COMPMID-406: Port CLActivationLayer to use QS8/QS16.

Change-Id: Ia4114984c38e1d2027ad97335b3c6c11f5754e23
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78727
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

9 files changed, 309 insertions, 95 deletions

diff --git a/arm_compute/core/CL/kernels/CLActivationLayerKernel.h b/arm_compute/core/CL/kernels/CLActivationLayerKernel.h
index df22574de8..a06f2fa0ae 100644
--- a/arm_compute/core/CL/kernels/CLActivationLayerKernel.h
+++ b/arm_compute/core/CL/kernels/CLActivationLayerKernel.h
@@ -51,8 +51,8 @@ public:
      * @note If the output tensor is a nullptr, the activation function will be performed in-place
      *
      * @param[in, out] input    Source tensor. In case of @p output tensor = nullptr, this tensor will store the result
-     *                          of the activation function. Data types supported: F16/F32.
-     * @param[out]     output   Destination tensor. Data type should match the input data type.
+     *                          of the activation function. Data types supported: QS8/QS16/F16/F32.
+     * @param[out]     output   Destination tensor. Data type supported: same as @p input
      * @param[in]      act_info Activation layer information.
      */
     void configure(ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info);
diff --git a/arm_compute/core/NEON/NEFixedPoint.inl b/arm_compute/core/NEON/NEFixedPoint.inl
index f62a338a61..4e862ba387 100644
--- a/arm_compute/core/NEON/NEFixedPoint.inl
+++ b/arm_compute/core/NEON/NEFixedPoint.inl
@@ -21,6 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
+#include <limits>
 
 namespace arm_compute
 {
@@ -1196,7 +1197,7 @@ inline qint8x16_t vqrecipq_qs8(qint8x16_t a, int fixed_point_position)
     const qint8x16_t shift_value = vqnegq_s8(vqsubq_s8(vdupq_n_s8(8), vqaddq_s8(vclzq_s8(a), vdupq_n_s8(fixed_point_position))));
     const qint8x16_t temp        = vqshlq_s8(a, shift_value);
 
-    qint8x16_t x = vqsubq_qs8(const_48_over_17, vmulq_qs8(temp, const_32_over_17, fixed_point_position));
+    qint8x16_t x = vqsubq_qs8(const_48_over_17, vqmulq_qs8(temp, const_32_over_17, fixed_point_position));
 
     // Set initial guess to one if x > 1
     uint8x16_t set_one = vcgtq_s8(x, const_one);
@@ -1234,7 +1235,8 @@ inline qint16x8_t vqrecipq_qs16(qint16x8_t a, int fixed_point_position)
     x = vqaddq_s16(x, vqmulq_qs16(x, vqsubq_s16(const_one, vqmulq_qs16(temp, x, fixed_point_position)), fixed_point_position));
     x = vqaddq_s16(x, vqmulq_qs16(x, vqsubq_s16(const_one, vqmulq_qs16(temp, x, fixed_point_position)), fixed_point_position));
 
-    return vqshlq_s16(x, shift_value);
+    // Saturate result in case of overflow
+    return vbslq_s16(vceqq_s16(a, vdupq_n_s16(0)), vdupq_n_s16(std::numeric_limits<int16_t>::max()), vqshlq_s16(x, shift_value));
 }
 
 inline qint8x8_t vdiv_qs8(qint8x8_t a, qint8x8_t b, int fixed_point_position)
diff --git a/arm_compute/runtime/CL/functions/CLActivationLayer.h b/arm_compute/runtime/CL/functions/CLActivationLayer.h
index 3028afb25b..a1aeb193d1 100644
--- a/arm_compute/runtime/CL/functions/CLActivationLayer.h
+++ b/arm_compute/runtime/CL/functions/CLActivationLayer.h
@@ -44,8 +44,8 @@ public:
      * @note If the output tensor is a nullptr, the activation function will be performed in-place
      *
      * @param[in, out] input    Source tensor. In case of @p output tensor = nullptr, this tensor will store the result
-     *                          of the activation function. Data types supported: F16/F32.
-     * @param[out]     output   Destination tensor. Data type should match the input data type.
+     *                          of the activation function. Data types supported: QS8/QS16/F16/F32.
+     * @param[out]     output   Destination tensor. Data type supported: same as @p input
      * @param[in]      act_info Activation layer parameters.
      */
     void configure(ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info);
diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl
index 721c43c017..5f812cf5b3 100644
--- a/src/core/CL/cl_kernels/activation_layer.cl
+++ b/src/core/CL/cl_kernels/activation_layer.cl
@@ -23,16 +23,99 @@
  */
 #include "helpers.h"
 
+#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+
+#if defined(FIXED_POINT_POSITION)
+#include "fixed_point.h"
+
+#define CONST_ONE (1 << FIXED_POINT_POSITION)
+#define ABS_OP(a) ABS_SAT_OP_EXPAND((a), DATA_TYPE, VEC_SIZE)
+#define ADD_OP(a, b) ADD_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE)
+#define SUB_OP(a, b) SUB_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE)
+#define MUL_OP(a, b) MUL_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
+#define MLA_OP(a, b, c) MLA_SAT_OP_EXPAND((a), (b), (c), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
+#define DIV_OP(a, b) DIV_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
+#define EXP_OP(a) EXP_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
+#define LOG_OP(a) LOG_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
+#define SQRT_OP(a) DIV_OP(CONST_ONE, INVSQRT_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION))
+#define TANH_OP(a) TANH_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
+
+#else /* FIXED_POINT_POSITION */
+
+#define CONST_ONE (1.f)
+#define ABS_OP(a) fabs((a))
+#define ADD_OP(a, b) ((a) + (b))
+#define SUB_OP(a, b) ((a) - (b))
+#define MUL_OP(a, b) ((a) * (b))
+#define MLA_OP(a, b, c) ((b) * (c) + (a))
+#define DIV_OP(a, b) ((a) / (b))
+#define EXP_OP(a) exp((a))
+#define LOG_OP(a) log((a))
+#define SQRT_OP(a) sqrt((a))
+#define TANH_OP(a) tanh((a))
+
+#endif /* FIXED_POINT_POSITION */
+
+// Logistic Activation
+inline TYPE logistic_op(TYPE x)
+{
+    return DIV_OP(CONST_ONE, ADD_OP(CONST_ONE, EXP_OP(-x)));
+}
+// Hyperbolic Tangent Activation
+inline TYPE tanh_op(TYPE x)
+{
+    return MUL_OP((TYPE)A_VAL, TANH_OP(MUL_OP((TYPE)B_VAL, x)));
+}
+// RELU Tangent Activation
+inline TYPE relu_op(TYPE x)
+{
+    return max(0, x);
+}
+// Bounded RELU Activation
+inline TYPE brelu_op(TYPE x)
+{
+    return min((TYPE)A_VAL, max(0, x));
+}
+// Soft RELU Activation
+inline TYPE srelu_op(TYPE x)
+{
+    return LOG_OP(ADD_OP(CONST_ONE, EXP_OP(x)));
+}
+// Absolute Activation
+inline TYPE abs_op(TYPE x)
+{
+    return ABS_OP(x);
+}
+// Square Activation
+inline TYPE square_op(TYPE x)
+{
+    return MUL_OP(x, x);
+}
+// Square-root Activation
+inline TYPE sqrt_op(TYPE x)
+{
+    return SQRT_OP(x);
+}
+// Linear Activation
+inline TYPE linear_op(TYPE x)
+{
+    return MLA_OP((TYPE)B_VAL, (TYPE)A_VAL, x);
+}
+
+#define ACTIVATION_OP2(op, x) op##_op(x)
+#define ACTIVATION_OP(op, x) ACTIVATION_OP2(op, x)
+
 /** This performs an activation function floating point inputs.
  *
  * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
- * @note Activation function should be given as a preprocessor argument using -DNAME. e.g. -DTANH
- * @note Distinction between floating point and integer is done using -DTYPE_FP and -DTYPE_INT preprocessor argument
- * @note A, B variables required by some activation functions are set using -DA= and -DB= respectively.
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH
+ * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
+ * @note In case of fixed point calculations the fixed point position is passed using -DFIXED_POINT_POSITION=position. e.g. -DFIXED_POINT_POSITION=3.
  *
- * @param[in]  input_ptr                            Pointer to the source image. Supported data types: F16, F32
+ * @param[in]  input_ptr                            Pointer to the source image. Supported data types: QS8/QS16/F16/F32
  * @param[in]  input_stride_x                       Stride of the source image in X dimension (in bytes)
  * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  input_stride_y                       Stride of the source image in Y dimension (in bytes)
@@ -40,7 +123,7 @@
  * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
  * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source image
- * @param[out] output_ptr                           Pointer to the destination image. Supported data types: F16, F32
+ * @param[out] output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
  * @param[in]  output_stride_x                      Stride of the destination image in X dimension (in bytes)
  * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  output_stride_y                      Stride of the destination image in Y dimension (in bytes)
@@ -66,34 +149,12 @@ __kernel void activation_layer(
 #endif /* IN_PLACE */
 
     // Load data
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    data = vload16(0, (__global DATA_TYPE *)input.ptr);
+    TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
 
     // Perform activation
-#ifdef LOGISTIC
-    data = 1 / (1 + exp(-data));
-#elif defined(TANH)
-    data            = (VEC_DATA_TYPE(DATA_TYPE, 16))A * tanh((VEC_DATA_TYPE(DATA_TYPE, 16))B * data);
-#elif defined(RELU)
-    data = max(0, data);
-#elif defined(BRELU)
-    data = min((VEC_DATA_TYPE(DATA_TYPE, 16))A, max(0, data));
-#elif defined(SRELU)
-    data = log(1 + exp(data));
-#elif defined(ABS)
-#ifdef TYPE_INT
-    data = abs(data);
-#else  /* TYPE_INT */
-    data = fabs(data);
-#endif /* TYPE_INT */
-#elif defined(SQUARE)
-    data = data * data;
-#elif defined(SQRT)
-    data = sqrt(data);
-#elif defined(LINEAR)
-    data = (VEC_DATA_TYPE(DATA_TYPE, 16))A * data + (VEC_DATA_TYPE(DATA_TYPE, 16))B;
-#endif /* switch TANH, RELU, BRELU, SRELU, ABS, SQUARE, SQRT, LINEAR */
+    data = ACTIVATION_OP(ACT, data);
 
     // Store result
-    vstore16(data, 0, (__global DATA_TYPE *)output.ptr);
+    VSTORE(VEC_SIZE)
+    (data, 0, (__global DATA_TYPE *)output.ptr);
 }
diff --git a/src/core/CL/cl_kernels/fixed_point.h b/src/core/CL/cl_kernels/fixed_point.h
index 5d340c4e95..bb534f5a51 100644
--- a/src/core/CL/cl_kernels/fixed_point.h
+++ b/src/core/CL/cl_kernels/fixed_point.h
@@ -99,6 +99,24 @@ TYPE_ALIAS(int, qs32)
 #define CONVERT_SAT_STR(x, type) CONVERT_SAT_STR2(x, type, type##_TYPE)
 #define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type)
 
+/** Computes saturating absolute value of fixed point vector.
+  *
+  * @param[in] type the actual data type.
+  *
+  * @return The result of the fixed point absolute value.
+  */
+#define ABSQ_SAT_IMPL(type)                  \
+    inline type abs_##type##_sat(type VopA)  \
+    {                                        \
+        return CONVERT_SAT(abs(VopA), type); \
+    }
+
+ABSQ_SAT_IMPL(qs8x16)
+ABSQ_SAT_IMPL(qs16x8)
+
+#define ABS_SAT_OP_EXPAND_STR(a, type, size) abs_##type##x##size##_sat((a))
+#define ABS_SAT_OP_EXPAND(a, type, size) ABS_SAT_OP_EXPAND_STR(a, type, size)
+
 /** Computes max of fixed point types.
   *
   * @param[in] type the actual data type.
@@ -280,6 +298,7 @@ MLALQ_SAT_IMPL(qs16x8, qs32x8)
     }
 
 DIVQ_SAT_IMPL(qs8, qs8x16, qs16x16)
+DIVQ_SAT_IMPL(qs16, qs16x8, qs32x8)
 DIVQ_SAT_IMPL(qs16, qs16x16, qs32x16)
 
 #define DIV_SAT_OP_EXPAND_STR(a, b, type, size, position) div_sat_##type##x##size((a), (b), (position))
@@ -287,34 +306,37 @@ DIVQ_SAT_IMPL(qs16, qs16x16, qs32x16)
 
 /** Saturate exponential of a fixed point vector
   *
+  * @note Implemented approach uses taylor polynomial to approximate the exponential function.
+  *
   * @param[in] stype the actual scalar data type.
   * @param[in] type  the actual data type.
   * @param[in] size  the number of the calculated elements.
   *
   * @return The result of the fixed point exponential. The result is saturated in case of overflow
   */
-#define EXPQ_IMPL(stype, type, size)                                                                                     \
-    inline type exp_sat_##type(type VopA, int fixed_point_position)                                                      \
-    {                                                                                                                    \
-        type const_one = (type)(1 << (fixed_point_position));                                                            \
-        type ln2       = (type)((((0x58B9 >> (14 - fixed_point_position))) + 1) >> 1);                                   \
-        type inv_ln2   = (type)((((0x38AA >> (14 - fixed_point_position)) + 1) >> 1)) | const_one;                       \
-        type A         = (type)(((0x7FBA >> (14 - fixed_point_position)) + 1) >> 1);                                     \
-        type B         = (type)(((0x3FE9 >> (14 - fixed_point_position)) + 1) >> 1);                                     \
-        type C         = (type)(((0x1693 >> (14 - fixed_point_position)) + 1) >> 1);                                     \
-        type D         = (type)(((0x0592 >> (14 - fixed_point_position)) + 1) >> 1);                                     \
-        type m         = MUL_SAT_OP_EXPAND(VopA, inv_ln2, stype, size, fixed_point_position);                            \
-        type dec_m     = m >> (type)fixed_point_position;                                                                \
-        type alpha     = MUL_SAT_OP_EXPAND(dec_m << (type)fixed_point_position, ln2, stype, size, fixed_point_position); \
-        alpha          = CONVERT(abs_diff(VopA, alpha), type);                                                           \
-        type sum       = add_sat(MUL_SAT_OP_EXPAND(alpha, D, stype, size, fixed_point_position), C);                     \
-        sum            = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), B);                   \
-        sum            = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), A);                   \
-        sum            = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), const_one);           \
-        return select(select(sum << dec_m, sum >> -dec_m, dec_m < (type)0), (type)stype##_MAX, clz(sum) <= dec_m);       \
+#define EXPQ_IMPL(stype, type, size)                                                                                                              \
+    inline type exp_sat_##type(type VopA, int fixed_point_position)                                                                               \
+    {                                                                                                                                             \
+        type const_one = (type)(1 << (fixed_point_position));                                                                                     \
+        type ln2       = (type)((((0x58B9 >> (14 - fixed_point_position))) + 1) >> 1);                                                            \
+        type inv_ln2   = (type)((((0x38AA >> (14 - fixed_point_position)) + 1) >> 1)) | const_one;                                                \
+        type A         = (type)(((0x7FBA >> (14 - fixed_point_position)) + 1) >> 1);                                                              \
+        type B         = (type)(((0x3FE9 >> (14 - fixed_point_position)) + 1) >> 1);                                                              \
+        type C         = (type)(((0x1693 >> (14 - fixed_point_position)) + 1) >> 1);                                                              \
+        type D         = (type)(((0x0592 >> (14 - fixed_point_position)) + 1) >> 1);                                                              \
+        type m         = MUL_SAT_OP_EXPAND(VopA, inv_ln2, stype, size, fixed_point_position);                                                     \
+        type dec_m     = m >> (type)fixed_point_position;                                                                                         \
+        type alpha     = MUL_SAT_OP_EXPAND(dec_m << (type)fixed_point_position, ln2, stype, size, fixed_point_position);                          \
+        alpha          = CONVERT(abs_diff(VopA, alpha), type);                                                                                    \
+        type sum       = add_sat(MUL_SAT_OP_EXPAND(alpha, D, stype, size, fixed_point_position), C);                                              \
+        sum            = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), B);                                            \
+        sum            = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), A);                                            \
+        sum            = add_sat(MUL_SAT_OP_EXPAND(alpha, sum, stype, size, fixed_point_position), const_one);                                    \
+        return select((type)stype##_MAX, select(sum << dec_m, sum >> -dec_m, dec_m < (type)0), clz(sum) > dec_m); /* Saturate result if needed */ \
     }
 
 EXPQ_IMPL(qs8, qs8x16, 16)
+EXPQ_IMPL(qs16, qs16x8, 8)
 EXPQ_IMPL(qs16, qs16x16, 16)
 
 #define EXP_OP_EXPAND_STR(a, type, size, position) exp_sat_##type##x##size((a), (position))
@@ -322,6 +344,8 @@ EXPQ_IMPL(qs16, qs16x16, 16)
 
 /** Saturate logarithm of a fixed point vector
   *
+  * @note Implemented approach uses taylor polynomial to approximate the logarithm function.
+  *
   * @param[in] stype the actual scalar data type.
   * @param[in] type  the actual data type.
   * @param[in] size  the number of the calculated elements.
@@ -332,11 +356,11 @@ EXPQ_IMPL(qs16, qs16x16, 16)
     inline type log_sat_##type(type VopA, int fixed_point_position)                                                                        \
     {                                                                                                                                      \
         type const_one = (type)(1 << (fixed_point_position));                                                                              \
-        type ln2       = (type)(0x58B9 >> (15 - fixed_point_position));                                                                    \
-        type A         = (type)(0x5C0F >> (14 - fixed_point_position));                                                                    \
-        type B         = -(type)(0x56AE >> (15 - fixed_point_position));                                                                   \
-        type C         = (type)(0x2933 >> (15 - fixed_point_position));                                                                    \
-        type D         = -(type)(0x0AA7 >> (15 - fixed_point_position));                                                                   \
+        type ln2       = (type)(0x58B9 >> (15 - fixed_point_position));  /* 1.4384189 */                                                   \
+        type A         = (type)(0x5C0F >> (14 - fixed_point_position));  /* 1.4384189 */                                                   \
+        type B         = -(type)(0x56AE >> (15 - fixed_point_position)); /* -0.6771900 */                                                  \
+        type C         = (type)(0x2933 >> (15 - fixed_point_position));  /* 0.3218538 */                                                   \
+        type D         = -(type)(0x0AA7 >> (15 - fixed_point_position)); /* -0.0832229 */                                                  \
         type inter_a   = select(VopA, DIV_SAT_OP_EXPAND(const_one, VopA, stype, size, fixed_point_position), VopA < const_one);            \
         type shift_val = (type)(15 - stype##_SHIFT) - clz(inter_a >> (type)fixed_point_position);                                          \
         inter_a        = inter_a >> shift_val;                                                                                             \
@@ -346,16 +370,19 @@ EXPQ_IMPL(qs16, qs16x16, 16)
         sum            = add_sat(MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position), A);                                   \
         sum            = MUL_SAT_OP_EXPAND(inter_a, sum, stype, size, fixed_point_position);                                               \
         sum            = MUL_SAT_OP_EXPAND(add_sat(sum, shift_val << (type)fixed_point_position), ln2, stype, size, fixed_point_position); \
-        return select(select(sum, -sum, VopA < const_one), (type)0, VopA < (type)0);                                                       \
+        return select(select(sum, -sum, VopA < const_one), (type)0, VopA < (type)0); /* Saturate result if needed */                       \
     }
 
 LOGQ_IMPL(qs8, qs8x16, 16)
+LOGQ_IMPL(qs16, qs16x8, 8)
 
 #define LOG_OP_EXPAND_STR(a, type, size, position) log_sat_##type##x##size((a), (position))
 #define LOG_OP_EXPAND(a, type, size, position) LOG_OP_EXPAND_STR(a, type, size, position)
 
 /** Saturate inverse square root of a fixed point vector
   *
+  * @note Implemented approach uses Newton's method to approximate the inverse square root function.
+  *
   * @param[in] stype the actual scalar data type.
   * @param[in] type  the actual data type.
   * @param[in] size  the number of the calculated elements.
@@ -367,20 +394,53 @@ LOGQ_IMPL(qs8, qs8x16, 16)
     {                                                                                                                                                                                                                                  \
         type const_three = (type)(3 << (fixed_point_position));                                                                                                                                                                        \
         type shift_value = (type)(16 - stype##_SHIFT) - (clz(VopA) + (type)fixed_point_position);                                                                                                                                      \
-        type temp        = select(VopA >> shift_value, VopA << (-shift_value), shift_value < (type)0);                                                                                                                                 \
+        type temp        = select(VopA >> shift_value, select((type)stype##_MAX, VopA << (-shift_value), clz(VopA) > (-shift_value)), shift_value < (type)0);                                                                          \
         type x           = temp;                                                                                                                                                                                                       \
         x                = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \
         x                = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \
         x                = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1; \
-        type res         = select(x >> (shift_value >> 1), x << ((-shift_value) >> 1), shift_value < (type)0);                                                                                                                         \
-        return select(res, stype##_MAX, res < (type)0);                                                                                                                                                                                \
+        if(sizeof((stype)(1)) > 1) /* Perform more iterations if datatype is QS16 */                                                                                                                                                   \
+        {                                                                                                                                                                                                                              \
+            x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1;            \
+            x = MUL_SAT_OP_EXPAND(x, sub_sat(const_three, MUL_SAT_OP_EXPAND(MUL_SAT_OP_EXPAND(x, x, stype, size, fixed_point_position), temp, stype, size, fixed_point_position)), stype, size, fixed_point_position) >> 1;            \
+        }                                                                                                                                                                                                                              \
+        type shift_value2 = select(shift_value >> 1, (-shift_value) >> 1, shift_value < (type)0);                                                                                                                                      \
+        return select(x >> shift_value2, select((type)stype##_MAX, x << shift_value2, clz(x) > shift_value2), shift_value < (type)0); /* Saturate result if needed */                                                                  \
     }
 
 INVSQRTQ_IMPL(qs8, qs8x16, 16)
+INVSQRTQ_IMPL(qs16, qs16x8, 8)
 
 #define INVSQRT_OP_EXPAND_STR(a, type, size, position) invsqrt_sat_##type##x##size((a), (position))
 #define INVSQRT_OP_EXPAND(a, type, size, position) INVSQRT_OP_EXPAND_STR(a, type, size, position)
 
+/** Saturate hyperbolic tangent of a fixed point vector
+  *
+  * tanh(x) = (e^2x - 1)/(e^2x + 1)
+  *
+  * @param[in] stype the actual scalar data type.
+  * @param[in] type  the actual data type.
+  * @param[in] size  the number of the calculated elements.
+  *
+  * @return The result of the fixed point hyperbolic tangent. The result is saturated in case of overflow
+  */
+#define TANHQ_IMPL(stype, type, size)                                                                                                             \
+    inline type tanh_sat_##type(type VopA, int fixed_point_position)                                                                              \
+    {                                                                                                                                             \
+        type const_one = (type)(1 << (fixed_point_position));                                                                                     \
+        type const_two = (type)(2 << (fixed_point_position));                                                                                     \
+        type exp2x     = EXP_OP_EXPAND(MUL_SAT_OP_EXPAND(const_two, VopA, stype, size, fixed_point_position), stype, size, fixed_point_position); \
+        type num       = SUB_SAT_OP_EXPAND(exp2x, const_one, stype, size);                                                                        \
+        type den       = ADD_SAT_OP_EXPAND(exp2x, const_one, stype, size);                                                                        \
+        return DIV_SAT_OP_EXPAND(num, den, stype, size, fixed_point_position);                                                                    \
+    }
+
+TANHQ_IMPL(qs8, qs8x16, 16)
+TANHQ_IMPL(qs16, qs16x8, 8)
+
+#define TANH_OP_EXPAND_STR(a, type, size, position) tanh_sat_##type##x##size((a), (position))
+#define TANH_OP_EXPAND(a, type, size, position) TANH_OP_EXPAND_STR(a, type, size, position)
+
 #define floatx16 float16
 #define float16_TYPE float16
 
diff --git a/src/core/CL/kernels/CLActivationLayerKernel.cpp b/src/core/CL/kernels/CLActivationLayerKernel.cpp
index fda69b0b94..18202c1c5b 100644
--- a/src/core/CL/kernels/CLActivationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLActivationLayerKernel.cpp
@@ -26,6 +26,7 @@
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/FixedPoint.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/IAccessWindow.h"
 #include "arm_compute/core/TensorInfo.h"
@@ -33,6 +34,10 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
 
+#include "support/ToolchainSupport.h"
+
+#include <cmath>
+
 using namespace arm_compute;
 
 CLActivationLayerKernel::CLActivationLayerKernel()
@@ -42,7 +47,7 @@ CLActivationLayerKernel::CLActivationLayerKernel()
 
 void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
 
     if(output != nullptr)
     {
@@ -54,20 +59,33 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act
         ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
     }
 
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+    const int          fixed_point_position              = input->info()->fixed_point_position();
+    float              a_const                           = act_info.a();
+    float              b_const                           = act_info.b();
+    if(is_data_type_fixed_point(input->info()->data_type()))
+    {
+        a_const = static_cast<int>(lround(a_const * (1 << fixed_point_position)));
+        b_const = static_cast<int>(lround(b_const * (1 << fixed_point_position)));
+    }
+
     // Set build options
     std::set<std::string> build_opts;
-    build_opts.insert(("-D" + string_from_activation_func(act_info.activation())));
-    build_opts.insert(("-D" + ((is_data_type_float(input->info()->data_type())) ? std::string("TYPE_FP") : std::string("TYPE_INT"))));
-    build_opts.insert(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
-    build_opts.insert(("-DA=" + support::cpp11::to_string(act_info.a())));
-    build_opts.insert(("-DB=" + support::cpp11::to_string(act_info.b())));
-    build_opts.insert(output == nullptr ? "-DIN_PLACE" : "");
+    build_opts.emplace(("-DACT=" + lower_string(string_from_activation_func(act_info.activation()))));
+    build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
+    build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
+    build_opts.emplace(("-DA_VAL=" + support::cpp11::to_string(a_const)));
+    build_opts.emplace(("-DB_VAL=" + support::cpp11::to_string(b_const)));
+    build_opts.emplace(output == nullptr ? "-DIN_PLACE" : "");
+    if(is_data_type_fixed_point(input->info()->data_type()))
+    {
+        build_opts.emplace(("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(fixed_point_position)));
+    }
 
     // Create kernel
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("activation_layer", build_opts));
 
     // Make sure _kernel is initialized before calling the parent's configure
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
 
     _input  = input;
     _output = output;
@@ -77,12 +95,9 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act
 
     if(output != nullptr)
     {
+        AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
         AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-
-        update_window_and_padding(win,
-                                  AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration),
-                                  output_access);
-
+        update_window_and_padding(win, input_access, output_access);
         output_access.set_valid_region(win, input->info()->valid_region());
     }
     else
diff --git a/tests/validation/CL/ActivationLayer.cpp b/tests/validation/CL/ActivationLayer.cpp
index 7286b93485..ac1da5c8b4 100644
--- a/tests/validation/CL/ActivationLayer.cpp
+++ b/tests/validation/CL/ActivationLayer.cpp
@@ -124,7 +124,14 @@ CLTensor compute_activation_layer(bool in_place, const TensorShape &shape, DataT
     {
         int min_bound = 0;
         int max_bound = 0;
-        std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int8_t>(act_info.activation(), fixed_point_position);
+        if(dt == DataType::QS8)
+        {
+            std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int8_t>(act_info.activation(), fixed_point_position);
+        }
+        else
+        {
+            std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int16_t>(act_info.activation(), fixed_point_position);
+        }
         std::uniform_int_distribution<> distribution(min_bound, max_bound);
         library->fill(CLAccessor(src), distribution, 0);
     }
@@ -148,7 +155,7 @@ BOOST_AUTO_TEST_SUITE(CL)
 BOOST_AUTO_TEST_SUITE(ActivationLayer)
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly"))
-BOOST_DATA_TEST_CASE(Configuration, boost::unit_test::data::make({ false, true }) * (SmallShapes() + LargeShapes()) * CNNFloatDataTypes(), in_place, shape, dt)
+BOOST_DATA_TEST_CASE(Configuration, boost::unit_test::data::make({ false, true }) * (SmallShapes() + LargeShapes()) * CNNDataTypes(), in_place, shape, dt)
 {
     // Set fixed point position data type allowed
     const int fixed_point_position = (arm_compute::is_data_type_fixed_point(dt)) ? 3 : 0;
@@ -182,7 +189,8 @@ BOOST_DATA_TEST_CASE(Configuration, boost::unit_test::data::make({ false, true }
     }
 
     // Validate padding
-    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    const int         step    = 16 / arm_compute::data_size_from_type(dt);
+    const PaddingSize padding = PaddingCalculator(shape.x(), step).required_padding();
     validate(src.info()->padding(), padding);
 
     if(!in_place)
@@ -193,10 +201,11 @@ BOOST_DATA_TEST_CASE(Configuration, boost::unit_test::data::make({ false, true }
 
 BOOST_AUTO_TEST_SUITE(Float)
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
-BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * SmallShapes() * CNNFloatDataTypes() * ActivationFunctions(), in_place, shape, dt, act_function)
+BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * SmallShapes() * CNNFloatDataTypes() * ActivationFunctions() * boost::unit_test::data::make({ 0.5f, 1.f }),
+                     in_place, shape, dt, act_function, alpha_beta)
 {
     // Create activation layer info
-    ActivationLayerInfo act_info(act_function, 1.f, 1.f);
+    ActivationLayerInfo act_info(act_function, alpha_beta, alpha_beta);
 
     // Compute function
     CLTensor dst = compute_activation_layer(in_place, shape, dt, act_info);
@@ -209,10 +218,11 @@ BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * S
 }
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
-BOOST_DATA_TEST_CASE(RunLarge, boost::unit_test::data::make({ false, true }) * LargeShapes() * CNNFloatDataTypes() * ActivationFunctions(), in_place, shape, dt, act_function)
+BOOST_DATA_TEST_CASE(RunLarge, boost::unit_test::data::make({ false, true }) * LargeShapes() * CNNFloatDataTypes() * ActivationFunctions() * boost::unit_test::data::make({ 0.5f, 1.f }),
+                     in_place, shape, dt, act_function, alpha_beta)
 {
     // Create activation layer info
-    ActivationLayerInfo act_info(act_function, 1.f, 1.f);
+    ActivationLayerInfo act_info(act_function, alpha_beta, alpha_beta);
 
     // Compute function
     CLTensor dst = compute_activation_layer(in_place, shape, dt, act_info);
@@ -225,6 +235,49 @@ BOOST_DATA_TEST_CASE(RunLarge, boost::unit_test::data::make({ false, true }) * L
 }
 BOOST_AUTO_TEST_SUITE_END()
 
+/** @note We test for fixed point precision [3,5] because [1,2] and [6,7] ranges
+ *        cause overflowing issues in most of the transcendentals functions.
+ */
+BOOST_AUTO_TEST_SUITE(Quantized)
+BOOST_AUTO_TEST_SUITE(QS8)
+BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
+BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * SmallShapes() * ActivationFunctions() * boost::unit_test::data::xrange(3, 6, 1) * boost::unit_test::data::make({ 0.5f, 1.f }),
+                     in_place, shape, act_function, fixed_point_position, alpha_beta)
+{
+    // Create activation layer info
+    ActivationLayerInfo act_info(act_function, alpha_beta, alpha_beta);
+
+    // Compute function
+    CLTensor dst = compute_activation_layer(in_place, shape, DataType::QS8, act_info, fixed_point_position);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_activation_layer(shape, DataType::QS8, act_info, fixed_point_position);
+
+    // Validate output
+    validate(CLAccessor(dst), ref_dst, activation_layer_tolerance(act_function, fixed_point_position));
+}
+BOOST_AUTO_TEST_SUITE_END()
+
+BOOST_AUTO_TEST_SUITE(QS16)
+BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
+BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * SmallShapes() * ActivationFunctions() * boost::unit_test::data::xrange(3, 14, 1) * boost::unit_test::data::make({ 0.5f, 1.f }),
+                     in_place, shape, act_function, fixed_point_position, alpha_beta)
+{
+    // Create activation layer info
+    ActivationLayerInfo act_info(act_function, alpha_beta, alpha_beta);
+
+    // Compute function
+    CLTensor dst = compute_activation_layer(in_place, shape, DataType::QS16, act_info, fixed_point_position);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_activation_layer(shape, DataType::QS16, act_info, fixed_point_position);
+
+    // Validate output
+    validate(CLAccessor(dst), ref_dst, activation_layer_tolerance(act_function, fixed_point_position));
+}
+BOOST_AUTO_TEST_SUITE_END()
+BOOST_AUTO_TEST_SUITE_END()
+
 BOOST_AUTO_TEST_SUITE_END()
 BOOST_AUTO_TEST_SUITE_END()
-#endif /* DOXYGEN_SKIP_THIS */
+#endif /* DOXYGEN_SKIP_THIS */
\ No newline at end of file
diff --git a/tests/validation/NEON/ActivationLayer.cpp b/tests/validation/NEON/ActivationLayer.cpp
index 71dfcdc4e2..2b24fd5175 100644
--- a/tests/validation/NEON/ActivationLayer.cpp
+++ b/tests/validation/NEON/ActivationLayer.cpp
@@ -53,12 +53,13 @@ namespace
 {
 /** Define tolerance of the activation layer
  *
+ * @param[in] dt                   The data type used.
  * @param[in] activation           The activation function used.
  * @param[in] fixed_point_position Number of bits for the fractional part..
  *
  * @return Tolerance depending on the activation function.
  */
-float activation_layer_tolerance(ActivationLayerInfo::ActivationFunction activation, int fixed_point_position = 0)
+float activation_layer_tolerance(DataType dt, ActivationLayerInfo::ActivationFunction activation, int fixed_point_position = 0)
 {
     switch(activation)
     {
@@ -66,7 +67,15 @@ float activation_layer_tolerance(ActivationLayerInfo::ActivationFunction activat
         case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
         case ActivationLayerInfo::ActivationFunction::SQRT:
         case ActivationLayerInfo::ActivationFunction::TANH:
-            return (fixed_point_position != 0) ? 5.f : 0.00001f;
+            switch(dt)
+            {
+                case DataType::QS8:
+                    return 5.f;
+                case DataType::QS16:
+                    return 11.f;
+                default:
+                    return 0.00001f;
+            }
             break;
         default:
             return 0.f;
@@ -124,7 +133,14 @@ Tensor compute_activation_layer(bool in_place, const TensorShape &shape, DataTyp
     {
         int min_bound = 0;
         int max_bound = 0;
-        std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int8_t>(act_info.activation(), fixed_point_position);
+        if(dt == DataType::QS8)
+        {
+            std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int8_t>(act_info.activation(), fixed_point_position);
+        }
+        else
+        {
+            std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int16_t>(act_info.activation(), fixed_point_position);
+        }
         std::uniform_int_distribution<> distribution(min_bound, max_bound);
         library->fill(NEAccessor(src), distribution, 0);
     }
@@ -206,7 +222,7 @@ BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * S
     RawTensor ref_dst = Reference::compute_reference_activation_layer(shape, dt, act_info);
 
     // Validate output
-    validate(NEAccessor(dst), ref_dst, activation_layer_tolerance(act_function));
+    validate(NEAccessor(dst), ref_dst, activation_layer_tolerance(dt, act_function));
 }
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
@@ -223,7 +239,7 @@ BOOST_DATA_TEST_CASE(RunLarge, boost::unit_test::data::make({ false, true }) * L
     RawTensor ref_dst = Reference::compute_reference_activation_layer(shape, dt, act_info);
 
     // Validate output
-    validate(NEAccessor(dst), ref_dst, activation_layer_tolerance(act_function));
+    validate(NEAccessor(dst), ref_dst, activation_layer_tolerance(dt, act_function));
 }
 BOOST_AUTO_TEST_SUITE_END()
 
@@ -246,13 +262,13 @@ BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * S
     RawTensor ref_dst = Reference::compute_reference_activation_layer(shape, DataType::QS8, act_info, fixed_point_position);
 
     // Validate output
-    validate(NEAccessor(dst), ref_dst, activation_layer_tolerance(act_function, fixed_point_position));
+    validate(NEAccessor(dst), ref_dst, activation_layer_tolerance(DataType::QS8, act_function, fixed_point_position));
 }
 BOOST_AUTO_TEST_SUITE_END()
 
 BOOST_AUTO_TEST_SUITE(QS16)
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
-BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * SmallShapes() * ActivationFunctions() * boost::unit_test::data::xrange(3, 6, 1) * boost::unit_test::data::make({ 0.5f, 1.f }),
+BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * SmallShapes() * ActivationFunctions() * boost::unit_test::data::xrange(3, 14, 1) * boost::unit_test::data::make({ 0.5f, 1.f }),
                      in_place, shape, act_function, fixed_point_position, alpha_beta)
 {
     // Create activation layer info
@@ -265,7 +281,7 @@ BOOST_DATA_TEST_CASE(RunSmall, boost::unit_test::data::make({ false, true }) * S
     RawTensor ref_dst = Reference::compute_reference_activation_layer(shape, DataType::QS16, act_info, fixed_point_position);
 
     // Validate output
-    validate(NEAccessor(dst), ref_dst, activation_layer_tolerance(act_function, fixed_point_position));
+    validate(NEAccessor(dst), ref_dst, activation_layer_tolerance(DataType::QS16, act_function, fixed_point_position));
 }
 BOOST_AUTO_TEST_SUITE_END()
 
diff --git a/tests/validation/Reference.cpp b/tests/validation/Reference.cpp
index 0a57fc0ea5..1b941870ba 100644
--- a/tests/validation/Reference.cpp
+++ b/tests/validation/Reference.cpp
@@ -459,7 +459,14 @@ RawTensor Reference::compute_reference_activation_layer(const TensorShape &shape
     {
         int min_bound = 0;
         int max_bound = 0;
-        std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int8_t>(act_info.activation(), fixed_point_position);
+        if(dt == DataType::QS8)
+        {
+            std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int8_t>(act_info.activation(), fixed_point_position);
+        }
+        else
+        {
+            std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<int16_t>(act_info.activation(), fixed_point_position);
+        }
         std::uniform_int_distribution<> distribution(min_bound, max_bound);
         library->fill(ref_src, distribution, 0);
     }