From b91e34c9837756c9ee45917e13fb6a6cb901f795 Mon Sep 17 00:00:00 2001
From: Michalis Spyrou <michalis.spyrou@arm.com>
Date: Wed, 20 Dec 2017 15:50:55 +0000
Subject: COMPMID-746 Allow NEDirectConvolution to work without biases for QS.

Renamed BiasAccumulateKernel to OutputStage. If no bias is provided
when the input is quantized, the kernel simply downscales the input.
Throw error if no bias is provided and input is floating point.

Change-Id: I645a4ee9c6014b0547778fdd92c9ec72ef2f0aab
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/114158
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
---
 .../NEON/functions/NEDirectConvolutionLayer.cpp    | 29 +++++++++-------------
 1 file changed, 12 insertions(+), 17 deletions(-)

(limited to 'src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp')

diff --git a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
index ef5d987832..c26c99a0f8 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)), _accumulate_bias_kernel(), _conv_kernel(), _input_border_handler(), _accumulator(), _has_bias(false)
+    : _memory_group(std::move(memory_manager)), _output_stage_kernel(), _conv_kernel(), _input_border_handler(), _accumulator(), _has_bias(false), _is_fixed_point(false)
 {
 }
 
@@ -50,17 +50,16 @@ void NEDirectConvolutionLayer::configure(ITensor *input, const ITensor *weights,
     _has_bias = (bias != nullptr);
 
     // Allocate the intermediate accumulator tensor in case of fixed point input
-    if(is_data_type_fixed_point(input->info()->data_type()))
+    _is_fixed_point = is_data_type_fixed_point(input->info()->data_type());
+    if(_is_fixed_point)
     {
         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);
-        // TODO (COMPMID-746): Fix accumulate biases to just down-cast when no bias is provided
-        if(_has_bias)
-        {
-            _accumulate_bias_kernel.configure(&_accumulator, bias, output);
-        }
+
+        // When no bias is provided, we need to downscale the accumulator tensor
+        _output_stage_kernel.configure(&_accumulator, bias, output);
         _accumulator.allocator()->allocate();
     }
     else
@@ -68,7 +67,7 @@ void NEDirectConvolutionLayer::configure(ITensor *input, const ITensor *weights,
         _conv_kernel.configure(input, weights, output, conv_info);
         if(_has_bias)
         {
-            _accumulate_bias_kernel.configure(output, bias);
+            _output_stage_kernel.configure(output, bias);
         }
     }
 
@@ -91,20 +90,17 @@ Status NEDirectConvolutionLayer::validate(const ITensorInfo *input, const ITenso
     // Validate Convolution kernel
     ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerKernel::validate(input, weights, &accumulator, conv_info));
 
-    // Validate bias
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((bias == nullptr) && is_data_type_fixed_point(data_type),
-                                    "Biases should be provided for fixed point inputs");
     if(bias != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, bias);
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->dimension(0) != weights->dimension(3),
                                         "Biases size and number of input feature maps should match");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->num_dimensions() > 1, "Biases should be one dimensional");
-
-        // Validate bias kernel
-        ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerBiasAccumulateKernel::validate(&accumulator, bias, output));
     }
 
+    // Validate bias kernel
+    ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&accumulator, bias, output));
+
     return Status{};
 }
 
@@ -115,10 +111,9 @@ void NEDirectConvolutionLayer::run()
     _memory_group.acquire();
 
     NEScheduler::get().schedule(&_conv_kernel, Window::DimZ);
-    if(_has_bias)
+    if(_has_bias || _is_fixed_point)
     {
-        NEScheduler::get().schedule(&_accumulate_bias_kernel, Window::DimY);
+        NEScheduler::get().schedule(&_output_stage_kernel, Window::DimY);
     }
-
     _memory_group.release();
 }
-- 
cgit v1.2.1