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

Example 1 with CudaResource

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

the class ImgBandBiasLayer 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 leftData = input.getData();
    @Nonnull final int[] inputDimensions = leftData.getDimensions();
    final int length = leftData.length();
    if (3 != inputDimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(inputDimensions));
    }
    // assert !right.isAlive();
    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, inputDimensions[2], inputDimensions[1], inputDimensions[0], inputDimensions[2] * inputDimensions[1] * inputDimensions[0], inputDimensions[1] * inputDimensions[0], inputDimensions[0], 1);
        @Nullable final CudaTensor inputTensor = gpu.getTensor(leftData, precision, MemoryType.Device, false);
        CudaMemory biasMem = gpu.allocate(bias.length() * precision.size, MemoryType.Device, true).write(precision, bias.getData());
        int[] biasDim = bias.getDimensions();
        CudaDevice.CudaTensorDescriptor biasDescriptor = gpu.newTensorDescriptor(precision, 1, biasDim[2], biasDim[1], biasDim[0], biasDim[2] * biasDim[1] * biasDim[0], biasDim[1] * biasDim[0], biasDim[0], 1);
        // assert lPtr.size == rPtr.size;
        @Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * outputDescriptor.nStride * length, MemoryType.Managed.normalize(), true);
        CudaMemory inputMemory = inputTensor.getMemory(gpu);
        CudaSystem.handle(gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputMemory.getPtr(), precision.getPointer(1.0), biasDescriptor.getPtr(), biasMem.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        inputMemory.dirty();
        biasMem.dirty();
        outputPtr.dirty();
        inputMemory.freeRef();
        biasMem.freeRef();
        biasDescriptor.freeRef();
        inputTensor.freeRef();
        opDescriptor.freeRef();
        CudaTensor cudaTensor = CudaTensor.wrap(outputPtr, outputDescriptor, precision);
        return CudaTensorList.wrap(cudaTensor, length, inputDimensions, precision);
    }, leftData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        if (!isFrozen()) {
            @Nonnull double[] biasDelta = CudaSystem.run(gpu -> {
                @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
                CudaMemory biasMem = gpu.allocate(bias.length() * precision.size, MemoryType.Device, true).write(precision, bias.getData());
                int[] biasDim = bias.getDimensions();
                CudaDevice.CudaTensorDescriptor biasDescriptor = gpu.newTensorDescriptor(precision, 1, biasDim[2], biasDim[1], biasDim[0], biasDim[2] * biasDim[1] * biasDim[0], biasDim[1] * biasDim[0], biasDim[0], 1);
                CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                gpu.cudnnConvolutionBackwardBias(precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), precision.getPointer(0.0), biasDescriptor.getPtr(), biasMem.getPtr());
                assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
                biasMem.dirty();
                double[] biasV = new double[bias.length()];
                biasMem.read(precision, biasV);
                Stream.<ReferenceCounting>of(biasMem, deltaTensorMemory, deltaTensor, biasDescriptor).forEach(ReferenceCounting::freeRef);
                return biasV;
            }, delta);
            buffer.get(ImgBandBiasLayer.this, bias).addInPlace(biasDelta).freeRef();
        }
        if (input.isAlive()) {
            input.accumulate(buffer, delta);
        } else {
            delta.freeRef();
        }
    }) {

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

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

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

Example 2 with CudaResource

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

the class BandAvgReducerLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    final Result input = inObj[0];
    TensorList inputData = input.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    int length = inputData.length();
    final int bands = inputSize[2];
    CudaTensorList result = CudaSystem.run(gpu -> {
        CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, bands, 1, 1);
        long size = (long) precision.size * outputDescriptor.nStride * length;
        @Nonnull final CudaMemory outputPtr = 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);
        CudaMemory inputMemory = inputTensor.getMemory(gpu);
        @Nonnull final CudaMemory workspacePtr = gpu.allocate(inputMemory.size, MemoryType.Device, true);
        @Nonnull final CudaMemory indexPtr = gpu.allocate(12 * length, MemoryType.Device, false);
        gpu.cudnnReduceTensor(reduceTensorDescriptor.getPtr(), indexPtr.getPtr(), indexPtr.size, workspacePtr.getPtr(), workspacePtr.size, precision.getPointer(alpha), inputTensor.descriptor.getPtr(), inputMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr());
        outputPtr.dirty();
        inputMemory.dirty();
        Stream.of(inputMemory, inputTensor, reduceTensorDescriptor, workspacePtr, indexPtr, inputData).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(CudaTensor.wrap(outputPtr, outputDescriptor, precision), length, new int[] { 1, 1, bands }, precision);
    });
    int pixels = inputSize[0] * inputSize[1];
    return new Result(result, (DeltaSet<Layer> ctx, TensorList delta) -> {
        TensorList passback;
        passback = TensorArray.wrap(delta.stream().map(x -> {
            Tensor tensor = new Tensor(inputSize[0], inputSize[1], inputSize[2]).setByCoord(c -> x.get(c.getCoords()[2]) * alpha / pixels);
            x.freeRef();
            return tensor;
        }).toArray(i -> new Tensor[i]));
        // passback = CudaSystem.run(gpu -> {
        // CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
        // @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision,
        // length, inputSize[2], inputSize[1], inputSize[0]);
        // @Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * outputDescriptor.nStride * length, MemoryType.Device, true);
        // CudaMemory deltaMemory = deltaTensor.getMemory(gpu);
        // @Nonnull final CudaDevice.CudaTensorDescriptor inputDescriptor = gpu.newTensorDescriptor(precision,
        // 1, 1, inputSize[1], inputSize[0]);
        // for(int batch=0;batch<length;batch++){
        // Tensor tensor = delta.get(batch);
        // for(int band=0;band<bands;band++){
        // int i = batch * bands + band;
        // CudaMemory img = outputPtr.withByteOffset(precision.size * i * outputDescriptor.cStride);
        // CudaMemory val = deltaMemory.withByteOffset(precision.size * i);
        // gpu.cudnnSetTensor(inputDescriptor.getPtr(), img.getPtr(), precision.getPointer(tensor.get(band) / outputDescriptor.cStride));
        // img.freeRef();
        // val.freeRef();
        // outputPtr.dirty().synchronize();
        // }
        // }
        // Stream.of(deltaMemory, deltaTensor, inputDescriptor).forEach(ReferenceCounting::freeRef);
        // return CudaTensorList.wrap(CudaTensor.wrap(outputPtr, outputDescriptor, precision), length, inputSize, precision);
        // });
        input.accumulate(ctx, passback);
    }) {

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

Example 3 with CudaResource

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

the class BinarySumLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (inObj.length == 1) {
        if (rightFactor != 1)
            throw new IllegalStateException();
        if (leftFactor != 1)
            throw new IllegalStateException();
        return inObj[0];
    }
    if (inObj.length > 2) {
        if (rightFactor != 1)
            throw new IllegalStateException();
        if (leftFactor != 1)
            throw new IllegalStateException();
        return Arrays.stream(inObj).reduce((a, b) -> evalAndFree(a, b)).get();
    }
    assert (inObj.length == 2);
    final TensorList leftData = inObj[0].getData();
    final TensorList rightData = inObj[1].getData();
    int[] leftDimensions = leftData.getDimensions();
    if (3 < leftDimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(leftDimensions));
    }
    @Nonnull final int[] dimensions = { leftDimensions.length < 1 ? 0 : leftDimensions[0], leftDimensions.length < 2 ? 1 : leftDimensions[1], leftDimensions.length < 3 ? 1 : leftDimensions[2] };
    final int length = leftData.length();
    if (length != rightData.length())
        throw new IllegalArgumentException();
    if (3 != dimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(dimensions));
    }
    for (int i = 1; i < inObj.length; i++) {
        if (Tensor.length(dimensions) != Tensor.length(inObj[i].getData().getDimensions())) {
            throw new IllegalArgumentException(Arrays.toString(dimensions) + " != " + Arrays.toString(inObj[i].getData().getDimensions()));
        }
    }
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    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, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
        // .getDenseAndFree(gpu);//.moveTo(gpu.getDeviceNumber());
        @Nullable final CudaTensor lPtr = gpu.getTensor(leftData, precision, MemoryType.Device, false);
        // .getDenseAndFree(gpu);//.moveTo(gpu.getDeviceNumber());
        @Nullable final CudaTensor rPtr = gpu.getTensor(rightData, precision, MemoryType.Device, false);
        @Nonnull final CudaMemory outputPtr = gpu.allocate(precision.size * Tensor.length(dimensions) * length, MemoryType.Managed, true);
        CudaMemory lPtrMemory = lPtr.getMemory(gpu);
        CudaMemory rPtrMemory = rPtr.getMemory(gpu);
        gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(leftFactor), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(rightFactor), rPtr.descriptor.getPtr(), rPtrMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr());
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        lPtrMemory.dirty();
        rPtrMemory.dirty();
        outputPtr.dirty();
        rPtrMemory.freeRef();
        lPtrMemory.freeRef();
        CudaTensor cudaTensor = CudaTensor.wrap(outputPtr, outputDescriptor, precision);
        Stream.<ReferenceCounting>of(opDescriptor, lPtr, rPtr).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(cudaTensor, length, dimensions, precision);
    }, leftData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        Runnable a = () -> {
            if (inObj[0].isAlive()) {
                CudaTensorList tensorList = CudaSystem.run(gpu -> {
                    @Nullable final CudaTensor lPtr = gpu.getTensor(delta, precision, MemoryType.Device, false);
                    @Nonnull final CudaMemory passbackPtr = gpu.allocate(precision.size * Tensor.length(dimensions) * length, MemoryType.Managed.normalize(), true);
                    @Nonnull final CudaDevice.CudaTensorDescriptor passbackDescriptor = gpu.newTensorDescriptor(precision, length, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
                    CudaMemory lPtrMemory = lPtr.getMemory(gpu);
                    gpu.cudnnTransformTensor(precision.getPointer(leftFactor), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(0.0), passbackDescriptor.getPtr(), passbackPtr.getPtr());
                    assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
                    passbackPtr.dirty();
                    lPtrMemory.dirty();
                    lPtrMemory.freeRef();
                    CudaTensor cudaTensor = CudaTensor.wrap(passbackPtr, passbackDescriptor, precision);
                    lPtr.freeRef();
                    return CudaTensorList.wrap(cudaTensor, length, dimensions, precision);
                }, delta);
                inObj[0].accumulate(buffer, tensorList);
            }
        };
        Runnable b = () -> {
            if (inObj[1].isAlive()) {
                CudaTensorList tensorList = CudaSystem.run(gpu -> {
                    @Nullable final CudaTensor lPtr = gpu.getTensor(delta, precision, MemoryType.Device, false);
                    @Nonnull final CudaMemory outputPtr = gpu.allocate(precision.size * Tensor.length(dimensions) * length, MemoryType.Managed.normalize(), true);
                    @Nonnull final CudaDevice.CudaTensorDescriptor passbackDescriptor = gpu.newTensorDescriptor(precision, length, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
                    CudaMemory lPtrMemory = lPtr.getMemory(gpu);
                    gpu.cudnnTransformTensor(precision.getPointer(rightFactor), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(0.0), passbackDescriptor.getPtr(), outputPtr.getPtr());
                    outputPtr.dirty();
                    lPtrMemory.dirty();
                    lPtrMemory.freeRef();
                    CudaTensor cudaTensor = CudaTensor.wrap(outputPtr, passbackDescriptor, precision);
                    lPtr.freeRef();
                    return CudaTensorList.wrap(cudaTensor, length, dimensions, precision);
                }, delta);
                inObj[1].accumulate(buffer, tensorList);
            }
        };
        if (CoreSettings.INSTANCE.isSingleThreaded())
            TestUtil.runAllSerial(a, b);
        else
            TestUtil.runAllParallel(a, b);
    }) {

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

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

Example 4 with CudaResource

use of com.simiacryptus.mindseye.lang.cudnn.CudaResource 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 5 with CudaResource

use of com.simiacryptus.mindseye.lang.cudnn.CudaResource 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();
        }
    };
}