private static void initCL(CLDevice.Type clType) throws Exception {
/** * Hole OpenCL-Plattformen z.B. AMD APP, NVIDIA CUDA ** */
platforms = JavaCL.listPlatforms();
/** * Hole OpenCL-Device des geforderten Typs z.B. GPU, CPU ** */
EnumSet<CLDevice.Type> types = EnumSet.of(clType);
devices = new ArrayList<CLDevice>();
CLDevice[] devTmp;
for (CLPlatform platform : platforms) {
devTmp = platform.listDevices(types, true);
devices.addAll(Arrays.asList(devTmp));
}
/** * Erstelle OpenCL-Context und CommandQueue ** */
devTmp = new CLDevice[devices.size()];
context = JavaCL.createContext(null, devices.toArray(devTmp));
cmdQ = context.createDefaultQueue(QueueProperties.ProfilingEnable);
/** * OpenCL-Quellcode einlesen ** */
String src = readFile(KERNEL_PATH);
// String src = KERNEL_SRC;
/** * OpenCL-Programm aus Quellcode erstellen ** */
program = context.createProgram(src);
try {
program.build();
} catch (CLBuildException err) {
Logger.logError(CLAZZ, "Build log for \"" + devices.get(0) + "\n" + err.getMessage());
throw err;
}
/** * OpenCL-Kernel laden ** */
kernel = program.createKernel("addVec");
}
public OpenCLInnerLoop(int numThreads) {
// choose device
List<CLDevice> devices = new ArrayList<CLDevice>();
System.out.println();
for (CLPlatform platform : JavaCL.listPlatforms()) {
for (CLDevice device : platform.listAllDevices(true)) {
System.out.println("Type: " + device.getType());
System.out.println("Vendor: " + device.getVendor());
System.out.println("Name: " + device.getName());
System.out.println("Compute units: " + device.getMaxComputeUnits());
System.out.println("Global mem: " + device.getGlobalMemSize() / 1e6 + "MB");
System.out.println("Driver version: " + device.getDriverVersion());
System.out.println();
devices.add(device);
}
}
if (context == null) {
for (CLDevice device : devices) {
if (device.getVendor().toLowerCase().contains("intel")
&& device.getType().contains(CLDevice.Type.GPU)
&& device.getMaxComputeUnits() >= 140
&& device.getGlobalMemSize() > 512e6) {
this.context = JavaCL.createContext(null, device);
}
}
}
// if (context == null) {
// for (CLDevice device : devices) {
// if (device.getVendor().toLowerCase().contains("nvidia") &&
// device.getType().contains(CLDevice.Type.GPU) && device.getMaxComputeUnits() >= 8 &&
// device.getGlobalMemSize() > 1e9 &&
// !device.getPlatform().getName().toLowerCase().contains("apple")) {
// this.context = JavaCL.createContext(null, device);
// }
// }
// }
if (context == null) {
this.context = JavaCL.createBestContext(DeviceFeature.CPU);
}
if (context.getDevices()[0].getType().contains(CLDevice.Type.GPU)
&& context.getDevices()[0].getVendor().toLowerCase().contains("nvidia")) {
this.blockSizeX = NVIDIA_GPU_BLOCK_SIZE_X;
this.rollX = NVIDIA_GPU_ROLL_X;
this.blockSizeY = NVIDIA_GPU_BLOCK_SIZE_Y;
} else if (context.getDevices()[0].getType().contains(CLDevice.Type.GPU)
&& context.getDevices()[0].getVendor().toLowerCase().contains("intel")) {
this.blockSizeX = INTEL_GPU_BLOCK_SIZE_X;
this.rollX = INTEL_GPU_ROLL_X;
this.blockSizeY = INTEL_GPU_BLOCK_SIZE_Y;
} else if (context.getDevices()[0].getType().contains(CLDevice.Type.CPU)) {
this.blockSizeX = CPU_BLOCK_SIZE_X;
this.rollX = CPU_ROLL_X;
this.blockSizeY = CPU_BLOCK_SIZE_Y;
}
System.out.println("Using context:");
System.out.println(context.toString());
System.out.println("Block size x: " + blockSizeX);
System.out.println("Roll x: " + rollX);
System.out.println("Block size y: " + blockSizeY);
this.context.setCacheBinaries(false);
this.queue = context.createDefaultQueue();
this.program = context.createProgram(kernelSrc());
this.program.addBuildOption("-cl-fast-relaxed-math");
this.program.addBuildOption("-cl-mad-enable");
this.program.addBuildOption("-cl-unsafe-math-optimizations");
this.program.addBuildOption("-cl-fast-relaxed-math");
this.program.addBuildOption("-cl-single-precision-constant");
this.program.build();
this.pc = new PointerCapturer();
this.numThreads = numThreads;
}