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

Example 21 with CudaMemory

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

the class SimpleConvolutionLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().eval(inObj);
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    @Nonnull final int[] kernelSize = kernel.getDimensions();
    final int[] outputSize = getOutputSize(inputSize);
    final int length = inputData.length();
    kernel.addRef();
    SimpleConvolutionLayer.this.addRef();
    return new Result(CudaSystem.run(gpu -> {
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        final CudaResource<cudnnFilterDescriptor> filterDescriptor = gpu.newFilterDescriptor(precision, cudnnTensorFormat.CUDNN_TENSOR_NCHW, outputSize[2], inputSize[2], kernelSize[1], kernelSize[0]);
        final CudaResource<cudnnConvolutionDescriptor> convolutionDescriptor = gpu.newConvolutions2dDescriptor(cudnnConvolutionMode.CUDNN_CONVOLUTION, precision, paddingY, paddingX, strideY, strideX, 1, 1);
        final int[] outputDims = IntStream.of(reverse(CudaSystem.getOutputDims(inputTensor.descriptor.getPtr(), filterDescriptor.getPtr(), convolutionDescriptor.getPtr()))).limit(3).toArray();
        final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, outputDims[2], outputDims[1], outputDims[0], outputDims[2] * outputDims[1] * outputDims[0], outputDims[1] * outputDims[0], outputDims[0], 1);
        final int forwardAlgorithm = getForwardAlgorithm(gpu, inputTensor, filterDescriptor, convolutionDescriptor, outputDescriptor);
        final CudaMemory forwardWorkspace = gpu.allocateForwardWorkspace(inputTensor.descriptor.getPtr(), filterDescriptor.getPtr(), convolutionDescriptor.getPtr(), outputDescriptor.getPtr(), forwardAlgorithm, 1);
        try {
            assert 0 < kernel.getData().length;
            assert kernelSize[0] * kernelSize[1] * kernelSize[2] == kernel.getData().length;
            @Nonnull CudaMemory filterPtr = getCudaFilter(gpu);
            @Nonnull final CudaMemory outputBuffer = gpu.allocate((long) Tensor.length(outputDims) * length * precision.size, MemoryType.Managed.normalize(), true);
            CudaMemory inputTensorMemory = inputTensor.getMemory(gpu);
            // inputTensorMemory.synchronize();
            CudaSystem.handle(gpu.cudnnConvolutionForward(precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputTensorMemory.getPtr(), filterDescriptor.getPtr(), filterPtr.getPtr(), convolutionDescriptor.getPtr(), forwardAlgorithm, null == forwardWorkspace ? null : forwardWorkspace.getPtr(), null == forwardWorkspace ? 0 : forwardWorkspace.size, precision.getPointer(0.0), outputDescriptor.getPtr(), outputBuffer.getPtr()));
            assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
            forwardWorkspace.dirty();
            filterPtr.dirty();
            outputBuffer.dirty();
            inputTensorMemory.dirty();
            // inputTensorMemory.synchronize();
            inputTensorMemory.freeRef();
            filterPtr.freeRef();
            outputDescriptor.addRef();
            return CudaTensorList.wrap(CudaTensor.wrap(outputBuffer, outputDescriptor, precision), length, outputDims, precision);
        } catch (@Nonnull final Throwable e) {
            throw new ComponentException(String.format("Error in convolution %s x %s", Arrays.toString(inputSize), Arrays.toString(kernelSize)), e);
        } finally {
            Stream.of(inputTensor, filterDescriptor, outputDescriptor, forwardWorkspace, convolutionDescriptor).forEach(ReferenceCounting::freeRef);
        }
    }, inputData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        delta.assertAlive();
        buffer.assertAlive();
        inputData.assertAlive();
        assert delta.length() == length;
        delta.addRef();
        Runnable learnFn = () -> {
            if (!isFrozen()) {
                @Nonnull final Tensor weightGradient = CudaSystem.run(gpu -> {
                    @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
                    delta.freeRef();
                    @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
                    final CudaResource<cudnnFilterDescriptor> filterDescriptor = gpu.newFilterDescriptor(precision, cudnnTensorFormat.CUDNN_TENSOR_NCHW, outputSize[2], inputSize[2], kernelSize[1], kernelSize[0]);
                    final CudaResource<cudnnConvolutionDescriptor> convolutionDescriptor = gpu.newConvolutions2dDescriptor(cudnnConvolutionMode.CUDNN_CONVOLUTION, precision, paddingY, paddingX, strideY, strideX, 1, 1);
                    final int backwardFilterAlgorithm = getBackwardFilterAlgorithm(gpu, deltaTensor, inputTensor, filterDescriptor, convolutionDescriptor);
                    final CudaMemory backwardsFilterWorkSpace = gpu.allocateBackwardFilterWorkspace(inputTensor.descriptor.getPtr(), filterDescriptor.getPtr(), convolutionDescriptor.getPtr(), deltaTensor.descriptor.getPtr(), backwardFilterAlgorithm, 1);
                    @Nonnull CudaMemory filterPtr = gpu.allocate((long) kernel.length() * precision.size, MemoryType.Device, true);
                    try {
                        CudaMemory inputTensorMemory = inputTensor.getMemory(gpu);
                        CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu, MemoryType.Managed.normalize());
                        // inputTensorMemory.synchronize();
                        CudaSystem.handle(gpu.cudnnConvolutionBackwardFilter(precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputTensorMemory.getPtr(), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), convolutionDescriptor.getPtr(), backwardFilterAlgorithm, backwardsFilterWorkSpace.getPtr(), backwardsFilterWorkSpace.size, precision.getPointer(0.0), filterDescriptor.getPtr(), filterPtr.getPtr()));
                        filterPtr.dirty();
                        deltaTensorMemory.dirty();
                        inputTensorMemory.dirty();
                        backwardsFilterWorkSpace.dirty();
                        // backwardsFilterWorkSpace.synchronize();
                        inputTensorMemory.freeRef();
                        deltaTensorMemory.freeRef();
                        return filterPtr.read(precision, kernel.getDimensions());
                    } catch (@Nonnull final Throwable e) {
                        throw new ComponentException(String.format("Error in convolution %s x %s => %s", Arrays.toString(inputSize), Arrays.toString(kernelSize), Arrays.toString(outputSize)), e);
                    } finally {
                        inputTensor.freeRef();
                        filterPtr.freeRef();
                        deltaTensor.freeRef();
                        Stream.of(filterDescriptor, convolutionDescriptor, backwardsFilterWorkSpace).forEach(ReferenceCounting::freeRef);
                    }
                }, delta);
                buffer.get(SimpleConvolutionLayer.this, kernel.getData()).addInPlace(weightGradient.getData()).freeRef();
                weightGradient.freeRef();
                clearCudaFilters();
            } else {
                delta.freeRef();
            }
        };
        Runnable backpropFn = () -> {
            if (input.isAlive()) {
                final TensorList inputBufferTensors = CudaSystem.run(gpu -> {
                    final CudaDevice.CudaTensorDescriptor inputDescriptor = gpu.newTensorDescriptor(precision, length, inputSize[2], inputSize[1], inputSize[0], inputSize[2] * inputSize[1] * inputSize[0], inputSize[1] * inputSize[0], inputSize[0], 1);
                    final CudaResource<cudnnFilterDescriptor> filterDescriptor = gpu.newFilterDescriptor(precision, cudnnTensorFormat.CUDNN_TENSOR_NCHW, outputSize[2], inputSize[2], kernelSize[1], kernelSize[0]);
                    final CudaResource<cudnnConvolutionDescriptor> convolutionDescriptor = gpu.newConvolutions2dDescriptor(cudnnConvolutionMode.CUDNN_CONVOLUTION, precision, paddingY, paddingX, strideY, strideX, 1, 1);
                    @Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
                    delta.freeRef();
                    final int backwardDataAlgorithm = getBackwardDataAlgorithm(gpu, inputDescriptor, filterDescriptor, convolutionDescriptor, deltaTensor);
                    final CudaMemory backwardsDataWorkSpace = gpu.allocateBackwardDataWorkspace(inputDescriptor.getPtr(), filterDescriptor.getPtr(), convolutionDescriptor.getPtr(), deltaTensor.descriptor.getPtr(), backwardDataAlgorithm, 1);
                    @Nonnull final CudaMemory filterPtr = getCudaFilter(gpu);
                    try {
                        @Nonnull final CudaMemory passbackMemory = gpu.allocate((long) Tensor.length(inputData.getDimensions()) * length * precision.size, MemoryType.Managed.normalize(), true);
                        CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                        // deltaTensorMemory.synchronize();
                        CudaSystem.handle(gpu.cudnnConvolutionBackwardData(precision.getPointer(1.0), filterDescriptor.getPtr(), filterPtr.getPtr(), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), convolutionDescriptor.getPtr(), backwardDataAlgorithm, backwardsDataWorkSpace.getPtr(), backwardsDataWorkSpace.size, precision.getPointer(0.0), inputDescriptor.getPtr(), passbackMemory.getPtr()));
                        passbackMemory.dirty();
                        backwardsDataWorkSpace.dirty();
                        deltaTensorMemory.dirty();
                        // deltaTensorMemory.synchronize();
                        filterPtr.dirty();
                        deltaTensorMemory.freeRef();
                        inputDescriptor.addRef();
                        return CudaTensorList.wrap(CudaTensor.wrap(passbackMemory, inputDescriptor, precision), length, inputSize, precision);
                    } catch (@Nonnull final Throwable e) {
                        throw new ComponentException(String.format("Error in convolution %s x %s => %s", Arrays.toString(inputSize), Arrays.toString(kernelSize), Arrays.toString(outputSize)), e);
                    } finally {
                        filterPtr.freeRef();
                        deltaTensor.freeRef();
                        Stream.of(inputDescriptor, filterDescriptor, convolutionDescriptor, backwardsDataWorkSpace).forEach(ReferenceCounting::freeRef);
                    }
                }, delta);
                if (null != inputBufferTensors) {
                    input.accumulate(buffer, inputBufferTensors);
                }
            } else {
                delta.freeRef();
            }
        };
        Stream.of(learnFn, backpropFn).forEach(Runnable::run);
    }) {

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

        @Override
        protected void _free() {
            kernel.freeRef();
            inputData.freeRef();
            Arrays.stream(inObj).forEach(ReferenceCounting::freeRef);
            SimpleConvolutionLayer.this.freeRef();
        }

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

Example 22 with CudaMemory

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

the class SimpleConvolutionLayer method getCudaFilter_cached.

@Nonnull
private CudaMemory getCudaFilter_cached(final CudaDevice deviceNumber) {
    CudaMemory cudaMemory;
    if (gpuFilters.containsKey(deviceNumber.getDeviceId())) {
        cudaMemory = gpuFilters.get(deviceNumber.getDeviceId());
    } else {
        double[] data = kernel.getData();
        cudaMemory = deviceNumber.allocate((long) data.length * precision.size, MemoryType.Device, true).write(precision, data);
        CudaMemory replaced = gpuFilters.put(deviceNumber.getDeviceId(), cudaMemory);
        if (null != replaced)
            replaced.freeRef();
    }
    cudaMemory.addRef();
    return cudaMemory;
}

Example 23 with CudaMemory

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

the class SoftmaxActivationLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    final Result inputResult = inObj[0];
    final TensorList inputData = inputResult.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    @Nonnull final int[] outputSize = inputSize;
    final int length = inputData.length();
    final int inputDims = Tensor.length(inputSize);
    try {
        final CudaTensor outPtr = CudaSystem.run(gpu -> {
            @Nullable CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
            final CudaTensor outputTensor;
            if (1 == inputData.currentRefCount() && 1 == inputTensor.currentRefCount()) {
                outputTensor = inputTensor;
                outputTensor.addRef();
            } else {
                @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, inputSize[2], inputSize[1], inputSize[0], inputSize[2] * inputSize[1] * inputSize[0], inputSize[1] * inputSize[0], inputSize[0], 1);
                @Nonnull final CudaMemory outputData = gpu.allocate(precision.size * 1l * inputDims * length, MemoryType.Managed.normalize(), true);
                outputTensor = CudaTensor.wrap(outputData, outputDescriptor, precision);
            }
            try {
                CudaMemory inputMemory = inputTensor.getMemory(gpu);
                CudaMemory outputMemory = outputTensor.getMemory(gpu);
                CudaSystem.handle(gpu.cudnnSoftmaxForward(algorithm.code, mode.code, precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputMemory.getPtr(), precision.getPointer(0.0), outputTensor.descriptor.getPtr(), outputMemory.getPtr()));
                assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
                inputMemory.dirty();
                outputMemory.dirty();
                outputMemory.freeRef();
                inputMemory.freeRef();
                return outputTensor;
            } catch (@Nonnull final Throwable e) {
                throw new ComponentException("Error apply " + Arrays.toString(inputSize), e);
            } finally {
                inputTensor.freeRef();
            }
        }, inputData);
        return new Result(CudaTensorList.create(outPtr, length, outputSize, precision), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
            if (inputResult.isAlive()) {
                final TensorList data = CudaSystem.run(gpu -> {
                    @Nullable CudaTensor inputTensor;
                    synchronized (gpu) {
                        inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, true);
                    }
                    @Nullable CudaTensor deltaTensor;
                    synchronized (gpu) {
                        deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
                    }
                    outPtr.addRef();
                    CudaTensor localOut = outPtr.getDenseAndFree(gpu);
                    delta.freeRef();
                    CudaTensor passbackTensor;
                    passbackTensor = CudaTensor.wrap(gpu.allocate((long) Tensor.length(inputSize) * length * precision.size, MemoryType.Managed.normalize(), false), gpu.newTensorDescriptor(precision, delta.length(), inputSize[2], inputSize[1], inputSize[0], inputSize[2] * inputSize[1] * inputSize[0], inputSize[1] * inputSize[0], inputSize[0], 1), precision);
                    try {
                        CudaMemory localOutMemory = localOut.getMemory(gpu);
                        CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
                        CudaMemory inputMemory = inputTensor.getMemory(gpu);
                        CudaMemory passbackMemory = passbackTensor.getMemory(gpu);
                        CudaSystem.handle(gpu.cudnnSoftmaxBackward(algorithm.code, mode.code, precision.getPointer(1.0), localOut.descriptor.getPtr(), localOutMemory.getPtr(), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), precision.getPointer(0.0), passbackTensor.descriptor.getPtr(), passbackMemory.getPtr()));
                        localOutMemory.dirty();
                        deltaTensorMemory.dirty();
                        passbackMemory.dirty();
                        localOutMemory.freeRef();
                        deltaTensorMemory.freeRef();
                        inputMemory.freeRef();
                        passbackMemory.freeRef();
                    } catch (@Nonnull final Throwable e) {
                        throw new ComponentException("Error apply " + Arrays.toString(inputSize), e);
                    } finally {
                        localOut.freeRef();
                        inputTensor.freeRef();
                        deltaTensor.freeRef();
                    }
                    return CudaTensorList.wrap(passbackTensor, length, inputSize, precision);
                }, delta);
                inputResult.accumulate(buffer, data);
            } else {
                delta.freeRef();
            }
        }) {

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

            @Override
            protected void _free() {
                inputData.freeRef();
                outPtr.freeRef();
                inputResult.freeRef();
            }

            @Override
            public boolean isAlive() {
                return inputResult.isAlive() || !isFrozen();
            }
        };
    } catch (@Nonnull final Throwable e) {
        throw new ComponentException("Error apply image res " + Arrays.toString(inputSize), e);
    }
}

Example 24 with CudaMemory

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

the class SumReducerLayer 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_ADD, 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());
        inputMemory.dirty();
        outputMemory.dirty();
        workspacePtr.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(tensor.get(0));
            tensor.freeRef();
            return tensor1;
        }).toArray(i -> new Tensor[i]));
        input.accumulate(ctx, passback);
    }) {

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

Example 25 with CudaMemory

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

the class ValueLayer method toDevice.

/**
 * To device cuda tensor list.
 *
 * @param data      the data
 * @param precision the precision
 * @return the cuda tensor list
 */
public CudaTensorList toDevice(final Tensor data, final Precision precision) {
    return CudaSystem.run(gpu -> {
        CudaMemory cudaMemory = gpu.allocate(data.length() * precision.size, MemoryType.Managed, true);
        cudaMemory.write(precision, data.getData());
        int[] dimensions = data.getDimensions();
        CudaDevice.CudaTensorDescriptor tensorDescriptor = gpu.newTensorDescriptor(precision, 1, dimensions[2], dimensions[1], dimensions[0]);
        return CudaTensorList.wrap(CudaTensor.wrap(cudaMemory, tensorDescriptor, precision), 1, dimensions, precision);
    });
}