@SuppressWarnings("unchecked")
public void startup(
float[][] whiteTemplates,
float[][] blackTemplates,
int[] templateNumIndices,
int[] templateIndicesOffsets,
int minTemplateWidth,
int maxTemplateWidth,
int maxSequenceLength,
int totalTemplateNumIndices) {
this.templateNumIndices = templateNumIndices;
this.templateIndicesOffsets = templateIndicesOffsets;
this.maxTemplateWidth = maxTemplateWidth;
this.minTemplateWidth = minTemplateWidth;
// Allocate the device input data
int extendedMaxSeqLength =
(blockSizeX * rollX) * (int) Math.ceil(((double) maxSequenceLength) / (blockSizeX * rollX));
this.d_Ow =
context.createFloatBuffer(
Usage.Input,
(extendedMaxSeqLength + maxTemplateWidth - 1) * CharacterTemplate.LINE_HEIGHT);
this.d_Ob =
context.createFloatBuffer(
Usage.Input,
(extendedMaxSeqLength + maxTemplateWidth - 1) * CharacterTemplate.LINE_HEIGHT);
this.d_scores =
context.createFloatBuffer(Usage.Output, maxSequenceLength * totalTemplateNumIndices);
int numTemplateWidths = (maxTemplateWidth - minTemplateWidth) + 1;
this.d_Tw = new CLBuffer[numTemplateWidths];
this.d_Tb = new CLBuffer[numTemplateWidths];
for (int tw = minTemplateWidth; tw <= maxTemplateWidth; ++tw) {
if (templateNumIndices[tw - minTemplateWidth] > 0) {
d_Tw[tw - minTemplateWidth] =
context.createFloatBuffer(Usage.Input, whiteTemplates[tw - minTemplateWidth].length);
d_Tw[tw - minTemplateWidth].write(
queue,
pc.capture(Pointer.pointerToFloats(whiteTemplates[tw - minTemplateWidth])),
false);
d_Tb[tw - minTemplateWidth] =
context.createFloatBuffer(Usage.Input, whiteTemplates[tw - minTemplateWidth].length);
d_Tb[tw - minTemplateWidth].write(
queue,
pc.capture(Pointer.pointerToFloats(blackTemplates[tw - minTemplateWidth])),
false);
}
}
}
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 static synchronized boolean addVec(
CLDevice.Type clType, int[] vecC, int[] vecA, int[] vecB) {
try {
if ((Logger.getLogMask() & Level.DEFAULT.DEBUG.getLevel().getValue())
== Level.DEFAULT.DEBUG.getLevel().getValue()) {
Logger.logDebug(CLAZZ, "addVec() vecA: " + Convert.toString(vecA));
Logger.logDebug(CLAZZ, "addVec() vecB: " + Convert.toString(vecB));
}
/** * Initialisiere OpenCL-Objekte ** */
initCL(clType);
/** * Ausgabe von Informationen ueber gewaehltes OpenCL-Device ** */
Logger.logInfo(CLAZZ, "max compute units: " + devices.get(0).getMaxComputeUnits());
Logger.logInfo(CLAZZ, "max work group sizes: " + devices.get(0).getMaxWorkGroupSize());
Logger.logInfo(
CLAZZ, "max global mem size (KB): " + devices.get(0).getGlobalMemSize() / 1024);
Logger.logInfo(CLAZZ, "max local mem size (KB): " + devices.get(0).getLocalMemSize() / 1024);
/** * Erstellen und Vorbereiten der Daten ** */
IntBuffer tmpBuffer =
ByteBuffer.allocateDirect(vecA.length * Integer.SIZE)
.order(context.getByteOrder())
.asIntBuffer();
tmpBuffer.put(vecA);
CLBuffer<IntBuffer> aBuffer = context.createBuffer(CLMem.Usage.Input, tmpBuffer, true);
tmpBuffer.clear();
tmpBuffer.put(vecB);
CLBuffer<IntBuffer> bBuffer = context.createBuffer(CLMem.Usage.Input, tmpBuffer, true);
CLBuffer<IntBuffer> cBuffer =
context.createBuffer(CLMem.Usage.Output, vecC.length, IntBuffer.class);
/** * Kernel-Argumente setzen ** */
kernel.setArg(0, cBuffer);
kernel.setArg(1, aBuffer);
kernel.setArg(2, bBuffer);
kernel.setArg(3, vecC.length);
/** * Kernel ausfuehren und auf Abarbeitung warten ** */
CLEvent event = kernel.enqueueNDRange(cmdQ, new int[] {vecC.length}, new CLEvent[0]);
event.waitFor();
cmdQ.finish();
/** * Daten vom OpenCL-Device holen ** */
cBuffer.read(cmdQ, tmpBuffer, true, new CLEvent[0]);
tmpBuffer.clear();
tmpBuffer.get(vecC);
if ((Logger.getLogMask() & Level.DEFAULT.DEBUG.getLevel().getValue())
== Level.DEFAULT.DEBUG.getLevel().getValue()) {
Logger.logDebug(CLAZZ, "addVec() vecC: " + Convert.toString(vecC));
}
} catch (CLException err) {
Logger.logError(CLAZZ, "OpenCL error:\n" + err.getMessage() + "():" + err.getCode());
err.printStackTrace();
return EXIT_FAILURE;
} catch (Exception err) {
Logger.logError(CLAZZ, "Error:\n" + err.getMessage() + "()");
err.printStackTrace();
return EXIT_FAILURE;
}
return EXIT_SUCCESS;
}
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;
}
