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

Example 16 with CudaMemory

use of com.simiacryptus.mindseye.lang.cudnn.CudaMemory 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 17 with CudaMemory

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

the class ImgTileAssemblyLayer method copy.

/**
 * Copy.
 *
 * @param gpu                   the gpu
 * @param length                the length
 * @param sourceDimensions      the length in
 * @param source                the input buffer
 * @param destinationDimensions the length out
 * @param destination           the output buffer
 * @param positionX             the position x
 * @param positionY             the position y
 * @return the int [ ]
 */
public int[] copy(@Nonnull CudnnHandle gpu, int length, @Nonnull int[] sourceDimensions, @Nonnull CudaTensor source, @Nonnull int[] destinationDimensions, @Nonnull CudaMemory destination, int positionX, int positionY) {
    if (3 != sourceDimensions.length)
        throw new IllegalArgumentException("inputDimensions.length");
    if (3 != destinationDimensions.length)
        throw new IllegalArgumentException("dimOut.length");
    int bands = sourceDimensions[2];
    if (bands != destinationDimensions[2])
        throw new IllegalArgumentException(String.format("%d != %d", bands, destinationDimensions[2]));
    // log.info(String.format("offset=%d,%d", offsetX, offsetY));
    @Nonnull final int[] viewDim = getViewDimensions(sourceDimensions, destinationDimensions, new int[] { positionX, positionY, 0 });
    @Nonnull final CudaDevice.CudaTensorDescriptor sourceViewDescriptor = gpu.newTensorDescriptor(// 
    precision, // 
    length, // 
    viewDim[2], // 
    viewDim[1], // 
    viewDim[0], // 
    source.descriptor.nStride, // 
    source.descriptor.cStride, // 
    source.descriptor.hStride, source.descriptor.wStride);
    @Nonnull final CudaDevice.CudaTensorDescriptor destinationViewDescriptor = gpu.newTensorDescriptor(// 
    precision, // 
    length, // 
    viewDim[2], // 
    viewDim[1], // 
    viewDim[0], // 
    destinationDimensions[2] * destinationDimensions[1] * destinationDimensions[0], // 
    destinationDimensions[1] * destinationDimensions[0], // 
    destinationDimensions[0], 1);
    int sourceOffset = 0;
    int destinationOffset = 0;
    if (positionX > 0) {
        destinationOffset += Math.abs(positionX);
    } else {
        sourceOffset += source.descriptor.wStride * Math.abs(positionX);
    }
    if (positionY > 0) {
        destinationOffset += destinationDimensions[0] * Math.abs((positionY));
    } else {
        sourceOffset += source.descriptor.hStride * (Math.abs(positionY));
    }
    assert sourceOffset >= 0;
    assert destinationOffset >= 0;
    assert sourceOffset + Tensor.length(viewDim) <= (source.descriptor.nStride * length);
    assert destinationOffset + Tensor.length(viewDim) <= Tensor.length(destinationDimensions);
    CudaMemory sourceMemory = source.getMemory(gpu);
    CudaSystem.handle(gpu.cudnnTransformTensor(precision.getPointer(1.0), sourceViewDescriptor.getPtr(), sourceMemory.getPtr().withByteOffset(sourceOffset * precision.size), precision.getPointer(1.0), destinationViewDescriptor.getPtr(), destination.getPtr().withByteOffset(destinationOffset * precision.size)));
    assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
    sourceMemory.dirty();
    destination.dirty();
    sourceMemory.freeRef();
    Arrays.stream(new ReferenceCounting[] { sourceViewDescriptor, destinationViewDescriptor }).forEach(ReferenceCounting::freeRef);
    return viewDim;
}

Example 18 with CudaMemory

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

the class ImgTileAssemblyLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    if (1 == inObj.length) {
        return inObj[0];
    }
    int[] inputDimensions = inObj[0].getData().getDimensions();
    assert 3 == inputDimensions.length;
    final int length = inObj[0].getData().length();
    int[] outputDims = getOutputDims(inObj);
    final TensorList outputData = CudaSystem.run(gpu -> {
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        assert outputDims[0] > 0;
        assert outputDims[1] > 0;
        assert outputDims[2] > 0;
        @Nonnull final CudaMemory outputBuffer = gpu.allocate((long) length * outputDims[2] * outputDims[1] * outputDims[0] * precision.size, MemoryType.Managed.normalize(), false);
        int totalWidth = 0;
        int totalHeight = 0;
        int inputIndex = 0;
        List<CopyParams> copies = new ArrayList<>();
        for (int row = 0; row < rows; row++) {
            int positionX = 0;
            int rowHeight = 0;
            for (int col = 0; col < columns; col++) {
                int[] tileDimensions = inObj[inputIndex].getData().getDimensions();
                rowHeight = Math.max(rowHeight, tileDimensions[1]);
                copies.add(new CopyParams(gpu, inObj, outputBuffer, length, outputDims, tileDimensions, inputIndex, positionX, totalHeight));
                positionX += tileDimensions[0];
                inputIndex += 1;
                assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
            }
            totalHeight += rowHeight;
            totalWidth = Math.max(totalWidth, positionX);
        }
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        Stream<CopyParams> stream = copies.stream();
        if (!CoreSettings.INSTANCE.isSingleThreaded() && parallel)
            stream = stream.parallel();
        stream.forEach(this::copy);
        Arrays.stream(inObj).forEach(r -> r.getData().freeRef());
        CudaDevice.CudaTensorDescriptor descriptor = gpu.newTensorDescriptor(precision, length, outputDims[2], outputDims[1], outputDims[0]);
        CudaTensor ptr = CudaTensor.wrap(outputBuffer, descriptor, precision);
        return CudaTensorList.wrap(ptr, length, outputDims, precision);
    }, Arrays.stream(inObj).map(Result::getData).toArray());
    return new Result(outputData, (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList error) -> {
        if (!Arrays.equals(error.getDimensions(), outputData.getDimensions())) {
            throw new AssertionError(Arrays.toString(error.getDimensions()) + " != " + Arrays.toString(outputData.getDimensions()));
        }
        if (error.length() != outputData.length()) {
            throw new AssertionError(error.length() + " != " + outputData.length());
        }
        assert error.length() == length;
        int totalHeight = 0;
        int inputIndex = 0;
        List<BackpropParams> tasks = new ArrayList<>();
        for (int row = 0; row < rows; row++) {
            int positionX = 0;
            int rowHeight = 0;
            for (int col = 0; col < columns; col++) {
                Result in = inObj[inputIndex];
                int[] tileDimensions = in.getData().getDimensions();
                rowHeight = Math.max(rowHeight, tileDimensions[1]);
                if (inObj[inputIndex].isAlive()) {
                    tasks.add(new BackpropParams(inObj, buffer, error, outputDims, tileDimensions, length, positionX, totalHeight, inputIndex));
                }
                positionX += tileDimensions[0];
                inputIndex += 1;
            }
            totalHeight += rowHeight;
        }
        Stream<BackpropParams> stream = tasks.stream();
        if (!CoreSettings.INSTANCE.isSingleThreaded() && parallel)
            stream = stream.parallel();
        stream.forEach(this::backprop);
    }) {

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

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

Example 19 with CudaMemory

use of com.simiacryptus.mindseye.lang.cudnn.CudaMemory 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 20 with CudaMemory

use of com.simiacryptus.mindseye.lang.cudnn.CudaMemory 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;
        }
    };
}