!14376 [MSLITE][DEVELOP] clear static check warnings in module of lite runtime op

From: @yangruoqi713
Reviewed-by: @zhang_xue_tong,@zhanghaibo5,@hangangqiang
Signed-off-by: @zhanghaibo5

mindspore/lite/nnacl/fp16/activation_fp16.c

@@ -195,7 +195,7 @@ int GeluFp16(const float16_t *src, int length, float16_t *dst, bool approximate)
     int C8 = UP_ROUND(length, C8NUM);
     for (; i < C8; i += C8NUM) {
       float16x8_t in = vld1q_f16(src + i);
-      float16x8_t res = 0.5 * in * (1.0 + MS_ERFX8_F16(in / (float16_t)1.4142135623730951f));
+      const float16x8_t res = 0.5 * in * (1.0 + MS_ERFX8_F16(in / (float16_t)1.4142135623730951f));
       vst1q_f16(dst + i, res);
     }
 #endif

mindspore/lite/nnacl/fp16/exp_fp16.h

@@ -59,7 +59,7 @@ static inline void single_exp_fp16(float16_t src, float16_t *dst) {
   int integer = (float)src / param[0];
   float decimal = (float)src - integer * param[0];
   int int_exp = (integer + 127) << 23;
-  float decimal_exp =
+  const float decimal_exp =
     1.0f + decimal * (1.0f + decimal * (0.5f + decimal * (param[3] + decimal * (param[2] + decimal * param[1]))));
   *dst = (float16_t)(*((float *)&int_exp) * decimal_exp);
 }

mindspore/lite/nnacl/fp32/conv_depthwise_fp32.c

@@ -339,7 +339,7 @@ bool CheckConvDw1DWinograd(const ConvParameter *conv_param, int thread_num) {
          conv_param->stride_h_ == 1 && conv_param->dilation_h_ == 1 && conv_param->dilation_w_ == 1 &&
          conv_param->pad_u_ == 1 && conv_param->pad_d_ == 1 && conv_param->pad_l_ == 1 && conv_param->pad_r_ == 1 &&
          conv_param->input_channel_ == conv_param->output_channel_ &&
-         conv_param->output_h_ / thread_num >= 4;  // better had more than 4 rows for each thread
+         conv_param->output_h_ >= thread_num * 4;  // better had more than 4 rows for each thread
 }
 
 void ConvDw3x3RowLeft(const float *src, float *line, int lw, int channel) {

mindspore/lite/nnacl/fp32_grad/layernorm_grad.c

@@ -20,7 +20,7 @@
 void LayerNormGrad(const float *x, const float *dy, const float *var, const float *mean, const float *gamma,
                    int param_num, int param_size, int block_num, int block_size, float *dx, float *dg, float *db) {
   // var is actually layer_norm forward output var
-  float eps = 1e-12;
+  const float eps = 1e-12;
   const float *var_sqrt_rev = var;
   for (size_t i = 0; i < param_num; ++i) {
     float dgamma = 0.0f;
@@ -37,23 +37,23 @@ void LayerNormGrad(const float *x, const float *dy, const float *var, const floa
     float sum1 = 0.0f;
     float sum2 = 0.0f;
     float sum3 = 0.0f;
-    for (size_t j = i * block_size; j < (i + 1) * block_size; ++j) {
+    for (size_t j = 0; j < block_size; ++j) {
-      int param_shift = j % param_num;
+      int index = i * block_size + j;
-      int norm_shift = (int)(j / block_size);
+      float dxm = x[index] - mean[i];
-      float dxm = x[j] - mean[norm_shift];
+      int param_shift = index % param_num;
-      float dyg = dy[j] * gamma[param_shift];
+      float dyg = dy[index] * gamma[param_shift];
-      sum1 += -0.5f * dyg * dxm * pow(var_sqrt_rev[norm_shift] + eps, -1.5);
+      sum1 += -0.5f * dyg * dxm * pow(var_sqrt_rev[i] + eps, -1.5);
       sum2 += dyg;
       sum3 += -2.0f * dxm;
     }
-    for (size_t j = i * block_size; j < (i + 1) * block_size; ++j) {
+    for (size_t j = 0; j < block_size; ++j) {
-      int param_shift = j % param_num;
+      int index = i * block_size + j;
-      int norm_shift = (int)(j / block_size);
+      float var_sqrt = pow(var_sqrt_rev[i] + eps, -0.5);
-      float var_sqrt = pow(var_sqrt_rev[norm_shift] + eps, -0.5);
+      int param_shift = index % param_num;
-      float dx1 = dy[j] * gamma[param_shift] * var_sqrt;
+      float dx1 = dy[index] * gamma[param_shift] * var_sqrt;
-      float dx2 = sum1 * 2.0f / block_size * (x[j] - mean[norm_shift]);
+      float dx2 = sum1 * 2.0f / block_size * (x[index] - mean[i]);
       float dx3 = (-1.0f * var_sqrt * sum2 + (1.0f / block_size) * sum1 * sum3) * (1.0f / block_size);
-      dx[j] = dx1 + dx2 + dx3;
+      dx[index] = dx1 + dx2 + dx3;
     }
   }
 }

mindspore/lite/nnacl/fp32_grad/pooling_grad.c

@@ -144,8 +144,7 @@ void MaxPoolingGrad(const float *input_ptr, const float *dy_ptr, float *output_p
               int xw = yw * stride_w + kw - pad_w;
               int val_idx = (xw + in_w * xh) * channel + ic;
 #ifdef ENABLE_ARM
-              unsigned int val_idx_vec[] = {val_idx, val_idx + 1, val_idx + 2, val_idx + 3};
+              uint32x4_t index = {val_idx, val_idx + 1, val_idx + 2, val_idx + 3};
-              uint32x4_t index = vld1q_u32(val_idx_vec);
               float32x4_t in = vld1q_f32(inPtr + val_idx);
               max_idx = MaxIndex(in, &max_val, index, max_idx);
 #else

mindspore/lite/nnacl/fp32_grad/resize_grad.c

@@ -52,17 +52,17 @@ void ResizeBiLinearGrad(float *in_addr, float *out_addr, int batch_size, int cha
       size_t h = i / param->in_width_;
       size_t w = i % param->in_width_;
       for (int32_t c = 0; c < channel; ++c) {
-        float in_y = (float)h * param->height_scale_;
+        const float in_y = (float)h * param->height_scale_;
         size_t top_y_index = MSMAX((size_t)(floorf(in_y)), (size_t)(0));
         size_t bottom_y_index = MSMIN((size_t)(ceilf(in_y)), param->out_height_ - 1);
-        float y_lerp = in_y - floorf(in_y);
+        const float y_lerp = in_y - floorf(in_y);
-        float inverse_y_lerp = 1.0 - y_lerp;
+        const float inverse_y_lerp = 1.0 - y_lerp;
 
-        float in_x = (float)w * param->width_scale_;
+        const float in_x = (float)w * param->width_scale_;
         size_t left_x_index = MSMAX((size_t)(floorf(in_x)), (size_t)(0));
         size_t right_x_index = MSMIN((size_t)(ceilf(in_x)), param->out_width_ - 1);
-        float x_lerp = in_x - floorf(in_x);
+        const float x_lerp = in_x - floorf(in_x);
-        float inverse_x_lerp = 1.0 - x_lerp;
+        const float inverse_x_lerp = 1.0 - x_lerp;
 
         size_t in_offset = h * (param->in_width_ * channel) + (w * channel) + c;
         size_t out_offset_top_y_left_x = top_y_index * (param->out_width_ * channel) + (left_x_index * channel) + c;

mindspore/lite/nnacl/fp32_grad/unsorted_segment_sum.c

@@ -18,6 +18,9 @@
 
 int UnsortedSegmentSum(const float *input, int unit_num, int input_dim1, const int *indices, float *output,
                        int output_dim0, int output_dim1) {
+  if (input_dim1 == 0) {
+    return NNACL_ERR;
+  }
   for (int i = 0; i < unit_num; ++i) {
     int j = i / input_dim1;
     int k = i % input_dim1;

mindspore/lite/nnacl/infer/common_infer.c

@@ -368,6 +368,9 @@ int FftInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC **ou
 }
 
 int VectorCInit(VectorC *vc, size_t per_malloc_size) {
+  if (per_malloc_size == 0) {
+    return NNACL_ERR;
+  }
   vc->data_ = (int *)malloc(per_malloc_size * sizeof(int));
   if (vc->data_ == NULL) {
     return NNACL_ERR;

mindspore/lite/nnacl/infer/prior_box_infer.c

@@ -63,10 +63,10 @@ int PriorBoxInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC
   size_t min_sizes_size = param->min_sizes_size;
   size_t max_sizes_size = param->max_sizes_size;
   int32_t num_priors_box = min_sizes_size * different_aspect_ratios_size + max_sizes_size;
-  int kPriorBoxPoints = 4;
+  const int kPriorBoxPoints = 4;
-  int kPriorBoxN = 1;
+  const int kPriorBoxN = 1;
-  int kPriorBoxW = 1;
+  const int kPriorBoxW = 1;
-  int kPriorBoxC = 2;
+  const int kPriorBoxC = 2;
 
   int32_t h = GetHeight(input) * GetWidth(input) * num_priors_box * kPriorBoxPoints;
   output->shape_size_ = 4;

mindspore/lite/nnacl/infer/reduce_infer.c

@@ -68,7 +68,7 @@ int ReduceInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC *
   }
   bool keep_dims = param->keep_dims_;
   int out_shape[MAX_SHAPE_SIZE];
-  size_t out_shape_size = 0;
+  const size_t out_shape_size = 0;
   // get axes from input tensor
   const TensorC *axes_input = inputs[1];
   if (axes_input->shape_size_ == 1 && axes_input->shape_[0] == 0) {

mindspore/lite/nnacl/infer/reshape_infer.c

@@ -75,6 +75,9 @@ int CalNewShape(const TensorC *in_tensor, int *out_shape, size_t out_shape_size)
 
 int CalShapeByType(const TensorC *const *inputs, size_t shape_size, int *out_shape, size_t *out_shape_size) {
   const TensorC *shape_tensor = inputs[1];
+  if (shape_size == 0) {
+    return NNACL_ERR;
+  }
   switch (shape_tensor->data_type_) {
     case kNumberTypeInt8: {
       int8_t *data = (int8_t *)(shape_tensor->data_);

mindspore/lite/nnacl/infer/strided_slice_infer.c

@@ -59,8 +59,7 @@ int HandleAxesCheckNull(const TensorC *input_tensor, const TensorC *begin_tensor
   return NNACL_OK;
 }
 
-int HandleAxesInputExist(const TensorC *const *inputs, int *ndim_, int *in_shape_, int *begins_, int *strides_,
+int HandleAxesInputExist(const TensorC *const *inputs, int *ndim, int *in_shape, int *begins, int *strides, int *ends) {
-                         int *ends_) {
   const TensorC *input_tensor = inputs[0];
   const TensorC *begin_tensor = inputs[1];
   int *begin_data = (int *)(begin_tensor->data_);
@@ -73,7 +72,7 @@ int HandleAxesInputExist(const TensorC *const *inputs, int *ndim_, int *in_shape
   }
 
   // when input contains axes, begins, ends, strides will be expand to the same length as input rank
-  *ndim_ = (int)(input_tensor->shape_size_);
+  *ndim = (int)(input_tensor->shape_size_);
   int begin_ndim = GetElementNum(begin_tensor);
 
   int *axes_data = NULL;
@@ -111,20 +110,20 @@ int HandleAxesInputExist(const TensorC *const *inputs, int *ndim_, int *in_shape
     }
     for (int i = 0; i < begin_ndim; ++i) {
       if (axes[i] < 0) {
-        axes[i] += *ndim_;
+        axes[i] += *ndim;
       }
     }
   }
 
-  for (size_t i = 0; i < *ndim_; i++) {
+  for (size_t i = 0; i < *ndim; i++) {
-    in_shape_[i] = 0;
+    in_shape[i] = 0;
-    begins_[i] = 0;
+    begins[i] = 0;
-    strides_[i] = 0;
+    strides[i] = 0;
   }
-  for (size_t i = 0; i < *ndim_; ++i) {
+  for (size_t i = 0; i < *ndim; ++i) {
-    in_shape_[i] = input_tensor->shape_[i];
+    in_shape[i] = input_tensor->shape_[i];
   }
-  for (size_t i = 0; i < *ndim_; ++i) {
+  for (size_t i = 0; i < *ndim; ++i) {
     int axes_it = 0;
     for (size_t j = 0; j < begin_ndim; j++) {
       if (axes[j] == i) {
@@ -137,13 +136,13 @@ int HandleAxesInputExist(const TensorC *const *inputs, int *ndim_, int *in_shape
     if (axes_it != begin_ndim) {
       int axis = axes_it;
       // begins or ends exceed limit will be set to limit
-      begins_[i] = imax(imin(begin_data[axis], input_tensor->shape_[i] - 1), -input_tensor->shape_[i]);
+      begins[i] = imax(imin(begin_data[axis], input_tensor->shape_[i] - 1), -input_tensor->shape_[i]);
-      ends_[i] = imax(imin(end_data[axis], input_tensor->shape_[i]), -input_tensor->shape_[i] - 1);
+      ends[i] = imax(imin(end_data[axis], input_tensor->shape_[i]), -input_tensor->shape_[i] - 1);
-      strides_[i] = stride_data[axis];
+      strides[i] = stride_data[axis];
     } else {
-      begins_[i] = 0;
+      begins[i] = 0;
-      ends_[i] = input_tensor->shape_[i];
+      ends[i] = input_tensor->shape_[i];
-      strides_[i] = 1;
+      strides[i] = 1;
     }
   }
   return NNACL_OK;
@@ -174,18 +173,18 @@ void Bit2Vector(StridedSliceTransferBuffer *transfer_buffer, StridedSliceParamet
   }
 }
 
-void ApplyNewAxisMask(StridedSliceTransferBuffer *transfer_buffer, StridedSliceParameter *param, int *in_shape_,
+void ApplyNewAxisMask(StridedSliceTransferBuffer *transfer_buffer, StridedSliceParameter *param, int *in_shape,
-                      size_t *in_shape_size) {
+                      size_t *out_shape_size) {
   for (size_t i = 0; i < transfer_buffer->new_axis_mask_size_; i++) {
     if (transfer_buffer->new_axis_mask_[i]) {
       transfer_buffer->ndim_ += 1;
-      ShapeInsert(in_shape_, in_shape_size, i, 1);
+      ShapeInsert(in_shape, out_shape_size, i, 1);
       transfer_buffer->begins_[i] = 0;
       transfer_buffer->ends_[i] = 1;
       transfer_buffer->strides_[i] = 1;
 
       ShapePush(transfer_buffer->begins_, &transfer_buffer->begins_size_, 0);
-      ShapePush(transfer_buffer->ends_, &transfer_buffer->ends_size_, in_shape_[transfer_buffer->ndim_ - 1]);
+      ShapePush(transfer_buffer->ends_, &transfer_buffer->ends_size_, in_shape[transfer_buffer->ndim_ - 1]);
       ShapePush(transfer_buffer->strides_, &transfer_buffer->strides_size_, 1);
 
       transfer_buffer->begins_mask_[i] = false;
@@ -204,33 +203,45 @@ void ApplyBeginMask(StridedSliceTransferBuffer *transfer_buffer) {
   }
 }
 
-void ApplyEndMask(StridedSliceTransferBuffer *transfer_buffer, int *in_shape_) {
+int ApplyEndMask(StridedSliceTransferBuffer *transfer_buffer, const int *in_shape, size_t in_shape_size) {
   for (int i = 0; i < transfer_buffer->ndim_; i++) {
     if (transfer_buffer->ends_mask_[i]) {
-      transfer_buffer->ends_[i] = in_shape_[i];
+      if (i >= in_shape_size) {
+        return NNACL_ERR;
+      }
+      transfer_buffer->ends_[i] = in_shape[i];
     }
   }
+  return NNACL_OK;
 }
 
-void ApplyEllipsisMask(StridedSliceTransferBuffer *transfer_buffer, int *in_shape_) {
+int ApplyEllipsisMask(StridedSliceTransferBuffer *transfer_buffer, const int *in_shape, size_t in_shape_size) {
   for (size_t i = 0; i < transfer_buffer->ellipsis_mask_size_; i++) {
     if (transfer_buffer->ellipsis_mask_[i]) {
+      if (i >= in_shape_size) {
+        return NNACL_ERR;
+      }
       transfer_buffer->begins_[i] = 0;
-      transfer_buffer->ends_[i] = in_shape_[i];
+      transfer_buffer->ends_[i] = in_shape[i];
       break;
     }
   }
+  return NNACL_OK;
 }
 
-void TransIndexToPositive(StridedSliceTransferBuffer *transfer_buffer, int *in_shape_) {
+int TransIndexToPositive(StridedSliceTransferBuffer *transfer_buffer, const int *in_shape, size_t in_shape_size) {
-  for (int i = 0; i < (int)(transfer_buffer->begins_size_); ++i) {
+  for (size_t i = 0; i < transfer_buffer->begins_size_; i++) {
+    if (i >= in_shape_size) {
+      return NNACL_ERR;
+    }
     if (transfer_buffer->begins_[i] < 0) {
-      transfer_buffer->begins_[i] += in_shape_[i];
+      transfer_buffer->begins_[i] += in_shape[i];
     }
     if (transfer_buffer->ends_[i] < 0) {
-      transfer_buffer->ends_[i] += in_shape_[i];
+      transfer_buffer->ends_[i] += in_shape[i];
     }
   }
+  return NNACL_OK;
 }
 
 void ApplyShrinkMask(StridedSliceTransferBuffer *transfer_buffer, int *output_shape, size_t *output_shape_size) {
@@ -251,20 +262,25 @@ void ApplyShrinkMask(StridedSliceTransferBuffer *transfer_buffer, int *output_sh
   }
 }
 
-void TransferBuffer2Param(StridedSliceTransferBuffer *transfer_buffer, StridedSliceParameter *param, int *in_shape_) {
+int TransferBuffer2Param(StridedSliceTransferBuffer *transfer_buffer, StridedSliceParameter *param, const int *in_shape,
+                         size_t in_shape_size) {
+  if (transfer_buffer->ndim_ >= in_shape_size || param->in_shape_length_ >= in_shape_size) {
+    return NNACL_ERR;
+  }
   for (int i = 0; i < transfer_buffer->ndim_; i++) {
     param->begins_[i] = transfer_buffer->begins_[i];
     param->ends_[i] = transfer_buffer->ends_[i];
-    param->in_shape_[i] = in_shape_[i];
+    param->in_shape_[i] = in_shape[i];
     param->strides_[i] = transfer_buffer->strides_[i];
   }
 
   for (int i = transfer_buffer->ndim_; i < param->in_shape_length_; i++) {
     param->begins_[i] = 0;
-    param->ends_[i] = in_shape_[i];
+    param->ends_[i] = in_shape[i];
-    param->in_shape_[i] = in_shape_[i];
+    param->in_shape_[i] = in_shape[i];
     param->strides_[i] = 1;
   }
+  return NNACL_OK;
 }
 
 void InitStridedSliceTransferBuffer(StridedSliceTransferBuffer *transfer_buffer) {
@@ -302,9 +318,9 @@ int StridedSliceInferShape(const TensorC *const *inputs, size_t inputs_size, Ten
     return NNACL_INFER_INVALID;
   }
 
-  int in_shape_[MAX_SHAPE_SIZE];
+  int in_shape[MAX_SHAPE_SIZE];
   size_t in_shape_size = 0;
-  ShapeSet(in_shape_, &in_shape_size, input->shape_, input->shape_size_);
+  ShapeSet(in_shape, &in_shape_size, input->shape_, input->shape_size_);
 
   StridedSliceTransferBuffer transfer_buffer;
   InitStridedSliceTransferBuffer(&transfer_buffer);
@@ -345,7 +361,7 @@ int StridedSliceInferShape(const TensorC *const *inputs, size_t inputs_size, Ten
   }
 
   if (inputs_size == 5) {
-    int ret = HandleAxesInputExist(inputs, &transfer_buffer.ndim_, in_shape_, transfer_buffer.begins_,
+    int ret = HandleAxesInputExist(inputs, &transfer_buffer.ndim_, in_shape, transfer_buffer.begins_,
                                    transfer_buffer.strides_, transfer_buffer.ends_);
     if (ret != NNACL_OK) {
       return ret;
@@ -355,15 +371,24 @@ int StridedSliceInferShape(const TensorC *const *inputs, size_t inputs_size, Ten
   // set all mask to original input shape
   SetMaskSize(&transfer_buffer);
   Bit2Vector(&transfer_buffer, param);
-  ApplyNewAxisMask(&transfer_buffer, param, in_shape_, &in_shape_size);
+  ApplyNewAxisMask(&transfer_buffer, param, in_shape, &in_shape_size);
   ApplyBeginMask(&transfer_buffer);
-  ApplyEndMask(&transfer_buffer, in_shape_);
+  int ret = ApplyEndMask(&transfer_buffer, in_shape, MAX_SHAPE_SIZE);
-  ApplyEllipsisMask(&transfer_buffer, in_shape_);
+  if (ret != NNACL_OK) {
+    return ret;
+  }
+  ret = ApplyEllipsisMask(&transfer_buffer, in_shape, MAX_SHAPE_SIZE);
+  if (ret != NNACL_OK) {
+    return ret;
+  }
 
   int output_shape[MAX_SHAPE_SIZE];
   size_t output_shape_size = 0;
-  ShapeSet(output_shape, &output_shape_size, in_shape_, in_shape_size);
+  ShapeSet(output_shape, &output_shape_size, in_shape, in_shape_size);
-  TransIndexToPositive(&transfer_buffer, in_shape_);
+  ret = TransIndexToPositive(&transfer_buffer, in_shape, MAX_SHAPE_SIZE);
+  if (ret != NNACL_OK) {
+    return ret;
+  }
   for (int i = 0; i < transfer_buffer.ndim_; i++) {
     if (transfer_buffer.strides_[i] == 0) {
       return NNACL_ERR;
@@ -374,8 +399,10 @@ int StridedSliceInferShape(const TensorC *const *inputs, size_t inputs_size, Ten
   }
   ApplyShrinkMask(&transfer_buffer, output_shape, &output_shape_size);
   SetShapeArray(outputs[0], output_shape, output_shape_size);
-  TransferBuffer2Param(&transfer_buffer, param, in_shape_);
+  ret = TransferBuffer2Param(&transfer_buffer, param, in_shape, MAX_SHAPE_SIZE);
-
+  if (ret != NNACL_OK) {
+    return ret;
+  }
   return NNACL_OK;
 }
 

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

@@ -31,6 +31,7 @@ OpParameter *PopulateSoftmaxParameter(const void *prim) {
   auto prim_softmax = primitive->value_as_Softmax();
   if (prim_softmax->axis()->size() != 1) {
     MS_LOG(ERROR) << "axis number invalid!number: " << prim_softmax->axis()->size();
+    free(softmax_param);
     return nullptr;
   }
   softmax_param->axis_ = prim_softmax->axis()->data()[0];

mindspore/lite/src/ops/populate/v0/split_populate_v0.cc

@@ -34,6 +34,7 @@ OpParameter *PopulateSplitParameter(const void *prim) {
   split_param->num_split_ = split_prim->numberSplit();
   if (split_param->num_split_ > std::numeric_limits<int>::max() / static_cast<int>(sizeof(int))) {
     MS_LOG(ERROR) << "The value of split_param->num_split_ is too big";
+    free(split_param);
     return nullptr;
   }
   int *split_sizes = reinterpret_cast<int *>(malloc(split_param->num_split_ * sizeof(int)));

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_creator_manager.h

@@ -51,7 +51,7 @@ inline ConvParameter *CreateNewConvParameter(ConvParameter *parameter) {
 
 inline void FreeMemory(ConvParameter *conv_param, const std::vector<lite::Tensor *> &new_inputs,
                        const std::vector<lite::Tensor *> &new_outputs) {
-  if (conv_param) {
+  if (conv_param != nullptr) {
     free(conv_param);
   }
   for (auto &in_tensor : new_inputs) {

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.cc

@@ -146,7 +146,7 @@ kernel::LiteKernel *ConvolutionDelegateCPUKernel::CpuConvFp32KernelSelect() {
     }
   }
 
-  if (kernel) {
+  if (kernel != nullptr) {
     auto ret = kernel->Init();
     if (ret != RET_OK) {
       MS_LOG(ERROR) << "conv kernel init failed.";

mindspore/lite/src/runtime/kernel/npu/convolution_npu.cc

@@ -112,12 +112,6 @@ kernel::LiteKernel *NpuConvKernelCreator(const std::vector<lite::Tensor *> &inpu
                                          const lite::InnerContext *ctx, const kernel::KernelKey &desc) {
   MS_ASSERT(op_parameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_Conv2DFusion);
-  if (inputs[0]->Size() > NPU_MEMORY_MAX) {
-    MS_LOG(ERROR) << "Npu does not support input tensor size greater than 200MB";
-    free(op_parameter);
-    return nullptr;
-  }
-
   auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter);
   kernel::NPUKernel *kernel = nullptr;
   if (conv_param->group_ == 1) {

mindspore/lite/src/runtime/kernel/npu/npu_kernel.h

@@ -27,7 +27,6 @@ using mindspore::kernel::LiteKernel;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
 namespace mindspore::kernel {
-#define NPU_MEMORY_MAX 200 * 1024 * 1024
 class NPUKernel : public LiteKernel {
  public:
   NPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs,
@@ -63,11 +62,6 @@ kernel::LiteKernel *NPUKernelCreator(const std::vector<lite::Tensor *> &inputs,
     free(op_parameter);
     return nullptr;
   }
-  if (inputs[0]->Size() > NPU_MEMORY_MAX) {
-    MS_LOG(ERROR) << "Npu does not support input tensor size greater than 200MB";
-    free(op_parameter);
-    return nullptr;
-  }
   auto *kernel = new (std::nothrow) T(op_parameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "kernel " << op_parameter->name_ << "is nullptr.";