Skip to content

Commit

Permalink
Speed optimization: do MUCH LESS Groestl in Jackpot, throw away 75% o…
Browse files Browse the repository at this point in the history 
…f the hashes. More speed ;-)
  • Loading branch information
@cbuchner1
cbuchner1 committed May 5, 2014
1 parent 48996fa commit be044f3
Show file tree
Hide file tree
Showing 4 changed files with 363 additions and 132 deletions.
257 changes: 177 additions & 80 deletions JHA/cuda_jha_compactionTest.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,69 +14,54 @@ extern int device_map[8];
static cudaDeviceProp props[8];

static uint32_t *d_tempBranch1Nonces[8];
static uint32_t *d_tempBranch2Nonces[8];
static size_t *d_numValid[8];
static size_t *h_numValid[8];
static uint32_t *d_numValid[8];
static uint32_t *h_numValid[8];

static uint32_t *d_partSum1[8], *d_partSum2[8]; // 2x partielle summen
static uint32_t *d_validTemp1[8], *d_validTemp2[8];

// Zwischenspeicher
static uint32_t *d_tempBranchAllNonces[8];
static uint32_t *d_partSum[2][8]; // für bis zu vier partielle Summen

// aus heavy.cu
extern cudaError_t MyStreamSynchronize(cudaStream_t stream, int situation, int thr_id);

// True/False tester
typedef uint32_t(*cuda_compactTestFunction_t)(uint32_t *inpHash);

__device__ uint32_t JackpotTrueTest(uint32_t *inpHash)
{
uint32_t tmp = inpHash[0] & 0x01;
return (tmp == 1);
}

__device__ uint32_t JackpotFalseTest(uint32_t *inpHash)
{
uint32_t tmp = inpHash[0] & 0x01;
return (tmp == 0);
}

__device__ cuda_compactTestFunction_t d_JackpotTrueFunction = JackpotTrueTest, d_JackpotFalseFunction = JackpotFalseTest;
cuda_compactTestFunction_t h_JackpotTrueFunction[8], h_JackpotFalseFunction[8];

// Setup-Funktionen
__host__ void jackpot_compactTest_cpu_init(int thr_id, int threads)
{
cudaGetDeviceProperties(&props[thr_id], device_map[thr_id]);

cudaMemcpyFromSymbol(&h_JackpotTrueFunction[thr_id], d_JackpotTrueFunction, sizeof(cuda_compactTestFunction_t));
cudaMemcpyFromSymbol(&h_JackpotFalseFunction[thr_id], d_JackpotFalseFunction, sizeof(cuda_compactTestFunction_t));

// wir brauchen auch Speicherplatz auf dem Device
cudaMalloc(&d_tempBranchAllNonces[thr_id], sizeof(uint32_t) * threads);
cudaMalloc(&d_tempBranch1Nonces[thr_id], sizeof(uint32_t) * threads);
cudaMalloc(&d_tempBranch2Nonces[thr_id], sizeof(uint32_t) * threads);
cudaMalloc(&d_numValid[thr_id], 2*sizeof(size_t));
cudaMallocHost(&h_numValid[thr_id], 2*sizeof(size_t));
cudaMalloc(&d_tempBranch1Nonces[thr_id], sizeof(uint32_t) * threads * 2);
cudaMalloc(&d_numValid[thr_id], 2*sizeof(uint32_t));
cudaMallocHost(&h_numValid[thr_id], 2*sizeof(uint32_t));

uint32_t s1;
s1 = threads / 256;
s1 = (threads / 256) * 2;

cudaMalloc(&d_partSum1[thr_id], sizeof(uint32_t) * s1); // BLOCKSIZE (Threads/Block)
cudaMalloc(&d_partSum2[thr_id], sizeof(uint32_t) * s1); // BLOCKSIZE (Threads/Block)

cudaMalloc(&d_validTemp1[thr_id], sizeof(uint32_t) * threads); // BLOCKSIZE (Threads/Block)
cudaMalloc(&d_validTemp2[thr_id], sizeof(uint32_t) * threads); // BLOCKSIZE (Threads/Block)
}

// Die Testfunktion (zum Erstellen der TestMap)
__global__ void jackpot_compactTest_gpu_TEST_64(int threads, uint32_t startNounce, uint32_t *inpHashes, uint32_t *d_noncesFull,
uint32_t *d_nonces1, uint32_t *d_nonces2,
uint32_t *d_validT1, uint32_t *d_validT2)
{
int thread = (blockDim.x * blockIdx.x + threadIdx.x);
if (thread < threads)
{
// bestimme den aktuellen Zähler
uint32_t nounce = startNounce + thread;
uint32_t *inpHash = &inpHashes[16 * thread];

uint32_t tmp = inpHash[0] & 0x01;
uint32_t val1 = (tmp == 1);
uint32_t val2 = (tmp == 0);

d_nonces1[thread] = val1;
d_validT1[thread] = val1;
d_nonces2[thread] = val2;
d_validT2[thread] = val2;
d_noncesFull[thread] = nounce;
}
cudaMalloc(&d_partSum[0][thr_id], sizeof(uint32_t) * s1); // BLOCKSIZE (Threads/Block)
cudaMalloc(&d_partSum[1][thr_id], sizeof(uint32_t) * s1); // BLOCKSIZE (Threads/Block)
}

// Die Summenfunktion (vom NVIDIA SDK)
__global__ void jackpot_compactTest_gpu_SCAN(uint32_t *data, int width, uint32_t *partial_sums=NULL)
__global__ void jackpot_compactTest_gpu_SCAN(uint32_t *data, int width, uint32_t *partial_sums=NULL, cuda_compactTestFunction_t testFunc=NULL, int threads=0, uint32_t startNounce=0, uint32_t *inpHashes=NULL, uint32_t *d_validNonceTable=NULL)
{
extern __shared__ uint32_t sums[];
int id = ((blockIdx.x * blockDim.x) + threadIdx.x);
Expand All @@ -86,10 +71,38 @@ __global__ void jackpot_compactTest_gpu_SCAN(uint32_t *data, int width, uint32_t
//int warp_id = threadIdx.x / warpSize;
int warp_id = threadIdx.x / width;

sums[lane_id] = 0;

// Below is the basic structure of using a shfl instruction
// for a scan.
// Record "value" as a variable - we accumulate it along the way
uint32_t value = data[id];
uint32_t value;
if(testFunc != NULL)
{
if (id < threads)
{
uint32_t *inpHash;
if(d_validNonceTable == NULL)
{
// keine Nonce-Liste
inpHash = &inpHashes[id<<4];
}else
{
// Nonce-Liste verfügbar
int nonce = d_validNonceTable[id] - startNounce;
inpHash = &inpHashes[nonce<<4];
}
value = (*testFunc)(inpHash);
}else
{
value = 0;
}
}else
{
value = data[id];
}

__syncthreads();

// Now accumulate in log steps up the chain
// compute sums, with another thread's value who is
Expand Down Expand Up @@ -177,79 +190,163 @@ __global__ void jackpot_compactTest_gpu_ADD(uint32_t *data, uint32_t *partial_su
}

// Der Scatter
__global__ void jackpot_compactTest_gpu_SCATTER(uint32_t *data, uint32_t *valid, uint32_t *sum, uint32_t *outp)
__global__ void jackpot_compactTest_gpu_SCATTER(uint32_t *sum, uint32_t *outp, cuda_compactTestFunction_t testFunc, int threads=0, uint32_t startNounce=0, uint32_t *inpHashes=NULL, uint32_t *d_validNonceTable=NULL)
{
int id = ((blockIdx.x * blockDim.x) + threadIdx.x);
if( valid[id] )
uint32_t actNounce = id;
uint32_t value;
if (id < threads)
{
// uint32_t nounce = startNounce + id;
uint32_t *inpHash;
if(d_validNonceTable == NULL)
{
// keine Nonce-Liste
inpHash = &inpHashes[id<<4];
}else
{
// Nonce-Liste verfügbar
int nonce = d_validNonceTable[id] - startNounce;
actNounce = nonce;
inpHash = &inpHashes[nonce<<4];
}

value = (*testFunc)(inpHash);
}else
{
value = 0;
}

if( value )
{
int idx = sum[id];
if(idx > 0)
outp[idx-1] = data[id];
outp[idx-1] = startNounce + actNounce;
}
}

__host__ static uint32_t jackpot_compactTest_roundUpExp(uint32_t val)
{
if(val == 0)
return 0;

uint32_t mask = 0x80000000;
while( (val & mask) == 0 ) mask = mask >> 1;

if( (val & (~mask)) != 0 )
return mask << 1;

return mask;
}

__host__ void jackpot_compactTest_cpu_singleCompaction(int thr_id, int threads, uint32_t *nrm,
uint32_t *d_nonces1, cuda_compactTestFunction_t function,
uint32_t startNounce, uint32_t *inpHashes, uint32_t *d_validNonceTable)
{
int orgThreads = threads;
threads = (int)jackpot_compactTest_roundUpExp((uint32_t)threads);
// threadsPerBlock ausrechnen
int blockSize = 256;
int nSummen = threads / blockSize;

int thr1 = (threads+blockSize-1) / blockSize;
int thr2 = threads / (blockSize*blockSize);
int blockSize2 = (nSummen < blockSize) ? nSummen : blockSize;
int thr3 = (nSummen + blockSize2-1) / blockSize2;

bool callThrid = (thr2 > 0) ? true : false;

// Erster Initialscan
jackpot_compactTest_gpu_SCAN<<<thr1,blockSize, 32*sizeof(uint32_t)>>>(
d_tempBranch1Nonces[thr_id], 32, d_partSum[0][thr_id], function, orgThreads, startNounce, inpHashes, d_validNonceTable);

// weitere Scans
if(callThrid)
{
jackpot_compactTest_gpu_SCAN<<<thr2,blockSize, 32*sizeof(uint32_t)>>>(d_partSum[0][thr_id], 32, d_partSum[1][thr_id]);
jackpot_compactTest_gpu_SCAN<<<1, thr2, 32*sizeof(uint32_t)>>>(d_partSum[1][thr_id], (thr2>32) ? 32 : thr2);
}else
{
jackpot_compactTest_gpu_SCAN<<<thr3,blockSize2, 32*sizeof(uint32_t)>>>(d_partSum[0][thr_id], (blockSize2>32) ? 32 : blockSize2);
}

// Sync + Anzahl merken
cudaStreamSynchronize(NULL);

if(callThrid)
cudaMemcpy(nrm, &(d_partSum[1][thr_id])[thr2-1], sizeof(uint32_t), cudaMemcpyDeviceToHost);
else
cudaMemcpy(nrm, &(d_partSum[0][thr_id])[nSummen-1], sizeof(uint32_t), cudaMemcpyDeviceToHost);


// Addieren
if(callThrid)
{
jackpot_compactTest_gpu_ADD<<<thr2-1, blockSize>>>(d_partSum[0][thr_id]+blockSize, d_partSum[1][thr_id], blockSize*thr2);
}
jackpot_compactTest_gpu_ADD<<<thr1-1, blockSize>>>(d_tempBranch1Nonces[thr_id]+blockSize, d_partSum[0][thr_id], threads);

// Scatter
jackpot_compactTest_gpu_SCATTER<<<thr1,blockSize,0>>>(d_tempBranch1Nonces[thr_id], d_nonces1,
function, orgThreads, startNounce, inpHashes, d_validNonceTable);

// Sync
cudaStreamSynchronize(NULL);
}

////// ACHTUNG: Diese funktion geht aktuell nur mit threads > 65536 (Am besten 256 * 1024 oder 256*2048)
__host__ void jackpot_compactTest_cpu_dualCompaction(int thr_id, int threads, size_t *nrm,
uint32_t *d_nonces1, uint32_t *d_nonces2)
__host__ void jackpot_compactTest_cpu_dualCompaction(int thr_id, int threads, uint32_t *nrm,
uint32_t *d_nonces1, uint32_t *d_nonces2,
uint32_t startNounce, uint32_t *inpHashes, uint32_t *d_validNonceTable)
{
jackpot_compactTest_cpu_singleCompaction(thr_id, threads, &nrm[0], d_nonces1, h_JackpotTrueFunction[thr_id], startNounce, inpHashes, d_validNonceTable);
jackpot_compactTest_cpu_singleCompaction(thr_id, threads, &nrm[1], d_nonces2, h_JackpotFalseFunction[thr_id], startNounce, inpHashes, d_validNonceTable);

/*
// threadsPerBlock ausrechnen
int blockSize = 256;
int thr1 = threads / blockSize;
int thr2 = threads / (blockSize*blockSize);
// 1
jackpot_compactTest_gpu_SCAN<<<thr1,blockSize, 8*sizeof(uint32_t)>>>(d_tempBranch1Nonces[thr_id], 32, d_partSum1[thr_id]);
jackpot_compactTest_gpu_SCAN<<<thr2,blockSize, 8*sizeof(uint32_t)>>>(d_partSum1[thr_id], 32, d_partSum2[thr_id]);
jackpot_compactTest_gpu_SCAN<<<1, thr2, 8*sizeof(uint32_t)>>>(d_partSum2[thr_id], (thr2>32) ? 32 : thr2);
jackpot_compactTest_gpu_SCAN<<<thr1,blockSize, 32*sizeof(uint32_t)>>>(d_tempBranch1Nonces[thr_id], 32, d_partSum1[thr_id], h_JackpotTrueFunction[thr_id], threads, startNounce, inpHashes);
jackpot_compactTest_gpu_SCAN<<<thr2,blockSize, 32*sizeof(uint32_t)>>>(d_partSum1[thr_id], 32, d_partSum2[thr_id]);
jackpot_compactTest_gpu_SCAN<<<1, thr2, 32*sizeof(uint32_t)>>>(d_partSum2[thr_id], (thr2>32) ? 32 : thr2);
cudaStreamSynchronize(NULL);
cudaMemcpy(&nrm[0], &(d_partSum2[thr_id])[thr2-1], sizeof(uint32_t), cudaMemcpyDeviceToHost);
jackpot_compactTest_gpu_ADD<<<thr2-1, blockSize>>>(d_partSum1[thr_id]+blockSize, d_partSum2[thr_id], blockSize*thr2);
jackpot_compactTest_gpu_ADD<<<thr1-1, blockSize>>>(d_tempBranch1Nonces[thr_id]+blockSize, d_partSum1[thr_id], threads);
// 2
jackpot_compactTest_gpu_SCAN<<<thr1,blockSize, 8*sizeof(uint32_t)>>>(d_tempBranch2Nonces[thr_id], 32, d_partSum1[thr_id]);
jackpot_compactTest_gpu_SCAN<<<thr2,blockSize, 8*sizeof(uint32_t)>>>(d_partSum1[thr_id], 32, d_partSum2[thr_id]);
jackpot_compactTest_gpu_SCAN<<<1, thr2, 8*sizeof(uint32_t)>>>(d_partSum2[thr_id], (thr2>32) ? 32 : thr2);
jackpot_compactTest_gpu_SCAN<<<thr1,blockSize, 32*sizeof(uint32_t)>>>(d_tempBranch2Nonces[thr_id], 32, d_partSum1[thr_id], h_JackpotFalseFunction[thr_id], threads, startNounce, inpHashes);
jackpot_compactTest_gpu_SCAN<<<thr2,blockSize, 32*sizeof(uint32_t)>>>(d_partSum1[thr_id], 32, d_partSum2[thr_id]);
jackpot_compactTest_gpu_SCAN<<<1, thr2, 32*sizeof(uint32_t)>>>(d_partSum2[thr_id], (thr2>32) ? 32 : thr2);
cudaStreamSynchronize(NULL);
cudaMemcpy(&nrm[1], &(d_partSum2[thr_id])[thr2-1], sizeof(uint32_t), cudaMemcpyDeviceToHost);
jackpot_compactTest_gpu_ADD<<<thr2-1, blockSize>>>(d_partSum1[thr_id]+blockSize, d_partSum2[thr_id], blockSize*thr2);
jackpot_compactTest_gpu_ADD<<<thr1-1, blockSize>>>(d_tempBranch2Nonces[thr_id]+blockSize, d_partSum1[thr_id], threads);
// Hier ist noch eine Besonderheit: in d_tempBranch1Nonces sind die element von 1...nrm1 die Interessanten
// Schritt 3: Scatter
jackpot_compactTest_gpu_SCATTER<<<thr1,blockSize,0>>>(d_tempBranchAllNonces[thr_id], d_validTemp1[thr_id], d_tempBranch1Nonces[thr_id], d_nonces1);
jackpot_compactTest_gpu_SCATTER<<<thr1,blockSize,0>>>(d_tempBranchAllNonces[thr_id], d_validTemp2[thr_id], d_tempBranch2Nonces[thr_id], d_nonces2);
jackpot_compactTest_gpu_SCATTER<<<thr1,blockSize,0>>>(d_tempBranch1Nonces[thr_id], d_nonces1, h_JackpotTrueFunction[thr_id], threads, startNounce, inpHashes);
jackpot_compactTest_gpu_SCATTER<<<thr1,blockSize,0>>>(d_tempBranch2Nonces[thr_id], d_nonces2, h_JackpotFalseFunction[thr_id], threads, startNounce, inpHashes);
cudaStreamSynchronize(NULL);
*/
}

__host__ void jackpot_compactTest_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *inpHashes,
__host__ void jackpot_compactTest_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *inpHashes, uint32_t *d_validNonceTable,
uint32_t *d_nonces1, size_t *nrm1,
uint32_t *d_nonces2, size_t *nrm2,
int order)
{
// Compute 3.x und 5.x Geräte am besten mit 768 Threads ansteuern,
// alle anderen mit 512 Threads.
//int threadsperblock = (props[thr_id].major >= 3) ? 768 : 512;
int threadsperblock = 256;

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

size_t shared_size = 0;

// fprintf(stderr, "threads=%d, %d blocks, %d threads per block, %d bytes shared\n", threads, grid.x, block.x, shared_size);

// Schritt 1: Prüfen der Bedingung und Speicherung in d_tempBranch1/2Nonces
jackpot_compactTest_gpu_TEST_64<<<grid, block, shared_size>>>(threads, startNounce, inpHashes, d_tempBranchAllNonces[thr_id],
d_tempBranch1Nonces[thr_id], d_tempBranch2Nonces[thr_id],
d_validTemp1[thr_id], d_validTemp2[thr_id]);
// Wenn validNonceTable genutzt wird, dann werden auch nur die Nonces betrachtet, die dort enthalten sind
// "threads" ist in diesem Fall auf die Länge dieses Array's zu setzen!

// Strategisches Sleep Kommando zur Senkung der CPU Last
jackpot_compactTest_cpu_dualCompaction(thr_id, threads,
h_numValid[thr_id], d_nonces1, d_nonces2);
h_numValid[thr_id], d_nonces1, d_nonces2,
startNounce, inpHashes, d_validNonceTable);

cudaStreamSynchronize(NULL); // Das original braucht zwar etwas CPU-Last, ist an dieser Stelle aber evtl besser
*nrm1 = h_numValid[thr_id][0];
*nrm2 = h_numValid[thr_id][1];
*nrm1 = (size_t)h_numValid[thr_id][0];
*nrm2 = (size_t)h_numValid[thr_id][1];
}
Loading

0 comments on commit be044f3

Please sign in to comment.