Fix prefast warnings (#13017)

Fix prefast warnings:
[C26451](https://learn.microsoft.com/en-us/cpp/code-quality/C26451?view=msvc-170)
[C26436](https://learn.microsoft.com/en-us/cpp/code-quality/c26436?view=msvc-170)
[C26814](https://learn.microsoft.com/en-us/cpp/code-quality/C26814?view=msvc-170)

onnxruntime/contrib_ops/cpu/bert/attention_cpu_base.h

@@ -144,7 +144,9 @@ class AttentionCPUBase : public AttentionBase {
 
           // Broadcast mask data: (Bx)SxS* -> (BxNx)SxS*
           if (mask_data != nullptr) {
-            memcpy(output, mask_data + mask_offset, sequence_length * all_sequence_length * sizeof(T));
+            memcpy(output,
+                   mask_data + mask_offset,
+                   static_cast<size_t>(sequence_length) * all_sequence_length * sizeof(T));
           }
 
           const T* k = K + input_chunk_length * i;

onnxruntime/contrib_ops/cpu/bert/attention_helper.h

@@ -91,7 +91,7 @@ void PrepareMask(const int32_t* mask_index,
             p_mask[s_i * all_sequence_length + m_i] += static_cast<T>(-10000.0f);
           }
         }
-        p_mask += sequence_length * all_sequence_length;
+        p_mask += static_cast<size_t>(sequence_length) * all_sequence_length;
       }
     }
 

onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_base.h

@@ -141,11 +141,13 @@ class BeamSearchBase : public GenerateBase  {
     parameters_->ParseFromInputs(&context);
   }
 
+  ~BeamSearchBase() override = default;
+
   // Initialize by validating all the inputs, and allocating the output tensors.
-  Status Initialize();
+  Status Initialize() override;
 
   // Validate inputs.
-  Status CheckInputs(const OpKernelContextInternal& context);
+  Status CheckInputs(const OpKernelContextInternal& context) override;
 
  protected:
   // Process logits and append next tokens to sequences.

onnxruntime/contrib_ops/cpu/transformers/generate_impl_base.h

@@ -56,6 +56,8 @@ class GenerateBase {
                          ->GetAllocator(0, OrtMemTypeDefault);
   }
 
+  virtual ~GenerateBase() = default;
+
   // Initialize by validating all the inputs, and allocating the output tensors.
   virtual Status Initialize() = 0;
 

onnxruntime/contrib_ops/cpu/transformers/generation_device_helper.cc

@@ -449,7 +449,7 @@ Status GreedySearchProcessLogits(
   const Tensor& input = next_token_scores_value.Get<Tensor>();
 
   constexpr int axis = 1;
-  const unsigned top_k = static_cast<unsigned>(1);
+  constexpr unsigned top_k = 1;
   constexpr bool largest = true;
   constexpr bool sorted = false;
 
@@ -771,7 +771,7 @@ Status UpdateDecoderFeeds(
 #endif
 
   // Update past state
-  ORT_ENFORCE(last_outputs.size() >= static_cast<size_t>(1 + num_present_tensors));
+  ORT_ENFORCE(last_outputs.size() >= static_cast<size_t>(1) + num_present_tensors);
   // TODO(tianleiwu): remove num_beams==1 once GreedySearch operator is available.
   if (num_beams == 1) {
     // feed present_* output to past_* inputs one by one

onnxruntime/contrib_ops/cpu/transformers/greedy_search_impl_base.h

@@ -92,11 +92,13 @@ class GreedySearchBase : public GenerateBase {
     parameters_->ParseFromInputs(&context);
   }
 
+  ~GreedySearchBase() override = default;
+
   // Initialize by validating all the inputs, and allocating the output tensors.
-  Status Initialize();
+  Status Initialize() override;
 
   // Validate inputs.
-  Status CheckInputs(const OpKernelContextInternal& context);
+  Status CheckInputs(const OpKernelContextInternal& context) override;
 
  protected:
   // Process logits and append next tokens to sequences.

onnxruntime/contrib_ops/cpu/transformers/greedy_search_impl_gpt.h

@@ -225,7 +225,9 @@ Status GreedySearchGpt<T>::Execute(const FeedsFetchesManager& feeds_fetches_mana
   // Copy the sequences to output
   gsl::span<int32_t> output = output_sequences->MutableDataAsSpan<int32_t>();
   for (int batch_id = 0; batch_id < parameters->batch_size; ++batch_id) {
-    auto batch_output = output.subspan(batch_id * parameters->max_length, parameters->max_length);
+    auto batch_output = output.subspan(
+      static_cast<size_t>(batch_id) * parameters->max_length,
+      parameters->max_length);
     gsl::span<const int32_t> sequence_source = greedy_state.sequences.GetSequence(batch_id);
     gsl::copy(sequence_source, batch_output);
   }

onnxruntime/contrib_ops/cpu/transformers/subgraph_t5_decoder.cc

@@ -161,7 +161,7 @@ Status T5DecoderSubgraph::CreateInitialFeeds(
   // When first_past_input_index_ == 3, the encoder_hidden_states and past states are copied from the second output
   // of encoder.
   // When first_past_input_index_ == 2, the past states are copied from the second output of encoder.
-  for (size_t j = 4 - first_past_input_index_; j < encoder_fetches.size(); j++) {
+  for (size_t j = static_cast<size_t>(4) - first_past_input_index_; j < encoder_fetches.size(); j++) {
     if (j == 1) {
       ORT_RETURN_IF(has_hidden_state_ == false, "Invalid hidden_states expension: has_hidden_state_ == false");
       OrtValue expanded_hidden_states;

onnxruntime/contrib_ops/cuda/bert/attention.cc

@@ -123,7 +123,8 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
   int m = batch_size * sequence_length;
   int n = 3 * hidden_size;
   int k = input_hidden_size;
-  auto gemm_buffer = GetScratchBuffer<T>(batch_size * sequence_length * 3 * hidden_size * element_size);
+  size_t gemm_buffer_size = static_cast<size_t>(batch_size) * sequence_length * 3 * hidden_size * element_size;
+  auto gemm_buffer = GetScratchBuffer<T>(gemm_buffer_size);
 
   typedef typename ToCudaType<T>::MappedType CudaT;
   CudaT one = ToCudaType<T>::FromFloat(1.0f);

onnxruntime/contrib_ops/cuda/bert/longformer_attention.cc

@@ -79,7 +79,7 @@ Status LongformerAttention<T>::ComputeInternal(OpKernelContext* context) const {
 
   // TODO(tianleiwu): only calculate global index once per model instead of once per LongformerAttention node.
   // Build Global Index
-  auto global_index_buffer = GetScratchBuffer<int>(batch_size * sequence_length);
+  auto global_index_buffer = GetScratchBuffer<int>(static_cast<size_t>(batch_size) * sequence_length);
   auto batch_global_num_buffer = GetScratchBuffer<int>(batch_size);
 
   size_t global_scratch_bytes = GetGlobalScratchSize(sequence_length);

onnxruntime/contrib_ops/cuda/transformers/generation_device_helper.cc

@@ -249,7 +249,7 @@ Status ProcessLogits(const OrtValue& logits,                                 //
     const CudaT* current_logits = logits_data + (input_length - 1) * vocab_size;
     for (int i = 0; i < batch_beam_size; i++) {
       gsl::span<const T> source(reinterpret_cast<const T*>(current_logits), vocab_size);
-      gsl::span<T> target = next_token_logits.subspan(i * vocab_size, vocab_size);
+      gsl::span<T> target = next_token_logits.subspan(static_cast<size_t>(i) * vocab_size, vocab_size);
       CUDA_RETURN_IF_ERROR(cudaMemcpyAsync(target.data(), source.data(), sizeof(T) * vocab_size,
                                            cudaMemcpyDeviceToDevice, cuda_stream));
       if (logits_batch_size == batch_beam_size) {

onnxruntime/core/providers/cuda/generator/range.cc

@@ -64,7 +64,8 @@ static Status ComputeRange(cudaStream_t stream, OpKernelContext* ctx) {
     return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, "delta in Range operator can not be zero!");
   }
 
-  int count = static_cast<int>(ceil(1.0 * (limit - start) / delta));
+  double num = (static_cast<double>(limit) - static_cast<double>(start)) / static_cast<double>(delta);
+  int count = static_cast<int>(ceil(num));
   if (count <= 0)
     count = 0;
   TensorShape shape = {static_cast<int64_t>(count)};