COMPMID-998: Release unused trainable parameters.

Change-Id: I361a520f34080016a25bc86e1e6789777c5152c1
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/124432
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

1 files changed, 16 insertions, 7 deletions

diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 112af60f35..8d7c92bdf1 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -64,7 +64,7 @@ void CLDepthwiseConvolutionLayer3x3::run()
 
 CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer()
     : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(),
-      _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_quantized(false)
+      _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_first_run(true), _is_quantized(false), _original_weights(nullptr)
 {
 }
 
@@ -78,7 +78,9 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w
     const size_t weights_h = weights->info()->dimension(1);
     const size_t weights_z = weights->info()->dimension(2);
 
-    _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+    _is_first_run     = true;
+    _original_weights = weights;
+    _is_quantized     = is_data_type_quantized_asymmetric(input->info()->data_type());
 
     bool            append_bias = (biases != nullptr) && !_is_quantized;
     const GPUTarget gpu_target  = CLScheduler::get().target();
@@ -154,16 +156,23 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w
 
 void CLDepthwiseConvolutionLayer::run()
 {
-    CLScheduler::get().enqueue(_im2col_kernel);
+    // Run weights reshaping (Runs once for every configure)
+    if(_is_first_run)
+    {
+        ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 
-    CLScheduler::get().enqueue(_weights_reshape_kernel);
+        CLScheduler::get().enqueue(_weights_reshape_kernel);
+        CLScheduler::get().enqueue(_v2mm_weights_fill_border);
+        _is_first_run = false;
+
+        // Mark original weights tensor as unused
+        _original_weights->mark_as_unused();
+    }
 
+    CLScheduler::get().enqueue(_im2col_kernel);
     CLScheduler::get().enqueue(_v2mm_input_fill_border);
-    CLScheduler::get().enqueue(_v2mm_weights_fill_border);
     CLScheduler::get().enqueue(_v2mm_kernel);
-
     CLScheduler::get().enqueue(_vector_to_tensor_kernel);
-
     if(_is_quantized)
     {
         CLScheduler::get().enqueue(_output_stage_kernel);