COMPMID-1188 - Fixed performance degradation with GEMM3D

The previous implementation of GEMM3D degradated the performance when the
input had to be reinterpreted as 3D. However if both input and output have to be
reinterpreted as 3D, we can skip the offset calculation for that specific case
and run the multi GEMM approach

Change-Id: I0d5d48add2c6ccdebfbb268ea199dd181101f3aa
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/142872
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

4 files changed, 26 insertions, 9 deletions

diff --git a/examples/graph_squeezenet_v1_1.cpp b/examples/graph_squeezenet_v1_1.cpp
index bf6324698f..e52062927b 100644
--- a/examples/graph_squeezenet_v1_1.cpp
+++ b/examples/graph_squeezenet_v1_1.cpp
@@ -87,7 +87,6 @@ public:
         graph << common_params.target
               << common_params.fast_math_hint
               << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor)))
-              << ConvolutionMethod::Direct
               << ConvolutionLayer(
                   3U, 3U, 64U,
                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_w.npy", weights_layout),
@@ -95,7 +94,6 @@ public:
                   PadStrideInfo(2, 2, 0, 0))
               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
-              << ConvolutionMethod::Default
               << ConvolutionLayer(
                   1U, 1U, 16U,
                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_w.npy", weights_layout),
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index c9e6bb34b2..79e2f8b11a 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -57,6 +57,7 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_interleaved_transposed && reshape_info.reinterpret_input_as_3d(), "The input tensor cannot be reinterpreted as 3D if is_interleaved_transposed is true");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 2 && reshape_info.reinterpret_input_as_3d(), "The input1 tensor cannot have more than 2 dimensions if input0 has to be reinterpreted as 3D");
 
     if(!is_interleaved_transposed)
     {
@@ -109,13 +110,23 @@ inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *inpu
     const DataType data_type                           = input0->data_type();
     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
+    bool           reinterpret_input_as_3d             = reshape_info.reinterpret_input_as_3d();
+    bool           reinterpret_output_as_3d            = (reshape_info.depth_output_gemm3d() != 1);
+
+    // In case both input and output have to be reinterpreted as 3D tensors,
+    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+    if(reinterpret_input_as_3d == reinterpret_output_as_3d)
+    {
+        reinterpret_input_as_3d  = false;
+        reinterpret_output_as_3d = false;
+    }
 
     // Output tensor auto inizialitation if not yet initialized
     auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info)));
 
     TensorInfo tmp_info(*output);
 
-    if(reshape_info.depth_output_gemm3d() != 1)
+    if(reinterpret_output_as_3d)
     {
         // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
         // the window needs to be constructed on the 2D collapsed version of the tensor
@@ -162,7 +173,7 @@ inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *inpu
 
         // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor
         // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
-        const int m          = reshape_info.reinterpret_input_as_3d() ? input0->tensor_shape()[1] * input0->tensor_shape()[2] : input0->tensor_shape()[1];
+        const int m          = reinterpret_input_as_3d ? input0->tensor_shape()[1] * input0->tensor_shape()[2] : input0->tensor_shape()[1];
         const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;
 
         // Create kernels according to the architecture, data type and input size.
@@ -219,6 +230,14 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
     _reinterpret_input_as_3d  = reshape_info.reinterpret_input_as_3d();
     _reinterpret_output_as_3d = (reshape_info.depth_output_gemm3d() != 1);
 
+    // In case both input and output have to be reinterpreted as 3D tensors,
+    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+    if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
+    {
+        _reinterpret_input_as_3d  = false;
+        _reinterpret_output_as_3d = false;
+    }
+
     // Check if we need to slide the matrix B
     const unsigned int num_dimensions_input0 = _reinterpret_input_as_3d ? _input0->info()->num_dimensions() - 1 : _input0->info()->num_dimensions();
 
diff --git a/tests/datasets/LargeGEMMDataset.h b/tests/datasets/LargeGEMMDataset.h
index 9c45cb0657..bbf362ca21 100644
--- a/tests/datasets/LargeGEMMDataset.h
+++ b/tests/datasets/LargeGEMMDataset.h
@@ -71,11 +71,11 @@ public:
     {
         add_config(TensorShape(923U, 143U, 3U), TensorShape(871U, 923U), TensorShape(871U, 143U, 3U), TensorShape(871U, 143U, 3U), 1.0f, 0.0f);
         add_config(TensorShape(681U, 205U, 5U), TensorShape(213U, 681U), TensorShape(213U, 205U, 5U), TensorShape(213U, 205U, 5U), 1.0f, 0.0f);
-        add_config(TensorShape(364U, 605U, 5U), TensorShape(96U, 364U), TensorShape(96U, 605U, 5U), TensorShape(96U, 605U, 5U), 1.0f, 0.0f);
+        add_config(TensorShape(364U, 605U, 5U), TensorShape(96U, 364U), TensorShape(96U, 605U, 5U), TensorShape(96U, 605U, 5U), 0.2f, 1.2f);
         add_config(TensorShape(1201U, 243U, 3U), TensorShape(128U, 1201U), TensorShape(128U, 243U, 3U), TensorShape(128U, 243U, 3U), 1.0f, 0.0f);
-        add_config(TensorShape(2305U, 13U, 13U), TensorShape(384U, 2305U), TensorShape(384U, 13U, 13U), TensorShape(384U, 13U, 13U), 1.0f, 0.0f);
-        add_config(TensorShape(1729U, 85U, 2U, 2U), TensorShape(192U, 1729U, 2U), TensorShape(192U, 85U, 2U, 2U), TensorShape(192U, 85U, 2U, 2U), 1.0f, 0.0f);
-        add_config(TensorShape(1729U, 17U, 10U, 3U), TensorShape(128U, 1729U, 3U), TensorShape(128U, 17U, 10U, 3U), TensorShape(128U, 17U, 10U, 3U), 1.0f, 0.0f);
+        add_config(TensorShape(2305U, 13U, 13U), TensorShape(384U, 2305U), TensorShape(384U, 13U, 13U), TensorShape(384U, 13U, 13U), 0.4f, 0.7f);
+        add_config(TensorShape(1729U, 85U, 2U, 2U), TensorShape(192U, 1729U), TensorShape(192U, 85U, 2U, 2U), TensorShape(192U, 85U, 2U, 2U), 1.0f, 0.0f);
+        add_config(TensorShape(1729U, 17U, 10U, 3U), TensorShape(128U, 1729U), TensorShape(128U, 17U, 10U, 3U), TensorShape(128U, 17U, 10U, 3U), 1.0f, 0.3f);
     }
 };
 } // namespace datasets
diff --git a/tests/datasets/SmallGEMMDataset.h b/tests/datasets/SmallGEMMDataset.h
index e108fcc1ca..15a3504e85 100644
--- a/tests/datasets/SmallGEMMDataset.h
+++ b/tests/datasets/SmallGEMMDataset.h
@@ -73,7 +73,7 @@ public:
         add_config(TensorShape(38U, 12U, 2U), TensorShape(21U, 38U), TensorShape(21U, 12U, 2U), TensorShape(21U, 12U, 2U), 0.2f, 1.2f);
         add_config(TensorShape(32U, 1U, 4U, 3U), TensorShape(17U, 32U), TensorShape(17U, 1U, 4U, 3U), TensorShape(17U, 1U, 4U, 3U), 0.4f, 0.7f);
         add_config(TensorShape(16U, 16U, 3U, 2U), TensorShape(8U, 16U), TensorShape(8U, 16U, 3U, 2U), TensorShape(8U, 16U, 3U, 2U), 1.0f, 0.0f);
-        add_config(TensorShape(16U, 16U, 5U, 3U), TensorShape(8U, 16U, 3U), TensorShape(8U, 16U, 5U, 3U), TensorShape(8U, 16U, 5U, 3U), 1.0f, 0.0f);
+        add_config(TensorShape(16U, 16U, 5U, 3U), TensorShape(8U, 16U), TensorShape(8U, 16U, 5U, 3U), TensorShape(8U, 16U, 5U, 3U), 1.0f, 0.3f);
     }
 };
 } // namespace datasets