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>

6 files changed, 21 insertions, 39 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