@ -5,280 +5,189 @@
@@ -5,280 +5,189 @@
#include <unistd.h>
#include "cryptonight.h"
#define LONG_SHL_IDX 19
#define LONG_LOOPS32 0x80000
#define LONG_SHL_IDX 19U
#define LONG_LOOPS32 0x80000U
#include "cn_aes.cuh"
#define MUL_SUM_XOR_DST(a,c,dst) { \
uint64_t hi, lo = cuda_mul128(((uint64_t *)a)[0], ((uint64_t *)dst)[0], &hi) + ((uint64_t *)c)[1]; \
hi += ((uint64_t *)c)[0]; \
((uint64_t *)c)[0] = ((uint64_t *)dst)[0] ^ hi; \
((uint64_t *)c)[1] = ((uint64_t *)dst)[1] ^ lo; \
((uint64_t *)dst)[0] = hi; \
((uint64_t *)dst)[1] = lo; }
__device__ __forceinline__ uint64_t cuda_mul128(uint64_t multiplier, uint64_t multiplicand, uint64_t* product_hi)
{
*product_hi = __umul64hi(multiplier, multiplicand);
return(multiplier * multiplicand);
}
__global__
void cryptonight_core_gpu_phase1(int threads, uint32_t * __restrict__ long_state, uint32_t * __restrict__ ctx_state, uint32_t * __restrict__ ctx_key1)
//__launch_bounds__(128, 9) // 56 registers
void cryptonight_core_gpu_phase1(const uint32_t threads, uint32_t * long_state, uint32_t * const ctx_state, uint32_t * ctx_key1)
{
__shared__ uint32_t __align__(16) sharedMemory[1024];
__shared__ __align__(16) uint32_t sharedMemory[1024];
cn_aes_gpu_init(sharedMemory);
const int thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 3;
const int sub = (threadIdx.x & 7) << 2; // 0 4 8 ... 28
const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 3;
const uint32_t sub = (threadIdx.x & 7) << 2; // 0 4 8 ... 28
if(thread < threads)
{
const int oft = thread * 50 + sub + 16; // not aligned 16!
const int long_oft = (thread << LONG_SHL_IDX) + sub;
uint32_t __align__(16) key[40];
uint32_t __align__(16) text[4];
const uint32_t long_oft = (thread << LONG_SHL_IDX) + sub;
ulonglong2 text = AS_UL2(&ctx_state[thread * 52U + sub + 16U]);
AS_UINT2(&text[0]) = AS_UINT2(&ctx_state[oft]);
AS_UINT2(&text[2]) = AS_UINT2(&ctx_state[oft + 2]);
const uint32_t* ctx_key = &ctx_key1[thread * 40U];
uint32_t key[40];
#pragma unroll 10 // copy 160 bytes
for (uint32_t i = 0; i < 40U; i += 4U)
AS_UINT4(&key[i]) = AS_UINT4(&ctx_key[i]);
// copy 160 bytes
#pragma unroll
for (int i = 0; i < 40; i += 4)
AS_UINT4(&key[i]) = AS_UINT4(ctx_key1 + thread * 40 + i);
__syncthreads();
for(int i = 0; i < LONG_LOOPS32; i += 32) {
cn_aes_pseudo_round_mut(sharedMemory, text, key);
AS_UINT4(&long_state[long_oft + i]) = AS_UINT4(text);
__threadfence_block();
for(uint32_t i = 0; i < LONG_LOOPS32; i += 32U) {
cn_aes_pseudo_round_mut(sharedMemory, (uint32_t*) &text, key);
AS_UL2(&long_state[long_oft + i]) = text;
}
}
}
__global__
void cryptonight_core_gpu_phase2(const int threads, const int bfactor, const int partidx, uint32_t * d_long_state, uint32_t * d_ctx_a, uint32_t * d_ctx_b)
{
__shared__ uint32_t __align__(16) sharedMemory[1024];
cn_aes_gpu_init(sharedMemory);
__syncthreads();
#if 0 && __CUDA_ARCH__ >= 300
const int thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 2;
const int sub = threadIdx.x & 3;
static __forceinline__ __device__ ulonglong2 operator ^ (const ulonglong2 &a, const ulonglong2 &b) {
return make_ulonglong2(a.x ^ b.x, a.y ^ b.y);
}
static __forceinline__ __device__ uint4 operator ^ (const uint4 &a, const uint4 &b) {
return make_uint4(a.x ^ b.x, a.y ^ b.y, a.z ^ b.z, a.w ^ b.w);
}
if(thread < threads )
__device__ __forceinline__ ulonglong2 cuda_mul128(const uint64_t multiplier, const uint64_t multiplicand)
{
const int batchsize = ITER >> (2 + bfactor);
const int start = partidx * batchsize;
const int end = start + batchsize;
uint32_t * __restrict__ long_state = &d_long_state[thread << 19];
uint32_t * __restrict__ ctx_a = d_ctx_a + thread * 4;
uint32_t * __restrict__ ctx_b = d_ctx_b + thread * 4;
uint32_t a, b, c, x[4];
uint32_t t1[4], t2[4], res;
uint64_t reshi, reslo;
int j;
a = ctx_a[sub];
b = ctx_b[sub];
#pragma unroll 8
for(int i = start; i < end; ++i)
{
//j = ((uint32_t *)a)[0] & 0x1FFFF0;
j = (__shfl((int)a, 0, 4) & E2I_MASK1) >> 2;
//cn_aes_single_round(sharedMemory, &long_state[j], c, a);
x[0] = long_state[j + sub];
x[1] = __shfl((int)x[0], sub + 1, 4);
x[2] = __shfl((int)x[0], sub + 2, 4);
x[3] = __shfl((int)x[0], sub + 3, 4);
c = a ^
t_fn0(x[0] & 0xff) ^
t_fn1((x[1] >> 8) & 0xff) ^
t_fn2((x[2] >> 16) & 0xff) ^
t_fn3((x[3] >> 24) & 0xff);
//XOR_BLOCKS_DST(c, b, &long_state[j]);
long_state[j + sub] = c ^ b;
//MUL_SUM_XOR_DST(c, a, &long_state[((uint32_t *)c)[0] & 0x1FFFF0]);
j = (__shfl((int)c, 0, 4) & E2I_MASK1) >> 2;
#pragma unroll
for(int k = 0; k < 2; k++)
t1[k] = __shfl((int)c, k, 4);
#pragma unroll
for(int k = 0; k < 4; k++)
t2[k] = __shfl((int)a, k, 4);
asm(
"mad.lo.u64 %0, %2, %3, %4;\n\t"
"mad.hi.u64 %1, %2, %3, %5;\n\t"
: "=l"(reslo), "=l"(reshi)
: "l"(((uint64_t *)t1)[0]), "l"(((uint64_t *)long_state)[j >> 1]), "l"(((uint64_t *)t2)[1]), "l"(((uint64_t *)t2)[0]));
res = (sub & 2 ? reslo : reshi) >> (sub & 1 ? 32 : 0);
a = long_state[j + sub] ^ res;
long_state[j + sub] = res;
//j = ((uint32_t *)a)[0] & 0x1FFFF0;
j = (__shfl((int)a, 0, 4) & E2I_MASK1) >> 2;
//cn_aes_single_round(sharedMemory, &long_state[j], b, a);
x[0] = long_state[j + sub];
x[1] = __shfl((int)x[0], sub + 1, 4);
x[2] = __shfl((int)x[0], sub + 2, 4);
x[3] = __shfl((int)x[0], sub + 3, 4);
b = a ^
t_fn0(x[0] & 0xff) ^
t_fn1((x[1] >> 8) & 0xff) ^
t_fn2((x[2] >> 16) & 0xff) ^
t_fn3((x[3] >> 24) & 0xff);
//XOR_BLOCKS_DST(b, c, &long_state[j]);
long_state[j + sub] = c ^ b;
//MUL_SUM_XOR_DST(b, a, &long_state[((uint32_t *)b)[0] & 0x1FFFF0]);
j = (__shfl((int)b, 0, 4) & E2I_MASK1) >> 2;
#pragma unroll
for(int k = 0; k < 2; k++)
t1[k] = __shfl((int)b, k, 4);
#pragma unroll
for(int k = 0; k < 4; k++)
t2[k] = __shfl((int)a, k, 4);
asm(
"mad.lo.u64 %0, %2, %3, %4;\n\t"
"mad.hi.u64 %1, %2, %3, %5;\n\t"
: "=l"(reslo), "=l"(reshi)
: "l"(((uint64_t *)t1)[0]), "l"(((uint64_t *)long_state)[j >> 1]), "l"(((uint64_t *)t2)[1]), "l"(((uint64_t *)t2)[0]));
res = (sub & 2 ? reslo : reshi) >> (sub & 1 ? 32 : 0);
a = long_state[j + sub] ^ res;
long_state[j + sub] = res;
ulonglong2 product;
product.x = __umul64hi(multiplier, multiplicand);
product.y = multiplier * multiplicand;
return product;
}
if(bfactor > 0) {
ctx_a[sub] = a;
ctx_b[sub] = b;
static __forceinline__ __device__ void operator += (ulonglong2 &a, const ulonglong2 b) {
a.x += b.x; a.y += b.y;
}
#undef MUL_SUM_XOR_DST
__device__ __forceinline__ void MUL_SUM_XOR_DST(const uint64_t m, uint4 &a, void* far_dst)
{
ulonglong2 d = AS_UL2(far_dst);
ulonglong2 p = cuda_mul128(m, d.x);
p += AS_UL2(&a);
AS_UL2(&a) = p ^ d;
AS_UL2(far_dst) = p;
}
#else
__global__
#if __CUDA_ARCH__ >= 500
//__launch_bounds__(128,12) /* force 40 regs to allow -l ...x32 */
#endif
void cryptonight_core_gpu_phase2(const uint32_t threads, const uint32_t bfactor, const uint32_t partidx,
uint32_t * d_long_state, uint32_t * d_ctx_a, uint32_t * d_ctx_b)
{
__shared__ __align__(16) uint32_t sharedMemory[1024];
// cn_aes_gpu_init(sharedMemory);
// __syncthreads();
const int thread = blockDim.x * blockIdx.x + threadIdx.x;
const uint32_t thread = blockDim.x * blockIdx.x + threadIdx.x;
if (thread < threads)
{
const int batchsize = ITER >> (2 + bfactor);
const int start = partidx * batchsize;
const int end = start + batchsize;
const off_t longptr = (off_t)thread << LONG_SHL_IDX;
const uint32_t batchsize = ITER >> (2U + bfactor);
const uint32_t start = partidx * batchsize;
const uint32_t end = start + batchsize;
const uint32_t longptr = thread << LONG_SHL_IDX;
uint32_t * long_state = &d_long_state[longptr];
uint64_t * ctx_a = (uint64_t*)(&d_ctx_a[thread * 4 ]);
uint64_t * ctx_b = (uint64_t*)(&d_ctx_b[thread * 4 ]);
uint4 A = AS_UINT4(ctx_a);
void * ctx_a = (void*)(&d_ctx_a[thread << 2U]);
void * ctx_b = (void*)(&d_ctx_b[thread << 2U]);
uint4 A = AS_UINT4(ctx_a); // ld.global.u32.v4
uint4 B = AS_UINT4(ctx_b);
uint32_t* a = (uint32_t*)&A;
uint32_t* b = (uint32_t*)&B;
for (int i = start; i < end; i++) // end = 262144
{
uint32_t c[4];
uint32_t j = (a[0] >> 2) & E2I_MASK2;
cn_aes_single_round(sharedMemory, &long_state[j], c, a);
XOR_BLOCKS_DST(c, b, &long_state[j]);
MUL_SUM_XOR_DST(c, a, &long_state[(c[0] >> 2) & E2I_MASK2]);
j = (a[0] >> 2) & E2I_MASK2;
cn_aes_single_round(sharedMemory, &long_state[j], b, a);
XOR_BLOCKS_DST(b, c, &long_state[j]);
MUL_SUM_XOR_DST(b, a, &long_state[(b[0] >> 2) & E2I_MASK2]);
uint4 C;
uint32_t j = (A.x >> 2) & E2I_MASK2;
cn_aes_single_round_b((uint8_t*)sharedMemory, &long_state[j], A, &C);
AS_UINT4(&long_state[j]) = C ^ B; // // st.global.u32.v4
MUL_SUM_XOR_DST((AS_UL2(&C)).x, A, &long_state[(C.x >> 2U) & E2I_MASK2]);
j = (A.x >> 2) & E2I_MASK2;
cn_aes_single_round_b((uint8_t*)sharedMemory, &long_state[j], A, &B);
AS_UINT4(&long_state[j]) = C ^ B;
MUL_SUM_XOR_DST((AS_UL2(&B)).x, A, &long_state[(B.x >> 2U) & E2I_MASK2]);
}
if (bfactor > 0 ) {
if (bfactor) {
AS_UINT4(ctx_a) = A;
AS_UINT4(ctx_b) = B;
}
}
#endif // __CUDA_ARCH__ >= 300
}
__global__
void cryptonight_core_gpu_phase3(int threads, const uint32_t * __restrict__ long_state, uint32_t * ctx_state, uint32_t * __restrict__ ctx_key2)
void cryptonight_core_gpu_phase3(const uint32_t threads, const uint32_t * __restrict__ long_state, uint32_t * ctx_state, uint32_t * __restrict__ ctx_key2)
{
__shared__ uint32_t __align__(16) sharedMemory[1024];
__shared__ __align__(16) uint32_t sharedMemory[1024];
cn_aes_gpu_init(sharedMemory);
//cn_aes_gpu_init(sharedMemory);
const int thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 3;
const int sub = (threadIdx.x & 7) << 2;
const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 3U;
const uint32_t sub = (threadIdx.x & 7U) << 2U;
if(thread < threads)
{
const int long_oft = (thread << LONG_SHL_IDX) + sub;
const int oft = thread * 50 + sub + 16 ;
const uint32_t long_oft = (thread << LONG_SHL_IDX) + sub;
const uint32_t st_oft = thread * 52U + sub + 16U;
uint32_t __align__(16) key[40];
uint32_t __align__(8) text[4];
ulonglong2 text = AS_UL2(&ctx_state[st_oft]);
// copy 160 bytes
#pragma unroll
for (int i = 0; i < 40; i += 4)
AS_UINT4(&key[i]) = AS_UINT4(ctx_key2 + thread * 40 + i);
AS_UINT2(&text[0]) = AS_UINT2(&ctx_state[oft+0]);
AS_UINT2(&text[2]) = AS_UINT2(&ctx_state[oft+2]);
__syncthreads();
for(int i = 0; i < LONG_LOOPS32; i += 32)
uint32_t key[40];
const uint32_t* ctx_key = &ctx_key2[thread * 40U];
#pragma unroll 10
for (uint32_t i = 0; i < 40U; i += 4U)
AS_UL2(&key[i]) = AS_UL2(&ctx_key[i]);
//__syncthreads();
for(uint32_t i = 0; i < LONG_LOOPS32; i += 32U)
{
uint32_t __align__(16) st[4];
AS_UINT4(st) = AS_UINT4(&long_state[long_oft + i]);
#pragma unroll
for(int j = 0; j < 4; j++)
text[j] ^= st[j];
cn_aes_pseudo_round_mut(sharedMemory, text, key);
ulonglong2 st = AS_UL2(&long_state[long_oft + i]);
text = text ^ st;
cn_aes_pseudo_round_mut(sharedMemory, (uint32_t*) (&text), key);
}
AS_UINT2(&ctx_state[oft+0]) = AS_UINT2(&text[0]);
AS_UINT2(&ctx_state[oft+2]) = AS_UINT2(&text[2]);
AS_UL2(&ctx_state[st_oft]) = text;
}
}
extern int device_bfactor[MAX_GPUS];
__host__
void cryptonight_core_cpu_hash(int thr_id, int blocks, int threads, uint32_t *d_long_state, uint32_t *d_ctx_state, uint32_t *d_ctx_a, uint32_t *d_ctx_b, uint32_t *d_ctx_key1, uint32_t *d_ctx_key2)
void cryptonight_core_cpu_hash(int thr_id, int blocks, int threads, uint32_t *d_long_state, uint64_t *d_ctx_state,
uint32_t *d_ctx_a, uint32_t *d_ctx_b, uint32_t *d_ctx_key1, uint32_t *d_ctx_key2)
{
dim3 grid(blocks);
dim3 block(threads);
dim3 block2(threads << 1);
dim3 block4(threads << 2);
dim3 block8(threads << 3);
const int bfactor = device_bfactor[thr_id];
const int bsleep = bfactor ? 100 : 0;
const uint32_t bfactor = (uint32_t) device_bfactor[thr_id];
const uint32_t partcount = 1 << bfactor;
const uint32_t throughput = (uint32_t) (blocks*threads);
int i, partcount = 1 << bfactor;
int dev_id = device_map[thr_id];
const int bsleep = bfactor ? 100 : 0;
const int dev_id = device_map[thr_id];
int i;
cryptonight_core_gpu_phase1 <<<grid, block8 >>>(blocks*threads, d_long_state, d_ctx_state, d_ctx_key1);
cryptonight_core_gpu_phase1 <<<grid, block8, 4096>>> (throughput, d_long_state, (uint32_t*) d_ctx_state, d_ctx_key1);
exit_if_cudaerror(thr_id, __FUNCTION__, __LINE__);
if(partcount > 1) usleep(bsleep);
for(i = 0; i < partcount; i++)
{
cryptonight_core_gpu_phase2 <<<grid, (device_sm[dev_id] >= 300 ? block4 : block)>>>(blocks*threads, bfactor, i, d_long_state, d_ctx_a, d_ctx_b);
dim3 b = device_sm[dev_id] >= 300 ? block4 : block;
cryptonight_core_gpu_phase2 <<<grid, b, 4096>>> (throughput, bfactor, i, d_long_state, d_ctx_a, d_ctx_b);
exit_if_cudaerror(thr_id, __FUNCTION__, __LINE__);
if(partcount > 1) usleep(bsleep);
}
cryptonight_core_gpu_phase3 <<<grid, block8 >>>(blocks*threads, d_long_state, d_ctx_state, d_ctx_key2);
cryptonight_core_gpu_phase3 <<<grid, block8, 4096>>> (throughput, d_long_state, (uint32_t*) d_ctx_state, d_ctx_key2);
exit_if_cudaerror(thr_id, __FUNCTION__, __LINE__);
}
Tanguy Pruvot
8 years ago