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

Example 11 with Precision

use of com.simiacryptus.mindseye.lang.cudnn.Precision 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 12 with Precision

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

Example 13 with Precision

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

the class ImgConcatLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    // assert Arrays.stream(this.bias).allMatch(Double::isFinite);
    // assert Arrays.stream(inObj).flatMapToDouble(input->input.data.stream().flatMapToDouble(x-> Arrays.stream(x.getData()))).allMatch(v->Double.isFinite(v));
    int[] dimensions = inObj[0].getData().getDimensions();
    assert 3 == dimensions.length;
    @Nonnull final int[] outputDimensions = Arrays.copyOf(dimensions, dimensions.length);
    final int length = inObj[0].getData().length();
    assert Arrays.stream(inObj).allMatch(x -> {
        @Nonnull int[] d = x.getData().getDimensions();
        return 3 == d.length && d[0] == outputDimensions[0] && d[1] == outputDimensions[1] && x.getData().length() == length;
    });
    outputDimensions[2] = Arrays.stream(inObj).mapToInt(x -> x.getData().getDimensions()[2]).sum();
    if (0 < maxBands && outputDimensions[2] > maxBands) {
        outputDimensions[2] = maxBands;
    }
    return new Result(CudaSystem.run(gpu -> {
        final long outputSize = ((long) length * outputDimensions[2] * outputDimensions[1] * outputDimensions[0] * precision.size);
        @Nonnull final CudaMemory cudaOutput = gpu.allocate(outputSize, MemoryType.Managed.normalize(), true);
        IntStream stream = IntStream.range(0, inObj.length);
        // if (!CoreSettings.INSTANCE.isConservative() && parallel) stream = stream.parallel();
        stream.forEach(i -> {
            assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
            final TensorList input = inObj[i].getData();
            @Nonnull final int[] inputDimensions = input.getDimensions();
            assert inputDimensions[0] == outputDimensions[0];
            assert inputDimensions[1] == outputDimensions[1];
            int bandOffset = IntStream.range(0, i).map(j -> inObj[j].getData().getDimensions()[2]).sum();
            if (maxBands > 0)
                bandOffset = Math.min(bandOffset, maxBands);
            int inputBands = inputDimensions[2];
            if (maxBands > 0)
                inputBands = Math.min(inputBands, maxBands - bandOffset);
            if (inputBands > 0) {
                @Nullable final CudaTensor cudaInput = gpu.getTensor(input, precision, MemoryType.Device, false);
                assert inputBands > 0;
                assert maxBands <= 0 || inputBands <= maxBands;
                assert inputBands <= inputDimensions[2];
                @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(// 
                precision, // 
                length, // 
                inputBands, // 
                outputDimensions[1], // 
                outputDimensions[0], // 
                outputDimensions[2] * outputDimensions[1] * outputDimensions[0], // 
                outputDimensions[1] * outputDimensions[0], // 
                outputDimensions[0], 1);
                @Nonnull final CudaDevice.CudaTensorDescriptor inputDescriptor = gpu.newTensorDescriptor(// 
                precision, // 
                length, // 
                inputBands, // 
                inputDimensions[1], // 
                inputDimensions[0], // 
                cudaInput.descriptor.nStride, // 
                cudaInput.descriptor.cStride, // 
                cudaInput.descriptor.hStride, cudaInput.descriptor.wStride);
                int byteOffset = outputDescriptor.cStride * bandOffset * precision.size;
                CudaMemory cudaInputMemory = cudaInput.getMemory(gpu);
                gpu.cudnnTransformTensor(precision.getPointer(1.0), inputDescriptor.getPtr(), cudaInputMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), cudaOutput.getPtr().withByteOffset(byteOffset));
                assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
                cudaInputMemory.dirty();
                cudaOutput.dirty();
                cudaInputMemory.freeRef();
                Stream.<ReferenceCounting>of(cudaInput, outputDescriptor, inputDescriptor).forEach(ReferenceCounting::freeRef);
            }
        });
        CudaDevice.CudaTensorDescriptor outDesc = gpu.newTensorDescriptor(precision, length, outputDimensions[2], outputDimensions[1], outputDimensions[0]);
        return CudaTensorList.wrap(CudaTensor.wrap(cudaOutput, outDesc, precision), length, outputDimensions, precision);
    }, Arrays.stream(inObj).map(Result::getData).toArray()), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        assert delta.getDimensions()[0] == outputDimensions[0];
        assert delta.getDimensions()[1] == outputDimensions[1];
        assert delta.getDimensions()[2] == outputDimensions[2];
        if (!Arrays.equals(delta.getDimensions(), outputDimensions)) {
            throw new AssertionError(Arrays.toString(delta.getDimensions()) + " != " + Arrays.toString(outputDimensions));
        }
        // outputBuffer.freeRef();
        // assert error.stream().flatMapToDouble(x-> Arrays.stream(x.getData())).allMatch(Double::isFinite);
        @Nonnull IntStream stream = IntStream.range(0, inObj.length);
        if (!CoreSettings.INSTANCE.isSingleThreaded() && parallel)
            stream = stream.parallel();
        stream.forEach(i -> {
            final Result input = inObj[i];
            int[] inputDimentions = input.getData().getDimensions();
            assert 3 == inputDimentions.length;
            assert delta.length() == input.getData().length();
            assert inputDimentions[0] == outputDimensions[0];
            assert inputDimentions[1] == outputDimensions[1];
            int bandOffset = IntStream.range(0, i).map(j -> inObj[j].getData().getDimensions()[2]).sum();
            int inputBands = maxBands <= 0 ? inputDimentions[2] : Math.min(inputDimentions[2], maxBands - bandOffset);
            if (inputBands > 0 && input.isAlive()) {
                assert inputBands <= inputDimentions[2];
                assert inputBands <= outputDimensions[2];
                final TensorList passbackTensorList = CudaSystem.run(gpu -> {
                    final CudaTensor result;
                    synchronized (gpu) {
                        result = gpu.getTensor(delta, precision, MemoryType.Device, true);
                    }
                    @Nullable final CudaTensor cudaDelta = result;
                    CudaMemory cudaDeltaMemory = cudaDelta.getMemory(gpu);
                    try {
                        if (inputDimentions[2] == inputBands) {
                            @Nonnull final CudaDevice.CudaTensorDescriptor viewDescriptor = gpu.newTensorDescriptor(// 
                            precision, // 
                            length, // 
                            inputDimentions[2], // 
                            inputDimentions[1], // 
                            inputDimentions[0], // 
                            cudaDelta.descriptor.nStride, // 
                            cudaDelta.descriptor.cStride, // 
                            cudaDelta.descriptor.hStride, cudaDelta.descriptor.wStride);
                            int byteOffset = cudaDelta.descriptor.cStride * bandOffset * precision.size;
                            CudaMemory ptr = cudaDeltaMemory.withByteOffset(byteOffset);
                            CudaTensor cudaTensor = CudaTensor.wrap(ptr, viewDescriptor, precision);
                            Stream.<ReferenceCounting>of(cudaDelta).forEach(ReferenceCounting::freeRef);
                            return CudaTensorList.wrap(cudaTensor, length, inputDimentions, precision);
                        } else {
                            @Nonnull final CudaDevice.CudaTensorDescriptor passbackTransferDescriptor = gpu.newTensorDescriptor(// 
                            precision, // 
                            length, // 
                            inputBands, // 
                            inputDimentions[1], // 
                            inputDimentions[0], // 
                            inputDimentions[2] * inputDimentions[1] * inputDimentions[0], // 
                            inputDimentions[1] * inputDimentions[0], // 
                            inputDimentions[0], 1);
                            @Nonnull final CudaDevice.CudaTensorDescriptor passbackDescriptor = gpu.newTensorDescriptor(// 
                            precision, // 
                            length, // 
                            inputDimentions[2], // 
                            inputDimentions[1], // 
                            inputDimentions[0], // 
                            inputDimentions[2] * inputDimentions[1] * inputDimentions[0], // 
                            inputDimentions[1] * inputDimentions[0], // 
                            inputDimentions[0], 1);
                            @Nonnull final CudaDevice.CudaTensorDescriptor deltaViewDescriptor = gpu.newTensorDescriptor(// 
                            precision, // 
                            length, // 
                            inputBands, // 
                            inputDimentions[1], // 
                            inputDimentions[0], // 
                            cudaDelta.descriptor.nStride, // 
                            cudaDelta.descriptor.cStride, // 
                            cudaDelta.descriptor.hStride, cudaDelta.descriptor.wStride);
                            @Nonnull final CudaMemory cudaBackprop = gpu.allocate((long) passbackDescriptor.nStride * length * precision.size, MemoryType.Managed.normalize(), inputBands == inputDimentions[2]);
                            int byteOffset = cudaDelta.descriptor.cStride * bandOffset * precision.size;
                            gpu.cudnnTransformTensor(precision.getPointer(1.0), deltaViewDescriptor.getPtr(), cudaDeltaMemory.getPtr().withByteOffset(byteOffset), precision.getPointer(0.0), passbackTransferDescriptor.getPtr(), cudaBackprop.getPtr());
                            cudaBackprop.dirty();
                            cudaDeltaMemory.dirty();
                            Stream.<ReferenceCounting>of(cudaDelta, deltaViewDescriptor, passbackTransferDescriptor).forEach(ReferenceCounting::freeRef);
                            return CudaTensorList.wrap(CudaTensor.wrap(cudaBackprop, passbackDescriptor, precision), length, inputDimentions, precision);
                        }
                    } finally {
                        cudaDeltaMemory.freeRef();
                    }
                });
                input.accumulate(buffer, passbackTensorList);
            }
        // assert passbackTensorList.stream().flatMapToDouble(x-> Arrays.stream(x.getData())).allMatch(v->Double.isFinite(v));
        });
    }) {

        @Override
        protected void _free() {
            for (@Nonnull Result result : inObj) {
                result.freeRef();
                result.getData().freeRef();
            }
        }

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

Example 14 with Precision

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

the class ImgLinearSubnetLayer method getJson.

@Nonnull
@Override
public JsonObject getJson(Map<CharSequence, byte[]> resources, DataSerializer dataSerializer) {
    @Nonnull final JsonObject json = super.getJsonStub();
    json.addProperty("precision", precision.name());
    json.addProperty("parallel", isParallel());
    JsonArray jsonArray = new JsonArray();
    legs.stream().map(x -> x.getJson(resources, dataSerializer)).forEach(jsonArray::add);
    json.add("legs", jsonArray);
    return json;
}

Example 15 with Precision

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