COMPMID-2608: Enable quantization with multiplier greater than 1 on NEON

Change-Id: Ib2b0c9ac88fc2b645f478c9981f71ee28f2c77fd
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2425
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, 42 insertions, 25 deletions

diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index ca4fe732a7..ddcc71f466 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -60,16 +60,6 @@ Status validate_arguments_optimized(const ITensorInfo *input, const ITensorInfo
 
     const bool is_quantized = (!is_data_type_quantized_per_channel(weights->data_type())) && is_data_type_quantized_asymmetric(input->data_type());
 
-    if(is_quantized)
-    {
-        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
-
-        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
-        ARM_COMPUTE_UNUSED(multiplier);
-        ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
-    }
     if(!NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(input, weights, conv_info, depth_multiplier, dilation))
     {
         TensorInfo accumulator = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
@@ -205,7 +195,7 @@ void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::
         float   multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
         int32_t output_multiplier;
         int32_t output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+        quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
         _output_stage_kernel.configure(&_accumulator, biases, _is_nchw ? output : &_permuted_output, output_multiplier, output_shift, oq_info.offset);
         _accumulator.allocator()->allocate();
     }
diff --git a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
index 322bb2c425..65538848df 100644
--- a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
@@ -31,8 +31,8 @@
 #include <cmath>
 #include <tuple>
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 NEDirectConvolutionLayer::NEDirectConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _output_stage_kernel(), _conv_kernel(), _input_border_handler(), _activationlayer_function(), _accumulator(), _has_bias(false),
       _is_activationlayer_enabled(false), _dim_split(Window::DimZ)
@@ -118,3 +118,4 @@ void NEDirectConvolutionLayer::run()
         _activationlayer_function.run();
     }
 }
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index b3b90f8599..01746eb3db 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -33,7 +33,8 @@
 #include <algorithm>
 #include <cmath>
 
-using namespace arm_compute;
+namespace arm_compute
+{
 using namespace arm_compute::misc::shape_calculator;
 
 namespace
@@ -258,7 +259,7 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
         float   multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
         int32_t output_multiplier;
         int32_t output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+        quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
         _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output_multiplier, output_shift, oq_info.offset);
         _gemmlowp_output.allocator()->allocate();
     }
@@ -352,13 +353,14 @@ Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn
     // Validate output stage for asymmetric quantized types
     if(is_quantized)
     {
-        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 float                   multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
-
-        ARM_COMPUTE_UNUSED(multiplier);
-        ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f);
+        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
+        const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
+        const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
+
+        float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale;
+        int   output_multiplier;
+        int   output_shift;
+        ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
         ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(&gemmlowp_output, biases, output));
     }
 
@@ -475,4 +477,5 @@ void NEFullyConnectedLayer::prepare()
 
         _is_prepared = true;
     }
-}
\ No newline at end of file
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index 0507c6b2bd..920917a58b 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -154,7 +154,7 @@ void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *w
         output_info.gemmlowp_min_bound       = min_activation;
         output_info.gemmlowp_max_bound       = max_activation;
         output_info.is_quantized_per_channel = (weights->info()->data_type() == DataType::QSYMM8_PER_CHANNEL);
-        quantization::calculate_quantized_multipliers_less_than_one(iqinfo, wqinfo, oqinfo, output_info);
+        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));
 
@@ -217,7 +217,7 @@ Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
         output_info.gemmlowp_min_bound       = min_activation;
         output_info.gemmlowp_max_bound       = max_activation;
         output_info.is_quantized_per_channel = (weights->data_type() == DataType::QSYMM8_PER_CHANNEL);
-        ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multipliers_less_than_one(iqinfo, wqinfo, oqinfo, output_info));
+        ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info));
 
         // Perform validation step on GEMMLowp
         std::unique_ptr<ITensorInfo> input_qa   = input->clone();
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index e36cb3d399..440f043527 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -344,6 +344,19 @@ Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
     {
         run_optimised             = bool(NEGEMMAssemblyDispatch::validate(a_to_use, b, c, output, gemm_info));
         run_optimised_requantized = run_optimised;
+
+        const UniformQuantizationInfo a_qinfo      = a_to_use->quantization_info().uniform();
+        const QuantizationInfo        b_qinfo      = b->quantization_info();
+        const UniformQuantizationInfo output_qinfo = output->quantization_info().uniform();
+        for(auto const s : b_qinfo.scale())
+        {
+            const float fmultipler = a_qinfo.scale * s / output_qinfo.scale;
+            if(fmultipler > 1.f)
+            {
+                run_optimised_requantized = false;
+                break;
+            }
+        }
     }
     else
     {
diff --git a/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp b/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
index 3235eee19a..142f873ef4 100644
--- a/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
+++ b/src/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.cpp
@@ -437,6 +437,16 @@ Status NEDepthwiseConvolutionAssemblyDispatch::validate(const ITensorInfo
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
 
+    // The uniform quantization case will only have 1 scale value in the weights quantization info
+    const UniformQuantizationInfo input_qinfo   = input->quantization_info().uniform();
+    const QuantizationInfo        weights_qinfo = weights->quantization_info();
+    const UniformQuantizationInfo output_qinfo  = output->quantization_info().uniform();
+    for(auto const s : weights_qinfo.scale())
+    {
+        const float fmultipler = input_qinfo.scale * s / output_qinfo.scale;
+        ARM_COMPUTE_RETURN_ERROR_ON(fmultipler > 1.f);
+    }
+
     return Status{};
 }