package com.mbien.opencl; import com.sun.gluegen.runtime.BufferFactory; import com.sun.gluegen.runtime.CPU; import com.sun.gluegen.runtime.PointerBuffer; import java.nio.ByteBuffer; import java.nio.ByteOrder; import java.nio.IntBuffer; import java.util.Collections; import java.util.HashMap; import java.util.HashSet; import java.util.Set; import java.util.Map; import static com.mbien.opencl.CLException.*; import static com.mbien.opencl.CL.*; /** * Represents a OpenCL program executed on one or more {@link CLDevice}s. * A CLProgram must be build using one of the build methods before creating {@link CLKernel}s. * @see CLContext#createProgram(java.io.InputStream) * @see CLContext#createProgram(java.lang.String) * @see CLContext#createProgram(java.util.Map) * @author Michael Bien */ public class CLProgram extends CLObject implements CLResource { private final Set kernels; private Map buildStatusMap; private boolean executable; private boolean released; private CLProgram(CLContext context, long id) { super(context, id); this.kernels = new HashSet(); } static CLProgram create(CLContext context, String src) { IntBuffer status = BufferFactory.newDirectByteBuffer(4).asIntBuffer(); // Create the program long id = context.cl.clCreateProgramWithSource(context.ID, 1, new String[] {src}, PointerBuffer.allocateDirect(1).put(src.length()), status); checkForError(status.get(), "can not create program with source"); return new CLProgram(context, id); } static CLProgram create(CLContext context, Map binaries) { PointerBuffer devices = PointerBuffer.allocateDirect(binaries.size()); PointerBuffer lengths = PointerBuffer.allocateDirect(binaries.size()); ByteBuffer[] codeBuffers = new ByteBuffer[binaries.size()]; int i = 0; Set keys = binaries.keySet(); for (CLDevice device : keys) { byte[] bytes = binaries.get(device); devices.put(device.ID); lengths.put(bytes.length); codeBuffers[i] = BufferFactory.newDirectByteBuffer(bytes.length).put(bytes); codeBuffers[i].rewind(); i++; } devices.rewind(); lengths.rewind(); IntBuffer err = BufferFactory.newDirectByteBuffer(4).asIntBuffer(); // IntBuffer status = BufferFactory.newDirectByteBuffer(binaries.size()*4).asIntBuffer(); long id = context.cl.clCreateProgramWithBinary(context.ID, devices.capacity(), devices, lengths, codeBuffers, /*status*/null, err); // while(status.remaining() != 0) { // checkForError(status.get(), "unable to load binaries on all devices"); // } checkForError(err.get(), "can not create program with binary"); return new CLProgram(context, id); } private void initBuildStatus() { if(buildStatusMap == null) { Map map = new HashMap(); CLDevice[] devices = getCLDevices(); for (CLDevice device : devices) { Status status = getBuildStatus(device); if(status == Status.BUILD_SUCCESS) { executable = true; } map.put(device, status); } this.buildStatusMap = Collections.unmodifiableMap(map); } } private String getBuildInfoString(long device, int flag) { if(released) { return ""; } PointerBuffer pb = PointerBuffer.allocateDirect(1); int ret = cl.clGetProgramBuildInfo(ID, device, flag, 0, null, pb); checkForError(ret, "on clGetProgramBuildInfo"); ByteBuffer bb = ByteBuffer.allocateDirect((int)pb.get(0)).order(ByteOrder.nativeOrder()); ret = cl.clGetProgramBuildInfo(ID, device, flag, bb.capacity(), bb, null); checkForError(ret, "on clGetProgramBuildInfo"); return CLUtils.clString2JavaString(bb, (int)pb.get(0)); } private String getProgramInfoString(int flag) { if(released) { return ""; } PointerBuffer pb = PointerBuffer.allocateDirect(1); int ret = cl.clGetProgramInfo(ID, flag, 0, null, pb); checkForError(ret, "on clGetProgramInfo"); ByteBuffer bb = ByteBuffer.allocateDirect((int)pb.get(0)).order(ByteOrder.nativeOrder()); ret = cl.clGetProgramInfo(ID, flag, bb.capacity(), bb, null); checkForError(ret, "on clGetProgramInfo"); return CLUtils.clString2JavaString(bb, (int)pb.get(0)); } // private int getProgramInfoInt(int flag) { // // ByteBuffer bb = ByteBuffer.allocateDirect(4).order(ByteOrder.nativeOrder()); // // int ret = cl.clGetProgramInfo(programID, flag, bb.capacity(), bb, null, 0); // checkForError(ret, ""); // // return bb.getInt(); // } private int getBuildInfoInt(long device, int flag) { ByteBuffer bb = ByteBuffer.allocateDirect(4).order(ByteOrder.nativeOrder()); int ret = cl.clGetProgramBuildInfo(ID, device, flag, bb.capacity(), bb, null); checkForError(ret, "error on clGetProgramBuildInfo"); return bb.getInt(); } /** * Builds this program for all devices associated with the context. * @return this */ public CLProgram build() { build(null, (CLDevice[])null); return this; } /** * Builds this program for the given devices. * @return this * @param devices A list of devices this program should be build on or null for all devices of its context. */ public CLProgram build(CLDevice... devices) { build(null, devices); return this; } /** * Builds this program for all devices associated with the context using the specified build options. * @see CompilerOptions * @return this */ public CLProgram build(String options) { build(options, (CLDevice[])null); return this; } /** * Builds this program for all devices associated with the context using the specified build options. * @see CompilerOptions * @return this */ public CLProgram build(String... options) { build(optionsOf(options), (CLDevice[])null); return this; } /** * Builds this program for the given devices and with the specified build options. In case this program was * already built and there are kernels associated with this program they will be released first before rebuild. * @see CompilerOptions * @return this * @param devices A list of devices this program should be build on or null for all devices of its context. */ public CLProgram build(String options, CLDevice... devices) { if(!kernels.isEmpty()) { //No changes to the program executable are allowed while there are //kernel objects associated with a program object. releaseKernels(); } PointerBuffer deviceIDs = null; int count = 0; if(devices != null) { deviceIDs = PointerBuffer.allocateDirect(devices.length); for (int i = 0; i < devices.length; i++) { deviceIDs.put(i, devices[i].ID); } deviceIDs.rewind(); count = devices.length; } // invalidate build status buildStatusMap = null; executable = false; // Build the program int ret = cl.clBuildProgram(ID, count, deviceIDs, options, null, null); if(ret != CL_SUCCESS) { throw newException(ret, "\n"+getBuildLog()); } return this; } /** * Creates a kernel with the specified kernel name. */ public CLKernel createCLKernel(String kernelName) { if(released) { return null; } int[] err = new int[1]; long id = cl.clCreateKernel(ID, kernelName, err, 0); checkForError(err[0], "unable to create Kernel with name: "+kernelName); CLKernel kernel = new CLKernel(this, id); kernels.add(kernel); return kernel; } /** * Creates n instances of a kernel with the specified kernel name. */ /* public CLKernel[] createCLKernels(String kernelName, int instanceCount) { if(released) { return new CLKernel[0]; } CLKernel[] newKernels = new CLKernel[instanceCount]; for (int i = 0; i < newKernels.length; i++) { newKernels[i] = createCLKernel(kernelName); } return newKernels; } */ /** * Creates all kernels of this program and stores them a Map with the kernel name as key. */ public Map createCLKernels() { if(released) { return Collections.emptyMap(); } HashMap newKernels = new HashMap(); IntBuffer numKernels = BufferFactory.newDirectByteBuffer(4).asIntBuffer(); int ret = cl.clCreateKernelsInProgram(ID, 0, null, numKernels); checkForError(ret, "can not create kernels for program"); if(numKernels.get(0) > 0) { PointerBuffer kernelIDs = PointerBuffer.allocateDirect(numKernels.get(0)); ret = cl.clCreateKernelsInProgram(ID, kernelIDs.capacity(), kernelIDs, null); checkForError(ret, "can not create kernels for program"); for (int i = 0; i < kernelIDs.capacity(); i++) { CLKernel kernel = new CLKernel(this, kernelIDs.get(i)); kernels.add(kernel); newKernels.put(kernel.name, kernel); } }else{ initBuildStatus(); if(!isExecutable()) { // It is illegal to create kernels from a not executable program. // For consistency between AMD and NVIDIA drivers throw an exception at this point. throw newException(CL_INVALID_PROGRAM_EXECUTABLE, "can not initialize kernels, program is not executable. status: "+buildStatusMap); } } return newKernels; } void onKernelReleased(CLKernel kernel) { this.kernels.remove(kernel); } /** * Releases this program with its kernels. */ public void release() { releaseKernels(); executable = false; released = true; buildStatusMap = null; int ret = cl.clReleaseProgram(ID); context.onProgramReleased(this); checkForError(ret, "can not release program"); } private void releaseKernels() { if(!kernels.isEmpty()) { // copy to array to prevent concurrent modification exception CLKernel[] array = kernels.toArray(new CLKernel[kernels.size()]); for (CLKernel kernel : array) { kernel.release(); } } } /** * Returns all devices associated with this program. */ public CLDevice[] getCLDevices() { if(released) { return new CLDevice[0]; } PointerBuffer pb = PointerBuffer.allocateDirect(1); int ret = cl.clGetProgramInfo(ID, CL_PROGRAM_DEVICES, 0, null, pb); checkForError(ret, "on clGetProgramInfo"); ByteBuffer bb = ByteBuffer.allocateDirect((int) pb.get(0)).order(ByteOrder.nativeOrder()); ret = cl.clGetProgramInfo(ID, CL_PROGRAM_DEVICES, bb.capacity(), bb, null); checkForError(ret, "on clGetProgramInfo"); int count = bb.capacity() / (CPU.is32Bit()?4:8); CLDevice[] devices = new CLDevice[count]; for (int i = 0; i < count; i++) { devices[i] = context.getDevice(CPU.is32Bit()?bb.getInt():bb.getLong()); } return devices; } /** * Returns the build log of this program on all devices. The contents of the log are * implementation dependent. */ public String getBuildLog() { if(released) { return ""; } StringBuilder sb = new StringBuilder(); CLDevice[] devices = getCLDevices(); for (int i = 0; i < devices.length; i++) { CLDevice device = devices[i]; sb.append(device).append(" build log:\n"); String log = getBuildLog(device).trim(); sb.append(log.isEmpty()?" ":log); if(i != devices.length-1) sb.append("\n"); } return sb.toString(); } /** * Returns the build status enum of this program for each device as Map. */ public Map getBuildStatus() { if(released) { return Collections.emptyMap(); } initBuildStatus(); return buildStatusMap; } /** * Returns true if the build status 'BUILD_SUCCESS' for at least one device * of this program exists. */ public boolean isExecutable() { if(released) { return false; } initBuildStatus(); return executable; } /** * Returns the build log for this program on the specified device. The contents * of the log are implementation dependent log can be an empty String. */ public String getBuildLog(CLDevice device) { return getBuildInfoString(device.ID, CL_PROGRAM_BUILD_LOG); } /** * Returns the build status enum for this program on the specified device. */ public Status getBuildStatus(CLDevice device) { if(released) { return Status.BUILD_NONE; } int clStatus = getBuildInfoInt(device.ID, CL_PROGRAM_BUILD_STATUS); return Status.valueOf(clStatus); } /** * Returns the source code of this program. Note: sources are not cached, * each call of this method calls into Open */ public String getSource() { return getProgramInfoString(CL_PROGRAM_SOURCE); } /** * Returns the binaries for this program in a Map containing the device as key * and the program binaries as value. */ public Map getBinaries() { if(!isExecutable()) { return Collections.emptyMap(); } CLDevice[] devices = getCLDevices(); ByteBuffer sizes = ByteBuffer.allocateDirect(8*devices.length).order(ByteOrder.nativeOrder()); int ret = cl.clGetProgramInfo(ID, CL_PROGRAM_BINARY_SIZES, sizes.capacity(), sizes, null); checkForError(ret, "on clGetProgramInfo"); int binariesSize = 0; while(sizes.remaining() != 0) { int size = (int) sizes.getLong(); binariesSize += size; } ByteBuffer binaries = ByteBuffer.allocateDirect(binariesSize).order(ByteOrder.nativeOrder()); long address = InternalBufferUtil.getDirectBufferAddress(binaries); PointerBuffer addresses = PointerBuffer.allocateDirect(sizes.capacity()); sizes.rewind(); while(sizes.remaining() != 0) { addresses.put(address); address += sizes.getLong(); } ret = cl.clGetProgramInfo(ID, CL_PROGRAM_BINARIES, addresses.capacity(), addresses.getBuffer(), null); checkForError(ret, "on clGetProgramInfo"); Map map = new HashMap(); sizes.rewind(); for (int i = 0; i < devices.length; i++) { byte[] bytes = new byte[(int)sizes.getLong()]; binaries.get(bytes); map.put(devices[i], bytes); } return map; } /** * Utility method which builds a properly seperated option string. */ public static String optionsOf(String... options) { StringBuilder sb = new StringBuilder(options.length * 24); for (int i = 0; i < options.length; i++) { sb.append(options[i]); if(i!= options.length-1) sb.append(" "); } return sb.toString(); } /** * Utility method for defining macros as build options (Returns "-D name"). */ public static String define(String name) { return "-D "+name; } /** * Utility method for defining macros as build options (Returns "-D name=value"). */ public static String define(String name, String value) { return "-D "+name+"="+value; } @Override public String toString() { return "CLProgram [id: " + ID + " status: "+getBuildStatus()+"]"; } @Override public boolean equals(Object obj) { if (obj == null) { return false; } if (getClass() != obj.getClass()) { return false; } final CLProgram other = (CLProgram) obj; if (this.ID != other.ID) { return false; } if (!this.context.equals(other.context)) { return false; } return true; } @Override public int hashCode() { int hash = 7; hash = 37 * hash + (this.context != null ? this.context.hashCode() : 0); hash = 37 * hash + (int) (this.ID ^ (this.ID >>> 32)); return hash; } public enum Status { BUILD_SUCCESS(CL_BUILD_SUCCESS), BUILD_NONE(CL_BUILD_NONE), BUILD_IN_PROGRESS(CL_BUILD_IN_PROGRESS), BUILD_ERROR(CL_BUILD_ERROR); /** * Value of wrapped OpenCL device type. */ public final int STATUS; private Status(int status) { this.STATUS = status; } public static Status valueOf(int clBuildStatus) { switch(clBuildStatus) { case(CL_BUILD_SUCCESS): return BUILD_SUCCESS; case(CL_BUILD_NONE): return BUILD_NONE; case(CL_BUILD_IN_PROGRESS): return BUILD_IN_PROGRESS; case(CL_BUILD_ERROR): return BUILD_ERROR; // is this a standard state? // case (CL_BUILD_PROGRAM_FAILURE): // return BUILD_PROGRAM_FAILURE; } return null; } } /** * Common compiler options for the OpenCL compiler. */ public interface CompilerOptions { /** * Treat double precision floating-point constant as single precision constant. */ public final static String SINGLE_PRECISION_CONSTANTS = "-cl-single-precision-constant"; /** * This option controls how single precision and double precision denormalized numbers are handled. * If specified as a build option, the single precision denormalized numbers may be flushed to zero * and if the optional extension for double precision is supported, double precision denormalized numbers * may also be flushed to zero. This is intended to be a performance hint and the OpenCL compiler can choose * not to flush denorms to zero if the device supports single precision (or double precision) denormalized numbers.
* This option is ignored for single precision numbers if the device does not support single precision denormalized * numbers i.e. {@link CLDevice.FPConfig#DENORM} is not present in the set returned by {@link CLDevice#getSingleFPConfig()}
* This option is ignored for double precision numbers if the device does not support double precision or if it does support * double precison but {@link CLDevice.FPConfig#DENORM} is not present in the set returned by {@link CLDevice#getDoubleFPConfig()}.
* This flag only applies for scalar and vector single precision floating-point variables and computations on * these floating-point variables inside a program. It does not apply to reading from or writing to image objects. */ public final static String DENORMS_ARE_ZERO = "-cl-denorms-are-zero"; /** * This option disables all optimizations. The default is optimizations are enabled. */ public final static String DISABLE_OPT = "-cl-opt-disable"; /** * This option allows the compiler to assume the strictest aliasing rules. */ public final static String STRICT_ALIASING = "-cl-strict-aliasing"; /** * Allow a * b + c to be replaced by a mad. The mad computes a * b + c with reduced accuracy. * For example, some OpenCL devices implement mad as truncate the result of a * b before adding it to c. */ public final static String ENABLE_MAD = "-cl-mad-enable"; /** * Allow optimizations for floating-point arithmetic that ignore the signedness of zero. * IEEE 754 arithmetic specifies the behavior of distinct +0.0 and -0.0 values, which then prohibits * simplification of expressions such as x+0.0 or 0.0*x (even with -cl-finite-math-only ({@link #FINITE_MATH_ONLY})). * This option implies that the sign of a zero result isn't significant. */ public final static String NO_SIGNED_ZEROS = "-cl-no-signed-zeros"; /** * Allow optimizations for floating-point arithmetic that
* (a) assume that arguments and results are valid,
* (b) may violate IEEE 754 standard and
* (c) may violate the OpenCL numerical compliance requirements as defined in section * 7.4 for single-precision floating-point, section 9.3.9 for double-precision floating-point, * and edge case behavior in section 7.5. * This option includes the -cl-no-signed-zeros ({@link #NO_SIGNED_ZEROS}) * and -cl-mad-enable ({@link #ENABLE_MAD}) options. */ public final static String UNSAFE_MATH = "-cl-unsafe-math-optimizations"; /** * Allow optimizations for floating-point arithmetic that assume that arguments and results are not NaNs or ±∞. * This option may violate the OpenCL numerical compliance requirements defined in in section 7.4 for * single-precision floating-point, section 9.3.9 for double-precision floating-point, and edge case behavior in section 7.5. */ public final static String FINITE_MATH_ONLY = "-cl-finite-math-only"; /** * Sets the optimization options -cl-finite-math-only ({@link #FINITE_MATH_ONLY}) and -cl-unsafe-math-optimizations ({@link #UNSAFE_MATH}). * This allows optimizations for floating-point arithmetic that may violate the IEEE 754 * standard and the OpenCL numerical compliance requirements defined in the specification * in section 7.4 for single-precision floating-point, section 9.3.9 for double-precision * floating-point, and edge case behavior in section 7.5. This option causes the preprocessor * macro __FAST_RELAXED_MATH__ to be defined in the OpenCL program. */ public final static String FAST_RELAXED_MATH = "-cl-fast-relaxed-math"; /** * Inhibit all warning messages. */ public final static String DISABLE_WARNINGS = "-w"; /** * Make all warnings into errors. */ public final static String WARNINGS_ARE_ERRORS = "-Werror"; } }