COMPMID-970 : Remove QS8 / QS16 support

Removed fixed point related code.

Change-Id: I487acf138dace3b0450e0d72ca7071eaec254566
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137678
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

8 files changed, 25 insertions, 56 deletions

diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
index 8051d6da0e..fda9f57499 100644
--- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -63,18 +63,15 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
                                                     info.pad().first, info.pad().second, inner_border_right, inner_border_top, stride_x, stride_y);
 
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, bias);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights, bias);
 
     if(bias != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, bias);
     }
 
     if(output->tensor_shape().total_size() > 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
 
         const TensorShape output_shape = deconvolution_output_shape(out_dims, input->tensor_shape(), weights->tensor_shape());
 
@@ -117,8 +114,7 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
 
     // configure scale function
     // Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor
-    const TensorInfo scale_out_info(compute_deconvolution_shape(*input->info(), stride_x, stride_y, inner_border_right, inner_border_top, info), 1, input->info()->data_type(),
-                                    input->info()->fixed_point_position());
+    const TensorInfo scale_out_info(compute_deconvolution_shape(*input->info(), stride_x, stride_y, inner_border_right, inner_border_top, info), 1, input->info()->data_type());
     _scaled_output.allocator()->init(scale_out_info);
 
     // setup the function to convolve the upscaled output
diff --git a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
index 930f8d5a26..3d47ec2ac2 100644
--- a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
@@ -52,7 +52,7 @@ void NEDepthConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, IT
     TensorShape output_shape = calculate_depth_concatenate_shape(inputs_vector);
 
     // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type(), inputs_vector[0]->info()->fixed_point_position());
+    auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type());
 
     unsigned int depth_offset = 0;
     for(unsigned int i = 0; i < _num_inputs; ++i)
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 83c3e217f3..1d65dde2a6 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -88,7 +88,7 @@ void NEDepthwiseConvolutionLayer3x3::configure(ITensor *input, const ITensor *we
     }
     else
     {
-        // Allocate the intermediate accumulator tensor in case of fixed point input
+        // Allocate the intermediate accumulator tensor in case of quantized input
         if(_is_quantized)
         {
             _accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, DataType::S32));
diff --git a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
index 445864c2a9..40e40c8ffa 100644
--- a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
@@ -34,7 +34,7 @@
 using 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_fixed_point(false),
+    : _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)
 {
 }
@@ -54,26 +54,10 @@ void NEDirectConvolutionLayer::configure(ITensor *input, const ITensor *weights,
     // Check if bias should be added in the convolution result
     _has_bias = (bias != nullptr);
 
-    // Allocate the intermediate accumulator tensor in case of fixed point input
-    _is_fixed_point = is_data_type_fixed_point(input->info()->data_type());
-    if(_is_fixed_point)
+    _conv_kernel.configure(input, weights, output, conv_info);
+    if(_has_bias)
     {
-        const DataType promoted_dt = (input->info()->data_type() == DataType::QS8) ? DataType::QS16 : DataType::QS32;
-        _accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, promoted_dt, output->info()->fixed_point_position()));
-        _memory_group.manage(&_accumulator);
-        _conv_kernel.configure(input, weights, &_accumulator, conv_info);
-
-        // When no bias is provided, we need to downscale the accumulator tensor
-        _output_stage_kernel.configure(&_accumulator, bias, output);
-        _accumulator.allocator()->allocate();
-    }
-    else
-    {
-        _conv_kernel.configure(input, weights, output, conv_info);
-        if(_has_bias)
-        {
-            _output_stage_kernel.configure(output, bias);
-        }
+        _output_stage_kernel.configure(output, bias);
     }
 
     // Add zero padding XY
@@ -92,12 +76,7 @@ Status NEDirectConvolutionLayer::validate(const ITensorInfo *input, const ITenso
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
 
-    DataType data_type = output->data_type();
-    if(is_data_type_fixed_point(data_type))
-    {
-        // Promote data type in case of fixed point
-        data_type = ((data_type == DataType::QS8) ? DataType::QS16 : DataType::QS32);
-    }
+    DataType   data_type = output->data_type();
     TensorInfo accumulator(output->clone()->set_is_resizable(true).reset_padding().set_data_type(data_type));
 
     // Validate Convolution kernel
@@ -129,7 +108,7 @@ void NEDirectConvolutionLayer::run()
     _memory_group.acquire();
 
     NEScheduler::get().schedule(&_conv_kernel, _dim_split);
-    if(_has_bias || _is_fixed_point)
+    if(_has_bias)
     {
         NEScheduler::get().schedule(&_output_stage_kernel, Window::DimY);
     }
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index 5b9f182bcb..3126823e9c 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -83,7 +83,7 @@ void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITenso
 
 Status NEFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output, bool transpose_weights, bool is_batched_fc_layer)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 2);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!transpose_weights && !is_batched_fc_layer, "Configuration transpose_weights=false & is_batched_fc_layer=false not supported");
 
@@ -233,9 +233,8 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
 
 Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights, bool are_weights_reshaped)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, weights, output);
 
     const int    num_batch_dimensions = std::max(0, static_cast<int>(output->tensor_shape().num_dimensions()) - 1);
     const int    num_input_dimensions = input->tensor_shape().num_dimensions() - num_batch_dimensions;
diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp
index a98309d304..795ffc5d1c 100644
--- a/src/runtime/NEON/functions/NEGEMM.cpp
+++ b/src/runtime/NEON/functions/NEGEMM.cpp
@@ -46,7 +46,7 @@ NEGEMM::NEGEMM(std::shared_ptr<IMemoryManager> memory_manager)
 
 void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16, DataType::QS8, DataType::QS16);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, d);
     ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
     ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
@@ -54,7 +54,7 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
 
     if(c != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16, DataType::QS8, DataType::QS16);
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c);
         ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A");
         ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix B");
@@ -103,8 +103,8 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
             shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w);
             shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w)));
 
-            TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position());
-            TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), a->info()->fixed_point_position());
+            TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type());
+            TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type());
 
             _tmp_a.allocator()->init(info_a);
             _tmp_b.allocator()->init(info_b);
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index d4400b8864..94ef4e7b32 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -90,12 +90,11 @@ void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const I
 
 Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
     if(!is_data_type_quantized_asymmetric(weights->data_type()))
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
     }
     // Check if bias are present, if yes they will be embedded to the weights matrix
     const bool append_bias = (biases != nullptr);
@@ -104,7 +103,6 @@ Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, co
     {
         ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(weights->data_type()));
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases);
         ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
@@ -173,9 +171,8 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf
                                       unsigned int &mat_weights_cols, unsigned int &mat_weights_rows,
                                       unsigned int &conv_w, unsigned int &conv_h, const Size2D &dilation)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights);
 
     DataLayout data_layout = input->data_layout();
@@ -201,7 +198,6 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf
         {
             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
         }
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
         ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3));
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
@@ -287,10 +283,9 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
 
     ARM_COMPUTE_ERROR_THROW_ON(status);
 
-    _is_prepared                            = false;
-    _original_weights                       = weights;
-    const unsigned int fixed_point_position = input->info()->fixed_point_position();
-    const ITensor     *biases_to_use        = (_append_bias) ? biases : nullptr;
+    _is_prepared                 = false;
+    _original_weights            = weights;
+    const ITensor *biases_to_use = (_append_bias) ? biases : nullptr;
 
     bool run_optimised = dt == DataType::F32;
 
@@ -300,7 +295,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
         TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows };
 
         // Create tensor to store the reshaped weights
-        _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position));
+        _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt));
         _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */);
         weights = &_weights_reshaped;
     }
@@ -336,7 +331,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
             }
 
             // Create tensor to store the reshaped weights
-            _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position));
+            _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt));
             _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, _is_interleaved /* 1xW transpose */);
             weights = &_weights_reshaped;
         }
@@ -372,7 +367,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
         shape_gemm.set(1, mat_input_rows);
         const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt;
         // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
-        TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position());
+        TensorInfo info_gemm(shape_gemm, 1, gemm_data_type);
         info_gemm.set_quantization_info(output->info()->quantization_info());
         _gemm_output.allocator()->init(info_gemm);
         _memory_group.manage(&_gemm_output);
diff --git a/src/runtime/NEON/functions/NENormalizationLayer.cpp b/src/runtime/NEON/functions/NENormalizationLayer.cpp
index af98ac1f17..f00114f930 100644
--- a/src/runtime/NEON/functions/NENormalizationLayer.cpp
+++ b/src/runtime/NEON/functions/NENormalizationLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,7 +41,7 @@ void NENormalizationLayer::configure(const ITensor *input, ITensor *output, cons
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
-    TensorInfo tensor_info(input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position());
+    TensorInfo tensor_info(input->info()->tensor_shape(), 1, input->info()->data_type());
     _input_squared.allocator()->init(tensor_info);
 
     // Manage intermediate buffers