COMPMID-470: Neon Deconvolution.

Implemented by up-sampling the input with zeros insertions between the input samples and convolving the Deconvolution kernels on the up-sampled result.

The upsampling is performed by the function NEDeconvolutionLayerUpsample.
Convolving is done by NEDirectConvolutionLayer.

Change-Id: I25f7ba7c6b99cd9310797972ede40aeff4a54900
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85319
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 441 insertions, 0 deletions

diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h
index 6b3b5c748f..86ed2b2ad7 100644
--- a/tests/datasets/ShapeDatasets.h
+++ b/tests/datasets/ShapeDatasets.h
@@ -198,6 +198,21 @@ public:
     }
 };
 
+/** Data set containing small tensor shapes for deconvolution. */
+class SmallDeconvolutionShapes final : public ShapeDataset
+{
+public:
+    SmallDeconvolutionShapes()
+        : ShapeDataset("InputShape",
+    {
+        TensorShape{ 2U, 3U, 3U, 2U },
+                     TensorShape{ 5U, 5U, 3U },
+                     TensorShape{ 11U, 13U, 4U, 3U }
+    })
+    {
+    }
+};
+
 /** Data set containing small tensor shapes for direct convolution. */
 class SmallDirectConvolutionShapes final : public ShapeDataset
 {
diff --git a/tests/validation/CPP/DeconvolutionLayer.cpp b/tests/validation/CPP/DeconvolutionLayer.cpp
new file mode 100644
index 0000000000..34f3d10edb
--- /dev/null
+++ b/tests/validation/CPP/DeconvolutionLayer.cpp
@@ -0,0 +1,108 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "ConvolutionLayer.h"
+
+#include "tests/validation/FixedPoint.h"
+#include "tests/validation/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T>
+SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<T> &bias, const TensorShape &output_shape,
+                                    const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a)
+{
+    // Create reference
+    TensorShape scaled_shape = src.shape();
+    scaled_shape.set(0, output_shape.x());
+    scaled_shape.set(1, output_shape.y());
+    SimpleTensor<T> scaled{ scaled_shape, src.data_type(), 1, src.fixed_point_position() };
+
+    const int          width_in      = src.shape().x();
+    const int          height_in     = src.shape().y();
+    const int          width_scaled  = scaled.shape().x();
+    const int          height_scaled = scaled.shape().y();
+    const int          num_2d_slices = src.shape().total_size() / (width_in * height_in);
+    const auto         width_ratio   = static_cast<float>(width_in) / static_cast<float>(width_scaled);
+    const auto         height_ratio  = static_cast<float>(height_in) / static_cast<float>(height_scaled);
+    const int          ax            = a.first;  // The number of zeros added to right edge of the input.
+    const int          ay            = a.second; // The number of zeros added to bottom edge of the input.
+    const unsigned int kernel_size   = weights.shape().x();
+    ARM_COMPUTE_ERROR_ON(info.pad().first > (kernel_size - 1));
+    const int transposed_convolution_padx = kernel_size - info.pad().first - 1;
+    const int transposed_convolution_pady = kernel_size - info.pad().second - 1;
+    const int stridex                     = info.stride().first;
+    const int stridey                     = info.stride().second;
+
+    for(int j = 0; j < scaled.num_elements(); ++j)
+    {
+        scaled[j] = T(0);
+    }
+
+    for(int slice = 0; slice < num_2d_slices; ++slice)
+    {
+        const int offset_slice_in  = slice * width_in * height_in;
+        const int offset_slice_out = slice * width_scaled * height_scaled;
+        for(int yi = ay; yi < height_scaled; yi += stridey)
+        {
+            for(int xi = transposed_convolution_padx; xi < width_scaled; xi += stridex)
+            {
+                const float x_src     = (xi + 0.5f) * width_ratio - 0.5f;
+                const float y_src     = (yi + 0.5f) * height_ratio - 0.5f;
+                T          *out       = scaled.data() + offset_slice_out + xi + yi * width_scaled;
+                const bool  in_bounds = x_src > -1 && y_src > -1 && x_src < width_in && y_src < height_in;
+                const bool  in_axy    = xi < transposed_convolution_padx || xi >= (width_scaled - ax)  // this is checking if the x coordinate is in the padded left/right area
+                                        || yi < ay || yi >= (height_scaled - transposed_convolution_pady); // like above but top and bottom padding in the upscaled XY plane
+                if(!in_axy)
+                {
+                    if(in_bounds)
+                    {
+                        const int in_scaled_x = support::cpp11::round(x_src);
+                        const int in_scaled_y = support::cpp11::round(y_src);
+                        const T *in          = src.data() + offset_slice_in + in_scaled_x + in_scaled_y * width_in;
+                        *out                  = *in;
+                    }
+                    else
+                    {
+                        *out = T(0);
+                    }
+                }
+            }
+        }
+    }
+    const PadStrideInfo conv_info(1, 1, 1, 1, DimensionRoundingType::CEIL);
+    return convolution_layer(scaled, weights, bias, output_shape, conv_info);
+}
+
+template SimpleTensor<float> deconvolution_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, const TensorShape &output_shape,
+                                                 const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/CPP/DeconvolutionLayer.h b/tests/validation/CPP/DeconvolutionLayer.h
new file mode 100644
index 0000000000..8222e32027
--- /dev/null
+++ b/tests/validation/CPP/DeconvolutionLayer.h
@@ -0,0 +1,55 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_TEST_DECONVOLUTION_LAYER_H__
+#define __ARM_COMPUTE_TEST_DECONVOLUTION_LAYER_H__
+
+#include "tests/SimpleTensor.h"
+#include "tests/validation/Helpers.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+/** Deconvolution reference implementation.
+ *
+ * src              Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs. Data types supported: F32.
+ * weights          The 4d weights with dimensions [width, height, OFM, IFM]. Data type supported: Same as @p input.
+ * bias             Optional, ignored if NULL. The biases have one dimension. Data type supported: Same as @p input.
+ * output_shape     Output tensor shape. The output has the same number of dimensions as the @p input.
+ * info             Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
+ * a                The number of zeros added to right edge of the input.
+ *
+ */
+template <typename T>
+SimpleTensor<T> deconvolution_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<T> &bias, const TensorShape &output_shape, const PadStrideInfo &info,
+                                    const std::pair<unsigned int, unsigned int> &a);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_TEST_DECONVOLUTION_LAYER_H__ */
diff --git a/tests/validation/NEON/DeconvolutionLayer.cpp b/tests/validation/NEON/DeconvolutionLayer.cpp
new file mode 100644
index 0000000000..751a96558a
--- /dev/null
+++ b/tests/validation/NEON/DeconvolutionLayer.cpp
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/DeconvolutionLayerFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */
+
+const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0,
+                     2)
+                     * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("ax", 1, 3) * framework::dataset::make("ay", 1, 3) * framework::dataset::make("NumKernels", { 1, 3 })
+                     *framework::dataset::make("ux", 1, 4) *framework::dataset::make("uy", 1, 4);
+
+const auto data1x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0,
+                     1)
+                     * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("ax", 1, 3) * framework::dataset::make("ay", 1, 3) * framework::dataset::make("NumKernels", { 1, 3 })
+                     *framework::dataset::make("ux", 1, 4) *framework::dataset::make("uy", 1, 4);
+
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(DeconvolutionLayer)
+
+template <typename T>
+using NEDeconvolutionLayerFixture3x3 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 3, 3>;
+
+template <typename T>
+using NEDeconvolutionLayerFixture1x1 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 1, 1>;
+
+TEST_SUITE(Float)
+
+TEST_SUITE(FP32)
+TEST_SUITE(W3x3)
+
+FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL, combine(data3x3, framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fp32);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(W1x1)
+FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerFixture1x1<float>, framework::DatasetMode::ALL, combine(data1x1, framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fp32);
+}
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/DeconvolutionLayerFixture.h b/tests/validation/fixtures/DeconvolutionLayerFixture.h
new file mode 100644
index 0000000000..8dff97d83f
--- /dev/null
+++ b/tests/validation/fixtures/DeconvolutionLayerFixture.h
@@ -0,0 +1,168 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "tests/AssetsLibrary.h"
+#include "tests/Globals.h"
+#include "tests/IAccessor.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Fixture.h"
+#include "tests/validation/CPP/DeconvolutionLayer.h"
+#include "tests/validation/Helpers.h"
+
+#include <random>
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DeconvolutionLayerFixtureBase : public framework::Fixture
+{
+public:
+    /*
+     *
+     * @param[in] a The number of zeros added to right and bottom edges of the input.
+     * @param[in] u How much to scale the X and Y axis.
+     */
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info,
+               const std::pair<unsigned int, unsigned int> &a, const std::pair<unsigned int, unsigned int> &u, DataType data_type, int fractional_bits)
+    {
+        _fractional_bits = fractional_bits;
+        _data_type       = data_type;
+
+        _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, a, u, data_type, fractional_bits);
+        _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, a, data_type, fractional_bits);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::F32:
+            {
+                std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            default:
+                library->fill_tensor_uniform(tensor, i);
+        }
+    }
+    /*
+     *
+     * @param[in] a The number of zeros added to right and bottom edges of the input.
+     * @param[in] u How much to scale the X and Y axis.
+     */
+    TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape,
+                              const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> &a, const std::pair<float, float> &u, DataType data_type, int fixed_point_position)
+    {
+        // Create tensors
+        TensorType src     = create_tensor<TensorType>(input_shape, data_type, 1, fixed_point_position);
+        TensorType weights = create_tensor<TensorType>(weights_shape, data_type, 1, fixed_point_position);
+        TensorType bias    = create_tensor<TensorType>(bias_shape, data_type, 1, fixed_point_position);
+        TensorType dst     = create_tensor<TensorType>(output_shape, data_type, 1, fixed_point_position);
+
+        // Create and configure function
+        FunctionType conv;
+        conv.configure(&src, &weights, &bias, &dst, info, a.first, a.second, u.first, u.second);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        weights.allocator()->allocate();
+        bias.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src), 0);
+        fill(AccessorType(weights), 1);
+        fill(AccessorType(bias), 2);
+
+        // Compute NEConvolutionLayer function
+        conv.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape,
+                                      const PadStrideInfo &info, const std::pair<unsigned int, unsigned int> a, DataType data_type, int fixed_point_position)
+    {
+        // Create reference
+        SimpleTensor<T> src{ input_shape, data_type, 1, fixed_point_position };
+        SimpleTensor<T> weights{ weights_shape, data_type, 1, fixed_point_position };
+        SimpleTensor<T> bias{ bias_shape, data_type, 1, fixed_point_position };
+
+        // Fill reference
+        fill(src, 0);
+        fill(weights, 1);
+        fill(bias, 2);
+
+        return reference::deconvolution_layer<T>(src, weights, bias, output_shape, info, a);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+    int             _fractional_bits{};
+    DataType        _data_type{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T, unsigned int kernel_size_x, unsigned int kernel_size_y>
+class DeconvolutionValidationFixture : public DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, unsigned int sx, unsigned int sy, unsigned int padx, unsigned int pady,
+               unsigned int ax, unsigned int ay, unsigned int ux, unsigned int uy, unsigned int num_kernels, DataType data_type)
+    {
+        ARM_COMPUTE_ERROR_ON_MSG(kernel_size_x != kernel_size_y, "Only square kernels supported");
+        const TensorShape   weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels);
+        const TensorShape   bias_shape(num_kernels);
+        const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL);
+        const std::pair<unsigned int, unsigned int> a(ax, ay);
+        const std::pair<float, float>               u(ux, uy);
+        auto out_dim = deconvolution_output_dimensions(input_shape.x(), input_shape.y(), kernel_size_x, kernel_size_y, padx, pady, a.first, a.second, u.first, u.second,
+                                                       DimensionRoundingType::CEIL);
+        TensorShape output_shape = deconvolution_output_shape(out_dim, input_shape, weights_shape);
+        DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, a, u, data_type, 0);
+    }
+};
+
+} // namespace validation
+} // namespace test
+} // namespace arm_compute