COMPMID-421: Added FP16 support in ActivationLayer.

Change-Id: I7ba573b19d56e3c87996edb5218a00e5bfca451e
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79755
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

3 files changed, 53 insertions, 0 deletions

diff --git a/arm_compute/core/NEON/NEMath.h b/arm_compute/core/NEON/NEMath.h
index 8dd9d609e7..b467a600d6 100644
--- a/arm_compute/core/NEON/NEMath.h
+++ b/arm_compute/core/NEON/NEMath.h
@@ -93,6 +93,24 @@ float32x4_t vtanhq_f32(float32x4_t val);
 float32x4_t vpowq_f32(float32x4_t val, float32x4_t n);
 
 #ifdef ARM_COMPUTE_ENABLE_FP16
+/** Calculate hyperbolic tangent.
+ *
+ * tanh(x) = (e^2x - 1)/(e^2x + 1)
+ *
+ * @note We clamp x to [-5,5] to avoid overflowing issues.
+ *
+ * @param[in] val Input vector value in F32 format.
+ *
+ * @return The calculated Hyperbolic Tangent.
+ */
+float16x8_t vtanhq_f16(float16x8_t val);
+/** Calculate inverse square root.
+ *
+ * @param[in] x Input value.
+ *
+ * @return The calculated inverse square root.
+ */
+float16x8_t vinvsqrtq_f16(float16x8_t x);
 /** Calculate exponential
  *
  * @param[in] x Input vector value in F16 format.
diff --git a/arm_compute/core/NEON/NEMath.inl b/arm_compute/core/NEON/NEMath.inl
index c73c54501f..1d90029147 100644
--- a/arm_compute/core/NEON/NEMath.inl
+++ b/arm_compute/core/NEON/NEMath.inl
@@ -172,6 +172,14 @@ const std::array<float16x8_t, 8> log_tab_f16 =
         vdupq_n_f16(0.0141278216615f),
     }
 };
+inline float16x8_t vinvsqrtq_f16(float16x8_t x)
+{
+    float16x8_t sqrt_reciprocal = vrsqrteq_f16(x);
+    sqrt_reciprocal             = vmulq_f16(vrsqrtsq_f16(vmulq_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
+    sqrt_reciprocal             = vmulq_f16(vrsqrtsq_f16(vmulq_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
+
+    return sqrt_reciprocal;
+}
 
 inline float16x8_t vinvq_f16(float16x8_t x)
 {
@@ -181,6 +189,21 @@ inline float16x8_t vinvq_f16(float16x8_t x)
     return recip;
 }
 
+inline float16x8_t vtanhq_f16(float16x8_t val)
+{
+    const float16x8_t CONST_1        = vdupq_n_f16(1.f);
+    const float16x8_t CONST_2        = vdupq_n_f16(2.f);
+    const float16x8_t CONST_MIN_TANH = vdupq_n_f16(-10.f);
+    const float16x8_t CONST_MAX_TANH = vdupq_n_f16(10.f);
+
+    const float16x8_t x     = vminq_f16(vmaxq_f16(val, CONST_MIN_TANH), CONST_MAX_TANH);
+    const float16x8_t exp2x = vexpq_f16(vmulq_f16(CONST_2, x));
+    const float16x8_t num   = vsubq_f16(exp2x, CONST_1);
+    const float16x8_t den   = vaddq_f16(exp2x, CONST_1);
+    const float16x8_t tanh  = vmulq_f16(num, vinvq_f16(den));
+    return tanh;
+}
+
 inline float16x8_t vtaylor_polyq_f16(float16x8_t x, const std::array<float16x8_t, 8> &coeffs)
 {
     const float16x8_t A   = vaddq_f16(coeffs[0], vmulq_f16(coeffs[4], x));
diff --git a/arm_compute/core/NEON/kernels/NEActivationLayerKernel.h b/arm_compute/core/NEON/kernels/NEActivationLayerKernel.h
index e995f1e5e0..2c88debfb4 100644
--- a/arm_compute/core/NEON/kernels/NEActivationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEActivationLayerKernel.h
@@ -27,6 +27,10 @@
 #include "arm_compute/core/FixedPoint.h"
 #include "arm_compute/core/NEON/INEKernel.h"
 
+#ifdef ARM_COMPUTE_ENABLE_FP16
+#include <arm_fp16.h>
+#endif /* ARM_COMPUTE_ENABLE_FP16 */
+
 namespace arm_compute
 {
 class ITensor;
@@ -72,6 +76,14 @@ private:
      */
     template <ActivationLayerInfo::ActivationFunction F, typename T>
     typename std::enable_if<std::is_same<T, float>::value, void>::type activation(const Window &window);
+#ifdef ARM_COMPUTE_ENABLE_FP16
+    /** Function to apply an activation function on a tensor.
+     *
+     *  @param[in] window Region on which to execute the kernel
+     */
+    template <ActivationLayerInfo::ActivationFunction F, typename T>
+    typename std::enable_if<std::is_same<T, float16_t>::value, void>::type activation(const Window &window);
+#endif /* ARM_COMPUTE_ENABLE_FP16 */
     /** Function to apply an activation function on a tensor.
      *
      *  @param[in] window Region on which to execute the kernel