Repository: systemml Updated Branches: refs/heads/master 3cbd9d5ab -> be2b3e220
[SYSTEMML-445] Extend shadow buffer for double precision - This commit also prepares SystemML for very low precision. Project: http://git-wip-us.apache.org/repos/asf/systemml/repo Commit: http://git-wip-us.apache.org/repos/asf/systemml/commit/be2b3e22 Tree: http://git-wip-us.apache.org/repos/asf/systemml/tree/be2b3e22 Diff: http://git-wip-us.apache.org/repos/asf/systemml/diff/be2b3e22 Branch: refs/heads/master Commit: be2b3e220401c0244bb5df33ddfa8125996066b6 Parents: 3cbd9d5 Author: Niketan Pansare <npan...@us.ibm.com> Authored: Thu Nov 1 05:05:10 2018 -0700 Committer: Niketan Pansare <npan...@us.ibm.com> Committed: Thu Nov 1 17:36:03 2018 +0530 ---------------------------------------------------------------------- .../instructions/gpu/context/ShadowBuffer.java | 98 ++++++++++++++------ 1 file changed, 72 insertions(+), 26 deletions(-) ---------------------------------------------------------------------- http://git-wip-us.apache.org/repos/asf/systemml/blob/be2b3e22/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/ShadowBuffer.java ---------------------------------------------------------------------- diff --git a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/ShadowBuffer.java b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/ShadowBuffer.java index 88ea972..1aeec6f 100644 --- a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/ShadowBuffer.java +++ b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/ShadowBuffer.java @@ -22,9 +22,9 @@ import static jcuda.runtime.JCuda.cudaMemcpy; import org.apache.commons.logging.Log; import org.apache.commons.logging.LogFactory; -import org.apache.sysml.api.DMLScript; import org.apache.sysml.conf.ConfigurationManager; import org.apache.sysml.conf.DMLConfig; +import org.apache.sysml.runtime.DMLRuntimeException; import org.apache.sysml.runtime.controlprogram.parfor.stat.InfrastructureAnalyzer; import org.apache.sysml.runtime.matrix.data.LibMatrixCUDA; import org.apache.sysml.runtime.matrix.data.MatrixBlock; @@ -36,21 +36,17 @@ import jcuda.Sizeof; public class ShadowBuffer { private static final Log LOG = LogFactory.getLog(ShadowBuffer.class.getName()); - GPUObject gpuObj; - float[] shadowPointer = null; + private GPUObject gpuObj; + // shadowPointer can be double[], float[] or short[]. + private Object shadowPointer = null; private static boolean _warnedAboutShadowBuffer = false; private static long EVICTION_SHADOW_BUFFER_CURR_BYTES = 0; private static long EVICTION_SHADOW_BUFFER_MAX_BYTES; static { - if(DMLScript.FLOATING_POINT_PRECISION.equals("double")) { - EVICTION_SHADOW_BUFFER_MAX_BYTES = 0; - } - else { - double shadowBufferSize = ConfigurationManager.getDMLConfig().getDoubleValue(DMLConfig.EVICTION_SHADOW_BUFFERSIZE); - if(shadowBufferSize < 0 || shadowBufferSize > 1) - throw new RuntimeException("Incorrect value (" + shadowBufferSize + ") for the configuration:" + DMLConfig.EVICTION_SHADOW_BUFFERSIZE); - EVICTION_SHADOW_BUFFER_MAX_BYTES = (long) (((double)InfrastructureAnalyzer.getLocalMaxMemory())*shadowBufferSize); - } + double shadowBufferSize = ConfigurationManager.getDMLConfig().getDoubleValue(DMLConfig.EVICTION_SHADOW_BUFFERSIZE); + if(shadowBufferSize < 0 || shadowBufferSize > 1) + throw new RuntimeException("Incorrect value (" + shadowBufferSize + ") for the configuration:" + DMLConfig.EVICTION_SHADOW_BUFFERSIZE); + EVICTION_SHADOW_BUFFER_MAX_BYTES = (long) (((double)InfrastructureAnalyzer.getLocalMaxMemory())*shadowBufferSize); } public ShadowBuffer(GPUObject gpuObj) { @@ -73,9 +69,21 @@ public class ShadowBuffer { public void moveFromDevice(String instName) { long start = ConfigurationManager.isStatistics() ? System.nanoTime() : 0; int numElems = GPUObject.toIntExact(gpuObj.mat.getNumRows()*gpuObj.mat.getNumColumns()); - shadowPointer = new float[numElems]; - EVICTION_SHADOW_BUFFER_CURR_BYTES += getSizeOfFloat(shadowPointer.length); - cudaMemcpy(Pointer.to(shadowPointer), gpuObj.jcudaDenseMatrixPtr, getSizeOfDataType(numElems), jcuda.runtime.cudaMemcpyKind.cudaMemcpyDeviceToHost); + if(LibMatrixCUDA.sizeOfDataType == Sizeof.DOUBLE) { + shadowPointer = new double[numElems]; + } + else if(LibMatrixCUDA.sizeOfDataType == Sizeof.FLOAT) { + shadowPointer = new float[numElems]; + } + else if(LibMatrixCUDA.sizeOfDataType == Sizeof.SHORT) { + shadowPointer = new short[numElems]; + } + else { + throw new DMLRuntimeException("Unsupported datatype"); + } + long numBytes = getNumBytesOfShadowBuffer(); + EVICTION_SHADOW_BUFFER_CURR_BYTES += numBytes; + cudaMemcpy(getHostShadowPointer(), gpuObj.jcudaDenseMatrixPtr, numBytes, jcuda.runtime.cudaMemcpyKind.cudaMemcpyDeviceToHost); gpuObj.getGPUContext().cudaFreeHelper(instName, gpuObj.jcudaDenseMatrixPtr, true); gpuObj.jcudaDenseMatrixPtr = null; if (ConfigurationManager.isStatistics()) { @@ -87,12 +95,35 @@ public class ShadowBuffer { } } - private long getSizeOfFloat(long numElems) { - return numElems*Sizeof.FLOAT; + private long getNumBytesOfShadowBuffer() { + long numElems = 0; + switch(LibMatrixCUDA.sizeOfDataType) { + case Sizeof.DOUBLE: + numElems = ((double[])shadowPointer).length; + break; + case Sizeof.FLOAT: + numElems = ((float[])shadowPointer).length; + break; + case Sizeof.SHORT: + numElems = ((short[])shadowPointer).length; + break; + default: + throw new DMLRuntimeException("Unsupported datatype of size:" + LibMatrixCUDA.sizeOfDataType); + } + return numElems*LibMatrixCUDA.sizeOfDataType; } - private long getSizeOfDataType(long numElems) { - return numElems*LibMatrixCUDA.sizeOfDataType; + private Pointer getHostShadowPointer() { + switch(LibMatrixCUDA.sizeOfDataType) { + case Sizeof.DOUBLE: + return Pointer.to((double[])shadowPointer); + case Sizeof.FLOAT: + return Pointer.to((float[])shadowPointer); + case Sizeof.SHORT: + return Pointer.to((short[])shadowPointer); + default: + throw new DMLRuntimeException("Unsupported datatype of size:" + LibMatrixCUDA.sizeOfDataType); + } } /** @@ -103,9 +134,24 @@ public class ShadowBuffer { MatrixBlock tmp = new MatrixBlock(GPUObject.toIntExact(gpuObj.mat.getNumRows()), GPUObject.toIntExact(gpuObj.mat.getNumColumns()), false); tmp.allocateDenseBlock(); double [] tmpArr = tmp.getDenseBlockValues(); - for(int i = 0; i < shadowPointer.length; i++) { - tmpArr[i] = shadowPointer[i]; + if(LibMatrixCUDA.sizeOfDataType == Sizeof.DOUBLE) { + double[] sArr = ((double[])shadowPointer); + System.arraycopy(sArr, 0, tmpArr, 0, sArr.length); + } + else if(LibMatrixCUDA.sizeOfDataType == Sizeof.FLOAT) { + float[] sArr = ((float[])shadowPointer); + for(int i = 0; i < sArr.length; i++) { + tmpArr[i] = sArr[i]; + } + } + else if(LibMatrixCUDA.sizeOfDataType == Sizeof.SHORT) { + // short[] sArr = ((short[])shadowPointer); + throw new DMLRuntimeException("Unsupported operation: moveToHost for half precision"); + } + else { + throw new DMLRuntimeException("Unsupported datatype of size:" + LibMatrixCUDA.sizeOfDataType); } + gpuObj.mat.acquireModify(tmp); gpuObj.mat.release(); clearShadowPointer(); @@ -125,9 +171,9 @@ public class ShadowBuffer { */ public void moveToDevice() { long start = ConfigurationManager.isStatistics() ? System.nanoTime() : 0; - long numBytes = getSizeOfDataType(shadowPointer.length); + long numBytes = getNumBytesOfShadowBuffer(); gpuObj.jcudaDenseMatrixPtr = gpuObj.getGPUContext().allocate(null, numBytes); - cudaMemcpy(gpuObj.jcudaDenseMatrixPtr, Pointer.to(shadowPointer), numBytes, jcuda.runtime.cudaMemcpyKind.cudaMemcpyHostToDevice); + cudaMemcpy(gpuObj.jcudaDenseMatrixPtr, getHostShadowPointer(), numBytes, jcuda.runtime.cudaMemcpyKind.cudaMemcpyHostToDevice); clearShadowPointer(); if (ConfigurationManager.isStatistics()) { long totalTime = System.nanoTime() - start; @@ -144,8 +190,8 @@ public class ShadowBuffer { * @return true if the given GPU object is eligible to be shadow buffered */ public boolean isEligibleForBuffering(boolean isEviction, boolean eagerDelete) { - if(LibMatrixCUDA.sizeOfDataType == jcuda.Sizeof.FLOAT && isEviction && eagerDelete && !gpuObj.isDensePointerNull()) { - long numBytes = getSizeOfFloat(gpuObj.mat.getNumRows()*gpuObj.mat.getNumColumns()); + if(isEviction && eagerDelete && !gpuObj.isDensePointerNull()) { + long numBytes = gpuObj.mat.getNumRows()*gpuObj.mat.getNumColumns()*LibMatrixCUDA.sizeOfDataType; boolean ret = EVICTION_SHADOW_BUFFER_CURR_BYTES + numBytes <= EVICTION_SHADOW_BUFFER_MAX_BYTES; if(!ret && !_warnedAboutShadowBuffer) { LOG.warn("Shadow buffer is full, so using CP bufferpool instead. Consider increasing sysml.gpu.eviction.shadow.bufferSize."); @@ -163,7 +209,7 @@ public class ShadowBuffer { */ public void clearShadowPointer() { if(shadowPointer != null) { - EVICTION_SHADOW_BUFFER_CURR_BYTES -= getSizeOfFloat(shadowPointer.length); + EVICTION_SHADOW_BUFFER_CURR_BYTES -= getNumBytesOfShadowBuffer(); } shadowPointer = null; }
