COMPMID-400 Add support for 16 bit fixed point arithmetic.

Change-Id: Iebfaef1b219d80d6362b7fd4b1357612b31e43cb
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/77749
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

1 files changed, 113 insertions, 2 deletions

diff --git a/arm_compute/core/FixedPoint.h b/arm_compute/core/FixedPoint.h
index 925b4949a3..da304c6329 100644
--- a/arm_compute/core/FixedPoint.h
+++ b/arm_compute/core/FixedPoint.h
@@ -33,12 +33,21 @@ using qint32_t = int32_t; /**< 32 bit fixed point scalar value */
 /** 8 bit fixed point scalar saturating shift left
  *
  * @param[in] a     First 8 bit fixed point input
- * @param[in] shift Shift amount
+ * @param[in] shift Shift amount (positive only values)
  *
  * @return The result of the 8 bit fixed point shift. The result is saturated in case of overflow
  */
 qint8_t sqshl_qs8(qint8_t a, int shift);
 
+/** 16 bit fixed point scalar saturating shift left
+ *
+ * @param[in] a     First 16 bit fixed point input
+ * @param[in] shift Shift amount (positive only values)
+ *
+ * @return The result of the 16 bit fixed point shift. The result is saturated in case of overflow
+ */
+qint16_t sqshl_qs16(qint16_t a, int shift);
+
 /** 8 bit fixed point scalar absolute value
  *
  * @param[in] a 8 bit fixed point input
@@ -47,6 +56,14 @@ qint8_t sqshl_qs8(qint8_t a, int shift);
  */
 qint8_t sabs_qs8(qint8_t a);
 
+/** 16 bit fixed point scalar absolute value
+ *
+ * @param[in] a 16 bit fixed point input
+ *
+ * @return The result of the 16 bit fixed point absolute value
+ */
+qint16_t sabs_qs16(qint16_t a);
+
 /** 8 bit fixed point scalar add
  *
  * @param[in] a First 8 bit fixed point input
@@ -56,6 +73,15 @@ qint8_t sabs_qs8(qint8_t a);
  */
 qint8_t sadd_qs8(qint8_t a, qint8_t b);
 
+/** 16 bit fixed point scalar add
+ *
+ * @param[in] a First 16 bit fixed point input
+ * @param[in] b Second 16 bit fixed point input
+ *
+ * @return The result of the 16 bit fixed point addition
+ */
+qint16_t sadd_qs16(qint16_t a, qint16_t b);
+
 /** 8 bit fixed point scalar saturating add
  *
  * @param[in] a First 8 bit fixed point input
@@ -83,6 +109,15 @@ qint16_t sqadd_qs16(qint16_t a, qint16_t b);
  */
 qint8_t ssub_qs8(qint8_t a, qint8_t b);
 
+/** 16 bit fixed point scalar subtraction
+ *
+ * @param[in] a First 16 bit fixed point input
+ * @param[in] b Second 16 bit fixed point input
+ *
+ * @return The result of the 16 bit fixed point subtraction
+ */
+qint16_t ssub_qs16(qint16_t a, qint16_t b);
+
 /** 8 bit fixed point scalar saturating subtraction
  *
  * @param[in] a First 8 bit fixed point input
@@ -92,6 +127,15 @@ qint8_t ssub_qs8(qint8_t a, qint8_t b);
  */
 qint8_t sqsub_qs8(qint8_t a, qint8_t b);
 
+/** 16 bit fixed point scalar saturating subtraction
+ *
+ * @param[in] a First 16 bit fixed point input
+ * @param[in] b Second 16 bit fixed point input
+ *
+ * @return The result of the 16 bit fixed point subtraction. The result is saturated in case of overflow
+ */
+qint16_t sqsub_qs16(qint16_t a, qint16_t b);
+
 /** 8 bit fixed point scalar multiply
  *
  * @param[in] a                    First 8 bit fixed point input
@@ -102,6 +146,16 @@ qint8_t sqsub_qs8(qint8_t a, qint8_t b);
  */
 qint8_t smul_qs8(qint8_t a, qint8_t b, int fixed_point_position);
 
+/** 16 bit fixed point scalar multiply
+ *
+ * @param[in] a                    First 16 bit fixed point input
+ * @param[in] b                    Second 16 bit fixed point input
+ * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number
+ *
+ * @return The result of the 16 bit fixed point multiplication.
+ */
+qint16_t smul_qs16(qint16_t a, qint16_t b, int fixed_point_position);
+
 /** 8 bit fixed point scalar saturating multiply
  *
  * @param[in] a                    First 8 bit fixed point input
@@ -112,6 +166,16 @@ qint8_t smul_qs8(qint8_t a, qint8_t b, int fixed_point_position);
  */
 qint8_t sqmul_qs8(qint8_t a, qint8_t b, int fixed_point_position);
 
+/** 16 bit fixed point scalar saturating multiply
+ *
+ * @param[in] a                    First 16 bit fixed point input
+ * @param[in] b                    Second 16 bit fixed point input
+ * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number
+ *
+ * @return The result of the 16 bit fixed point multiplication. The result is saturated in case of overflow
+ */
+qint16_t sqmul_qs16(qint16_t a, qint16_t b, int fixed_point_position);
+
 /** 8 bit fixed point scalar multiply long
  *
  * @param[in] a                    First 8 bit fixed point input
@@ -122,6 +186,16 @@ qint8_t sqmul_qs8(qint8_t a, qint8_t b, int fixed_point_position);
  */
 qint16_t sqmull_qs8(qint8_t a, qint8_t b, int fixed_point_position);
 
+/** 16 bit fixed point scalar multiply long
+ *
+ * @param[in] a                    First 16 bit fixed point input
+ * @param[in] b                    Second 16 bit fixed point input
+ * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number
+ *
+ * @return The result of the 16 bit fixed point multiplication long. The result is saturated in case of overflow
+ */
+qint32_t sqmull_qs16(qint16_t a, qint16_t b, int fixed_point_position);
+
 /** 16 bit fixed point scalar saturating multiply
 *
 * @param[in] a                    First 16 bit fixed point input
@@ -141,6 +215,15 @@ qint16_t sqmul_qs16(qint16_t a, qint16_t b, int fixed_point_position);
 */
 qint8_t sinvsqrt_qs8(qint8_t a, int fixed_point_position);
 
+/** 16 bit fixed point scalar inverse square root
+*
+* @param[in] a                    16 bit fixed point input
+* @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number
+*
+* @return The result of the 16 bit fixed point inverse square root.
+*/
+qint16_t sinvsqrt_qs16(qint16_t a, int fixed_point_position);
+
 /** 8 bit fixed point scalar division
 *
 * @param[in] a                    First 8 bit fixed point input
@@ -151,6 +234,16 @@ qint8_t sinvsqrt_qs8(qint8_t a, int fixed_point_position);
 */
 qint8_t sdiv_qs8(qint8_t a, qint8_t b, int fixed_point_position);
 
+/** 16 bit fixed point scalar division
+*
+* @param[in] a                    First 16 bit fixed point input
+* @param[in] b                    Second 16 bit fixed point input
+* @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number
+*
+* @return The result of the 16 bit fixed point division.
+*/
+qint16_t sdiv_qs16(qint16_t a, qint16_t b, int fixed_point_position);
+
 /** 8 bit fixed point scalar exponential
 *
 * @param[in] a                    8 bit fixed point input
@@ -160,6 +253,15 @@ qint8_t sdiv_qs8(qint8_t a, qint8_t b, int fixed_point_position);
 */
 qint8_t sexp_qs8(qint8_t a, int fixed_point_position);
 
+/** 16 bit fixed point scalar exponential
+*
+* @param[in] a                    16 bit fixed point input
+* @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number
+*
+* @return The result of the 16 bit fixed point exponential.
+*/
+qint16_t sexp_qs16(qint16_t a, int fixed_point_position);
+
 /** 8 bit fixed point scalar logarithm
 *
 * @param[in] a                    8 bit fixed point input
@@ -169,6 +271,15 @@ qint8_t sexp_qs8(qint8_t a, int fixed_point_position);
 */
 qint8_t slog_qs8(qint8_t a, int fixed_point_position);
 
+/** 16 bit fixed point scalar logarithm
+*
+* @param[in] a                    16 bit fixed point input
+* @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number
+*
+* @return The result of the 16 bit fixed point logarithm.
+*/
+qint16_t slog_qs16(qint16_t a, int fixed_point_position);
+
 /** Convert an 8 bit fixed point to float
  *
  * @param[in] a                    Input to convert
@@ -203,7 +314,7 @@ float scvt_f32_qs16(qint16_t a, int fixed_point_position);
  *
  * @return The result of the conversion float -> 16 bit fixed point
  */
-qint8_t scvt_qs16_f32(float a, int fixed_point_position);
+qint16_t scvt_qs16_f32(float a, int fixed_point_position);
 
 /** Scalar saturating move and narrow.
  *