COMPMID-2408: Add QSYMM16 support for ElementwiseAddition for NEON

Change-Id: I22991e9369ffba9b51a94522ff4977933e887b94
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1352
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>

5 files changed, 71 insertions, 9 deletions

diff --git a/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h b/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
index 872c3a5b6b..958c02d516 100644
--- a/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
+++ b/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
@@ -64,18 +64,19 @@ public:
      *   - (F16,F16)         -> F16
      *   - (F32,F32)         -> F32
      *   - (QASYMM8,QASYMM8) -> QASYMM8
+     *   - (QSYMM16,QSYMM16) -> QSYMM16
      *
-     * @param[in]  input1 An input tensor. Data types supported: U8/QASYMM8/S16/F16/F32
-     * @param[in]  input2 An input tensor. Data types supported: U8/QASYMM8/S16/F16/F32
-     * @param[out] output The output tensor. Data types supported: U8/QASYMM8/S16/F16/F32.
+     * @param[in]  input1 An input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32
+     * @param[in]  input2 An input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32
+     * @param[out] output The output tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32.
      * @param[in]  policy Overflow policy.
      */
     void configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticAdditionKernel
      *
-     * @param[in] input1 An input tensor. Data types supported: U8/QASYMM8/S16/F16/F32
-     * @param[in] input2 An input tensor. Data types supported: U8/QASYMM8/S16/F16/F32
-     * @param[in] output The output tensor. Data types supported: U8/QASYMM8/S16/F16/F32.
+     * @param[in] input1 An input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32
+     * @param[in] input2 An input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32
+     * @param[in] output The output tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32.
      * @param[in] policy Overflow policy.
      *
      * @return a status
@@ -88,9 +89,9 @@ public:
 private:
     /** Common signature for all the specialised add functions
      *
-     * @param[in]  input1 An input tensor. Data types supported: U8/QASYMM8/S16/F16/F32
-     * @param[in]  input2 An input tensor. Data types supported: U8/QASYMM8/S16/F16/F32
-     * @param[out] output The output tensor. Data types supported: U8/QASYMM8/S16/F16/F32.
+     * @param[in]  input1 An input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32
+     * @param[in]  input2 An input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32
+     * @param[out] output The output tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32.
      * @param[in]  policy Overflow policy.
      * @param[in]  window Region on which to execute the kernel.
      */
diff --git a/arm_compute/core/PixelValue.h b/arm_compute/core/PixelValue.h
index 4bdcad61a2..4a07fa6c5d 100644
--- a/arm_compute/core/PixelValue.h
+++ b/arm_compute/core/PixelValue.h
@@ -68,6 +68,9 @@ public:
             case DataType::S16:
                 value.s16 = static_cast<int16_t>(v);
                 break;
+            case DataType::QSYMM16:
+                value.s16 = quantize_qsymm16(static_cast<int16_t>(v), qinfo);
+                break;
             case DataType::U32:
                 value.u32 = static_cast<uint32_t>(v);
                 break;
diff --git a/arm_compute/core/QuantizationInfo.h b/arm_compute/core/QuantizationInfo.h
index 94f7e76c3e..06c9b61154 100644
--- a/arm_compute/core/QuantizationInfo.h
+++ b/arm_compute/core/QuantizationInfo.h
@@ -25,6 +25,7 @@
 #define __ARM_COMPUTE_QUANTIZATION_INFO_H__
 
 #include "arm_compute/core/Rounding.h"
+#include "utils/misc/Utility.h"
 
 #include <cstddef>
 #include <vector>
@@ -255,5 +256,57 @@ inline float dequantize_qsymm8(int8_t value, const QuantizationInfo &qinfo)
 {
     return value * qinfo.uniform().scale;
 }
+
+/** Quantize a value given a 16-bit symmetric quantization scheme
+ *
+ * @param[in] value           Value to quantize
+ * @param[in] qinfo           Quantization information to use for quantizing
+ * @param[in] rounding_policy (Optional) Rounding policy to use. Default: nearest up
+ *
+ * @return Quantized value
+ */
+inline int16_t quantize_qsymm16(float value, const UniformQuantizationInfo &qinfo, RoundingPolicy rounding_policy = RoundingPolicy::TO_NEAREST_UP)
+{
+    int quantized = arm_compute::round(value / qinfo.scale, rounding_policy);
+    quantized     = arm_compute::utility::clamp<int, int16_t>(quantized);
+    return quantized;
+}
+
+/** Dequantize a value given a 16-bit symmetric quantization scheme
+ *
+ * @param[in] value Value to dequantize
+ * @param[in] qinfo Quantization information to use for dequantizing
+ *
+ * @return Dequantized value
+ */
+inline float dequantize_qsymm16(int16_t value, const UniformQuantizationInfo &qinfo)
+{
+    return value * qinfo.scale;
+}
+
+/** Quantize a value given a 16-bit symmetric quantization scheme
+ *
+ * @param[in] value Value to quantize
+ * @param[in] qinfo Quantization information to use for quantizing
+ *
+ * @return Quantized value
+ */
+inline int16_t quantize_qsymm16(float value, const QuantizationInfo &qinfo)
+{
+    return quantize_qsymm16(value, qinfo.uniform());
+}
+
+/** Dequantize a value given a 16-bit symmetric quantization scheme
+ *
+ * @param[in] value Value to dequantize
+ * @param[in] qinfo Quantization information to use for dequantizing
+ *
+ * @return Dequantized value
+ */
+inline float dequantize_qsymm16(int16_t value, const QuantizationInfo &qinfo)
+{
+    return dequantize_qsymm16(value, qinfo.uniform());
+}
+
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_QUANTIZATION_INFO_H__ */
\ No newline at end of file
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 1a49624113..ad679d6786 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -80,6 +80,7 @@ enum class DataType
     QSYMM8_PER_CHANNEL, /**< quantized, symmetric per channel fixed-point 8-bit number */
     U16,                /**< unsigned 16-bit number */
     S16,                /**< signed 16-bit number */
+    QSYMM16,            /**< quantized, symmetric fixed-point 16-bit number */
     U32,                /**< unsigned 32-bit number */
     S32,                /**< signed 32-bit number */
     U64,                /**< unsigned 64-bit number */
diff --git a/arm_compute/core/Utils.h b/arm_compute/core/Utils.h
index 8630eeee23..b711451453 100644
--- a/arm_compute/core/Utils.h
+++ b/arm_compute/core/Utils.h
@@ -117,6 +117,7 @@ inline size_t data_size_from_type(DataType data_type)
             return 1;
         case DataType::U16:
         case DataType::S16:
+        case DataType::QSYMM16:
         case DataType::F16:
             return 2;
         case DataType::F32:
@@ -191,6 +192,7 @@ inline size_t element_size_from_data_type(DataType dt)
             return 1;
         case DataType::U16:
         case DataType::S16:
+        case DataType::QSYMM16:
         case DataType::F16:
             return 2;
         case DataType::U32:
@@ -528,6 +530,7 @@ inline DataType get_promoted_data_type(DataType dt)
         case DataType::QSYMM8:
         case DataType::QASYMM8:
         case DataType::QSYMM8_PER_CHANNEL:
+        case DataType::QSYMM16:
         case DataType::F16:
         case DataType::U32:
         case DataType::S32:
@@ -1008,6 +1011,7 @@ inline bool is_data_type_quantized(DataType dt)
         case DataType::QSYMM8:
         case DataType::QASYMM8:
         case DataType::QSYMM8_PER_CHANNEL:
+        case DataType::QSYMM16:
             return true;
         default:
             return false;