COMPMID-417: Add Leaky RELU support for both NEON/CL.

-Adds parametrizable leaky relu (x>0) ? x : a*x.

Change-Id: Ief19a435b5832a30b56f4aaaf55125787addee94
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/80575
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

7 files changed, 77 insertions, 34 deletions

diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 2bd449c5c6..765cae4ad4 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -505,6 +505,7 @@ public:
         TANH,         /**< Hyperbolic tangent */
         RELU,         /**< Rectifier */
         BOUNDED_RELU, /**< Bounded Rectifier */
+        LEAKY_RELU,   /**< Leaky Rectifier */
         SOFT_RELU,    /**< Soft Rectifier */
         ABS,          /**< Absolute */
         SQUARE,       /**< Square */
diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl
index 5f812cf5b3..9f958610d6 100644
--- a/src/core/CL/cl_kernels/activation_layer.cl
+++ b/src/core/CL/cl_kernels/activation_layer.cl
@@ -76,6 +76,11 @@ inline TYPE brelu_op(TYPE x)
 {
     return min((TYPE)A_VAL, max(0, x));
 }
+// Leaky RELU Activation
+inline TYPE lrelu_op(TYPE x)
+{
+    return select(MUL_OP((TYPE)A_VAL, x), x, x > (TYPE)0);
+}
 // Soft RELU Activation
 inline TYPE srelu_op(TYPE x)
 {
diff --git a/src/core/NEON/kernels/NEActivationLayerKernel.cpp b/src/core/NEON/kernels/NEActivationLayerKernel.cpp
index f530413453..70b7057fcd 100644
--- a/src/core/NEON/kernels/NEActivationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEActivationLayerKernel.cpp
@@ -73,6 +73,7 @@ void NEActivationLayerKernel::configure(ITensor *input, ITensor *output, Activat
         { ActivationFunction::LOGISTIC, &NEActivationLayerKernel::activation<ActivationFunction::LOGISTIC, float> },
         { ActivationFunction::RELU, &NEActivationLayerKernel::activation<ActivationFunction::RELU, float> },
         { ActivationFunction::BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::BOUNDED_RELU, float> },
+        { ActivationFunction::LEAKY_RELU, &NEActivationLayerKernel::activation<ActivationFunction::LEAKY_RELU, float> },
         { ActivationFunction::SOFT_RELU, &NEActivationLayerKernel::activation<ActivationFunction::SOFT_RELU, float> },
         { ActivationFunction::SQRT, &NEActivationLayerKernel::activation<ActivationFunction::SQRT, float> },
         { ActivationFunction::SQUARE, &NEActivationLayerKernel::activation<ActivationFunction::SQUARE, float> },
@@ -86,6 +87,7 @@ void NEActivationLayerKernel::configure(ITensor *input, ITensor *output, Activat
         { ActivationFunction::LOGISTIC, &NEActivationLayerKernel::activation<ActivationFunction::LOGISTIC, qint8_t> },
         { ActivationFunction::RELU, &NEActivationLayerKernel::activation<ActivationFunction::RELU, qint8_t> },
         { ActivationFunction::BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::BOUNDED_RELU, qint8_t> },
+        { ActivationFunction::LEAKY_RELU, &NEActivationLayerKernel::activation<ActivationFunction::LEAKY_RELU, qint8_t> },
         { ActivationFunction::SOFT_RELU, &NEActivationLayerKernel::activation<ActivationFunction::SOFT_RELU, qint8_t> },
         { ActivationFunction::SQRT, &NEActivationLayerKernel::activation<ActivationFunction::SQRT, qint8_t> },
         { ActivationFunction::SQUARE, &NEActivationLayerKernel::activation<ActivationFunction::SQUARE, qint8_t> },
@@ -99,6 +101,7 @@ void NEActivationLayerKernel::configure(ITensor *input, ITensor *output, Activat
         { ActivationFunction::LOGISTIC, &NEActivationLayerKernel::activation<ActivationFunction::LOGISTIC, qint16_t> },
         { ActivationFunction::RELU, &NEActivationLayerKernel::activation<ActivationFunction::RELU, qint16_t> },
         { ActivationFunction::BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::BOUNDED_RELU, qint16_t> },
+        { ActivationFunction::LEAKY_RELU, &NEActivationLayerKernel::activation<ActivationFunction::LEAKY_RELU, qint16_t> },
         { ActivationFunction::SOFT_RELU, &NEActivationLayerKernel::activation<ActivationFunction::SOFT_RELU, qint16_t> },
         { ActivationFunction::SQRT, &NEActivationLayerKernel::activation<ActivationFunction::SQRT, qint16_t> },
         { ActivationFunction::SQUARE, &NEActivationLayerKernel::activation<ActivationFunction::SQUARE, qint16_t> },
@@ -177,17 +180,6 @@ typename std::enable_if<std::is_same<T, float>::value, void>::type NEActivationL
                     }
                 };
                 break;
-            case ActivationFunction::BOUNDED_RELU:
-                tmp =
-                {
-                    {
-                        vminq_f32(a, vmaxq_f32(CONST_0, in.val[0])),
-                        vminq_f32(a, vmaxq_f32(CONST_0, in.val[1])),
-                        vminq_f32(a, vmaxq_f32(CONST_0, in.val[2])),
-                        vminq_f32(a, vmaxq_f32(CONST_0, in.val[3])),
-                    }
-                };
-                break;
             case ActivationFunction::LINEAR:
                 tmp =
                 {
@@ -221,6 +213,28 @@ typename std::enable_if<std::is_same<T, float>::value, void>::type NEActivationL
                     }
                 };
                 break;
+            case ActivationFunction::BOUNDED_RELU:
+                tmp =
+                {
+                    {
+                        vminq_f32(a, vmaxq_f32(CONST_0, in.val[0])),
+                        vminq_f32(a, vmaxq_f32(CONST_0, in.val[1])),
+                        vminq_f32(a, vmaxq_f32(CONST_0, in.val[2])),
+                        vminq_f32(a, vmaxq_f32(CONST_0, in.val[3])),
+                    }
+                };
+                break;
+            case ActivationFunction::LEAKY_RELU:
+                tmp =
+                {
+                    {
+                        vbslq_f32(vcgtq_f32(in.val[0], CONST_0), in.val[0], vmulq_f32(a, in.val[0])),
+                        vbslq_f32(vcgtq_f32(in.val[1], CONST_0), in.val[1], vmulq_f32(a, in.val[1])),
+                        vbslq_f32(vcgtq_f32(in.val[2], CONST_0), in.val[2], vmulq_f32(a, in.val[2])),
+                        vbslq_f32(vcgtq_f32(in.val[3], CONST_0), in.val[3], vmulq_f32(a, in.val[3])),
+                    }
+                };
+                break;
             case ActivationFunction::SOFT_RELU:
                 tmp =
                 {
@@ -299,9 +313,6 @@ typename std::enable_if<std::is_same<T, int8_t>::value, void>::type NEActivation
             case ActivationFunction::ABS:
                 tmp = vqabsq_qs8(in);
                 break;
-            case ActivationFunction::BOUNDED_RELU:
-                tmp = vminq_qs8(a, vmaxq_qs8(CONST_0, in));
-                break;
             case ActivationFunction::LINEAR:
                 tmp = vqmlaq_qs8(b, a, in, fixed_point_position);
                 break;
@@ -311,6 +322,12 @@ typename std::enable_if<std::is_same<T, int8_t>::value, void>::type NEActivation
             case ActivationFunction::RELU:
                 tmp = vmaxq_qs8(CONST_0, in);
                 break;
+            case ActivationFunction::BOUNDED_RELU:
+                tmp = vminq_qs8(a, vmaxq_qs8(CONST_0, in));
+                break;
+            case ActivationFunction::LEAKY_RELU:
+                tmp = vbslq_s8(vcgtq_s8(in, CONST_0), in, vmulq_qs8(a, in, fixed_point_position));
+                break;
             case ActivationFunction::SOFT_RELU:
                 tmp = vlogq_qs8(vqaddq_qs8(CONST_1, vqexpq_qs8(in, fixed_point_position)), fixed_point_position);
                 break;
@@ -363,15 +380,6 @@ typename std::enable_if<std::is_same<T, int16_t>::value, void>::type NEActivatio
                     }
                 };
                 break;
-            case ActivationFunction::BOUNDED_RELU:
-                tmp =
-                {
-                    {
-                        vminq_qs16(a, vmaxq_qs16(CONST_0, in.val[0])),
-                        vminq_qs16(a, vmaxq_qs16(CONST_0, in.val[1])),
-                    }
-                };
-                break;
             case ActivationFunction::LINEAR:
                 tmp =
                 {
@@ -399,6 +407,24 @@ typename std::enable_if<std::is_same<T, int16_t>::value, void>::type NEActivatio
                     }
                 };
                 break;
+            case ActivationFunction::BOUNDED_RELU:
+                tmp =
+                {
+                    {
+                        vminq_qs16(a, vmaxq_qs16(CONST_0, in.val[0])),
+                        vminq_qs16(a, vmaxq_qs16(CONST_0, in.val[1])),
+                    }
+                };
+                break;
+            case ActivationFunction::LEAKY_RELU:
+                tmp =
+                {
+                    {
+                        vbslq_s16(vcgtq_s16(in.val[0], CONST_0), in.val[0], vmulq_qs16(a, in.val[0], fixed_point_position)),
+                        vbslq_s16(vcgtq_s16(in.val[1], CONST_0), in.val[1], vmulq_qs16(a, in.val[1], fixed_point_position)),
+                    }
+                };
+                break;
             case ActivationFunction::SOFT_RELU:
                 tmp =
                 {
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index 11b41aa178..cf8e1940ec 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -156,6 +156,7 @@ const std::string &arm_compute::string_from_activation_func(ActivationLayerInfo:
         { ActivationLayerInfo::ActivationFunction::LOGISTIC, "LOGISTIC" },
         { ActivationLayerInfo::ActivationFunction::RELU, "RELU" },
         { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, "BRELU" },
+        { ActivationLayerInfo::ActivationFunction::LEAKY_RELU, "LRELU" },
         { ActivationLayerInfo::ActivationFunction::SOFT_RELU, "SRELU" },
         { ActivationLayerInfo::ActivationFunction::SQRT, "SQRT" },
         { ActivationLayerInfo::ActivationFunction::SQUARE, "SQUARE" },
diff --git a/tests/TypePrinter.h b/tests/TypePrinter.h
index ff9863e1fb..c4f3495761 100644
--- a/tests/TypePrinter.h
+++ b/tests/TypePrinter.h
@@ -197,9 +197,6 @@ inline ::std::ostream &operator<<(::std::ostream &os, const ActivationLayerInfo:
         case ActivationLayerInfo::ActivationFunction::ABS:
             os << "ABS";
             break;
-        case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
-            os << "BOUNDED_RELU";
-            break;
         case ActivationLayerInfo::ActivationFunction::LINEAR:
             os << "LINEAR";
             break;
@@ -209,6 +206,12 @@ inline ::std::ostream &operator<<(::std::ostream &os, const ActivationLayerInfo:
         case ActivationLayerInfo::ActivationFunction::RELU:
             os << "RELU";
             break;
+        case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+            os << "BOUNDED_RELU";
+            break;
+        case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+            os << "LEAKY_RELU";
+            break;
         case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
             os << "SOFT_RELU";
             break;
diff --git a/tests/dataset/ActivationFunctionDataset.h b/tests/dataset/ActivationFunctionDataset.h
index bc0e011bde..e6c196560b 100644
--- a/tests/dataset/ActivationFunctionDataset.h
+++ b/tests/dataset/ActivationFunctionDataset.h
@@ -40,17 +40,18 @@ namespace test
  * Can be used as input for Boost data test cases to automatically run a test
  * case on all activation functions.
  */
-class ActivationFunctions final : public GenericDataset<ActivationLayerInfo::ActivationFunction, 9>
+class ActivationFunctions final : public GenericDataset<ActivationLayerInfo::ActivationFunction, 10>
 {
 public:
     ActivationFunctions()
         : GenericDataset
     {
         ActivationLayerInfo::ActivationFunction::ABS,
-        ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
         ActivationLayerInfo::ActivationFunction::LINEAR,
         ActivationLayerInfo::ActivationFunction::LOGISTIC,
         ActivationLayerInfo::ActivationFunction::RELU,
+        ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
+        ActivationLayerInfo::ActivationFunction::LEAKY_RELU,
         ActivationLayerInfo::ActivationFunction::SOFT_RELU,
         ActivationLayerInfo::ActivationFunction::SQRT,
         ActivationLayerInfo::ActivationFunction::SQUARE,
diff --git a/tests/validation/TensorOperations.h b/tests/validation/TensorOperations.h
index e2747249b4..27c50cf6d2 100644
--- a/tests/validation/TensorOperations.h
+++ b/tests/validation/TensorOperations.h
@@ -868,9 +868,6 @@ void activation_layer(const Tensor<T> &in, Tensor<T> &out, ActivationLayerInfo a
             case ActivationLayerInfo::ActivationFunction::ABS:
                 out[i] = std::abs(x);
                 break;
-            case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
-                out[i] = std::min<T>(a, std::max<T>(0, x));
-                break;
             case ActivationLayerInfo::ActivationFunction::LINEAR:
                 out[i] = a * x + b;
                 break;
@@ -880,6 +877,12 @@ void activation_layer(const Tensor<T> &in, Tensor<T> &out, ActivationLayerInfo a
             case ActivationLayerInfo::ActivationFunction::RELU:
                 out[i] = std::max<T>(0, x);
                 break;
+            case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                out[i] = std::min<T>(a, std::max<T>(0, x));
+                break;
+            case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+                out[i] = (x > 0) ? x : a * x;
+                break;
             case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
                 out[i] = std::log(static_cast<T>(1) + std::exp(x));
                 break;
@@ -919,9 +922,6 @@ void activation_layer(const Tensor<T> &in, Tensor<T> &out, ActivationLayerInfo a
             case ActivationLayerInfo::ActivationFunction::ABS:
                 out[i] = abs(x).raw();
                 break;
-            case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
-                out[i] = min(a, max(const_0, x)).raw();
-                break;
             case ActivationLayerInfo::ActivationFunction::LINEAR:
                 out[i] = add(b, mul(a, x)).raw();
                 break;
@@ -931,6 +931,12 @@ void activation_layer(const Tensor<T> &in, Tensor<T> &out, ActivationLayerInfo a
             case ActivationLayerInfo::ActivationFunction::RELU:
                 out[i] = max(const_0, x).raw();
                 break;
+            case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                out[i] = min(a, max(const_0, x)).raw();
+                break;
+            case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+                out[i] = (x > const_0) ? x.raw() : mul(a, x).raw();
+                break;
             case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
                 out[i] = log(const_1 + exp(x)).raw();
                 break;