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

Example 1 with CudaTensor

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

the class GramianLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(final Result... inObj) {
    assert 1 == inObj.length;
    TensorList inputData = inObj[0].getData();
    int[] inputDimensions = inputData.getDimensions();
    assert 3 == inputDimensions.length;
    return new Result(CudaSystem.run(gpu -> {
        CudaTensor tensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        CudaTensorList output = getOutput(gpu, tensor);
        tensor.freeRef();
        return output;
    }, inputData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        @Nonnull final int[] outputDimensions = { 1, 1, inputDimensions[2] * inputDimensions[2] };
        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 -> {
                @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
                CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
                delta.freeRef();
                CudaTensorList feedback = getFeedback(gpu, inputTensor, deltaTensor);
                deltaTensor.freeRef();
                inputTensor.freeRef();
                return feedback;
            }, delta);
            inObj[0].accumulate(buffer, passbackTensorList);
        } else {
            delta.freeRef();
        }
    }) {

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

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

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

Example 2 with CudaTensor

use of com.simiacryptus.mindseye.lang.cudnn.CudaTensor 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 3 with CudaTensor

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

the class ImgLinearSubnetLayer 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 length = inputData.length();
    int maxBand = legs.stream().mapToInt(x -> x.toBand).max().getAsInt();
    assert maxBand == inputDims[2] : maxBand + " != " + inputDims[2];
    assert IntStream.range(0, maxBand).allMatch(i -> 1 == legs.stream().filter(x -> x.fromBand <= i && x.toBand > i).count());
    CudaTensor passback = CudaSystem.run(gpu -> {
        return CudaTensor.wrap(gpu.allocate(inputData.getElements() * precision.size, MemoryType.Device, true), gpu.newTensorDescriptor(precision, length, inputDims[2], inputDims[1], inputDims[0]), precision);
    });
    try {
        AtomicInteger counter = new AtomicInteger(0);
        SumInputsLayer sumInputsLayer = new SumInputsLayer();
        try {
            Result[] legResults = legs.stream().map(leg -> {
                passback.addRef();
                ImgBandSelectLayer imgBandSelectLayer = new ImgBandSelectLayer(leg.fromBand, leg.toBand);
                input.addRef();
                TensorList legData = imgBandSelectLayer.eval(input).getDataAndFree();
                imgBandSelectLayer.freeRef();
                return leg.inner.evalAndFree(new Result(legData, (DeltaSet<Layer> ctx, TensorList delta) -> {
                    int[] outputDimensions = delta.getDimensions();
                    int[] inputDimensions = inputDims;
                    synchronized (passback) {
                        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);
                            final int byteOffset = viewDescriptor.cStride * leg.fromBand * precision.size;
                            assert delta.length() == inputData.length();
                            assert passback.getDeviceId() == gpu.getDeviceId();
                            // assert error.stream().flatMapToDouble(x-> Arrays.stream(x.getData())).allMatch(Double::isFinite);
                            @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
                            @Nonnull final CudaMemory passbackBuffer = passback.getMemory(gpu);
                            CudaMemory errorPtrMemory = deltaTensor.getMemory(gpu);
                            passbackBuffer.synchronize();
                            gpu.cudnnTransformTensor(precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), errorPtrMemory.getPtr(), precision.getPointer(0.0), viewDescriptor.getPtr(), passbackBuffer.getPtr().withByteOffset(byteOffset));
                            errorPtrMemory.dirty();
                            passbackBuffer.dirty();
                            Stream.<ReferenceCounting>of(deltaTensor, viewDescriptor, passbackBuffer, errorPtrMemory).forEach(ReferenceCounting::freeRef);
                        }, passback);
                    }
                    if (counter.incrementAndGet() >= legs.size()) {
                        counter.set(0);
                        input.accumulate(ctx, CudaTensorList.create(passback, length, inputDims, precision));
                    }
                }) {

                    @Override
                    protected void _free() {
                        super._free();
                        input.freeRef();
                        passback.freeRef();
                    }
                });
            }).toArray(i -> new Result[i]);
            return sumInputsLayer.setParallel(parallel).setPrecision(precision).evalAndFree(legResults);
        } finally {
            sumInputsLayer.freeRef();
            input.freeRef();
        }
    } finally {
        passback.freeRef();
    }
}

Example 4 with CudaTensor

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

the class ImgTileAssemblyLayer method backprop.

/**
 * Backprop.
 *
 * @param backpropParams the backprop params
 */
public void backprop(final BackpropParams backpropParams) {
    final TensorList passbackTensorList = CudaSystem.run(gpu -> {
        CudaTensor ptr = copy(gpu, backpropParams.error, backpropParams.tileDimensions, backpropParams.outputDims, backpropParams.length, -backpropParams.positionX, -backpropParams.totalHeight);
        return CudaTensorList.wrap(ptr, backpropParams.length, backpropParams.tileDimensions, precision);
    }, backpropParams.error);
    backpropParams.inObj[backpropParams.inputIndex].accumulate(backpropParams.buffer, passbackTensorList);
}

Example 5 with CudaTensor

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

the class ImgTileAssemblyLayer method copy.

/**
 * Copy cuda tensor.
 *
 * @param gpu            the gpu
 * @param error          the error
 * @param tileDimensions the tile dimensions
 * @param outputDims     the output dims
 * @param length         the length
 * @param positionX      the position x
 * @param positionY      the position y
 * @return the cuda tensor
 */
public CudaTensor copy(final CudnnHandle gpu, final TensorList error, final int[] tileDimensions, final int[] outputDims, final int length, final int positionX, final int positionY) {
    @Nullable final CudaTensor errorPtr = gpu.getTensor(error, precision, MemoryType.Device, false);
    @Nonnull final CudaMemory passbackBuffer = gpu.allocate((long) length * tileDimensions[2] * tileDimensions[1] * tileDimensions[0] * precision.size, MemoryType.Managed.normalize(), false);
    copy(gpu, length, outputDims, errorPtr, tileDimensions, passbackBuffer, positionX, positionY);
    errorPtr.freeRef();
    CudaDevice.CudaTensorDescriptor descriptor = gpu.newTensorDescriptor(precision, length, tileDimensions[2], tileDimensions[1], tileDimensions[0]);
    return CudaTensor.wrap(passbackBuffer, descriptor, precision);
}