// creates neural network; also calls function to initialize network data
NeuralNetwork makeneuralnetwork(int inputs, int hiddens, unsigned short skip, int fn, double *fnpars, double learnr, int bias,
double ho_low, double ho_high, double ih_low, double ih_high, double skip_low, double skip_high)
{
NeuralNetwork netwk;
if (bias)
{
netwk.hbias = generateRand(BIAS_LOW, BIAS_HIGH); // hidden node bias - imposed in this system that all hidden nodes have the same bias
netwk.obias = generateRand(BIAS_LOW, BIAS_HIGH); // output node bias
}
else
{
netwk.hbias = 0;
netwk.obias = 0;
}
netwk.activationFunction = fn;
// array of activation function parameters
netwk.fnpars = fnpars;
// lr_ih, lr_ho, lr_skip: learning rate for the input->hidden, hidden->output, skip layer weights
netwk.lr_ih = netwk.lr_ho = netwk.lr_skip = learnr;
netwk.data = makenetworkdat(inputs, hiddens, skip, ho_low, ho_high, ih_low, ih_high, skip_low, skip_high);
return netwk;
}
// Returns a rand number from the specified interval of absolute values with a random sign.
float generateRandomSignInterval(float start, float end)
{
float sign = 1.;
if( generateRand(-1., 1.) < 0)
sign = -1.;
return generateRand(sign*start, sign*end);
}
NetworkData makenetworkdat(int inputs, int hiddens, unsigned short skip, double ho_low, double ho_high,
double ih_low, double ih_high, double skip_low, double skip_high)
{
NetworkData dat;
int i, j;
// numbers of hidden, input, and output nodes; whether to use skip layer or not
dat.hiddenNodes = hiddens;
dat.inputNodes = inputs;
dat.outputNodes = 1;
dat.useSkipLayer = skip;
// values in the input layer
dat.inputLayer = calloc(inputs, sizeof(double));
// values in the hidden layer
dat.hiddenLayer = calloc(hiddens, sizeof(double));
for (i=0; i<hiddens; ++i)
{
dat.hiddenLayer[i] = 0.0;
}
// values in the output layer - only one node in output layer here
dat.outputLayer = 0.0;
//input->hidden layer weights
dat.weights_IH = calloc(hiddens, sizeof(double*));
//hidden->output layer weights
dat.weights_HO = calloc(hiddens, sizeof(double));
//input->output layer weights (the skip layer)
dat.weights_skip = calloc(inputs, sizeof(double));
// each hidden node has a weighted connection to an (the) output node, and to each input node
for (i=0; i<hiddens; ++i)
{
dat.weights_HO[i] = generateRand(ho_low, ho_high);
dat.weights_IH[i] = calloc(inputs, sizeof(double));
for(j=0; j < dat.inputNodes; j++)
dat.weights_IH[i][j] = generateRand(ih_low, ih_high);
}
// each input node has a connection to an (the) output node
if (skip)
for(j=0; j < dat.inputNodes; j++)
dat.weights_skip[j] = generateRand(skip_low, skip_high);
else
for(j=0; j < dat.inputNodes; j++)
dat.weights_skip[j] = 0;
dat.patternerr = 0.0;
return dat;
}
//return bool: if is multiple of maxNumThread
//if yes, info[0]: number of blocks, info[1] = maxNumThread
//if no, info[0]: number of blocks except of the last block, info[1]: number of thread in the last block
void testReduceImpl( int rLen, int OPERATOR, int numThreadPB , int numMaxBlock)
{
int _CPU_GPU=0;
int result=0;
int memSize = sizeof(Record)*rLen;
void * h_Rin;
HOST_MALLOC(h_Rin, memSize );
generateRand((Record *)h_Rin, TEST_MAX - 11111, rLen, 0 );
void* h_Rout;
unsigned int numResult = 0;
cl_mem d_Rin=NULL;
cl_mem d_Rout;
CL_MALLOC( &d_Rin, memSize );
cl_writebuffer( d_Rin, h_Rin, memSize,0);
numResult= CL_AggMaxOnly( d_Rin, rLen, &d_Rout, numThreadPB, numMaxBlock,_CPU_GPU);
HOST_MALLOC(h_Rout, sizeof(Record)*numResult );
cl_readbuffer( h_Rout, d_Rout, sizeof(Record)*numResult,_CPU_GPU);
//validateReduce((Record *)h_Rin, rLen,((Record *)h_Rout)[0].y, OPERATOR );
HOST_FREE( h_Rin );
HOST_FREE( h_Rout );
CL_FREE( d_Rin );
CL_FREE( d_Rout );
printf("testReduceFinish\n");
}
void testAggAfterGroupByImpl( int rLen, int OPERATOR, int numThread, int numBlock)
{
int _CPU_GPU=0;
int memSize = sizeof(Record)*rLen;
void* h_Rin;
void* h_Rout;
void* h_Sin;
int* h_startPos;
HOST_MALLOC( h_Rin, memSize );
HOST_MALLOC( h_Rout, memSize );
HOST_MALLOC( h_Sin, memSize );
generateRand((Record *) h_Rin, 50, rLen, 0 );
generateRand((Record *) h_Sin, TEST_MAX, rLen, 0 );
int numGroup = 0;
numGroup = CL_GroupBy((Record *)h_Rin, rLen,(Record *) h_Rout, &h_startPos, numThread, numBlock,_CPU_GPU);
validateGroupBy((Record*) h_Rin, rLen, (Record*)h_Rout,h_startPos, numGroup );
void* h_aggResults;
HOST_MALLOC(h_aggResults, sizeof(int)*numGroup );
switch(OPERATOR){
case REDUCE_MAX:
{
CL_agg_max_afterGroupBy((Record *)h_Rout,rLen,h_startPos,numGroup,(Record *)h_Sin,(int *)h_aggResults,numThread,_CPU_GPU);
break;
}
case REDUCE_MIN:
{
CL_agg_min_afterGroupBy((Record *)h_Rout,rLen,h_startPos,numGroup,(Record *)h_Sin,(int *)h_aggResults,numThread,_CPU_GPU);
break;
}
case REDUCE_SUM:
{
CL_agg_sum_afterGroupBy((Record *)h_Rout,rLen,h_startPos,numGroup,(Record *)h_Sin,(int *)h_aggResults,numThread,_CPU_GPU);
break;
}
case REDUCE_AVERAGE:
{
CL_agg_avg_afterGroupBy((Record *)h_Rout,rLen,h_startPos,numGroup,(Record *)h_Sin,(int *)h_aggResults,numThread,_CPU_GPU);
break;
}
}
validateAggAfterGroupBy((Record*) h_Rin, rLen, (int*)h_startPos, numGroup,(Record*) h_Sin, (int *)h_aggResults, OPERATOR);
}
/**
* Inserts random numbers between 0.0 and 0.1 as Weight values.
*/
int NetInit( PtNet Net )
{
int i, j, k;
/* Verifying possible errors: */
if (Net==NULL)
return -2;
/* Initializing Weights and Gradients: */
for( i=Net->Layers-2 ; i>=0 ; i-- )
for( j=0 ; j<=Net->NLayers[i+1] ; j++ )
for( k=0 ; k<=Net->NLayers[i] ; k++ )
{
/* Inserting random values for the Weights: */
Net->Weights[i][j][k] = generateRand()/Net->NLayers[i];
/*printf("Net->Weights: %lf\n", Net->Weights[i][j][k]);*/
/* Inserting 0.0 in Gradients: */
Net->Gradients[i][j][k] = 0.0;
}
/* Inserting 0.0 in the neurons's output values: */
for( i=0 ; i<Net->Layers ; i++ )
for ( j=0 ; j<Net->NLayers[i] ; j++ )
Net->Neurons[i][j] = generateRand();
return 0;
}
void testGroupByImpl( int rLen, int numThread, int numBlock)
{
int _CPU_GPU=0;
int memSize = sizeof(Record)*rLen;
void* h_Rin;
HOST_MALLOC(h_Rin, memSize );
void* h_Rout;
HOST_MALLOC(h_Rout, memSize );
generateRand((Record *)h_Rin, 64, rLen, 0 );
int* h_startPos;
int numGroup = 0;
//group by
numGroup=CL_GroupBy((Record *) h_Rin, rLen, (Record*) h_Rout, &h_startPos, numThread, numBlock,_CPU_GPU);
//copy back
validateGroupBy( (Record*)h_Rin, rLen, (Record*)h_Rout, h_startPos, numGroup );
HOST_FREE(h_startPos);
HOST_FREE( h_Rin );
HOST_FREE( h_Rout );
}