codex clean && fuzz bugfix

mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/infer/bias_grad_infer.c

@@ -27,17 +27,11 @@ int BiasGradInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC
   const TensorC *in0 = inputs[0];
   TensorC *out = outputs[0];
 
-  if (in0->shape_size_ > MAX_SHAPE_SIZE) {
+  if (in0->shape_size_ > MAX_SHAPE_SIZE || in0->shape_size_ < 1) {
     return NNACL_INPUT_TENSOR_ERROR;
   }
-  int inshape[MAX_SHAPE_SIZE];
-  size_t inshape_size = 0;
-  ShapeSet(inshape, &inshape_size, in0->shape_, in0->shape_size_);
-  size_t ndim = inshape_size;
-  MS_CHECK_TRUE_RET(ndim - 1 <= MAX_SHAPE_SIZE, NNACL_ERR);
-  for (size_t i = 0; i < ndim - 1; i++) {
-    inshape[i] = 1;
-  }
+  int inshape[] = {in0->shape_[in0->shape_size_ - 1]};
+  size_t inshape_size = 1;
   SetDataTypeFormat(out, in0);
   SetShapeArray(out, inshape, inshape_size);
 

mindspore/lite/src/common/tensor_util.cc

@@ -16,6 +16,7 @@
 
 #include "src/common/tensor_util.h"
 #include <algorithm>
+#include <unordered_map>
 #include "schema/model_generated.h"
 #include "include/errorcode.h"
 #include "src/common/log_adapter.h"
@@ -265,6 +266,23 @@ int CheckTensorsInvalid(const std::vector<Tensor *> &tensors) {
   return RET_OK;
 }
 
+int CheckGraphInputShapes(const std::vector<Tensor *> &inputs,
+                          const std::unordered_map<Tensor *, std::vector<int>> &input_shape_map) {
+  for (const auto input : inputs) {
+    MS_CHECK_TRUE_MSG(input != nullptr, RET_ERROR, "graph input tensor is nullptr.");
+    if (input_shape_map.find(input) == input_shape_map.end()) {
+      MS_LOG(ERROR) << "can't find " << input->tensor_name() << " in input_shape_map";
+      return RET_ERROR;
+    }
+    if (!input_shape_map.at(input).empty() && input_shape_map.at(input) != input->shape()) {
+      MS_LOG(ERROR) << "graph input:" << input->tensor_name()
+                    << " shape has been illegally modified, please modify the input shape with method Resize().";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
 std::vector<mindspore::MSTensor> LiteTensorsToMSTensors(const std::vector<lite::Tensor *> &lite_tensors) {
   std::vector<mindspore::MSTensor> tensors;
   std::transform(lite_tensors.begin(), lite_tensors.end(), std::back_inserter(tensors), [](lite::Tensor *tensor) {

mindspore/lite/src/common/tensor_util.h

@@ -17,7 +17,7 @@
 #ifndef MINDSPORE_LITE_SRC_COMMON_TENSOR_UTIL_H_
 #define MINDSPORE_LITE_SRC_COMMON_TENSOR_UTIL_H_
 #include <vector>
-
+#include <unordered_map>
 #include <memory>
 #include "src/tensor.h"
 #include "nnacl/tensor_c.h"
@@ -43,6 +43,8 @@ int GenerateInTensorC(const OpParameter *const parameter, const std::vector<lite
 int GenerateOutTensorC(const OpParameter *const parameter, const std::vector<lite::Tensor *> &outputs,
                        std::vector<TensorC *> *out_tensor_c);
 int CheckTensorsInvalid(const std::vector<Tensor *> &tensors);
+int CheckGraphInputShapes(const std::vector<Tensor *> &inputs,
+                          const std::unordered_map<Tensor *, std::vector<int>> &input_shape_map);
 std::vector<mindspore::MSTensor> LiteTensorsToMSTensors(const std::vector<lite::Tensor *> &lite_tensors);
 void MoveCommonTensorData(Tensor *dst_tensor, Tensor *src_tensor);
 void MoveTensorData(Tensor *dst_tensor, Tensor *src_tensor);

mindspore/lite/src/lite_session.cc

@@ -416,6 +416,7 @@ void LiteSession::InitGraphOutputTensors(const lite::Model *model) {
 void LiteSession::InitGraphInputMap(const lite::Model *model) {
   MS_ASSERT(model != nullptr);
   MS_ASSERT(this->input_map_.empty());
+  MS_ASSERT(this->input_shape_map_.empty());
   auto graph_input_node_indexes = GetGraphInputNodes(model);
   auto graph_in_size = model->input_indices_.size();
   for (auto in_node_index : graph_input_node_indexes) {
@@ -443,6 +444,7 @@ void LiteSession::InitGraphInputMap(const lite::Model *model) {
       }
       auto tensor_name = in_node->name_ + std::to_string(i);
       this->input_map_[tensor_name] = in_tensor;
+      this->input_shape_map_[in_tensor] = in_tensor->shape();
       if (!in_tensor->tensor_name().empty()) {
         this->input_map_[in_tensor->tensor_name()] = in_tensor;
       }
@@ -861,6 +863,12 @@ int LiteSession::RunGraph(const KernelCallBack &before, const KernelCallBack &af
     MS_LOG(ERROR) << "CheckInputs failed.";
     return ret;
   }
+  ret = CheckGraphInputShapes(inputs_, input_shape_map_);
+  if (ret != RET_OK) {
+    is_running_.store(false);
+    MS_LOG(ERROR) << "Check graph input shapes failed.";
+    return ret;
+  }
   MS_ASSERT(this->context_ != nullptr);
   if (before == nullptr && after == nullptr) {
     ret = executor_->Run(this->inputs_, this->outputs_, this->kernels_);
@@ -1089,6 +1097,7 @@ LiteSession::~LiteSession() {
 
   // Tensor * in input_map output_map are freed in tensors
   input_map_.clear();
+  input_shape_map_.clear();
   output_node_map_.clear();
   output_tensor_map_.clear();
   input_vec_.clear();
@@ -1143,6 +1152,19 @@ std::unordered_map<std::string, mindspore::tensor::MSTensor *> LiteSession::GetO
   return this->output_tensor_map_;
 }
 
+int LiteSession::UpdateInputShapeMap() {
+  for (auto input : inputs_) {
+    MS_CHECK_TRUE_MSG(input != nullptr, RET_ERROR, "graph input tensor is nullptr.");
+    if (input_shape_map_.find(input) != input_shape_map_.end()) {
+      input_shape_map_.at(input) = input->shape();
+    } else {
+      MS_LOG(ERROR) << "can't find " << input->tensor_name() << " in input_shape_map";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
 int LiteSession::ResizeInputs(const std::vector<mindspore::tensor::MSTensor *> &inputs,
                               const std::vector<std::vector<int>> &dims) {
   if (inputs.size() != inputs_.size()) {
@@ -1317,6 +1339,11 @@ int LiteSession::Resize(const std::vector<mindspore::tensor::MSTensor *> &inputs
 #if defined(LINUX_RUNTIME)
   (void)malloc_trim(0);
 #endif
+  ret = UpdateInputShapeMap();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "update input shape map failed.";
+    return RET_ERROR;
+  }
   return RET_OK;
 }
 

mindspore/lite/src/lite_session.h

@@ -104,6 +104,7 @@ class LiteSession : public session::LiteSession {
   void InitGraphOutputNodeMap(const lite::Model *model);
   void InitGraphOutputTensorMap(const lite::Model *model);
   void AdjustModelOutputTensorInitRefCount(const lite::Model *model);
+  int UpdateInputShapeMap();
   int ResizeInputs(const std::vector<mindspore::tensor::MSTensor *> &inputs, const std::vector<std::vector<int>> &dims);
   int SetAllocatorForDelegateKernels(const kernel::LiteKernel *kernel);
   int PrepareKernels(const Model *model);
@@ -160,6 +161,8 @@ class LiteSession : public session::LiteSession {
   std::vector<mindspore::tensor::MSTensor *> input_vec_;
   // graph input tensor name -- input tensors
   std::unordered_map<std::string, mindspore::tensor::MSTensor *> input_map_;
+  // graph input tensor -- input tensor shape
+  std::unordered_map<Tensor *, std::vector<int>> input_shape_map_;
   // graph output node name -- output tensors
   std::unordered_map<std::string, std::vector<mindspore::tensor::MSTensor *>> output_node_map_;
 

mindspore/lite/src/ops/populate/conv2d_populate.cc

@@ -29,8 +29,12 @@ int SetPadAndAct(schema::PadMode pad_mode, schema::ActivationType act_type, Conv
     case schema::PadMode_VALID:
       param->pad_mode_ = Pad_valid;
       break;
-    default:
+    case schema::PadMode_PAD:
       param->pad_mode_ = Pad_pad;
+      break;
+    default:
+      MS_LOG(ERROR) << "Pad mode does not support, " << pad_mode;
+      return RET_NOT_SUPPORT;
   }
 
   switch (act_type) {
@@ -54,12 +58,9 @@ int SetPadAndAct(schema::PadMode pad_mode, schema::ActivationType act_type, Conv
 
 OpParameter *PopulateConvParameter(const void *prim) {
   auto primitive = static_cast<const schema::Primitive *>(prim);
-  MS_ASSERT(primitive != nullptr);
+  MS_CHECK_TRUE_MSG(primitive != nullptr, nullptr, "primitive is nullptr.");
   auto value = primitive->value_as_Conv2DFusion();
-  if (value == nullptr) {
-    MS_LOG(ERROR) << "value is nullptr";
-    return nullptr;
-  }
+  MS_CHECK_TRUE_MSG(value != nullptr, nullptr, "value is nullptr.");
 
   auto *param = reinterpret_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
   if (param == nullptr) {
@@ -95,7 +96,14 @@ OpParameter *PopulateConvParameter(const void *prim) {
     free(param);
     return nullptr;
   }
-
+  for (size_t i = 0; i <= 1; i++) {
+    auto stride_item = *(stride->begin() + i);
+    if (stride_item < 0 || stride_item > static_cast<int64_t>(INT32_MAX)) {
+      MS_LOG(ERROR) << "strides has invalid num.";
+      free(param);
+      return nullptr;
+    }
+  }
   param->group_ = static_cast<int>(value->group());
   param->stride_h_ = static_cast<int>(*(stride->begin()));
   param->stride_w_ = static_cast<int>(*(stride->begin() + 1));
@@ -105,6 +113,14 @@ OpParameter *PopulateConvParameter(const void *prim) {
     param->pad_l_ = 0;
     param->pad_r_ = 0;
   } else {
+    for (size_t i = 0; i <= kOffsetThree; i++) {
+      auto pad_item = *(pad_list->begin() + i);
+      if (pad_item < 0 || pad_item > static_cast<int64_t>(INT32_MAX)) {
+        MS_LOG(ERROR) << "pad list has invalid num.";
+        free(param);
+        return nullptr;
+      }
+    }
     param->pad_u_ = static_cast<int>(*(pad_list->begin()));
     param->pad_d_ = static_cast<int>(*(pad_list->begin() + 1));
     param->pad_l_ = static_cast<int>(*(pad_list->begin() + kOffsetTwo));

mindspore/lite/src/runtime/kernel/arm/fp32/addn_fp32.cc

@@ -93,6 +93,11 @@ int AddNCPUKernel::Run() {
     }
     return RET_OK;
   }
+  if (std::any_of(in_tensors_.begin(), in_tensors_.end(),
+                  [this](const lite::Tensor *input) { return input->shape() != out_tensors_.at(0)->shape(); })) {
+    MS_LOG(ERROR) << "all inputs should have the same shape of output.";
+    return RET_ERROR;
+  }
   in1_addr_ = input0_data;
   in2_addr_ = input1_data;
   out_addr_ = output_data;

mindspore/lite/src/runtime/kernel/arm/fp32/scale_fp32.cc

@@ -27,6 +27,24 @@ using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_ScaleFusion;
 
 namespace mindspore::kernel {
+namespace {
+int CheckInputsOutputsDataType(const std::vector<lite::Tensor *> &in_tensors,
+                               const std::vector<lite::Tensor *> &out_tensors) {
+  if (std::any_of(in_tensors.begin(), in_tensors.end(), [](const lite::Tensor *input) {
+        return input->data_type() != kNumberTypeFloat && input->data_type() != kNumberTypeFloat32;
+      })) {
+    MS_LOG(ERROR) << "scale op input data type should float32";
+    return RET_ERROR;
+  }
+  if (std::any_of(out_tensors.begin(), out_tensors.end(), [](const lite::Tensor *output) {
+        return output->data_type() != kNumberTypeFloat && output->data_type() != kNumberTypeFloat32;
+      })) {
+    MS_LOG(ERROR) << "scale op output data type should float32";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+}  // namespace
 ScaleCPUKernel::~ScaleCPUKernel() {
   if (scale_param_->const_scale_) {
     if (scale_ != nullptr) {
@@ -121,7 +139,12 @@ int ScaleCPUKernel::CalculateParameter() {
 int ScaleCPUKernel::Prepare() {
   CHECK_LESS_RETURN(in_tensors_.size(), C2NUM);
   CHECK_LESS_RETURN(out_tensors_.size(), 1);
-  auto ret = InitScaleOffset();
+  auto ret = CheckInputsOutputsDataType(in_tensors_, out_tensors_);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Scale inputs or outputs data type is invalid.";
+    return RET_ERROR;
+  }
+  ret = InitScaleOffset();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Scale fp32 InitScaleOffset failed.";
     return RET_ERROR;

mindspore/lite/tools/converter/parser/tflite/tflite_argmax_parser.cc

@@ -32,6 +32,7 @@ ops::PrimitiveC *TfliteArgmaxParser::Parse(const std::unique_ptr<tflite::Operato
   prim->set_out_max_value(false);
   prim->set_top_k(1);
 
+  MS_CHECK_TRUE_MSG(tflite_op->inputs.size() >= kInputSize1, nullptr, "argmax input size should be greater than 1.");
   const auto &axis_tensor = tflite_subgraph->tensors.at(tflite_op->inputs[1]);
   MS_CHECK_TRUE_MSG(axis_tensor != nullptr, nullptr, "axis_tensor is nullptr");
   const auto &buf_data = tflite_model->buffers.at(axis_tensor->buffer);

mindspore/lite/tools/converter/parser/tflite/tflite_conv_parser.cc

@@ -27,6 +27,70 @@ constexpr int kWeightChannelOut = 0;
 constexpr int kWeightKernelH = 1;
 constexpr int kWeightKernelW = 2;
 constexpr int kWeightChannelIn = 3;
+STATUS GetConvPaddingParam(const std::unique_ptr<tflite::TensorT> &tensor, mindspore::PadMode pad_mode,
+                           const ops::Conv2DFusion *conv_prim, std::vector<int64_t> *params) {
+  MSLITE_CHECK_PTR(tensor);
+  MSLITE_CHECK_PTR(params);
+  MSLITE_CHECK_PTR(conv_prim);
+  if (tensor->shape.empty()) {
+    MS_LOG(DEBUG) << "the tensor's shape is dynamic, which obtain only when running.";
+    return RET_NO_CHANGE;
+  }
+  int pad_u = 0;
+  int pad_d = 0;
+  int pad_l = 0;
+  int pad_r = 0;
+  if (pad_mode == mindspore::PadMode::SAME) {
+    auto shape = tensor->shape;
+    MS_CHECK_TRUE_RET(shape.size() == DIMENSION_4D, RET_ERROR);
+    int input_h = shape.at(kNHWC_H);
+    int input_w = shape.at(kNHWC_W);
+    auto strides = conv_prim->get_stride();
+    MS_CHECK_TRUE_MSG(strides.size() > 1, RET_ERROR, "conv stride param is invalid.");
+    auto dilates = conv_prim->get_dilation();
+    MS_CHECK_TRUE_MSG(dilates.size() > 1, RET_ERROR, "conv dilation param is invalid.");
+    auto kernel_size = conv_prim->get_kernel_size();
+    MS_CHECK_TRUE_MSG(kernel_size.size() > 1, RET_ERROR, "conv kernel_size param is invalid.");
+    int stride_h = strides[0];
+    int stride_w = strides[1];
+    int dilate_h = dilates[0];
+    int dilate_w = dilates[1];
+    int kernel_h = kernel_size[0];
+    int kernel_w = kernel_size[1];
+    MS_CHECK_TRUE_MSG(stride_h != 0, RET_ERROR, "stride_h shouldn't be 0");
+    MS_CHECK_TRUE_MSG(stride_w != 0, RET_ERROR, "stride_w shouldn't be 0");
+    int output_w = ceil(static_cast<float>(input_w) / static_cast<float>(stride_w));
+    int output_h = ceil(static_cast<float>(input_h) / static_cast<float>(stride_h));
+    if (INT_MUL_OVERFLOW(output_h - 1, stride_h) || INT_MUL_OVERFLOW(kernel_h - 1, dilate_h)) {
+      MS_LOG(ERROR) << "int mul overflow";
+      return RET_ERROR;
+    }
+    int pad_h_all = ((output_h - 1) * stride_h + (kernel_h - 1) * dilate_h + 1 - input_h);
+    if (INT_MUL_OVERFLOW(output_w - 1, stride_w) || INT_MUL_OVERFLOW(kernel_w - 1, dilate_w)) {
+      MS_LOG(ERROR) << "int mul overflow";
+      return RET_ERROR;
+    }
+    int pad_w_all = ((output_w - 1) * stride_w + (kernel_w - 1) * dilate_w + 1 - input_w);
+    if (pad_h_all < 0) {
+      pad_u = pad_d = 0;
+    } else {
+      pad_u = pad_h_all / 2;
+      pad_d = pad_h_all - pad_u;
+    }
+    if (pad_w_all < 0) {
+      pad_l = pad_r = 0;
+    } else {
+      pad_l = pad_w_all / 2;
+      pad_r = pad_w_all - pad_l;
+    }
+  }
+
+  params->emplace_back(pad_u);
+  params->emplace_back(pad_d);
+  params->emplace_back(pad_l);
+  params->emplace_back(pad_r);
+  return RET_OK;
+}
 }  // namespace
 ops::PrimitiveC *TfliteConvParser::Parse(const std::unique_ptr<tflite::OperatorT> &tflite_op,
                                          const std::unique_ptr<tflite::SubGraphT> &tflite_subgraph,
@@ -72,8 +136,7 @@ ops::PrimitiveC *TfliteConvParser::Parse(const std::unique_ptr<tflite::OperatorT
   // calculate pad params
   const auto &dataTensor = tflite_subgraph->tensors.at(tflite_op->inputs[0]);
   std::vector<int64_t> params;
-  int status = getPaddingParam(dataTensor, padMode, tflite_attr->stride_h, tflite_attr->stride_w,
-                               weight_shape[kWeightKernelH], weight_shape[kWeightKernelW], &params);
+  int status = GetConvPaddingParam(dataTensor, padMode, prim.get(), &params);
   if (status != RET_OK && status != RET_NO_CHANGE) {
     MS_LOG(ERROR) << "get padding params failed";
     return nullptr;
@@ -146,8 +209,7 @@ ops::PrimitiveC *TfliteDepthwiseConv2DParser::Parse(const std::unique_ptr<tflite
 
   // calculate pad params
   std::vector<int64_t> params;
-  int status = getPaddingParam(data_tensor, padMode, tflite_attr->stride_h, tflite_attr->stride_w,
-                               weight_shape[kWeightKernelH], weight_shape[kWeightKernelW], &params);
+  int status = GetConvPaddingParam(data_tensor, padMode, prim.get(), &params);
   if (status != RET_OK && status != RET_NO_CHANGE) {
     MS_LOG(ERROR) << "get padding params failed";
     return nullptr;