Mercurial > hg > truffle
view graal/com.oracle.graal.hotspot.ptx/src/com/oracle/graal/hotspot/ptx/PTXWrapperBuilder.java @ 13949:f694daada5bf
mx jmh: display the number of benchmarks that will run
author | Gilles Duboscq <duboscq@ssw.jku.at> |
---|---|
date | Thu, 13 Feb 2014 17:03:20 +0100 |
parents | 36226874c146 |
children | 958c99d0790c |
line wrap: on
line source
/* * Copyright (c) 2012, Oracle and/or its affiliates. All rights reserved. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This code is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 only, as * published by the Free Software Foundation. * * This code is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License * version 2 for more details (a copy is included in the LICENSE file that * accompanied this code). * * You should have received a copy of the GNU General Public License version * 2 along with this work; if not, write to the Free Software Foundation, * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. * * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA * or visit www.oracle.com if you need additional information or have any * questions. */ package com.oracle.graal.hotspot.ptx; import static com.oracle.graal.api.meta.DeoptimizationAction.*; import static com.oracle.graal.api.meta.DeoptimizationReason.*; import static com.oracle.graal.api.meta.LocationIdentity.*; import static com.oracle.graal.api.meta.MetaUtil.*; import static com.oracle.graal.asm.NumUtil.*; import static com.oracle.graal.hotspot.ptx.PTXHotSpotBackend.*; import static com.oracle.graal.hotspot.ptx.PTXWrapperBuilder.LaunchArg.*; import static com.oracle.graal.hotspot.replacements.HotSpotReplacementsUtil.*; import static com.oracle.graal.nodes.ConstantNode.*; import static java.lang.reflect.Modifier.*; import java.util.*; import com.oracle.graal.api.meta.*; import com.oracle.graal.debug.*; import com.oracle.graal.graph.*; import com.oracle.graal.hotspot.meta.*; import com.oracle.graal.hotspot.nodes.*; import com.oracle.graal.hotspot.stubs.*; import com.oracle.graal.java.*; import com.oracle.graal.lir.ptx.*; import com.oracle.graal.nodes.*; import com.oracle.graal.nodes.HeapAccess.BarrierType; import com.oracle.graal.nodes.calc.*; import com.oracle.graal.nodes.extended.*; import com.oracle.graal.nodes.java.*; import com.oracle.graal.nodes.type.*; import com.oracle.graal.replacements.*; import com.oracle.graal.replacements.nodes.*; import com.oracle.graal.word.*; /** * Utility for building a graph that "wraps" a compiled PTX kernel. Such a wrapper handles the * transition from the host CPU to the GPU and back. The wrapper allocate 3 on-stack buffers: * <ul> * <li>PARAMS: a buffer for the kernel parameters and one word for the on-device address of the * return value (if any).</li> * <li>PINNED: a buffer into which the address of pinned objects is saved.</li> * <li>OBJECT_OFFSETS: the offsets of the object values in PARAMS.</li> * </ul> * * * The PARAMS buffer is the {@code CU_LAUNCH_PARAM_BUFFER_POINTER} buffer passed in the * {@code extra} argument to the {@code cuLaunchKernel} function. This buffer contains the * parameters to the call. The buffer is word aligned and each parameter is aligned in the buffer * according to its data size. The wrapper copies the incoming arguments into the buffer as is. The * native {@link PTXHotSpotBackend#CALL_KERNEL callKernel} function will pin the memory for each * object parameter (using {@code cuMemHostRegister}) and then replace the object pointer in PARAMS * with an on-device pointer to the object's memory (see {@code cuMemHostGetDevicePointer}). The * function saves pinned object pointer into PINNED so that it can unpinned once the kernel returns. * The object pointers in PARAMS are specified by OBJECT_OFFSETS. * <p> * As a concrete example, for a kernel whose Java method signature is: * * <pre> * static int kernel(int p1, short p2, Object p3, long p4) * </pre> * * the graph created is shown below as psuedo-code: * * <pre> * int kernel_wrapper(int p1, short p2, oop p3, long p4) { * address kernelAddr = kernel.start; * if (kernelAddr == 0) { * deopt(InvalidateRecompile, RuntimeConstraint); * } * byte PARAMS[32]; * word PINNED[1]; // note: no refmap * int OBJECT_OFFSETS[1] = {8}; * ((int*) PARAMS)[0] = p1; * ((short*) PARAMS)[2] = p2; * ((word*) PARAMS)[1] = p3; * ((long*) PARAMS)[2] = p4; * int result = CALL_KERNEL(THREAD_REGISTER, KERNEL_ENTRY_POINT, 1, 1, 1, PARAMS, 32, 1, OBJECT_OFFSETS, PINNED, 4); * if (clearPendingException(thread)) { * deopt(None, RuntimeConstraint); * } * return result; * } * </pre> * <p> * The generated graph includes a reference to the {@link HotSpotNmethod} for the kernel. There must * be another reference to the same {@link HotSpotNmethod} object to ensure that the nmethod is not * unloaded by the next full GC. Currently, these extra "keep-alive" references are maintained by * {@link PTXHotSpotBackend}. * <p> * The PTX runtime code called by the wrapper blocks GC while the kernel is executing (cf * GetPrimitiveArrayCritical/ReleasePrimitiveArrayCritical JNI functions). This ensures objects can * be safely passed to kernels but should be replaced with a lighter weight mechanism at some point. */ public class PTXWrapperBuilder extends GraphKit { /** * The incoming Java arguments to the method. */ ParameterNode[] javaParameters; /** * The size of the buffer holding the kernel parameters and the extra word for storing the * pointer to device memory for the return value. * * @see LaunchArg#ParametersAndReturnValueBufferSize */ int bufSize; /** * Offset of each Java argument in the kernel parameters buffer. */ int[] javaParameterOffsetsInKernelParametersBuffer; /** * Constants denoting the arguments to {@link PTXHotSpotBackend#CALL_KERNEL}. */ // @formatter:off enum LaunchArg { Thread, Kernel, DimX, DimY, DimZ, ParametersAndReturnValueBuffer, ParametersAndReturnValueBufferSize, ObjectParametersCount, ObjectParametersOffsets, PinnedObjects, EncodedReturnTypeSize } // @formatter:on /** * Creates the graph implementing the CPU to GPU transition. * * @param method a method that has been compiled to GPU binary code * @param kernel the installed GPU binary for {@code method} * @see PTXWrapperBuilder */ public PTXWrapperBuilder(ResolvedJavaMethod method, HotSpotNmethod kernel, HotSpotProviders providers) { super(new StructuredGraph(method), providers); int wordSize = providers.getCodeCache().getTarget().wordSize; int intSize = Integer.SIZE / Byte.SIZE; Kind wordKind = providers.getCodeCache().getTarget().wordKind; Signature sig = method.getSignature(); boolean isStatic = isStatic(method.getModifiers()); int sigCount = sig.getParameterCount(false); javaParameters = new ParameterNode[(!isStatic ? 1 : 0) + sigCount]; javaParameterOffsetsInKernelParametersBuffer = new int[javaParameters.length]; int javaParametersIndex = 0; Kind returnKind = sig.getReturnKind(); List<Integer> objectSlots = new ArrayList<>(javaParameters.length); if (!isStatic) { allocateParameter(Kind.Object, javaParametersIndex++, objectSlots, wordSize); } for (int sigIndex = 0; sigIndex < sigCount; sigIndex++) { Kind kind = sig.getParameterKind(sigIndex); allocateParameter(kind, javaParametersIndex++, objectSlots, wordSize); } bufSize = roundUp(bufSize, wordSize); // Add slot for the device memory pointer. The kernel writes a // pointer in this slot that points to the return value. int encodedReturnTypeSize = 0; if (returnKind != Kind.Void) { bufSize += wordSize; if (returnKind == Kind.Object) { encodedReturnTypeSize = -wordSize; } else { encodedReturnTypeSize = returnKind.getBitCount() / Byte.SIZE; } } InvokeNode kernelStart = createInvoke(getClass(), "getKernelStart", ConstantNode.forObject(kernel, providers.getMetaAccess(), getGraph())); AllocaNode buf = append(new AllocaNode(bufSize / wordSize, new BitSet())); ValueNode objectParametersOffsets; ValueNode pinnedObjects; ConstantNode nullWord = ConstantNode.forIntegerKind(wordKind, 0L, getGraph()); if (objectSlots.isEmpty()) { objectParametersOffsets = ConstantNode.forLong(0, getGraph()); pinnedObjects = ConstantNode.forLong(0, getGraph()); } else { int intsPerWord = wordSize / intSize; int slots = roundUp(objectSlots.size(), intsPerWord); objectParametersOffsets = append(new AllocaNode(slots, new BitSet())); // No refmap for pinned objects list since kernel execution is (currently) GC unsafe pinnedObjects = append(new AllocaNode(objectSlots.size(), new BitSet())); // Initialize the object parameter offsets array int index = 0; for (int slot : objectSlots) { int offset = slot * wordSize; LocationNode location = ConstantLocationNode.create(FINAL_LOCATION, Kind.Int, index * intSize, getGraph()); append(new WriteNode(objectParametersOffsets, ConstantNode.forInt(offset, getGraph()), location, BarrierType.NONE, false, false)); index++; } } Map<LaunchArg, ValueNode> args = new EnumMap<>(LaunchArg.class); args.put(Thread, append(new ReadRegisterNode(providers.getRegisters().getThreadRegister(), true, false))); args.put(Kernel, kernelStart); args.put(DimX, forInt(1, getGraph())); args.put(DimY, forInt(1, getGraph())); args.put(DimZ, forInt(1, getGraph())); args.put(ParametersAndReturnValueBuffer, buf); args.put(ParametersAndReturnValueBufferSize, forInt(bufSize, getGraph())); args.put(ObjectParametersCount, forInt(objectSlots.size(), getGraph())); args.put(ObjectParametersOffsets, objectParametersOffsets); args.put(PinnedObjects, pinnedObjects); args.put(EncodedReturnTypeSize, forInt(encodedReturnTypeSize, getGraph())); int sigIndex = isStatic ? 0 : -1; for (javaParametersIndex = 0; javaParametersIndex < javaParameters.length; javaParametersIndex++) { ParameterNode javaParameter = javaParameters[javaParametersIndex]; int javaParameterOffset = javaParameterOffsetsInKernelParametersBuffer[javaParametersIndex]; LocationNode location = ConstantLocationNode.create(FINAL_LOCATION, javaParameter.kind(), javaParameterOffset, getGraph()); append(new WriteNode(buf, javaParameter, location, BarrierType.NONE, false, false)); updateDimArg(method, providers, sig, sigIndex++, args, javaParameter); } if (returnKind != Kind.Void) { LocationNode location = ConstantLocationNode.create(FINAL_LOCATION, wordKind, bufSize - wordSize, getGraph()); append(new WriteNode(buf, nullWord, location, BarrierType.NONE, false, false)); } FrameStateBuilder fsb = new FrameStateBuilder(method, getGraph(), true); FrameState fs = fsb.create(0); getGraph().start().setStateAfter(fs); ValueNode[] launchArgsArray = args.values().toArray(new ValueNode[args.size()]); ForeignCallNode result = append(new ForeignCallNode(providers.getForeignCalls(), CALL_KERNEL, launchArgsArray)); result.setDeoptimizationState(fs); InvokeNode getObjectResult = null; ValueNode returnValue; switch (returnKind) { case Void: returnValue = null; break; case Boolean: case Byte: case Short: case Char: case Int: returnValue = unique(new ConvertNode(Kind.Long, Kind.Int, result)); break; case Long: returnValue = result; break; case Float: { ValueNode asInt = unique(new ConvertNode(Kind.Long, Kind.Int, result)); returnValue = unique(new ReinterpretNode(Kind.Float, asInt)); break; } case Double: returnValue = unique(new ReinterpretNode(returnKind, result)); break; case Object: getObjectResult = createInvoke(getClass(), "getObjectResult", args.get(Thread)); returnValue = append(getObjectResult); break; default: throw new GraalInternalError("%s return kind not supported", returnKind); } append(new ReturnNode(returnValue)); if (Debug.isDumpEnabled()) { Debug.dump(getGraph(), "Initial kernel launch graph"); } rewriteWordTypes(); inlineInvokes(); if (Debug.isDumpEnabled()) { Debug.dump(getGraph(), "Kernel launch graph before compilation"); } } /** * Computes offset and size of space in PARAMS for a Java parameter. * * @param kind the kind of the parameter * @param javaParametersIndex the index of the Java parameter */ private void allocateParameter(Kind kind, int javaParametersIndex, List<Integer> objectSlots, int wordSize) { int kindByteSize = kind == Kind.Object ? wordSize : kind.getBitCount() / Byte.SIZE; bufSize = roundUp(bufSize, kindByteSize); javaParameterOffsetsInKernelParametersBuffer[javaParametersIndex] = bufSize; Stamp stamp; if (kind == Kind.Object) { stamp = StampFactory.object(); int slot = bufSize / wordSize; objectSlots.add(slot); } else { stamp = StampFactory.forKind(kind); } javaParameters[javaParametersIndex] = unique(new ParameterNode(javaParametersIndex, stamp)); bufSize += kindByteSize; } /** * Updates the {@code dimX}, {@code dimY} or {@code dimZ} argument passed to the kernel if * {@code javaParameter} is annotated with {@link ParallelOver}. */ private void updateDimArg(ResolvedJavaMethod method, HotSpotProviders providers, Signature sig, int sigIndex, Map<LaunchArg, ValueNode> launchArgs, ParameterNode javaParameter) { if (sigIndex >= 0) { ParallelOver parallelOver = getParameterAnnotation(ParallelOver.class, sigIndex, method); if (parallelOver != null && sig.getParameterType(sigIndex, method.getDeclaringClass()).equals(providers.getMetaAccess().lookupJavaType(int[].class))) { ArrayLengthNode dimension = append(new ArrayLengthNode(javaParameter)); LaunchArg argKey = LaunchArg.valueOf(LaunchArg.class, "Dim" + parallelOver.dimension()); ValueNode existing = launchArgs.put(argKey, dimension); if (existing != null && existing instanceof ArrayLengthNode) { throw new GraalInternalError("@" + ParallelOver.class.getSimpleName() + " with dimension=" + parallelOver.dimension() + " applied to multiple parameters"); } } } } /** * Snippet invoked to get the {@linkplain HotSpotNmethod#getStart() entry point} of the kernel, * deoptimizing if the kernel is invalid. */ @Snippet private static long getKernelStart(HotSpotNmethod ptxKernel) { long start = ptxKernel.getStart(); if (start == 0L) { DeoptimizeNode.deopt(InvalidateRecompile, RuntimeConstraint); } return start; } /** * Snippet invoked upon return from the kernel to retrieve an object return value from the * thread local used for communicating object return values from VM calls. */ @Snippet private static Object getObjectResult(Word thread) { return getAndClearObjectResult(thread); } }