, ..., 0x299f3350, 0x082efa98, 0xec4e6c89}

//#pragma unroll
for (int i = 0; i < 14; i++) {
/* column step */
Expand All @@ -177,11 +166,10 @@ void blake256_compress(uint32_t *h, uint32_t *block, uint8_t ((*sigma)[16]), con
//#pragma unroll 16
for(int i = 0; i < 16; i++)
h[i % 8] ^= v[i];

//second H0 = 0x0c7b1594 ... H7 = 0x9051b305
}

#if __CUDA_ARCH__ >= 200
/* memory should be aligned to use __nvvm_memset */
#if (__NV_POINTER_SIZE == 64)
# define SZCT uint64_t
#else
Expand All @@ -196,9 +184,9 @@ void blake256_gpu_hash_80(int threads, uint32_t startNounce, void *outputHash)
int thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
{
uint32_t /* __align__(16) */ h[8];
uint32_t /* __align__(16) */ msg[16];
const uint32_t nounce = startNounce + thread;
uint32_t /* __align__(8) */ msg[16];
uint32_t h[8];

#pragma unroll
for(int i=0; i<8; i++)
Expand All @@ -209,35 +197,32 @@ void blake256_gpu_hash_80(int threads, uint32_t startNounce, void *outputHash)
// ------ Close: Bytes 64 to 80 ------

#if 0 /* __CUDA_ARCH__ >= 200 */
__nvvm_memset((uint8_t*)(&msg[4]), 0, sizeof(msg)-16, 16);
__nvvm_memset((uint8_t*)(&msg[4]), 0, sizeof(msg)-16, 8);
#else
msg[5] = 0;
msg[5] = 0; // uchar[17 to 55]
msg[6] = 0;
msg[7] = 0;
msg[8] = 0;
msg[9] = 0;
msg[10] = 0;
msg[11] = 0;
msg[12] = 0;

msg[14] = 0;
#endif
msg[0] = c_PaddedMessage80[16];
msg[1] = c_PaddedMessage80[17];
msg[2] = c_PaddedMessage80[18];
msg[3] = cuda_swab32(nounce); // here or at 80 ?

msg[4] = 0x80; // uchar[16] after buffer
msg[13] = 0x01000000; //((uint8_t*)msg)[55] = 1; // uchar[17 to 55]
msg[15] = 0x80020000; // 60-63 0x280
msg[3] = nounce; /* our tested value */
msg[4] = 0x80000000; //cuda_swab32(0x80U);

//h => {0xb5bfb2f9, 0x14cfcc63, 0xb85c549c, 0xc9b4184e, 0x67dfc6ce, 0x29e9904b, 0xd59ee74e, 0xfaa9c653}
//msg {0, 0, 0, 0, 0x80, 0...}
msg[13] = 1;
msg[15] = 0x280; // 60-63

blake256_compress(h, msg, c_sigma, c_u256, 0x280); // or 0x80
//h => {0x0c7b1594, 0x52328517, 0x463db487, 0xdf5e39b7, 0x1322afaf, 0x14ed562c, 0xe9d18d7d, 0x9051b305}

uint32_t *outHash = (uint32_t*) outputHash + 16*thread; // 16 = 4 x sizeof(uint32)
//#pragma unroll
uint32_t *outHash = (uint32_t*) outputHash + thread;
//#pragma unroll 8
for (int i=0; i < 8; i++) {
outHash[i] = cuda_swab32(h[i]);
}
Expand All @@ -247,7 +232,7 @@ void blake256_gpu_hash_80(int threads, uint32_t startNounce, void *outputHash)
__host__
void blake256_cpu_hash_80(int thr_id, int threads, uint32_t startNounce, uint32_t *d_outputHash, int order)
{
const int threadsperblock = 256;
const int threadsperblock = 128;

dim3 grid((threads + threadsperblock-1)/threadsperblock);
dim3 block(threadsperblock);
Expand All @@ -265,28 +250,25 @@ void gpu_check_hash_64(int threads, uint32_t startNounce, uint32_t *g_nonceVecto
int thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
{
uint32_t nounce = (g_nonceVector != NULL) ? g_nonceVector[thread] : (startNounce + thread);
const uint32_t nounce = g_nonceVector ? g_nonceVector[thread] : (startNounce + thread);

int hashPosition = nounce - startNounce;
uint32_t *inpHash = &g_hash[16 * hashPosition];
uint32_t *inpHash = &g_hash[hashPosition];
uint32_t hash[8];

#pragma unroll 8
for (int i=0; i < 8; i++)
hash[i] = inpHash[i];

int i, position = -1;
bool rc = true;

#pragma unroll 8
/* to enhance ? */
int i, rc = 1, position = -1;
for (i = 7; i >= 0; i--) {
// rc &= (hash[i] <= pTarget[i]);
if (hash[i] > pTarget[i] && position < i) {
position = i;
rc = false;
rc = false; position = i;
}
if (hash[i] < pTarget[i] && position < i) {
position = i;
rc = true;
rc = true; position = i;
}
}

Expand All @@ -298,8 +280,8 @@ void gpu_check_hash_64(int threads, uint32_t startNounce, uint32_t *g_nonceVecto
__host__
uint32_t cpu_check_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_inputHash, int order)
{
const int threadsperblock = 128;
uint32_t result = 0xffffffff;
const int threadsperblock = 256;

cudaMemset(d_resNounce[thr_id], 0xff, sizeof(uint32_t));

Expand All @@ -309,14 +291,12 @@ uint32_t cpu_check_hash_64(int thr_id, int threads, uint32_t startNounce, uint32
size_t shared_size = 0;

gpu_check_hash_64 <<<grid, block, shared_size>>>(threads, startNounce, d_nonceVector, d_inputHash, d_resNounce[thr_id]);

MyStreamSynchronize(NULL, order, thr_id);

CUDA_SAFE_CALL(cudaMemcpy(h_resNounce[thr_id], d_resNounce[thr_id], sizeof(uint32_t), cudaMemcpyDeviceToHost));

// cudaMemcpy() is asynch!
cudaThreadSynchronize();
result = *h_resNounce[thr_id];
if (cudaSuccess == cudaMemcpy(h_resNounce[thr_id], d_resNounce[thr_id], sizeof(uint32_t), cudaMemcpyDeviceToHost)) {
cudaThreadSynchronize();
result = *h_resNounce[thr_id];
}

return result;
}
Expand All @@ -325,9 +305,9 @@ __host__
void blake256_cpu_init(int thr_id)
{
CUDA_SAFE_CALL(cudaMemcpyToSymbol(c_sigma, host_sigma, sizeof(host_sigma), 0, cudaMemcpyHostToDevice));

CUDA_SAFE_CALL(cudaMallocHost(&h_resNounce[thr_id], 1*sizeof(uint32_t)));
CUDA_SAFE_CALL(cudaMalloc(&d_resNounce[thr_id], 1*sizeof(uint32_t)));
CUDA_SAFE_CALL(cudaMallocHost(&h_resNounce[thr_id], sizeof(uint32_t)));
CUDA_SAFE_CALL(cudaMalloc(&d_resNounce[thr_id], sizeof(uint32_t)));
init_made = true;
}

__host__
Expand All @@ -336,8 +316,6 @@ void blake256_cpu_setBlock_80(uint32_t *pdata, const void *ptarget)
uint32_t PaddedMessage[32];
memcpy(PaddedMessage, pdata, 80);
memset(&PaddedMessage[20], 0, 48);
//for (int i=0; i<20; i++)
// PaddedMessage[i] = cuda_swab32(pdata[i]);

CUDA_SAFE_CALL(cudaMemcpyToSymbol(pTarget, ptarget, 32, 0, cudaMemcpyHostToDevice));
CUDA_SAFE_CALL(cudaMemcpyToSymbol(c_PaddedMessage80, PaddedMessage, sizeof(PaddedMessage), 0, cudaMemcpyHostToDevice));
Expand All @@ -348,19 +326,19 @@ void blake256_cpu_setBlock_80(uint32_t *pdata, const void *ptarget)
extern "C" int scanhash_blake32(int thr_id, uint32_t *pdata, const uint32_t *ptarget,
uint32_t max_nonce, unsigned long *hashes_done)
{
uint32_t endiandata[20];
const uint32_t first_nonce = pdata[19];
const int throughput = 256*256*2;
const int throughput = 128 * 2048;
static bool init[8] = {0,0,0,0,0,0,0,0};
uint32_t endiandata[20];
uint32_t Htarg = ptarget[7];
int rc = 0;

if (opt_benchmark)
((uint32_t*)ptarget)[7] = 0x00000f;

uint32_t Htarg = ptarget[7];
((uint32_t*)ptarget)[7] = Htarg = 0x00000f;

if (!init[thr_id]) {
CUDA_SAFE_CALL(cudaSetDevice(device_map[thr_id]));
CUDA_SAFE_CALL(cudaMalloc(&d_hash[thr_id], 16 * sizeof(uint32_t) * throughput));
CUDA_SAFE_CALL(cudaMalloc(&d_hash[thr_id], 32 * throughput));

blake256_cpu_init(thr_id);

Expand All @@ -375,11 +353,11 @@ extern "C" int scanhash_blake32(int thr_id, uint32_t *pdata, const uint32_t *pta
blake32hash(vhash, pdata);
#endif

blake256_cpu_setBlock_80(pdata, (void*)ptarget);

for (int k=0; k < 20; k++)
be32enc(&endiandata[k], pdata[k]);

blake256_cpu_setBlock_80(endiandata, (void*)ptarget);

do {
int order = 0;
uint32_t foundNonce;
Expand All @@ -401,14 +379,14 @@ extern "C" int scanhash_blake32(int thr_id, uint32_t *pdata, const uint32_t *pta

blake32hash(vhashcpu, endiandata);

if (opt_debug)
applog(LOG_DEBUG, "foundNonce = %08x",foundNonce);
//if (opt_debug)
// applog(LOG_DEBUG, "foundNonce = %08x",foundNonce);

if (vhashcpu[7] <= Htarg && fulltest(vhashcpu, ptarget))
{
pdata[19] = foundNonce;
*hashes_done = pdata[19] - first_nonce + 1;
return 1;
rc = 1;
goto exit_scan;
} else {
applog(LOG_INFO, "GPU #%d: result for nonce %08x does not validate on CPU!", thr_id, foundNonce);
}
Expand All @@ -418,77 +396,12 @@ extern "C" int scanhash_blake32(int thr_id, uint32_t *pdata, const uint32_t *pta

} while (pdata[19] < max_nonce && !work_restart[thr_id].restart);

exit_scan:
*hashes_done = pdata[19] - first_nonce + 1;
return 0;
}

//#define DEBUG_ALGO

__host__
int scanhash_blake256_cpu(int thr_id, uint32_t *pdata, const uint32_t *ptarget,
uint32_t max_nonce, uint64_t *hashes_done)
{
uint32_t n = pdata[19] - 1;
const uint32_t first_nonce = pdata[19];
const uint32_t Htarg = ptarget[7];

uint32_t __align__(32) hash64[8];
uint32_t endiandata[32];

uint64_t htmax[] = {
0,
0xF,
0xFF,
0xFFF,
0xFFFF,
0x10000000
};
uint32_t masks[] = {
0xFFFFFFFF,
0xFFFFFFF0,
0xFFFFFF00,
0xFFFFF000,
0xFFFF0000,
0
};

// we need bigendian data...
for (int kk=0; kk < 32; kk++) {
be32enc(&endiandata[kk], ((uint32_t*)pdata)[kk]);
};
#ifdef DEBUG_ALGO
if (Htarg != 0)
printf("[%d] Htarg=%X\n", thr_id, Htarg);
#endif
for (int m=0; m < 6; m++) {
if (Htarg <= htmax[m]) {
uint32_t mask = masks[m];
do {
pdata[19] = ++n;
be32enc(&endiandata[19], n);
blake32hash(hash64, endiandata);
#ifndef DEBUG_ALGO
if ((!(hash64[7] & mask)) && fulltest(hash64, ptarget)) {
*hashes_done = n - first_nonce + 1;
return true;
}
#else
if (!(n % 0x1000) && !thr_id) printf(".");
if (!(hash64[7] & mask)) {
printf("[%d]",thr_id);
if (fulltest(hash64, ptarget)) {
*hashes_done = n - first_nonce + 1;
return true;
}
}
#endif
} while (n < max_nonce && !work_restart[thr_id].restart);
// see blake.c if else to understand the loop on htmax => mask
break;
}
if (init_made && opt_debug && h_resNounce[thr_id]) {
// made auto ???
//applog(LOG_DEBUG, "%08x", h_resNounce[thr_id]);
//cudaFreeHost(h_resNounce[thr_id]);
}

*hashes_done = n - first_nonce + 1;
pdata[19] = n;
return 0;
return rc;
}
Loading

0 comments on commit 0aeac87

Please sign in to comment.