COMPMID-1386: Add support for converting weights for CL.

Change-Id: I62e3ead903366baeeb1488f233a9b8b0c388c9de
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140403
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

6 files changed, 55 insertions, 12 deletions

diff --git a/tests/validation/CL/FullyConnectedLayer.cpp b/tests/validation/CL/FullyConnectedLayer.cpp
index 9958a88419..cd050e378e 100644
--- a/tests/validation/CL/FullyConnectedLayer.cpp
+++ b/tests/validation/CL/FullyConnectedLayer.cpp
@@ -96,9 +96,14 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
     const QuantizationInfo src_quantization_info     = src.info()->quantization_info();
     const QuantizationInfo weights_quantization_info = weights.info()->quantization_info();
 
+    // Create Fully Connected layer info
+    FullyConnectedLayerInfo fc_info;
+    fc_info.transpose_weights    = transpose_weights;
+    fc_info.are_weights_reshaped = !reshape_weights;
+
     // Create and configure function.
     CLFullyConnectedLayer fc;
-    fc.configure(&src, &weights, &bias, &dst, transpose_weights, !reshape_weights);
+    fc.configure(&src, &weights, &bias, &dst, fc_info);
 
     // Validate valid region
     const ValidRegion dst_valid_region = shape_to_valid_region(dst_shape);
@@ -141,12 +146,16 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(
     framework::dataset::make("Expected", { false, true, true, false, false })),
     input_info, weights_info, bias_info, output_info, transpose_weights, reshaped_weights, expected)
 {
+    // Create Fully Connected layer info
+    FullyConnectedLayerInfo fc_info;
+    fc_info.transpose_weights = transpose_weights;
+    fc_info.are_weights_reshaped = reshaped_weights;
+
     Status status = CLFullyConnectedLayer::validate(&input_info.clone()->set_is_resizable(false),
                                                     &weights_info.clone()->set_is_resizable(false),
                                                     &bias_info.clone()->set_is_resizable(false),
                                                     &output_info.clone()->set_is_resizable(false),
-                                                    transpose_weights,
-                                                    reshaped_weights);
+                                                    fc_info);
     ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
 }
 // clang-format on
diff --git a/tests/validation/GLES_COMPUTE/FullyConnectedLayer.cpp b/tests/validation/GLES_COMPUTE/FullyConnectedLayer.cpp
index 49716dc946..c82a8a1a43 100644
--- a/tests/validation/GLES_COMPUTE/FullyConnectedLayer.cpp
+++ b/tests/validation/GLES_COMPUTE/FullyConnectedLayer.cpp
@@ -87,9 +87,14 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
     ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
 
+    // Create Fully Connected layer info
+    FullyConnectedLayerInfo fc_info;
+    fc_info.transpose_weights    = transpose_weights;
+    fc_info.are_weights_reshaped = !reshape_weights;
+
     // Create and configure function.
     GCFullyConnectedLayer fc;
-    fc.configure(&src, &weights, &bias, &dst, transpose_weights, !reshape_weights);
+    fc.configure(&src, &weights, &bias, &dst, fc_info);
 
     // Validate valid region
     const ValidRegion dst_valid_region = shape_to_valid_region(dst_shape);
diff --git a/tests/validation/NEON/FullyConnectedLayer.cpp b/tests/validation/NEON/FullyConnectedLayer.cpp
index 174778b8ef..80fdf1784e 100644
--- a/tests/validation/NEON/FullyConnectedLayer.cpp
+++ b/tests/validation/NEON/FullyConnectedLayer.cpp
@@ -99,9 +99,14 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
     ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
 
+    // Create Fully Connected layer info
+    FullyConnectedLayerInfo fc_info;
+    fc_info.transpose_weights    = transpose_weights;
+    fc_info.are_weights_reshaped = !reshape_weights;
+
     // Create and configure function.
     NEFullyConnectedLayer fc;
-    fc.configure(&src, &weights, &bias, &dst, transpose_weights, !reshape_weights);
+    fc.configure(&src, &weights, &bias, &dst, fc_info);
 
     // Validate valid region
     const ValidRegion dst_valid_region = shape_to_valid_region(dst_shape);
@@ -144,7 +149,12 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(
     framework::dataset::make("Expected", { false, true, true, false, false, true })),
     input_info, weights_info, bias_info, output_info, transpose_weights, reshaped_weights, expected)
 {
-    Status status = NEFullyConnectedLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), transpose_weights, reshaped_weights);
+    // Create Fully Connected layer info
+    FullyConnectedLayerInfo fc_info;
+    fc_info.transpose_weights = transpose_weights;
+    fc_info.are_weights_reshaped = reshaped_weights;
+
+    Status status = NEFullyConnectedLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), fc_info);
     ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
 }
 // clang-format on
diff --git a/tests/validation/fixtures/FullyConnectedLayerFixture.h b/tests/validation/fixtures/FullyConnectedLayerFixture.h
index 895e43b735..18321480f8 100644
--- a/tests/validation/fixtures/FullyConnectedLayerFixture.h
+++ b/tests/validation/fixtures/FullyConnectedLayerFixture.h
@@ -130,9 +130,14 @@ protected:
         TensorType bias    = create_tensor<TensorType>(bias_shape, _bias_data_type, 1, _quantization_info);
         TensorType dst     = create_tensor<TensorType>(output_shape, _data_type, 1, _quantization_info);
 
+        // Create Fully Connected layer info
+        FullyConnectedLayerInfo fc_info;
+        fc_info.transpose_weights    = transpose_weights;
+        fc_info.are_weights_reshaped = !reshape_weights;
+
         // Create and configure function.
         FunctionType fc;
-        fc.configure(&src, &weights, &bias, &dst, transpose_weights, !reshape_weights);
+        fc.configure(&src, &weights, &bias, &dst, fc_info);
 
         ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
diff --git a/tests/validation/fixtures/UNIT/MemoryManagerFixture.h b/tests/validation/fixtures/UNIT/MemoryManagerFixture.h
index 21ad42bf77..d8e2b0b427 100644
--- a/tests/validation/fixtures/UNIT/MemoryManagerFixture.h
+++ b/tests/validation/fixtures/UNIT/MemoryManagerFixture.h
@@ -239,9 +239,13 @@ protected:
         dst.allocator()->info().set_tensor_shape(TensorShape(24U, _cur_batches)).set_is_resizable(true).extend_padding(new_dst_padding);
         dst.allocator()->info().set_is_resizable(false);
 
+        // Configure FC info
+        FullyConnectedLayerInfo fc_info;
+        fc_info.retain_internal_weights = true;
+
         // Configure functions (2nd iteration)
-        fc_layer_1.configure(&src, &w1, &b1, &fc1, true, false, true);
-        fc_layer_2.configure(&fc1, &w2, &b2, &dst, true, false, true);
+        fc_layer_1.configure(&src, &w1, &b1, &fc1, fc_info);
+        fc_layer_2.configure(&fc1, &w2, &b2, &dst, fc_info);
 
         // Fill tensors (2nd iteration)
         fill(AccessorType(src), 5);
@@ -357,6 +361,10 @@ protected:
         // Get padding requirements
         auto fc_padding = fc.allocator()->info().padding();
 
+        // Configure FC info
+        FullyConnectedLayerInfo fc_info;
+        fc_info.retain_internal_weights = true;
+
         // Run rest iterations
         for(int i = _max_batches; i >= static_cast<int>(_cur_batches); --i)
         {
@@ -368,7 +376,7 @@ protected:
             dst.allocator()->info().set_tensor_shape(TensorShape(8U, i));
 
             // Configure functions
-            fc_layer.configure(&src, &w, &b, &fc, true, false, true);
+            fc_layer.configure(&src, &w, &b, &fc, fc_info);
             smx_layer.configure(&fc, &dst);
 
             // Fill tensors
diff --git a/tests/validation/reference/ConvertFullyConnectedWeights.cpp b/tests/validation/reference/ConvertFullyConnectedWeights.cpp
index b0f537fa0c..e27846c726 100644
--- a/tests/validation/reference/ConvertFullyConnectedWeights.cpp
+++ b/tests/validation/reference/ConvertFullyConnectedWeights.cpp
@@ -36,9 +36,15 @@ SimpleTensor<T> convert_fully_connected_weights(const SimpleTensor<T> &src, cons
 {
     SimpleTensor<T> dst(src.shape(), src.data_type());
 
+    const DataLayout original_input_data_layout = (training_data_layout == DataLayout::NCHW) ? DataLayout::NHWC : DataLayout::NCHW;
+
+    const int width_idx   = get_data_layout_dimension_index(original_input_data_layout, DataLayoutDimension::WIDTH);
+    const int height_idx  = get_data_layout_dimension_index(original_input_data_layout, DataLayoutDimension::HEIGHT);
+    const int channel_idx = get_data_layout_dimension_index(original_input_data_layout, DataLayoutDimension::CHANNEL);
+
     const bool         is_nchw_to_nhwc           = training_data_layout == DataLayout::NCHW;
-    const unsigned int num_elems_per_input_plane = original_input_shape.x() * original_input_shape.y();
-    const unsigned int num_channels              = original_input_shape.z();
+    const unsigned int num_elems_per_input_plane = original_input_shape[width_idx] * original_input_shape[height_idx];
+    const unsigned int num_channels              = original_input_shape[channel_idx];
     const unsigned int factor_1                  = is_nchw_to_nhwc ? num_elems_per_input_plane : num_channels;
     const unsigned int factor_2                  = is_nchw_to_nhwc ? num_channels : num_elems_per_input_plane;