COMPMID-: Implement L2NormalizationLayer.

Change-Id: I2223850b84826a87650729d287c57c22bcab7462
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/82767
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

4 files changed, 416 insertions, 0 deletions

diff --git a/src/core/NEON/kernels/NEL2NormalizeKernel.cpp b/src/core/NEON/kernels/NEL2NormalizeKernel.cpp
new file mode 100644
index 0000000000..f3f58b6eb3
--- /dev/null
+++ b/src/core/NEON/kernels/NEL2NormalizeKernel.cpp
@@ -0,0 +1,125 @@
+/*
+ * 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/NEON/kernels/NEL2NormalizeKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/NEMath.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <arm_neon.h>
+#include <cmath>
+
+using namespace arm_compute;
+
+namespace
+{
+void l2_normalize_X(const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window)
+{
+    Window window_sum(window);
+    window_sum.set(Window::DimX, Window::Dimension(0, 0, 0));
+
+    Window in_slice  = window.first_slice_window_1D();
+    Window sum_slice = window_sum.first_slice_window_1D();
+
+    do
+    {
+        Iterator input_it(in, in_slice);
+        Iterator sum_it(sum, sum_slice);
+        Iterator output_it(out, in_slice);
+
+        const float       sum_value           = *reinterpret_cast<const float *>(sum_it.ptr());
+        const float32x4_t vec_normalize_value = vdupq_n_f32(1.f / std::sqrt(std::max(sum_value, epsilon)));
+
+        execute_window_loop(in_slice, [&](const Coordinates & id)
+        {
+            const auto in_ptr  = reinterpret_cast<const float *>(input_it.ptr());
+            const auto out_ptr = reinterpret_cast<float *>(output_it.ptr());
+
+            vst1q_f32(out_ptr, vmulq_f32(vld1q_f32(in_ptr), vec_normalize_value));
+        },
+        input_it, output_it);
+    }
+    while(window.slide_window_slice_1D(in_slice) && window.slide_window_slice_1D(sum_slice));
+}
+} // namespace
+
+NEL2NormalizeKernel::NEL2NormalizeKernel()
+    : _input(nullptr), _sum(nullptr), _output(nullptr), _axis(0), _epsilon(1e-12)
+{
+}
+
+void NEL2NormalizeKernel::configure(const ITensor *input, const ITensor *sum, ITensor *output, unsigned int axis, float epsilon)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output);
+    ARM_COMPUTE_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Normalization axis greater than max number of dimensions");
+    ARM_COMPUTE_ERROR_ON_MSG(axis > 0, "Unsupported normalization axis, Supported axis is 0");
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position());
+
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, sum);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
+
+    unsigned int num_elems_processed_per_iteration     = 16 / data_size_from_type(input->info()->data_type());
+    unsigned int num_elems_processed_per_iteration_sum = (axis == 0) ? 1 : num_elems_processed_per_iteration;
+
+    _input   = input;
+    _sum     = sum;
+    _output  = output;
+    _axis    = axis;
+    _epsilon = epsilon;
+
+    // Configure kernel window
+    Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal sum_access(sum->info(), 0, num_elems_processed_per_iteration_sum);
+    AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
+
+    update_window_and_padding(win, input_access, sum_access, output_access);
+
+    output_access.set_valid_region(win, input->info()->valid_region());
+
+    INEKernel::configure(win);
+}
+
+void NEL2NormalizeKernel::run(const Window &window)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    switch(_axis)
+    {
+        case 0:
+            l2_normalize_X(_input, _sum, _output, _epsilon, window);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported normalization axis");
+    }
+}
diff --git a/src/core/NEON/kernels/NEReductionOperationKernel.cpp b/src/core/NEON/kernels/NEReductionOperationKernel.cpp
new file mode 100644
index 0000000000..5c8954609b
--- /dev/null
+++ b/src/core/NEON/kernels/NEReductionOperationKernel.cpp
@@ -0,0 +1,157 @@
+/*
+ * 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/NEON/kernels/NEReductionOperationKernel.h"
+
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/NEON/NEMath.h"
+#include "arm_compute/core/Validate.h"
+
+#include <arm_neon.h>
+
+using namespace arm_compute;
+
+namespace
+{
+template <class F>
+class Reducer
+{
+public:
+    static void reduceX(const Window &window, const ITensor *input, ITensor *output, F f)
+    {
+        // Set out window
+        Window out_window(window);
+        out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
+
+        // Get first input and output slices
+        Window in_slice  = window.first_slice_window_1D();
+        Window out_slice = out_window.first_slice_window_1D();
+
+        do
+        {
+            Iterator in(input, in_slice);
+            Iterator out(output, out_slice);
+
+            f(in, out, in_slice, out_slice);
+        }
+        while(window.slide_window_slice_1D(in_slice) && window.slide_window_slice_1D(out_slice));
+    }
+};
+
+struct SumsqOpX
+{
+    inline void operator()(Iterator &input, Iterator &output, Window &in_slice, Window &out_slice)
+    {
+        ARM_COMPUTE_UNUSED(out_slice);
+        float32x4_t vec_sum_value = vdupq_n_f32(0.f);
+
+        execute_window_loop(in_slice, [&](const Coordinates & id)
+        {
+            const auto        in_ptr       = reinterpret_cast<const float *>(input.ptr());
+            const float32x4_t vec_elements = vld1q_f32(in_ptr);
+            vec_sum_value                  = vaddq_f32(vmulq_f32(vec_elements, vec_elements), vec_sum_value);
+        },
+        input);
+
+        float32x2_t carry_addition = vpadd_f32(vget_high_f32(vec_sum_value), vget_low_f32(vec_sum_value));
+        carry_addition             = vpadd_f32(carry_addition, carry_addition);
+
+        *(reinterpret_cast<float *>(output.ptr())) = vget_lane_f32(carry_addition, 0);
+    }
+};
+
+void reduce_sumsq(const Window &window, const ITensor *input, ITensor *output, unsigned int axis)
+{
+    switch(axis)
+    {
+        case 0:
+            return Reducer<SumsqOpX>::reduceX(window, input, output, SumsqOpX());
+        default:
+            ARM_COMPUTE_ERROR("Unsupported reduction axis");
+    }
+}
+} // namespace
+
+NEReductionOperationKernel::NEReductionOperationKernel()
+    : _input(nullptr), _output(nullptr), _reduction_axis(0), _op(ReductionOperation::SUM_SQUARE), _border_size()
+{
+}
+
+BorderSize NEReductionOperationKernel::border_size() const
+{
+    return _border_size;
+}
+
+void NEReductionOperationKernel::configure(const ITensor *input, ITensor *output, unsigned int axis, ReductionOperation op)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
+    ARM_COMPUTE_ERROR_ON_MSG(axis > 0, "Unsupported reduction axis, Supported axis is 0");
+
+    // Calculate output shape and set if empty
+    TensorShape output_shape{ input->info()->tensor_shape() };
+    output_shape.set(axis, 1);
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
+
+    unsigned int num_elems_processed_per_iteration = 16 / data_size_from_type(input->info()->data_type());
+
+    _input       = input;
+    _output      = output;
+    _border_size = (axis == 0) ? BorderSize(0, num_elems_processed_per_iteration - (input->info()->dimension(0) % num_elems_processed_per_iteration), 0, 0) : BorderSize();
+    _op          = op;
+
+    // Configure kernel window
+    Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
+
+    update_window_and_padding(win, input_access, output_access);
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEReductionOperationKernel::run(const Window &window)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    switch(_op)
+    {
+        case ReductionOperation::SUM_SQUARE:
+            reduce_sumsq(window, _input, _output, _reduction_axis);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported reduction operation.");
+    }
+}
diff --git a/src/runtime/NEON/functions/NEL2Normalize.cpp b/src/runtime/NEON/functions/NEL2Normalize.cpp
new file mode 100644
index 0000000000..378d78e3f3
--- /dev/null
+++ b/src/runtime/NEON/functions/NEL2Normalize.cpp
@@ -0,0 +1,50 @@
+/*
+ * 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/runtime/NEON/functions/NEL2Normalize.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+
+using namespace arm_compute;
+
+NEL2Normalize::NEL2Normalize()
+    : _reduce_func(), _normalize_kernel(), _sumsq()
+{
+}
+
+void NEL2Normalize::configure(ITensor *input, ITensor *output, unsigned int axis, float epsilon)
+{
+    // Configure Kernels
+    _reduce_func.configure(input, &_sumsq, axis, ReductionOperation::SUM_SQUARE);
+    _normalize_kernel.configure(input, &_sumsq, output, axis, epsilon);
+
+    // Allocate intermediate tensors
+    _sumsq.allocator()->allocate();
+}
+
+void NEL2Normalize::run()
+{
+    _reduce_func.run();
+    NEScheduler::get().schedule(&_normalize_kernel, Window::DimY);
+}
diff --git a/src/runtime/NEON/functions/NEReductionOperation.cpp b/src/runtime/NEON/functions/NEReductionOperation.cpp
new file mode 100644
index 0000000000..45c3e5dbc9
--- /dev/null
+++ b/src/runtime/NEON/functions/NEReductionOperation.cpp
@@ -0,0 +1,84 @@
+/*
+ * 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/runtime/NEON/functions/NEReductionOperation.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+
+using namespace arm_compute;
+
+namespace
+{
+/** Define dimension to split the window
+ *
+ * @param[in] axis Reduction axis
+ *
+ * @return The dimension to split the window
+ */
+size_t reduction_window_split_dimension(unsigned int axis)
+{
+    switch(axis)
+    {
+        case 0:
+            return Window::DimY;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported reduction axis");
+    }
+}
+BorderMode reduction_operation_border_mode(ReductionOperation op)
+{
+    switch(op)
+    {
+        case ReductionOperation::SUM_SQUARE:
+            return BorderMode::CONSTANT;
+        default:
+            return BorderMode::CONSTANT;
+    }
+}
+} // namespace
+
+NEReductionOperation::NEReductionOperation()
+    : _reduction_kernel(), _fill_border_kernel(), _window_split(0)
+{
+}
+
+void NEReductionOperation::configure(ITensor *input, ITensor *output, unsigned int axis, ReductionOperation op)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+
+    // Configure reduction kernel
+    _reduction_kernel.configure(input, output, axis, op);
+    _window_split = reduction_window_split_dimension(axis);
+
+    // Configure fill border kernel
+    BorderSize fill_border_size = (axis == 0) ? _reduction_kernel.border_size() : BorderSize();
+    BorderMode fill_border_mode = reduction_operation_border_mode(op);
+    _fill_border_kernel.configure(input, fill_border_size, fill_border_mode, PixelValue(0));
+}
+
+void NEReductionOperation::run()
+{
+    NEScheduler::get().schedule(&_fill_border_kernel, Window::DimY);
+    NEScheduler::get().schedule(&_reduction_kernel, _window_split);
+}