Examples with CLCommandQueue com.jogamp.opencl.CLCommandQueue used on opensource projects

Example 1 with CLCommandQueue

use of com.jogamp.opencl.CLCommandQueue in project ffx by mjschnie.

the class Complex3DOpenCL method run.

@Override
public void run() {
    CLContext context = null;
    try {
        // Choose a platform.
        CLPlatform[] platforms = CLPlatform.listCLPlatforms();
        CLPlatform platform = platforms[0];
        // Prefer NV
        try {
            for (CLPlatform p : platforms) {
                if (p.getICDSuffix().equals("NV")) {
                    platform = p;
                    break;
                }
            }
        } catch (Exception e) {
        // ignore.
        }
        logger.info(String.format("   Platform: %s", platform));
        // Choose a device.
        CLDevice[] devices = platform.listCLDevices(CLDevice.Type.ACCELERATOR, CLDevice.Type.GPU);
        CLDevice device = devices[0];
        for (CLDevice dev : devices) {
            if (dev.getVendor().startsWith("NV")) {
                device = dev;
                break;
            }
        }
        logger.info(String.format("   Device:   %s", device));
        // Initialize the OpenCL Context
        context = CLContext.create(device);
        CLCommandQueue queue = device.createCommandQueue();
        // Allocate memory on the device.
        int bufferSize = len * 2;
        clData = context.createDoubleBuffer(bufferSize, Mem.READ_WRITE);
        DoubleBuffer doubleBuffer = clData.getBuffer();
        int MB = 1024 * 1024;
        logger.info(String.format("   FFT data buffer        [direct: %b, write: %b, size: %d MB]", doubleBuffer.isDirect(), !doubleBuffer.isReadOnly(), clData.getCLSize() / MB));
        clRecip = context.createDoubleBuffer(len, Mem.READ_WRITE);
        doubleBuffer = clRecip.getBuffer();
        logger.info(String.format("   Reciprocal data buffer [direct: %b, write: %b, size: %d MB]", doubleBuffer.isDirect(), !doubleBuffer.isReadOnly(), clRecip.getCLSize() / MB));
        // Initialize the OpenCL FFT library.
        setup();
        int[] dims = { nX, nY, nZ };
        planHandle = createDefaultPlan(context, Complex3DOpenCL_DIMENSION.Complex3DOpenCL_3D, dims);
        // Initialize the Reciprocal Space Multitply Kernal
        URL source = getClass().getClassLoader().getResource("ffx/numerics/fft/VectorMultiply.cl");
        InputStream input = source.openStream();
        CLProgram program = context.createProgram(input).build();
        // Get a reference to the kernel function with the name 'VectorMultiply'
        CLKernel kernel = program.createCLKernel("VectorMultiply");
        int localWorkSize = Math.min(device.getMaxWorkGroupSize(), 128);
        int globalWorkSize = roundUp(localWorkSize, len);
        synchronized (this) {
            while (!free) {
                if (mode != null) {
                    switch(mode) {
                        case RECIP:
                            clRecip.getBuffer().put(recip).rewind();
                            queue.putWriteBuffer(clRecip, true);
                            break;
                        case FFT:
                            clData.getBuffer().rewind();
                            queue.putWriteBuffer(clData, true);
                            if (!transferOnly) {
                                executeTransform(Complex3DOpenCL_DIRECTION.FORWARD, queue, clData, clData);
                                queue.finish();
                            }
                            clData.getBuffer().rewind();
                            queue.putReadBuffer(clData, true);
                            clData.getBuffer().rewind();
                            queue.finish();
                            break;
                        case CONVOLUTION:
                            queue.putWriteBuffer(clData, true);
                            // Forward FFT
                            if (!transferOnly) {
                                // long time = -System.nanoTime();
                                executeTransform(Complex3DOpenCL_DIRECTION.FORWARD, queue, clData, clData);
                                // Reciprocal Space Multiply
                                kernel.rewind().putArgs(clData, clRecip).putArg(len);
                                queue.put1DRangeKernel(kernel, 0, globalWorkSize, localWorkSize);
                                queue.putBarrier();
                                // Backward FFT
                                executeTransform(Complex3DOpenCL_DIRECTION.BACKWARD, queue, clData, clData);
                            // time += System.nanoTime();
                            // logger.info(String.format(" Compute Time %6.3f sec", time * 1.0e-9));
                            }
                            queue.putReadBuffer(clData, true);
                            break;
                        case IFFT:
                            queue.putWriteBuffer(clData, true);
                            if (!transferOnly) {
                                executeTransform(Complex3DOpenCL_DIRECTION.BACKWARD, queue, clData, clData);
                            }
                            queue.putReadBuffer(clData, true);
                    }
                    // Reset the mode to null and notify the calling thread.
                    mode = null;
                    notify();
                }
                // The OpenCL thread will wait until it's notified again.
                try {
                    wait();
                } catch (InterruptedException e) {
                    logger.severe(e.toString());
                }
            }
            queue.finish();
            clData.release();
            clRecip.release();
            destroyPlan();
            teardown();
            dead = true;
            notify();
        }
    } catch (IOException e) {
        logger.warning(e.toString());
    } finally {
        if (context != null) {
            context.release();
        }
    }
    logger.info(" OpenCL FFT/convolution thread is done.");
}