Retain weights in ClGemm when reconfiguring the operator with retention

- Reduces the size of the WeightRetentions tests
- Ensure the weights are retained the only the Src,Dst are updated

Resolves: COMPMID-4775

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: If12daa4bd4bf89ec28faa743fb7291895cbe7b7c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6121
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 23 insertions, 14 deletions

diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index c719a667a7..6836c12022 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -76,7 +76,7 @@ void CLFullyConnectedLayer::configure(const CLCompileContext &compile_context, c
 
     _impl->op               = std::make_unique<opencl::ClFullyConnected>();
     _impl->original_weights = weights;
-    _impl->is_prepared      = false;
+    _impl->is_prepared      = fc_info.retain_internal_weights;
 
     _impl->op->configure(compile_context, input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), fc_info);
 
@@ -85,9 +85,17 @@ void CLFullyConnectedLayer::configure(const CLCompileContext &compile_context, c
         _impl->weights_manager->manage(weights);
     }
 
-    _impl->aux_mem_req = _impl->op->workspace();
-    _impl->run_pack    = { { ACL_SRC_0, input }, { ACL_SRC_1, weights }, { ACL_SRC_2, biases }, { ACL_DST, output } };
-    _impl->workspace   = manage_workspace<CLTensor>(_impl->aux_mem_req, _impl->memory_group, _impl->run_pack, _impl->run_pack);
+    if(!_impl->is_prepared)
+    {
+        _impl->aux_mem_req = _impl->op->workspace();
+        _impl->run_pack    = { { ACL_SRC_0, input }, { ACL_SRC_1, weights }, { ACL_SRC_2, biases }, { ACL_DST, output } };
+        _impl->workspace   = manage_workspace<CLTensor>(_impl->aux_mem_req, _impl->memory_group, _impl->run_pack, _impl->run_pack);
+    }
+    else
+    {
+        _impl->run_pack.add_tensor(ACL_SRC_0, input);
+        _impl->run_pack.add_tensor(ACL_DST, output);
+    }
 }
 
 Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
diff --git a/src/runtime/gpu/cl/operators/ClGemm.cpp b/src/runtime/gpu/cl/operators/ClGemm.cpp
index 2792dc470d..59bbabba26 100644
--- a/src/runtime/gpu/cl/operators/ClGemm.cpp
+++ b/src/runtime/gpu/cl/operators/ClGemm.cpp
@@ -564,6 +564,7 @@ void ClGemm::configure(const CLCompileContext &compile_context, ITensorInfo *a,
 
     // Check if we need to reshape the matrix B only on the first run
     _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
+    _is_prepared                 = gemm_info.retain_internal_weights();
 
     bool               reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
     const unsigned int m                       = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
diff --git a/tests/validation/fixtures/UNIT/WeightsRetentionFixture.h b/tests/validation/fixtures/UNIT/WeightsRetentionFixture.h
index af9f776ebc..f5e6071340 100644
--- a/tests/validation/fixtures/UNIT/WeightsRetentionFixture.h
+++ b/tests/validation/fixtures/UNIT/WeightsRetentionFixture.h
@@ -74,10 +74,10 @@ protected:
     TensorType compute_target()
     {
         // Create tensors
-        TensorType w1  = create_tensor<TensorType>(TensorShape(180000U, 150U), DataType::F32, 1);
-        TensorType b1  = create_tensor<TensorType>(TensorShape(150U), DataType::F32, 1);
-        TensorType src = create_tensor<TensorType>(TensorShape(1U, 150U, 1200U, _max_batches), DataType::F32, 1);
-        TensorType dst = create_tensor<TensorType>(TensorShape(150U, _max_batches), DataType::F32, 1);
+        TensorType w1  = create_tensor<TensorType>(TensorShape(6000U, 15U), DataType::F32, 1);
+        TensorType b1  = create_tensor<TensorType>(TensorShape(15U), DataType::F32, 1);
+        TensorType src = create_tensor<TensorType>(TensorShape(1U, 15U, 400U, _max_batches), DataType::F32, 1);
+        TensorType dst = create_tensor<TensorType>(TensorShape(15U, _max_batches), DataType::F32, 1);
 
         // Create and configure function
         FullyConnectedFunction fc_layer_1;
@@ -105,9 +105,9 @@ protected:
         int  diff            = _max_batches - _cur_batches;
         auto new_src_padding = PaddingSize(src_padding.top, src_padding.right, src_padding.bottom + diff, src_padding.left);
         auto new_dst_padding = PaddingSize(dst_padding.top, dst_padding.right, dst_padding.bottom + diff, dst_padding.left);
-        src.allocator()->info().set_tensor_shape(TensorShape(1U, 150U, 1200U, _cur_batches)).set_is_resizable(true).extend_padding(new_src_padding);
+        src.allocator()->info().set_tensor_shape(TensorShape(1U, 15U, 400U, _cur_batches)).set_is_resizable(true).extend_padding(new_src_padding);
         src.allocator()->info().set_is_resizable(false);
-        dst.allocator()->info().set_tensor_shape(TensorShape(150U, _cur_batches)).set_is_resizable(true).extend_padding(new_dst_padding);
+        dst.allocator()->info().set_tensor_shape(TensorShape(15U, _cur_batches)).set_is_resizable(true).extend_padding(new_dst_padding);
         dst.allocator()->info().set_is_resizable(false);
 
         // Configure FC info
@@ -129,16 +129,16 @@ protected:
     SimpleTensor<T> compute_reference()
     {
         // Create reference
-        SimpleTensor<T> w1{ TensorShape(180000U, 150U), DataType::F32 };
-        SimpleTensor<T> b1{ TensorShape(150U), DataType::F32 };
-        SimpleTensor<T> src{ TensorShape(1U, 150U, 1200U, _cur_batches), DataType::F32 };
+        SimpleTensor<T> w1{ TensorShape(6000U, 15U), DataType::F32 };
+        SimpleTensor<T> b1{ TensorShape(15U), DataType::F32 };
+        SimpleTensor<T> src{ TensorShape(1U, 15U, 400U, _cur_batches), DataType::F32 };
 
         // Fill reference
         fill(src, 5);
         fill(w1, 1);
         fill(b1, 2);
 
-        return reference::fully_connected_layer(src, w1, b1, TensorShape(150U, _cur_batches));
+        return reference::fully_connected_layer(src, w1, b1, TensorShape(15U, _cur_batches));
     }
 
 protected: