COMPMID-2317: Add new QASYMM16 data type

Change-Id: I9fbbba633f10d542474a08b1898150b9799b7ae5
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1805
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

10 files changed, 96 insertions, 15 deletions

diff --git a/arm_compute/core/PixelValue.h b/arm_compute/core/PixelValue.h
index 4a07fa6c5d..e237498920 100644
--- a/arm_compute/core/PixelValue.h
+++ b/arm_compute/core/PixelValue.h
@@ -43,7 +43,7 @@ public:
      *
      * @param[in] v        int value.
      * @param[in] datatype DataType that @p v have to be stored
-     * @param[in] qinfo    (Optional) QuantizationInfo to apply in case of QASYMM8 datatype to @p v
+     * @param[in] qinfo    (Optional) QuantizationInfo to apply in case of quantized data types to @p v
      */
     PixelValue(uint64_t v, DataType datatype, QuantizationInfo qinfo = QuantizationInfo())
         : PixelValue()
@@ -68,6 +68,9 @@ public:
             case DataType::S16:
                 value.s16 = static_cast<int16_t>(v);
                 break;
+            case DataType::QASYMM16:
+                value.u16 = quantize_qasymm16(static_cast<uint16_t>(v), qinfo);
+                break;
             case DataType::QSYMM16:
                 value.s16 = quantize_qsymm16(static_cast<int16_t>(v), qinfo);
                 break;
diff --git a/arm_compute/core/QuantizationInfo.h b/arm_compute/core/QuantizationInfo.h
index 1517d48381..5e6e5b349c 100644
--- a/arm_compute/core/QuantizationInfo.h
+++ b/arm_compute/core/QuantizationInfo.h
@@ -32,8 +32,9 @@
 
 namespace arm_compute
 {
-using qasymm8_t = uint8_t; /**< 8 bit quantized asymmetric scalar value */
-using qsymm16_t = int16_t; /**< 16 bit quantized symmetric scalar value */
+using qasymm8_t  = uint8_t;  /**< 8 bit quantized asymmetric scalar value */
+using qsymm16_t  = int16_t;  /**< 16 bit quantized symmetric scalar value */
+using qasymm16_t = uint16_t; /**< 16 bit quantized asymmetric scalar value */
 
 /** Quantization info when assuming per layer quantization */
 struct UniformQuantizationInfo
@@ -204,7 +205,7 @@ inline bool operator!=(const UniformQuantizationInfo &lhs, const UniformQuantiza
     return !(operator==(lhs, rhs));
 }
 
-/** Quantize a value given a asymmetric quantization scheme
+/** Quantize a value given a 8-bit asymmetric quantization scheme
  *
  * @param[in] value           Value to quantize
  * @param[in] qinfo           Quantization information to use for quantizing
@@ -219,7 +220,7 @@ inline uint8_t quantize_qasymm8(float value, const UniformQuantizationInfo &qinf
     return quantized;
 }
 
-/** Quantize a value given a asymmetric quantization scheme
+/** Quantize a value given a 8-bit asymmetric quantization scheme
  *
  * @param[in] value           Value to quantize
  * @param[in] qinfo           Quantization information to use for quantizing
@@ -235,7 +236,7 @@ inline uint8_t quantize_qasymm8(float value, const QuantizationInfo &qinfo, Roun
     return quantized;
 }
 
-/** Quantize a value given a symmetric quantization scheme
+/** Quantize a value given a 8-bit symmetric quantization scheme
  *
  * @param[in] value Value to quantize
  * @param[in] qinfo Quantization information to use for quantizing
@@ -249,7 +250,7 @@ inline int8_t quantize_qsymm8(float value, const QuantizationInfo &qinfo)
     return quantized;
 }
 
-/** Dequantize a value given a asymmetric quantization scheme
+/** Dequantize a value given a 8-bit asymmetric quantization scheme
  *
  * @param[in] value Value to dequantize
  * @param[in] qinfo Quantization information to use for dequantizing
@@ -261,7 +262,7 @@ inline float dequantize_qasymm8(uint8_t value, const UniformQuantizationInfo &qi
     return (static_cast<int>(value) - qinfo.offset) * qinfo.scale;
 }
 
-/** Dequantize a value given a asymmetric quantization scheme
+/** Dequantize a value given a 8-bit asymmetric quantization scheme
  *
  * @param[in] value Value to dequantize
  * @param[in] qinfo Quantization information to use for dequantizing
@@ -274,7 +275,7 @@ inline float dequantize_qasymm8(uint8_t value, const QuantizationInfo &qinfo)
     return (static_cast<int>(value) - uqinfo.offset) * uqinfo.scale;
 }
 
-/** Dequantize a value given an asymmetric quantization scheme
+/** Dequantize a value given an 8-bit asymmetric quantization scheme
  *
  * @param[in] value  Value to dequantize
  * @param[in] scale  Scale to use for dequantization
@@ -287,7 +288,7 @@ inline float dequantize(uint8_t value, float scale, int32_t offset)
     return (static_cast<int>(value) - offset) * scale;
 }
 
-/** Dequantize a value given a symmetric quantization scheme
+/** Dequantize a value given a 8-bit symmetric quantization scheme
  *
  * @param[in] value Value to dequantize
  * @param[in] qinfo Quantization information to use for dequantizing
@@ -299,7 +300,7 @@ inline float dequantize_qsymm8(int8_t value, const UniformQuantizationInfo &qinf
     return value * qinfo.scale;
 }
 
-/** Dequantize a value given a symmetric quantization scheme
+/** Dequantize a value given a 8-bit symmetric quantization scheme
  *
  * @param[in] value Value to dequantize
  * @param[in] scale Scale to use for dequantization
@@ -311,7 +312,7 @@ inline float dequantize(int8_t value, float scale)
     return value * scale;
 }
 
-/** Dequantize a value given a symmetric quantization scheme
+/** Dequantize a value given a 16-bit symmetric quantization scheme
  *
  * @param[in] value Value to dequantize
  * @param[in] scale Scale to use for dequantization
@@ -323,6 +324,19 @@ inline float dequantize(int16_t value, float scale)
     return value * scale;
 }
 
+/** Dequantize a value given a 16-bit asymmetric quantization scheme
+ *
+ * @param[in] value  Value to dequantize
+ * @param[in] scale  Scale to use for dequantization
+ * @param[in] offset Zero-offset to use for dequantization
+ *
+ * @return Dequantized value
+ */
+inline float dequantize(uint16_t value, float scale, int32_t offset)
+{
+    return (static_cast<int>(value) - offset) * scale;
+}
+
 /** Quantize a value given a 16-bit symmetric quantization scheme
  *
  * @param[in] value           Value to quantize
@@ -373,5 +387,56 @@ inline float dequantize_qsymm16(int16_t value, const QuantizationInfo &qinfo)
 {
     return dequantize_qsymm16(value, qinfo.uniform());
 }
+
+/** Quantize a value given a 16-bit asymmetric 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 uint16_t quantize_qasymm16(float value, const UniformQuantizationInfo &qinfo, RoundingPolicy rounding_policy = RoundingPolicy::TO_NEAREST_UP)
+{
+    int quantized = arm_compute::round(value / qinfo.scale, rounding_policy) + qinfo.offset;
+    quantized     = arm_compute::utility::clamp<int, uint16_t>(quantized);
+    return quantized;
+}
+
+/** Dequantize a value given a 16-bit asymmetric quantization scheme
+ *
+ * @param[in] value Value to dequantize
+ * @param[in] qinfo Quantization information to use for dequantizing
+ *
+ * @return Dequantized value
+ */
+inline float dequantize_qasymm16(uint16_t value, const UniformQuantizationInfo &qinfo)
+{
+    return (static_cast<int>(value) - qinfo.offset) * qinfo.scale;
+}
+
+/** Quantize a value given a 16-bit asymmetric quantization scheme
+ *
+ * @param[in] value Value to quantize
+ * @param[in] qinfo Quantization information to use for quantizing
+ *
+ * @return Quantized value
+ */
+inline uint16_t quantize_qasymm16(float value, const QuantizationInfo &qinfo)
+{
+    return quantize_qasymm16(value, qinfo.uniform());
+}
+
+/** Dequantize a value given a 16-bit asymmetric quantization scheme
+ *
+ * @param[in] value Value to dequantize
+ * @param[in] qinfo Quantization information to use for dequantizing
+ *
+ * @return Dequantized value
+ */
+inline float dequantize_qasymm16(uint16_t value, const QuantizationInfo &qinfo)
+{
+    return dequantize_qasymm16(value, qinfo.uniform());
+}
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_QUANTIZATION_INFO_H__ */
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 59e39958b6..7f60638d05 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -83,6 +83,7 @@ enum class DataType
     U16,                 /**< unsigned 16-bit number */
     S16,                 /**< signed 16-bit number */
     QSYMM16,             /**< quantized, symmetric fixed-point 16-bit number */
+    QASYMM16,            /**< quantized, asymmetric 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 eb4cf05ae9..6ffab4b9ea 100644
--- a/arm_compute/core/Utils.h
+++ b/arm_compute/core/Utils.h
@@ -120,6 +120,7 @@ inline size_t data_size_from_type(DataType data_type)
         case DataType::U16:
         case DataType::S16:
         case DataType::QSYMM16:
+        case DataType::QASYMM16:
         case DataType::F16:
             return 2;
         case DataType::F32:
@@ -195,6 +196,7 @@ inline size_t element_size_from_data_type(DataType dt)
         case DataType::U16:
         case DataType::S16:
         case DataType::QSYMM16:
+        case DataType::QASYMM16:
         case DataType::F16:
             return 2;
         case DataType::U32:
@@ -532,7 +534,9 @@ inline DataType get_promoted_data_type(DataType dt)
         case DataType::QSYMM8:
         case DataType::QASYMM8:
         case DataType::QSYMM8_PER_CHANNEL:
+        case DataType::QASYMM8_PER_CHANNEL:
         case DataType::QSYMM16:
+        case DataType::QASYMM16:
         case DataType::F16:
         case DataType::U32:
         case DataType::S32:
@@ -1017,6 +1021,7 @@ inline bool is_data_type_quantized(DataType dt)
         case DataType::QSYMM8_PER_CHANNEL:
         case DataType::QASYMM8_PER_CHANNEL:
         case DataType::QSYMM16:
+        case DataType::QASYMM16:
             return true;
         default:
             return false;
@@ -1035,6 +1040,7 @@ inline bool is_data_type_quantized_asymmetric(DataType dt)
     {
         case DataType::QASYMM8:
         case DataType::QASYMM8_PER_CHANNEL:
+        case DataType::QASYMM16:
             return true;
         default:
             return false;
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index d5d9d10e6b..122373f5c6 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -163,6 +163,7 @@ const std::string &arm_compute::string_from_data_type(DataType dt)
         { DataType::QASYMM8_PER_CHANNEL, "QASYMM8_PER_CHANNEL" },
         { DataType::QASYMM8, "QASYMM8" },
         { DataType::QSYMM16, "QSYMM16" },
+        { DataType::QASYMM16, "QASYMM16" },
     };
 
     return dt_map[dt];
diff --git a/src/runtime/CL/functions/CLLSTMLayerQuantized.cpp b/src/runtime/CL/functions/CLLSTMLayerQuantized.cpp
index e0006a77d0..da19cc1cc3 100644
--- a/src/runtime/CL/functions/CLLSTMLayerQuantized.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayerQuantized.cpp
@@ -345,7 +345,7 @@ void CLLSTMLayerQuantized::run()
     _tanh_output_state.run();
     _mul_output_state_tmp_output_gate.run();
 
-    // Requantize output state from QSYMM16 to QASYMM16
+    // Requantize output state from QSYMM16 to QASYMM8
     _dequantize.run();
     _quantize.run();
 }
@@ -394,4 +394,4 @@ void CLLSTMLayerQuantized::prepare()
     }
 }
 
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
index 6cfa9887ff..96c776141a 100644
--- a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
+++ b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp
@@ -324,7 +324,7 @@ void NELSTMLayerQuantized::run()
     _tanh_output_state.run();
     _mul3.run();
 
-    // Requantize output state from QSYMM16 to QASYMM16
+    // Requantize output state from QSYMM16 to QASYMM8
     _dequantize.run();
     _quantize.run();
 }
diff --git a/tests/Utils.h b/tests/Utils.h
index 81bc2663de..f88b01dc40 100644
--- a/tests/Utils.h
+++ b/tests/Utils.h
@@ -361,6 +361,7 @@ void store_value_with_data_type(void *ptr, T value, DataType data_type)
             *reinterpret_cast<int8_t *>(ptr) = value;
             break;
         case DataType::U16:
+        case DataType::QASYMM16:
             *reinterpret_cast<uint16_t *>(ptr) = value;
             break;
         case DataType::S16:
diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h
index 904115360a..41bfd1d6c8 100644
--- a/utils/TypePrinter.h
+++ b/utils/TypePrinter.h
@@ -646,6 +646,9 @@ inline ::std::ostream &operator<<(::std::ostream &os, const DataType &data_type)
         case DataType::QSYMM16:
             os << "QSYMM16";
             break;
+        case DataType::QASYMM16:
+            os << "QASYMM16";
+            break;
         case DataType::U32:
             os << "U32";
             break;
diff --git a/utils/Utils.h b/utils/Utils.h
index 8605f4e3e1..ec08896257 100644
--- a/utils/Utils.h
+++ b/utils/Utils.h
@@ -173,6 +173,7 @@ inline std::string get_typestring(DataType data_type)
         case DataType::QSYMM8_PER_CHANNEL:
             return no_endianness + "i" + support::cpp11::to_string(sizeof(int8_t));
         case DataType::U16:
+        case DataType::QASYMM16:
             return endianness + "u" + support::cpp11::to_string(sizeof(uint16_t));
         case DataType::S16:
         case DataType::QSYMM16: