void initAligned()
{
weightsAligned.reset(connsCount);
srcConnsAligned.reset(connsCount);
tgtConnsAligned.reset(connsCount);
for(int i = 0; i < connsCount; i++)
{
weightsAligned[i] = weights[i];
srcConnsAligned[i] = srcConns[i];
tgtConnsAligned[i] = tgtConns[i];
}
endConnsAligned.reset(layersCount);
endNodesAligned.reset(layersCount);
for(int i = 0; i < layersCount; i++)
{
endConnsAligned[i] = endConns[i];
endNodesAligned[i] = endNodes[i];
}
outputsAligned.reset(outputsCount);
for(int i = 0; i < outputsCount; i++)
{
outputsAligned[i] = 0;
}
outputIdsAligned.reset(outputIds.size());
for(int i = 0; i < outputIds.size(); i++)
{
outputIdsAligned[i] = outputIds[i];
}
}
// Free the resources allocated during initialization
void cleanup() {
for(unsigned i = 0; i < num_devices; ++i) {
if(kernel && kernel[i]) {
clReleaseKernel(kernel[i]);
}
if(queue && queue[i]) {
clReleaseCommandQueue(queue[i]);
}
#if USE_SVM_API == 0
if(input_a_buf && input_a_buf[i]) {
clReleaseMemObject(input_a_buf[i]);
}
if(input_b_buf && input_b_buf[i]) {
clReleaseMemObject(input_b_buf[i]);
}
if(output_buf && output_buf[i]) {
clReleaseMemObject(output_buf[i]);
}
#else
if(input_a[i].get())
input_a[i].reset();
if(output[i].get())
output[i].reset();
#endif /* USE_SVM_API == 0 */
}
#if USE_SVM_API == 1
if(input_b.get())
input_b.reset();
#endif /* USE_SVM_API == 1 */
if(program) {
clReleaseProgram(program);
}
if(context) {
clReleaseContext(context);
}
}
// Initialize the data for the problem. Requires num_devices to be known.
void init_problem() {
if(num_devices == 0) {
checkError(-1, "No devices");
}
// Generate input matrices A and B. For matrix A, we divide up the host
// buffers so that the buffers are aligned for each device. The whole of
// matrix B is used by each device, so it does not need to be divided.
printf("Generating input matrices\n");
input_a.reset(num_devices);
output.reset(num_devices);
#if USE_SVM_API == 0
for(unsigned i = 0; i < num_devices; ++i) {
input_a[i].reset(rows_per_device[i] * A_width);
output[i].reset(rows_per_device[i] * C_width);
for(unsigned j = 0; j < rows_per_device[i] * A_width; ++j) {
input_a[i][j] = rand_float();
}
}
input_b.reset(B_height * B_width);
for(unsigned i = 0; i < B_height * B_width; ++i) {
input_b[i] = rand_float();
}
#else
for(unsigned i = 0; i < num_devices; ++i) {
input_a[i].reset(context, rows_per_device[i] * A_width);
output[i].reset(context, rows_per_device[i] * C_width);
cl_int status;
status = clEnqueueSVMMap(queue[i], CL_TRUE, CL_MAP_WRITE,
(void *)input_a[i], rows_per_device[i] * A_width * sizeof(float), 0, NULL, NULL);
checkError(status, "Failed to map input A");
for(unsigned j = 0; j < rows_per_device[i] * A_width; ++j) {
input_a[i][j] = rand_float();
}
status = clEnqueueSVMUnmap(queue[i], (void *)input_a[i], 0, NULL, NULL);
checkError(status, "Failed to unmap input A");
}
input_b.reset(context, B_height * B_width);
cl_int status;
for (unsigned i = 0; i < num_devices; ++i) {
status = clEnqueueSVMMap(queue[i], CL_TRUE, CL_MAP_WRITE,
(void *)input_b, B_height * B_width * sizeof(float), 0, NULL, NULL);
checkError(status, "Failed to map input B");
}
for(unsigned i = 0; i < B_height * B_width; ++i) {
input_b[i] = rand_float();
}
for (unsigned i = 0; i < num_devices; ++i) {
status = clEnqueueSVMUnmap(queue[i], (void *)input_b, 0, NULL, NULL);
checkError(status, "Failed to unmap input B");
}
#endif /* USE_SVM_API == 0 */
}
bool initProblem()
{
string genome_path("K:\\nn\\opencl\\altera\\fnn\\simple_genome.xml");
ifstream infile(genome_path);
if (!infile.is_open()) {
printf("Failed to open genome file.");
return false;
}
string line;
getline(infile, line);
nodesCount = atoi(line.c_str());
getline(infile, line);
inputsCount = atoi(line.c_str());
getline(infile, line);
outputsCount = atoi(line.c_str());
getline(infile, line);
layersCount = atoi(line.c_str());
getline(infile, line);
string buf; // Have a buffer string
stringstream ssWeights(line); // Insert the string into a stream
while (ssWeights >> buf)
weights.push_back(strtod(buf.c_str(), 0));
connsCount = weights.size();
// source connections
getline(infile, line);
stringstream ssSrcConns(line);
while (ssSrcConns >> buf)
srcConns.push_back(atoi(buf.c_str()));
// target connections
getline(infile, line);
stringstream ssTgtConns(line);
while (ssTgtConns >> buf)
tgtConns.push_back(atoi(buf.c_str()));
// layers end connections
getline(infile, line);
stringstream ssEndConns(line);
while (ssEndConns >> buf)
endConns.push_back(atoi(buf.c_str()));
// layers end nodes
getline(infile, line);
stringstream ssEndNodes(line);
while (ssEndNodes >> buf)
endNodes.push_back(atoi(buf.c_str()));
// output ids
getline(infile, line);
stringstream ssOut(line);
while (ssOut >> buf)
outputIds.push_back(atoi(buf.c_str()));
string validation_set("K:\\nn\\opencl\\altera\\fnn\\cross-validate.txt");
ifstream vs_infile(validation_set);
if (!vs_infile.is_open()) {
printf("Failed to open validation dataset.");
return false;
}
float feature;
while(getline(vs_infile, line)) {
stringstream ssIrisDatasetLine(line);
vector<float> row;
while (ssIrisDatasetLine >> feature)
{
//printf((line + "\n").c_str());
row.push_back(feature);
//printf("Add item %f from row %s \n", feature, line.c_str());
if(ssIrisDatasetLine.peek() == ';')
{
ssIrisDatasetLine.ignore();
}
}
validationSet.push_back(row);
}
if(validationSet.size() == 0)
{
printf("Failed to parse validation dataset.");
return false;
}
//printf("First row size %llu", validationSet[0].size());
inputValuesAligned.reset(validationSet[0].size() - 3);
for(int i = 0; i < 4; i++)
{
inputValuesAligned[i] = validationSet[0][i];
}
initAligned();
return true;
}
