Fix NEGEMMLowpMatrixMultiplyCore nonfused activation

* Include the missing activation info when passing it from
NEGEMMConvolutionLayer to NEGEMMLowpMatrixMultiplyCore

* Fix the logics for running the separate NEActivationLayer when it
cannot be fused into NEGEMMAssemblyDispatch

Resolves COMPMID-4039

Change-Id: Id6b3a698333f278ee65f195853c7d3b5c7adec72
Signed-off-by: SiCong Li <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4638
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 10 insertions, 11 deletions

diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index a3bdde24b0..8fc788c402 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -169,7 +169,7 @@ void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *w
         output_info.is_quantized_per_channel = (weights->info()->data_type() == DataType::QSYMM8_PER_CHANNEL);
         quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info);
 
-        _mm_gemmlowp.configure(input, weights, biases, output, GEMMInfo(false, false, true, gemm_3d_depth, _skip_im2col, false, output_info));
+        _mm_gemmlowp.configure(input, weights, biases, output, GEMMInfo(false, false, true, gemm_3d_depth, _skip_im2col, false, output_info, false, false, act_info));
 
         // Revert back QuantizatioInfo as input and weights could be used in other convolution layers
         input->info()->set_quantization_info(iqinfo);
@@ -232,7 +232,7 @@ Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
         std::unique_ptr<ITensorInfo> weights_qa = weights->clone();
         input_qa->set_quantization_info(QuantizationInfo(iqinfo.uniform().scale, -iqinfo.uniform().offset));
         weights_qa->set_quantization_info(QuantizationInfo(wqinfo.uniform().scale, -wqinfo.uniform().offset));
-        return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, output, GEMMInfo(false, false, true, gemm_3d_depth, skip_im2col, false, output_info));
+        return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, output, GEMMInfo(false, false, true, gemm_3d_depth, skip_im2col, false, output_info, false, false, act_info));
     }
     else
     {
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index d8f9d08c13..50c7fe4c66 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -257,14 +257,13 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
             _offset_contribution_kernel = std::make_unique<NEGEMMLowpOffsetContributionKernel>();
             _offset_contribution_kernel->configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, a_to_use->info()->dimension(0), _a_offset, _b_offset);
         }
-
-        // Configure activation
-        const ActivationLayerInfo &activation = gemm_info.activation_info();
-        _run_activation                       = activation.enabled() && (!_assembly_path || (_assembly_path && !NEGEMMAssemblyDispatch::is_activation_supported(activation)));
-        if(_run_activation)
-        {
-            _activation_func.configure(output, nullptr, activation);
-        }
+    }
+    // Configure activation
+    const ActivationLayerInfo &activation = gemm_info.activation_info();
+    _run_activation                       = activation.enabled() && (!_assembly_path || !NEGEMMAssemblyDispatch::is_activation_supported(activation));
+    if(_run_activation)
+    {
+        _activation_func.configure(output, nullptr, activation);
     }
 
     // Allocate tensors
@@ -564,7 +563,7 @@ void NEGEMMLowpMatrixMultiplyCore::run()
     }
 
     // Run fused activation unless already run in the fused assembly
-    if(_run_activation && !_fused_assembly_path)
+    if(_run_activation)
     {
         _activation_func.run();
     }