COMPMID-897 Merge batch normalization with bounded relu

Change-Id: I9a607fe620f795cdea1a99fdd3f5f8c2fc76f980
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/119234
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

2 files changed, 33 insertions, 9 deletions

diff --git a/src/core/GLES_COMPUTE/cs_shaders/batchnormalization_layer.cs b/src/core/GLES_COMPUTE/cs_shaders/batchnormalization_layer.cs
index 53fb51557c..7629b255b7 100644
--- a/src/core/GLES_COMPUTE/cs_shaders/batchnormalization_layer.cs
+++ b/src/core/GLES_COMPUTE/cs_shaders/batchnormalization_layer.cs
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,6 +36,16 @@ precision mediump float;
 #define INVSQRT_OP(a) inversesqrt((a))
 #define SQCVT_SAT(a) (a)
 
+#if defined(LU_BRELU)
+#define ACTIVATION_FUNC(x) min(max(x, float(B_VAL)), float(A_VAL))
+#elif defined(BRELU)
+#define ACTIVATION_FUNC(x) min(max(x, float(0)), float(A_VAL))
+#elif defined(RELU)
+#define ACTIVATION_FUNC(x) max(x, float(0))
+#else /* defined(FUSED_ACT) */
+#define ACTIVATION_FUNC(x) (x)
+#endif /* defined(FUSED_ACT) */
+
 /** Apply batch normalization.
  *
  * @note The data type must be passed at compile time using "#define DATA_TYPE_NAME". e.g. "#define DATA_TYPE_FP32"
@@ -102,7 +112,7 @@ void main(void)
     gamma_param = LOAD(gamma_ptr, TENSOR_OFFSET_ADVANCE_IN_BYTES(gamma_iter, current_slice * beta_attrs.stride_x));
     beta_param  = LOAD(beta_ptr, TENSOR_OFFSET_ADVANCE_IN_BYTES(beta_iter, current_slice * beta_attrs.stride_x));
 
-    STORE_CURRENT_ITEM(dst_ptr, dst_iter, ADD_OP(MUL_OP(gamma_param, x_bar), beta_param));
+    STORE_CURRENT_ITEM(dst_ptr, dst_iter, ACTIVATION_FUNC(ADD_OP(MUL_OP(gamma_param, x_bar), beta_param)));
 }
 
 #elif defined(DATA_TYPE_FP16)
@@ -148,7 +158,7 @@ void main(void)
 
         gamma_param = unpacked_s[3].x;
         beta_param  = unpacked_s[4].x;
-        result      = ADD_OP(MUL_OP(gamma_param, x_bar), beta_param);
+        result      = ACTIVATION_FUNC(ADD_OP(MUL_OP(gamma_param, x_bar), beta_param));
 
         STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, result);
     }
@@ -163,7 +173,7 @@ void main(void)
 
         gamma_param = unpacked_s[3].y;
         beta_param  = unpacked_s[4].y;
-        result      = ADD_OP(MUL_OP(gamma_param, x_bar), beta_param);
+        result      = ACTIVATION_FUNC(ADD_OP(MUL_OP(gamma_param, x_bar), beta_param));
 
         STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, result);
     }
@@ -178,7 +188,7 @@ void main(void)
 
         gamma_param = unpacked_s[3].z;
         beta_param  = unpacked_s[4].z;
-        result      = ADD_OP(MUL_OP(gamma_param, x_bar), beta_param);
+        result      = ACTIVATION_FUNC(ADD_OP(MUL_OP(gamma_param, x_bar), beta_param));
 
         STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, result);
     }
@@ -193,7 +203,7 @@ void main(void)
 
         gamma_param = unpacked_s[3].w;
         beta_param  = unpacked_s[4].w;
-        result      = ADD_OP(MUL_OP(gamma_param, x_bar), beta_param);
+        result      = ACTIVATION_FUNC(ADD_OP(MUL_OP(gamma_param, x_bar), beta_param));
 
         STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, result);
     }
diff --git a/src/core/GLES_COMPUTE/kernels/GCBatchNormalizationLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCBatchNormalizationLayerKernel.cpp
index dee2a5579b..a41b62fbab 100644
--- a/src/core/GLES_COMPUTE/kernels/GCBatchNormalizationLayerKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCBatchNormalizationLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -42,7 +42,7 @@ GCBatchNormalizationLayerKernel::GCBatchNormalizationLayerKernel()
 }
 
 void GCBatchNormalizationLayerKernel::configure(const IGCTensor *input, IGCTensor *output, const IGCTensor *mean, const IGCTensor *var, const IGCTensor *beta, const IGCTensor *gamma,
-                                                float epsilon)
+                                                float epsilon, ActivationLayerInfo act_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
@@ -54,7 +54,14 @@ void GCBatchNormalizationLayerKernel::configure(const IGCTensor *input, IGCTenso
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output, mean, var, beta, gamma);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(mean, var, beta, gamma);
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != mean->info()->dimension(0));
+    if(act_info.enabled())
+    {
+        ARM_COMPUTE_ERROR_ON(input->info()->data_type() != DataType::F32 && input->info()->data_type() != DataType::F16);
+        ARM_COMPUTE_ERROR_ON(act_info.activation() != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::RELU
+                             && act_info.activation() != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU
+                             && act_info.activation() != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU);
+        ARM_COMPUTE_ERROR_ON(act_info.b() > act_info.a());
+    }
 
     _input   = input;
     _output  = output;
@@ -79,6 +86,13 @@ void GCBatchNormalizationLayerKernel::configure(const IGCTensor *input, IGCTenso
     build_opts.emplace(("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)));
     build_opts.emplace(("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)));
 
+    if(act_info.enabled())
+    {
+        build_opts.emplace("#define " + string_from_activation_func(act_info.activation()));
+        build_opts.emplace("#define A_VAL " + float_to_string_with_full_precision(act_info.a()));
+        build_opts.emplace("#define B_VAL " + float_to_string_with_full_precision(act_info.b()));
+    }
+
     // Create kernel
     _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("batchnormalization_layer", build_opts));