Examples with CUstream_st org.bytedeco.javacpp.cuda.CUstream_st used on opensource projects

Example 1 with CUstream_st

use of org.bytedeco.javacpp.cuda.CUstream_st in project nd4j by deeplearning4j.

the class JcublasLapack method dsyev.

public int dsyev(char _jobz, char _uplo, int N, INDArray A, INDArray R) {
    int status = -1;
    int jobz = _jobz == 'V' ? CUSOLVER_EIG_MODE_VECTOR : CUSOLVER_EIG_MODE_NOVECTOR;
    int uplo = _uplo == 'L' ? CUBLAS_FILL_MODE_LOWER : CUBLAS_FILL_MODE_UPPER;
    if (Nd4j.dataType() != DataBuffer.Type.DOUBLE)
        log.warn("DOUBLE dsyev called in FLOAT environment");
    INDArray a = A;
    if (A.ordering() == 'c')
        a = A.dup('f');
    int M = A.rows();
    if (Nd4j.getExecutioner() instanceof GridExecutioner)
        ((GridExecutioner) Nd4j.getExecutioner()).flushQueue();
    // Get context for current thread
    CudaContext ctx = (CudaContext) allocator.getDeviceContext().getContext();
    // setup the solver handles for cuSolver calls
    cusolverDnHandle_t handle = ctx.getSolverHandle();
    cusolverDnContext solverDn = new cusolverDnContext(handle);
    // synchronized on the solver
    synchronized (handle) {
        status = cusolverDnSetStream(new cusolverDnContext(handle), new CUstream_st(ctx.getOldStream()));
        if (status == 0) {
            // transfer the INDArray into GPU memory
            CublasPointer xAPointer = new CublasPointer(a, ctx);
            CublasPointer xRPointer = new CublasPointer(R, ctx);
            // this output - indicates how much memory we'll need for the real operation
            DataBuffer worksizeBuffer = Nd4j.getDataBufferFactory().createInt(1);
            status = cusolverDnDsyevd_bufferSize(solverDn, jobz, uplo, M, (DoublePointer) xAPointer.getDevicePointer(), M, (DoublePointer) xRPointer.getDevicePointer(), (IntPointer) worksizeBuffer.addressPointer());
            if (status == CUSOLVER_STATUS_SUCCESS) {
                int worksize = worksizeBuffer.getInt(0);
                // allocate memory for the workspace, the non-converging row buffer and a return code
                Pointer workspace = new Workspace(worksize * Nd4j.sizeOfDataType());
                INDArray INFO = Nd4j.createArrayFromShapeBuffer(Nd4j.getDataBufferFactory().createInt(1), Nd4j.getShapeInfoProvider().createShapeInformation(new int[] { 1, 1 }));
                // Do the actual decomp
                status = cusolverDnDsyevd(solverDn, jobz, uplo, M, (DoublePointer) xAPointer.getDevicePointer(), M, (DoublePointer) xRPointer.getDevicePointer(), new CudaPointer(workspace).asDoublePointer(), worksize, new CudaPointer(allocator.getPointer(INFO, ctx)).asIntPointer());
                allocator.registerAction(ctx, INFO);
                if (status == 0)
                    status = INFO.getInt(0);
            }
        }
    }
    if (status == 0) {
        allocator.registerAction(ctx, R);
        allocator.registerAction(ctx, a);
        if (a != A)
            A.assign(a);
    }
    return status;
}

Example 2 with CUstream_st

use of org.bytedeco.javacpp.cuda.CUstream_st in project nd4j by deeplearning4j.

the class JcublasLapack method spotrf.

// =========================
// CHOLESKY DECOMP
@Override
public void spotrf(byte uplo, int N, INDArray A, INDArray INFO) {
    INDArray a = A;
    if (Nd4j.dataType() != DataBuffer.Type.FLOAT)
        log.warn("DOUBLE potrf called in FLOAT environment");
    if (A.ordering() == 'c')
        a = A.dup('f');
    if (Nd4j.getExecutioner() instanceof GridExecutioner)
        ((GridExecutioner) Nd4j.getExecutioner()).flushQueue();
    // Get context for current thread
    CudaContext ctx = (CudaContext) allocator.getDeviceContext().getContext();
    // setup the solver handles for cuSolver calls
    cusolverDnHandle_t handle = ctx.getSolverHandle();
    cusolverDnContext solverDn = new cusolverDnContext(handle);
    // synchronized on the solver
    synchronized (handle) {
        int result = cusolverDnSetStream(new cusolverDnContext(handle), new CUstream_st(ctx.getOldStream()));
        if (result != 0)
            throw new BlasException("solverSetStream failed");
        // transfer the INDArray into GPU memory
        CublasPointer xAPointer = new CublasPointer(a, ctx);
        // this output - indicates how much memory we'll need for the real operation
        DataBuffer worksizeBuffer = Nd4j.getDataBufferFactory().createInt(1);
        int stat = cusolverDnSpotrf_bufferSize(solverDn, uplo, N, (FloatPointer) xAPointer.getDevicePointer(), N, // we intentionally use host pointer here
        (IntPointer) worksizeBuffer.addressPointer());
        if (stat != CUSOLVER_STATUS_SUCCESS) {
            throw new BlasException("cusolverDnSpotrf_bufferSize failed", stat);
        }
        int worksize = worksizeBuffer.getInt(0);
        // Now allocate memory for the workspace, the permutation matrix and a return code
        Pointer workspace = new Workspace(worksize * Nd4j.sizeOfDataType());
        // Do the actual decomp
        stat = cusolverDnSpotrf(solverDn, uplo, N, (FloatPointer) xAPointer.getDevicePointer(), N, new CudaPointer(workspace).asFloatPointer(), worksize, new CudaPointer(allocator.getPointer(INFO, ctx)).asIntPointer());
        if (stat != CUSOLVER_STATUS_SUCCESS) {
            throw new BlasException("cusolverDnSpotrf failed", stat);
        }
    }
    allocator.registerAction(ctx, a);
    allocator.registerAction(ctx, INFO);
    if (a != A)
        A.assign(a);
    if (uplo == 'U') {
        A.assign(A.transpose());
        INDArrayIndex[] ix = new INDArrayIndex[2];
        for (int i = 1; i < Math.min(A.rows(), A.columns()); i++) {
            ix[0] = NDArrayIndex.point(i);
            ix[1] = NDArrayIndex.interval(0, i);
            A.put(ix, 0);
        }
    } else {
        INDArrayIndex[] ix = new INDArrayIndex[2];
        for (int i = 0; i < Math.min(A.rows(), A.columns() - 1); i++) {
            ix[0] = NDArrayIndex.point(i);
            ix[1] = NDArrayIndex.interval(i + 1, A.columns());
            A.put(ix, 0);
        }
    }
    log.info("A: {}", A);
}

Example 3 with CUstream_st

use of org.bytedeco.javacpp.cuda.CUstream_st in project nd4j by deeplearning4j.

the class JcublasLapack method sgesvd.

@Override
public void sgesvd(byte jobu, byte jobvt, int M, int N, INDArray A, INDArray S, INDArray U, INDArray VT, INDArray INFO) {
    if (Nd4j.dataType() != DataBuffer.Type.FLOAT)
        log.warn("FLOAT gesvd called in DOUBLE environment");
    INDArray a = A;
    INDArray u = U;
    INDArray vt = VT;
    // we should transpose & adjust outputs if M<N
    // cuda has a limitation, but it's OK we know
    // A = U S V'
    // transpose multiply rules give us ...
    // A' = V S' U'
    boolean hadToTransposeA = false;
    if (M < N) {
        hadToTransposeA = true;
        int tmp1 = N;
        N = M;
        M = tmp1;
        a = A.transpose().dup('f');
        u = VT.dup('f');
        vt = U.dup('f');
    } else {
        // cuda requires column ordering - we'll register a warning in case
        if (A.ordering() == 'c')
            a = A.dup('f');
        if (U != null && U.ordering() == 'c')
            u = U.dup('f');
        if (VT != null && VT.ordering() == 'c')
            vt = VT.dup('f');
    }
    if (Nd4j.getExecutioner() instanceof GridExecutioner)
        ((GridExecutioner) Nd4j.getExecutioner()).flushQueue();
    // Get context for current thread
    CudaContext ctx = (CudaContext) allocator.getDeviceContext().getContext();
    // setup the solver handles for cuSolver calls
    cusolverDnHandle_t handle = ctx.getSolverHandle();
    cusolverDnContext solverDn = new cusolverDnContext(handle);
    // synchronized on the solver
    synchronized (handle) {
        int result = cusolverDnSetStream(new cusolverDnContext(handle), new CUstream_st(ctx.getOldStream()));
        if (result != 0)
            throw new BlasException("solverSetStream failed");
        // transfer the INDArray into GPU memory
        CublasPointer xAPointer = new CublasPointer(a, ctx);
        // this output - indicates how much memory we'll need for the real operation
        DataBuffer worksizeBuffer = Nd4j.getDataBufferFactory().createInt(1);
        int stat = cusolverDnSgesvd_bufferSize(// we intentionally use host pointer here
        solverDn, // we intentionally use host pointer here
        M, // we intentionally use host pointer here
        N, // we intentionally use host pointer here
        (IntPointer) worksizeBuffer.addressPointer());
        if (stat != CUSOLVER_STATUS_SUCCESS) {
            throw new BlasException("cusolverDnSgesvd_bufferSize failed", stat);
        }
        int worksize = worksizeBuffer.getInt(0);
        Pointer workspace = new Workspace(worksize * Nd4j.sizeOfDataType());
        DataBuffer rwork = Nd4j.getDataBufferFactory().createFloat((M < N ? M : N) - 1);
        // Do the actual decomp
        stat = cusolverDnSgesvd(solverDn, jobu, jobvt, M, N, (FloatPointer) xAPointer.getDevicePointer(), M, new CudaPointer(allocator.getPointer(S, ctx)).asFloatPointer(), U == null ? null : new CudaPointer(allocator.getPointer(u, ctx)).asFloatPointer(), M, VT == null ? null : new CudaPointer(allocator.getPointer(vt, ctx)).asFloatPointer(), N, new CudaPointer(workspace).asFloatPointer(), worksize, new CudaPointer(allocator.getPointer(rwork, ctx)).asFloatPointer(), new CudaPointer(allocator.getPointer(INFO, ctx)).asIntPointer());
        if (stat != CUSOLVER_STATUS_SUCCESS) {
            throw new BlasException("cusolverDnSgesvd failed", stat);
        }
    }
    allocator.registerAction(ctx, INFO);
    allocator.registerAction(ctx, S);
    if (U != null)
        allocator.registerAction(ctx, u);
    if (VT != null)
        allocator.registerAction(ctx, vt);
    // if we transposed A then swap & transpose U & V'
    if (hadToTransposeA) {
        U.assign(vt.transpose());
        VT.assign(u.transpose());
    } else {
        if (u != U)
            U.assign(u);
        if (vt != VT)
            VT.assign(vt);
    }
}

Example 4 with CUstream_st

use of org.bytedeco.javacpp.cuda.CUstream_st in project nd4j by deeplearning4j.

the class JcublasLapack method dpotrf.

@Override
public void dpotrf(byte uplo, int N, INDArray A, INDArray INFO) {
    INDArray a = A;
    if (Nd4j.dataType() != DataBuffer.Type.DOUBLE)
        log.warn("FLOAT potrf called in DOUBLE environment");
    if (A.ordering() == 'c')
        a = A.dup('f');
    if (Nd4j.getExecutioner() instanceof GridExecutioner)
        ((GridExecutioner) Nd4j.getExecutioner()).flushQueue();
    // Get context for current thread
    CudaContext ctx = (CudaContext) allocator.getDeviceContext().getContext();
    // setup the solver handles for cuSolver calls
    cusolverDnHandle_t handle = ctx.getSolverHandle();
    cusolverDnContext solverDn = new cusolverDnContext(handle);
    // synchronized on the solver
    synchronized (handle) {
        int result = cusolverDnSetStream(new cusolverDnContext(handle), new CUstream_st(ctx.getOldStream()));
        if (result != 0)
            throw new BlasException("solverSetStream failed");
        // transfer the INDArray into GPU memory
        CublasPointer xAPointer = new CublasPointer(a, ctx);
        // this output - indicates how much memory we'll need for the real operation
        DataBuffer worksizeBuffer = Nd4j.getDataBufferFactory().createInt(1);
        int stat = cusolverDnDpotrf_bufferSize(solverDn, uplo, N, (DoublePointer) xAPointer.getDevicePointer(), N, // we intentionally use host pointer here
        (IntPointer) worksizeBuffer.addressPointer());
        if (stat != CUSOLVER_STATUS_SUCCESS) {
            throw new BlasException("cusolverDnDpotrf_bufferSize failed", stat);
        }
        int worksize = worksizeBuffer.getInt(0);
        // Now allocate memory for the workspace, the permutation matrix and a return code
        Pointer workspace = new Workspace(worksize * Nd4j.sizeOfDataType());
        // Do the actual decomp
        stat = cusolverDnDpotrf(solverDn, uplo, N, (DoublePointer) xAPointer.getDevicePointer(), N, new CudaPointer(workspace).asDoublePointer(), worksize, new CudaPointer(allocator.getPointer(INFO, ctx)).asIntPointer());
        if (stat != CUSOLVER_STATUS_SUCCESS) {
            throw new BlasException("cusolverDnDpotrf failed", stat);
        }
    }
    allocator.registerAction(ctx, a);
    allocator.registerAction(ctx, INFO);
    if (a != A)
        A.assign(a);
    if (uplo == 'U') {
        A.assign(A.transpose());
        INDArrayIndex[] ix = new INDArrayIndex[2];
        for (int i = 1; i < Math.min(A.rows(), A.columns()); i++) {
            ix[0] = NDArrayIndex.point(i);
            ix[1] = NDArrayIndex.interval(0, i);
            A.put(ix, 0);
        }
    } else {
        INDArrayIndex[] ix = new INDArrayIndex[2];
        for (int i = 0; i < Math.min(A.rows(), A.columns() - 1); i++) {
            ix[0] = NDArrayIndex.point(i);
            ix[1] = NDArrayIndex.interval(i + 1, A.columns());
            A.put(ix, 0);
        }
    }
    log.info("A: {}", A);
}

Example 5 with CUstream_st

use of org.bytedeco.javacpp.cuda.CUstream_st in project nd4j by deeplearning4j.

the class JcublasLevel2 method sgemv.

@Override
protected void sgemv(char order, char TransA, int M, int N, float alpha, INDArray A, int lda, INDArray X, int incX, float beta, INDArray Y, int incY) {
    if (Nd4j.dataType() != DataBuffer.Type.FLOAT)
        logger.warn("FLOAT gemv called");
    Nd4j.getExecutioner().push();
    CudaContext ctx = allocator.getFlowController().prepareAction(Y, A, X);
    CublasPointer cAPointer = new CublasPointer(A, ctx);
    CublasPointer cBPointer = new CublasPointer(X, ctx);
    CublasPointer cCPointer = new CublasPointer(Y, ctx);
    cublasHandle_t handle = ctx.getHandle();
    synchronized (handle) {
        cublasSetStream_v2(new cublasContext(handle), new CUstream_st(ctx.getOldStream()));
        cublasSgemv_v2(new cublasContext(handle), convertTranspose(TransA), M, N, new FloatPointer(alpha), (FloatPointer) cAPointer.getDevicePointer(), lda, (FloatPointer) cBPointer.getDevicePointer(), incX, new FloatPointer(beta), (FloatPointer) cCPointer.getDevicePointer(), incY);
    }
    allocator.registerAction(ctx, Y, A, X);
    OpExecutionerUtil.checkForAny(Y);
}