COMPIMID-553: MobileNet use case.

Change-Id: I1181abbd5785065f3d57e91844376a4b110938a9
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110701
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>

8 files changed, 55 insertions, 22 deletions

diff --git a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
index 38e367dfb7..e8882b9daf 100644
--- a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
@@ -130,7 +130,8 @@ void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *out
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batchnormalization_layer", build_opts));
 
     // Set kernel static arguments
-    unsigned int idx = 2 * num_arguments_per_3D_tensor() + 4 * num_arguments_per_1D_tensor(); // Skip the input and output parameters
+    unsigned int include_output = (output != nullptr) ? 1 : 0;
+    unsigned int idx            = (1 + include_output) * num_arguments_per_3D_tensor() + 4 * num_arguments_per_1D_tensor(); // Skip the input and output parameters
     _kernel.setArg<cl_float>(idx++, _epsilon);
 
     // Configure kernel window
@@ -160,7 +161,8 @@ void CLBatchNormalizationLayerKernel::run(const Window &window, cl::CommandQueue
     Window vector_slice = window.first_slice_window_1D();
     vector_slice.set(Window::DimX, Window::Dimension(0, 0, 0));
 
-    unsigned int idx = 2 * num_arguments_per_3D_tensor();
+    unsigned int include_output = (_output != nullptr) ? 1 : 0;
+    unsigned int idx            = (1 + include_output) * num_arguments_per_3D_tensor();
     add_1D_tensor_argument(idx, _mean, vector_slice);
     add_1D_tensor_argument(idx, _var, vector_slice);
     add_1D_tensor_argument(idx, _beta, vector_slice);
diff --git a/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
index be8fae2885..e86c55fbc0 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
@@ -37,6 +37,29 @@
 
 using namespace arm_compute;
 
+namespace
+{
+/** Calculates expected output shape dimension
+ *
+ * @param[in] Input shape
+ *
+ * @return Expected output shape
+ */
+TensorShape get_output_shape(TensorShape input_shape, TensorShape weights_shape, PadStrideInfo conv_info)
+{
+    unsigned int output_width  = 0;
+    unsigned int output_height = 0;
+
+    std::tie(output_width, output_height) = scaled_dimensions(input_shape.x(), input_shape.y(), weights_shape.x(), weights_shape.y(), conv_info);
+
+    TensorShape output_shape = input_shape;
+    output_shape.set(0, output_width);
+    output_shape.set(1, output_height);
+
+    return output_shape;
+}
+} // namespace
+
 CLDepthwiseConvolution3x3Kernel::CLDepthwiseConvolution3x3Kernel()
     : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_x(0), _conv_stride_y(0), _conv_pad_left(0), _conv_pad_top(0)
 {
@@ -50,9 +73,7 @@ BorderSize CLDepthwiseConvolution3x3Kernel::border_size() const
 void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
 
     if(biases != nullptr)
@@ -69,13 +90,18 @@ void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, const IC
         ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
     }
 
-    std::pair<unsigned int, unsigned int> expected_output = scaled_dimensions(input->info()->tensor_shape().x(), input->info()->tensor_shape().y(),
-                                                                              weights->info()->tensor_shape().x(), weights->info()->tensor_shape().y(),
-                                                                              conv_info);
+    // Get convolved dimensions
+    TensorShape output_shape = get_output_shape(input->info()->tensor_shape(), weights->info()->tensor_shape(), conv_info);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(),
+                       output_shape,
+                       1,
+                       input->info()->data_type(),
+                       input->info()->fixed_point_position(),
+                       input->info()->quantization_info());
 
-    ARM_COMPUTE_UNUSED(expected_output);
-    ARM_COMPUTE_ERROR_ON(expected_output.first != output->info()->tensor_shape().x());
-    ARM_COMPUTE_ERROR_ON(expected_output.second != output->info()->tensor_shape().y());
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
 
     _input         = input;
     _output        = output;
diff --git a/src/graph/nodes/ConvolutionLayer.cpp b/src/graph/nodes/ConvolutionLayer.cpp
index a7236fc78a..ae4a8d7e6b 100644
--- a/src/graph/nodes/ConvolutionLayer.cpp
+++ b/src/graph/nodes/ConvolutionLayer.cpp
@@ -189,7 +189,7 @@ std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_node(Graph
                                      in->info()->data_type(),
                                      in->info()->fixed_point_position()));
     }
-    if(_biases.tensor() == nullptr)
+    if(_biases.has_accessor() && _biases.tensor() == nullptr)
     {
         _biases.set_info(TensorInfo(TensorShape(_ofm), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
     }
@@ -200,11 +200,14 @@ std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_node(Graph
 
     // Check if the weights and biases are loaded
     bool weights_are_loaded = _weights.tensor() != nullptr;
-    bool biases_are_loaded  = _weights.tensor() != nullptr;
+    bool biases_are_loaded  = _biases.has_accessor() ? _biases.tensor() != nullptr : true;
 
     // Set bias and weights target
     _weights.set_target(_target_hint);
-    _biases.set_target(_target_hint);
+    if(_biases.has_accessor())
+    {
+        _biases.set_target(_target_hint);
+    }
 
     // Calculate output shape
     TensorShape output_shape = calculate_convolution_layer_output_shape(in->info()->tensor_shape(), _weights.info().tensor_shape(), _conv_info);
diff --git a/src/graph/nodes/DepthwiseConvolutionLayer.cpp b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
index 1c006d61db..ceac2a2def 100644
--- a/src/graph/nodes/DepthwiseConvolutionLayer.cpp
+++ b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
@@ -51,10 +51,13 @@ std::unique_ptr<arm_compute::IFunction> DepthwiseConvolutionLayer::instantiate_n
     }
 
     bool weights_is_loaded = _weights.tensor() != nullptr;
-    bool biases_is_loaded  = _biases.has_accessor() ? _biases.tensor() != nullptr : false;
+    bool biases_is_loaded  = _biases.has_accessor() ? _biases.tensor() != nullptr : true;
 
     _weights.set_target(_target_hint);
-    _biases.set_target(_target_hint);
+    if(_biases.has_accessor())
+    {
+        _biases.set_target(_target_hint);
+    }
 
     // Create node context
     NodeContext node_ctx(OperationType::DepthwiseConvolutionLayer);
diff --git a/src/graph/nodes/ReshapeLayer.cpp b/src/graph/nodes/ReshapeLayer.cpp
index 4967534879..bbe0739e64 100644
--- a/src/graph/nodes/ReshapeLayer.cpp
+++ b/src/graph/nodes/ReshapeLayer.cpp
@@ -47,11 +47,11 @@ std::unique_ptr<arm_compute::IFunction> ReshapeLayer::instantiate_node(GraphCont
     arm_compute::auto_init_if_empty(*out->info(), _shape, 1, in->info()->data_type(), in->info()->fixed_point_position());
 
     // Create node context
-    NodeContext node_ctx(OperationType::QuantizationLayer);
+    NodeContext node_ctx(OperationType::ReshapeLayer);
     node_ctx.set_target(_target_hint);
     node_ctx.add_input(in);
     node_ctx.add_output(out);
 
     // Get function
-    return OperationRegistry::get().find_operation(OperationType::QuantizationLayer, _target_hint)->configure(node_ctx);
+    return OperationRegistry::get().find_operation(OperationType::ReshapeLayer, _target_hint)->configure(node_ctx);
 }
diff --git a/src/graph/operations/CLSimpleOperations.cpp b/src/graph/operations/CLSimpleOperations.cpp
index 881f4910ad..647f88f0e2 100644
--- a/src/graph/operations/CLSimpleOperations.cpp
+++ b/src/graph/operations/CLSimpleOperations.cpp
@@ -138,7 +138,7 @@ REGISTER_SIMPLE_OPERATION(CLDepthConvertLayerOperation, OPENCL, OperationType::D
 /* DepthwiseConvolutionLayer Layer */
 REGISTER_SIMPLE_OPERATION(CLDepthwiseConvolutionOperation, OPENCL, OperationType::DepthwiseConvolutionLayer)
 {
-    ARM_COMPUTE_ERROR_ON(ctx.num_inputs() != 2 || ctx.num_inputs() != 3);
+    ARM_COMPUTE_ERROR_ON(ctx.num_inputs() != 2 && ctx.num_inputs() != 3);
     ARM_COMPUTE_ERROR_ON(ctx.num_outputs() != 1);
     ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ICLTensor *>(ctx.input(0)) == nullptr);
     ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ICLTensor *>(ctx.output(0)) == nullptr);
diff --git a/src/graph/operations/NESimpleOperations.cpp b/src/graph/operations/NESimpleOperations.cpp
index c77aeeca11..f234341cec 100644
--- a/src/graph/operations/NESimpleOperations.cpp
+++ b/src/graph/operations/NESimpleOperations.cpp
@@ -138,7 +138,7 @@ REGISTER_SIMPLE_OPERATION(NEDepthConvertLayerOperation, NEON, OperationType::Dep
 /* DepthwiseConvolutionLayer Layer */
 REGISTER_SIMPLE_OPERATION(NEDepthwiseConvolutionOperation, NEON, OperationType::DepthwiseConvolutionLayer)
 {
-    ARM_COMPUTE_ERROR_ON(ctx.num_inputs() != 2 || ctx.num_inputs() != 3);
+    ARM_COMPUTE_ERROR_ON(ctx.num_inputs() != 2 && ctx.num_inputs() != 3);
     ARM_COMPUTE_ERROR_ON(ctx.num_outputs() != 1);
     ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ITensor *>(ctx.input(0)) == nullptr);
     ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ITensor *>(ctx.output(0)) == nullptr);
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolution.cpp b/src/runtime/CL/functions/CLDepthwiseConvolution.cpp
index a701391c44..81149508dd 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolution.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolution.cpp
@@ -38,8 +38,7 @@ CLDepthwiseConvolution3x3::CLDepthwiseConvolution3x3()
 void CLDepthwiseConvolution3x3::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
 
     _kernel.set_target(CLScheduler::get().target());
     _kernel.configure(input, weights, biases, output, conv_info);