Fix all 64/32 bit warning in core/common_runtime.

Change: 152141388

tensorflow/core/common_runtime/bfc_allocator.cc

@@ -453,8 +453,8 @@ void BFCAllocator::RemoveFreeChunkIterFromBin(
 void BFCAllocator::RemoveFreeChunkFromBin(BFCAllocator::ChunkHandle h) {
   Chunk* c = ChunkFromHandle(h);
   CHECK(!c->in_use() && (c->bin_num != kInvalidBinNum));
-  int count = BinFromIndex(c->bin_num)->free_chunks.erase(h);
+  CHECK_GT(BinFromIndex(c->bin_num)->free_chunks.erase(h), 0)
-  CHECK(count > 0) << "Could not find chunk in bin";
+      << "Could not find chunk in bin";
   c->bin_num = kInvalidBinNum;
 }
 

tensorflow/core/common_runtime/bfc_allocator.h

@@ -78,7 +78,7 @@ class BFCAllocator : public VisitableAllocator {
 
   // A ChunkHandle is an index into the chunks_ vector in BFCAllocator
   // kInvalidChunkHandle means an invalid chunk
-  typedef int ChunkHandle;
+  typedef size_t ChunkHandle;
   static const int kInvalidChunkHandle = -1;
 
   typedef int BinNum;

tensorflow/core/common_runtime/device_mgr.cc

@@ -44,7 +44,7 @@ DeviceMgr::~DeviceMgr() {
 }
 
 StringPiece DeviceMgr::CopyToBackingStore(StringPiece s) {
-  int n = s.size();
+  size_t n = s.size();
   char* space = name_backing_store_.Alloc(n);
   memcpy(space, s.data(), n);
   return StringPiece(space, n);

tensorflow/core/common_runtime/direct_session.cc

@@ -427,7 +427,7 @@ Status DirectSession::Run(const RunOptions& run_options,
   TF_RETURN_IF_ERROR(SendInputs(inputs, executors_and_keys, run_state.rendez));
 
   // Start parallel Executors.
-  const int num_executors = executors_and_keys->items.size();
+  const size_t num_executors = executors_and_keys->items.size();
   ExecutorBarrier* barrier = new ExecutorBarrier(
       num_executors, run_state.rendez, [&run_state](const Status& ret) {
         {
@@ -458,7 +458,7 @@ Status DirectSession::Run(const RunOptions& run_options,
         options_.config.graph_options().build_cost_model();
     const int64 build_cost_model_after =
         options_.config.graph_options().build_cost_model_after();
-    int measure_step_count = executor_step_count - build_cost_model_after;
+    int64 measure_step_count = executor_step_count - build_cost_model_after;
     if (measure_step_count >= 0) {
       update_cost_model =
           ((measure_step_count + 1) % build_cost_model_every == 0);
@@ -611,7 +611,7 @@ Status DirectSession::PRunSetup(const std::vector<string>& input_names,
   }
 
   // Start parallel Executors.
-  const int num_executors = executors_and_keys->items.size();
+  const size_t num_executors = executors_and_keys->items.size();
   ExecutorBarrier* barrier = new ExecutorBarrier(
       num_executors, run_state->rendez, [run_state](const Status& ret) {
         if (!ret.ok()) {

tensorflow/core/common_runtime/executor.cc

@@ -232,7 +232,7 @@ struct NodeItem {
   int input_start = 0;
 
   // Number of output edges.
-  int num_output_edges;
+  size_t num_output_edges;
 
   PendingCounts::Handle pending_id;
 
@@ -307,7 +307,7 @@ class GraphView {
   void Initialize(const Graph* g);
   Status SetAllocAttrs(const Graph* g, const Device* device);
 
-  NodeItem* node(int id) const {
+  NodeItem* node(size_t id) const {
     DCHECK_GE(id, 0);
     DCHECK_LT(id, num_nodes_);
     uint32 offset = node_offsets_[id];
@@ -454,7 +454,7 @@ GraphView::~GraphView() {
 }
 
 size_t GraphView::NodeItemBytes(const Node* n) {
-  const int num_output_edges = n->out_edges().size();
+  const size_t num_output_edges = n->out_edges().size();
   const int num_inputs = n->num_inputs();
   const int num_outputs = n->num_outputs();
 
@@ -500,11 +500,11 @@ char* GraphView::InitializeNode(char* ptr, const Node* n) {
   // pointers). Casting to int64 is needed on 32bit CPU to avoid comparing
   // values as "int" vs "size_t" in CHECK_LE.
   CHECK_LE(static_cast<int64>(ptr - space_), kuint32max);
-  const uint32 offset = ptr - space_;
+  const uint32 offset = static_cast<uint32>(ptr - space_);
   node_offsets_[id] = offset;
   ptr += bytes;
 
-  const int num_output_edges = n->out_edges().size();
+  const size_t num_output_edges = n->out_edges().size();
   const int num_inputs = n->num_inputs();
   const int num_outputs = n->num_outputs();
 
@@ -580,9 +580,10 @@ void GraphView::Initialize(const Graph* g) {
   CHECK_EQ(ptr, space_ + total_bytes);
 }
 
-void GetMaxPendingCounts(const Node* n, int* max_pending, int* max_dead_count) {
+void GetMaxPendingCounts(const Node* n, size_t* max_pending,
-  const int num_in_edges = n->in_edges().size();
+                         size_t* max_dead_count) {
-  int initial_count;
+  const size_t num_in_edges = n->in_edges().size();
+  size_t initial_count;
   if (IsMerge(n)) {
     // merge waits all control inputs so we initialize the pending
     // count to be the number of control edges.
@@ -626,8 +627,7 @@ Status ExecutorImpl::Initialize() {
     FrameInfo* frame_info = EnsureFrameInfo(frame_name);
 
     // See if this node is a root node, and if so, add to root_nodes_.
-    const int num_in_edges = n->in_edges().size();
+    if (n->in_edges().empty()) {
-    if (num_in_edges == 0) {
       root_nodes_.push_back(n);
     }
 
@@ -659,7 +659,7 @@ Status ExecutorImpl::Initialize() {
     // pending counts data structure, and allocate a handle in
     // that frame's pending counts data structure that has enough
     // space to store these maximal count values.
-    int max_pending, max_dead;
+    size_t max_pending, max_dead;
     GetMaxPendingCounts(n, &max_pending, &max_dead);
     item->pending_id =
         frame_info->pending_counts_layout.CreateHandle(max_pending, max_dead);
@@ -896,7 +896,7 @@ class ExecutorState {
     Entry* input_tensors;
 
     // The number of outstanding ops for each iteration.
-    int outstanding_ops;
+    size_t outstanding_ops;
 
     // The number of outstanding frames for each iteration.
     int outstanding_frame_count;
@@ -1037,13 +1037,13 @@ class ExecutorState {
 
     inline IterationState* GetIteration(int64 iter)
         EXCLUSIVE_LOCKS_REQUIRED(mu) {
-      int index = iter % iterations.size();
+      size_t index = iter % iterations.size();
       return iterations[index];
     }
 
     inline void SetIteration(int64 iter, IterationState* state)
         EXCLUSIVE_LOCKS_REQUIRED(mu) {
-      int index = iter % iterations.size();
+      size_t index = iter % iterations.size();
       DCHECK(state == nullptr || iterations[index] == nullptr);
       iterations[index] = state;
     }
@@ -1404,7 +1404,7 @@ void ExecutorImpl::InitializePending(const Graph* graph,
   for (const Node* n : graph->nodes()) {
     const int id = n->id();
     const string& name = cf_info.frame_names[id];
-    int max_pending, max_dead;
+    size_t max_pending, max_dead;
     GetMaxPendingCounts(n, &max_pending, &max_dead);
     const NodeItem* item = gview_.node(id);
     PendingCounts* counts = EnsureFrameInfo(name)->pending_counts;
@@ -2027,7 +2027,7 @@ bool ExecutorState::NodeDone(const Status& s, const Node* node,
   }
 
   bool completed = false;
-  int ready_size = ready.size();
+  size_t ready_size = ready.size();
   if (ready_size == 0 || !s.ok()) {
     completed = (num_outstanding_ops_.fetch_sub(1) == 1);
   } else if (ready_size > 1) {
@@ -2375,10 +2375,10 @@ void ExecutorState::FrameState::ActivateNodes(const NodeItem* item,
                                               TaggedNodeSeq* ready) {
   const GraphView& gview = executor->gview_;
   IterationState* iter_state = GetIteration(iter);
-  const int num_output_edges = item->num_output_edges;
+  const size_t num_output_edges = item->num_output_edges;
   const EdgeInfo* edges = item->output_edge_list();
   Entry* input_tensors = iter_state->input_tensors;
-  for (int out_index = 0; out_index < num_output_edges; out_index++) {
+  for (size_t out_index = 0; out_index < num_output_edges; out_index++) {
     const EdgeInfo& e = edges[out_index];
     const int dst_id = e.dst_id;
     const NodeItem* dst_item = gview.node(dst_id);

tensorflow/core/common_runtime/executor.h

@@ -162,7 +162,7 @@ class ExecutorBarrier {
   //
   // 'done' is called after the last executor completes, and
   // ExecutorBarrier is deleted.
-  ExecutorBarrier(int num, Rendezvous* r, StatusCallback done)
+  ExecutorBarrier(size_t num, Rendezvous* r, StatusCallback done)
       : rendez_(r), done_cb_(done), pending_(num) {}
 
   ~ExecutorBarrier() {}

tensorflow/core/common_runtime/function.cc

@@ -274,8 +274,9 @@ class CallOp : public AsyncOpKernel {
                if (!status.ok()) {
                  ctx->SetStatus(status);
                } else {
-                 CHECK_EQ(rets->size(), ctx->num_outputs());
+                 const int ret_size = static_cast<int>(rets->size());
-                 for (size_t i = 0; i < rets->size(); ++i) {
+                 CHECK_EQ(ret_size, ctx->num_outputs());
+                 for (int i = 0; i < ret_size; ++i) {
                    ctx->set_output(i, (*rets)[i]);
                  }
                }
@@ -1000,7 +1001,7 @@ string NewName(const Node* n, bool pretty) {
 void ToGraphDef(const Graph* g, GraphDef* gdef, bool pretty) {
   // We visit nodes in forward topological sort order, which is a
   // possible execution order of the graph.
-  std::vector<int> pending(g->num_node_ids());
+  std::vector<size_t> pending(g->num_node_ids());
   std::deque<const Node*> ready;
   for (const Node* n : g->nodes()) {
     pending[n->id()] = n->in_edges().size();
@@ -1154,7 +1155,7 @@ FunctionBody* SymbolicGradientHelper::Compute() {
 
   Graph* g = gbody_->graph;
 
-  const int num_y = gbody_->ret_nodes.size();
+  const int num_y = static_cast<int>(gbody_->ret_nodes.size());
 
   // Populate 'y_node_outputs_' with node function body outputs.
   // Populate 'y_grad_nodes' with initial gradient nodes for each return node of
@@ -1169,7 +1170,7 @@ FunctionBody* SymbolicGradientHelper::Compute() {
     y_node_outputs.push_back({y, 0});
     DCHECK_EQ(y->type_string(), kRetOp);
     const DataType dtype = y->input_type(0);
-    const int index = gbody_->arg_nodes.size();
+    const int index = static_cast<int>(gbody_->arg_nodes.size());
     Node* dy = AddArg(g, dtype, index);
     gbody_->arg_types.push_back(dtype);
     gbody_->arg_nodes.push_back(dy);
@@ -1177,7 +1178,7 @@ FunctionBody* SymbolicGradientHelper::Compute() {
   }
 
   // Populate 'x_nodes' with function args (excluding 'y_grad_node_outputs').
-  const int num_x = fbody_->arg_nodes.size();
+  const size_t num_x = fbody_->arg_nodes.size();
   std::vector<NodeOut> x_node_outputs;
   x_node_outputs.reserve(num_x);
   for (size_t i = 0; i < fbody_->arg_nodes.size(); ++i) {
@@ -1200,7 +1201,8 @@ FunctionBody* SymbolicGradientHelper::Compute() {
   gbody_->ret_nodes.clear();
   // Add new return nodes to the function gradient body for each node
   // in 'x_grad_nodes'.
-  for (size_t i = 0; i < fbody_->arg_types.size(); ++i) {
+  const int arg_types_size = static_cast<int>(fbody_->arg_types.size());
+  for (int i = 0; i < arg_types_size; ++i) {
     Endpoint grad = {x_grad_node_outputs[i].node, x_grad_node_outputs[i].index};
     Node* ret = AddRet(g, grad, i);
     gbody_->ret_nodes.push_back(ret);

tensorflow/core/common_runtime/gpu/gpu_stream_util.cc

@@ -82,7 +82,7 @@ Status AssignStreams(const Graph* graph, const AssignStreamsOpts& opts,
     // Determine a suitable stream to use.
     int stream_id = highest_stream_id + 1;
     for (const Edge* e : n->in_edges()) {
-      const int fanout = e->src()->out_edges().size();
+      const size_t fanout = e->src()->out_edges().size();
       if (fanout == 1) {
         stream_id = (*node_to_stream_id)[e->src()->id()];
         break;

tensorflow/core/common_runtime/gpu/process_state.cc

@@ -191,7 +191,7 @@ Allocator* ProcessState::GetCUDAHostAllocator(int numa_node) {
   // example, process_state could maybe save the first stream executor
   // it knows is valid.
   gpu::StreamExecutor* se = nullptr;
-  for (size_t i = 0; i < gpu_allocators_.size(); ++i) {
+  for (int i = 0; i < static_cast<int>(gpu_allocators_.size()); ++i) {
     if (gpu_allocators_[i] != nullptr) {
       se = GPUMachineManager()->ExecutorForDevice(i).ValueOrDie();
       break;

tensorflow/core/common_runtime/pending_counts.h

@@ -69,7 +69,7 @@ class PendingCounts {
   // to retrieve the count data for this node.
   class Layout {
    public:
-    Handle CreateHandle(int max_pending_count, int max_dead_count);
+    Handle CreateHandle(size_t max_pending_count, size_t max_dead_count);
 
    private:
     friend class PendingCounts;
@@ -91,7 +91,7 @@ class PendingCounts {
 
   ~PendingCounts() { delete[] bytes_; }
 
-  void set_initial_count(Handle h, int pending_count) {
+  void set_initial_count(Handle h, size_t pending_count) {
     if (h.is_large_) {
       LargeCounts* c = Large(h);
       c->pending = pending_count;
@@ -306,7 +306,7 @@ class PendingCounts {
 };
 
 inline PendingCounts::Handle PendingCounts::Layout::CreateHandle(
-    int max_pending_count, int max_dead_count) {
+    size_t max_pending_count, size_t max_dead_count) {
   Handle result;
   if ((max_pending_count > kMaxCountForPackedCounts) ||
       (max_dead_count > kMaxCountForPackedCounts)) {

tensorflow/core/util/util.cc

@@ -85,7 +85,7 @@ void MovingAverage::AddValue(double v) {
 
 static char hex_char[] = "0123456789abcdef";
 
-string PrintMemory(const char* ptr, int n) {
+string PrintMemory(const char* ptr, size_t n) {
   string ret;
   ret.resize(n * 3);
   for (int i = 0; i < n; ++i) {

tensorflow/core/util/util.h

@@ -49,7 +49,7 @@ class MovingAverage {
 
 // Returns a string printing bytes in ptr[0..n).  The output looks
 // like "00 01 ef cd cd ef".
-string PrintMemory(const char* ptr, int n);
+string PrintMemory(const char* ptr, size_t n);
 
 // Given a flattened index into a tensor, computes a string s so that
 // StrAppend("tensor", s) is a Python indexing expression.  E.g.,