clang-tidy

Author: shmsong

torch/csrc/jit/codegen/cuda/fusion.cpp

@@ -280,9 +280,11 @@ void Fusion::swapInputsTo(const std::vector<Val*>& inputs, bool update_tv_mem) {
   inputs_ = inputs;
   for (auto in : inputs_) {
     in->setIsFusionInput(true);
-    if (update_tv_mem && in->getValType().value() == ValType::TensorView) {
-      auto tv = in->as<TensorView>();
-      tv->setMemoryType(MemoryType::Global);
+    if (update_tv_mem) {
+      if (in->getValType().value() == ValType::TensorView) {
+        auto tv = in->as<TensorView>();
+        tv->setMemoryType(MemoryType::Global);
+      }
     }
   }
   all_tv_uses_valid_ = false;
@@ -297,9 +299,11 @@ void Fusion::swapOutputsTo(
   outputs_ = outputs;
   for (auto out : outputs_) {
     out->setIsFusionOutput(true);
-    if (update_tv_mem && out->getValType().value() == ValType::TensorView) {
-      auto tv = out->as<TensorView>();
-      tv->setMemoryType(MemoryType::Global);
+    if (update_tv_mem) {
+      if (out->getValType().value() == ValType::TensorView) {
+        auto tv = out->as<TensorView>();
+        tv->setMemoryType(MemoryType::Global);
+      }
     }
   }
   all_tv_uses_valid_ = false;

torch/csrc/jit/codegen/cuda/fusion_segmenter.h

@@ -232,8 +232,12 @@ class TORCH_CUDA_CU_API FusionSegmentGuard : public NonCopyable {
     fusion_->swapOutputsTo(new_outputs_, update_tv_mem_);
   }
 
+  // TODO: clang-tidy doesn't like the possibility of exceptions
+  //  from use of c10::optional and TORCH_INTERNAL_ASSERT, even though
+  //  there shouldn't be anyway this is throwing. Would
+  //  need an assertion free path for this in another cleanup.
+  // NOLINTNEXTLINE(bugprone-exception-escape)
   ~FusionSegmentGuard() {
-    FUSER_PERF_SCOPE("~Segmenter::FusionSegmentGuard");
     if (fusion_ == nullptr) {
       return;
     }

torch/csrc/jit/codegen/cuda/kernel_cache.cpp

@@ -345,7 +345,7 @@ FusionKernelRuntime* FusionExecutorCache::getKernelRuntimeFor(
         return kernel_runtime->getMaybeHeuristicsFor(inputs);
       });
 
-  FusionKernelRuntime* kernel_runtime;
+  FusionKernelRuntime* kernel_runtime = nullptr;
   if (reuse_it != kernel_runtimes.end()) {
     kernel_runtime = reuse_it->get();
   } else {
@@ -520,13 +520,13 @@ void FusionKernelRuntime::prepareRuntimeOrder() {
   std::unordered_set<Val*> available_input;
 
   // setup the order tensor dimensions are bound
-  for (size_t i = 0; i < segmented_fusion_->inputs().size(); i++) {
+  for (size_t i : c10::irange(segmented_fusion_->inputs().size())) {
     auto input_val = segmented_fusion_->inputs()[i];
     available_input.insert(input_val);
 
     if (auto input_tv = dynamic_cast<TensorView*>(input_val)) {
       auto root_dom = TensorDomain::noReductions(input_tv->getRootDomain());
-      for (size_t dim = 0; dim < root_dom.size(); dim++) {
+      for (size_t dim : c10::irange(root_dom.size())) {
         const auto extent = root_dom[dim]->extent();
         available_input.insert(extent);
         runtime_workspace_.group_extent_binding_order.push_back(extent);
@@ -559,8 +559,7 @@ void FusionKernelRuntime::prepareRuntimeOrder() {
         const auto& group_outputs = group->outputs();
 
         // Insert graph segment output to tensor map
-        for (size_t group_out_i = 0; group_out_i < group_outputs.size();
-             group_out_i++) {
+        for (size_t group_out_i : c10::irange(group_outputs.size())) {
           available_input.insert(group_outputs[group_out_i]);
         }
         group_ran[group_i] = true;