private void cpuRun() {
int num_cores = Runtime.getRuntime().availableProcessors();
m_cpuWatch.start();
List<MatrixCpuThread> threads = new ArrayList<MatrixCpuThread>();
for (int i = 0; i < num_cores; ++i) {
MatrixCpuThread thread =
new MatrixCpuThread(
m_a, m_b, m_ccpu, i, m_blockSize, m_gridSize * m_blockIters, num_cores);
threads.add(thread);
}
for (int i = 0; i < num_cores; ++i) {
MatrixCpuThread thread = threads.get(i);
thread.join();
}
m_cpuWatch.stop();
System.out.println("avg cpu time: " + m_cpuWatch.getAverageTime() + " ms");
}
private void gpuRun() {
m_gpuWatch.start();
MatrixKernel matrix_kernel =
new MatrixKernel(m_a, m_b, m_cgpu, m_blockSize, m_gridSize, m_blockIters);
Rootbeer rootbeer = new Rootbeer();
rootbeer.setThreadConfig(1024, m_gridSize);
rootbeer.runAll(matrix_kernel);
m_gpuWatch.stop();
System.out.println("avg gpu time: " + m_gpuWatch.getAverageTime() + " ms");
/*
int sum = 0;
for(Calculation calc : matrix_kernel.m_calcs){
if(calc == null){
continue;
}
System.out.println(" calc:");
System.out.println(" k: "+calc.m_invalidIndexK);
System.out.println(" row: "+calc.m_invalidIndexRow);
System.out.println(" col: "+calc.m_invalidIndexCol);
System.out.println(" a_value: "+calc.m_invalidAValue);
System.out.println(" b_value: "+calc.m_invalidBValue);
System.out.println(" prev_a: "+calc.m_invalidPrevA);
System.out.println(" prev_b: "+calc.m_invalidPrevB);
System.out.println(" m: "+calc.m_invalidIndexM);
System.out.println(" sub_matrix_row: "+calc.m_invalidSubMatrixRow);
System.out.println(" sub_matrix_col: "+calc.m_invalidSubMatrixCol);
sum += (calc.m_invalidAValue * calc.m_invalidBValue);
}
System.out.println("SUM: "+sum);
*/
List<StatsRow> stats = rootbeer.getStats();
for (StatsRow row : stats) {
System.out.println(" StatsRow:");
System.out.println(" init time: " + row.getInitTime());
System.out.println(" serial time: " + row.getSerializationTime());
System.out.println(" exec time: " + row.getExecutionTime());
System.out.println(" deserial time: " + row.getDeserializationTime());
System.out.println(" num blocks: " + row.getNumBlocks());
System.out.println(" num threads: " + row.getNumThreads());
}
}
private void test(TestSerialization creator) {
int i = 0;
try {
Rootbeer rootbeer = new Rootbeer();
List<Kernel> known_good_items = creator.create();
List<Kernel> testing_items = creator.create();
Stopwatch watch = new Stopwatch();
watch.start();
rootbeer.runAll(testing_items);
m_passed = true;
watch.stop();
m_gpuTime = watch.elapsedTimeMillis();
watch.start();
for (i = 0; i < known_good_items.size(); ++i) {
Kernel known_good_item = known_good_items.get(i);
known_good_item.gpuMethod();
}
watch.stop();
m_cpuTime = watch.elapsedTimeMillis();
for (i = 0; i < known_good_items.size(); ++i) {
Kernel known_good_item = known_good_items.get(i);
Kernel testing_item = testing_items.get(i);
if (!creator.compare(known_good_item, testing_item)) {
m_message = "Compare failed at: " + i;
m_passed = false;
return;
}
}
} catch (Throwable ex) {
ex.printStackTrace();
m_message = "Exception thrown at index: " + i + "\n";
ByteArrayOutputStream os = new ByteArrayOutputStream();
PrintWriter writer = new PrintWriter(os);
ex.printStackTrace(writer);
writer.flush();
writer.close();
m_message += os.toString();
m_passed = false;
}
}