Fix incorrect memory handling in ported functions

Details of the functions:
- ClSoftmax
- CpuSoftmax
- CpuPool2d

Change-Id: Icd2c14d5df010c3b2301e2693ce6f414d7c61916
Resolves: COMPMID-4404
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5797
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 187 insertions, 55 deletions

diff --git a/src/runtime/cpu/operators/CpuPool2d.cpp b/src/runtime/cpu/operators/CpuPool2d.cpp
index b225199c40..e746c8fb3b 100644
--- a/src/runtime/cpu/operators/CpuPool2d.cpp
+++ b/src/runtime/cpu/operators/CpuPool2d.cpp
@@ -30,6 +30,8 @@
 #include "src/core/cpu/kernels/CpuPool2dKernel.h"
 #include "src/core/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h"
 
+using namespace arm_compute::experimental;
+
 namespace arm_compute
 {
 namespace cpu
@@ -40,7 +42,7 @@ CpuPool2d::CpuPool2d()
       _asm_glue(),
       _is_global_pooling_layer(false),
       _data_layout(DataLayout::NCHW),
-      _mem_req()
+      _aux_mem(1)
 {
 }
 
@@ -71,7 +73,7 @@ void CpuPool2d::configure(ITensorInfo *src, ITensorInfo *dst, const PoolingLayer
         // Get kernel's memory requirements
         constexpr size_t alignment      = 4096;
         const size_t     workspace_size = pooling_wrapper->get_working_size(num_threads);
-        _mem_req.push_back({ TensorType::ACL_INT_0, workspace_size, alignment });
+        _aux_mem[0]                     = MemoryInfo(TensorType::ACL_INT_0, MemoryLifetime::Temporary, workspace_size, alignment);
 
         _asm_glue = std::move(pooling_wrapper);
     }
@@ -150,7 +152,7 @@ void CpuPool2d::run(ITensorPack &tensors)
 
 experimental::MemoryRequirements CpuPool2d::workspace() const
 {
-    return _mem_req;
+    return _aux_mem;
 }
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/runtime/cpu/operators/CpuPool2d.h b/src/runtime/cpu/operators/CpuPool2d.h
index ae3d115dfc..68416b5cfc 100644
--- a/src/runtime/cpu/operators/CpuPool2d.h
+++ b/src/runtime/cpu/operators/CpuPool2d.h
@@ -80,7 +80,7 @@ private:
 
     bool                             _is_global_pooling_layer;
     DataLayout                       _data_layout;
-    experimental::MemoryRequirements _mem_req;
+    experimental::MemoryRequirements _aux_mem{};
 };
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/runtime/cpu/operators/CpuSoftmax.cpp b/src/runtime/cpu/operators/CpuSoftmax.cpp
index 0e1bcd5c69..e17925ee50 100644
--- a/src/runtime/cpu/operators/CpuSoftmax.cpp
+++ b/src/runtime/cpu/operators/CpuSoftmax.cpp
@@ -29,7 +29,11 @@
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 #include "src/core/cpu/kernels/CpuSoftmaxKernel.h"
+#include "src/core/helpers/MemoryHelpers.h"
 #include "src/core/helpers/SoftmaxHelpers.h"
+#include "src/runtime/cpu/utils/CpuAuxTensorHandler.h"
+
+using namespace arm_compute::experimental;
 
 namespace arm_compute
 {
@@ -37,7 +41,16 @@ namespace cpu
 {
 template <bool IS_LOG>
 CpuSoftmaxGeneric<IS_LOG>::CpuSoftmaxGeneric()
-    : _permute_input(), _permute_output(), _max_kernel(), _softmax_kernel(), _max(nullptr), _tmp(nullptr), _input_permuted(nullptr), _output_permuted(nullptr), _needs_permute(false)
+    : _permute_input(),
+      _permute_output(),
+      _max_kernel(),
+      _softmax_kernel(),
+      _max(),
+      _tmp(),
+      _input_permuted(),
+      _output_permuted(),
+      _needs_permute(false),
+      _aux_mem(InternalTensorIdx::COUNT)
 {
 }
 
@@ -54,13 +67,12 @@ void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *d
 
     if(_needs_permute)
     {
-        _input_permuted = std::make_unique<TensorInfo>();
-        _permute_input.configure(src, _input_permuted.get(), softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
+        _permute_input.configure(src, &_input_permuted, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
     }
 
     // We want to deal with a 2D input. Either it is the permuted version of the original input (4D case)
     // or it is the original input case (2D case)
-    const ITensorInfo *tmp_input = (_needs_permute ? _input_permuted.get() : src);
+    const ITensorInfo *tmp_input = (_needs_permute ? &_input_permuted : src);
 
     // Create intermediate tensors shapes
     TensorShape max_sum_shape = tmp_input->tensor_shape();
@@ -71,31 +83,35 @@ void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *d
     TensorInfo       max_info(tmp_input->clone()->set_tensor_shape(max_sum_shape));
 
     // Init intermediate tensors
-    _max = std::make_unique<TensorInfo>(max_info);
-    _tmp = std::make_unique<TensorInfo>(tensor_info_tmp);
+    _max = TensorInfo(max_info);
+    _tmp = TensorInfo(tensor_info_tmp);
 
     // Configure kernels
     auto mk = std::make_unique<kernels::CpuLogits1DMaxKernel>();
-    mk->configure(tmp_input, _max.get());
+    mk->configure(tmp_input, &_max);
     _max_kernel = std::move(mk);
 
     auto sm = std::make_unique<kernels::CpuLogits1DSoftmaxKernel<IS_LOG>>();
     if(_needs_permute)
     {
-        _output_permuted = std::make_unique<TensorInfo>();
-
         // The normalization kernel stores the result in a permuted output tensor
-        sm->configure(tmp_input, _max.get(), _output_permuted.get(), beta, _tmp.get());
+        sm->configure(tmp_input, &_max, &_output_permuted, beta, &_tmp);
 
         // Re-permute the permuted output into the requested (4D) output
-        _permute_output.configure(_output_permuted.get(), dst, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
+        _permute_output.configure(&_output_permuted, dst, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
     }
     else
     {
         // Softmax 2D case
-        sm->configure(tmp_input, _max.get(), dst, beta, _tmp.get());
+        sm->configure(tmp_input, &_max, dst, beta, &_tmp);
     }
     _softmax_kernel = std::move(sm);
+
+    _aux_mem[InternalTensorIdx::MAX] = MemoryInfo(offset_int_vec(InternalTensorIdx::MAX), MemoryLifetime::Temporary, _max.total_size());
+    _aux_mem[InternalTensorIdx::TMP] = MemoryInfo(offset_int_vec(InternalTensorIdx::TMP), MemoryLifetime::Temporary, _tmp.total_size());
+
+    _aux_mem[InternalTensorIdx::PERMUTED_SRC] = MemoryInfo(offset_int_vec(InternalTensorIdx::PERMUTED_SRC), MemoryLifetime::Temporary, _input_permuted.total_size());
+    _aux_mem[InternalTensorIdx::PERMUTED_DST] = MemoryInfo(offset_int_vec(InternalTensorIdx::PERMUTED_DST), MemoryLifetime::Temporary, _output_permuted.total_size());
 }
 
 template <bool IS_LOG>
@@ -141,42 +157,54 @@ void CpuSoftmaxGeneric<IS_LOG>::run(ITensorPack &tensors)
 {
     ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
 
+    auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto dst = tensors.get_tensor(TensorType::ACL_DST);
+
+    CpuAuxTensorHandler tmp(offset_int_vec(InternalTensorIdx::TMP), _tmp, tensors, false);
+    CpuAuxTensorHandler max(offset_int_vec(InternalTensorIdx::MAX), _max, tensors, false);
+
+    CpuAuxTensorHandler input_permuted(offset_int_vec(InternalTensorIdx::PERMUTED_SRC), _input_permuted, tensors, false);
+    CpuAuxTensorHandler output_permuted(offset_int_vec(InternalTensorIdx::PERMUTED_DST), _output_permuted, tensors, false);
+
     ITensorPack max_pack;
     ITensorPack softmax_pack;
 
     if(_needs_permute)
     {
-        ITensorPack permute_in_pack;
-        permute_in_pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC));
-        permute_in_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_INT_2));
+        ITensorPack permute_in_pack = { { TensorType::ACL_SRC, src }, { TensorType::ACL_DST, input_permuted.get() } };
         _permute_input.run(permute_in_pack);
 
-        max_pack.add_tensor(TensorType::ACL_SRC, tensors.get_tensor(ACL_INT_2));
+        max_pack = { { TensorType::ACL_SRC, input_permuted.get() }, { TensorType::ACL_DST, max.get() } };
 
-        softmax_pack.add_tensor(TensorType::ACL_SRC_0, tensors.get_tensor(ACL_INT_2));
-        softmax_pack.add_tensor(TensorType::ACL_SRC_1, tensors.get_tensor(ACL_INT_1));
-        softmax_pack.add_tensor(TensorType::ACL_DST_0, tensors.get_tensor(ACL_INT_3));
-        softmax_pack.add_tensor(TensorType::ACL_DST_1, tensors.get_tensor(ACL_INT_0));
+        softmax_pack =
+        {
+            { TensorType::ACL_SRC_0, input_permuted.get() },
+            { TensorType::ACL_SRC_1, max.get() },
+            { TensorType::ACL_DST_0, output_permuted.get() },
+            { TensorType::ACL_DST_1, tmp.get() }
+        };
     }
     else
     {
-        max_pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC));
-        softmax_pack.add_tensor(TensorType::ACL_SRC_0, tensors.get_const_tensor(ACL_SRC));
-        softmax_pack.add_tensor(TensorType::ACL_SRC_1, tensors.get_tensor(ACL_INT_1));
-        softmax_pack.add_tensor(TensorType::ACL_DST_0, tensors.get_tensor(ACL_DST));
-        softmax_pack.add_tensor(TensorType::ACL_DST_1, tensors.get_tensor(ACL_INT_0));
+        max_pack = { { TensorType::ACL_SRC, src }, { TensorType::ACL_DST, max.get() } };
+
+        softmax_pack =
+        {
+            { TensorType::ACL_SRC_0, src },
+            { TensorType::ACL_SRC_1, max.get() },
+            { TensorType::ACL_DST_0, dst },
+            { TensorType::ACL_DST_1, tmp.get() }
+        };
     }
 
-    max_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_INT_1));
-
     NEScheduler::get().schedule_op(_max_kernel.get(), Window::DimY, _max_kernel->window(), max_pack);
     NEScheduler::get().schedule_op(_softmax_kernel.get(), Window::DimY, _softmax_kernel->window(), softmax_pack);
 
     if(_needs_permute)
     {
         ITensorPack permute_out_pack;
-        permute_out_pack.add_tensor(TensorType::ACL_SRC, tensors.get_tensor(ACL_INT_3));
-        permute_out_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_DST));
+        permute_out_pack.add_tensor(TensorType::ACL_SRC, output_permuted.get());
+        permute_out_pack.add_tensor(TensorType::ACL_DST, dst);
         _permute_output.run(permute_out_pack);
     }
 }
@@ -184,18 +212,7 @@ void CpuSoftmaxGeneric<IS_LOG>::run(ITensorPack &tensors)
 template <bool                   IS_LOG>
 experimental::MemoryRequirements CpuSoftmaxGeneric<IS_LOG>::workspace() const
 {
-    experimental::MemoryRequirements req{};
-
-    req.push_back({ TensorType::ACL_INT_0, _tmp->total_size(), 0 });
-    req.push_back({ TensorType::ACL_INT_1, _max->total_size(), 0 });
-
-    if(_needs_permute)
-    {
-        req.push_back({ TensorType::ACL_INT_2, _input_permuted->total_size(), 0 });
-        req.push_back({ TensorType::ACL_INT_3, _output_permuted->total_size(), 0 });
-    }
-
-    return req;
+    return _aux_mem;
 }
 
 template class CpuSoftmaxGeneric<false>;
diff --git a/src/runtime/cpu/operators/CpuSoftmax.h b/src/runtime/cpu/operators/CpuSoftmax.h
index 9f18e0e4c5..38817977b3 100644
--- a/src/runtime/cpu/operators/CpuSoftmax.h
+++ b/src/runtime/cpu/operators/CpuSoftmax.h
@@ -24,7 +24,7 @@
 #ifndef ARM_COMPUTE_CPU_SOFTMAX_H
 #define ARM_COMPUTE_CPU_SOFTMAX_H
 
-#include "arm_compute/core/ITensorInfo.h"
+#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/experimental/Types.h"
 #include "src/core/cpu/ICpuKernel.h"
 #include "src/runtime/cpu/ICpuOperator.h"
@@ -87,15 +87,27 @@ public:
     experimental::MemoryRequirements workspace() const override;
 
 private:
-    CpuPermute                   _permute_input;
-    CpuPermute                   _permute_output;
-    std::unique_ptr<ICpuKernel>  _max_kernel;
-    std::unique_ptr<ICpuKernel>  _softmax_kernel;
-    std::unique_ptr<ITensorInfo> _max;
-    std::unique_ptr<ITensorInfo> _tmp;
-    std::unique_ptr<ITensorInfo> _input_permuted;
-    std::unique_ptr<ITensorInfo> _output_permuted;
-    bool                         _needs_permute;
+    enum InternalTensorIdx
+    {
+        MAX = 0,
+        TMP,
+        PERMUTED_SRC,
+        PERMUTED_DST,
+        COUNT
+    };
+
+    CpuPermute                  _permute_input;
+    CpuPermute                  _permute_output;
+    std::unique_ptr<ICpuKernel> _max_kernel;
+    std::unique_ptr<ICpuKernel> _softmax_kernel;
+
+    TensorInfo _max;
+    TensorInfo _tmp;
+    TensorInfo _input_permuted;
+    TensorInfo _output_permuted;
+
+    bool                             _needs_permute;
+    experimental::MemoryRequirements _aux_mem{};
 };
 using CpuSoftmax    = CpuSoftmaxGeneric<false>;
 using CpuLogSoftmax = CpuSoftmaxGeneric<true>;
diff --git a/src/runtime/cpu/utils/CpuAuxTensorHandler.h b/src/runtime/cpu/utils/CpuAuxTensorHandler.h
new file mode 100644
index 0000000000..644018a718
--- /dev/null
+++ b/src/runtime/cpu/utils/CpuAuxTensorHandler.h
@@ -0,0 +1,101 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CPU_UTILS_CPU_AUX_TENSOR_HANDLER_H
+#define ARM_COMPUTE_CPU_UTILS_CPU_AUX_TENSOR_HANDLER_H
+
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/runtime/Tensor.h"
+
+#include "support/Cast.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+/* Tensor handler to wrap and handle tensor allocations on workspace buffers */
+class CpuAuxTensorHandler
+{
+public:
+    CpuAuxTensorHandler(int slot_id, TensorInfo &info, ITensorPack &pack, bool pack_inject = false)
+        : _tensor()
+    {
+        _tensor.allocator()->soft_init(info);
+
+        ITensor *packed_tensor = utils::cast::polymorphic_downcast<ITensor *>(pack.get_tensor(slot_id));
+        if((packed_tensor == nullptr) || (info.total_size() > packed_tensor->info()->total_size()))
+        {
+            _tensor.allocator()->allocate();
+            if(pack_inject)
+            {
+                pack.add_tensor(slot_id, &_tensor);
+                _injected_tensor_pack = &pack;
+                _injected_slot_id     = slot_id;
+            }
+        }
+        else
+        {
+            _tensor.allocator()->import_memory(packed_tensor->buffer());
+        }
+    }
+
+    CpuAuxTensorHandler(TensorInfo &info, ITensor &tensor)
+        : _tensor()
+    {
+        _tensor.allocator()->soft_init(info);
+        if(info.total_size() <= tensor.info()->total_size())
+        {
+            _tensor.allocator()->import_memory(tensor.buffer());
+        }
+    }
+
+    CpuAuxTensorHandler(const CpuAuxTensorHandler &) = delete;
+    CpuAuxTensorHandler &operator=(const CpuAuxTensorHandler) = delete;
+
+    ~CpuAuxTensorHandler()
+    {
+        if(_injected_tensor_pack)
+        {
+            _injected_tensor_pack->remove_tensor(_injected_slot_id);
+        }
+    }
+
+    ITensor *get()
+    {
+        return &_tensor;
+    }
+
+    ITensor *operator()()
+    {
+        return &_tensor;
+    }
+
+private:
+    Tensor       _tensor{};
+    ITensorPack *_injected_tensor_pack{ nullptr };
+    int          _injected_slot_id{ TensorType::ACL_UNKNOWN };
+};
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_UTILS_CPU_AUX_TENSOR_HANDLER_H */
\ No newline at end of file