COMPMID-1843: Implement NECrop

Change-Id: I27e8b1a00c2315c72106e8e596f84ad48fb770e3
Signed-off-by: George Wort <george.wort@arm.com>
Reviewed-on: https://review.mlplatform.org/c/648
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>

6 files changed, 266 insertions, 21 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 8b37b2f603..f1d94c89db 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -46,6 +46,7 @@
 #include "arm_compute/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h"
 #include "arm_compute/core/NEON/kernels/NEConvolutionKernel.h"
 #include "arm_compute/core/NEON/kernels/NECopyKernel.h"
+#include "arm_compute/core/NEON/kernels/NECropKernel.h"
 #include "arm_compute/core/NEON/kernels/NECumulativeDistributionKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDepthConcatenateLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEDepthConvertLayerKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NECropKernel.h b/arm_compute/core/NEON/kernels/NECropKernel.h
new file mode 100644
index 0000000000..6713a40c86
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NECropKernel.h
@@ -0,0 +1,123 @@
+/*
+ * Copyright (c) 2019 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_NEON_CROP_KERNEL_H__
+#define __ARM_COMPUTE_NEON_CROP_KERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Types.h"
+
+#include <cstdint>
+#include <map>
+
+namespace arm_compute
+{
+// Forward declarations
+class ITensor;
+
+/** Interface for the kernel to perform tensor cropping */
+class NECropKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NECropKernel";
+    }
+    /** Default constructor */
+    NECropKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NECropKernel(const NECropKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NECropKernel &operator=(const NECropKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NECropKernel(NECropKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NECropKernel &operator=(NECropKernel &&) = default;
+    /** Default destructor */
+    ~NECropKernel() = default;
+    /** Configure kernel
+     *
+     * @note Supported tensor rank: up to 4
+     * @note Padding not supported.
+     *
+     * @param[in]  input               Source tensor. Data type supported: U16/S16/U32/S32/F16/F32. Data layouts supported: NHWC.
+     * @param[in]  crop_boxes          Tensor containing all possible boxes used to crop the image, each represented by 4 normalized values.
+     *                                 Data type supported: F32
+     * @param[in]  box_ind             One dimensional tensor mapping the @p crop_box_ind to the index of the 3D image in @p input.
+     *                                 Data type supported: F32
+     * @param[out] output              Destination tensor. Data type supported: F32
+     * @param[in]  crop_box_ind        Index of the crop box to be used from @p crop_boxes. Default is 0.
+     * @param[in]  extrapolation_value Value to be used for values outside of the image. Default is 0.
+     */
+    void configure(const ITensor *input, const ITensor *crop_boxes, const ITensor *box_ind, ITensor *output, uint32_t crop_box_ind = 0, float extrapolation_value = 0);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CLStridedSliceKernel
+     *
+     * @note Supported tensor rank: up to 4
+     * @note Padding not supported.
+     *
+     * @param[in] input               Source tensor info. Data type supported: U16/S16/U32/S32/F16/F32. Data layouts supported: NHWC.
+     * @param[in] crop_boxes          Tensor info for tensor containing all possible boxes used to crop the image. Data type supported: F32
+     * @param[in] box_ind             Tensor info for the one dimensional tensor mapping the @p crop_box_ind to the index of the 3D image
+     *                                in @p input. Data type supported: F32
+     * @param[in] output              Destination tensor. Data type supported: F32
+     * @param[in] crop_box_ind        Index of the crop box to be used from @p crop_boxes. Default is 0.
+     * @param[in] extrapolation_value Value to be used for values outside of the image. Default is 0.
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *crop_boxes, const ITensorInfo *box_ind, const ITensorInfo *output, uint32_t crop_box_ind = 0, float extrapolation_value = 0);
+
+    /** Configure output tensor's shape as this can only be determined at runtime. */
+    void configure_output_shape();
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+    /** Function to use for in bounds crop for the particular tensor types passed to configure() */
+    using InBoundsCropFunction = void(const ITensor *, const ITensor *, float *, Coordinates, int32_t, int32_t, int32_t);
+
+private:
+    const ITensor *_input;
+    const ITensor *_crop_boxes;
+    const ITensor *_box_ind;
+    ITensor       *_output;
+
+    Coordinates _start;
+    Coordinates _end;
+    uint32_t    _crop_box_ind;
+    float       _extrapolation_value;
+    /** The number of rows out of bounds at the start and end of output. */
+    uint32_t _rows_out_of_bounds[2];
+    /** The number of columns out of bounds at the start and end of output. */
+    uint32_t _cols_out_of_bounds[2];
+
+    std::pair<NECropKernel::InBoundsCropFunction *, NECropKernel::InBoundsCropFunction *> _in_bounds_crop_functions;
+    NECropKernel::InBoundsCropFunction *_in_bounds_crop_function;
+
+    using CropFunction = void(const ITensor *, const ITensor *, Coordinates, float, const uint32_t *, const uint32_t *,
+                              NECropKernel::InBoundsCropFunction *);
+
+    NECropKernel::CropFunction *_crop_function;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NEON_CROP_KERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEScaleKernel.h b/arm_compute/core/NEON/kernels/NEScaleKernel.h
index 83d99643dc..b132bb57b6 100644
--- a/arm_compute/core/NEON/kernels/NEScaleKernel.h
+++ b/arm_compute/core/NEON/kernels/NEScaleKernel.h
@@ -55,33 +55,41 @@ public:
     /** Initialise the kernel's inputs, output and interpolation policy
      *
      * @note dx, dy and offsets have the same dimensions (width and height) of the output tensor
+     * @note Using @p policy Area only supports data layout NCHW and input data type U8.
      *
-     * @param[in]  input           Source tensor. Data types supported: U8/S16/F16/F32.
-     * @param[in]  dx              Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
-     * @param[in]  dy              Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
-     * @param[in]  offsets         Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
-     * @param[out] output          Destination tensor. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
-     * @param[in]  policy          Interpolation type to use
-     * @param[in]  border_mode     Border mode policy
-     * @param[in]  sampling_policy (Optional) Sampling policy used by the interpolation. Defaults to @ref SamplingPolicy::CENTER
+     * @param[in]  input                 Source tensor. Data types supported: U8/S16/F16/F32.
+     * @param[in]  dx                    Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
+     * @param[in]  dy                    Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
+     * @param[in]  offsets               Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
+     * @param[out] output                Destination tensor. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
+     * @param[in]  policy                Interpolation type to use
+     * @param[in]  border_mode           Border mode policy
+     * @param[in]  constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT and use_padding is set to false.
+     * @param[in]  sampling_policy       (Optional) Sampling policy used by the interpolation. Defaults to @ref SamplingPolicy::CENTER
+     * @param[in]  use_padding           (Optional) Is padding in use or not. Defaults to true.
      */
     void configure(const ITensor *input, const ITensor *dx, const ITensor *dy, const ITensor *offsets, ITensor *output,
-                   InterpolationPolicy policy, BorderMode border_mode, SamplingPolicy sampling_policy = SamplingPolicy::CENTER);
+                   InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value = PixelValue(),
+                   SamplingPolicy sampling_policy = SamplingPolicy::CENTER, bool use_padding = true);
     /** Static function to check if given info will lead to a valid configuration of @ref NEScaleKernel
      *
      * @note dx, dy and offsets have the same dimensions (width and height) of the output tensor
+     * @note Using @p policy Area only supports data layout NCHW and input data type U8.
      *
-     * @param[in] input           Source tensor. Data types supported: U8/S16/F16/F32.
-     * @param[in] dx              Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
-     * @param[in] dy              Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
-     * @param[in] offsets         Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
-     * @param[in] output          Destination tensor. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
-     * @param[in] policy          Interpolation type to use
-     * @param[in] border_mode     Border mode policy
-     * @param[in] sampling_policy (Optional) Sampling policy used by the interpolation. Defaults to @ref SamplingPolicy::CENTER
+     * @param[in] input                 Source tensor. Data types supported: U8/S16/F16/F32.
+     * @param[in] dx                    Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
+     * @param[in] dy                    Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
+     * @param[in] offsets               Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
+     * @param[in] output                Destination tensor. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. scaling is only performed within the XY-plane.
+     * @param[in] policy                Interpolation type to use
+     * @param[in] border_mode           Border mode policy
+     * @param[in] constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT and use_padding is set to false.
+     * @param[in] sampling_policy       (Optional) Sampling policy used by the interpolation. Defaults to @ref SamplingPolicy::CENTER
+     * @param[in] use_padding           (Optional) Is padding in use or not. Defaults to true.
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets, ITensorInfo *output,
-                           InterpolationPolicy policy, BorderMode border_mode, SamplingPolicy sampling_policy = SamplingPolicy::CENTER);
+                           InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value = PixelValue(),
+                           SamplingPolicy sampling_policy = SamplingPolicy::CENTER, bool use_padding = true);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
@@ -110,7 +118,9 @@ private:
     InterpolationPolicy _policy;
     BorderSize          _border_size;
     BorderMode          _border_mode;
+    PixelValue          _constant_border_value;
     float               _sampling_offset;
+    bool                _use_padding;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_NESCALEKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 15ce4e3d66..432c751308 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -49,6 +49,7 @@
 #include "arm_compute/runtime/NEON/functions/NEConvolution.h"
 #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
 #include "arm_compute/runtime/NEON/functions/NECopy.h"
+#include "arm_compute/runtime/NEON/functions/NECropResize.h"
 #include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEDepthConvertLayer.h"
diff --git a/arm_compute/runtime/NEON/functions/NECropResize.h b/arm_compute/runtime/NEON/functions/NECropResize.h
new file mode 100644
index 0000000000..e790e68b5f
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NECropResize.h
@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2019 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_NEON_CROP_RESIZE_H__
+#define __ARM_COMPUTE_NEON_CROP_RESIZE_H__
+
+#include "arm_compute/core/NEON/kernels/NECropKernel.h"
+#include "arm_compute/runtime/NEON/functions/NEScale.h"
+
+#include <cstdint>
+#include <memory>
+
+namespace arm_compute
+{
+// Forward Declarations
+class ITensor;
+
+/** Function to perform cropping and resizing */
+class NECropResize : public IFunction
+{
+public:
+    /** Default constructor */
+    NECropResize();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NECropResize(const NECropResize &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NECropResize &operator=(const NECropResize &) = delete;
+    /** Allow instances of this class to be moved */
+    NECropResize(NECropResize &&) = default;
+    /** Allow instances of this class to be moved */
+    NECropResize &operator=(NECropResize &&) = default;
+    /** Default destructor */
+    virtual ~NECropResize() = default;
+
+    /** Configure kernel
+     *
+     * @note Supported tensor rank: up to 4
+     * @note Box indices may be outside of the bounds, in which case @p extrapolation_value is used.
+     * @note Start and end indices of boxes are inclusive.
+     *
+     * @param[in]  input               Source tensor containing N batches of 3D images to be cropped. Data type supported: U16/S16/U32/S32/F16/F32
+     * @param[in]  boxes               Tensor containing the boxes used to crop the images. Data type supported: F32
+     * @param[in]  box_ind             One dimensional tensor containing the batch index of the 3D image in @p input that the corresponding
+     *                                 box in @p boxes will be applied to. Data type supported: F32
+     * @param[out] output              Destination tensor containing a cropped and resized image for each box in @p boxes. Data type supported: F32
+     * @param[in]  crop_size           The dimensions that each cropped image will be resized to.
+     * @param[in]  method              The policy to be used when resizing image. Default is bilinear.
+     * @param[in]  extrapolation_value Value to be used for values outside of the image for cropping and resizing. Default is 0.
+     */
+    void configure(const ITensor *input, const ITensor *boxes, const ITensor *box_ind, ITensor *output, Coordinates2D crop_size,
+                   InterpolationPolicy method = InterpolationPolicy::BILINEAR, float extrapolation_value = 0);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref NESlice
+     *
+     * @note Supported tensor rank: up to 4
+     * @note Box indices may be outside of the bounds, in which case @p extrapolation_value is used.
+     * @note Start and end indices of boxes are inclusive.
+     *
+     * @param[in] input               Source tensor containing N batches of 3D images to be cropped. Data type supported: U16/S16/U32/S32/F16/F32
+     * @param[in] boxes               Tensor info for the tensor containing the boxes used to crop the images. Data type supported: F32
+     * @param[in] box_ind             Tensor info for the one dimensional tensor containing the batch index of the 3D image in @p input
+     *                                that the corresponding box in @p boxes will be applied to. Data type supported: F32
+     * @param[in] output              Tensor info for the destination tensor containing a cropped and resized image for each box in @p boxes.
+     *                                Data type supported: F32
+     * @param[in] crop_size           The dimensions that each cropped image will be resized to.
+     * @param[in] method              The policy to be used when resizing image. Default is bilinear.
+     * @param[in] extrapolation_value Value to be used for values outside of the image for cropping and resizing. Default is 0.
+     *
+     * @return A status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *boxes, const ITensorInfo *box_ind, const ITensorInfo *output,
+                           Coordinates2D crop_size, InterpolationPolicy method, float extrapolation_value);
+
+    void run() override;
+
+    ITensor            *_output;
+    size_t              _num_boxes;
+    InterpolationPolicy _method;
+    float               _extrapolation_value;
+
+    std::unique_ptr<NECropKernel[]> _crop;
+    std::unique_ptr<NEScale[]>      _scale;
+    std::unique_ptr<Tensor[]>       _crop_results{ nullptr };
+    std::unique_ptr<Tensor[]>       _scaled_results{ nullptr };
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEON_CROP_RESIZE_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEScale.h b/arm_compute/runtime/NEON/functions/NEScale.h
index d59e3cccb6..d7dfbbfc9f 100644
--- a/arm_compute/runtime/NEON/functions/NEScale.h
+++ b/arm_compute/runtime/NEON/functions/NEScale.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -53,9 +53,10 @@ public:
      * @param[in]      border_mode           Strategy to use for borders.
      * @param[in]      constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT.
      * @param[in]      sampling_policy       (Optional) Sampling policy used by the interpolation. Defaults to @ref SamplingPolicy::CENTER
+     * @param[in]      use_padding           (Optional) Is padding in use or not. Defaults to true.
      */
     void configure(ITensor *input, ITensor *output, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value = PixelValue(),
-                   SamplingPolicy sampling_policy = SamplingPolicy::CENTER);
+                   SamplingPolicy sampling_policy = SamplingPolicy::CENTER, bool use_padding = true);
     /** Static function to check if given info will lead to a valid configuration of @ref NEScale
      *
      * @param[in] input                 Source tensor. Data type supported: U8/S16/F16/F32. (Written to only for @p border_mode != UNDEFINED)
@@ -64,11 +65,12 @@ public:
      * @param[in] border_mode           Strategy to use for borders.
      * @param[in] constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT.
      * @param[in] sampling_policy       (Optional) Sampling policy used by the interpolation. Defaults to @ref SamplingPolicy::CENTER
+     * @param[in] use_padding           (Optional) Is padding in use or not. Defaults to true.
      *
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, InterpolationPolicy policy, BorderMode border_mode,
-                           PixelValue constant_border_value = PixelValue(), SamplingPolicy sampling_policy = SamplingPolicy::CENTER);
+                           PixelValue constant_border_value = PixelValue(), SamplingPolicy sampling_policy = SamplingPolicy::CENTER, bool use_padding = true);
 
     // Inherited methods overridden:
     void run() override;
@@ -79,6 +81,7 @@ private:
     Tensor             _dy;             /**< Element's distance between the Y real coordinate and the smallest Y following integer */
     NEScaleKernel      _scale_kernel;   /**< Kernel to perform the scaling */
     NEFillBorderKernel _border_handler; /**< kernel to handle tensor borders */
+    bool               _use_padding;    /**< Is padding used on the tensors */
 };
 }
 #endif /*__ARM_COMPUTE_NESCALEIMAGE_H__ */