Fix allocation of prepare tensor on ClWinogradConv2d

Preparation tensor was allocated during execution. Avoid if GEMM used by
the Winograd function is reshaped.

Resolves: COMPMID-4640

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: Id5b7e6bbfe9f69661baa9d42698cdc2cc1b422b2
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5991
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 15 insertions, 14 deletions

diff --git a/src/runtime/gpu/cl/operators/ClWinogradConv2d.cpp b/src/runtime/gpu/cl/operators/ClWinogradConv2d.cpp
index 2ca1ff59df..07f90ddaef 100644
--- a/src/runtime/gpu/cl/operators/ClWinogradConv2d.cpp
+++ b/src/runtime/gpu/cl/operators/ClWinogradConv2d.cpp
@@ -233,37 +233,32 @@ Status ClWinogradConv2d::validate(const ITensorInfo *src, const ITensorInfo *wei
 
 void ClWinogradConv2d::run(ITensorPack &tensors)
 {
-    prepare(tensors);
+    const bool is_gemm_reshaped = _aux_mem[3].lifetime == MemoryLifetime::Prepare;
 
     auto src    = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
     auto biases = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
     auto dst    = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
 
     CLAuxTensorHandler input0(offset_int_vec(2), _input0, tensors, true);
-    CLAuxTensorHandler input1(offset_int_vec(3), _input1, tensors, true);
+    CLAuxTensorHandler input1(offset_int_vec(3), _input1, tensors, true, is_gemm_reshaped);
     CLAuxTensorHandler batched_mm_output(offset_int_vec(4), _batched_mm_output, tensors, true);
 
+    prepare(tensors);
+
     // Run input transform
     ITensorPack pack_it
     {
         { TensorType::ACL_SRC, src },
         { TensorType::ACL_DST, input0.get() },
     };
-    CLScheduler::get().enqueue_op(_border_handler, pack_it);
-    CLScheduler::get().enqueue_op(*_input_transform, pack_it);
+    CLScheduler::get().enqueue_op(_border_handler, pack_it, false);
+    CLScheduler::get().enqueue_op(*_input_transform, pack_it, false);
 
     // Run batched matrix multiplication
     ITensorPack pack_mm = tensors;
     pack_mm.add_const_tensor(TensorType::ACL_SRC_0, input0.get());
     pack_mm.add_tensor(TensorType::ACL_DST, batched_mm_output.get());
-    if(_aux_mem[3].lifetime == MemoryLifetime::Prepare)
-    {
-        pack_mm.remove_tensor(TensorType::ACL_SRC_1);
-    }
-    else
-    {
-        pack_mm.add_const_tensor(TensorType::ACL_SRC_1, input1.get());
-    }
+    is_gemm_reshaped ? pack_mm.remove_tensor(TensorType::ACL_SRC_1) : pack_mm.add_const_tensor(TensorType::ACL_SRC_1, input1.get());
     _batched_mm.run(pack_mm);
 
     // Run output transform
diff --git a/src/runtime/gpu/cl/utils/ClAuxTensorHandler.h b/src/runtime/gpu/cl/utils/ClAuxTensorHandler.h
index 1cf717cf6f..af383489a1 100644
--- a/src/runtime/gpu/cl/utils/ClAuxTensorHandler.h
+++ b/src/runtime/gpu/cl/utils/ClAuxTensorHandler.h
@@ -28,6 +28,7 @@
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 
+#include "src/common/utils/Log.h"
 #include "support/Cast.h"
 
 namespace arm_compute
@@ -38,7 +39,7 @@ namespace opencl
 class CLAuxTensorHandler
 {
 public:
-    CLAuxTensorHandler(int slot_id, TensorInfo &info, ITensorPack &pack, bool pack_inject = false)
+    CLAuxTensorHandler(int slot_id, TensorInfo &info, ITensorPack &pack, bool pack_inject = false, bool bypass_alloc = false)
         : _tensor()
     {
         if(info.total_size() == 0)
@@ -50,7 +51,12 @@ public:
         ICLTensor *packed_tensor = utils::cast::polymorphic_downcast<ICLTensor *>(pack.get_tensor(slot_id));
         if((packed_tensor == nullptr) || (info.total_size() > packed_tensor->info()->total_size()))
         {
-            _tensor.allocator()->allocate();
+            if(!bypass_alloc)
+            {
+                _tensor.allocator()->allocate();
+                ARM_COMPUTE_LOG_INFO_WITH_FUNCNAME_ACL("Allocating auxiliary tensor");
+            }
+
             if(pack_inject)
             {
                 pack.add_tensor(slot_id, &_tensor);