commit 52ba29e936b8e711e8acdfe819e36f884d4f3fe1
author Jenkins <bsgcomp@arm.com> Wed Aug 29 15:32:11 2018 +0000
committer Michele Di Giorgio <michele.digiorgio@arm.com> Thu Aug 30 17:15:09 2018 +0100
tree 23a8e15208ad7fcf07a37ed9119979cde01dcdc8
parent e2542c9f35ca427286822cd0c9296f49914f78b0

arm_compute v18.08

src/runtime/CL/functions/CLLocallyConnectedLayer.cpp

diff --git a/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp b/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
index 986fe00..40bf032 100644
--- a/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
@@ -48,7 +48,10 @@
     // Get convolved dimensions
     unsigned int conv_w = 0;
     unsigned int conv_h = 0;
-    std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
+    std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0),
+                                                 input->dimension(1),
+                                                 kernel_width,
+                                                 kernel_height,
                                                  conv_info);
 
     const size_t mat_weights_cols = weights->dimension(3);
@@ -61,9 +64,12 @@
     const size_t mat_input_rows = conv_w * conv_h;
 
     shape_im2col = input->tensor_shape();
+    if(shape_im2col.num_dimensions() >= 3)
+    {
+        shape_im2col.remove_dimension(2);
+    }
     shape_im2col.set(0, mat_input_cols);
     shape_im2col.set(1, mat_input_rows);
-    shape_im2col.set(2, 1);
 
     shape_gemm = shape_im2col;
     shape_gemm.set(0, mat_weights_cols);
@@ -73,7 +79,7 @@
 
 CLLocallyConnectedLayer::CLLocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _weights_reshape_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _weights_reshaped(), _gemm_output(),
-      _is_first_run(false), _original_weights(nullptr)
+      _is_prepared(false), _original_weights(nullptr)
 {
 }
 
@@ -128,7 +134,7 @@
 
     bool _has_bias    = (biases != nullptr);
     _original_weights = weights;
-    _is_first_run     = true;
+    _is_prepared      = false;
 
     const unsigned int kernel_width  = weights->info()->dimension(0);
     const unsigned int kernel_height = weights->info()->dimension(1);
@@ -160,24 +166,15 @@
     _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
 
     // Allocate intermediate tensors
-    _weights_reshaped.allocator()->allocate();
     _input_im2col_reshaped.allocator()->allocate();
     _gemm_output.allocator()->allocate();
+
+    CLScheduler::get().tune_kernel_static(_input_im2col_kernel);
 }
 
 void CLLocallyConnectedLayer::run()
 {
-    // Run weights reshaping (Runs once for every configure)
-    if(_is_first_run)
-    {
-        ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
-
-        _is_first_run = false;
-        CLScheduler::get().enqueue(_weights_reshape_kernel);
-
-        // Mark original weights tensor as unused
-        _original_weights->mark_as_unused();
-    }
+    prepare();
 
     _memory_group.acquire();
 
@@ -192,3 +189,19 @@
 
     _memory_group.release();
 }
+
+void CLLocallyConnectedLayer::prepare()
+{
+    if(!_is_prepared)
+    {
+        ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
+        // Run weights reshaping and mark original weights tensor as unused
+        _weights_reshaped.allocator()->allocate();
+        CLScheduler::get().enqueue(_weights_reshape_kernel);
+        _original_weights->mark_as_unused();
+
+        CLScheduler::get().queue().finish();
+        _is_prepared = true;
+    }
+}