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

Example 11 with CudaMemory

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

the class PoolingLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    final int poolDims = 2;
    @Nonnull final int[] windowSize = { windowX, windowY };
    @Nonnull final int[] padding = { paddingX, paddingY };
    @Nonnull final int[] stride = { strideX, strideY };
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    final int length = inputData.length();
    final int inputDims = Tensor.length(inputSize);
    @Nonnull final int[] outputSize = new int[4];
    final CudaTensor outputData = CudaSystem.run(gpu -> {
        try {
            gpu.initThread();
            @Nonnull final CudaResource<cudnnPoolingDescriptor> poolingDesc = gpu.createPoolingDescriptor(mode.id, poolDims, windowSize, padding, stride);
            @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
            CudaSystem.handle(CudaSystem.cudnnGetPoolingNdForwardOutputDim(poolingDesc.getPtr(), inputTensor.descriptor.getPtr(), 4, outputSize));
            assert inputSize[2] == outputSize[1];
            @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, outputSize[0], outputSize[1], outputSize[2], outputSize[3], outputSize[1] * outputSize[2] * outputSize[3], outputSize[2] * outputSize[3], outputSize[3], 1);
            @Nonnull final CudaMemory outputTensor = gpu.allocate((long) precision.size * Tensor.length(outputSize), MemoryType.Managed.normalize(), true);
            CudaMemory inputDataMemory = inputTensor.getMemory(gpu);
            CudaSystem.handle(gpu.cudnnPoolingForward(poolingDesc.getPtr(), precision.getPointer(alpha), inputTensor.descriptor.getPtr(), inputDataMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputTensor.getPtr()));
            assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
            inputDataMemory.dirty();
            outputTensor.dirty();
            Stream.<ReferenceCounting>of(inputTensor, poolingDesc, inputDataMemory).forEach(ReferenceCounting::freeRef);
            return CudaTensor.wrap(outputTensor, outputDescriptor, precision);
        } catch (@Nonnull final Throwable e) {
            throw new ComponentException("Error", e);
        }
    }, inputData);
    return new Result(CudaTensorList.create(outputData, length, new int[] { outputSize[3], outputSize[2], outputSize[1] }, precision), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList error) -> {
        assert error.length() == inputData.length();
        if (input.isAlive()) {
            TensorList data = CudaSystem.run(gpu -> {
                @Nonnull final CudaDevice.CudaTensorDescriptor passbackDescriptor = 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 CudaResource<cudnnPoolingDescriptor> poolingDesc = gpu.createPoolingDescriptor(mode.id, poolDims, windowSize, padding, stride);
                @Nullable final CudaTensor inputTensor;
                synchronized (gpu) {
                    inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, true);
                }
                @Nullable final CudaTensor errorPtr;
                synchronized (gpu) {
                    errorPtr = gpu.getTensor(error, precision, MemoryType.Device, true);
                }
                @Nonnull final CudaMemory passbackBuffer = gpu.allocate((long) inputDims * precision.size * length, MemoryType.Managed.normalize(), true);
                CudaMemory outputDataMemory = outputData.getMemory(gpu);
                CudaMemory errorPtrMemory = errorPtr.getMemory(gpu);
                CudaMemory inputDataMemory = inputTensor.getMemory(gpu);
                CudaSystem.handle(gpu.cudnnPoolingBackward(poolingDesc.getPtr(), precision.getPointer(this.alpha), outputData.descriptor.getPtr(), outputDataMemory.getPtr(), errorPtr.descriptor.getPtr(), errorPtrMemory.getPtr(), inputTensor.descriptor.getPtr(), inputDataMemory.getPtr(), precision.getPointer(0.0), passbackDescriptor.getPtr(), passbackBuffer.getPtr()));
                outputDataMemory.dirty();
                errorPtrMemory.dirty();
                inputDataMemory.dirty();
                passbackBuffer.dirty();
                Stream.<ReferenceCounting>of(errorPtr, inputTensor, poolingDesc, outputDataMemory, errorPtrMemory, inputDataMemory).forEach(ReferenceCounting::freeRef);
                return CudaTensorList.wrap(CudaTensor.wrap(passbackBuffer, passbackDescriptor, precision), length, inputSize, precision);
            }, error);
            input.accumulate(buffer, data);
        }
    }) {

        @Override
        protected void _free() {
            Arrays.stream(inObj).forEach(nnResult -> nnResult.freeRef());
            inputData.freeRef();
            outputData.freeRef();
        }

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

Example 12 with CudaMemory

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

the class SquareActivationLayer 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);
    }
    Result input = inObj[0];
    final TensorList inputData = input.getData();
    @Nonnull final int[] dimensions = inputData.getDimensions();
    final int length = inputData.length();
    if (3 != dimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(dimensions));
    }
    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);
        @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        // assert inputTensor.size == rPtr.size;
        @Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * outputDescriptor.nStride * length, MemoryType.Device, true);
        CudaMemory lPtrMemory = inputTensor.getMemory(gpu);
        CudaSystem.handle(gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(alpha), inputTensor.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(1.0), inputTensor.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        outputPtr.dirty();
        lPtrMemory.dirty();
        outputPtr.dirty();
        lPtrMemory.freeRef();
        inputTensor.freeRef();
        opDescriptor.freeRef();
        CudaTensor cudaTensor = CudaTensor.wrap(outputPtr, outputDescriptor, precision);
        return CudaTensorList.wrap(cudaTensor, length, dimensions, precision);
    }, inputData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        if (input.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, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
                @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
                delta.freeRef();
                @Nullable final CudaTensor inputTensor = gpu.getTensor(input.getData(), precision, MemoryType.Device, false);
                // assert deltaTensor.size == inputTensor.size;
                @Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * outputDescriptor.nStride * length, MemoryType.Device, true);
                CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                CudaMemory rightTensorMemory = inputTensor.getMemory(gpu);
                CudaSystem.handle(gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(2), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), precision.getPointer(alpha), inputTensor.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, inputTensor, opDescriptor, outputDescriptor }).forEach(ReferenceCounting::freeRef);
                outputPtr.freeRef();
                return CudaTensorList.wrap(cudaTensor, length, dimensions, precision);
            }, delta);
            input.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() {
            inputData.freeRef();
            input.freeRef();
        }

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

Example 13 with CudaMemory

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

the class AvgReducerLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    int length = inputData.length();
    CudaTensorList result = CudaSystem.run(gpu -> {
        CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        inputData.freeRef();
        CudaMemory inputMemory = inputTensor.getMemory(gpu);
        @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, 1, 1, 1);
        long size = (long) precision.size * outputDescriptor.nStride * length;
        @Nonnull final CudaMemory outputMemory = gpu.allocate(size, MemoryType.Managed, true);
        CudaResource<cudnnReduceTensorDescriptor> reduceTensorDescriptor = gpu.cudnnCreateReduceTensorDescriptor(cudnnReduceTensorOp.CUDNN_REDUCE_TENSOR_AVG, precision.code, cudnnNanPropagation.CUDNN_NOT_PROPAGATE_NAN, cudnnReduceTensorIndices.CUDNN_REDUCE_TENSOR_NO_INDICES, cudnnIndicesType.CUDNN_32BIT_INDICES);
        @Nonnull final CudaMemory workspacePtr = gpu.allocate(inputMemory.size, MemoryType.Device, true);
        @Nonnull final CudaMemory indexPtr = gpu.allocate(12 * length, MemoryType.Device, false);
        // outputPtr.synchronize();
        gpu.cudnnReduceTensor(reduceTensorDescriptor.getPtr(), indexPtr.getPtr(), indexPtr.size, workspacePtr.getPtr(), workspacePtr.size, precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputMemory.getPtr());
        outputMemory.dirty();
        inputMemory.dirty();
        Stream.of(inputTensor, inputMemory, reduceTensorDescriptor, workspacePtr, indexPtr).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(CudaTensor.wrap(outputMemory, outputDescriptor, precision), length, new int[] { 1, 1, 1 }, precision);
    });
    return new Result(result, (DeltaSet<Layer> ctx, TensorList delta) -> {
        // Not supported by CuDNN?
        // CudaTensorList passback = CudaSystem.run(gpu -> {
        // CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
        // CudaMemory deltaMemory = deltaTensor.getMemory(gpu);
        // 
        // @Nonnull final CudaDevice.CudaTensorDescriptor passbackDescriptor1 = gpu.newTensorDescriptor(
        // precision, length, inputSize[2], inputSize[1], inputSize[0]
        // );
        // @Nonnull final CudaMemory passbackPtr1 = gpu.allocate((long) precision.size * passbackDescriptor1.nStride * length, MemoryType.Device, false);
        // gpu.cudnnAddTensor(precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), deltaMemory.getPtr(),
        // precision.getPointer(1.0), passbackDescriptor1.getPtr(), passbackPtr1.getPtr());
        // passbackPtr1.dirty();
        // 
        // Stream.of(deltaTensor, deltaMemory, passbackDescriptor1, passbackPtr1).forEach(ReferenceCounting::freeRef);
        // return CudaTensorList.wrap(CudaTensor.wrap(passbackPtr1, passbackDescriptor1, precision), length, inputSize, precision);
        // });
        TensorList passback = TensorArray.wrap(IntStream.range(0, length).mapToObj(i -> {
            Tensor tensor = delta.get(i);
            Tensor tensor1 = new Tensor(inputSize).setAll((double) tensor.get(0) / Tensor.length(inputSize));
            tensor.freeRef();
            return tensor1;
        }).toArray(i -> new Tensor[i]));
        input.accumulate(ctx, passback);
    }) {

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

Example 14 with CudaMemory

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

the class GateBiasLayer 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_ADD, 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()) {
            delta.addRef();
            left.accumulate(buffer, delta);
        }
        if (right.isAlive()) {
            @Nonnull TensorList data = CudaSystem.run(gpu -> {
                // assert deltaTensor.size == rightTensor.size;
                if (Arrays.equals(rightDimensions, leftDimensions) && length == rightData.length()) {
                    assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
                    delta.addRef();
                    return delta;
                } 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);
                    @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
                    CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                    @Nonnull final CudaMemory workspacePtr = gpu.allocate(deltaTensorMemory.size, MemoryType.Device, true);
                    @Nonnull final CudaMemory indexPtr = gpu.allocate(12 * delta.length(), MemoryType.Device, false);
                    delta.freeRef();
                    // outputPtr.synchronize();
                    gpu.cudnnReduceTensor(reduceTensorDescriptor.getPtr(), indexPtr.getPtr(), indexPtr.size, workspacePtr.getPtr(), workspacePtr.size, precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), precision.getPointer(0.0), reducedOutputDescriptor.getPtr(), reducedOutputPtr.getPtr());
                    reducedOutputPtr.dirty();
                    deltaTensorMemory.dirty();
                    Stream.of(deltaTensorMemory, deltaTensor, reduceTensorDescriptor, workspacePtr, indexPtr).forEach(ReferenceCounting::freeRef);
                    return CudaTensorList.wrap(CudaTensor.wrap(reducedOutputPtr, reducedOutputDescriptor, precision), rightData.length(), rightDimensions, precision);
                }
            }, delta);
            right.accumulate(buffer, data);
        } else {
            delta.freeRef();
        }
    }) {

        @Override
        public final void accumulate(DeltaSet<Layer> buffer, 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 15 with CudaMemory

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

the class ImgBandSelectLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    // assert Arrays.stream(inObj).flatMapToDouble(input->input.data.stream().flatMapToDouble(x-> Arrays.stream(x.getData()))).allMatch(v->Double.isFinite(v));
    assert getFrom() < getTo();
    assert getFrom() >= 0;
    assert getTo() > 0;
    assert 1 == inObj.length;
    assert 3 == inObj[0].getData().getDimensions().length;
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().eval(inObj);
    final TensorList inputData = inObj[0].getData();
    @Nonnull final int[] inputDimensions = inputData.getDimensions();
    final int length = inputData.length();
    @Nonnull final int[] outputDimensions = Arrays.copyOf(inputDimensions, 3);
    outputDimensions[2] = getTo() - getFrom();
    long size = (length * outputDimensions[2] * outputDimensions[1] * outputDimensions[0] * precision.size);
    return new Result(CudaSystem.run(gpu -> {
        @Nullable final CudaTensor cudaInput = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        inputData.freeRef();
        final int byteOffset = cudaInput.descriptor.cStride * getFrom() * precision.size;
        @Nonnull final CudaDevice.CudaTensorDescriptor inputDescriptor = gpu.newTensorDescriptor(// 
        precision, // 
        length, // 
        outputDimensions[2], // 
        outputDimensions[1], // 
        outputDimensions[0], // 
        cudaInput.descriptor.nStride, // 
        cudaInput.descriptor.cStride, // 
        cudaInput.descriptor.hStride, cudaInput.descriptor.wStride);
        CudaMemory cudaInputMemory = cudaInput.getMemory(gpu);
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        CudaTensor cudaTensor = CudaTensor.wrap(cudaInputMemory.withByteOffset(byteOffset), inputDescriptor, precision);
        Stream.<ReferenceCounting>of(cudaInput, cudaInputMemory).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(cudaTensor, length, outputDimensions, precision);
    }, inputData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        if (!Arrays.equals(delta.getDimensions(), outputDimensions)) {
            throw new AssertionError(Arrays.toString(delta.getDimensions()) + " != " + Arrays.toString(outputDimensions));
        }
        if (inObj[0].isAlive()) {
            final TensorList passbackTensorList = CudaSystem.run(gpu -> {
                @Nonnull final CudaDevice.CudaTensorDescriptor viewDescriptor = gpu.newTensorDescriptor(// 
                precision, // 
                length, // 
                outputDimensions[2], // 
                outputDimensions[1], // 
                outputDimensions[0], // 
                inputDimensions[2] * inputDimensions[1] * inputDimensions[0], // 
                inputDimensions[1] * inputDimensions[0], // 
                inputDimensions[0], 1);
                @Nonnull final CudaDevice.CudaTensorDescriptor inputDescriptor = gpu.newTensorDescriptor(// 
                precision, // 
                length, // 
                inputDimensions[2], // 
                inputDimensions[1], // 
                inputDimensions[0], // 
                inputDimensions[2] * inputDimensions[1] * inputDimensions[0], // 
                inputDimensions[1] * inputDimensions[0], // 
                inputDimensions[0], 1);
                final int byteOffset = viewDescriptor.cStride * getFrom() * precision.size;
                assert delta.length() == length;
                // assert error.stream().flatMapToDouble(x-> Arrays.stream(x.getData())).allMatch(Double::isFinite);
                @Nullable final CudaTensor errorPtr = gpu.getTensor(delta, precision, MemoryType.Device, false);
                delta.freeRef();
                long size1 = (length * inputDimensions[2] * inputDimensions[1] * inputDimensions[0] * precision.size);
                @Nonnull final CudaMemory passbackBuffer = gpu.allocate(size1, MemoryType.Managed.normalize(), false);
                CudaMemory errorPtrMemory = errorPtr.getMemory(gpu);
                gpu.cudnnTransformTensor(precision.getPointer(1.0), errorPtr.descriptor.getPtr(), errorPtrMemory.getPtr(), precision.getPointer(0.0), viewDescriptor.getPtr(), passbackBuffer.getPtr().withByteOffset(byteOffset));
                errorPtrMemory.dirty();
                passbackBuffer.dirty();
                errorPtrMemory.freeRef();
                CudaTensor cudaTensor = CudaTensor.wrap(passbackBuffer, inputDescriptor, precision);
                Stream.<ReferenceCounting>of(errorPtr, viewDescriptor).forEach(ReferenceCounting::freeRef);
                return CudaTensorList.wrap(cudaTensor, length, inputDimensions, precision);
            // assert passbackTensorList.stream().flatMapToDouble(x-> Arrays.stream(x.getData())).allMatch(v->Double.isFinite(v));
            }, delta);
            inObj[0].accumulate(buffer, passbackTensorList);
        } else {
            delta.freeRef();
        }
    }) {

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

        @Override
        protected void _free() {
            Arrays.stream(inObj).forEach(nnResult -> nnResult.freeRef());
        }

        @Override
        public boolean isAlive() {
            return Arrays.stream(inObj).anyMatch(x -> x.isAlive());
        }
    };
}