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>

5 files changed, 14 insertions, 25 deletions

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>;
 }