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 void compute(
final float[] scores,
final float[] whiteObservations,
final float[] blackObservations,
final int sequenceLength) {
int gridSizeX = (int) Math.ceil(((double) sequenceLength) / (blockSizeX * rollX));
int extendedSeqLength = gridSizeX * (blockSizeX * rollX);
d_Ow.write(
queue,
pc.capture(
Pointer.pointerToFloats(
CudaUtil.extendWithZeros(
whiteObservations,
(extendedSeqLength + maxTemplateWidth - 1) * CharacterTemplate.LINE_HEIGHT))),
false);
d_Ob.write(
queue,
pc.capture(
Pointer.pointerToFloats(
CudaUtil.extendWithZeros(
blackObservations,
(extendedSeqLength + maxTemplateWidth - 1) * CharacterTemplate.LINE_HEIGHT))),
false);
queue.enqueueBarrier();
for (int tw = minTemplateWidth; tw <= maxTemplateWidth; ++tw) {
if (templateNumIndices[tw - minTemplateWidth] > 0) {
int gridSizeY =
(int) Math.ceil(((double) templateNumIndices[tw - minTemplateWidth]) / blockSizeY);
CLKernel computeKernel = program.createKernel("compute_emissions_" + tw);
computeKernel.setArgs(
templateIndicesOffsets[tw - minTemplateWidth] * sequenceLength,
sequenceLength,
templateNumIndices[tw - minTemplateWidth],
d_Tw[tw - minTemplateWidth],
d_Tb[tw - minTemplateWidth],
d_Ow,
d_Ob,
d_scores);
computeKernel.enqueueNDRange(
queue,
new int[] {gridSizeX * blockSizeX, gridSizeY * blockSizeY},
new int[] {blockSizeX, blockSizeY});
}
}
queue.enqueueBarrier();
d_scores.read(queue).getFloats(scores);
}
