Browse Source

x15: optimize by 2ms (39ms to 37)

and clean whirlpool ifdefs, cost too much to keep both methods
master
Tanguy Pruvot 10 years ago
parent
commit
7d430edc25
  1. 149
      x15/cuda_x15_whirlpool.cu

149
x15/cuda_x15_whirlpool.cu

@ -2214,7 +2214,7 @@ static uint64_t table_skew(uint64_t val, int num) { @@ -2214,7 +2214,7 @@ static uint64_t table_skew(uint64_t val, int num) {
}
__device__ __forceinline__
static uint64_t ROUND_ELT_SMALL(const uint64_t* __restrict__ sharedMemory,uint64_t in[8],
static uint64_t ROUND_ELT(const uint64_t* __restrict__ sharedMemory, uint64_t in[8],
int i0,int i1,int i2,int i3,int i4,int i5,int i6,int i7)
{
uint32_t idx0, idx1, idx2, idx3, idx4, idx5, idx6, idx7;
@ -2239,29 +2239,7 @@ static uint64_t ROUND_ELT_SMALL(const uint64_t* __restrict__ sharedMemory,uint64 @@ -2239,29 +2239,7 @@ static uint64_t ROUND_ELT_SMALL(const uint64_t* __restrict__ sharedMemory,uint64
);
}
#define ROUND_SMALL(table, in, out, c0, c1, c2, c3, c4, c5, c6, c7) { \
out ## 0 = xor1(ROUND_ELT_SMALL(table, in, 0, 7, 6, 5, 4, 3, 2, 1), c0); \
out ## 1 = xor1(ROUND_ELT_SMALL(table, in, 1, 0, 7, 6, 5, 4, 3, 2), c1); \
out ## 2 = xor1(ROUND_ELT_SMALL(table, in, 2, 1, 0, 7, 6, 5, 4, 3), c2); \
out ## 3 = xor1(ROUND_ELT_SMALL(table, in, 3, 2, 1, 0, 7, 6, 5, 4), c3); \
out ## 4 = xor1(ROUND_ELT_SMALL(table, in, 4, 3, 2, 1, 0, 7, 6, 5), c4); \
out ## 5 = xor1(ROUND_ELT_SMALL(table, in, 5, 4, 3, 2, 1, 0, 7, 6), c5); \
out ## 6 = xor1(ROUND_ELT_SMALL(table, in, 6, 5, 4, 3, 2, 1, 0, 7), c6); \
out ## 7 = xor1(ROUND_ELT_SMALL(table, in, 7, 6, 5, 4, 3, 2, 1, 0), c7); \
}
#define ROUND_KSCHED_SMALL(table, in, out, c) \
ROUND_SMALL(table, in, out, c, 0, 0, 0, 0, 0, 0, 0) \
TRANSFER(in, out)
#define ROUND_WENC_SMALL(table, in, key, out) \
ROUND_SMALL(table, in, out, key[0], key[1], key[2],key[3], key[4], key[5], key[6], key[7]) \
TRANSFER(in, out)
#else
# define ROUND_KSCHED_SMALL(table, in, out, c)
# define ROUND_WENC_SMALL(table, in, key, out)
#endif
__device__ __forceinline__
static uint64_t ROUND_ELT(const uint64_t* __restrict__ sharedMemory, uint64_t in[8],
@ -2281,6 +2259,8 @@ static uint64_t ROUND_ELT(const uint64_t* __restrict__ sharedMemory, uint64_t in @@ -2281,6 +2259,8 @@ static uint64_t ROUND_ELT(const uint64_t* __restrict__ sharedMemory, uint64_t in
sharedMemory[idx4],sharedMemory[idx5],sharedMemory[idx6],sharedMemory[idx7]);
}
#endif /* USE_ALL_TABLES */
#define ROUND(table, in, out, c0, c1, c2, c3, c4, c5, c6, c7) { \
out ## 0 = xor1(ROUND_ELT(table, in, 0, 7, 6, 5, 4, 3, 2, 1), c0); \
out ## 1 = xor1(ROUND_ELT(table, in, 1, 0, 7, 6, 5, 4, 3, 2), c1); \
@ -2308,7 +2288,7 @@ void whirlpool512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHas @@ -2308,7 +2288,7 @@ void whirlpool512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHas
if (threadIdx.x < 256) {
sharedMemory[threadIdx.x] = mixTob0Tox[threadIdx.x];
if (USE_ALL_TABLES) {
#if USE_ALL_TABLES
sharedMemory[threadIdx.x+256] = mixTob1Tox[threadIdx.x];
sharedMemory[threadIdx.x+512] = mixTob2Tox[threadIdx.x];
sharedMemory[threadIdx.x+768] = mixTob3Tox[threadIdx.x];
@ -2316,7 +2296,7 @@ void whirlpool512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHas @@ -2316,7 +2296,7 @@ void whirlpool512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHas
sharedMemory[threadIdx.x+1280] = mixTob5Tox[threadIdx.x];
sharedMemory[threadIdx.x+1536] = mixTob6Tox[threadIdx.x];
sharedMemory[threadIdx.x+1792] = mixTob7Tox[threadIdx.x];
}
#endif
}
int thread = (blockDim.x * blockIdx.x + threadIdx.x);
@ -2343,13 +2323,8 @@ void whirlpool512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHas @@ -2343,13 +2323,8 @@ void whirlpool512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHas
#pragma unroll 10
for (unsigned r=0; r < 10; r++) {
uint64_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
if (USE_ALL_TABLES) {
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
} else {
ROUND_KSCHED_SMALL(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC_SMALL(sharedMemory, n, h, tmp);
}
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
}
#pragma unroll 8
@ -2374,13 +2349,8 @@ void whirlpool512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHas @@ -2374,13 +2349,8 @@ void whirlpool512_gpu_hash_80(int threads, uint32_t startNounce, void *outputHas
#pragma unroll 10
for (unsigned r=0; r < 10; r++) {
uint64_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
if (USE_ALL_TABLES) {
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
} else {
ROUND_KSCHED_SMALL(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC_SMALL(sharedMemory, n, h, tmp);
}
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
}
state[0] = xor3(state[0], n[0], c_PaddedMessage80[8]);
@ -2412,7 +2382,7 @@ void whirlpool512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_has @@ -2412,7 +2382,7 @@ void whirlpool512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_has
if (threadIdx.x < 256) {
sharedMemory[threadIdx.x] = mixTob0Tox[threadIdx.x];
if (USE_ALL_TABLES) {
#if USE_ALL_TABLES
sharedMemory[threadIdx.x+256] = mixTob1Tox[threadIdx.x];
sharedMemory[threadIdx.x+512] = mixTob2Tox[threadIdx.x];
sharedMemory[threadIdx.x+768] = mixTob3Tox[threadIdx.x];
@ -2420,53 +2390,32 @@ void whirlpool512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_has @@ -2420,53 +2390,32 @@ void whirlpool512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_has
sharedMemory[threadIdx.x+1280] = mixTob5Tox[threadIdx.x];
sharedMemory[threadIdx.x+1536] = mixTob6Tox[threadIdx.x];
sharedMemory[threadIdx.x+1792] = mixTob7Tox[threadIdx.x];
}
#endif
}
int thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
{
uint32_t nounce = (g_nonceVector != NULL) ? g_nonceVector[thread] : (startNounce + thread);
int hashPosition = nounce - startNounce;
uint32_t *inpHash = (uint32_t*)&g_hash[8 * hashPosition];
union {
uint8_t h1[64];
uint32_t h4[16];
uint64_t h8[8];
} hash;
#pragma unroll 16
for (int i=0;i<16;i++) {
hash.h4[i]= inpHash[i];
}
uint64_t state[8];
uint64_t n[8];
uint64_t h[8];
uint32_t hashPosition = (nounce - startNounce) << 3;
uint64_t hash[8], state[8], n[8], h[8] = {0,0,0,0, 0,0,0,0};
uint8_t i;
#pragma unroll 8
for (int i=0;i<8;i++) {
n[i] = hash.h8[i];
h[i] = 0;
n[i] = xor1(n[i],h[i]);
for (i=0; i<8; i++) {
n[i] = hash[i] = g_hash[hashPosition + i];
}
#pragma unroll 10
for (unsigned r=0; r < 10; r++) {
for (i=0; i < 10; i++) {
uint64_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
if (USE_ALL_TABLES) {
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
} else {
ROUND_KSCHED_SMALL(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC_SMALL(sharedMemory, n, h, tmp);
}
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[i]);
ROUND_WENC(sharedMemory, n, h, tmp);
}
#pragma unroll 8
for (int i=0; i<8; i++) {
state[i] = xor1(n[i],hash.h8[i]);
for (i=0; i<8; i++) {
state[i] = xor1(n[i], hash[i]);
n[i]=0;
}
@ -2474,21 +2423,16 @@ void whirlpool512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_has @@ -2474,21 +2423,16 @@ void whirlpool512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_has
n[7] = 0x2000000000000;
#pragma unroll 8
for (int i=0; i < 8; i++) {
for (i=0; i < 8; i++) {
h[i] = state[i];
n[i] = xor1(n[i], h[i]);
}
#pragma unroll 10
for (unsigned r=0; r < 10; r++) {
for (i=0; i < 10; i++) {
uint64_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
if (USE_ALL_TABLES) {
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
} else {
ROUND_KSCHED_SMALL(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC_SMALL(sharedMemory, n, h, tmp);
}
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[i]);
ROUND_WENC(sharedMemory, n, h, tmp);
}
state[0] = xor3(state[0], n[0], 0x80);
@ -2501,12 +2445,8 @@ void whirlpool512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_has @@ -2501,12 +2445,8 @@ void whirlpool512_gpu_hash_64(int threads, uint32_t startNounce, uint64_t *g_has
state[7] = xor3(state[7], n[7], 0x2000000000000);
#pragma unroll 8
for (unsigned i = 0; i < 8; i++)
hash.h8[i] = state[i];
#pragma unroll 16
for (int u = 0; u < 16; u ++)
inpHash[u] = hash.h4[u];
for (i=0; i < 8; i++)
g_hash[hashPosition + i] = state[i];
}
}
@ -2518,7 +2458,7 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t * @@ -2518,7 +2458,7 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t *
if (threadIdx.x < 256)
{
sharedMemory[threadIdx.x] = mixTob0Tox[threadIdx.x];
if (USE_ALL_TABLES) {
#if USE_ALL_TABLES
sharedMemory[threadIdx.x+256] = mixTob1Tox[threadIdx.x];
sharedMemory[threadIdx.x+512] = mixTob2Tox[threadIdx.x];
sharedMemory[threadIdx.x+768] = mixTob3Tox[threadIdx.x];
@ -2526,7 +2466,7 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t * @@ -2526,7 +2466,7 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t *
sharedMemory[threadIdx.x+1280] = mixTob5Tox[threadIdx.x];
sharedMemory[threadIdx.x+1536] = mixTob6Tox[threadIdx.x];
sharedMemory[threadIdx.x+1792] = mixTob7Tox[threadIdx.x];
}
#endif
}
int thread = (blockDim.x * blockIdx.x + threadIdx.x);
@ -2537,7 +2477,6 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t * @@ -2537,7 +2477,6 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t *
int hashPosition = nounce - startNounce;
uint32_t *inpHash = (uint32_t*)&g_hash[8 * hashPosition];
union {
uint8_t h1[64];
uint32_t h4[16];
uint64_t h8[8];
} hash;
@ -2555,24 +2494,19 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t * @@ -2555,24 +2494,19 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t *
for (int i=0; i<8; i++) {
n[i] = hash.h8[i];
h[i] = 0;
n[i] = xor1(n[i],h[i]);
n[i] = xor1(n[i], h[i]);
}
#pragma unroll 10
for (unsigned r=0; r < 10; r++) {
uint64_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
if (USE_ALL_TABLES) {
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
} else {
ROUND_KSCHED_SMALL(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC_SMALL(sharedMemory, n, h, tmp);
}
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
}
#pragma unroll 8
for (int i=0; i<8; i++) {
state[i] = xor1(n[i],hash.h8[i]);
state[i] = xor1(n[i], hash.h8[i]);
n[i]=0;
}
@ -2588,13 +2522,8 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t * @@ -2588,13 +2522,8 @@ void whirlpool512_gpu_finalhash_64(int threads, uint32_t startNounce, uint64_t *
#pragma unroll 10
for (unsigned r=0; r < 10; r++) {
uint64_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
if (USE_ALL_TABLES) {
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
} else {
ROUND_KSCHED_SMALL(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC_SMALL(sharedMemory, n, h, tmp);
}
ROUND_KSCHED(sharedMemory, h, tmp, InitVector_RC[r]);
ROUND_WENC(sharedMemory, n, h, tmp);
}
state[0] = xor3(state[0], n[0], 0x80);
@ -2634,6 +2563,7 @@ extern void x15_whirlpool_cpu_init(int thr_id, int threads, int mode) @@ -2634,6 +2563,7 @@ extern void x15_whirlpool_cpu_init(int thr_id, int threads, int mode)
case 0: /* x15 with rotated T1-T7 (based on T0) */
cudaMemcpyToSymbol(InitVector_RC, plain_RC, sizeof(plain_RC), 0, cudaMemcpyHostToDevice);
cudaMemcpyToSymbol(mixTob0Tox, plain_T0, sizeof(plain_T0), 0, cudaMemcpyHostToDevice);
#if USE_ALL_TABLES
cudaMemcpyToSymbol(mixTob1Tox, plain_T1, (256*8), 0, cudaMemcpyHostToDevice);
cudaMemcpyToSymbol(mixTob2Tox, plain_T2, (256*8), 0, cudaMemcpyHostToDevice);
cudaMemcpyToSymbol(mixTob3Tox, plain_T3, (256*8), 0, cudaMemcpyHostToDevice);
@ -2641,6 +2571,7 @@ extern void x15_whirlpool_cpu_init(int thr_id, int threads, int mode) @@ -2641,6 +2571,7 @@ extern void x15_whirlpool_cpu_init(int thr_id, int threads, int mode)
cudaMemcpyToSymbol(mixTob5Tox, plain_T5, (256*8), 0, cudaMemcpyHostToDevice);
cudaMemcpyToSymbol(mixTob6Tox, plain_T6, (256*8), 0, cudaMemcpyHostToDevice);
cudaMemcpyToSymbol(mixTob7Tox, plain_T7, (256*8), 0, cudaMemcpyHostToDevice);
#endif
break;
case 1: /* old (whirlcoin?) */
@ -2654,12 +2585,6 @@ extern void x15_whirlpool_cpu_init(int thr_id, int threads, int mode) @@ -2654,12 +2585,6 @@ extern void x15_whirlpool_cpu_init(int thr_id, int threads, int mode)
cudaMemcpyToSymbol(mixTob6Tox, old1_T6, (256*8), 0, cudaMemcpyHostToDevice);
cudaMemcpyToSymbol(mixTob7Tox, old1_T7, (256*8), 0, cudaMemcpyHostToDevice);
break;
#if !USE_ALL_TABLES
case 2: /* x15 without rotated T1-T7, slower but use less memory */
cudaMemcpyToSymbol(InitVector_RC, plain_RC, sizeof(plain_RC), 0, cudaMemcpyHostToDevice);
cudaMemcpyToSymbol(mixTob0Tox, plain_T0, sizeof(plain_T0), 0, cudaMemcpyHostToDevice);
break;
#endif
}
cudaMalloc(&d_WNonce[thr_id], sizeof(uint32_t));

Loading…
Cancel
Save