COMPMID-3097 Fuse activation with fully connected layer CL

Change-Id: I447030e69b9e565f2f81529a41af8c5e7ece7ecf
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2702
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

15 files changed, 86 insertions, 77 deletions

diff --git a/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp b/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp
index 5550003f33..b9563553b8 100644
--- a/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp
@@ -113,22 +113,9 @@ Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vecto
         ARM_COMPUTE_RETURN_ERROR_ON(output_stage.output_data_type != output->data_type());
         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mm_result, output);
-        PixelValue min_val{};
-        PixelValue max_val{};
-        std::tie(min_val, max_val) = get_min_max(output->data_type());
-        ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_max_bound > max_val.get<int32_t>());
-        ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound < min_val.get<int32_t>() || output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
-    }
-    else
-    {
-        // Output will be configured as depending on the chosen output data type in the output stage
-        PixelValue min_val{};
-        PixelValue max_val{};
-        std::tie(min_val, max_val) = get_min_max(output_stage.output_data_type);
-        ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_max_bound > max_val.get<int32_t>());
-        ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound < min_val.get<int32_t>() || output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
     }
 
+    ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(output_stage.gemmlowp_multipliers.size() != output_stage.gemmlowp_shifts.size(), "per channel quantization info is incorrect");
 
     return Status{};
@@ -248,8 +235,8 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const ICLTensor *m
     PixelValue min_val{};
     PixelValue max_val{};
     std::tie(min_val, max_val) = get_min_max(output->info()->data_type());
-    build_opts.add_option_if((min != min_val.get<int32_t>()) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
-    build_opts.add_option_if((max != max_val.get<int32_t>()) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    build_opts.add_option_if((min > min_val.get<int32_t>()), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max < max_val.get<int32_t>()), "-DMAX_BOUND=" + support::cpp11::to_string(max));
 
     std::string kernel_name("gemmlowp_offset_contribution");
     kernel_name += "_" + string_from_gemmlowp_output_stage(output_stage.type);
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
index 8720123366..1bc7fe3946 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -45,8 +45,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
                           int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 32767);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < -32768 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -137,8 +136,8 @@ void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(const
     CLBuildOptions build_opts;
     build_opts.add_option("-DRESULT_FIXEDPOINT_MULTIPLIER=" + support::cpp11::to_string(result_fixedpoint_multiplier));
     build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(result_shift));
-    build_opts.add_option_if((min != -32768) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
-    build_opts.add_option_if((max != 32767) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    build_opts.add_option_if((min > -32768), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max < 32767), "-DMAX_BOUND=" + support::cpp11::to_string(max));
     build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
 
     // Create kernel
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
index 8a5ce9fa87..e207fcb1b0 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
@@ -44,8 +44,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
                           int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 127);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < -128 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -136,8 +135,8 @@ void CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure(const I
     build_opts.add_option("-DRESULT_FIXEDPOINT_MULTIPLIER=" + support::cpp11::to_string(result_fixedpoint_multiplier));
     build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(result_shift));
     build_opts.add_option("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type()));
-    build_opts.add_option_if((min != -128) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
-    build_opts.add_option_if((max != 127) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    build_opts.add_option_if((min > -128), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max < 127), "-DMAX_BOUND=" + support::cpp11::to_string(max));
     build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
 
     // Create kernel
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index dc04fed96e..7601d7ee77 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -44,8 +44,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
                           int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -136,8 +135,8 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const
     build_opts.add_option("-DRESULT_FIXEDPOINT_MULTIPLIER=" + support::cpp11::to_string(result_fixedpoint_multiplier));
     build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(result_shift));
     build_opts.add_option("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type()));
-    build_opts.add_option_if((min != 0) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
-    build_opts.add_option_if((max != 255) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    build_opts.add_option_if((min > 0), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max < 255), "-DMAX_BOUND=" + support::cpp11::to_string(max));
     build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
 
     // Create kernel
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
index ae096f295c..dd1be748f5 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,8 +43,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
                           int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -132,8 +131,8 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::configure(const ICLTe
     CLBuildOptions build_opts;
     build_opts.add_option("-DREAL_MULTIPLIER=" + float_to_string_with_full_precision(multiplier));
     build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(offset));
-    build_opts.add_option_if((min != 0) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
-    build_opts.add_option_if((max != 255) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    build_opts.add_option_if((min > 0), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max < 255), "-DMAX_BOUND=" + support::cpp11::to_string(max));
     build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
 
     // Create kernel
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
index 8175f60275..7a22239a7c 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,8 +41,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -135,8 +134,8 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ICLTensor *i
     build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(result_offset));
     build_opts.add_option("-DRESULT_MULT_INT=" + support::cpp11::to_string(result_mult_int));
     build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(result_shift));
-    build_opts.add_option_if((min != 0) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
-    build_opts.add_option_if((max != 255) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    build_opts.add_option_if((min > 0), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+    build_opts.add_option_if((max < 255), "-DMAX_BOUND=" + support::cpp11::to_string(max));
     build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
 
     // Create kernel
diff --git a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp
index 5d2df6d2c9..31414e3f3f 100644
--- a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp
@@ -778,15 +778,8 @@ Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vecto
                           int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32);
-    if(output->data_type() == DataType::QASYMM8)
+    if(output->data_type() != DataType::QASYMM8)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_max_bound > 255);
-        ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound < 0);
-    }
-    else
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_max_bound > 127);
-        ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound < -128);
         ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) > 1 && output_stage.gemmlowp_multipliers.size() > 1 && b_offset != 0);
     }
     ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
@@ -914,7 +907,7 @@ get_configured_function(const ITensor *mm_result, const ITensor *vector_sum_row,
     std::tie(type_min, type_max) = get_min_max(output->info()->data_type());
     int32_t    type_min_int    = type_min.get<int32_t>();
     int32_t    type_max_int    = type_max.get<int32_t>();
-    const bool is_bounded_relu = !(output_stage.gemmlowp_min_bound == type_min_int && output_stage.gemmlowp_max_bound == type_max_int);
+    const bool is_bounded_relu = !(output_stage.gemmlowp_min_bound <= type_min_int && output_stage.gemmlowp_max_bound >= type_max_int);
 
     // Check if we need to perform fixed point requantization
     const bool is_fixed_point = output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN;
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
index bc513e6618..058007139d 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -46,8 +46,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 32767);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < -32768 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -213,7 +212,7 @@ void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(const
     INEKernel::configure(win_config.second);
 
     // Check if we need to clamp the result using min and max
-    const bool is_bounded_relu = ((min != max) && !(min == -32768 && max == 32767));
+    const bool is_bounded_relu = !(min <= -32768 && max >= 32767);
     _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<false>;
 }
 
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
index d24089d615..b8ca17ec3d 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -46,8 +46,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 127);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < -128 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -222,7 +221,7 @@ void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure(const I
     INEKernel::configure(win_config.second);
 
     // Check if we need to clamp the result using min and max
-    const bool is_bounded_relu = ((min != max) && !(min == -128 && max == 127));
+    const bool is_bounded_relu = !(min <= -128 && max >= 127);
     _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run<false>;
 }
 
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index bb0b86404e..4a9d2f7481 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -46,8 +46,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -224,7 +223,7 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const
     INEKernel::configure(win_config.second);
 
     // Check if we need to clamp the result using min and max
-    const bool is_bounded_relu = ((min != max) && !(min == 0 && max == 255));
+    const bool is_bounded_relu = !(min <= 0 && max >= 255);
     _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run<false>;
 }
 
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
index a221bd7925..a68e4e7efb 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,8 +43,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
-    ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
+    ARM_COMPUTE_RETURN_ERROR_ON(min > max);
 
     // Check biases if exist
     if(bias != nullptr)
@@ -324,7 +323,7 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ITensor *inp
     INEKernel::configure(win_config.second);
 
     // Check if we need to clamp the result using min and max
-    const bool is_bounded_relu = ((min != max) && !(min == 0 && max == 255));
+    const bool is_bounded_relu = !(min <= 0 && max >= 255);
     _func                      = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run<false>;
 }
 
diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp
index b7f081dc42..151a8bfa03 100644
--- a/src/graph/mutators/NodeFusionMutator.cpp
+++ b/src/graph/mutators/NodeFusionMutator.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -301,6 +301,10 @@ void NodeFusionMutator::mutate(Graph &g)
     {
         return true;
     };
+    auto cl_target_prec = [](INode & n)
+    {
+        return n.assigned_target() == Target::CL;
+    };
     auto qs8_prec = [&g](INode & n)
     {
         ARM_COMPUTE_ERROR_ON(n.output(0) == nullptr);
@@ -318,6 +322,7 @@ void NodeFusionMutator::mutate(Graph &g)
     detail::fuse_layer<BatchNormalizationLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<BatchNormalizationLayerNode>, supported_fused_activations);
     detail::fuse_layer<ConvolutionLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<ConvolutionLayerNode>, supported_fused_activations);
     detail::fuse_layer<DepthwiseConvolutionLayerNode, ActivationLayerNode>(g, qs8_prec, detail::fuse_node_with_activation<DepthwiseConvolutionLayerNode>, supported_fused_activations);
+    detail::fuse_layer<FullyConnectedLayerNode, ActivationLayerNode>(g, cl_target_prec, detail::fuse_node_with_activation<FullyConnectedLayerNode>, supported_fused_activations);
     detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization);
     detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization);
 }
diff --git a/src/graph/nodes/FullyConnectedLayer.cpp b/src/graph/nodes/FullyConnectedLayer.cpp
index 80fce7b8a1..34c432a1ce 100644
--- a/src/graph/nodes/FullyConnectedLayer.cpp
+++ b/src/graph/nodes/FullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -38,6 +38,11 @@ FullyConnectedLayerNode::FullyConnectedLayerNode(unsigned int num_outputs, Quant
     _outputs.resize(1, NullTensorID);
 }
 
+void FullyConnectedLayerNode::set_fused_activation(ActivationLayerInfo fused_activation)
+{
+    _info.activation_info = fused_activation;
+}
+
 TensorDescriptor FullyConnectedLayerNode::compute_weights_descriptor(const TensorDescriptor &input_descriptor,
                                                                      unsigned int            num_outputs,
                                                                      FullyConnectedLayerInfo fc_info,
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index dcaa12645e..9b7de8df1b 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -41,7 +41,7 @@ using namespace arm_compute::utils::cast;
 namespace
 {
 Status construct_gemmlowp_output_stage(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo &output,
-                                       GEMMLowpOutputStageInfo &gemmlowp_output_stage)
+                                       GEMMLowpOutputStageInfo &gemmlowp_output_stage, ActivationLayerInfo activation_info)
 {
     gemmlowp_output_stage.type                = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
     gemmlowp_output_stage.gemmlowp_offset     = 0;
@@ -53,13 +53,14 @@ Status construct_gemmlowp_output_stage(const ITensorInfo &input, const ITensorIn
     // Configure output stage for quantized case
     if(is_data_type_quantized_asymmetric(data_type))
     {
-        const UniformQuantizationInfo iq_info = input.quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = output.quantization_info().uniform();
+        const QuantizationInfo        oq_info = output.quantization_info();
+        const UniformQuantizationInfo iq_unif = input.quantization_info().uniform();
+        const UniformQuantizationInfo wq_unif = weights.quantization_info().uniform();
+        const UniformQuantizationInfo oq_unif = oq_info.uniform();
 
-        const auto output_quant_info = (output.total_size() == 0) ? iq_info : oq_info;
+        const auto output_quant_info = (output.total_size() == 0) ? iq_unif : oq_unif;
 
-        const float multiplier        = (iq_info.scale * wq_info.scale) / output_quant_info.scale;
+        const float multiplier        = (iq_unif.scale * wq_unif.scale) / output_quant_info.scale;
         int         output_multiplier = 0;
         int         output_shift      = 0;
         ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
@@ -68,6 +69,27 @@ Status construct_gemmlowp_output_stage(const ITensorInfo &input, const ITensorIn
         PixelValue type_max{};
         std::tie(type_min, type_max) = get_min_max(data_type);
 
+        if(activation_info.enabled())
+        {
+            switch(activation_info.activation())
+            {
+                case ActivationLayerInfo::ActivationFunction::RELU:
+                    type_min = PixelValue(oq_unif.offset);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                    type_min = PixelValue(oq_unif.offset);
+                    type_max = PixelValue(activation_info.a(), data_type, oq_info);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
+                    type_min = PixelValue(activation_info.b(), data_type, oq_info);
+                    type_max = PixelValue(activation_info.a(), data_type, oq_info);
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Activation function not supported.");
+                    break;
+            }
+        }
+
         // Set the GEMMLowp output stage info
         gemmlowp_output_stage.gemmlowp_offset     = output_quant_info.offset;
         gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
@@ -84,7 +106,7 @@ Status construct_gemmlowp_output_stage(const ITensorInfo &input, const ITensorIn
 Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo *bias, const ITensorInfo &output, const FullyConnectedLayerInfo &fc_info)
 {
     GEMMLowpOutputStageInfo gemmlowp_output_stage;
-    ARM_COMPUTE_RETURN_ON_ERROR(construct_gemmlowp_output_stage(input, weights, output, gemmlowp_output_stage));
+    ARM_COMPUTE_RETURN_ON_ERROR(construct_gemmlowp_output_stage(input, weights, output, gemmlowp_output_stage, fc_info.activation_info));
 
     const GEMMInfo &gemm_info = GEMMInfo(false,                           // is_a_reshaped
                                          false,                           // is_b_reshaped
@@ -144,7 +166,7 @@ CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> mem
 void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const FullyConnectedLayerInfo &fc_info)
 {
     GEMMLowpOutputStageInfo gemmlowp_output_stage;
-    construct_gemmlowp_output_stage(*input->info(), *weights->info(), *output->info(), gemmlowp_output_stage);
+    construct_gemmlowp_output_stage(*input->info(), *weights->info(), *output->info(), gemmlowp_output_stage, fc_info.activation_info);
 
     const GEMMInfo &gemm_info = GEMMInfo(false,                           // is_a_reshaped
                                          false,                           // is_b_reshaped
@@ -155,7 +177,7 @@ void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor
                                          gemmlowp_output_stage,           // gemmlowp_output_stage
                                          fc_info.fp_mixed_precision,      // fp_mixed_precision
                                          true,                            // broadcast_bias
-                                         ActivationLayerInfo());          // activation_info
+                                         fc_info.activation_info);        // activation_info
 
     if(_is_quantized)
     {
@@ -313,6 +335,8 @@ Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(fc_info.activation_info.enabled() && is_data_type_quantized(input->data_type()) && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::RELU
+                                && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU);
 
     bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
     bool is_fc_after_conv = true;
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index 682812b1c8..5398050533 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -333,8 +333,12 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
         gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0];
         gemmlowp_output_stage.gemmlowp_shift      = gemmlowp_output_stage.gemmlowp_shifts[0];
 
-        int min_activation = 0;
-        int max_activation = 0;
+        PixelValue min_val{};
+        PixelValue max_val{};
+        std::tie(min_val, max_val) = get_min_max(output->info()->data_type());
+
+        auto min_activation = min_val.get<int32_t>();
+        auto max_activation = max_val.get<int32_t>();
 
         const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
                                                                                    ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,