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

Example 16 with CudaTensor

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

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

the class ImgCropLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    assert 1 == inObj.length;
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    assert 3 == inputData.getDimensions().length;
    final int length = inputData.length();
    @Nonnull int[] dimIn = inputData.getDimensions();
    if (dimIn[0] == sizeX && dimIn[1] == sizeY) {
        return input;
    }
    @Nonnull final int[] dimOut = Arrays.copyOf(dimIn, 3);
    dimOut[0] = sizeX;
    dimOut[1] = sizeY;
    final TensorList outputData = CudaSystem.run(gpu -> {
        @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        inputData.freeRef();
        boolean dirty = dimOut[0] <= dimIn[0] && dimOut[1] <= dimIn[1];
        assert dimOut[0] > 0;
        assert dimOut[1] > 0;
        assert dimOut[2] > 0;
        CudaTensor cudaTensor = copy(gpu, inputTensor, length, dimIn, dimOut, dirty, precision);
        Stream.<ReferenceCounting>of(inputTensor).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(cudaTensor, length, dimOut, precision);
    }, inputData);
    return new Result(outputData, (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        if (!Arrays.equals(delta.getDimensions(), outputData.getDimensions())) {
            throw new AssertionError(Arrays.toString(delta.getDimensions()) + " != " + Arrays.toString(outputData.getDimensions()));
        }
        if (delta.length() != outputData.length()) {
            throw new AssertionError(delta.length() + " != " + outputData.length());
        }
        assert delta.length() == length;
        if (input.isAlive()) {
            final TensorList passbackTensorList = CudaSystem.run(gpu -> {
                @Nullable final CudaTensor errorPtr = gpu.getTensor(delta, precision, MemoryType.Device, false);
                delta.freeRef();
                boolean dirty = dimOut[0] >= dimIn[0] && dimOut[1] >= dimIn[1];
                CudaTensor cudaTensor = copy(gpu, errorPtr, length, dimOut, dimIn, dirty, precision);
                Stream.<ReferenceCounting>of(errorPtr).forEach(ReferenceCounting::freeRef);
                return CudaTensorList.wrap(cudaTensor, length, dimIn, precision);
            }, delta);
            input.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 18 with CudaTensor

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

the class ImgTileAssemblyLayer method copy.

/**
 * Copy.
 *
 * @param copyParams the copy params
 */
public void copy(final CopyParams copyParams) {
    CudnnHandle gpu = copyParams.gpu;
    gpu.initThread();
    assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
    @Nullable final CudaTensor inputBuffer = gpu.getTensor(copyParams.inObj[copyParams.inputIndex].getData(), precision, MemoryType.Device, false);
    copy(gpu, copyParams.length, copyParams.tileDimensions, inputBuffer, copyParams.outputDims, copyParams.outputBuffer, copyParams.positionX, copyParams.totalHeight);
    inputBuffer.freeRef();
}

Example 19 with CudaTensor

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

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

the class ImgTileSelectLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    assert 1 == inObj.length;
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    assert 3 == inputData.getDimensions().length;
    final int length = inputData.length();
    @Nonnull int[] dimIn = inputData.getDimensions();
    if (dimIn[0] == sizeY && dimIn[1] == sizeX) {
        return input;
    }
    @Nonnull final int[] dimOut = getViewDimensions(dimIn, new int[] { sizeY, sizeX, dimIn[2] }, new int[] { positionX, positionY, 0 });
    final TensorList outputData = CudaSystem.run(gpu -> {
        assert dimOut[0] > 0;
        assert dimOut[1] > 0;
        assert dimOut[2] > 0;
        boolean dirty = dimOut[0] == dimIn[0] && dimOut[1] == dimIn[1];
        CudaTensor cudaTensor = copy(gpu, inputData, dimIn, dimOut, precision, this.positionX, this.positionY, dirty);
        return CudaTensorList.wrap(cudaTensor, length, dimOut, precision);
    }, inputData);
    int[] outputDimensions = outputData.getDimensions();
    assert length == outputData.length();
    return new Result(outputData, (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList error) -> {
        if (!Arrays.equals(error.getDimensions(), outputDimensions)) {
            throw new AssertionError(Arrays.toString(error.getDimensions()) + " != " + Arrays.toString(outputDimensions));
        }
        if (error.length() != length) {
            throw new AssertionError(error.length() + " != " + length);
        }
        assert error.length() == inputData.length();
        if (input.isAlive()) {
            final TensorList passbackTensorList = CudaSystem.run(gpu -> {
                boolean dirty = dimOut[0] >= dimIn[0] && dimOut[1] >= dimIn[1];
                CudaTensor cudaTensor = copy(gpu, error, dimOut, dimIn, precision, -this.positionX, -this.positionY, dirty);
                return CudaTensorList.wrap(cudaTensor, length, dimIn, precision);
            }, error);
            input.accumulate(buffer, passbackTensorList);
        }
    }) {

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

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