COMPMID-1016: Optimize kernel reconfiguration

Optimizes kernel reconfiguration when memory manager is used.
Note that this works only if every sub-sequent reconfigurations leads to
sizes less than the first one.

Change-Id: I08898e99929c3756147a02979b726c2380b6e11d
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125114
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

4 files changed, 17 insertions, 4 deletions

diff --git a/src/runtime/GLES_COMPUTE/GCBufferAllocator.cpp b/src/runtime/GLES_COMPUTE/GCBufferAllocator.cpp
index 50e3cc7c1c..d8f6867634 100644
--- a/src/runtime/GLES_COMPUTE/GCBufferAllocator.cpp
+++ b/src/runtime/GLES_COMPUTE/GCBufferAllocator.cpp
@@ -29,8 +29,8 @@
 
 #include <cstddef>
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 void *GCBufferAllocator::allocate(size_t size, size_t alignment)
 {
     ARM_COMPUTE_UNUSED(alignment);
@@ -48,3 +48,4 @@ void GCBufferAllocator::free(void *ptr)
     auto *gl_buffer = reinterpret_cast<GLBufferWrapper *>(ptr);
     delete gl_buffer;
 }
+} // namespace arm_compute
diff --git a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
index 0f8f8e6c94..a300033bb2 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
@@ -79,7 +79,8 @@ void GCFullyConnectedLayer::configure_fc_fc(const IGCTensor *input, const IGCTen
     _mm_kernel.configure(input, weights, output, 1.0f, false);
 }
 
-void GCFullyConnectedLayer::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, bool transpose_weights, bool are_weights_reshaped)
+void GCFullyConnectedLayer::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output,
+                                      bool transpose_weights, bool are_weights_reshaped, bool retain_internal_weights)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
@@ -141,11 +142,14 @@ void GCFullyConnectedLayer::configure(const IGCTensor *input, const IGCTensor *w
     }
 
     // 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)
+    if(!_are_weights_reshaped && !retain_internal_weights)
     {
         // Allocate the tensor for the weights reshaped
         _reshape_weights_output.allocator()->allocate();
     }
+
+    ARM_COMPUTE_ERROR_ON(retain_internal_weights && _reshape_weights_output.gc_buffer() == 0);
+    _are_weights_reshaped = _are_weights_reshaped || retain_internal_weights;
 }
 
 void GCFullyConnectedLayer::run()
@@ -158,6 +162,7 @@ void GCFullyConnectedLayer::run()
     }
 
     _memory_group.acquire();
+
     // Linearize input if it comes from a convolutional layer
     if(_is_fc_after_conv)
     {
@@ -179,5 +184,6 @@ void GCFullyConnectedLayer::run()
 
         GCScheduler::get().dispatch(_accumulate_biases_kernel);
     }
+
     _memory_group.release();
 }
diff --git a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
index 46424a59f5..9c8568a329 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
@@ -92,6 +92,10 @@ void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *
 
         TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), b->info()->fixed_point_position());
         _tmp_b.allocator()->init(info_b);
+        if(!gemm_info.reshape_b_only_on_first_run())
+        {
+            _memory_group.manage(&_tmp_b);
+        }
 
         // Configure interleave kernel
         _interleave_kernel.configure(a, &_tmp_a);
diff --git a/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
index 13213d2b54..b2e69ee8c6 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCNormalizationLayer.cpp
@@ -57,10 +57,12 @@ void GCNormalizationLayer::configure(const IGCTensor *input, IGCTensor *output,
 void GCNormalizationLayer::run()
 {
     _memory_group.acquire();
+
     GCScheduler::get().dispatch(_multiply_kernel, false);
     GCScheduler::get().memory_barrier();
     GCScheduler::get().dispatch(_border_handler, false);
     GCScheduler::get().memory_barrier();
     GCScheduler::get().dispatch(_norm_kernel, true);
+
     _memory_group.release();
 }