Examples with CudaTensor com.simiacryptus.mindseye.lang.cudnn.CudaTensor used on opensource projects

Example 21 with CudaTensor

use of com.simiacryptus.mindseye.lang.cudnn.CudaTensor in project MindsEye by SimiaCryptus.

the class ImgTileSubnetLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    assert 1 == inObj.length;
    Result input = inObj[0];
    TensorList inputData = input.getData();
    @Nonnull final int[] inputDims = inputData.getDimensions();
    assert 3 == inputDims.length;
    int bands = inputDims[2];
    int length = inputData.length();
    CudaTensor passback = CudaSystem.run(gpu -> {
        return CudaTensor.wrap(gpu.allocate(inputData.getElements() * precision.size, MemoryType.Managed, true), gpu.newTensorDescriptor(precision, length, inputDims[2], inputDims[1], inputDims[0]), precision);
    });
    try {
        AtomicInteger counter = new AtomicInteger(0);
        int cols = (int) (Math.ceil((inputDims[0] - width) * 1.0 / strideX) + 1);
        int rows = (int) (Math.ceil((inputDims[1] - height) * 1.0 / strideY) + 1);
        if (cols == 1 && rows == 1)
            return getInner().evalAndFree(inObj);
        ArrayList<CudaTensor> tiles = new ArrayList<>();
        int[] tileDimensions = { width, height, bands };
        Result[][] tileResults = new Result[rows][];
        for (int row = 0; row < rows; row++) {
            tileResults[row] = new Result[cols];
            for (int col = 0; col < cols; col++) {
                int positionX = col * strideX;
                int positionY = row * strideY;
                assert positionX >= 0;
                assert positionY >= 0;
                assert positionX < inputDims[0];
                assert positionY < inputDims[1];
                CudaTensor tile = CudaSystem.run(gpu -> {
                    return ImgTileSelectLayer.copy(gpu, inputData, inputData.getDimensions(), tileDimensions, precision, positionX, positionY, true);
                });
                passback.addRef();
                tileResults[row][col] = getInner().evalAndFree(new Result(CudaTensorList.wrap(tile, length, tileDimensions, precision), (DeltaSet<Layer> ctx, TensorList delta) -> {
                    CudaSystem.run(gpu -> {
                        ImgTileSelectLayer.copy(gpu, delta, tileDimensions, -positionX, -positionY, precision, passback).freeRef();
                    });
                    if (counter.incrementAndGet() >= rows * cols) {
                        counter.set(0);
                        input.accumulate(ctx, CudaTensorList.create(passback, length, inputDims, precision));
                    }
                }) {

                    @Override
                    protected void _free() {
                        super._free();
                        passback.freeRef();
                    }
                });
            }
        }
        inputData.freeRef();
        logger.debug(String.format("Broke input %s into %s rows, %s cols", Arrays.toString(inputDims), rows, cols));
        Result result = new ImgTileAssemblyLayer(cols, rows).setParallel(parallel).setPrecision(precision).evalAndFree(Arrays.stream(tileResults).flatMap(Arrays::stream).toArray(i -> new Result[i]));
        return new Result(result.getData(), (ctx, delta) -> {
            result.accumulate(ctx, delta);
        }) {

            @Override
            public void accumulate(final DeltaSet<Layer> buffer, final TensorList delta) {
                getAccumulator().accept(buffer, delta);
            }

            @Override
            protected void _free() {
                super._free();
                result.freeRef();
                input.freeRef();
            }
        };
    } finally {
        passback.freeRef();
    }
}

Example 22 with CudaTensor

use of com.simiacryptus.mindseye.lang.cudnn.CudaTensor in project MindsEye by SimiaCryptus.

the class NProductLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    if (inObj.length <= 1) {
        throw new IllegalArgumentException("inObj.length=" + inObj.length);
    }
    @Nonnull final int[] dimensions = inObj[0].getData().getDimensions();
    final int length = inObj[0].getData().length();
    if (3 != dimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(dimensions));
    }
    for (int i = 1; i < inObj.length; i++) {
        TensorList data = inObj[i].getData();
        if (Tensor.length(dimensions) != Tensor.length(data.getDimensions())) {
            throw new IllegalArgumentException(Arrays.toString(dimensions) + " != " + Arrays.toString(data.getDimensions()));
        }
    }
    return new Result(CudaSystem.run(gpu -> {
        @Nonnull final CudaResource<cudnnOpTensorDescriptor> opDescriptor = gpu.newOpDescriptor(cudnnOpTensorOp.CUDNN_OP_TENSOR_MUL, precision);
        @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
        @Nonnull final TensorList result1 = Arrays.stream(inObj).map(x -> {
            TensorList data = x.getData();
            data.addRef();
            return data;
        }).reduce((l, r) -> {
            @Nullable final CudaTensor lPtr = gpu.getTensor(l, precision, MemoryType.Device, false);
            @Nullable final CudaTensor rPtr = gpu.getTensor(r, precision, MemoryType.Device, false);
            // assert lPtr.memory.size == rPtr.memory.size;
            @Nonnull final CudaMemory outputPtr = gpu.allocate((long) outputDescriptor.nStride * length * precision.size, MemoryType.Device, true);
            CudaMemory lPtrMemory = lPtr.getMemory(gpu);
            CudaMemory rPtrMemory = rPtr.getMemory(gpu);
            CudaSystem.handle(JCudnn.cudnnOpTensor(gpu.handle, opDescriptor.getPtr(), precision.getPointer(1.0), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(1.0), rPtr.descriptor.getPtr(), rPtrMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
            lPtrMemory.dirty();
            rPtrMemory.dirty();
            outputPtr.dirty();
            lPtrMemory.freeRef();
            rPtrMemory.freeRef();
            Arrays.stream(new ReferenceCounting[] { lPtr, rPtr, l, r }).forEach(ReferenceCounting::freeRef);
            outputDescriptor.addRef();
            return CudaTensorList.wrap(CudaTensor.wrap(outputPtr, outputDescriptor, precision), length, dimensions, precision);
        }).get();
        Arrays.stream(new ReferenceCounting[] { opDescriptor, outputDescriptor }).forEach(ReferenceCounting::freeRef);
        return result1;
    }, Arrays.stream(inObj).map(Result::getData).toArray()), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        for (int index = 0; index < inObj.length; index++) {
            final Result input = inObj[index];
            if (input.isAlive()) {
                final int _index = index;
                @Nonnull TensorList data = IntStream.range(0, inObj.length).mapToObj(i -> {
                    TensorList tensorList = i == _index ? delta : inObj[i].getData();
                    tensorList.addRef();
                    return tensorList;
                }).reduce((l, r) -> {
                    return CudaSystem.run(gpu -> {
                        @Nonnull final CudaResource<cudnnOpTensorDescriptor> opDescriptor = gpu.newOpDescriptor(cudnnOpTensorOp.CUDNN_OP_TENSOR_MUL, precision);
                        @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
                        @Nullable final CudaTensor lPtr = gpu.getTensor(l, precision, MemoryType.Device, false);
                        @Nullable final CudaTensor rPtr = gpu.getTensor(r, precision, MemoryType.Device, false);
                        // assert lPtr.memory.size == rPtr.memory.size;
                        @Nonnull final CudaMemory outputPtr = gpu.allocate((long) outputDescriptor.nStride * length * precision.size, MemoryType.Device, true);
                        CudaMemory lPtrMemory = lPtr.getMemory(gpu);
                        CudaMemory rPtrMemory = rPtr.getMemory(gpu);
                        CudaSystem.handle(JCudnn.cudnnOpTensor(gpu.handle, opDescriptor.getPtr(), precision.getPointer(1.0), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(1.0), rPtr.descriptor.getPtr(), rPtrMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
                        lPtrMemory.dirty();
                        rPtrMemory.dirty();
                        outputPtr.dirty();
                        lPtrMemory.freeRef();
                        rPtrMemory.freeRef();
                        Stream.of(lPtr, rPtr, opDescriptor, l, r).forEach(ReferenceCounting::freeRef);
                        return CudaTensorList.wrap(CudaTensor.wrap(outputPtr, outputDescriptor, precision), length, dimensions, precision);
                    }, l, r);
                }).get();
                input.accumulate(buffer, data);
            }
        }
        delta.freeRef();
    }) {

        @Override
        public final void accumulate(DeltaSet<Layer> buffer, TensorList delta) {
            getAccumulator().accept(buffer, delta);
        }

        @Override
        protected void _free() {
            Arrays.stream(inObj).forEach(nnResult -> nnResult.freeRef());
            for (int i = 0; i < inObj.length; i++) {
                inObj[i].getData().freeRef();
            }
        }

        @Override
        public boolean isAlive() {
            for (@Nonnull final Result element : inObj) if (element.isAlive()) {
                return true;
            }
            return false;
        }
    };
}

Example 23 with CudaTensor

use of com.simiacryptus.mindseye.lang.cudnn.CudaTensor in project MindsEye by SimiaCryptus.

the class ProductLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    if (inObj.length != 2) {
        throw new IllegalArgumentException("inObj.length=" + inObj.length);
    }
    Result left = inObj[0];
    Result right = inObj[1];
    final TensorList leftData = left.getData();
    final TensorList rightData = right.getData();
    @Nonnull final int[] leftDimensions = leftData.getDimensions();
    @Nonnull final int[] rightDimensions = rightData.getDimensions();
    final int length = leftData.length();
    if (3 != leftDimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(leftDimensions));
    }
    return new Result(CudaSystem.run(gpu -> {
        @Nonnull final CudaResource<cudnnOpTensorDescriptor> opDescriptor = gpu.newOpDescriptor(cudnnOpTensorOp.CUDNN_OP_TENSOR_MUL, precision);
        @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, leftDimensions[2], leftDimensions[1], leftDimensions[0], leftDimensions[2] * leftDimensions[1] * leftDimensions[0], leftDimensions[1] * leftDimensions[0], leftDimensions[0], 1);
        @Nullable final CudaTensor lPtr = gpu.getTensor(leftData, precision, MemoryType.Device, false);
        @Nullable final CudaTensor rPtr = gpu.getTensor(rightData, precision, MemoryType.Device, false);
        // assert lPtr.size == rPtr.size;
        @Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * outputDescriptor.nStride * length, MemoryType.Device, true);
        CudaMemory lPtrMemory = lPtr.getMemory(gpu);
        CudaMemory rPtrMemory = rPtr.getMemory(gpu);
        CudaSystem.handle(gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(1.0), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(1.0), rPtr.descriptor.getPtr(), rPtrMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        lPtrMemory.dirty();
        rPtrMemory.dirty();
        outputPtr.dirty();
        lPtrMemory.freeRef();
        rPtrMemory.freeRef();
        rPtr.freeRef();
        lPtr.freeRef();
        opDescriptor.freeRef();
        CudaTensor cudaTensor = CudaTensor.wrap(outputPtr, outputDescriptor, precision);
        return CudaTensorList.wrap(cudaTensor, length, leftDimensions, precision);
    }, leftData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        if (left.isAlive()) {
            @Nonnull TensorList data = CudaSystem.run(gpu -> {
                @Nonnull final CudaResource<cudnnOpTensorDescriptor> opDescriptor = gpu.newOpDescriptor(cudnnOpTensorOp.CUDNN_OP_TENSOR_MUL, precision);
                @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, leftDimensions[2], leftDimensions[1], leftDimensions[0], leftDimensions[2] * leftDimensions[1] * leftDimensions[0], leftDimensions[1] * leftDimensions[0], leftDimensions[0], 1);
                @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
                @Nullable final CudaTensor rightTensor = gpu.getTensor(right.getData(), precision, MemoryType.Device, false);
                // assert deltaTensor.size == rightTensor.size;
                @Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * outputDescriptor.nStride * length, MemoryType.Device, true);
                CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                CudaMemory rightTensorMemory = rightTensor.getMemory(gpu);
                CudaSystem.handle(gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), precision.getPointer(1.0), rightTensor.descriptor.getPtr(), rightTensorMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
                deltaTensorMemory.dirty();
                rightTensorMemory.dirty();
                outputPtr.dirty();
                deltaTensorMemory.freeRef();
                rightTensorMemory.freeRef();
                CudaTensor cudaTensor = new CudaTensor(outputPtr, outputDescriptor, precision);
                Arrays.stream(new ReferenceCounting[] { deltaTensor, rightTensor, opDescriptor, outputDescriptor }).forEach(ReferenceCounting::freeRef);
                outputPtr.freeRef();
                return CudaTensorList.wrap(cudaTensor, length, leftDimensions, precision);
            }, delta);
            left.accumulate(buffer, data);
        }
        if (right.isAlive()) {
            @Nonnull TensorList data = CudaSystem.run(gpu -> {
                @Nonnull final CudaResource<cudnnOpTensorDescriptor> opDescriptor = gpu.newOpDescriptor(cudnnOpTensorOp.CUDNN_OP_TENSOR_MUL, precision);
                @Nonnull final CudaDevice.CudaTensorDescriptor expandedDescriptor = gpu.newTensorDescriptor(precision, length, leftDimensions[2], leftDimensions[1], leftDimensions[0], leftDimensions[2] * leftDimensions[1] * leftDimensions[0], leftDimensions[1] * leftDimensions[0], leftDimensions[0], 1);
                @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
                delta.freeRef();
                @Nullable final CudaTensor leftTensor = gpu.getTensor(left.getData(), precision, MemoryType.Device, false);
                // assert deltaTensor.size == rightTensor.size;
                @Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * expandedDescriptor.nStride * length, MemoryType.Device, true);
                CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                CudaMemory leftTensorMemory = leftTensor.getMemory(gpu);
                CudaSystem.handle(gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), precision.getPointer(1.0), leftTensor.descriptor.getPtr(), leftTensorMemory.getPtr(), precision.getPointer(0.0), expandedDescriptor.getPtr(), outputPtr.getPtr()));
                deltaTensorMemory.dirty();
                leftTensorMemory.dirty();
                outputPtr.dirty();
                if (Arrays.equals(rightDimensions, leftDimensions) && length == rightData.length()) {
                    deltaTensorMemory.freeRef();
                    leftTensorMemory.freeRef();
                    assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
                    outputPtr.dirty();
                    CudaTensor cudaTensor = new CudaTensor(outputPtr, expandedDescriptor, precision);
                    Stream.of(deltaTensor, leftTensor, opDescriptor, expandedDescriptor, outputPtr).forEach(ReferenceCounting::freeRef);
                    CudaTensorList tensorList = CudaTensorList.wrap(cudaTensor, length, rightDimensions, precision);
                    return tensorList;
                } else {
                    @Nonnull final CudaDevice.CudaTensorDescriptor reducedOutputDescriptor = gpu.newTensorDescriptor(precision, rightData.length(), rightDimensions[2], rightDimensions[1], rightDimensions[0], rightDimensions[2] * rightDimensions[1] * rightDimensions[0], rightDimensions[1] * rightDimensions[0], rightDimensions[0], 1);
                    long size = (long) precision.size * reducedOutputDescriptor.nStride * rightData.length();
                    @Nonnull final CudaMemory reducedOutputPtr = gpu.allocate(size, MemoryType.Managed, true);
                    CudaResource<cudnnReduceTensorDescriptor> reduceTensorDescriptor = gpu.cudnnCreateReduceTensorDescriptor(cudnnReduceTensorOp.CUDNN_REDUCE_TENSOR_ADD, precision.code, cudnnNanPropagation.CUDNN_NOT_PROPAGATE_NAN, cudnnReduceTensorIndices.CUDNN_REDUCE_TENSOR_NO_INDICES, cudnnIndicesType.CUDNN_32BIT_INDICES);
                    @Nonnull final CudaMemory workspacePtr = gpu.allocate(outputPtr.size, MemoryType.Device, true);
                    @Nonnull final CudaMemory indexPtr = gpu.allocate(3, MemoryType.Device, false);
                    // outputPtr.synchronize();
                    gpu.cudnnReduceTensor(reduceTensorDescriptor.getPtr(), indexPtr.getPtr(), indexPtr.size, workspacePtr.getPtr(), workspacePtr.size, precision.getPointer(1.0), expandedDescriptor.getPtr(), outputPtr.getPtr(), precision.getPointer(0.0), reducedOutputDescriptor.getPtr(), reducedOutputPtr.getPtr());
                    reducedOutputPtr.dirty();
                    workspacePtr.dirty();
                    outputPtr.dirty();
                    deltaTensorMemory.freeRef();
                    leftTensorMemory.freeRef();
                    CudaTensor cudaTensor = new CudaTensor(reducedOutputPtr, reducedOutputDescriptor, precision);
                    Stream.of(deltaTensor, leftTensor, opDescriptor, expandedDescriptor, outputPtr, reducedOutputPtr, reducedOutputDescriptor, reduceTensorDescriptor, workspacePtr, indexPtr).forEach(ReferenceCounting::freeRef);
                    CudaTensorList tensorList = CudaTensorList.wrap(cudaTensor, rightData.length(), rightDimensions, precision);
                    return tensorList;
                }
            }, delta);
            right.accumulate(buffer, data);
        } else {
            delta.freeRef();
        }
    }) {

        @Override
        public void accumulate(final DeltaSet<Layer> buffer, final TensorList delta) {
            getAccumulator().accept(buffer, delta);
        }

        @Override
        protected void _free() {
            leftData.freeRef();
            rightData.freeRef();
            left.freeRef();
            right.freeRef();
        }

        @Override
        public boolean isAlive() {
            for (@Nonnull final Result element : inObj) if (element.isAlive()) {
                return true;
            }
            return false;
        }
    };
}

Example 24 with CudaTensor

use of com.simiacryptus.mindseye.lang.cudnn.CudaTensor in project MindsEye by SimiaCryptus.

the class SimpleConvolutionLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().eval(inObj);
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    @Nonnull final int[] kernelSize = kernel.getDimensions();
    final int[] outputSize = getOutputSize(inputSize);
    final int length = inputData.length();
    kernel.addRef();
    SimpleConvolutionLayer.this.addRef();
    return new Result(CudaSystem.run(gpu -> {
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        final CudaResource<cudnnFilterDescriptor> filterDescriptor = gpu.newFilterDescriptor(precision, cudnnTensorFormat.CUDNN_TENSOR_NCHW, outputSize[2], inputSize[2], kernelSize[1], kernelSize[0]);
        final CudaResource<cudnnConvolutionDescriptor> convolutionDescriptor = gpu.newConvolutions2dDescriptor(cudnnConvolutionMode.CUDNN_CONVOLUTION, precision, paddingY, paddingX, strideY, strideX, 1, 1);
        final int[] outputDims = IntStream.of(reverse(CudaSystem.getOutputDims(inputTensor.descriptor.getPtr(), filterDescriptor.getPtr(), convolutionDescriptor.getPtr()))).limit(3).toArray();
        final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, outputDims[2], outputDims[1], outputDims[0], outputDims[2] * outputDims[1] * outputDims[0], outputDims[1] * outputDims[0], outputDims[0], 1);
        final int forwardAlgorithm = getForwardAlgorithm(gpu, inputTensor, filterDescriptor, convolutionDescriptor, outputDescriptor);
        final CudaMemory forwardWorkspace = gpu.allocateForwardWorkspace(inputTensor.descriptor.getPtr(), filterDescriptor.getPtr(), convolutionDescriptor.getPtr(), outputDescriptor.getPtr(), forwardAlgorithm, 1);
        try {
            assert 0 < kernel.getData().length;
            assert kernelSize[0] * kernelSize[1] * kernelSize[2] == kernel.getData().length;
            @Nonnull CudaMemory filterPtr = getCudaFilter(gpu);
            @Nonnull final CudaMemory outputBuffer = gpu.allocate((long) Tensor.length(outputDims) * length * precision.size, MemoryType.Managed.normalize(), true);
            CudaMemory inputTensorMemory = inputTensor.getMemory(gpu);
            // inputTensorMemory.synchronize();
            CudaSystem.handle(gpu.cudnnConvolutionForward(precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputTensorMemory.getPtr(), filterDescriptor.getPtr(), filterPtr.getPtr(), convolutionDescriptor.getPtr(), forwardAlgorithm, null == forwardWorkspace ? null : forwardWorkspace.getPtr(), null == forwardWorkspace ? 0 : forwardWorkspace.size, precision.getPointer(0.0), outputDescriptor.getPtr(), outputBuffer.getPtr()));
            assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
            forwardWorkspace.dirty();
            filterPtr.dirty();
            outputBuffer.dirty();
            inputTensorMemory.dirty();
            // inputTensorMemory.synchronize();
            inputTensorMemory.freeRef();
            filterPtr.freeRef();
            outputDescriptor.addRef();
            return CudaTensorList.wrap(CudaTensor.wrap(outputBuffer, outputDescriptor, precision), length, outputDims, precision);
        } catch (@Nonnull final Throwable e) {
            throw new ComponentException(String.format("Error in convolution %s x %s", Arrays.toString(inputSize), Arrays.toString(kernelSize)), e);
        } finally {
            Stream.of(inputTensor, filterDescriptor, outputDescriptor, forwardWorkspace, convolutionDescriptor).forEach(ReferenceCounting::freeRef);
        }
    }, inputData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        delta.assertAlive();
        buffer.assertAlive();
        inputData.assertAlive();
        assert delta.length() == length;
        delta.addRef();
        Runnable learnFn = () -> {
            if (!isFrozen()) {
                @Nonnull final Tensor weightGradient = CudaSystem.run(gpu -> {
                    @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
                    delta.freeRef();
                    @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
                    final CudaResource<cudnnFilterDescriptor> filterDescriptor = gpu.newFilterDescriptor(precision, cudnnTensorFormat.CUDNN_TENSOR_NCHW, outputSize[2], inputSize[2], kernelSize[1], kernelSize[0]);
                    final CudaResource<cudnnConvolutionDescriptor> convolutionDescriptor = gpu.newConvolutions2dDescriptor(cudnnConvolutionMode.CUDNN_CONVOLUTION, precision, paddingY, paddingX, strideY, strideX, 1, 1);
                    final int backwardFilterAlgorithm = getBackwardFilterAlgorithm(gpu, deltaTensor, inputTensor, filterDescriptor, convolutionDescriptor);
                    final CudaMemory backwardsFilterWorkSpace = gpu.allocateBackwardFilterWorkspace(inputTensor.descriptor.getPtr(), filterDescriptor.getPtr(), convolutionDescriptor.getPtr(), deltaTensor.descriptor.getPtr(), backwardFilterAlgorithm, 1);
                    @Nonnull CudaMemory filterPtr = gpu.allocate((long) kernel.length() * precision.size, MemoryType.Device, true);
                    try {
                        CudaMemory inputTensorMemory = inputTensor.getMemory(gpu);
                        CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu, MemoryType.Managed.normalize());
                        // inputTensorMemory.synchronize();
                        CudaSystem.handle(gpu.cudnnConvolutionBackwardFilter(precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputTensorMemory.getPtr(), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), convolutionDescriptor.getPtr(), backwardFilterAlgorithm, backwardsFilterWorkSpace.getPtr(), backwardsFilterWorkSpace.size, precision.getPointer(0.0), filterDescriptor.getPtr(), filterPtr.getPtr()));
                        filterPtr.dirty();
                        deltaTensorMemory.dirty();
                        inputTensorMemory.dirty();
                        backwardsFilterWorkSpace.dirty();
                        // backwardsFilterWorkSpace.synchronize();
                        inputTensorMemory.freeRef();
                        deltaTensorMemory.freeRef();
                        return filterPtr.read(precision, kernel.getDimensions());
                    } catch (@Nonnull final Throwable e) {
                        throw new ComponentException(String.format("Error in convolution %s x %s => %s", Arrays.toString(inputSize), Arrays.toString(kernelSize), Arrays.toString(outputSize)), e);
                    } finally {
                        inputTensor.freeRef();
                        filterPtr.freeRef();
                        deltaTensor.freeRef();
                        Stream.of(filterDescriptor, convolutionDescriptor, backwardsFilterWorkSpace).forEach(ReferenceCounting::freeRef);
                    }
                }, delta);
                buffer.get(SimpleConvolutionLayer.this, kernel.getData()).addInPlace(weightGradient.getData()).freeRef();
                weightGradient.freeRef();
                clearCudaFilters();
            } else {
                delta.freeRef();
            }
        };
        Runnable backpropFn = () -> {
            if (input.isAlive()) {
                final TensorList inputBufferTensors = CudaSystem.run(gpu -> {
                    final CudaDevice.CudaTensorDescriptor inputDescriptor = gpu.newTensorDescriptor(precision, length, inputSize[2], inputSize[1], inputSize[0], inputSize[2] * inputSize[1] * inputSize[0], inputSize[1] * inputSize[0], inputSize[0], 1);
                    final CudaResource<cudnnFilterDescriptor> filterDescriptor = gpu.newFilterDescriptor(precision, cudnnTensorFormat.CUDNN_TENSOR_NCHW, outputSize[2], inputSize[2], kernelSize[1], kernelSize[0]);
                    final CudaResource<cudnnConvolutionDescriptor> convolutionDescriptor = gpu.newConvolutions2dDescriptor(cudnnConvolutionMode.CUDNN_CONVOLUTION, precision, paddingY, paddingX, strideY, strideX, 1, 1);
                    @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
                    delta.freeRef();
                    final int backwardDataAlgorithm = getBackwardDataAlgorithm(gpu, inputDescriptor, filterDescriptor, convolutionDescriptor, deltaTensor);
                    final CudaMemory backwardsDataWorkSpace = gpu.allocateBackwardDataWorkspace(inputDescriptor.getPtr(), filterDescriptor.getPtr(), convolutionDescriptor.getPtr(), deltaTensor.descriptor.getPtr(), backwardDataAlgorithm, 1);
                    @Nonnull final CudaMemory filterPtr = getCudaFilter(gpu);
                    try {
                        @Nonnull final CudaMemory passbackMemory = gpu.allocate((long) Tensor.length(inputData.getDimensions()) * length * precision.size, MemoryType.Managed.normalize(), true);
                        CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                        // deltaTensorMemory.synchronize();
                        CudaSystem.handle(gpu.cudnnConvolutionBackwardData(precision.getPointer(1.0), filterDescriptor.getPtr(), filterPtr.getPtr(), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), convolutionDescriptor.getPtr(), backwardDataAlgorithm, backwardsDataWorkSpace.getPtr(), backwardsDataWorkSpace.size, precision.getPointer(0.0), inputDescriptor.getPtr(), passbackMemory.getPtr()));
                        passbackMemory.dirty();
                        backwardsDataWorkSpace.dirty();
                        deltaTensorMemory.dirty();
                        // deltaTensorMemory.synchronize();
                        filterPtr.dirty();
                        deltaTensorMemory.freeRef();
                        inputDescriptor.addRef();
                        return CudaTensorList.wrap(CudaTensor.wrap(passbackMemory, inputDescriptor, precision), length, inputSize, precision);
                    } catch (@Nonnull final Throwable e) {
                        throw new ComponentException(String.format("Error in convolution %s x %s => %s", Arrays.toString(inputSize), Arrays.toString(kernelSize), Arrays.toString(outputSize)), e);
                    } finally {
                        filterPtr.freeRef();
                        deltaTensor.freeRef();
                        Stream.of(inputDescriptor, filterDescriptor, convolutionDescriptor, backwardsDataWorkSpace).forEach(ReferenceCounting::freeRef);
                    }
                }, delta);
                if (null != inputBufferTensors) {
                    input.accumulate(buffer, inputBufferTensors);
                }
            } else {
                delta.freeRef();
            }
        };
        Stream.of(learnFn, backpropFn).forEach(Runnable::run);
    }) {

        @Override
        public final void accumulate(DeltaSet<Layer> buffer, TensorList delta) {
            getAccumulator().accept(buffer, delta);
        }

        @Override
        protected void _free() {
            kernel.freeRef();
            inputData.freeRef();
            Arrays.stream(inObj).forEach(ReferenceCounting::freeRef);
            SimpleConvolutionLayer.this.freeRef();
        }

        @Override
        public boolean isAlive() {
            return input.isAlive() || !isFrozen();
        }
    };
}

Example 25 with CudaTensor

use of com.simiacryptus.mindseye.lang.cudnn.CudaTensor in project MindsEye by SimiaCryptus.

the class SoftmaxActivationLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    final Result inputResult = inObj[0];
    final TensorList inputData = inputResult.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    @Nonnull final int[] outputSize = inputSize;
    final int length = inputData.length();
    final int inputDims = Tensor.length(inputSize);
    try {
        final CudaTensor outPtr = CudaSystem.run(gpu -> {
            @Nullable CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
            final CudaTensor outputTensor;
            if (1 == inputData.currentRefCount() && 1 == inputTensor.currentRefCount()) {
                outputTensor = inputTensor;
                outputTensor.addRef();
            } else {
                @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, inputSize[2], inputSize[1], inputSize[0], inputSize[2] * inputSize[1] * inputSize[0], inputSize[1] * inputSize[0], inputSize[0], 1);
                @Nonnull final CudaMemory outputData = gpu.allocate(precision.size * 1l * inputDims * length, MemoryType.Managed.normalize(), true);
                outputTensor = CudaTensor.wrap(outputData, outputDescriptor, precision);
            }
            try {
                CudaMemory inputMemory = inputTensor.getMemory(gpu);
                CudaMemory outputMemory = outputTensor.getMemory(gpu);
                CudaSystem.handle(gpu.cudnnSoftmaxForward(algorithm.code, mode.code, precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputMemory.getPtr(), precision.getPointer(0.0), outputTensor.descriptor.getPtr(), outputMemory.getPtr()));
                assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
                inputMemory.dirty();
                outputMemory.dirty();
                outputMemory.freeRef();
                inputMemory.freeRef();
                return outputTensor;
            } catch (@Nonnull final Throwable e) {
                throw new ComponentException("Error apply " + Arrays.toString(inputSize), e);
            } finally {
                inputTensor.freeRef();
            }
        }, inputData);
        return new Result(CudaTensorList.create(outPtr, length, outputSize, precision), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
            if (inputResult.isAlive()) {
                final TensorList data = CudaSystem.run(gpu -> {
                    @Nullable CudaTensor inputTensor;
                    synchronized (gpu) {
                        inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, true);
                    }
                    @Nullable CudaTensor deltaTensor;
                    synchronized (gpu) {
                        deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
                    }
                    outPtr.addRef();
                    CudaTensor localOut = outPtr.getDenseAndFree(gpu);
                    delta.freeRef();
                    CudaTensor passbackTensor;
                    passbackTensor = CudaTensor.wrap(gpu.allocate((long) Tensor.length(inputSize) * length * precision.size, MemoryType.Managed.normalize(), false), gpu.newTensorDescriptor(precision, delta.length(), inputSize[2], inputSize[1], inputSize[0], inputSize[2] * inputSize[1] * inputSize[0], inputSize[1] * inputSize[0], inputSize[0], 1), precision);
                    try {
                        CudaMemory localOutMemory = localOut.getMemory(gpu);
                        CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                        CudaMemory inputMemory = inputTensor.getMemory(gpu);
                        CudaMemory passbackMemory = passbackTensor.getMemory(gpu);
                        CudaSystem.handle(gpu.cudnnSoftmaxBackward(algorithm.code, mode.code, precision.getPointer(1.0), localOut.descriptor.getPtr(), localOutMemory.getPtr(), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), precision.getPointer(0.0), passbackTensor.descriptor.getPtr(), passbackMemory.getPtr()));
                        localOutMemory.dirty();
                        deltaTensorMemory.dirty();
                        passbackMemory.dirty();
                        localOutMemory.freeRef();
                        deltaTensorMemory.freeRef();
                        inputMemory.freeRef();
                        passbackMemory.freeRef();
                    } catch (@Nonnull final Throwable e) {
                        throw new ComponentException("Error apply " + Arrays.toString(inputSize), e);
                    } finally {
                        localOut.freeRef();
                        inputTensor.freeRef();
                        deltaTensor.freeRef();
                    }
                    return CudaTensorList.wrap(passbackTensor, length, inputSize, precision);
                }, delta);
                inputResult.accumulate(buffer, data);
            } else {
                delta.freeRef();
            }
        }) {

            @Override
            public final void accumulate(DeltaSet<Layer> buffer, TensorList delta) {
                getAccumulator().accept(buffer, delta);
            }

            @Override
            protected void _free() {
                inputData.freeRef();
                outPtr.freeRef();
                inputResult.freeRef();
            }

            @Override
            public boolean isAlive() {
                return inputResult.isAlive() || !isFrozen();
            }
        };
    } catch (@Nonnull final Throwable e) {
        throw new ComponentException("Error apply image res " + Arrays.toString(inputSize), e);
    }
}