COMPMID-661: QASYMM8 support for fully connected layer.

Change-Id: I70e04d3a175ba366432ada98e9ca893c9f81b260
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111094
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 70 insertions, 27 deletions

diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 72d374e9c2..88aaf1cae8 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -25,6 +25,7 @@
 
 #include "arm_compute/core/Size2D.h"
 #include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "support/ToolchainSupport.h"
 
@@ -40,70 +41,87 @@ void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLT
 }
 
 CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _reshape_weights_output(),
-      _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false)
+    : _memory_group(memory_manager), _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _accumulate_biases_kernel(), _im2col_output(),
+      _gemmlowp_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false), _is_quantized(false)
 {
 }
 
-void CLFullyConnectedLayer::configure_conv_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, const GPUTarget gpu_target)
+void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed)
 {
-    ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
+    if(_is_quantized)
+    {
+        // Extract and negate input and weights offset
+        QuantizationInfo input_quantization_info   = input->info()->quantization_info();
+        QuantizationInfo weights_quantization_info = weights->info()->quantization_info();
+        input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset));
+        weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
+        // Configure gemmlowp function
+        _mm_gemmlowp.configure(input, weights, output);
+    }
+    else
+    {
+        // Configure matrix multiply kernel
+        _mm_kernel.set_target(CLScheduler::get().target());
+        _mm_kernel.configure(input, weights, output, 1.f, is_interleaved_transposed);
+    }
+}
 
-    const DataType dt                   = input->info()->data_type();
-    const int      fixed_point_position = input->info()->fixed_point_position();
+void CLFullyConnectedLayer::configure_conv_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
 
     // If the fully connected layer is called after a convolution layer, the input tensor must be linearized
 
     // Initialize output tensor for im2col
-    TensorShape shape_im2col;
-    shape_im2col.set(0, input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2));
-    shape_im2col.set(1, input->info()->dimension(3));
-    shape_im2col.set(2, input->info()->dimension(4));
-    shape_im2col.set(3, input->info()->dimension(5));
-    _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, dt, fixed_point_position));
+    TensorShape shape_im2col = input->info()->tensor_shape();
+    shape_im2col.collapse(3);
+    _im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
 
     // Configure im2col kernel
     _memory_group.manage(&_im2col_output);
     _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
 
     // Configure matrix multiply kernel
-    _mm_kernel.set_target(gpu_target);
-    _mm_kernel.configure(&_im2col_output, weights, output, 1.0f, false);
+    configure_mm(&_im2col_output, weights, output, false);
 
     // Allocate the output tensor for im2col once all the configure methods have been called
     _im2col_output.allocator()->allocate();
 }
 
-void CLFullyConnectedLayer::configure_fc_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, const GPUTarget gpu_target)
+void CLFullyConnectedLayer::configure_fc_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output)
 {
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
 
     // Configure matrix multiply kernel
-    _mm_kernel.set_target(gpu_target);
-    _mm_kernel.configure(input, weights, output, 1.0f, false);
+    configure_mm(input, weights, output, false);
 }
 
 void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 2);
 
     _are_weights_reshaped = transpose_weights ? are_weights_reshaped : true;
     _is_fc_after_conv     = true;
     _accumulate_biases    = false;
+    _is_quantized         = is_data_type_quantized_asymmetric(input->info()->data_type());
 
-    // Get GPU target
-    const GPUTarget gpu_target = CLScheduler::get().target();
+    // Configure gemmlowp output
+    if(_is_quantized)
+    {
+        _gemmlowp_output.allocator()->init(output->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
+    }
 
-    if(biases != nullptr)
+    // Configure accumulate biases kernel for non quantized asymmetric types
+    if(biases != nullptr && !_is_quantized)
     {
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
 
         _accumulate_biases = true;
 
         // Configure accumulate biases kernel
-        _accumulate_biases_kernel.set_target(gpu_target);
+        _accumulate_biases_kernel.set_target(CLScheduler::get().target());
         _accumulate_biases_kernel.configure(output, biases);
     }
 
@@ -137,15 +155,26 @@ void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *w
         _is_fc_after_conv = input->info()->num_dimensions() > 1;
     }
 
+    ICLTensor *tmp_output = (_is_quantized) ? &_gemmlowp_output : output;
     if(_is_fc_after_conv)
     {
         // Fully Connected layer after a Convolution Layer without batches
-        configure_conv_fc(input, weights_to_use, output, gpu_target);
+        configure_conv_fc(input, weights_to_use, tmp_output);
     }
     else
     {
         // Fully Connected layer after a Fully Connected Layer without batches
-        configure_fc_fc(input, weights_to_use, output, gpu_target);
+        configure_fc_fc(input, weights_to_use, tmp_output);
+    }
+
+    // Configure output stage for asymmetric quantized types
+    if(_is_quantized)
+    {
+        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale;
+        int   output_multiplier, output_shift;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+        _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output->info()->quantization_info().offset, output_multiplier, output_shift);
+        _gemmlowp_output.allocator()->allocate();
     }
 
     // Allocate the transpose tensor if the are_weights_reshaped flag is false and once all the configure methods have been called
@@ -174,12 +203,26 @@ void CLFullyConnectedLayer::run()
     }
 
     // Run matrix multiply
-    CLScheduler::get().enqueue(_mm_kernel, !_accumulate_biases);
+    if(_is_quantized)
+    {
+        _mm_gemmlowp.run();
+    }
+    else
+    {
+        CLScheduler::get().enqueue(_mm_kernel, !_accumulate_biases);
+    }
 
     // Accumulate biases if provided
-    if(_accumulate_biases)
+    if(_is_quantized)
+    {
+        _gemmlowp_output_stage.run();
+    }
+    else
     {
-        CLScheduler::get().enqueue(_accumulate_biases_kernel);
+        if(_accumulate_biases)
+        {
+            CLScheduler::get().enqueue(_accumulate_biases_kernel);
+        }
     }
 
     _memory_group.release();