saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-24 19:43:35 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Change Compute function to return a status code instead of an integer. (#1139)

Browse source
This commit is contained in:
Ryan Hill 2019-06-04 08:34:32 -07:00 committed by Changming Sun
parent c18de6817b
commit 148141dd5f
6 changed files with 22 additions and 28 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
include/onnxruntime/core/framework/execution_provider.h
View file

@ -28,7 +28,7 @@ typedef std::map<int, AllocatorPtr> AllocatorMap;
// if we are export the fused function to dll, the function will still in the same binary as lotus
// use std function to give execution provider some chance to capture some state.
using CreateFunctionStateFunc = std::function<int(ComputeContext*, FunctionState*)>;
using ComputeFunc = std::function<int(FunctionState, const OrtCustomOpApi*, OrtKernelContext*)>;
using ComputeFunc = std::function<Status(FunctionState, const OrtCustomOpApi*, OrtKernelContext*)>;
using DestroyFunctionStateFunc = std::function<void(FunctionState)>;
struct NodeComputeInfo {

4
include/onnxruntime/core/framework/func_api.h
View file

@ -19,8 +19,8 @@ typedef struct {
using FunctionState = void*;
// take the ComputeContext, and create a function state.
using CreateFunctionStateC = int (*)(ComputeContext*, FunctionState*);
// pass in the function state and input/output tensors, perform compute and return status code, 0 - succeed.
using ComputeFuncC = int (*)(FunctionState, const OrtCustomOpApi*, OrtKernelContext*);
// pass in the function state and input/output tensors, perform compute and return status
using ComputeFuncC = Status (*)(FunctionState, const OrtCustomOpApi*, OrtKernelContext*);
// release the function state.
using DestroyFunctionStateC = void (*)(FunctionState);
} // namespace onnxruntime

5
onnxruntime/core/framework/func_kernel.h
View file

@ -40,10 +40,7 @@ class FunctionKernel : public OpKernel {
virtual Status Compute(OpKernelContext* context) const override {
auto* context_internal = static_cast<OpKernelContextInternal*>(context);
int ret = func_(func_state_, &GetCustomOpApi(), reinterpret_cast<OrtKernelContext*>(context_internal));
if (ret != 0)
return Status(common::ONNXRUNTIME, common::FAIL, "FuncKernel call failed with error code: " + std::to_string(ret));
return Status::OK();
return func_(func_state_, &GetCustomOpApi(), reinterpret_cast<OrtKernelContext*>(context_internal));
}
private:

5
onnxruntime/core/providers/ngraph/ngraph_execution_provider.cc
View file

@ -547,10 +547,7 @@ Status NGRAPHExecutionProvider::Compile(const std::vector<onnxruntime::Node*>& f
compute_info.compute_func = [](FunctionState state, const OrtCustomOpApi* api, OrtKernelContext* context) {
onnxruntime::ngraph_ep::NGRAPHCustomOp* ng_custom_op = reinterpret_cast<onnxruntime::ngraph_ep::NGRAPHCustomOp*>(state);
const Status compute_status = ng_custom_op->Compute(api, context);
return compute_status == Status::OK() ? 0 : 1;
return ng_custom_op->Compute(api, context);
};
node_compute_funcs.push_back(compute_info);

28
onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
View file

@ -23,12 +23,12 @@ using namespace ::onnxruntime::logging;
namespace onnxruntime {
#define CHECK_CUDA(call) \
do { \
cudaError_t status = call; \
if (status != cudaSuccess) { \
return -1; \
} \
#define CHECK_CUDA(call) \
do { \
cudaError_t status = call; \
if (status != cudaSuccess) { \
return common::Status(common::ONNXRUNTIME, common::FAIL); \
} \
} while (0)
TensorrtExecutionProvider::TensorrtExecutionProvider()
@ -57,7 +57,7 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
std::unique_ptr<IndexedSubGraph> sub_graph = std::make_unique<IndexedSubGraph>();
// Find inputs and outputs of the subgraph
std::unordered_map<const NodeArg *, int> fused_inputs, fused_outputs, fused_outputs_to_add;
std::unordered_map<const NodeArg*, int> fused_inputs, fused_outputs, fused_outputs_to_add;
std::unordered_set<const NodeArg*> erased;
int input_order = 0;
int output_order = 0;
@ -114,7 +114,7 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
fused_outputs.insert(fused_outputs_to_add.begin(), fused_outputs_to_add.end());
// Sort inputs and outputs by the order they were added
std::multimap<int, const NodeArg *> inputs, outputs;
std::multimap<int, const NodeArg*> inputs, outputs;
for (auto it = fused_inputs.begin(), end = fused_inputs.end(); it != end; ++it) {
inputs.insert(std::pair<int, const NodeArg*>(it->second, it->first));
}
@ -168,7 +168,7 @@ SubGraphCollection_t TensorrtExecutionProvider::GetSupportedList(SubGraphCollect
//Add node and node args
for (const auto& index : group.first) {
const auto& node = graph.GetNode(node_index[index]);
std::vector<onnxruntime::NodeArg *> inputs, outputs;
std::vector<onnxruntime::NodeArg*> inputs, outputs;
for (auto input : node->InputDefs()) {
auto& n_input = graph_build.GetOrCreateNodeArg(input->Name(), input->TypeAsProto());
inputs.push_back(&n_input);
@ -223,7 +223,7 @@ TensorrtExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph,
onnxruntime::Model model(graph.Name(), true, ModelMetaData(), IOnnxRuntimeOpSchemaRegistryList(), graph.DomainToVersionMap());
onnxruntime::Graph& graph_build = model.MainGraph();
for (const auto& node : graph.Nodes()) {
std::vector<onnxruntime::NodeArg *> inputs, outputs;
std::vector<onnxruntime::NodeArg*> inputs, outputs;
for (auto input : node.InputDefs()) {
auto& n_input = graph_build.GetOrCreateNodeArg(input->Name(), input->TypeAsProto());
inputs.push_back(&n_input);
@ -485,7 +485,7 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<onnxruntime:
CHECK_CUDA(cudaMalloc(&buffers[i], input_size * sizeof(int32_t)));
CHECK_CUDA(cudaMemcpy(buffers[i], input, input_size * sizeof(int32_t), cudaMemcpyHostToDevice));
} else {
return static_cast<int>(common::StatusCode::NOT_IMPLEMENTED);
return common::Status(common::ONNXRUNTIME, common::NOT_IMPLEMENTED);
}
}
@ -498,7 +498,7 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<onnxruntime:
} else if (output_types[i] == ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32 || output_types[i] == ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64) {
CHECK_CUDA(cudaMalloc(&buffers[i + num_binding_inputs], batch_size * output_dim_sizes[i] * sizeof(int32_t)));
} else {
return static_cast<int>(common::StatusCode::NOT_IMPLEMENTED);
return common::Status(common::ONNXRUNTIME, common::NOT_IMPLEMENTED);
}
}
@ -528,7 +528,7 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<onnxruntime:
}
delete[] output;
} else {
return static_cast<int>(common::StatusCode::NOT_IMPLEMENTED);
return common::Status(common::ONNXRUNTIME, common::NOT_IMPLEMENTED);
}
}
@ -542,7 +542,7 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<onnxruntime:
CHECK_CUDA(cudaFree(buffers[i]));
}
return 0;
return Status::OK();
};
node_compute_funcs.push_back(compute_info);

6
onnxruntime/test/tvm/tvm_basic_test.cc
View file

@ -270,12 +270,12 @@ class FuseExecutionProviderX : public CPUExecutionProvider {
try {
evaluate_func_.CallPacked(tvm_args, &rvalue);
} catch (std::exception ex) {
return -1;
return Status(common::ONNXRUNTIME, common::FAIL); // TODO: Translate exception to error code
}
if (rvalue.type_code() != kNull) {
return -1; // TODO: get error code.
return Status(common::ONNXRUNTIME, common::FAIL); // TODO: get error code.
} else {
return 0;
return Status::OK();
}
};
node_compute_funcs.push_back(compute_info);