Examples with MatrixObject org.apache.sysml.runtime.controlprogram.caching.MatrixObject used on opensource projects

Example 81 with MatrixObject

use of org.apache.sysml.runtime.controlprogram.caching.MatrixObject in project incubator-systemml by apache.

the class ExecutionContext method getDenseMatrixOutputForGPUInstruction.

/**
 * Allocates a dense matrix on the GPU (for output)
 * @param varName	name of the output matrix (known by this {@link ExecutionContext})
 * @param numRows number of rows of matrix object
 * @param numCols number of columns of matrix object
 * @return a pair containing the wrapping {@link MatrixObject} and a boolean indicating whether a cuda memory allocation took place (as opposed to the space already being allocated)
 */
public Pair<MatrixObject, Boolean> getDenseMatrixOutputForGPUInstruction(String varName, long numRows, long numCols) {
    MatrixObject mo = allocateGPUMatrixObject(varName, numRows, numCols);
    boolean allocated = mo.getGPUObject(getGPUContext(0)).acquireDeviceModifyDense();
    mo.getMatrixCharacteristics().setNonZeros(-1);
    return new Pair<>(mo, allocated);
}

Example 82 with MatrixObject

use of org.apache.sysml.runtime.controlprogram.caching.MatrixObject in project incubator-systemml by apache.

the class ExternalFunctionInvocationInstruction method getInputObjects.

@SuppressWarnings("incomplete-switch")
private ArrayList<FunctionParameter> getInputObjects(CPOperand[] inputs, LocalVariableMap vars) {
    ArrayList<FunctionParameter> ret = new ArrayList<>();
    for (CPOperand input : inputs) {
        switch(input.getDataType()) {
            case MATRIX:
                MatrixObject mobj = (MatrixObject) vars.get(input.getName());
                ret.add(new Matrix(mobj, getMatrixValueType(input.getValueType())));
                break;
            case SCALAR:
                ScalarObject so = (ScalarObject) vars.get(input.getName());
                ret.add(new Scalar(getScalarValueType(input.getValueType()), so.getStringValue()));
                break;
            case OBJECT:
                ret.add(new BinaryObject(vars.get(input.getName())));
                break;
        }
    }
    return ret;
}

Example 83 with MatrixObject

use of org.apache.sysml.runtime.controlprogram.caching.MatrixObject in project incubator-systemml by apache.

the class ExternalFunctionInvocationInstruction method createOutputMatrixObject.

private MatrixObject createOutputMatrixObject(Matrix m) {
    MatrixObject ret = m.getMatrixObject();
    if (ret == null) {
        // otherwise, pass in-memory matrix from extfunct back to invoking program
        MatrixCharacteristics mc = new MatrixCharacteristics(m.getNumRows(), m.getNumCols(), ConfigurationManager.getBlocksize(), ConfigurationManager.getBlocksize());
        MetaDataFormat mfmd = new MetaDataFormat(mc, InputInfo.getMatchingOutputInfo(iinfo), iinfo);
        ret = new MatrixObject(Expression.ValueType.DOUBLE, m.getFilePath(), mfmd);
    }
    // for allowing in-memory packagesupport matrices w/o file names
    if (ret.getFileName().equals(Matrix.DEFAULT_FILENAME)) {
        ret.setFileName(createDefaultOutputFilePathAndName());
    }
    return ret;
}

Example 84 with MatrixObject

use of org.apache.sysml.runtime.controlprogram.caching.MatrixObject in project incubator-systemml by apache.

the class DynamicWriteMatrixCP method execute.

@Override
public void execute() {
    boolean success = false;
    try {
        Matrix mat = (Matrix) this.getFunctionInput(0);
        String fname = ((Scalar) this.getFunctionInput(1)).getValue();
        String format = ((Scalar) this.getFunctionInput(2)).getValue();
        MatrixObject mo = mat.getMatrixObject();
        MatrixCharacteristics mc = mo.getMatrixCharacteristics();
        OutputInfo oi = OutputInfo.stringToOutputInfo(format);
        MatrixBlock mb = mo.acquireRead();
        DataConverter.writeMatrixToHDFS(mb, fname, oi, mc);
        mo.release();
        success = true;
    } catch (Exception e) {
        throw new RuntimeException("Error executing dynamic write of matrix", e);
    }
    _success = new Scalar(ScalarValueType.Boolean, String.valueOf(success));
}

Example 85 with MatrixObject

use of org.apache.sysml.runtime.controlprogram.caching.MatrixObject in project incubator-systemml by apache.

the class LibMatrixCUDA method matmultTSMM.

// ********************************************************************/
// ************* End of DEEP LEARNING Operators ***********************/
// ********************************************************************/
// ********************************************************************/
// ********** TRANSPOSE SELF MATRIX MULTIPLY Functions ****************/
// ********************************************************************/
/**
 * Performs tsmm, A %*% A' or A' %*% A, on GPU by exploiting cublasDsyrk(...)
 * <p>
 * Memory Usage - If dense, input space - rows * cols, no intermediate memory, output - Max(rows*rows, cols*cols)
 * If sparse, calls matmult
 *
 * @param ec               execution context
 * @param gCtx             a valid {@link GPUContext}
 * @param instName         the invoking instruction's name for record {@link Statistics}.
 * @param left             input matrix, as in a tsmm expression like A %*% A' or A' %*% A, we just need to check whether the left one is transposed or not, I named it 'left'
 * @param outputName       output matrix name
 * @param isLeftTransposed if true, left transposed
 */
public static void matmultTSMM(ExecutionContext ec, GPUContext gCtx, String instName, MatrixObject left, String outputName, boolean isLeftTransposed) {
    if (LOG.isTraceEnabled()) {
        LOG.trace("GPU : matmultTSMM" + ", GPUContext=" + gCtx);
    }
    if (ec.getGPUContext(0) != gCtx)
        throw new DMLRuntimeException("GPU : Invalid internal state, the GPUContext set with the ExecutionContext is not the same used to run this LibMatrixCUDA function");
    if (isInSparseFormat(gCtx, left)) {
        // For sparse TSMM, invoke matmult (TODO: possible performance improvement)
        LibMatrixCuMatMult.matmult(ec, gCtx, instName, left, left, outputName, isLeftTransposed, !isLeftTransposed);
        return;
    }
    // Since CuBLAS expects inputs in column-major format,
    // reverse the order of matrix-multiplication and take care of dimension mismatch.
    int transa = isLeftTransposed ? cublasOperation.CUBLAS_OP_N : cublasOperation.CUBLAS_OP_T;
    // Note: the dimensions are swapped
    int m = toInt(isLeftTransposed ? left.getNumColumns() : left.getNumRows());
    int k = toInt(isLeftTransposed ? left.getNumRows() : left.getNumColumns());
    // For dense TSMM, exploit cublasDsyrk(...) and call custom kernel to flip the matrix
    // Allocated the dense output matrix
    MatrixObject output = getDenseMatrixOutputForGPUInstruction(ec, instName, outputName, m, m);
    if (m == -1)
        throw new DMLRuntimeException("Incorrect dimensions");
    int lda = toInt(isLeftTransposed ? m : k);
    int ldc = m;
    if (!left.getGPUObject(gCtx).isAllocated())
        throw new DMLRuntimeException("Input is not allocated:" + left.getGPUObject(gCtx).isAllocated());
    if (!output.getGPUObject(gCtx).isAllocated())
        throw new DMLRuntimeException("Output is not allocated:" + output.getGPUObject(gCtx).isAllocated());
    Pointer A = getDensePointer(gCtx, left, instName);
    Pointer C = getDensePointer(gCtx, output, instName);
    long t0 = 0, t1 = 0;
    if (DMLScript.FINEGRAINED_STATISTICS)
        t0 = System.nanoTime();
    cudaSupportFunctions.cublassyrk(getCublasHandle(gCtx), cublasFillMode.CUBLAS_FILL_MODE_LOWER, transa, m, k, one(), A, lda, zero(), C, ldc);
    if (DMLScript.FINEGRAINED_STATISTICS)
        GPUStatistics.maintainCPMiscTimes(instName, GPUInstruction.MISC_TIMER_SYRK_LIB, System.nanoTime() - t0);
    if (DMLScript.FINEGRAINED_STATISTICS)
        t1 = System.nanoTime();
    copyUpperToLowerTriangle(gCtx, instName, output);
    if (DMLScript.FINEGRAINED_STATISTICS)
        GPUStatistics.maintainCPMiscTimes(instName, GPUInstruction.MISC_TIMER_UPPER_TO_LOWER_TRIANGLE_KERNEL, System.nanoTime() - t1);
}