experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

tf2xla directory resolutions

Browse Source
This commit is contained in:
Taré Gaskin 2020-07-26 22:06:02 +00:00
parent 3bb28df8d4
commit ad58928e65
10 changed files with 46 additions and 39 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
tensorflow/compiler/tf2xla/const_analysis.cc
View File

@ -74,8 +74,7 @@ Status CondConstInputIndices(
*(fbody->graph), &compile_time_const_arg_indices,
/*compile_time_const_nodes=*/nullptr, flib_runtime));
}
for (int i = 0, iter_limit = compile_time_const_arg_indices.size();
i < iter_limit; i++) {
for (int i = 0, end = compile_time_const_arg_indices.size(); i < end; i++) {
if (compile_time_const_arg_indices[i]) {
// The 0th input is the pred or branch index, which is not passed to the
// branches. So the i'th input of a branch function corresponds to the

12
tensorflow/compiler/tf2xla/functionalize_cond.cc
View File

@ -224,7 +224,8 @@ string DebugString(const CondArgNodes& nodes) {
}
StateMap::CondId StateMap::LookupCondId(const Node* node) const {
if (node->id() < node_to_condid_map_.size())
const int64 node_to_condid_map_size = node_to_condid_map_.size();
if (node->id() < node_to_condid_map_size)
return node_to_condid_map_[node->id()];
return added_node_condid_mapping_.at(node->id());
}
@ -235,14 +236,16 @@ StateMap::CondId StateMap::GetCondId(const StateMap::CondState& state) {
}
void StateMap::ResetCondId(const Node* node, StateMap::CondId id) {
if (node->id() < node_to_condid_map_.size())
const int64 node_to_condid_map_size = node_to_condid_map_.size();
if (node->id() < node_to_condid_map_size)
node_to_condid_map_[node->id()] = id;
else
added_node_condid_mapping_[node->id()] = id;
}
StateMap::AncestorId StateMap::LookupAncestorId(const Node* node) const {
if (node->id() < node_to_ancestorid_map_.size())
const int64 node_to_ancestorid_map_size = node_to_ancestorid_map_.size();
if (node->id() < node_to_ancestorid_map_size)
return node_to_ancestorid_map_[node->id()];
return added_node_ancestorid_mapping_.at(node->id());
}
@ -254,7 +257,8 @@ StateMap::AncestorId StateMap::GetAncestorId(
}
void StateMap::ResetAncestorId(const Node* node, StateMap::AncestorId id) {
if (node->id() < node_to_ancestorid_map_.size())
const int64 node_to_ancestorid_map_size = node_to_ancestorid_map_.size();
if (node->id() < node_to_ancestorid_map_size)
node_to_ancestorid_map_[node->id()] = id;
else
added_node_ancestorid_mapping_[node->id()] = id;

10
tensorflow/compiler/tf2xla/functionalize_while.cc
View File

@ -130,7 +130,7 @@ Status BuildLoopCondition(const Graph& graph, WhileLoopFrame* frame,
std::vector<bool> squash_src_outputs(graph.num_node_ids(), false);
// Build one _Arg node for each Enter node.
for (int i = 0; i < frame->args.size(); ++i) {
for (int i = 0, end = frame->args.size(); i < end; ++i) {
const WhileLoopArg& arg = frame->args[i];
TF_ASSIGN_OR_RETURN(Node * arg_node,
@ -170,7 +170,7 @@ Status BuildLoopBody(const Graph& graph, WhileLoopFrame* frame,
std::vector<Node*> next_iterations;
next_iterations.reserve(frame->args.size());
arg_types->reserve(frame->args.size());
for (int i = 0; i < frame->args.size(); ++i) {
for (int i = 0, end = frame->args.size(); i < end; ++i) {
const WhileLoopArg& arg = frame->args[i];
DataType dtype = arg.enter->input_type(0);
@ -235,7 +235,7 @@ Status FunctionalizeLoop(Graph* graph, WhileLoopFrame* frame,
} else {
std::vector<const Edge*> edges(arg.enter->out_edges().begin(),
arg.enter->out_edges().end());
for (int i = 0; i < edges.size(); ++i) {
for (int i = 0, end = edges.size(); i < end; ++i) {
if (edges[i]->IsControlEdge() && edges[i]->dst()->IsSink()) {
continue;
}
@ -447,7 +447,7 @@ Status FunctionalizeLoop(Graph* graph, WhileLoopFrame* frame,
}
}
std::vector<NodeDefBuilder::NodeOut> inputs;
for (int i = 0; i < frame->args.size(); ++i) {
for (int i = 0, end = frame->args.size(); i < end; ++i) {
const WhileLoopArg& arg = frame->args[i];
const Edge* in_edge;
TF_RETURN_IF_ERROR(arg.enter->input_edge(0, &in_edge));
@ -463,7 +463,7 @@ Status FunctionalizeLoop(Graph* graph, WhileLoopFrame* frame,
TF_ASSIGN_OR_RETURN(Node * while_node, AddNodeDefToGraph(while_def, graph));
// Copies edges to the Enter nodes and from the Exit nodes onto the While.
for (int i = 0; i < frame->args.size(); ++i) {
for (int i = 0, end = frame->args.size(); i < end; ++i) {
const WhileLoopArg& arg = frame->args[i];
const Edge* in_edge;
TF_RETURN_IF_ERROR(arg.enter->input_edge(0, &in_edge));

7
tensorflow/compiler/tf2xla/graph_compiler.cc
View File

@ -65,7 +65,7 @@ Status PrepareArguments(XlaOpKernelContext* ctx, Graph* graph,
/*compile_time_const_nodes=*/nullptr, ctx->function_library()));
args->resize(expressions.size());
for (int i = 0, iter_limit = args->size(); i < iter_limit; ++i) {
for (int i = 0, end = args->size(); i < end; ++i) {
XlaCompiler::Argument& arg = (*args)[i];
arg.type = ctx->input_type(i);
arg.shape = ctx->InputShape(i);
@ -269,7 +269,7 @@ Status GraphCompiler::CompileFunctionalNode(Node* n,
TF_RET_CHECK(arguments.size() == expressions.size());
std::vector<xla::XlaOp> handles;
for (int64 i = 0, iter_limit = expressions.size(); i < iter_limit; ++i) {
for (int64 i = 0, end = expressions.size(); i < end; ++i) {
if (arguments[i].kind == XlaCompiler::Argument::kConstant) {
continue;
}
@ -313,8 +313,7 @@ Status GraphCompiler::CompileFunctionalNode(Node* n,
}
}
for (int64 i = 0, iter_limit = result.resource_updates.size(); i < iter_limit;
i++) {
for (int64 i = 0, end = result.resource_updates.size(); i < end; i++) {
if (result.resource_updates[i].modified) {
XlaResource* resource =
expressions[result.resource_updates[i].input_index]->resource();

2
tensorflow/compiler/tf2xla/lib/data_format.cc
View File

@ -66,7 +66,7 @@ xla::StatusOr<xla::XlaOp> Expand(xla::XlaOp input, int64 dim) {
// Move the newly created dimension to the end with a transpose.
std::vector<int64> permutation;
for (int64 i = 0, iter_limit = expanded_shape.size(); i != iter_limit; ++i) {
for (int64 i = 0, end = expanded_shape.size(); i != end; ++i) {
permutation.push_back(i);
if (i == dim) {
++i;

2
tensorflow/compiler/tf2xla/literal_util.cc
View File

@ -72,7 +72,7 @@ Status HostTensorsToBorrowingLiteralTuple(absl::Span<const Tensor> host_tensors,
buf_ptrs.reserve(host_tensors.size());
std::vector<xla::Shape> tensor_shapes(host_tensors.size());
for (int i = 0, iter_limit = host_tensors.size(); i < iter_limit; i++) {
for (int i = 0, end = host_tensors.size(); i < end; i++) {
// Validate runtime shapes and fail if it doesn't match the contract.
const Tensor* tensor = &host_tensors[i];
buf_ptrs.emplace_back(static_cast<const char*>(DMAHelper::base(tensor)));

20
tensorflow/compiler/tf2xla/rearrange_function_argument.cc
View File

@ -41,7 +41,7 @@ std::vector<DataType> ShuffleInputDataTypeAttribute(
const std::vector<DataType>& in_types,
const std::vector<int>& index_mapping) {
std::vector<DataType> result(index_mapping.size());
for (int i = 0; i < in_types.size(); i++) {
for (int i = 0, end = in_types.size(); i < end; i++) {
result[index_mapping.at(i)] = in_types[i];
}
return result;
@ -56,7 +56,7 @@ Status InputTypesNeedsRearrange(const std::vector<DataType>& in_types,
bool* need_rewrite, int* resource_input_count,
std::vector<int>* index_mapping) {
int first_resource_index = -1;
for (int i = 0; i < in_types.size(); i++) {
for (int i = 0, end = in_types.size(); i < end; i++) {
DataType type = in_types[i];
if (type == DT_RESOURCE) {
first_resource_index = i;
@ -70,7 +70,7 @@ Status InputTypesNeedsRearrange(const std::vector<DataType>& in_types,
}
*need_rewrite = false;
for (int i = first_resource_index + 1; i < in_types.size(); i++) {
for (int i = first_resource_index + 1, end = in_types.size(); i < end; i++) {
if (in_types[i] != DT_RESOURCE) {
*need_rewrite = true;
break;
@ -81,7 +81,7 @@ Status InputTypesNeedsRearrange(const std::vector<DataType>& in_types,
}
*resource_input_count = 0;
for (int i = 0; i < in_types.size(); i++) {
for (int i = 0, end = in_types.size(); i < end; i++) {
DataType type = in_types[i];
if (type == DT_RESOURCE) {
++(*resource_input_count);
@ -90,7 +90,7 @@ Status InputTypesNeedsRearrange(const std::vector<DataType>& in_types,
int non_resource_index = 0,
resource_index = in_types.size() - *resource_input_count;
index_mapping->resize(in_types.size());
for (int i = 0; i < in_types.size(); i++) {
for (int i = 0, end = in_types.size(); i < end; i++) {
if (in_types[i] != DT_RESOURCE) {
(*index_mapping)[i] = non_resource_index;
non_resource_index++;
@ -146,7 +146,7 @@ Status ReorderOutputEdges(Graph* g, Node* n, int input_count,
int dst_input = e->dst_input();
g->RemoveEdge(e);
if (new_src_output < input_count - resource_input_count) {
if (new_src_output < static_cast<int64>(input_count - resource_input_count)) {
g->AddEdge(n, new_src_output, dst, dst_input);
} else {
const Edge* input_edge;
@ -180,7 +180,7 @@ Status CalculateRetvalRearrange(
const gtl::InlinedVector<Node*, 4>& ret_nodes, // non-absl ok
std::map<int, int>* retval_index_mapping,
std::map<int, int>* resource_retval_to_arg) {
for (int i = 0; i < ret_nodes.size(); i++) {
for (int i = 0, end = ret_nodes.size(); i < end; i++) {
Node* n = ret_nodes[i];
DataType t;
TF_RETURN_IF_ERROR(GetNodeAttr(n->def(), "T", &t));
@ -261,7 +261,7 @@ Status RearrangeOutputEdges(Node* n, Graph* g,
void RearrangeRetvalNodes(
const gtl::InlinedVector<Node*, 4>& ret_nodes, // non-absl ok
Graph* g, const std::map<int, int>& retval_index_mapping) {
for (int i = 0; i < ret_nodes.size(); i++) {
for (int i = 0, end = ret_nodes.size(); i < end; i++) {
Node* n = ret_nodes[i];
auto iter = retval_index_mapping.find(i);
if (iter == retval_index_mapping.end()) {
@ -317,7 +317,7 @@ Status MaybeRewriteWhileNode(
// lambda resource_var1, resource_var2: [resource_var2, resource_var1],
// [resource_var1, resource_var2])
if (attr_name == "body") {
for (int i = 0; i < fbody->ret_nodes.size(); i++) {
for (int i = 0, end = fbody->ret_nodes.size(); i < end; i++) {
Node* n = fbody->ret_nodes[i];
DataType dtype;
TF_RETURN_IF_ERROR(GetNodeAttr(n->def(), "T", &dtype));
@ -349,7 +349,7 @@ Status MaybeRewriteWhileNode(
RearrangeArgNodes(&fbody->arg_nodes, index_mapping);
if (attr_name == "body") {
for (int i = 0; i < fbody->ret_nodes.size(); i++) {
for (int i = 0, end = fbody->ret_nodes.size(); i < end; i++) {
Node* n = fbody->ret_nodes[i];
int new_index = index_mapping.at(i);
if (new_index < types.size() - resource_input_count) {

2
tensorflow/compiler/tf2xla/tf2xla.cc
View File

@ -87,7 +87,7 @@ Status ConvertGraphToXla(std::unique_ptr<Graph> graph,
*computation = std::move(*result.computation);
int num_const_results = 0;
for (int i = 0, iter_limit = result.outputs.size(); i < iter_limit; ++i) {
for (int i = 0, end = result.outputs.size(); i < end; ++i) {
// Ending up with const results (i.e. output args) is an error, since it
// means that one or more fetches that the user specified will be dropped
// from the generated function. It's most likely a configuration error,

4
tensorflow/compiler/tf2xla/tf2xla_util.cc
View File

@ -143,7 +143,7 @@ Status ReplaceArgUsageWithConstNode(
usages.push_back({e->dst()->id(), e->dst_input()});
}
for (int i = 0; i < usages.size(); i++) {
for (int i = 0, end = usages.size(); i < end; i++) {
// Make a copy of `usage_node`, and change its input to const node.
Node* usage_node = g->FindNodeId(usages[i].dst_node_id);
NodeDef replace_def = usage_node->def();
@ -158,7 +158,7 @@ Status ReplaceArgUsageWithConstNode(
// Later entries in `usages` might have `usage_node` as dst node, but
// `usage_node` is removed. Replace such entries with `replace_node`.
for (int j = i + 1; j < usages.size(); j++) {
for (int j = i + 1, end = usages.size(); j < end; j++) {
if (usages[j].dst_node_id == usages[i].dst_node_id) {
usages[j].dst_node_id = replace_node->id();
}

23
tensorflow/compiler/tf2xla/xla_compiler.cc
View File

@ -64,7 +64,7 @@ Status CheckSignature(const DataTypeVector& types,
return errors::Internal("Compilation arguments have ", args.size(),
" elements while function has ", types.size());
}
for (int i = 0, iter_limit = types.size(); i < iter_limit; ++i) {
for (int i = 0, end = types.size(); i < end; ++i) {
// Don't perform type checks on resource variables and tensor
// lists (DT_VARIANT) as we have to trick the type system in order to
// plumb them through. DT_VARIANTS are wrapped in a DT_UINT8 tensor.
@ -192,7 +192,7 @@ Status BuildComputation(
// replicate sharding is used. The first element is the output index, second
// element is the sharding.
std::unordered_map<int, xla::OpSharding> retval_index_and_sharding;
for (int i = 0, iter_limit = retvals.size(); i < iter_limit; ++i) {
for (int i = 0, end = retvals.size(); i < end; ++i) {
XlaCompiler::OutputDescription& output = (*outputs)[i];
const XlaExpression& retval = retvals[i];
output.type = retval.dtype();
@ -362,7 +362,7 @@ Status BuildComputation(
xla::Shape shape = xla::ShapeUtil::MakeTupleShape(elem_shapes);
// Copy specified sharding from retval_index_and_sharding.
std::vector<xla::HloSharding> sharding_elems;
for (int i = 0, iter_limit = elems.size(); i < iter_limit; i++) {
for (int i = 0, end = elems.size(); i < end; i++) {
const auto& iter = retval_index_and_sharding.find(i);
TF_RET_CHECK(iter != retval_index_and_sharding.end());
const xla::OpSharding& sub_op_sharding = iter->second;
@ -707,7 +707,7 @@ Status XlaCompiler::CompileFunction(
// Set shapes for _Arg nodes. They are useful for constant folding (e.g. an
// Xla op requires a compile-time constant input, and that input is shape of
// an _Arg node.
for (int i = 0, iter_limit = args.size(); i < iter_limit; i++) {
for (int i = 0, end = args.size(); i < end; i++) {
// Skip resource variables and tensor lists.
DataType dtype;
TF_RETURN_IF_ERROR(GetNodeAttr(fbody->arg_nodes[i]->def(), "T", &dtype));
@ -949,7 +949,7 @@ Status XlaCompiler::BuildArguments(
// to the d'th XLA input. Note that the value -1 corresponds to constants, or
// other args that don't correspond to an input.
std::vector<int> arg_to_inputs(args.size(), -1);
for (int i = 0, iter_limit = input_to_args->size(); i < iter_limit; i++) {
for (int i = 0, end = input_to_args->size(); i < end; i++) {
arg_to_inputs[input_to_args->at(i)] = i;
}
@ -995,7 +995,7 @@ Status XlaCompiler::BuildArguments(
: it->second;
}
std::vector<bool> is_same_across_replicas;
for (int i = 0, iter_limit = input_to_args->size(); i < iter_limit; ++i) {
for (int i = 0, end = input_to_args->size(); i < end; ++i) {
// Add an entry to is_same_across_replicas for every leaf buffer.
is_same_across_replicas.insert(
is_same_across_replicas.end(),
@ -1011,7 +1011,7 @@ Status XlaCompiler::BuildArguments(
tuple = xla::Parameter(builder, 0, (*input_shapes)[0], "arg_tuple");
}
for (int i = 0, iter_limit = input_to_args->size(); i < iter_limit; ++i) {
for (int i = 0, end = input_to_args->size(); i < end; ++i) {
const XlaCompiler::Argument& arg = args[input_to_args->at(i)];
for (const auto& dim_and_arg_num : arg.dynamic_dim_to_arg_num_map) {
int dynamic_size_param_index = arg_to_inputs.at(dim_and_arg_num.second);
@ -1030,6 +1030,11 @@ Status XlaCompiler::BuildArguments(
xla::XlaScopedShardingAssignment assign_sharding(
builder, it == arg_shardings.end() ? absl::optional<xla::OpSharding>()
: it->second);
auto& arg = args[input_to_args->at(i)];
xla::OpMetadata arg_metadata;
arg_metadata.set_op_name(arg.node_name);
builder->SetOneShotOpMetadata(arg_metadata);
arg_handles[i] = xla::GetTupleElement(tuple, i);
}
} else {
@ -1052,7 +1057,7 @@ Status XlaCompiler::BuildArguments(
}
}
for (int i = 0, iter_limit = input_to_args->size(); i < iter_limit; ++i) {
for (int i = 0, end = input_to_args->size(); i < end; ++i) {
const XlaCompiler::Argument& arg = args[input_to_args->at(i)];
for (const auto& dim_and_arg_num : arg.dynamic_dim_to_arg_num_map) {
int dynamic_size_param_index = arg_to_inputs.at(dim_and_arg_num.second);
@ -1373,7 +1378,7 @@ void SetTransfer(const string& key, absl::Span<const DataType> types,
tf2xla::HostTransferMetadata* transfer) {
transfer->set_key(key);
CHECK(types.size() == shapes.size());
for (int i = 0, iter_limit = types.size(); i < iter_limit; ++i) {
for (int i = 0, end = types.size(); i < end; ++i) {
tf2xla::TensorMetadata* metadata = transfer->add_metadata();
metadata->set_type(types[i]);
shapes[i].AsProto(metadata->mutable_shape());