COMPMID-477 - Optimized batched case in CLConvolutionLayer

Change-Id: I4ef18f49f1da0cb816aaa0762466b940792c15ed
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/84162
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 71 insertions, 168 deletions

diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 66a858d3ed..f7cea551f6 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -26,217 +26,127 @@
 #include "arm_compute/core/Size2D.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
+#include "support/ToolchainSupport.h"
 
 #include <algorithm>
-#include <cmath>
 
-namespace arm_compute
+using namespace arm_compute;
+
+void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output)
 {
-CLFullyConnectedLayerReshapeWeights::CLFullyConnectedLayerReshapeWeights()
-    : _transpose_kernel(), _transpose1xW_kernel(), _transpose_output(), _transpose_weights(false), _is_batched_fc_layer(false)
+    auto k = arm_compute::support::cpp14::make_unique<CLTransposeKernel>();
+    k->configure(input, output);
+    _kernel = std::move(k);
+}
+
+CLFullyConnectedLayer::CLFullyConnectedLayer()
+    : _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)
 {
 }
 
-void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output, bool transpose_weights, bool is_batched_fc_layer)
+void CLFullyConnectedLayer::configure_conv_fc(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() > 2);
-    ARM_COMPUTE_ERROR_ON(output == nullptr);
-    ARM_COMPUTE_ERROR_ON(!transpose_weights && !is_batched_fc_layer);
+    ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
 
-    const DataType data_type            = input->info()->data_type();
+    const DataType dt                   = input->info()->data_type();
     const int      fixed_point_position = input->info()->fixed_point_position();
 
-    _transpose_weights   = transpose_weights;
-    _is_batched_fc_layer = is_batched_fc_layer;
+    // If the fully connected layer is called after a convolution layer, the input tensor must be linearized
 
-    // Check if we need to transpose the weights
-    if(_transpose_weights)
-    {
-        if(_is_batched_fc_layer)
-        {
-            // Initialize the output tensor for transpose
-            TensorShape shape_transposed(input->info()->dimension(1), input->info()->dimension(0));
-            _transpose_output.allocator()->init(TensorInfo(shape_transposed, 1, data_type, fixed_point_position));
-            _transpose_kernel.configure(input, &_transpose_output);
+    // 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));
 
-            // Configure transpose 1xW kernel
-            _transpose1xW_kernel.configure(&_transpose_output, output);
+    // Configure im2col kernel
+    _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
 
-            // Allocate temporary tensor used for transposing the weights
-            _transpose_output.allocator()->allocate();
-        }
-        else
-        {
-            _transpose_kernel.configure(input, output);
-        }
-    }
-    else
-    {
-        if(_is_batched_fc_layer)
-        {
-            // Configure transpose 1xW kernel
-            _transpose1xW_kernel.configure(input, output);
-        }
-        else
-        {
-            ARM_COMPUTE_ERROR("Configuration transpose_weights=false & is_batched_fc_layer=false not supported");
-        }
-    }
-}
-
-void CLFullyConnectedLayerReshapeWeights::run()
-{
-    if(_transpose_weights)
-    {
-        CLScheduler::get().enqueue(_transpose_kernel, _is_batched_fc_layer);
-    }
+    // Configure matrix multiply kernel
+    _mm_kernel.configure(&_im2col_output, weights, output, 1.0f, false);
 
-    if(_is_batched_fc_layer)
-    {
-        CLScheduler::get().enqueue(_transpose1xW_kernel);
-    }
+    // Allocate the output tensor for im2col once all the configure methods have been called
+    _im2col_output.allocator()->allocate();
 }
 
-CLFullyConnectedLayer::CLFullyConnectedLayer()
-    : _im2col_kernel(), _reshape_weights_kernel(), _interleave4x4_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _interleave4x4_output(), _reshape_weights_output(),
-      _are_weights_reshaped(false), _is_batched_fc_layer(false), _linearize_input(false), _accumulate_biases(false)
+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.configure(input, weights, output, 1.0f, false);
 }
 
 void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped)
 {
-    // With the Fully Connected layer we can have 4 different cases:
-    //  1) Convolution layer -> Fully Connected layer without batches
-    //  2) Fully Connected layer -> Fully Connected layer without batches
-    //  3) Convolution layer -> Fully Connected layer with batches
-    //  4) Fully Connected layer -> Fully Connected layer with batches
-
-    // Expected shape before transpose and reshaping
-    // Input: In x B (In and B can be multi-dimensional)
-    // Weights: flat(In) x Out
-    // Biases: Out
-    // Output: Out x B (B can be multi-dimensional)
-
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, weights, output);
+    ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() != 2);
 
-    const DataType data_type            = input->info()->data_type();
-    const int      fixed_point_position = input->info()->fixed_point_position();
-    const int      num_batch_dimensions = std::max(0, static_cast<int>(output->info()->tensor_shape().num_dimensions()) - 1);
-    const int      num_input_dimensions = input->info()->tensor_shape().num_dimensions() - num_batch_dimensions;
-    const size_t   linear_input_size    = input->info()->tensor_shape().total_size_lower(num_input_dimensions);
-
-    _linearize_input      = input->info()->tensor_shape().x() != linear_input_size;
-    _are_weights_reshaped = are_weights_reshaped;
-    _accumulate_biases    = biases != nullptr;
-    _is_batched_fc_layer  = num_batch_dimensions > 0;
-
-    // Check if number of batches match
-    ARM_COMPUTE_ERROR_ON(input->info()->tensor_shape().total_size_upper(num_input_dimensions) != output->info()->tensor_shape().total_size_upper(1));
-    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;
 
-    const size_t     interleave_width = 16 / input->info()->element_size();
-    const ICLTensor *weights_to_use   = weights;
-
-    if(!are_weights_reshaped && (transpose_weights || _is_batched_fc_layer))
+    if(biases != nullptr)
     {
-        weights_to_use = &_reshape_weights_output;
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+
+        _accumulate_biases = true;
 
-        TensorShape reshaped_weights_shape(weights->info()->tensor_shape());
+        // Configure accumulate biases kernel
+        _accumulate_biases_kernel.configure(output, biases);
+    }
 
-        // Transpose weights if the user hasn't done it
-        if(transpose_weights)
-        {
-            const size_t shape_x = reshaped_weights_shape.x();
-            reshaped_weights_shape.set(0, reshaped_weights_shape.y());
-            reshaped_weights_shape.set(1, shape_x);
-        }
+    // With the Fully Connected layer we can have 4 different cases:
+    //  1) Convolution layer -> Fully Connected layer without batches
+    //  2) Fully Connected layer -> Fully Connected layer without batches
+    //  3) Convolution layer -> Fully Connected layer with batches
+    //  4) Fully Connected layer -> Fully Connected layer with batches
 
-        // If the we run multiple batches we need 1xW transpose, too.
-        if(_is_batched_fc_layer)
-        {
-            const float shape_x = reshaped_weights_shape.x();
-            reshaped_weights_shape.set(0, reshaped_weights_shape.y() * interleave_width);
-            reshaped_weights_shape.set(1, static_cast<unsigned int>(std::ceil(shape_x / interleave_width)));
-        }
+    const ICLTensor *weights_to_use = weights;
 
-        _reshape_weights_output.allocator()->init(TensorInfo(reshaped_weights_shape, 1, data_type, fixed_point_position));
+    if(!_are_weights_reshaped)
+    {
+        weights_to_use = &_reshape_weights_output;
 
         // Reshape the weights
-        _reshape_weights_kernel.configure(weights, &_reshape_weights_output, transpose_weights, _is_batched_fc_layer);
+        _reshape_weights_kernel.configure(weights, &_reshape_weights_output);
     }
 
-    // Check correct shape of weights
-    if(_is_batched_fc_layer)
+    // Check if we have a fully connected layer with batches
+    const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
+
+    if(is_batched_fc_layer)
     {
-        // Transpose + Transpose1xW
-        ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().x() != linear_input_size * interleave_width);
-        ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().y() != static_cast<unsigned int>(std::ceil(static_cast<float>(output->info()->tensor_shape().x()) / interleave_width)));
+        _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3,
+                                                                                  input->info()->tensor_shape().cend(),
+                                                                                  output->info()->tensor_shape().cbegin() + 1));
     }
     else
     {
-        // Transpose
-        ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().x() != output->info()->tensor_shape().x());
-        ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().y() != linear_input_size);
+        _is_fc_after_conv = input->info()->num_dimensions() > 1;
     }
 
-    const ICLTensor *multiply_input = input;
-
-    if(_linearize_input)
+    if(_is_fc_after_conv)
     {
-        TensorShape shape_im2col(input->info()->tensor_shape());
-        shape_im2col.collapse(num_input_dimensions);
-        _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, data_type, fixed_point_position));
-
-        // Configure im2col kernel
-        _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
-
-        multiply_input = &_im2col_output;
+        // Fully Connected layer after a Convolution Layer without batches
+        configure_conv_fc(input, weights_to_use, output);
     }
-
-    if(_is_batched_fc_layer)
-    {
-        TensorShape shape_interleaved(multiply_input->info()->tensor_shape());
-        shape_interleaved.set(0, shape_interleaved.x() * 4);
-        shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
-        _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, data_type, fixed_point_position));
-
-        // Configure interleave4x4 kernel
-        _interleave4x4_kernel.configure(multiply_input, &_interleave4x4_output);
-
-        multiply_input = &_interleave4x4_output;
-    }
-
-    // Configure matrix multiply kernel
-    _mm_kernel.configure(multiply_input, weights_to_use, output, 1.0f);
-
-    if(_accumulate_biases)
+    else
     {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
-        ARM_COMPUTE_ERROR_ON(biases->info()->tensor_shape().x() != output->info()->tensor_shape().x());
-
-        // Configure accumulate biases kernel
-        _accumulate_biases_kernel.configure(output, biases);
+        // Fully Connected layer after a Fully Connected Layer without batches
+        configure_fc_fc(input, weights_to_use, output);
     }
 
     // Allocate the transpose tensor if the are_weights_reshaped flag is false and once all the configure methods have been called
-    if(!are_weights_reshaped && (transpose_weights || _is_batched_fc_layer))
+    if(!_are_weights_reshaped)
     {
         // Allocate the tensor for the weights reshaped
         _reshape_weights_output.allocator()->allocate();
     }
-
-    if(_linearize_input)
-    {
-        _im2col_output.allocator()->allocate();
-    }
-
-    if(_is_batched_fc_layer)
-    {
-        _interleave4x4_output.allocator()->allocate();
-    }
 }
 
 void CLFullyConnectedLayer::run()
@@ -249,17 +159,11 @@ void CLFullyConnectedLayer::run()
     }
 
     // Linearize input if it comes from a convolutional layer
-    if(_linearize_input)
+    if(_is_fc_after_conv)
     {
         CLScheduler::get().enqueue(_im2col_kernel, false);
     }
 
-    // Interleave input
-    if(_is_batched_fc_layer)
-    {
-        CLScheduler::get().enqueue(_interleave4x4_kernel, false);
-    }
-
     // Run matrix multiply
     CLScheduler::get().enqueue(_mm_kernel, !_accumulate_biases);
 
@@ -269,4 +173,3 @@ void CLFullyConnectedLayer::run()
         CLScheduler::get().enqueue(_accumulate_biases_kernel);
     }
 }
-} // namespace arm_compute