Skip to content

Commit

Permalink
groestl: explain code and improve perf on SM 2.x
Browse files Browse the repository at this point in the history 
Signed-off-by: Tanguy Pruvot <tanguy.pruvot@gmail.com>
  • Loading branch information
@tpruvot
tpruvot committed Jan 28, 2017
1 parent feb99d0 commit b7e8a38
Show file tree
Hide file tree
Showing 6 changed files with 641 additions and 409 deletions.
1 change: 1 addition & 0 deletions ccminer.vcxproj
View file
Original file line number Diff line number Diff line change
Expand Up @@ -353,6 +353,7 @@
<ClInclude Include="miner.h" />
<ClInclude Include="nvml.h" />
<ClInclude Include="quark\cuda_bmw512_sm3.cuh" />
<ClInclude Include="quark\cuda_quark_groestl512_sm2.cuh" />
<ClInclude Include="quark\cuda_quark_blake512_sp.cuh" />
<ClInclude Include="quark\cuda_skein512_sp.cuh" />
<ClInclude Include="res\resource.h" />
Expand Down
3 changes: 3 additions & 0 deletions ccminer.vcxproj.filters
View file
Original file line number Diff line number Diff line change
Expand Up @@ -512,6 +512,9 @@
<ClInclude Include="quark\cuda_bmw512_sm3.cuh">
<Filter>Source Files\CUDA\quark</Filter>
</ClInclude>
<ClInclude Include="quark\cuda_quark_groestl512_sm2.cuh">
<Filter>Source Files\CUDA\quark</Filter>
</ClInclude>
<ClInclude Include="x11\cuda_x11_simd512_sm2.cuh">
<Filter>Source Files\CUDA\x11</Filter>
</ClInclude>
Expand Down
234 changes: 167 additions & 67 deletions quark/cuda_quark_groestl512.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@
#include <memory.h>
#include <sys/types.h> // off_t

#include "cuda_helper.h"
#include <cuda_helper.h>

#ifdef __INTELLISENSE__
#define __CUDA_ARCH__ 500
Expand All @@ -14,98 +14,198 @@
#define THF 4U

#if __CUDA_ARCH__ >= 300
#include "quark/groestl_functions_quad.h"
#include "quark/groestl_transf_quad.h"
#include "groestl_functions_quad.h"
#include "groestl_transf_quad.h"
#endif

#include "quark/cuda_quark_groestl512_sm20.cu"
#define WANT_GROESTL80
#ifdef WANT_GROESTL80
__constant__ static uint32_t c_Message80[20];
#endif

#include "cuda_quark_groestl512_sm2.cuh"

__global__ __launch_bounds__(TPB, THF)
void quark_groestl512_gpu_hash_64_quad(const uint32_t threads, const uint32_t startNounce, uint32_t * g_hash, uint32_t * __restrict g_nonceVector)
{
#if __CUDA_ARCH__ >= 300
// durch 4 dividieren, weil jeweils 4 Threads zusammen ein Hash berechnen
const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 2;
if (thread < threads)
{
// GROESTL
uint32_t message[8];
uint32_t state[8];

uint32_t nounce = g_nonceVector ? g_nonceVector[thread] : (startNounce + thread);
off_t hashPosition = nounce - startNounce;
uint32_t *pHash = &g_hash[hashPosition << 4];

const uint32_t thr = threadIdx.x & 0x3; // % THF

#pragma unroll
for(int k=0;k<4;k++) message[k] = pHash[thr + (k * THF)];

#pragma unroll
for(int k=4;k<8;k++) message[k] = 0;

if (thr == 0) message[4] = 0x80U;
if (thr == 3) message[7] = 0x01000000U;

uint32_t msgBitsliced[8];
to_bitslice_quad(message, msgBitsliced);

groestl512_progressMessage_quad(state, msgBitsliced);

// Nur der erste von jeweils 4 Threads bekommt das Ergebns-Hash
uint32_t __align__(16) hash[16];
from_bitslice_quad(state, hash);

// uint4 = 4x4 uint32_t = 16 bytes
if (thr == 0) {
uint4 *phash = (uint4*) hash;
uint4 *outpt = (uint4*) pHash;
outpt[0] = phash[0];
outpt[1] = phash[1];
outpt[2] = phash[2];
outpt[3] = phash[3];
}
}

// BEWARE : 4-WAY CODE (one hash need 4 threads)
const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 2;

if (thread < threads)
{
uint32_t message[8];
uint32_t state[8];

uint32_t nounce = g_nonceVector ? g_nonceVector[thread] : (startNounce + thread);
off_t hashPosition = nounce - startNounce;
uint32_t *pHash = &g_hash[hashPosition << 4];

const uint32_t thr = threadIdx.x & 0x3; // % THF

/*| M0 M1 M2 M3 | M4 M5 M6 M7 | (input)
--|-------------|-------------|
T0| 0 4 8 12 | 80 |
T1| 1 5 13 | |
T2| 2 6 14 | |
T3| 3 7 15 | 01 |
--|-------------|-------------| */

#pragma unroll
for(int k=0;k<4;k++) message[k] = pHash[thr + (k * THF)];

#pragma unroll
for(int k=4;k<8;k++) message[k] = 0;

if (thr == 0) message[4] = 0x80U; // end of data tag
if (thr == 3) message[7] = 0x01000000U;

uint32_t msgBitsliced[8];
to_bitslice_quad(message, msgBitsliced);

groestl512_progressMessage_quad(state, msgBitsliced);

uint32_t hash[16];
from_bitslice_quad(state, hash);

// uint4 = 4x4 uint32_t = 16 bytes
if (thr == 0) {
uint4 *phash = (uint4*) hash;
uint4 *outpt = (uint4*) pHash;
outpt[0] = phash[0];
outpt[1] = phash[1];
outpt[2] = phash[2];
outpt[3] = phash[3];
}
}
#endif
}

__host__
void quark_groestl512_cpu_init(int thr_id, uint32_t threads)
{
int dev_id = device_map[thr_id];
cuda_get_arch(thr_id);
if (device_sm[dev_id] < 300 || cuda_arch[dev_id] < 300)
quark_groestl512_sm20_init(thr_id, threads);
int dev_id = device_map[thr_id];
cuda_get_arch(thr_id);
if (device_sm[dev_id] < 300 || cuda_arch[dev_id] < 300)
quark_groestl512_sm20_init(thr_id, threads);
}

__host__
void quark_groestl512_cpu_free(int thr_id)
{
int dev_id = device_map[thr_id];
if (device_sm[dev_id] < 300 || cuda_arch[dev_id] < 300)
quark_groestl512_sm20_free(thr_id);
int dev_id = device_map[thr_id];
if (device_sm[dev_id] < 300 || cuda_arch[dev_id] < 300)
quark_groestl512_sm20_free(thr_id);
}

__host__
void quark_groestl512_cpu_hash_64(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order)
{
uint32_t threadsperblock = TPB;
uint32_t threadsperblock = TPB;

// Compute 3.0 benutzt die registeroptimierte Quad Variante mit Warp Shuffle
// mit den Quad Funktionen brauchen wir jetzt 4 threads pro Hash, daher Faktor 4 bei der Blockzahl
const uint32_t factor = THF;
// Compute 3.0 benutzt die registeroptimierte Quad Variante mit Warp Shuffle
// mit den Quad Funktionen brauchen wir jetzt 4 threads pro Hash, daher Faktor 4 bei der Blockzahl
const uint32_t factor = THF;

// berechne wie viele Thread Blocks wir brauchen
dim3 grid(factor*((threads + threadsperblock-1)/threadsperblock));
dim3 block(threadsperblock);
dim3 grid(factor*((threads + threadsperblock-1)/threadsperblock));
dim3 block(threadsperblock);

int dev_id = device_map[thr_id];
int dev_id = device_map[thr_id];

if (device_sm[dev_id] >= 300 && cuda_arch[dev_id] >= 300)
quark_groestl512_gpu_hash_64_quad<<<grid, block>>>(threads, startNounce, d_hash, d_nonceVector);
else
quark_groestl512_sm20_hash_64(thr_id, threads, startNounce, d_nonceVector, d_hash, order);
if (device_sm[dev_id] >= 300 && cuda_arch[dev_id] >= 300)
quark_groestl512_gpu_hash_64_quad<<<grid, block>>>(threads, startNounce, d_hash, d_nonceVector);
else
quark_groestl512_sm20_hash_64(thr_id, threads, startNounce, d_nonceVector, d_hash, order);
}

// Strategisches Sleep Kommando zur Senkung der CPU Last
// MyStreamSynchronize(NULL, order, thr_id);
// --------------------------------------------------------------------------------------------------------------------------------------------

#ifdef WANT_GROESTL80

__host__
void groestl512_setBlock_80(int thr_id, uint32_t *endiandata)
{
cudaMemcpyToSymbol(c_Message80, endiandata, sizeof(c_Message80), 0, cudaMemcpyHostToDevice);
}

__global__ __launch_bounds__(TPB, THF)
void groestl512_gpu_hash_80_quad(const uint32_t threads, const uint32_t startNounce, uint32_t * g_outhash)
{
#if __CUDA_ARCH__ >= 300
// BEWARE : 4-WAY CODE (one hash need 4 threads)
const uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 2;
if (thread < threads)
{
const uint32_t thr = threadIdx.x & 0x3; // % THF

/*| M0 M1 M2 M3 M4 | M5 M6 M7 | (input)
--|----------------|----------|
T0| 0 4 8 12 16 | 80 |
T1| 1 5 17 | |
T2| 2 6 18 | |
T3| 3 7 Nc | 01 |
--|----------------|----------| TPR */

uint32_t message[8];

#pragma unroll 5
for(int k=0; k<5; k++) message[k] = c_Message80[thr + (k * THF)];

#pragma unroll 3
for(int k=5; k<8; k++) message[k] = 0;

if (thr == 0) message[5] = 0x80U;
if (thr == 3) {
message[4] = cuda_swab32(startNounce + thread);
message[7] = 0x01000000U;
}

uint32_t msgBitsliced[8];
to_bitslice_quad(message, msgBitsliced);

uint32_t state[8];
groestl512_progressMessage_quad(state, msgBitsliced);

uint32_t hash[16];
from_bitslice_quad(state, hash);

if (thr == 0) { /* 4 threads were done */
const off_t hashPosition = thread;
//if (!thread) hash[15] = 0xFFFFFFFF;
uint4 *outpt = (uint4*) &g_outhash[hashPosition << 4];
uint4 *phash = (uint4*) hash;
outpt[0] = phash[0];
outpt[1] = phash[1];
outpt[2] = phash[2];
outpt[3] = phash[3];
}
}
#endif
}

__host__
void groestl512_cuda_hash_80(const int thr_id, const uint32_t threads, const uint32_t startNounce, uint32_t *d_hash)
{
int dev_id = device_map[thr_id];

if (device_sm[dev_id] >= 300 && cuda_arch[dev_id] >= 300) {
const uint32_t threadsperblock = TPB;
const uint32_t factor = THF;

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

groestl512_gpu_hash_80_quad <<<grid, block>>> (threads, startNounce, d_hash);

} else {

const uint32_t threadsperblock = 256;
dim3 grid((threads + threadsperblock-1)/threadsperblock);
dim3 block(threadsperblock);

groestl512_gpu_hash_80_sm2 <<<grid, block>>> (threads, startNounce, d_hash);
}
}

#endif

0 comments on commit b7e8a38

Please sign in to comment.