COMPMID-401: Implement FixedPointPosition conversion for NEON.

Adds support of changing the fixed point position of a tensor in
DepthConvert.

Change-Id: Ic3b50a4628fac7497a0217d92941c9d6f64d21cb
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/80438
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

2 files changed, 36 insertions, 18 deletions

diff --git a/arm_compute/core/NEON/kernels/NEDepthConvertKernel.h b/arm_compute/core/NEON/kernels/NEDepthConvertKernel.h
index 0c5c29e4db..ad8d152bbf 100644
--- a/arm_compute/core/NEON/kernels/NEDepthConvertKernel.h
+++ b/arm_compute/core/NEON/kernels/NEDepthConvertKernel.h
@@ -24,7 +24,7 @@
 #ifndef __ARM_COMPUTE_DEPTHCONVERTKERNEL_H__
 #define __ARM_COMPUTE_DEPTHCONVERTKERNEL_H__
 
-#include "arm_compute/core/NEON/INESimpleKernel.h"
+#include "arm_compute/core/NEON/INEKernel.h"
 #include "arm_compute/core/Types.h"
 
 #include <cstdint>
@@ -34,35 +34,52 @@ namespace arm_compute
 class ITensor;
 
 /** Depth conversion kernel */
-class NEDepthConvertKernel : public INESimpleKernel
+class NEDepthConvertKernel : public INEKernel
 {
 public:
     /** Default constructor*/
     NEDepthConvertKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthConvertKernel(const NEDepthConvertKernel &) = delete;
+    /** Default move constructor */
+    NEDepthConvertKernel(NEDepthConvertKernel &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEDepthConvertKernel &operator=(const NEDepthConvertKernel &) = delete;
+    /** Default move assignment operator */
+    NEDepthConvertKernel &operator=(NEDepthConvertKernel &&) = default;
     /** Set the input and output of the kernel
      *
      * Valid conversions Input -> Output :
      *
-     *   - QS8 -> F32
+     *   - QS8 -> QS8, F32
      *   - U8 -> U16, S16, S32
      *   - U16 -> U8, U32
      *   - S16 -> U8, S32
+     *   - QS16 -> QS16, F32
      *   - F32 -> QS8
      *
+     * @warning In case of in-place fixed point position conversion make sure that configure has been called
+     *          before the updated tensor is used in other functions, as the TensorInfo of the tensor will be
+     *          altered. In-place is only supported for QS8 -> QS8, QS16 -> QS16.
      *
-     * @param[in]  input  The input tensor to convert. Data types supported: U8/QS8/U16/S16/F32.
-     * @param[out] output The output tensor. Data types supported: U8/QS8/U16/S16/U32/S32/F32.
-     * @param[in]  policy Conversion policy.
-     * @param[in]  shift  Value for down/up conversions. Must be 0 <= shift < 8.
+     * @param[in, out] input  The input tensor to convert (Written in case of in-place computation). Data types supported: U8/QS8/U16/S16/F32.
+     * @param[out]     output The output tensor. Can be null in case of in-place computation. Data types supported: U8/QS8/U16/S16/U32/S32/F32.
+     * @param[in]      policy Conversion policy.
+     * @param[in]      shift  (Optional) Value for down/up conversions. Must be 0 <= shift < 8.
+     *                         In case of fixed point position conversion, it specifies the new fixed point position, if operation is in-place.
      */
-    void configure(const ITensor *input, ITensor *output, ConvertPolicy policy, uint32_t shift);
+    void configure(ITensor *input, ITensor *output, ConvertPolicy policy, uint32_t shift = 0);
 
     // Inherited methods overridden:
     void run(const Window &window) override;
 
 private:
+    ITensor      *_input;
+    ITensor      *_output;
     ConvertPolicy _policy;
     uint32_t      _shift;
+    int           _fixed_point_position_input;
+    int           _fixed_point_position_output;
 };
 }
 #endif /*__ARM_COMPUTE_NEDEPTHCONVERTKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEDepthConvert.h b/arm_compute/runtime/NEON/functions/NEDepthConvert.h
index 47b3a7e6f6..37f7293fb3 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthConvert.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthConvert.h
@@ -45,24 +45,25 @@ public:
     const NEDepthConvert &operator=(const NEDepthConvert &) = delete;
     /** Initialize the function's source, destination
      *
-     * Input format must be different than output format.
-     *
      * Valid conversions Input -> Output :
-     *    QS8 -> F32
+     *    QS8 -> QS8, F32
      *    U8 -> U16, S16, S32
      *    U16 -> U8, U32
      *    S16 -> U8, S32
-     *    QS16 -> F32
+     *    QS16 -> QS16, F32
      *    F32 -> QS8, QS16
      *
+     * @warning In case of in-place fixed point position conversion make sure that configure has been called
+     *          before the updated tensor is used in other functions, as the TensorInfo of the tensor will be
+     *          altered. In-place is only supported for QS8 -> QS8, QS16 -> QS16.
      *
-     * @param[in]  input  The input tensor to convert. Data type supported: QS8/U8/U16/S16/QS16/F32.
-     * @param[out] output The output tensor. Data type supported: QS8/U8/U16/S16/QS16/U32/S32/F32.
-     * @param[in]  policy Conversion policy.
-     * @param[in]  shift  Value for down/up conversions. Must be 0 <= shift < 8.
-     *                    It is not used on fixed point conversion.
+     * @param[in, out] input  The input tensor to convert (Written in case of in-place computation). Data types supported: U8/QS8/U16/S16/F32.
+     * @param[out]     output The output tensor. Can be null in case of in-place computation. Data types supported: U8/QS8/U16/S16/U32/S32/F32.
+     * @param[in]      policy Conversion policy.
+     * @param[in]      shift  (Optional) Value for down/up conversions. Must be 0 <= shift < 8.
+     *                        In case of fixed point position conversion, it specifies the new fixed point position, if operation is in-place.
      */
-    void configure(const ITensor *input, ITensor *output, ConvertPolicy policy, uint32_t shift);
+    void configure(ITensor *input, ITensor *output, ConvertPolicy policy, uint32_t shift = 0);
 };
 }
 #endif /*__ARM_COMPUTE_NEDEPTHCONVERT_H__*/