[SDK] Fix asan errors (#308)

* fix several heap-use-after-free bugs

* rebase master

* add option for sanitizers

* rebase master

* fix conflicts

CMakeLists.txt

@@ -9,6 +9,7 @@ project(MMDeploy VERSION 0.1.0)
 
 set(CMAKE_CXX_STANDARD 17)
 
+
 set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
@@ -28,6 +29,20 @@ if (NOT CMAKE_BUILD_TYPE)
     set(CMAKE_BUILD_TYPE Release CACHE STRING "choose 'Release' as default build type" FORCE)
 endif ()
 
+# when CUDA devices are enabled, the environment variable ASAN_OPTIONS=protect_shadow_gap=0
+# must be set at runtime
+if (MMDEPLOY_ASAN_ENABLE)
+    add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-fsanitize=address>)
+    add_link_options(-fsanitize=address)
+endif ()
+
+# notice that ubsan has linker issues for ubuntu < 18.04, see
+# https://stackoverflow.com/questions/50024731/ld-unrecognized-option-push-state-no-as-needed
+if (MMDEPLOY_UBSAN_ENABLE)
+ add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-fsanitize=undefined>)
+ add_link_options(-fsanitize=undefined)
+endif ()
+
 include(${CMAKE_SOURCE_DIR}/cmake/common.cmake)
 # set INTERFACE target to gather linked modules
 add_library(MMDeployStaticModules INTERFACE)

csrc/codebase/mmcls/linear_cls.cpp

@@ -4,6 +4,7 @@
 
 #include "codebase/mmcls/mmcls.h"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "experimental/module_adapter.h"
 
@@ -29,29 +30,27 @@ class LinearClsHead : public MMClassification {
     assert(output_tensor.shape().size() >= 2);
     auto class_num = (int)output_tensor.shape()[1];
 
-    if (output_tensor.device().is_host()) {
-      vector<float> scores(output_tensor.data<float>(),
-                           output_tensor.data<float>() + output_tensor.size());
-      OUTCOME_TRY(stream().Wait());
-      return GetLabels(scores, class_num);
-    } else {
-      vector<float> scores(output_tensor.size());
-      OUTCOME_TRY(output_tensor.CopyTo(scores.data(), stream()));
-      OUTCOME_TRY(stream().Wait());
-      return GetLabels(scores, class_num);
+    if (output_tensor.data_type() != DataType::kFLOAT) {
+      return Status(eNotSupported);
     }
+
+    OUTCOME_TRY(auto _scores, MakeAvailableOnDevice(output_tensor, kHost, stream()));
+    OUTCOME_TRY(stream().Wait());
+
+    return GetLabels(_scores, class_num);
   }
 
  private:
-  Value GetLabels(const vector<float>& scores, int class_num) const {
+  Value GetLabels(const Tensor& scores, int class_num) const {
+    auto scores_data = scores.data<float>();
     ClassifyOutput output;
     output.labels.reserve(topk_);
     std::vector<int> idx(class_num);
     iota(begin(idx), end(idx), 0);
     partial_sort(begin(idx), begin(idx) + topk_, end(idx),
-                 [&](int i, int j) { return scores[i] > scores[j]; });
+                 [&](int i, int j) { return scores_data[i] > scores_data[j]; });
     for (int i = 0; i < topk_; ++i) {
-      auto label = ClassifyOutput::Label{idx[i], scores[idx[i]]};
+      auto label = ClassifyOutput::Label{idx[i], scores_data[idx[i]]};
       DEBUG("label_id: {}, score: {}", label.label_id, label.score);
       output.labels.push_back(label);
     }
@@ -59,6 +58,8 @@ class LinearClsHead : public MMClassification {
   }
 
  private:
+  static constexpr const auto kHost = Device{0};
+
   int topk_{1};
 };
 

csrc/codebase/mmdet/instance_segmentation.cpp

@@ -1,11 +1,11 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
 #include "core/registry.h"
+#include "core/utils/device_utils.h"
 #include "experimental/module_adapter.h"
 #include "object_detection.h"
 #include "opencv2/imgproc/imgproc.hpp"
 #include "preprocess/cpu/opencv_utils.h"
-#include "preprocess/transform/transform_utils.h"
 
 namespace mmdeploy::mmdet {
 

csrc/codebase/mmdet/object_detection.cpp

@@ -3,6 +3,7 @@
 #include "object_detection.h"
 
 #include "core/registry.h"
+#include "core/utils/device_utils.h"
 #include "experimental/module_adapter.h"
 
 using namespace std;
@@ -36,20 +37,13 @@ Result<Value> ResizeBBox::operator()(const Value& prep_res, const Value& infer_r
     // 'bboxes_number'
     assert(labels.shape().size() == 2);
 
-    if (dets.device().is_host()) {
-      OUTCOME_TRY(auto result, DispatchGetBBoxes(prep_res["img_metas"], dets, labels));
-      return to_value(result);
-    } else {
-      TensorDesc _dets_desc{Device{"cpu"}, dets.data_type(), dets.shape(), dets.name()};
-      TensorDesc _labels_desc{Device{"cpu"}, labels.data_type(), labels.shape(), labels.name()};
-      Tensor _dets(_dets_desc);
-      Tensor _labels(_labels_desc);
-      OUTCOME_TRY(dets.CopyTo(_dets, stream()));
-      OUTCOME_TRY(labels.CopyTo(_labels, stream()));
-      OUTCOME_TRY(stream().Wait());
-      OUTCOME_TRY(auto result, DispatchGetBBoxes(prep_res["img_metas"], _dets, _labels));
-      return to_value(result);
-    }
+    OUTCOME_TRY(auto _dets, MakeAvailableOnDevice(dets, kHost, stream()));
+    OUTCOME_TRY(auto _labels, MakeAvailableOnDevice(labels, kHost, stream()));
+    OUTCOME_TRY(stream().Wait());
+
+    OUTCOME_TRY(auto result, DispatchGetBBoxes(prep_res["img_metas"], _dets, _labels));
+    return to_value(result);
+
   } catch (...) {
     return Status(eFail);
   }

csrc/codebase/mmedit/restorer.cpp

@@ -1,11 +1,10 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
-#include <preprocess/transform/transform_utils.h>
-
 #include <opencv2/core.hpp>
 
 #include "codebase/mmedit/mmedit.h"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 
 namespace mmdeploy::mmedit {
 

csrc/codebase/mmocr/crnn.cpp

@@ -6,6 +6,7 @@
 #include "core/model.h"
 #include "core/registry.h"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "core/value.h"
 #include "experimental/module_adapter.h"
@@ -58,23 +59,17 @@ class CTCConvertor : public MMOCR {
   }
 
   Result<Value> operator()(const Value& _data, const Value& _prob) {
-    //    auto img = _data["img"].get<Tensor>();
-    //    WARN("img shape: {}", img.shape());
-
     auto d_conf = _prob["output"].get<Tensor>();
-    std::vector<float> h_conf;
 
-    float* data{};
-    if (d_conf.device() != kHost) {
-      h_conf.resize(d_conf.byte_size() / sizeof(float));
-      OUTCOME_TRY(d_conf.CopyTo(h_conf.data(), stream_));
-      OUTCOME_TRY(stream_.Wait());
-      data = h_conf.data();
-    } else {
-      OUTCOME_TRY(stream_.Wait());
-      data = d_conf.data<float>();
+    if (d_conf.data_type() != DataType::kFLOAT) {
+      return Status(eNotSupported);
     }
 
+    OUTCOME_TRY(auto h_conf, MakeAvailableOnDevice(d_conf, Device{0}, stream()));
+    OUTCOME_TRY(stream().Wait());
+
+    auto data = h_conf.data<float>();
+
     auto shape = d_conf.shape();
     auto w = static_cast<int>(shape[1]);
     auto c = static_cast<int>(shape[2]);

csrc/codebase/mmocr/dbnet.cpp

@@ -8,12 +8,12 @@
 #include "core/registry.h"
 #include "core/serialization.h"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "core/value.h"
 #include "experimental/module_adapter.h"
 #include "mmocr.h"
 #include "preprocess/cpu/opencv_utils.h"
-#include "preprocess/transform/transform_utils.h"
 
 namespace mmdeploy::mmocr {
 

csrc/codebase/mmocr/resize_ocr.cpp

@@ -5,11 +5,11 @@
 #include "archive/json_archive.h"
 #include "archive/value_archive.h"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "opencv2/imgproc.hpp"
 #include "preprocess/cpu/opencv_utils.h"
 #include "preprocess/transform/resize.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 

csrc/codebase/mmseg/segment.cpp

@@ -2,10 +2,10 @@
 
 #include "codebase/mmseg/mmseg.h"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "preprocess/cpu/opencv_utils.h"
 #include "preprocess/transform/transform.h"
-#include "preprocess/transform/transform_utils.h"
 
 namespace mmdeploy::mmseg {
 

csrc/core/CMakeLists.txt

@@ -33,6 +33,7 @@ set(SRCS
         operator.cpp
         status_code.cpp
         tensor.cpp
+        utils/device_utils.cpp
         utils/formatter.cpp
         utils/stacktrace.cpp)
 build_object_target(${CORE_OBJ} "${SRCS}")

csrc/core/tensor.h

@@ -19,9 +19,14 @@ struct TensorDesc {
   std::string name;
 };
 
-class Tensor final {
+class Tensor {
  public:
   Tensor() = default;
+  Tensor(const Tensor&) = default;
+  Tensor(Tensor&&) noexcept = default;
+  Tensor& operator=(const Tensor&) = default;
+  Tensor& operator=(Tensor&&) noexcept = default;
+
   Tensor(const TensorDesc& desc, Allocator allocator = {});  // NOLINT
   Tensor(const TensorDesc& desc, Buffer buffer);
   Tensor(const TensorDesc& desc, std::shared_ptr<void> data);

csrc/core/utils/device_utils.cpp

@@ -1,6 +1,6 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
-#include "transform_utils.h"
+#include "device_utils.h"
 namespace mmdeploy {
 
 Result<Mat> MakeAvailableOnDevice(const Mat& src, const Device& device, Stream& stream) {

csrc/device/device_allocator.h

@@ -126,8 +126,8 @@ class Tree : public AllocatorImpl {
     if (auto it = tree_.lower_bound(size); it != tree_.end()) {
       if (size * thresh_denominator_ >= it->first * thresh_numerator_) {
         Block block(it->second, it->first);
-        tree_.erase(it);
         tree_bytes_ -= it->first;
+        tree_.erase(it);
         return block;
       }
     }

csrc/net/net_module.cpp

@@ -169,10 +169,10 @@ struct NetModule::Impl {
       }
       if (output.size() > 1) {
         for (int i = 0; i < output.size(); ++i) {
-          output[i].insert({name, tmp.Slice(i)});
+          output[i].emplace(name, tmp.Slice(i));
         }
       } else {
-        output[0].insert({name, std::move(tmp)});
+        output[0].emplace(name, std::move(tmp));
       }
     }
 

csrc/preprocess/cpu/crop_impl.cpp

@@ -1,8 +1,8 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
+#include "core/utils/device_utils.h"
 #include "opencv_utils.h"
 #include "preprocess/transform/crop.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 

csrc/preprocess/cpu/image2tensor_impl.cpp

@@ -1,8 +1,8 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
+#include "core/utils/device_utils.h"
 #include "opencv_utils.h"
 #include "preprocess/transform/image2tensor.h"
-#include "preprocess/transform/transform_utils.h"
 
 namespace mmdeploy {
 namespace cpu {

csrc/preprocess/cpu/normalize_impl.cpp

@@ -1,9 +1,9 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
 #include "core/registry.h"
+#include "core/utils/device_utils.h"
 #include "opencv_utils.h"
 #include "preprocess/transform/normalize.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 

csrc/preprocess/cpu/pad_impl.cpp

@@ -1,8 +1,8 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
+#include "core/utils/device_utils.h"
 #include "opencv_utils.h"
 #include "preprocess/transform/pad.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 

csrc/preprocess/cpu/resize_impl.cpp

@@ -2,9 +2,9 @@
 
 #include "core/registry.h"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 #include "opencv_utils.h"
 #include "preprocess/transform/resize.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 

csrc/preprocess/cuda/crop_impl.cpp

@@ -2,8 +2,8 @@
 
 #include <cuda_runtime.h>
 
+#include "core/utils/device_utils.h"
 #include "preprocess/transform/crop.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 

csrc/preprocess/cuda/image2tensor_impl.cpp

@@ -2,8 +2,8 @@
 
 #include <cuda_runtime.h>
 
+#include "core/utils/device_utils.h"
 #include "preprocess/transform/image2tensor.h"
-#include "preprocess/transform/transform_utils.h"
 
 namespace mmdeploy {
 namespace cuda {

csrc/preprocess/cuda/load_impl.cpp

@@ -2,10 +2,10 @@
 
 #include <cuda_runtime.h>
 
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "ppl/cv/cuda/cvtcolor.h"
 #include "preprocess/transform/load.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 using namespace ppl::cv::cuda;

csrc/preprocess/cuda/normalize_impl.cpp

@@ -2,9 +2,9 @@
 
 #include <cuda_runtime.h>
 
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "preprocess/transform/normalize.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 

csrc/preprocess/cuda/pad_impl.cpp

@@ -1,9 +1,9 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "ppl/cv/cuda/copymakeborder.h"
 #include "preprocess/transform/pad.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 using namespace ppl::cv::cuda;

csrc/preprocess/cuda/resize_impl.cpp

@@ -1,9 +1,9 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
+#include "core/utils/device_utils.h"
 #include "core/utils/formatter.h"
 #include "ppl/cv/cuda/resize.h"
 #include "preprocess/transform/resize.h"
-#include "preprocess/transform/transform_utils.h"
 
 using namespace std;
 

csrc/preprocess/transform/CMakeLists.txt

@@ -14,8 +14,8 @@ set(SRCS
         normalize.cpp
         pad.cpp
         resize.cpp
-        transform.cpp
-        transform_utils.cpp)
+        transform.cpp)
+
 build_object_target(${TRANSFORM_OBJ} "${SRCS}")
 target_include_directories(${TRANSFORM_OBJ} PUBLIC
         $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/preprocess>)

tests/test_csrc/preprocess/test_crop.cpp

@@ -3,9 +3,9 @@
 
 #include "catch.hpp"
 #include "core/mat.h"
+#include "core/utils/device_utils.h"
 #include "preprocess/cpu/opencv_utils.h"
 #include "preprocess/transform/transform.h"
-#include "preprocess/transform/transform_utils.h"
 #include "test_resource.h"
 #include "test_utils.h"
 

tests/test_csrc/preprocess/test_image2tensor.cpp

@@ -1,9 +1,9 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 #include "catch.hpp"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 #include "preprocess/cpu/opencv_utils.h"
 #include "preprocess/transform/transform.h"
-#include "preprocess/transform/transform_utils.h"
 #include "test_resource.h"
 #include "test_utils.h"
 

tests/test_csrc/preprocess/test_load.cpp

@@ -3,9 +3,9 @@
 #include "catch.hpp"
 #include "core/mat.h"
 #include "core/tensor.h"
+#include "core/utils/device_utils.h"
 #include "preprocess/cpu/opencv_utils.h"
 #include "preprocess/transform/transform.h"
-#include "preprocess/transform/transform_utils.h"
 #include "test_resource.h"
 #include "test_utils.h"
 

tests/test_csrc/preprocess/test_normalize.cpp

@@ -2,9 +2,9 @@
 
 #include "catch.hpp"
 #include "core/mat.h"
+#include "core/utils/device_utils.h"
 #include "preprocess/cpu/opencv_utils.h"
 #include "preprocess/transform/transform.h"
-#include "preprocess/transform/transform_utils.h"
 #include "test_resource.h"
 #include "test_utils.h"
 

tests/test_csrc/preprocess/test_pad.cpp

@@ -2,9 +2,9 @@
 
 #include "catch.hpp"
 #include "core/mat.h"
+#include "core/utils/device_utils.h"
 #include "preprocess/cpu/opencv_utils.h"
 #include "preprocess/transform/transform.h"
-#include "preprocess/transform/transform_utils.h"
 #include "test_resource.h"
 #include "test_utils.h"
 

tests/test_csrc/preprocess/test_resize.cpp

@@ -2,9 +2,9 @@
 
 #include "catch.hpp"
 #include "core/mat.h"
+#include "core/utils/device_utils.h"
 #include "preprocess/cpu/opencv_utils.h"
 #include "preprocess/transform/transform.h"
-#include "preprocess/transform/transform_utils.h"
 #include "test_resource.h"
 #include "test_utils.h"